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/slab.h>
11 #include <linux/blkdev.h>
12 #include <linux/elevator.h>
13 #include <linux/jiffies.h>
14 #include <linux/rbtree.h>
15 #include <linux/ioprio.h>
16 #include <linux/blktrace_api.h>
17 #include "blk-cgroup.h"
22 /* max queue in one round of service */
23 static const int cfq_quantum
= 8;
24 static const int cfq_fifo_expire
[2] = { HZ
/ 4, HZ
/ 8 };
25 /* maximum backwards seek, in KiB */
26 static const int cfq_back_max
= 16 * 1024;
27 /* penalty of a backwards seek */
28 static const int cfq_back_penalty
= 2;
29 static const int cfq_slice_sync
= HZ
/ 10;
30 static int cfq_slice_async
= HZ
/ 25;
31 static const int cfq_slice_async_rq
= 2;
32 static int cfq_slice_idle
= HZ
/ 125;
33 static const int cfq_target_latency
= HZ
* 3/10; /* 300 ms */
34 static const int cfq_hist_divisor
= 4;
37 * offset from end of service tree
39 #define CFQ_IDLE_DELAY (HZ / 5)
42 * below this threshold, we consider thinktime immediate
44 #define CFQ_MIN_TT (2)
46 #define CFQ_SLICE_SCALE (5)
47 #define CFQ_HW_QUEUE_MIN (5)
48 #define CFQ_SERVICE_SHIFT 12
50 #define CFQQ_SEEK_THR (sector_t)(8 * 100)
51 #define CFQQ_CLOSE_THR (sector_t)(8 * 1024)
52 #define CFQQ_SECT_THR_NONROT (sector_t)(2 * 32)
53 #define CFQQ_SEEKY(cfqq) (hweight32(cfqq->seek_history) > 32/8)
56 ((struct cfq_io_context *) (rq)->elevator_private)
57 #define RQ_CFQQ(rq) (struct cfq_queue *) ((rq)->elevator_private2)
58 #define RQ_CFQG(rq) (struct cfq_group *) ((rq)->elevator_private3)
60 static struct kmem_cache
*cfq_pool
;
61 static struct kmem_cache
*cfq_ioc_pool
;
63 static DEFINE_PER_CPU(unsigned long, cfq_ioc_count
);
64 static struct completion
*ioc_gone
;
65 static DEFINE_SPINLOCK(ioc_gone_lock
);
67 #define CFQ_PRIO_LISTS IOPRIO_BE_NR
68 #define cfq_class_idle(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
69 #define cfq_class_rt(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
71 #define sample_valid(samples) ((samples) > 80)
72 #define rb_entry_cfqg(node) rb_entry((node), struct cfq_group, rb_node)
75 * Most of our rbtree usage is for sorting with min extraction, so
76 * if we cache the leftmost node we don't have to walk down the tree
77 * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should
78 * move this into the elevator for the rq sorting as well.
84 unsigned total_weight
;
86 struct rb_node
*active
;
88 #define CFQ_RB_ROOT (struct cfq_rb_root) { .rb = RB_ROOT, .left = NULL, \
89 .count = 0, .min_vdisktime = 0, }
92 * Per process-grouping structure
97 /* various state flags, see below */
100 struct cfq_data
*cfqd
;
101 /* service_tree member */
102 struct rb_node rb_node
;
103 /* service_tree key */
104 unsigned long rb_key
;
105 /* prio tree member */
106 struct rb_node p_node
;
107 /* prio tree root we belong to, if any */
108 struct rb_root
*p_root
;
109 /* sorted list of pending requests */
110 struct rb_root sort_list
;
111 /* if fifo isn't expired, next request to serve */
112 struct request
*next_rq
;
113 /* requests queued in sort_list */
115 /* currently allocated requests */
117 /* fifo list of requests in sort_list */
118 struct list_head fifo
;
120 /* time when queue got scheduled in to dispatch first request. */
121 unsigned long dispatch_start
;
122 unsigned int allocated_slice
;
123 unsigned int slice_dispatch
;
124 /* time when first request from queue completed and slice started. */
125 unsigned long slice_start
;
126 unsigned long slice_end
;
129 /* pending metadata requests */
131 /* number of requests that are on the dispatch list or inside driver */
134 /* io prio of this group */
135 unsigned short ioprio
, org_ioprio
;
136 unsigned short ioprio_class
, org_ioprio_class
;
141 sector_t last_request_pos
;
143 struct cfq_rb_root
*service_tree
;
144 struct cfq_queue
*new_cfqq
;
145 struct cfq_group
*cfqg
;
146 struct cfq_group
*orig_cfqg
;
150 * First index in the service_trees.
151 * IDLE is handled separately, so it has negative index
160 * Second index in the service_trees.
164 SYNC_NOIDLE_WORKLOAD
= 1,
168 /* This is per cgroup per device grouping structure */
170 /* group service_tree member */
171 struct rb_node rb_node
;
173 /* group service_tree key */
178 /* number of cfqq currently on this group */
181 /* Per group busy queus average. Useful for workload slice calc. */
182 unsigned int busy_queues_avg
[2];
184 * rr lists of queues with requests, onle rr for each priority class.
185 * Counts are embedded in the cfq_rb_root
187 struct cfq_rb_root service_trees
[2][3];
188 struct cfq_rb_root service_tree_idle
;
190 unsigned long saved_workload_slice
;
191 enum wl_type_t saved_workload
;
192 enum wl_prio_t saved_serving_prio
;
193 struct blkio_group blkg
;
194 #ifdef CONFIG_CFQ_GROUP_IOSCHED
195 struct hlist_node cfqd_node
;
201 * Per block device queue structure
204 struct request_queue
*queue
;
205 /* Root service tree for cfq_groups */
206 struct cfq_rb_root grp_service_tree
;
207 struct cfq_group root_group
;
210 * The priority currently being served
212 enum wl_prio_t serving_prio
;
213 enum wl_type_t serving_type
;
214 unsigned long workload_expires
;
215 struct cfq_group
*serving_group
;
216 bool noidle_tree_requires_idle
;
219 * Each priority tree is sorted by next_request position. These
220 * trees are used when determining if two or more queues are
221 * interleaving requests (see cfq_close_cooperator).
223 struct rb_root prio_trees
[CFQ_PRIO_LISTS
];
225 unsigned int busy_queues
;
231 * queue-depth detection
237 * -1 => indeterminate, (cfq will behave as if NCQ is present, to allow better detection)
238 * 1 => NCQ is present (hw_tag_est_depth is the estimated max depth)
241 int hw_tag_est_depth
;
242 unsigned int hw_tag_samples
;
245 * idle window management
247 struct timer_list idle_slice_timer
;
248 struct work_struct unplug_work
;
250 struct cfq_queue
*active_queue
;
251 struct cfq_io_context
*active_cic
;
254 * async queue for each priority case
256 struct cfq_queue
*async_cfqq
[2][IOPRIO_BE_NR
];
257 struct cfq_queue
*async_idle_cfqq
;
259 sector_t last_position
;
262 * tunables, see top of file
264 unsigned int cfq_quantum
;
265 unsigned int cfq_fifo_expire
[2];
266 unsigned int cfq_back_penalty
;
267 unsigned int cfq_back_max
;
268 unsigned int cfq_slice
[2];
269 unsigned int cfq_slice_async_rq
;
270 unsigned int cfq_slice_idle
;
271 unsigned int cfq_latency
;
272 unsigned int cfq_group_isolation
;
274 struct list_head cic_list
;
277 * Fallback dummy cfqq for extreme OOM conditions
279 struct cfq_queue oom_cfqq
;
281 unsigned long last_delayed_sync
;
283 /* List of cfq groups being managed on this device*/
284 struct hlist_head cfqg_list
;
288 static struct cfq_group
*cfq_get_next_cfqg(struct cfq_data
*cfqd
);
290 static struct cfq_rb_root
*service_tree_for(struct cfq_group
*cfqg
,
297 if (prio
== IDLE_WORKLOAD
)
298 return &cfqg
->service_tree_idle
;
300 return &cfqg
->service_trees
[prio
][type
];
303 enum cfqq_state_flags
{
304 CFQ_CFQQ_FLAG_on_rr
= 0, /* on round-robin busy list */
305 CFQ_CFQQ_FLAG_wait_request
, /* waiting for a request */
306 CFQ_CFQQ_FLAG_must_dispatch
, /* must be allowed a dispatch */
307 CFQ_CFQQ_FLAG_must_alloc_slice
, /* per-slice must_alloc flag */
308 CFQ_CFQQ_FLAG_fifo_expire
, /* FIFO checked in this slice */
309 CFQ_CFQQ_FLAG_idle_window
, /* slice idling enabled */
310 CFQ_CFQQ_FLAG_prio_changed
, /* task priority has changed */
311 CFQ_CFQQ_FLAG_slice_new
, /* no requests dispatched in slice */
312 CFQ_CFQQ_FLAG_sync
, /* synchronous queue */
313 CFQ_CFQQ_FLAG_coop
, /* cfqq is shared */
314 CFQ_CFQQ_FLAG_split_coop
, /* shared cfqq will be splitted */
315 CFQ_CFQQ_FLAG_deep
, /* sync cfqq experienced large depth */
316 CFQ_CFQQ_FLAG_wait_busy
, /* Waiting for next request */
319 #define CFQ_CFQQ_FNS(name) \
320 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \
322 (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name); \
324 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \
326 (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \
328 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \
330 return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \
334 CFQ_CFQQ_FNS(wait_request
);
335 CFQ_CFQQ_FNS(must_dispatch
);
336 CFQ_CFQQ_FNS(must_alloc_slice
);
337 CFQ_CFQQ_FNS(fifo_expire
);
338 CFQ_CFQQ_FNS(idle_window
);
339 CFQ_CFQQ_FNS(prio_changed
);
340 CFQ_CFQQ_FNS(slice_new
);
343 CFQ_CFQQ_FNS(split_coop
);
345 CFQ_CFQQ_FNS(wait_busy
);
348 #ifdef CONFIG_CFQ_GROUP_IOSCHED
349 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
350 blk_add_trace_msg((cfqd)->queue, "cfq%d%c %s " fmt, (cfqq)->pid, \
351 cfq_cfqq_sync((cfqq)) ? 'S' : 'A', \
352 blkg_path(&(cfqq)->cfqg->blkg), ##args);
354 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...) \
355 blk_add_trace_msg((cfqd)->queue, "%s " fmt, \
356 blkg_path(&(cfqg)->blkg), ##args); \
359 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
360 blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args)
361 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...) do {} while (0);
363 #define cfq_log(cfqd, fmt, args...) \
364 blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
366 /* Traverses through cfq group service trees */
367 #define for_each_cfqg_st(cfqg, i, j, st) \
368 for (i = 0; i <= IDLE_WORKLOAD; i++) \
369 for (j = 0, st = i < IDLE_WORKLOAD ? &cfqg->service_trees[i][j]\
370 : &cfqg->service_tree_idle; \
371 (i < IDLE_WORKLOAD && j <= SYNC_WORKLOAD) || \
372 (i == IDLE_WORKLOAD && j == 0); \
373 j++, st = i < IDLE_WORKLOAD ? \
374 &cfqg->service_trees[i][j]: NULL) \
377 static inline enum wl_prio_t cfqq_prio(struct cfq_queue *cfqq)
379 if (cfq_class_idle(cfqq
))
380 return IDLE_WORKLOAD
;
381 if (cfq_class_rt(cfqq
))
387 static enum wl_type_t
cfqq_type(struct cfq_queue
*cfqq
)
389 if (!cfq_cfqq_sync(cfqq
))
390 return ASYNC_WORKLOAD
;
391 if (!cfq_cfqq_idle_window(cfqq
))
392 return SYNC_NOIDLE_WORKLOAD
;
393 return SYNC_WORKLOAD
;
396 static inline int cfq_group_busy_queues_wl(enum wl_prio_t wl
,
397 struct cfq_data
*cfqd
,
398 struct cfq_group
*cfqg
)
400 if (wl
== IDLE_WORKLOAD
)
401 return cfqg
->service_tree_idle
.count
;
403 return cfqg
->service_trees
[wl
][ASYNC_WORKLOAD
].count
404 + cfqg
->service_trees
[wl
][SYNC_NOIDLE_WORKLOAD
].count
405 + cfqg
->service_trees
[wl
][SYNC_WORKLOAD
].count
;
408 static inline int cfqg_busy_async_queues(struct cfq_data
*cfqd
,
409 struct cfq_group
*cfqg
)
411 return cfqg
->service_trees
[RT_WORKLOAD
][ASYNC_WORKLOAD
].count
412 + cfqg
->service_trees
[BE_WORKLOAD
][ASYNC_WORKLOAD
].count
;
415 static void cfq_dispatch_insert(struct request_queue
*, struct request
*);
416 static struct cfq_queue
*cfq_get_queue(struct cfq_data
*, bool,
417 struct io_context
*, gfp_t
);
418 static struct cfq_io_context
*cfq_cic_lookup(struct cfq_data
*,
419 struct io_context
*);
421 static inline struct cfq_queue
*cic_to_cfqq(struct cfq_io_context
*cic
,
424 return cic
->cfqq
[is_sync
];
427 static inline void cic_set_cfqq(struct cfq_io_context
*cic
,
428 struct cfq_queue
*cfqq
, bool is_sync
)
430 cic
->cfqq
[is_sync
] = cfqq
;
433 #define CIC_DEAD_KEY 1ul
435 static inline void *cfqd_dead_key(struct cfq_data
*cfqd
)
437 return (void *)((unsigned long) cfqd
| CIC_DEAD_KEY
);
440 static inline struct cfq_data
*cic_to_cfqd(struct cfq_io_context
*cic
)
442 struct cfq_data
*cfqd
= cic
->key
;
444 if (unlikely((unsigned long) cfqd
& CIC_DEAD_KEY
))
451 * We regard a request as SYNC, if it's either a read or has the SYNC bit
452 * set (in which case it could also be direct WRITE).
454 static inline bool cfq_bio_sync(struct bio
*bio
)
456 return bio_data_dir(bio
) == READ
|| bio_rw_flagged(bio
, BIO_RW_SYNCIO
);
460 * scheduler run of queue, if there are requests pending and no one in the
461 * driver that will restart queueing
463 static inline void cfq_schedule_dispatch(struct cfq_data
*cfqd
)
465 if (cfqd
->busy_queues
) {
466 cfq_log(cfqd
, "schedule dispatch");
467 kblockd_schedule_work(cfqd
->queue
, &cfqd
->unplug_work
);
471 static int cfq_queue_empty(struct request_queue
*q
)
473 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
475 return !cfqd
->rq_queued
;
479 * Scale schedule slice based on io priority. Use the sync time slice only
480 * if a queue is marked sync and has sync io queued. A sync queue with async
481 * io only, should not get full sync slice length.
483 static inline int cfq_prio_slice(struct cfq_data
*cfqd
, bool sync
,
486 const int base_slice
= cfqd
->cfq_slice
[sync
];
488 WARN_ON(prio
>= IOPRIO_BE_NR
);
490 return base_slice
+ (base_slice
/CFQ_SLICE_SCALE
* (4 - prio
));
494 cfq_prio_to_slice(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
496 return cfq_prio_slice(cfqd
, cfq_cfqq_sync(cfqq
), cfqq
->ioprio
);
499 static inline u64
cfq_scale_slice(unsigned long delta
, struct cfq_group
*cfqg
)
501 u64 d
= delta
<< CFQ_SERVICE_SHIFT
;
503 d
= d
* BLKIO_WEIGHT_DEFAULT
;
504 do_div(d
, cfqg
->weight
);
508 static inline u64
max_vdisktime(u64 min_vdisktime
, u64 vdisktime
)
510 s64 delta
= (s64
)(vdisktime
- min_vdisktime
);
512 min_vdisktime
= vdisktime
;
514 return min_vdisktime
;
517 static inline u64
min_vdisktime(u64 min_vdisktime
, u64 vdisktime
)
519 s64 delta
= (s64
)(vdisktime
- min_vdisktime
);
521 min_vdisktime
= vdisktime
;
523 return min_vdisktime
;
526 static void update_min_vdisktime(struct cfq_rb_root
*st
)
528 u64 vdisktime
= st
->min_vdisktime
;
529 struct cfq_group
*cfqg
;
532 cfqg
= rb_entry_cfqg(st
->active
);
533 vdisktime
= cfqg
->vdisktime
;
537 cfqg
= rb_entry_cfqg(st
->left
);
538 vdisktime
= min_vdisktime(vdisktime
, cfqg
->vdisktime
);
541 st
->min_vdisktime
= max_vdisktime(st
->min_vdisktime
, vdisktime
);
545 * get averaged number of queues of RT/BE priority.
546 * average is updated, with a formula that gives more weight to higher numbers,
547 * to quickly follows sudden increases and decrease slowly
550 static inline unsigned cfq_group_get_avg_queues(struct cfq_data
*cfqd
,
551 struct cfq_group
*cfqg
, bool rt
)
553 unsigned min_q
, max_q
;
554 unsigned mult
= cfq_hist_divisor
- 1;
555 unsigned round
= cfq_hist_divisor
/ 2;
556 unsigned busy
= cfq_group_busy_queues_wl(rt
, cfqd
, cfqg
);
558 min_q
= min(cfqg
->busy_queues_avg
[rt
], busy
);
559 max_q
= max(cfqg
->busy_queues_avg
[rt
], busy
);
560 cfqg
->busy_queues_avg
[rt
] = (mult
* max_q
+ min_q
+ round
) /
562 return cfqg
->busy_queues_avg
[rt
];
565 static inline unsigned
566 cfq_group_slice(struct cfq_data
*cfqd
, struct cfq_group
*cfqg
)
568 struct cfq_rb_root
*st
= &cfqd
->grp_service_tree
;
570 return cfq_target_latency
* cfqg
->weight
/ st
->total_weight
;
574 cfq_set_prio_slice(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
576 unsigned slice
= cfq_prio_to_slice(cfqd
, cfqq
);
577 if (cfqd
->cfq_latency
) {
579 * interested queues (we consider only the ones with the same
580 * priority class in the cfq group)
582 unsigned iq
= cfq_group_get_avg_queues(cfqd
, cfqq
->cfqg
,
584 unsigned sync_slice
= cfqd
->cfq_slice
[1];
585 unsigned expect_latency
= sync_slice
* iq
;
586 unsigned group_slice
= cfq_group_slice(cfqd
, cfqq
->cfqg
);
588 if (expect_latency
> group_slice
) {
589 unsigned base_low_slice
= 2 * cfqd
->cfq_slice_idle
;
590 /* scale low_slice according to IO priority
591 * and sync vs async */
593 min(slice
, base_low_slice
* slice
/ sync_slice
);
594 /* the adapted slice value is scaled to fit all iqs
595 * into the target latency */
596 slice
= max(slice
* group_slice
/ expect_latency
,
600 cfqq
->slice_start
= jiffies
;
601 cfqq
->slice_end
= jiffies
+ slice
;
602 cfqq
->allocated_slice
= slice
;
603 cfq_log_cfqq(cfqd
, cfqq
, "set_slice=%lu", cfqq
->slice_end
- jiffies
);
607 * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
608 * isn't valid until the first request from the dispatch is activated
609 * and the slice time set.
611 static inline bool cfq_slice_used(struct cfq_queue
*cfqq
)
613 if (cfq_cfqq_slice_new(cfqq
))
615 if (time_before(jiffies
, cfqq
->slice_end
))
622 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
623 * We choose the request that is closest to the head right now. Distance
624 * behind the head is penalized and only allowed to a certain extent.
626 static struct request
*
627 cfq_choose_req(struct cfq_data
*cfqd
, struct request
*rq1
, struct request
*rq2
, sector_t last
)
629 sector_t s1
, s2
, d1
= 0, d2
= 0;
630 unsigned long back_max
;
631 #define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */
632 #define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */
633 unsigned wrap
= 0; /* bit mask: requests behind the disk head? */
635 if (rq1
== NULL
|| rq1
== rq2
)
640 if (rq_is_sync(rq1
) && !rq_is_sync(rq2
))
642 else if (rq_is_sync(rq2
) && !rq_is_sync(rq1
))
644 if (rq_is_meta(rq1
) && !rq_is_meta(rq2
))
646 else if (rq_is_meta(rq2
) && !rq_is_meta(rq1
))
649 s1
= blk_rq_pos(rq1
);
650 s2
= blk_rq_pos(rq2
);
653 * by definition, 1KiB is 2 sectors
655 back_max
= cfqd
->cfq_back_max
* 2;
658 * Strict one way elevator _except_ in the case where we allow
659 * short backward seeks which are biased as twice the cost of a
660 * similar forward seek.
664 else if (s1
+ back_max
>= last
)
665 d1
= (last
- s1
) * cfqd
->cfq_back_penalty
;
667 wrap
|= CFQ_RQ1_WRAP
;
671 else if (s2
+ back_max
>= last
)
672 d2
= (last
- s2
) * cfqd
->cfq_back_penalty
;
674 wrap
|= CFQ_RQ2_WRAP
;
676 /* Found required data */
679 * By doing switch() on the bit mask "wrap" we avoid having to
680 * check two variables for all permutations: --> faster!
683 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
699 case (CFQ_RQ1_WRAP
|CFQ_RQ2_WRAP
): /* both rqs wrapped */
702 * Since both rqs are wrapped,
703 * start with the one that's further behind head
704 * (--> only *one* back seek required),
705 * since back seek takes more time than forward.
715 * The below is leftmost cache rbtree addon
717 static struct cfq_queue
*cfq_rb_first(struct cfq_rb_root
*root
)
719 /* Service tree is empty */
724 root
->left
= rb_first(&root
->rb
);
727 return rb_entry(root
->left
, struct cfq_queue
, rb_node
);
732 static struct cfq_group
*cfq_rb_first_group(struct cfq_rb_root
*root
)
735 root
->left
= rb_first(&root
->rb
);
738 return rb_entry_cfqg(root
->left
);
743 static void rb_erase_init(struct rb_node
*n
, struct rb_root
*root
)
749 static void cfq_rb_erase(struct rb_node
*n
, struct cfq_rb_root
*root
)
753 rb_erase_init(n
, &root
->rb
);
758 * would be nice to take fifo expire time into account as well
760 static struct request
*
761 cfq_find_next_rq(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
762 struct request
*last
)
764 struct rb_node
*rbnext
= rb_next(&last
->rb_node
);
765 struct rb_node
*rbprev
= rb_prev(&last
->rb_node
);
766 struct request
*next
= NULL
, *prev
= NULL
;
768 BUG_ON(RB_EMPTY_NODE(&last
->rb_node
));
771 prev
= rb_entry_rq(rbprev
);
774 next
= rb_entry_rq(rbnext
);
776 rbnext
= rb_first(&cfqq
->sort_list
);
777 if (rbnext
&& rbnext
!= &last
->rb_node
)
778 next
= rb_entry_rq(rbnext
);
781 return cfq_choose_req(cfqd
, next
, prev
, blk_rq_pos(last
));
784 static unsigned long cfq_slice_offset(struct cfq_data
*cfqd
,
785 struct cfq_queue
*cfqq
)
788 * just an approximation, should be ok.
790 return (cfqq
->cfqg
->nr_cfqq
- 1) * (cfq_prio_slice(cfqd
, 1, 0) -
791 cfq_prio_slice(cfqd
, cfq_cfqq_sync(cfqq
), cfqq
->ioprio
));
795 cfqg_key(struct cfq_rb_root
*st
, struct cfq_group
*cfqg
)
797 return cfqg
->vdisktime
- st
->min_vdisktime
;
801 __cfq_group_service_tree_add(struct cfq_rb_root
*st
, struct cfq_group
*cfqg
)
803 struct rb_node
**node
= &st
->rb
.rb_node
;
804 struct rb_node
*parent
= NULL
;
805 struct cfq_group
*__cfqg
;
806 s64 key
= cfqg_key(st
, cfqg
);
809 while (*node
!= NULL
) {
811 __cfqg
= rb_entry_cfqg(parent
);
813 if (key
< cfqg_key(st
, __cfqg
))
814 node
= &parent
->rb_left
;
816 node
= &parent
->rb_right
;
822 st
->left
= &cfqg
->rb_node
;
824 rb_link_node(&cfqg
->rb_node
, parent
, node
);
825 rb_insert_color(&cfqg
->rb_node
, &st
->rb
);
829 cfq_group_service_tree_add(struct cfq_data
*cfqd
, struct cfq_group
*cfqg
)
831 struct cfq_rb_root
*st
= &cfqd
->grp_service_tree
;
832 struct cfq_group
*__cfqg
;
840 * Currently put the group at the end. Later implement something
841 * so that groups get lesser vtime based on their weights, so that
842 * if group does not loose all if it was not continously backlogged.
844 n
= rb_last(&st
->rb
);
846 __cfqg
= rb_entry_cfqg(n
);
847 cfqg
->vdisktime
= __cfqg
->vdisktime
+ CFQ_IDLE_DELAY
;
849 cfqg
->vdisktime
= st
->min_vdisktime
;
851 __cfq_group_service_tree_add(st
, cfqg
);
853 st
->total_weight
+= cfqg
->weight
;
857 cfq_group_service_tree_del(struct cfq_data
*cfqd
, struct cfq_group
*cfqg
)
859 struct cfq_rb_root
*st
= &cfqd
->grp_service_tree
;
861 if (st
->active
== &cfqg
->rb_node
)
864 BUG_ON(cfqg
->nr_cfqq
< 1);
867 /* If there are other cfq queues under this group, don't delete it */
871 cfq_log_cfqg(cfqd
, cfqg
, "del_from_rr group");
873 st
->total_weight
-= cfqg
->weight
;
874 if (!RB_EMPTY_NODE(&cfqg
->rb_node
))
875 cfq_rb_erase(&cfqg
->rb_node
, st
);
876 cfqg
->saved_workload_slice
= 0;
877 blkiocg_update_dequeue_stats(&cfqg
->blkg
, 1);
880 static inline unsigned int cfq_cfqq_slice_usage(struct cfq_queue
*cfqq
)
882 unsigned int slice_used
;
885 * Queue got expired before even a single request completed or
886 * got expired immediately after first request completion.
888 if (!cfqq
->slice_start
|| cfqq
->slice_start
== jiffies
) {
890 * Also charge the seek time incurred to the group, otherwise
891 * if there are mutiple queues in the group, each can dispatch
892 * a single request on seeky media and cause lots of seek time
893 * and group will never know it.
895 slice_used
= max_t(unsigned, (jiffies
- cfqq
->dispatch_start
),
898 slice_used
= jiffies
- cfqq
->slice_start
;
899 if (slice_used
> cfqq
->allocated_slice
)
900 slice_used
= cfqq
->allocated_slice
;
903 cfq_log_cfqq(cfqq
->cfqd
, cfqq
, "sl_used=%u", slice_used
);
907 static void cfq_group_served(struct cfq_data
*cfqd
, struct cfq_group
*cfqg
,
908 struct cfq_queue
*cfqq
)
910 struct cfq_rb_root
*st
= &cfqd
->grp_service_tree
;
911 unsigned int used_sl
, charge_sl
;
912 int nr_sync
= cfqg
->nr_cfqq
- cfqg_busy_async_queues(cfqd
, cfqg
)
913 - cfqg
->service_tree_idle
.count
;
916 used_sl
= charge_sl
= cfq_cfqq_slice_usage(cfqq
);
918 if (!cfq_cfqq_sync(cfqq
) && !nr_sync
)
919 charge_sl
= cfqq
->allocated_slice
;
921 /* Can't update vdisktime while group is on service tree */
922 cfq_rb_erase(&cfqg
->rb_node
, st
);
923 cfqg
->vdisktime
+= cfq_scale_slice(charge_sl
, cfqg
);
924 __cfq_group_service_tree_add(st
, cfqg
);
926 /* This group is being expired. Save the context */
927 if (time_after(cfqd
->workload_expires
, jiffies
)) {
928 cfqg
->saved_workload_slice
= cfqd
->workload_expires
930 cfqg
->saved_workload
= cfqd
->serving_type
;
931 cfqg
->saved_serving_prio
= cfqd
->serving_prio
;
933 cfqg
->saved_workload_slice
= 0;
935 cfq_log_cfqg(cfqd
, cfqg
, "served: vt=%llu min_vt=%llu", cfqg
->vdisktime
,
937 blkiocg_update_timeslice_used(&cfqg
->blkg
, used_sl
);
938 blkiocg_set_start_empty_time(&cfqg
->blkg
);
941 #ifdef CONFIG_CFQ_GROUP_IOSCHED
942 static inline struct cfq_group
*cfqg_of_blkg(struct blkio_group
*blkg
)
945 return container_of(blkg
, struct cfq_group
, blkg
);
950 cfq_update_blkio_group_weight(struct blkio_group
*blkg
, unsigned int weight
)
952 cfqg_of_blkg(blkg
)->weight
= weight
;
955 static struct cfq_group
*
956 cfq_find_alloc_cfqg(struct cfq_data
*cfqd
, struct cgroup
*cgroup
, int create
)
958 struct blkio_cgroup
*blkcg
= cgroup_to_blkio_cgroup(cgroup
);
959 struct cfq_group
*cfqg
= NULL
;
962 struct cfq_rb_root
*st
;
963 struct backing_dev_info
*bdi
= &cfqd
->queue
->backing_dev_info
;
964 unsigned int major
, minor
;
966 cfqg
= cfqg_of_blkg(blkiocg_lookup_group(blkcg
, key
));
967 if (cfqg
&& !cfqg
->blkg
.dev
&& bdi
->dev
&& dev_name(bdi
->dev
)) {
968 sscanf(dev_name(bdi
->dev
), "%u:%u", &major
, &minor
);
969 cfqg
->blkg
.dev
= MKDEV(major
, minor
);
975 cfqg
= kzalloc_node(sizeof(*cfqg
), GFP_ATOMIC
, cfqd
->queue
->node
);
979 for_each_cfqg_st(cfqg
, i
, j
, st
)
981 RB_CLEAR_NODE(&cfqg
->rb_node
);
984 * Take the initial reference that will be released on destroy
985 * This can be thought of a joint reference by cgroup and
986 * elevator which will be dropped by either elevator exit
987 * or cgroup deletion path depending on who is exiting first.
989 atomic_set(&cfqg
->ref
, 1);
991 /* Add group onto cgroup list */
992 sscanf(dev_name(bdi
->dev
), "%u:%u", &major
, &minor
);
993 blkiocg_add_blkio_group(blkcg
, &cfqg
->blkg
, (void *)cfqd
,
994 MKDEV(major
, minor
));
995 cfqg
->weight
= blkcg_get_weight(blkcg
, cfqg
->blkg
.dev
);
997 /* Add group on cfqd list */
998 hlist_add_head(&cfqg
->cfqd_node
, &cfqd
->cfqg_list
);
1005 * Search for the cfq group current task belongs to. If create = 1, then also
1006 * create the cfq group if it does not exist. request_queue lock must be held.
1008 static struct cfq_group
*cfq_get_cfqg(struct cfq_data
*cfqd
, int create
)
1010 struct cgroup
*cgroup
;
1011 struct cfq_group
*cfqg
= NULL
;
1014 cgroup
= task_cgroup(current
, blkio_subsys_id
);
1015 cfqg
= cfq_find_alloc_cfqg(cfqd
, cgroup
, create
);
1016 if (!cfqg
&& create
)
1017 cfqg
= &cfqd
->root_group
;
1022 static inline struct cfq_group
*cfq_ref_get_cfqg(struct cfq_group
*cfqg
)
1024 atomic_inc(&cfqg
->ref
);
1028 static void cfq_link_cfqq_cfqg(struct cfq_queue
*cfqq
, struct cfq_group
*cfqg
)
1030 /* Currently, all async queues are mapped to root group */
1031 if (!cfq_cfqq_sync(cfqq
))
1032 cfqg
= &cfqq
->cfqd
->root_group
;
1035 /* cfqq reference on cfqg */
1036 atomic_inc(&cfqq
->cfqg
->ref
);
1039 static void cfq_put_cfqg(struct cfq_group
*cfqg
)
1041 struct cfq_rb_root
*st
;
1044 BUG_ON(atomic_read(&cfqg
->ref
) <= 0);
1045 if (!atomic_dec_and_test(&cfqg
->ref
))
1047 for_each_cfqg_st(cfqg
, i
, j
, st
)
1048 BUG_ON(!RB_EMPTY_ROOT(&st
->rb
) || st
->active
!= NULL
);
1052 static void cfq_destroy_cfqg(struct cfq_data
*cfqd
, struct cfq_group
*cfqg
)
1054 /* Something wrong if we are trying to remove same group twice */
1055 BUG_ON(hlist_unhashed(&cfqg
->cfqd_node
));
1057 hlist_del_init(&cfqg
->cfqd_node
);
1060 * Put the reference taken at the time of creation so that when all
1061 * queues are gone, group can be destroyed.
1066 static void cfq_release_cfq_groups(struct cfq_data
*cfqd
)
1068 struct hlist_node
*pos
, *n
;
1069 struct cfq_group
*cfqg
;
1071 hlist_for_each_entry_safe(cfqg
, pos
, n
, &cfqd
->cfqg_list
, cfqd_node
) {
1073 * If cgroup removal path got to blk_group first and removed
1074 * it from cgroup list, then it will take care of destroying
1077 if (!blkiocg_del_blkio_group(&cfqg
->blkg
))
1078 cfq_destroy_cfqg(cfqd
, cfqg
);
1083 * Blk cgroup controller notification saying that blkio_group object is being
1084 * delinked as associated cgroup object is going away. That also means that
1085 * no new IO will come in this group. So get rid of this group as soon as
1086 * any pending IO in the group is finished.
1088 * This function is called under rcu_read_lock(). key is the rcu protected
1089 * pointer. That means "key" is a valid cfq_data pointer as long as we are rcu
1092 * "key" was fetched from blkio_group under blkio_cgroup->lock. That means
1093 * it should not be NULL as even if elevator was exiting, cgroup deltion
1094 * path got to it first.
1096 void cfq_unlink_blkio_group(void *key
, struct blkio_group
*blkg
)
1098 unsigned long flags
;
1099 struct cfq_data
*cfqd
= key
;
1101 spin_lock_irqsave(cfqd
->queue
->queue_lock
, flags
);
1102 cfq_destroy_cfqg(cfqd
, cfqg_of_blkg(blkg
));
1103 spin_unlock_irqrestore(cfqd
->queue
->queue_lock
, flags
);
1106 #else /* GROUP_IOSCHED */
1107 static struct cfq_group
*cfq_get_cfqg(struct cfq_data
*cfqd
, int create
)
1109 return &cfqd
->root_group
;
1112 static inline struct cfq_group
*cfq_ref_get_cfqg(struct cfq_group
*cfqg
)
1118 cfq_link_cfqq_cfqg(struct cfq_queue
*cfqq
, struct cfq_group
*cfqg
) {
1122 static void cfq_release_cfq_groups(struct cfq_data
*cfqd
) {}
1123 static inline void cfq_put_cfqg(struct cfq_group
*cfqg
) {}
1125 #endif /* GROUP_IOSCHED */
1128 * The cfqd->service_trees holds all pending cfq_queue's that have
1129 * requests waiting to be processed. It is sorted in the order that
1130 * we will service the queues.
1132 static void cfq_service_tree_add(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
1135 struct rb_node
**p
, *parent
;
1136 struct cfq_queue
*__cfqq
;
1137 unsigned long rb_key
;
1138 struct cfq_rb_root
*service_tree
;
1141 int group_changed
= 0;
1143 #ifdef CONFIG_CFQ_GROUP_IOSCHED
1144 if (!cfqd
->cfq_group_isolation
1145 && cfqq_type(cfqq
) == SYNC_NOIDLE_WORKLOAD
1146 && cfqq
->cfqg
&& cfqq
->cfqg
!= &cfqd
->root_group
) {
1147 /* Move this cfq to root group */
1148 cfq_log_cfqq(cfqd
, cfqq
, "moving to root group");
1149 if (!RB_EMPTY_NODE(&cfqq
->rb_node
))
1150 cfq_group_service_tree_del(cfqd
, cfqq
->cfqg
);
1151 cfqq
->orig_cfqg
= cfqq
->cfqg
;
1152 cfqq
->cfqg
= &cfqd
->root_group
;
1153 atomic_inc(&cfqd
->root_group
.ref
);
1155 } else if (!cfqd
->cfq_group_isolation
1156 && cfqq_type(cfqq
) == SYNC_WORKLOAD
&& cfqq
->orig_cfqg
) {
1157 /* cfqq is sequential now needs to go to its original group */
1158 BUG_ON(cfqq
->cfqg
!= &cfqd
->root_group
);
1159 if (!RB_EMPTY_NODE(&cfqq
->rb_node
))
1160 cfq_group_service_tree_del(cfqd
, cfqq
->cfqg
);
1161 cfq_put_cfqg(cfqq
->cfqg
);
1162 cfqq
->cfqg
= cfqq
->orig_cfqg
;
1163 cfqq
->orig_cfqg
= NULL
;
1165 cfq_log_cfqq(cfqd
, cfqq
, "moved to origin group");
1169 service_tree
= service_tree_for(cfqq
->cfqg
, cfqq_prio(cfqq
),
1171 if (cfq_class_idle(cfqq
)) {
1172 rb_key
= CFQ_IDLE_DELAY
;
1173 parent
= rb_last(&service_tree
->rb
);
1174 if (parent
&& parent
!= &cfqq
->rb_node
) {
1175 __cfqq
= rb_entry(parent
, struct cfq_queue
, rb_node
);
1176 rb_key
+= __cfqq
->rb_key
;
1179 } else if (!add_front
) {
1181 * Get our rb key offset. Subtract any residual slice
1182 * value carried from last service. A negative resid
1183 * count indicates slice overrun, and this should position
1184 * the next service time further away in the tree.
1186 rb_key
= cfq_slice_offset(cfqd
, cfqq
) + jiffies
;
1187 rb_key
-= cfqq
->slice_resid
;
1188 cfqq
->slice_resid
= 0;
1191 __cfqq
= cfq_rb_first(service_tree
);
1192 rb_key
+= __cfqq
? __cfqq
->rb_key
: jiffies
;
1195 if (!RB_EMPTY_NODE(&cfqq
->rb_node
)) {
1198 * same position, nothing more to do
1200 if (rb_key
== cfqq
->rb_key
&&
1201 cfqq
->service_tree
== service_tree
)
1204 cfq_rb_erase(&cfqq
->rb_node
, cfqq
->service_tree
);
1205 cfqq
->service_tree
= NULL
;
1210 cfqq
->service_tree
= service_tree
;
1211 p
= &service_tree
->rb
.rb_node
;
1216 __cfqq
= rb_entry(parent
, struct cfq_queue
, rb_node
);
1219 * sort by key, that represents service time.
1221 if (time_before(rb_key
, __cfqq
->rb_key
))
1224 n
= &(*p
)->rb_right
;
1232 service_tree
->left
= &cfqq
->rb_node
;
1234 cfqq
->rb_key
= rb_key
;
1235 rb_link_node(&cfqq
->rb_node
, parent
, p
);
1236 rb_insert_color(&cfqq
->rb_node
, &service_tree
->rb
);
1237 service_tree
->count
++;
1238 if ((add_front
|| !new_cfqq
) && !group_changed
)
1240 cfq_group_service_tree_add(cfqd
, cfqq
->cfqg
);
1243 static struct cfq_queue
*
1244 cfq_prio_tree_lookup(struct cfq_data
*cfqd
, struct rb_root
*root
,
1245 sector_t sector
, struct rb_node
**ret_parent
,
1246 struct rb_node
***rb_link
)
1248 struct rb_node
**p
, *parent
;
1249 struct cfq_queue
*cfqq
= NULL
;
1257 cfqq
= rb_entry(parent
, struct cfq_queue
, p_node
);
1260 * Sort strictly based on sector. Smallest to the left,
1261 * largest to the right.
1263 if (sector
> blk_rq_pos(cfqq
->next_rq
))
1264 n
= &(*p
)->rb_right
;
1265 else if (sector
< blk_rq_pos(cfqq
->next_rq
))
1273 *ret_parent
= parent
;
1279 static void cfq_prio_tree_add(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
1281 struct rb_node
**p
, *parent
;
1282 struct cfq_queue
*__cfqq
;
1285 rb_erase(&cfqq
->p_node
, cfqq
->p_root
);
1286 cfqq
->p_root
= NULL
;
1289 if (cfq_class_idle(cfqq
))
1294 cfqq
->p_root
= &cfqd
->prio_trees
[cfqq
->org_ioprio
];
1295 __cfqq
= cfq_prio_tree_lookup(cfqd
, cfqq
->p_root
,
1296 blk_rq_pos(cfqq
->next_rq
), &parent
, &p
);
1298 rb_link_node(&cfqq
->p_node
, parent
, p
);
1299 rb_insert_color(&cfqq
->p_node
, cfqq
->p_root
);
1301 cfqq
->p_root
= NULL
;
1305 * Update cfqq's position in the service tree.
1307 static void cfq_resort_rr_list(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
1310 * Resorting requires the cfqq to be on the RR list already.
1312 if (cfq_cfqq_on_rr(cfqq
)) {
1313 cfq_service_tree_add(cfqd
, cfqq
, 0);
1314 cfq_prio_tree_add(cfqd
, cfqq
);
1319 * add to busy list of queues for service, trying to be fair in ordering
1320 * the pending list according to last request service
1322 static void cfq_add_cfqq_rr(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
1324 cfq_log_cfqq(cfqd
, cfqq
, "add_to_rr");
1325 BUG_ON(cfq_cfqq_on_rr(cfqq
));
1326 cfq_mark_cfqq_on_rr(cfqq
);
1327 cfqd
->busy_queues
++;
1329 cfq_resort_rr_list(cfqd
, cfqq
);
1333 * Called when the cfqq no longer has requests pending, remove it from
1336 static void cfq_del_cfqq_rr(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
1338 cfq_log_cfqq(cfqd
, cfqq
, "del_from_rr");
1339 BUG_ON(!cfq_cfqq_on_rr(cfqq
));
1340 cfq_clear_cfqq_on_rr(cfqq
);
1342 if (!RB_EMPTY_NODE(&cfqq
->rb_node
)) {
1343 cfq_rb_erase(&cfqq
->rb_node
, cfqq
->service_tree
);
1344 cfqq
->service_tree
= NULL
;
1347 rb_erase(&cfqq
->p_node
, cfqq
->p_root
);
1348 cfqq
->p_root
= NULL
;
1351 cfq_group_service_tree_del(cfqd
, cfqq
->cfqg
);
1352 BUG_ON(!cfqd
->busy_queues
);
1353 cfqd
->busy_queues
--;
1357 * rb tree support functions
1359 static void cfq_del_rq_rb(struct request
*rq
)
1361 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
1362 const int sync
= rq_is_sync(rq
);
1364 BUG_ON(!cfqq
->queued
[sync
]);
1365 cfqq
->queued
[sync
]--;
1367 elv_rb_del(&cfqq
->sort_list
, rq
);
1369 if (cfq_cfqq_on_rr(cfqq
) && RB_EMPTY_ROOT(&cfqq
->sort_list
)) {
1371 * Queue will be deleted from service tree when we actually
1372 * expire it later. Right now just remove it from prio tree
1376 rb_erase(&cfqq
->p_node
, cfqq
->p_root
);
1377 cfqq
->p_root
= NULL
;
1382 static void cfq_add_rq_rb(struct request
*rq
)
1384 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
1385 struct cfq_data
*cfqd
= cfqq
->cfqd
;
1386 struct request
*__alias
, *prev
;
1388 cfqq
->queued
[rq_is_sync(rq
)]++;
1391 * looks a little odd, but the first insert might return an alias.
1392 * if that happens, put the alias on the dispatch list
1394 while ((__alias
= elv_rb_add(&cfqq
->sort_list
, rq
)) != NULL
)
1395 cfq_dispatch_insert(cfqd
->queue
, __alias
);
1397 if (!cfq_cfqq_on_rr(cfqq
))
1398 cfq_add_cfqq_rr(cfqd
, cfqq
);
1401 * check if this request is a better next-serve candidate
1403 prev
= cfqq
->next_rq
;
1404 cfqq
->next_rq
= cfq_choose_req(cfqd
, cfqq
->next_rq
, rq
, cfqd
->last_position
);
1407 * adjust priority tree position, if ->next_rq changes
1409 if (prev
!= cfqq
->next_rq
)
1410 cfq_prio_tree_add(cfqd
, cfqq
);
1412 BUG_ON(!cfqq
->next_rq
);
1415 static void cfq_reposition_rq_rb(struct cfq_queue
*cfqq
, struct request
*rq
)
1417 elv_rb_del(&cfqq
->sort_list
, rq
);
1418 cfqq
->queued
[rq_is_sync(rq
)]--;
1419 blkiocg_update_io_remove_stats(&(RQ_CFQG(rq
))->blkg
, rq_data_dir(rq
),
1422 blkiocg_update_io_add_stats(&(RQ_CFQG(rq
))->blkg
,
1423 &cfqq
->cfqd
->serving_group
->blkg
, rq_data_dir(rq
),
1427 static struct request
*
1428 cfq_find_rq_fmerge(struct cfq_data
*cfqd
, struct bio
*bio
)
1430 struct task_struct
*tsk
= current
;
1431 struct cfq_io_context
*cic
;
1432 struct cfq_queue
*cfqq
;
1434 cic
= cfq_cic_lookup(cfqd
, tsk
->io_context
);
1438 cfqq
= cic_to_cfqq(cic
, cfq_bio_sync(bio
));
1440 sector_t sector
= bio
->bi_sector
+ bio_sectors(bio
);
1442 return elv_rb_find(&cfqq
->sort_list
, sector
);
1448 static void cfq_activate_request(struct request_queue
*q
, struct request
*rq
)
1450 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
1452 cfqd
->rq_in_driver
++;
1453 cfq_log_cfqq(cfqd
, RQ_CFQQ(rq
), "activate rq, drv=%d",
1454 cfqd
->rq_in_driver
);
1456 cfqd
->last_position
= blk_rq_pos(rq
) + blk_rq_sectors(rq
);
1459 static void cfq_deactivate_request(struct request_queue
*q
, struct request
*rq
)
1461 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
1463 WARN_ON(!cfqd
->rq_in_driver
);
1464 cfqd
->rq_in_driver
--;
1465 cfq_log_cfqq(cfqd
, RQ_CFQQ(rq
), "deactivate rq, drv=%d",
1466 cfqd
->rq_in_driver
);
1469 static void cfq_remove_request(struct request
*rq
)
1471 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
1473 if (cfqq
->next_rq
== rq
)
1474 cfqq
->next_rq
= cfq_find_next_rq(cfqq
->cfqd
, cfqq
, rq
);
1476 list_del_init(&rq
->queuelist
);
1479 cfqq
->cfqd
->rq_queued
--;
1480 blkiocg_update_io_remove_stats(&(RQ_CFQG(rq
))->blkg
, rq_data_dir(rq
),
1482 if (rq_is_meta(rq
)) {
1483 WARN_ON(!cfqq
->meta_pending
);
1484 cfqq
->meta_pending
--;
1488 static int cfq_merge(struct request_queue
*q
, struct request
**req
,
1491 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
1492 struct request
*__rq
;
1494 __rq
= cfq_find_rq_fmerge(cfqd
, bio
);
1495 if (__rq
&& elv_rq_merge_ok(__rq
, bio
)) {
1497 return ELEVATOR_FRONT_MERGE
;
1500 return ELEVATOR_NO_MERGE
;
1503 static void cfq_merged_request(struct request_queue
*q
, struct request
*req
,
1506 if (type
== ELEVATOR_FRONT_MERGE
) {
1507 struct cfq_queue
*cfqq
= RQ_CFQQ(req
);
1509 cfq_reposition_rq_rb(cfqq
, req
);
1513 static void cfq_bio_merged(struct request_queue
*q
, struct request
*req
,
1516 blkiocg_update_io_merged_stats(&(RQ_CFQG(req
))->blkg
, bio_data_dir(bio
),
1521 cfq_merged_requests(struct request_queue
*q
, struct request
*rq
,
1522 struct request
*next
)
1524 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
1526 * reposition in fifo if next is older than rq
1528 if (!list_empty(&rq
->queuelist
) && !list_empty(&next
->queuelist
) &&
1529 time_before(rq_fifo_time(next
), rq_fifo_time(rq
))) {
1530 list_move(&rq
->queuelist
, &next
->queuelist
);
1531 rq_set_fifo_time(rq
, rq_fifo_time(next
));
1534 if (cfqq
->next_rq
== next
)
1536 cfq_remove_request(next
);
1537 blkiocg_update_io_merged_stats(&(RQ_CFQG(rq
))->blkg
, rq_data_dir(next
),
1541 static int cfq_allow_merge(struct request_queue
*q
, struct request
*rq
,
1544 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
1545 struct cfq_io_context
*cic
;
1546 struct cfq_queue
*cfqq
;
1549 * Disallow merge of a sync bio into an async request.
1551 if (cfq_bio_sync(bio
) && !rq_is_sync(rq
))
1555 * Lookup the cfqq that this bio will be queued with. Allow
1556 * merge only if rq is queued there.
1558 cic
= cfq_cic_lookup(cfqd
, current
->io_context
);
1562 cfqq
= cic_to_cfqq(cic
, cfq_bio_sync(bio
));
1563 return cfqq
== RQ_CFQQ(rq
);
1566 static inline void cfq_del_timer(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
1568 del_timer(&cfqd
->idle_slice_timer
);
1569 blkiocg_update_idle_time_stats(&cfqq
->cfqg
->blkg
);
1572 static void __cfq_set_active_queue(struct cfq_data
*cfqd
,
1573 struct cfq_queue
*cfqq
)
1576 cfq_log_cfqq(cfqd
, cfqq
, "set_active wl_prio:%d wl_type:%d",
1577 cfqd
->serving_prio
, cfqd
->serving_type
);
1578 blkiocg_update_avg_queue_size_stats(&cfqq
->cfqg
->blkg
);
1579 cfqq
->slice_start
= 0;
1580 cfqq
->dispatch_start
= jiffies
;
1581 cfqq
->allocated_slice
= 0;
1582 cfqq
->slice_end
= 0;
1583 cfqq
->slice_dispatch
= 0;
1585 cfq_clear_cfqq_wait_request(cfqq
);
1586 cfq_clear_cfqq_must_dispatch(cfqq
);
1587 cfq_clear_cfqq_must_alloc_slice(cfqq
);
1588 cfq_clear_cfqq_fifo_expire(cfqq
);
1589 cfq_mark_cfqq_slice_new(cfqq
);
1591 cfq_del_timer(cfqd
, cfqq
);
1594 cfqd
->active_queue
= cfqq
;
1598 * current cfqq expired its slice (or was too idle), select new one
1601 __cfq_slice_expired(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
1604 cfq_log_cfqq(cfqd
, cfqq
, "slice expired t=%d", timed_out
);
1606 if (cfq_cfqq_wait_request(cfqq
))
1607 cfq_del_timer(cfqd
, cfqq
);
1609 cfq_clear_cfqq_wait_request(cfqq
);
1610 cfq_clear_cfqq_wait_busy(cfqq
);
1613 * If this cfqq is shared between multiple processes, check to
1614 * make sure that those processes are still issuing I/Os within
1615 * the mean seek distance. If not, it may be time to break the
1616 * queues apart again.
1618 if (cfq_cfqq_coop(cfqq
) && CFQQ_SEEKY(cfqq
))
1619 cfq_mark_cfqq_split_coop(cfqq
);
1622 * store what was left of this slice, if the queue idled/timed out
1624 if (timed_out
&& !cfq_cfqq_slice_new(cfqq
)) {
1625 cfqq
->slice_resid
= cfqq
->slice_end
- jiffies
;
1626 cfq_log_cfqq(cfqd
, cfqq
, "resid=%ld", cfqq
->slice_resid
);
1629 cfq_group_served(cfqd
, cfqq
->cfqg
, cfqq
);
1631 if (cfq_cfqq_on_rr(cfqq
) && RB_EMPTY_ROOT(&cfqq
->sort_list
))
1632 cfq_del_cfqq_rr(cfqd
, cfqq
);
1634 cfq_resort_rr_list(cfqd
, cfqq
);
1636 if (cfqq
== cfqd
->active_queue
)
1637 cfqd
->active_queue
= NULL
;
1639 if (&cfqq
->cfqg
->rb_node
== cfqd
->grp_service_tree
.active
)
1640 cfqd
->grp_service_tree
.active
= NULL
;
1642 if (cfqd
->active_cic
) {
1643 put_io_context(cfqd
->active_cic
->ioc
);
1644 cfqd
->active_cic
= NULL
;
1648 static inline void cfq_slice_expired(struct cfq_data
*cfqd
, bool timed_out
)
1650 struct cfq_queue
*cfqq
= cfqd
->active_queue
;
1653 __cfq_slice_expired(cfqd
, cfqq
, timed_out
);
1657 * Get next queue for service. Unless we have a queue preemption,
1658 * we'll simply select the first cfqq in the service tree.
1660 static struct cfq_queue
*cfq_get_next_queue(struct cfq_data
*cfqd
)
1662 struct cfq_rb_root
*service_tree
=
1663 service_tree_for(cfqd
->serving_group
, cfqd
->serving_prio
,
1664 cfqd
->serving_type
);
1666 if (!cfqd
->rq_queued
)
1669 /* There is nothing to dispatch */
1672 if (RB_EMPTY_ROOT(&service_tree
->rb
))
1674 return cfq_rb_first(service_tree
);
1677 static struct cfq_queue
*cfq_get_next_queue_forced(struct cfq_data
*cfqd
)
1679 struct cfq_group
*cfqg
;
1680 struct cfq_queue
*cfqq
;
1682 struct cfq_rb_root
*st
;
1684 if (!cfqd
->rq_queued
)
1687 cfqg
= cfq_get_next_cfqg(cfqd
);
1691 for_each_cfqg_st(cfqg
, i
, j
, st
)
1692 if ((cfqq
= cfq_rb_first(st
)) != NULL
)
1698 * Get and set a new active queue for service.
1700 static struct cfq_queue
*cfq_set_active_queue(struct cfq_data
*cfqd
,
1701 struct cfq_queue
*cfqq
)
1704 cfqq
= cfq_get_next_queue(cfqd
);
1706 __cfq_set_active_queue(cfqd
, cfqq
);
1710 static inline sector_t
cfq_dist_from_last(struct cfq_data
*cfqd
,
1713 if (blk_rq_pos(rq
) >= cfqd
->last_position
)
1714 return blk_rq_pos(rq
) - cfqd
->last_position
;
1716 return cfqd
->last_position
- blk_rq_pos(rq
);
1719 static inline int cfq_rq_close(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
1722 return cfq_dist_from_last(cfqd
, rq
) <= CFQQ_CLOSE_THR
;
1725 static struct cfq_queue
*cfqq_close(struct cfq_data
*cfqd
,
1726 struct cfq_queue
*cur_cfqq
)
1728 struct rb_root
*root
= &cfqd
->prio_trees
[cur_cfqq
->org_ioprio
];
1729 struct rb_node
*parent
, *node
;
1730 struct cfq_queue
*__cfqq
;
1731 sector_t sector
= cfqd
->last_position
;
1733 if (RB_EMPTY_ROOT(root
))
1737 * First, if we find a request starting at the end of the last
1738 * request, choose it.
1740 __cfqq
= cfq_prio_tree_lookup(cfqd
, root
, sector
, &parent
, NULL
);
1745 * If the exact sector wasn't found, the parent of the NULL leaf
1746 * will contain the closest sector.
1748 __cfqq
= rb_entry(parent
, struct cfq_queue
, p_node
);
1749 if (cfq_rq_close(cfqd
, cur_cfqq
, __cfqq
->next_rq
))
1752 if (blk_rq_pos(__cfqq
->next_rq
) < sector
)
1753 node
= rb_next(&__cfqq
->p_node
);
1755 node
= rb_prev(&__cfqq
->p_node
);
1759 __cfqq
= rb_entry(node
, struct cfq_queue
, p_node
);
1760 if (cfq_rq_close(cfqd
, cur_cfqq
, __cfqq
->next_rq
))
1768 * cur_cfqq - passed in so that we don't decide that the current queue is
1769 * closely cooperating with itself.
1771 * So, basically we're assuming that that cur_cfqq has dispatched at least
1772 * one request, and that cfqd->last_position reflects a position on the disk
1773 * associated with the I/O issued by cur_cfqq. I'm not sure this is a valid
1776 static struct cfq_queue
*cfq_close_cooperator(struct cfq_data
*cfqd
,
1777 struct cfq_queue
*cur_cfqq
)
1779 struct cfq_queue
*cfqq
;
1781 if (cfq_class_idle(cur_cfqq
))
1783 if (!cfq_cfqq_sync(cur_cfqq
))
1785 if (CFQQ_SEEKY(cur_cfqq
))
1789 * Don't search priority tree if it's the only queue in the group.
1791 if (cur_cfqq
->cfqg
->nr_cfqq
== 1)
1795 * We should notice if some of the queues are cooperating, eg
1796 * working closely on the same area of the disk. In that case,
1797 * we can group them together and don't waste time idling.
1799 cfqq
= cfqq_close(cfqd
, cur_cfqq
);
1803 /* If new queue belongs to different cfq_group, don't choose it */
1804 if (cur_cfqq
->cfqg
!= cfqq
->cfqg
)
1808 * It only makes sense to merge sync queues.
1810 if (!cfq_cfqq_sync(cfqq
))
1812 if (CFQQ_SEEKY(cfqq
))
1816 * Do not merge queues of different priority classes
1818 if (cfq_class_rt(cfqq
) != cfq_class_rt(cur_cfqq
))
1825 * Determine whether we should enforce idle window for this queue.
1828 static bool cfq_should_idle(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
1830 enum wl_prio_t prio
= cfqq_prio(cfqq
);
1831 struct cfq_rb_root
*service_tree
= cfqq
->service_tree
;
1833 BUG_ON(!service_tree
);
1834 BUG_ON(!service_tree
->count
);
1836 /* We never do for idle class queues. */
1837 if (prio
== IDLE_WORKLOAD
)
1840 /* We do for queues that were marked with idle window flag. */
1841 if (cfq_cfqq_idle_window(cfqq
) &&
1842 !(blk_queue_nonrot(cfqd
->queue
) && cfqd
->hw_tag
))
1846 * Otherwise, we do only if they are the last ones
1847 * in their service tree.
1849 if (service_tree
->count
== 1 && cfq_cfqq_sync(cfqq
))
1851 cfq_log_cfqq(cfqd
, cfqq
, "Not idling. st->count:%d",
1852 service_tree
->count
);
1856 static void cfq_arm_slice_timer(struct cfq_data
*cfqd
)
1858 struct cfq_queue
*cfqq
= cfqd
->active_queue
;
1859 struct cfq_io_context
*cic
;
1863 * SSD device without seek penalty, disable idling. But only do so
1864 * for devices that support queuing, otherwise we still have a problem
1865 * with sync vs async workloads.
1867 if (blk_queue_nonrot(cfqd
->queue
) && cfqd
->hw_tag
)
1870 WARN_ON(!RB_EMPTY_ROOT(&cfqq
->sort_list
));
1871 WARN_ON(cfq_cfqq_slice_new(cfqq
));
1874 * idle is disabled, either manually or by past process history
1876 if (!cfqd
->cfq_slice_idle
|| !cfq_should_idle(cfqd
, cfqq
))
1880 * still active requests from this queue, don't idle
1882 if (cfqq
->dispatched
)
1886 * task has exited, don't wait
1888 cic
= cfqd
->active_cic
;
1889 if (!cic
|| !atomic_read(&cic
->ioc
->nr_tasks
))
1893 * If our average think time is larger than the remaining time
1894 * slice, then don't idle. This avoids overrunning the allotted
1897 if (sample_valid(cic
->ttime_samples
) &&
1898 (cfqq
->slice_end
- jiffies
< cic
->ttime_mean
)) {
1899 cfq_log_cfqq(cfqd
, cfqq
, "Not idling. think_time:%d",
1904 cfq_mark_cfqq_wait_request(cfqq
);
1906 sl
= cfqd
->cfq_slice_idle
;
1908 mod_timer(&cfqd
->idle_slice_timer
, jiffies
+ sl
);
1909 blkiocg_update_set_idle_time_stats(&cfqq
->cfqg
->blkg
);
1910 cfq_log_cfqq(cfqd
, cfqq
, "arm_idle: %lu", sl
);
1914 * Move request from internal lists to the request queue dispatch list.
1916 static void cfq_dispatch_insert(struct request_queue
*q
, struct request
*rq
)
1918 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
1919 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
1921 cfq_log_cfqq(cfqd
, cfqq
, "dispatch_insert");
1923 cfqq
->next_rq
= cfq_find_next_rq(cfqd
, cfqq
, rq
);
1924 cfq_remove_request(rq
);
1926 elv_dispatch_sort(q
, rq
);
1928 cfqd
->rq_in_flight
[cfq_cfqq_sync(cfqq
)]++;
1929 blkiocg_update_dispatch_stats(&cfqq
->cfqg
->blkg
, blk_rq_bytes(rq
),
1930 rq_data_dir(rq
), rq_is_sync(rq
));
1934 * return expired entry, or NULL to just start from scratch in rbtree
1936 static struct request
*cfq_check_fifo(struct cfq_queue
*cfqq
)
1938 struct request
*rq
= NULL
;
1940 if (cfq_cfqq_fifo_expire(cfqq
))
1943 cfq_mark_cfqq_fifo_expire(cfqq
);
1945 if (list_empty(&cfqq
->fifo
))
1948 rq
= rq_entry_fifo(cfqq
->fifo
.next
);
1949 if (time_before(jiffies
, rq_fifo_time(rq
)))
1952 cfq_log_cfqq(cfqq
->cfqd
, cfqq
, "fifo=%p", rq
);
1957 cfq_prio_to_maxrq(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
1959 const int base_rq
= cfqd
->cfq_slice_async_rq
;
1961 WARN_ON(cfqq
->ioprio
>= IOPRIO_BE_NR
);
1963 return 2 * (base_rq
+ base_rq
* (CFQ_PRIO_LISTS
- 1 - cfqq
->ioprio
));
1967 * Must be called with the queue_lock held.
1969 static int cfqq_process_refs(struct cfq_queue
*cfqq
)
1971 int process_refs
, io_refs
;
1973 io_refs
= cfqq
->allocated
[READ
] + cfqq
->allocated
[WRITE
];
1974 process_refs
= atomic_read(&cfqq
->ref
) - io_refs
;
1975 BUG_ON(process_refs
< 0);
1976 return process_refs
;
1979 static void cfq_setup_merge(struct cfq_queue
*cfqq
, struct cfq_queue
*new_cfqq
)
1981 int process_refs
, new_process_refs
;
1982 struct cfq_queue
*__cfqq
;
1984 /* Avoid a circular list and skip interim queue merges */
1985 while ((__cfqq
= new_cfqq
->new_cfqq
)) {
1991 process_refs
= cfqq_process_refs(cfqq
);
1993 * If the process for the cfqq has gone away, there is no
1994 * sense in merging the queues.
1996 if (process_refs
== 0)
2000 * Merge in the direction of the lesser amount of work.
2002 new_process_refs
= cfqq_process_refs(new_cfqq
);
2003 if (new_process_refs
>= process_refs
) {
2004 cfqq
->new_cfqq
= new_cfqq
;
2005 atomic_add(process_refs
, &new_cfqq
->ref
);
2007 new_cfqq
->new_cfqq
= cfqq
;
2008 atomic_add(new_process_refs
, &cfqq
->ref
);
2012 static enum wl_type_t
cfq_choose_wl(struct cfq_data
*cfqd
,
2013 struct cfq_group
*cfqg
, enum wl_prio_t prio
)
2015 struct cfq_queue
*queue
;
2017 bool key_valid
= false;
2018 unsigned long lowest_key
= 0;
2019 enum wl_type_t cur_best
= SYNC_NOIDLE_WORKLOAD
;
2021 for (i
= 0; i
<= SYNC_WORKLOAD
; ++i
) {
2022 /* select the one with lowest rb_key */
2023 queue
= cfq_rb_first(service_tree_for(cfqg
, prio
, i
));
2025 (!key_valid
|| time_before(queue
->rb_key
, lowest_key
))) {
2026 lowest_key
= queue
->rb_key
;
2035 static void choose_service_tree(struct cfq_data
*cfqd
, struct cfq_group
*cfqg
)
2039 struct cfq_rb_root
*st
;
2040 unsigned group_slice
;
2043 cfqd
->serving_prio
= IDLE_WORKLOAD
;
2044 cfqd
->workload_expires
= jiffies
+ 1;
2048 /* Choose next priority. RT > BE > IDLE */
2049 if (cfq_group_busy_queues_wl(RT_WORKLOAD
, cfqd
, cfqg
))
2050 cfqd
->serving_prio
= RT_WORKLOAD
;
2051 else if (cfq_group_busy_queues_wl(BE_WORKLOAD
, cfqd
, cfqg
))
2052 cfqd
->serving_prio
= BE_WORKLOAD
;
2054 cfqd
->serving_prio
= IDLE_WORKLOAD
;
2055 cfqd
->workload_expires
= jiffies
+ 1;
2060 * For RT and BE, we have to choose also the type
2061 * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
2064 st
= service_tree_for(cfqg
, cfqd
->serving_prio
, cfqd
->serving_type
);
2068 * check workload expiration, and that we still have other queues ready
2070 if (count
&& !time_after(jiffies
, cfqd
->workload_expires
))
2073 /* otherwise select new workload type */
2074 cfqd
->serving_type
=
2075 cfq_choose_wl(cfqd
, cfqg
, cfqd
->serving_prio
);
2076 st
= service_tree_for(cfqg
, cfqd
->serving_prio
, cfqd
->serving_type
);
2080 * the workload slice is computed as a fraction of target latency
2081 * proportional to the number of queues in that workload, over
2082 * all the queues in the same priority class
2084 group_slice
= cfq_group_slice(cfqd
, cfqg
);
2086 slice
= group_slice
* count
/
2087 max_t(unsigned, cfqg
->busy_queues_avg
[cfqd
->serving_prio
],
2088 cfq_group_busy_queues_wl(cfqd
->serving_prio
, cfqd
, cfqg
));
2090 if (cfqd
->serving_type
== ASYNC_WORKLOAD
) {
2094 * Async queues are currently system wide. Just taking
2095 * proportion of queues with-in same group will lead to higher
2096 * async ratio system wide as generally root group is going
2097 * to have higher weight. A more accurate thing would be to
2098 * calculate system wide asnc/sync ratio.
2100 tmp
= cfq_target_latency
* cfqg_busy_async_queues(cfqd
, cfqg
);
2101 tmp
= tmp
/cfqd
->busy_queues
;
2102 slice
= min_t(unsigned, slice
, tmp
);
2104 /* async workload slice is scaled down according to
2105 * the sync/async slice ratio. */
2106 slice
= slice
* cfqd
->cfq_slice
[0] / cfqd
->cfq_slice
[1];
2108 /* sync workload slice is at least 2 * cfq_slice_idle */
2109 slice
= max(slice
, 2 * cfqd
->cfq_slice_idle
);
2111 slice
= max_t(unsigned, slice
, CFQ_MIN_TT
);
2112 cfq_log(cfqd
, "workload slice:%d", slice
);
2113 cfqd
->workload_expires
= jiffies
+ slice
;
2114 cfqd
->noidle_tree_requires_idle
= false;
2117 static struct cfq_group
*cfq_get_next_cfqg(struct cfq_data
*cfqd
)
2119 struct cfq_rb_root
*st
= &cfqd
->grp_service_tree
;
2120 struct cfq_group
*cfqg
;
2122 if (RB_EMPTY_ROOT(&st
->rb
))
2124 cfqg
= cfq_rb_first_group(st
);
2125 st
->active
= &cfqg
->rb_node
;
2126 update_min_vdisktime(st
);
2130 static void cfq_choose_cfqg(struct cfq_data
*cfqd
)
2132 struct cfq_group
*cfqg
= cfq_get_next_cfqg(cfqd
);
2134 cfqd
->serving_group
= cfqg
;
2136 /* Restore the workload type data */
2137 if (cfqg
->saved_workload_slice
) {
2138 cfqd
->workload_expires
= jiffies
+ cfqg
->saved_workload_slice
;
2139 cfqd
->serving_type
= cfqg
->saved_workload
;
2140 cfqd
->serving_prio
= cfqg
->saved_serving_prio
;
2142 cfqd
->workload_expires
= jiffies
- 1;
2144 choose_service_tree(cfqd
, cfqg
);
2148 * Select a queue for service. If we have a current active queue,
2149 * check whether to continue servicing it, or retrieve and set a new one.
2151 static struct cfq_queue
*cfq_select_queue(struct cfq_data
*cfqd
)
2153 struct cfq_queue
*cfqq
, *new_cfqq
= NULL
;
2155 cfqq
= cfqd
->active_queue
;
2159 if (!cfqd
->rq_queued
)
2163 * We were waiting for group to get backlogged. Expire the queue
2165 if (cfq_cfqq_wait_busy(cfqq
) && !RB_EMPTY_ROOT(&cfqq
->sort_list
))
2169 * The active queue has run out of time, expire it and select new.
2171 if (cfq_slice_used(cfqq
) && !cfq_cfqq_must_dispatch(cfqq
)) {
2173 * If slice had not expired at the completion of last request
2174 * we might not have turned on wait_busy flag. Don't expire
2175 * the queue yet. Allow the group to get backlogged.
2177 * The very fact that we have used the slice, that means we
2178 * have been idling all along on this queue and it should be
2179 * ok to wait for this request to complete.
2181 if (cfqq
->cfqg
->nr_cfqq
== 1 && RB_EMPTY_ROOT(&cfqq
->sort_list
)
2182 && cfqq
->dispatched
&& cfq_should_idle(cfqd
, cfqq
)) {
2190 * The active queue has requests and isn't expired, allow it to
2193 if (!RB_EMPTY_ROOT(&cfqq
->sort_list
))
2197 * If another queue has a request waiting within our mean seek
2198 * distance, let it run. The expire code will check for close
2199 * cooperators and put the close queue at the front of the service
2200 * tree. If possible, merge the expiring queue with the new cfqq.
2202 new_cfqq
= cfq_close_cooperator(cfqd
, cfqq
);
2204 if (!cfqq
->new_cfqq
)
2205 cfq_setup_merge(cfqq
, new_cfqq
);
2210 * No requests pending. If the active queue still has requests in
2211 * flight or is idling for a new request, allow either of these
2212 * conditions to happen (or time out) before selecting a new queue.
2214 if (timer_pending(&cfqd
->idle_slice_timer
) ||
2215 (cfqq
->dispatched
&& cfq_should_idle(cfqd
, cfqq
))) {
2221 cfq_slice_expired(cfqd
, 0);
2224 * Current queue expired. Check if we have to switch to a new
2228 cfq_choose_cfqg(cfqd
);
2230 cfqq
= cfq_set_active_queue(cfqd
, new_cfqq
);
2235 static int __cfq_forced_dispatch_cfqq(struct cfq_queue
*cfqq
)
2239 while (cfqq
->next_rq
) {
2240 cfq_dispatch_insert(cfqq
->cfqd
->queue
, cfqq
->next_rq
);
2244 BUG_ON(!list_empty(&cfqq
->fifo
));
2246 /* By default cfqq is not expired if it is empty. Do it explicitly */
2247 __cfq_slice_expired(cfqq
->cfqd
, cfqq
, 0);
2252 * Drain our current requests. Used for barriers and when switching
2253 * io schedulers on-the-fly.
2255 static int cfq_forced_dispatch(struct cfq_data
*cfqd
)
2257 struct cfq_queue
*cfqq
;
2260 /* Expire the timeslice of the current active queue first */
2261 cfq_slice_expired(cfqd
, 0);
2262 while ((cfqq
= cfq_get_next_queue_forced(cfqd
)) != NULL
) {
2263 __cfq_set_active_queue(cfqd
, cfqq
);
2264 dispatched
+= __cfq_forced_dispatch_cfqq(cfqq
);
2267 BUG_ON(cfqd
->busy_queues
);
2269 cfq_log(cfqd
, "forced_dispatch=%d", dispatched
);
2273 static inline bool cfq_slice_used_soon(struct cfq_data
*cfqd
,
2274 struct cfq_queue
*cfqq
)
2276 /* the queue hasn't finished any request, can't estimate */
2277 if (cfq_cfqq_slice_new(cfqq
))
2279 if (time_after(jiffies
+ cfqd
->cfq_slice_idle
* cfqq
->dispatched
,
2286 static bool cfq_may_dispatch(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
2288 unsigned int max_dispatch
;
2291 * Drain async requests before we start sync IO
2293 if (cfq_should_idle(cfqd
, cfqq
) && cfqd
->rq_in_flight
[BLK_RW_ASYNC
])
2297 * If this is an async queue and we have sync IO in flight, let it wait
2299 if (cfqd
->rq_in_flight
[BLK_RW_SYNC
] && !cfq_cfqq_sync(cfqq
))
2302 max_dispatch
= max_t(unsigned int, cfqd
->cfq_quantum
/ 2, 1);
2303 if (cfq_class_idle(cfqq
))
2307 * Does this cfqq already have too much IO in flight?
2309 if (cfqq
->dispatched
>= max_dispatch
) {
2311 * idle queue must always only have a single IO in flight
2313 if (cfq_class_idle(cfqq
))
2317 * We have other queues, don't allow more IO from this one
2319 if (cfqd
->busy_queues
> 1 && cfq_slice_used_soon(cfqd
, cfqq
))
2323 * Sole queue user, no limit
2325 if (cfqd
->busy_queues
== 1)
2329 * Normally we start throttling cfqq when cfq_quantum/2
2330 * requests have been dispatched. But we can drive
2331 * deeper queue depths at the beginning of slice
2332 * subjected to upper limit of cfq_quantum.
2334 max_dispatch
= cfqd
->cfq_quantum
;
2338 * Async queues must wait a bit before being allowed dispatch.
2339 * We also ramp up the dispatch depth gradually for async IO,
2340 * based on the last sync IO we serviced
2342 if (!cfq_cfqq_sync(cfqq
) && cfqd
->cfq_latency
) {
2343 unsigned long last_sync
= jiffies
- cfqd
->last_delayed_sync
;
2346 depth
= last_sync
/ cfqd
->cfq_slice
[1];
2347 if (!depth
&& !cfqq
->dispatched
)
2349 if (depth
< max_dispatch
)
2350 max_dispatch
= depth
;
2354 * If we're below the current max, allow a dispatch
2356 return cfqq
->dispatched
< max_dispatch
;
2360 * Dispatch a request from cfqq, moving them to the request queue
2363 static bool cfq_dispatch_request(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
2367 BUG_ON(RB_EMPTY_ROOT(&cfqq
->sort_list
));
2369 if (!cfq_may_dispatch(cfqd
, cfqq
))
2373 * follow expired path, else get first next available
2375 rq
= cfq_check_fifo(cfqq
);
2380 * insert request into driver dispatch list
2382 cfq_dispatch_insert(cfqd
->queue
, rq
);
2384 if (!cfqd
->active_cic
) {
2385 struct cfq_io_context
*cic
= RQ_CIC(rq
);
2387 atomic_long_inc(&cic
->ioc
->refcount
);
2388 cfqd
->active_cic
= cic
;
2395 * Find the cfqq that we need to service and move a request from that to the
2398 static int cfq_dispatch_requests(struct request_queue
*q
, int force
)
2400 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
2401 struct cfq_queue
*cfqq
;
2403 if (!cfqd
->busy_queues
)
2406 if (unlikely(force
))
2407 return cfq_forced_dispatch(cfqd
);
2409 cfqq
= cfq_select_queue(cfqd
);
2414 * Dispatch a request from this cfqq, if it is allowed
2416 if (!cfq_dispatch_request(cfqd
, cfqq
))
2419 cfqq
->slice_dispatch
++;
2420 cfq_clear_cfqq_must_dispatch(cfqq
);
2423 * expire an async queue immediately if it has used up its slice. idle
2424 * queue always expire after 1 dispatch round.
2426 if (cfqd
->busy_queues
> 1 && ((!cfq_cfqq_sync(cfqq
) &&
2427 cfqq
->slice_dispatch
>= cfq_prio_to_maxrq(cfqd
, cfqq
)) ||
2428 cfq_class_idle(cfqq
))) {
2429 cfqq
->slice_end
= jiffies
+ 1;
2430 cfq_slice_expired(cfqd
, 0);
2433 cfq_log_cfqq(cfqd
, cfqq
, "dispatched a request");
2438 * task holds one reference to the queue, dropped when task exits. each rq
2439 * in-flight on this queue also holds a reference, dropped when rq is freed.
2441 * Each cfq queue took a reference on the parent group. Drop it now.
2442 * queue lock must be held here.
2444 static void cfq_put_queue(struct cfq_queue
*cfqq
)
2446 struct cfq_data
*cfqd
= cfqq
->cfqd
;
2447 struct cfq_group
*cfqg
, *orig_cfqg
;
2449 BUG_ON(atomic_read(&cfqq
->ref
) <= 0);
2451 if (!atomic_dec_and_test(&cfqq
->ref
))
2454 cfq_log_cfqq(cfqd
, cfqq
, "put_queue");
2455 BUG_ON(rb_first(&cfqq
->sort_list
));
2456 BUG_ON(cfqq
->allocated
[READ
] + cfqq
->allocated
[WRITE
]);
2458 orig_cfqg
= cfqq
->orig_cfqg
;
2460 if (unlikely(cfqd
->active_queue
== cfqq
)) {
2461 __cfq_slice_expired(cfqd
, cfqq
, 0);
2462 cfq_schedule_dispatch(cfqd
);
2465 BUG_ON(cfq_cfqq_on_rr(cfqq
));
2466 kmem_cache_free(cfq_pool
, cfqq
);
2469 cfq_put_cfqg(orig_cfqg
);
2473 * Must always be called with the rcu_read_lock() held
2476 __call_for_each_cic(struct io_context
*ioc
,
2477 void (*func
)(struct io_context
*, struct cfq_io_context
*))
2479 struct cfq_io_context
*cic
;
2480 struct hlist_node
*n
;
2482 hlist_for_each_entry_rcu(cic
, n
, &ioc
->cic_list
, cic_list
)
2487 * Call func for each cic attached to this ioc.
2490 call_for_each_cic(struct io_context
*ioc
,
2491 void (*func
)(struct io_context
*, struct cfq_io_context
*))
2494 __call_for_each_cic(ioc
, func
);
2498 static void cfq_cic_free_rcu(struct rcu_head
*head
)
2500 struct cfq_io_context
*cic
;
2502 cic
= container_of(head
, struct cfq_io_context
, rcu_head
);
2504 kmem_cache_free(cfq_ioc_pool
, cic
);
2505 elv_ioc_count_dec(cfq_ioc_count
);
2509 * CFQ scheduler is exiting, grab exit lock and check
2510 * the pending io context count. If it hits zero,
2511 * complete ioc_gone and set it back to NULL
2513 spin_lock(&ioc_gone_lock
);
2514 if (ioc_gone
&& !elv_ioc_count_read(cfq_ioc_count
)) {
2518 spin_unlock(&ioc_gone_lock
);
2522 static void cfq_cic_free(struct cfq_io_context
*cic
)
2524 call_rcu(&cic
->rcu_head
, cfq_cic_free_rcu
);
2527 static void cic_free_func(struct io_context
*ioc
, struct cfq_io_context
*cic
)
2529 unsigned long flags
;
2530 unsigned long dead_key
= (unsigned long) cic
->key
;
2532 BUG_ON(!(dead_key
& CIC_DEAD_KEY
));
2534 spin_lock_irqsave(&ioc
->lock
, flags
);
2535 radix_tree_delete(&ioc
->radix_root
, dead_key
& ~CIC_DEAD_KEY
);
2536 hlist_del_rcu(&cic
->cic_list
);
2537 spin_unlock_irqrestore(&ioc
->lock
, flags
);
2543 * Must be called with rcu_read_lock() held or preemption otherwise disabled.
2544 * Only two callers of this - ->dtor() which is called with the rcu_read_lock(),
2545 * and ->trim() which is called with the task lock held
2547 static void cfq_free_io_context(struct io_context
*ioc
)
2550 * ioc->refcount is zero here, or we are called from elv_unregister(),
2551 * so no more cic's are allowed to be linked into this ioc. So it
2552 * should be ok to iterate over the known list, we will see all cic's
2553 * since no new ones are added.
2555 __call_for_each_cic(ioc
, cic_free_func
);
2558 static void cfq_exit_cfqq(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
2560 struct cfq_queue
*__cfqq
, *next
;
2562 if (unlikely(cfqq
== cfqd
->active_queue
)) {
2563 __cfq_slice_expired(cfqd
, cfqq
, 0);
2564 cfq_schedule_dispatch(cfqd
);
2568 * If this queue was scheduled to merge with another queue, be
2569 * sure to drop the reference taken on that queue (and others in
2570 * the merge chain). See cfq_setup_merge and cfq_merge_cfqqs.
2572 __cfqq
= cfqq
->new_cfqq
;
2574 if (__cfqq
== cfqq
) {
2575 WARN(1, "cfqq->new_cfqq loop detected\n");
2578 next
= __cfqq
->new_cfqq
;
2579 cfq_put_queue(__cfqq
);
2583 cfq_put_queue(cfqq
);
2586 static void __cfq_exit_single_io_context(struct cfq_data
*cfqd
,
2587 struct cfq_io_context
*cic
)
2589 struct io_context
*ioc
= cic
->ioc
;
2591 list_del_init(&cic
->queue_list
);
2594 * Make sure dead mark is seen for dead queues
2597 cic
->key
= cfqd_dead_key(cfqd
);
2599 if (ioc
->ioc_data
== cic
)
2600 rcu_assign_pointer(ioc
->ioc_data
, NULL
);
2602 if (cic
->cfqq
[BLK_RW_ASYNC
]) {
2603 cfq_exit_cfqq(cfqd
, cic
->cfqq
[BLK_RW_ASYNC
]);
2604 cic
->cfqq
[BLK_RW_ASYNC
] = NULL
;
2607 if (cic
->cfqq
[BLK_RW_SYNC
]) {
2608 cfq_exit_cfqq(cfqd
, cic
->cfqq
[BLK_RW_SYNC
]);
2609 cic
->cfqq
[BLK_RW_SYNC
] = NULL
;
2613 static void cfq_exit_single_io_context(struct io_context
*ioc
,
2614 struct cfq_io_context
*cic
)
2616 struct cfq_data
*cfqd
= cic_to_cfqd(cic
);
2619 struct request_queue
*q
= cfqd
->queue
;
2620 unsigned long flags
;
2622 spin_lock_irqsave(q
->queue_lock
, flags
);
2625 * Ensure we get a fresh copy of the ->key to prevent
2626 * race between exiting task and queue
2628 smp_read_barrier_depends();
2629 if (cic
->key
== cfqd
)
2630 __cfq_exit_single_io_context(cfqd
, cic
);
2632 spin_unlock_irqrestore(q
->queue_lock
, flags
);
2637 * The process that ioc belongs to has exited, we need to clean up
2638 * and put the internal structures we have that belongs to that process.
2640 static void cfq_exit_io_context(struct io_context
*ioc
)
2642 call_for_each_cic(ioc
, cfq_exit_single_io_context
);
2645 static struct cfq_io_context
*
2646 cfq_alloc_io_context(struct cfq_data
*cfqd
, gfp_t gfp_mask
)
2648 struct cfq_io_context
*cic
;
2650 cic
= kmem_cache_alloc_node(cfq_ioc_pool
, gfp_mask
| __GFP_ZERO
,
2653 cic
->last_end_request
= jiffies
;
2654 INIT_LIST_HEAD(&cic
->queue_list
);
2655 INIT_HLIST_NODE(&cic
->cic_list
);
2656 cic
->dtor
= cfq_free_io_context
;
2657 cic
->exit
= cfq_exit_io_context
;
2658 elv_ioc_count_inc(cfq_ioc_count
);
2664 static void cfq_init_prio_data(struct cfq_queue
*cfqq
, struct io_context
*ioc
)
2666 struct task_struct
*tsk
= current
;
2669 if (!cfq_cfqq_prio_changed(cfqq
))
2672 ioprio_class
= IOPRIO_PRIO_CLASS(ioc
->ioprio
);
2673 switch (ioprio_class
) {
2675 printk(KERN_ERR
"cfq: bad prio %x\n", ioprio_class
);
2676 case IOPRIO_CLASS_NONE
:
2678 * no prio set, inherit CPU scheduling settings
2680 cfqq
->ioprio
= task_nice_ioprio(tsk
);
2681 cfqq
->ioprio_class
= task_nice_ioclass(tsk
);
2683 case IOPRIO_CLASS_RT
:
2684 cfqq
->ioprio
= task_ioprio(ioc
);
2685 cfqq
->ioprio_class
= IOPRIO_CLASS_RT
;
2687 case IOPRIO_CLASS_BE
:
2688 cfqq
->ioprio
= task_ioprio(ioc
);
2689 cfqq
->ioprio_class
= IOPRIO_CLASS_BE
;
2691 case IOPRIO_CLASS_IDLE
:
2692 cfqq
->ioprio_class
= IOPRIO_CLASS_IDLE
;
2694 cfq_clear_cfqq_idle_window(cfqq
);
2699 * keep track of original prio settings in case we have to temporarily
2700 * elevate the priority of this queue
2702 cfqq
->org_ioprio
= cfqq
->ioprio
;
2703 cfqq
->org_ioprio_class
= cfqq
->ioprio_class
;
2704 cfq_clear_cfqq_prio_changed(cfqq
);
2707 static void changed_ioprio(struct io_context
*ioc
, struct cfq_io_context
*cic
)
2709 struct cfq_data
*cfqd
= cic_to_cfqd(cic
);
2710 struct cfq_queue
*cfqq
;
2711 unsigned long flags
;
2713 if (unlikely(!cfqd
))
2716 spin_lock_irqsave(cfqd
->queue
->queue_lock
, flags
);
2718 cfqq
= cic
->cfqq
[BLK_RW_ASYNC
];
2720 struct cfq_queue
*new_cfqq
;
2721 new_cfqq
= cfq_get_queue(cfqd
, BLK_RW_ASYNC
, cic
->ioc
,
2724 cic
->cfqq
[BLK_RW_ASYNC
] = new_cfqq
;
2725 cfq_put_queue(cfqq
);
2729 cfqq
= cic
->cfqq
[BLK_RW_SYNC
];
2731 cfq_mark_cfqq_prio_changed(cfqq
);
2733 spin_unlock_irqrestore(cfqd
->queue
->queue_lock
, flags
);
2736 static void cfq_ioc_set_ioprio(struct io_context
*ioc
)
2738 call_for_each_cic(ioc
, changed_ioprio
);
2739 ioc
->ioprio_changed
= 0;
2742 static void cfq_init_cfqq(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
2743 pid_t pid
, bool is_sync
)
2745 RB_CLEAR_NODE(&cfqq
->rb_node
);
2746 RB_CLEAR_NODE(&cfqq
->p_node
);
2747 INIT_LIST_HEAD(&cfqq
->fifo
);
2749 atomic_set(&cfqq
->ref
, 0);
2752 cfq_mark_cfqq_prio_changed(cfqq
);
2755 if (!cfq_class_idle(cfqq
))
2756 cfq_mark_cfqq_idle_window(cfqq
);
2757 cfq_mark_cfqq_sync(cfqq
);
2762 #ifdef CONFIG_CFQ_GROUP_IOSCHED
2763 static void changed_cgroup(struct io_context
*ioc
, struct cfq_io_context
*cic
)
2765 struct cfq_queue
*sync_cfqq
= cic_to_cfqq(cic
, 1);
2766 struct cfq_data
*cfqd
= cic_to_cfqd(cic
);
2767 unsigned long flags
;
2768 struct request_queue
*q
;
2770 if (unlikely(!cfqd
))
2775 spin_lock_irqsave(q
->queue_lock
, flags
);
2779 * Drop reference to sync queue. A new sync queue will be
2780 * assigned in new group upon arrival of a fresh request.
2782 cfq_log_cfqq(cfqd
, sync_cfqq
, "changed cgroup");
2783 cic_set_cfqq(cic
, NULL
, 1);
2784 cfq_put_queue(sync_cfqq
);
2787 spin_unlock_irqrestore(q
->queue_lock
, flags
);
2790 static void cfq_ioc_set_cgroup(struct io_context
*ioc
)
2792 call_for_each_cic(ioc
, changed_cgroup
);
2793 ioc
->cgroup_changed
= 0;
2795 #endif /* CONFIG_CFQ_GROUP_IOSCHED */
2797 static struct cfq_queue
*
2798 cfq_find_alloc_queue(struct cfq_data
*cfqd
, bool is_sync
,
2799 struct io_context
*ioc
, gfp_t gfp_mask
)
2801 struct cfq_queue
*cfqq
, *new_cfqq
= NULL
;
2802 struct cfq_io_context
*cic
;
2803 struct cfq_group
*cfqg
;
2806 cfqg
= cfq_get_cfqg(cfqd
, 1);
2807 cic
= cfq_cic_lookup(cfqd
, ioc
);
2808 /* cic always exists here */
2809 cfqq
= cic_to_cfqq(cic
, is_sync
);
2812 * Always try a new alloc if we fell back to the OOM cfqq
2813 * originally, since it should just be a temporary situation.
2815 if (!cfqq
|| cfqq
== &cfqd
->oom_cfqq
) {
2820 } else if (gfp_mask
& __GFP_WAIT
) {
2821 spin_unlock_irq(cfqd
->queue
->queue_lock
);
2822 new_cfqq
= kmem_cache_alloc_node(cfq_pool
,
2823 gfp_mask
| __GFP_ZERO
,
2825 spin_lock_irq(cfqd
->queue
->queue_lock
);
2829 cfqq
= kmem_cache_alloc_node(cfq_pool
,
2830 gfp_mask
| __GFP_ZERO
,
2835 cfq_init_cfqq(cfqd
, cfqq
, current
->pid
, is_sync
);
2836 cfq_init_prio_data(cfqq
, ioc
);
2837 cfq_link_cfqq_cfqg(cfqq
, cfqg
);
2838 cfq_log_cfqq(cfqd
, cfqq
, "alloced");
2840 cfqq
= &cfqd
->oom_cfqq
;
2844 kmem_cache_free(cfq_pool
, new_cfqq
);
2849 static struct cfq_queue
**
2850 cfq_async_queue_prio(struct cfq_data
*cfqd
, int ioprio_class
, int ioprio
)
2852 switch (ioprio_class
) {
2853 case IOPRIO_CLASS_RT
:
2854 return &cfqd
->async_cfqq
[0][ioprio
];
2855 case IOPRIO_CLASS_BE
:
2856 return &cfqd
->async_cfqq
[1][ioprio
];
2857 case IOPRIO_CLASS_IDLE
:
2858 return &cfqd
->async_idle_cfqq
;
2864 static struct cfq_queue
*
2865 cfq_get_queue(struct cfq_data
*cfqd
, bool is_sync
, struct io_context
*ioc
,
2868 const int ioprio
= task_ioprio(ioc
);
2869 const int ioprio_class
= task_ioprio_class(ioc
);
2870 struct cfq_queue
**async_cfqq
= NULL
;
2871 struct cfq_queue
*cfqq
= NULL
;
2874 async_cfqq
= cfq_async_queue_prio(cfqd
, ioprio_class
, ioprio
);
2879 cfqq
= cfq_find_alloc_queue(cfqd
, is_sync
, ioc
, gfp_mask
);
2882 * pin the queue now that it's allocated, scheduler exit will prune it
2884 if (!is_sync
&& !(*async_cfqq
)) {
2885 atomic_inc(&cfqq
->ref
);
2889 atomic_inc(&cfqq
->ref
);
2894 * We drop cfq io contexts lazily, so we may find a dead one.
2897 cfq_drop_dead_cic(struct cfq_data
*cfqd
, struct io_context
*ioc
,
2898 struct cfq_io_context
*cic
)
2900 unsigned long flags
;
2902 WARN_ON(!list_empty(&cic
->queue_list
));
2903 BUG_ON(cic
->key
!= cfqd_dead_key(cfqd
));
2905 spin_lock_irqsave(&ioc
->lock
, flags
);
2907 BUG_ON(ioc
->ioc_data
== cic
);
2909 radix_tree_delete(&ioc
->radix_root
, (unsigned long) cfqd
);
2910 hlist_del_rcu(&cic
->cic_list
);
2911 spin_unlock_irqrestore(&ioc
->lock
, flags
);
2916 static struct cfq_io_context
*
2917 cfq_cic_lookup(struct cfq_data
*cfqd
, struct io_context
*ioc
)
2919 struct cfq_io_context
*cic
;
2920 unsigned long flags
;
2928 * we maintain a last-hit cache, to avoid browsing over the tree
2930 cic
= rcu_dereference(ioc
->ioc_data
);
2931 if (cic
&& cic
->key
== cfqd
) {
2937 cic
= radix_tree_lookup(&ioc
->radix_root
, (unsigned long) cfqd
);
2941 if (unlikely(cic
->key
!= cfqd
)) {
2942 cfq_drop_dead_cic(cfqd
, ioc
, cic
);
2947 spin_lock_irqsave(&ioc
->lock
, flags
);
2948 rcu_assign_pointer(ioc
->ioc_data
, cic
);
2949 spin_unlock_irqrestore(&ioc
->lock
, flags
);
2957 * Add cic into ioc, using cfqd as the search key. This enables us to lookup
2958 * the process specific cfq io context when entered from the block layer.
2959 * Also adds the cic to a per-cfqd list, used when this queue is removed.
2961 static int cfq_cic_link(struct cfq_data
*cfqd
, struct io_context
*ioc
,
2962 struct cfq_io_context
*cic
, gfp_t gfp_mask
)
2964 unsigned long flags
;
2967 ret
= radix_tree_preload(gfp_mask
);
2972 spin_lock_irqsave(&ioc
->lock
, flags
);
2973 ret
= radix_tree_insert(&ioc
->radix_root
,
2974 (unsigned long) cfqd
, cic
);
2976 hlist_add_head_rcu(&cic
->cic_list
, &ioc
->cic_list
);
2977 spin_unlock_irqrestore(&ioc
->lock
, flags
);
2979 radix_tree_preload_end();
2982 spin_lock_irqsave(cfqd
->queue
->queue_lock
, flags
);
2983 list_add(&cic
->queue_list
, &cfqd
->cic_list
);
2984 spin_unlock_irqrestore(cfqd
->queue
->queue_lock
, flags
);
2989 printk(KERN_ERR
"cfq: cic link failed!\n");
2995 * Setup general io context and cfq io context. There can be several cfq
2996 * io contexts per general io context, if this process is doing io to more
2997 * than one device managed by cfq.
2999 static struct cfq_io_context
*
3000 cfq_get_io_context(struct cfq_data
*cfqd
, gfp_t gfp_mask
)
3002 struct io_context
*ioc
= NULL
;
3003 struct cfq_io_context
*cic
;
3005 might_sleep_if(gfp_mask
& __GFP_WAIT
);
3007 ioc
= get_io_context(gfp_mask
, cfqd
->queue
->node
);
3011 cic
= cfq_cic_lookup(cfqd
, ioc
);
3015 cic
= cfq_alloc_io_context(cfqd
, gfp_mask
);
3019 if (cfq_cic_link(cfqd
, ioc
, cic
, gfp_mask
))
3023 smp_read_barrier_depends();
3024 if (unlikely(ioc
->ioprio_changed
))
3025 cfq_ioc_set_ioprio(ioc
);
3027 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3028 if (unlikely(ioc
->cgroup_changed
))
3029 cfq_ioc_set_cgroup(ioc
);
3035 put_io_context(ioc
);
3040 cfq_update_io_thinktime(struct cfq_data
*cfqd
, struct cfq_io_context
*cic
)
3042 unsigned long elapsed
= jiffies
- cic
->last_end_request
;
3043 unsigned long ttime
= min(elapsed
, 2UL * cfqd
->cfq_slice_idle
);
3045 cic
->ttime_samples
= (7*cic
->ttime_samples
+ 256) / 8;
3046 cic
->ttime_total
= (7*cic
->ttime_total
+ 256*ttime
) / 8;
3047 cic
->ttime_mean
= (cic
->ttime_total
+ 128) / cic
->ttime_samples
;
3051 cfq_update_io_seektime(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
3055 sector_t n_sec
= blk_rq_sectors(rq
);
3056 if (cfqq
->last_request_pos
) {
3057 if (cfqq
->last_request_pos
< blk_rq_pos(rq
))
3058 sdist
= blk_rq_pos(rq
) - cfqq
->last_request_pos
;
3060 sdist
= cfqq
->last_request_pos
- blk_rq_pos(rq
);
3063 cfqq
->seek_history
<<= 1;
3064 if (blk_queue_nonrot(cfqd
->queue
))
3065 cfqq
->seek_history
|= (n_sec
< CFQQ_SECT_THR_NONROT
);
3067 cfqq
->seek_history
|= (sdist
> CFQQ_SEEK_THR
);
3071 * Disable idle window if the process thinks too long or seeks so much that
3075 cfq_update_idle_window(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
3076 struct cfq_io_context
*cic
)
3078 int old_idle
, enable_idle
;
3081 * Don't idle for async or idle io prio class
3083 if (!cfq_cfqq_sync(cfqq
) || cfq_class_idle(cfqq
))
3086 enable_idle
= old_idle
= cfq_cfqq_idle_window(cfqq
);
3088 if (cfqq
->queued
[0] + cfqq
->queued
[1] >= 4)
3089 cfq_mark_cfqq_deep(cfqq
);
3091 if (!atomic_read(&cic
->ioc
->nr_tasks
) || !cfqd
->cfq_slice_idle
||
3092 (!cfq_cfqq_deep(cfqq
) && CFQQ_SEEKY(cfqq
)))
3094 else if (sample_valid(cic
->ttime_samples
)) {
3095 if (cic
->ttime_mean
> cfqd
->cfq_slice_idle
)
3101 if (old_idle
!= enable_idle
) {
3102 cfq_log_cfqq(cfqd
, cfqq
, "idle=%d", enable_idle
);
3104 cfq_mark_cfqq_idle_window(cfqq
);
3106 cfq_clear_cfqq_idle_window(cfqq
);
3111 * Check if new_cfqq should preempt the currently active queue. Return 0 for
3112 * no or if we aren't sure, a 1 will cause a preempt.
3115 cfq_should_preempt(struct cfq_data
*cfqd
, struct cfq_queue
*new_cfqq
,
3118 struct cfq_queue
*cfqq
;
3120 cfqq
= cfqd
->active_queue
;
3124 if (cfq_class_idle(new_cfqq
))
3127 if (cfq_class_idle(cfqq
))
3131 * Don't allow a non-RT request to preempt an ongoing RT cfqq timeslice.
3133 if (cfq_class_rt(cfqq
) && !cfq_class_rt(new_cfqq
))
3137 * if the new request is sync, but the currently running queue is
3138 * not, let the sync request have priority.
3140 if (rq_is_sync(rq
) && !cfq_cfqq_sync(cfqq
))
3143 if (new_cfqq
->cfqg
!= cfqq
->cfqg
)
3146 if (cfq_slice_used(cfqq
))
3149 /* Allow preemption only if we are idling on sync-noidle tree */
3150 if (cfqd
->serving_type
== SYNC_NOIDLE_WORKLOAD
&&
3151 cfqq_type(new_cfqq
) == SYNC_NOIDLE_WORKLOAD
&&
3152 new_cfqq
->service_tree
->count
== 2 &&
3153 RB_EMPTY_ROOT(&cfqq
->sort_list
))
3157 * So both queues are sync. Let the new request get disk time if
3158 * it's a metadata request and the current queue is doing regular IO.
3160 if (rq_is_meta(rq
) && !cfqq
->meta_pending
)
3164 * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
3166 if (cfq_class_rt(new_cfqq
) && !cfq_class_rt(cfqq
))
3169 if (!cfqd
->active_cic
|| !cfq_cfqq_wait_request(cfqq
))
3173 * if this request is as-good as one we would expect from the
3174 * current cfqq, let it preempt
3176 if (cfq_rq_close(cfqd
, cfqq
, rq
))
3183 * cfqq preempts the active queue. if we allowed preempt with no slice left,
3184 * let it have half of its nominal slice.
3186 static void cfq_preempt_queue(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
3188 cfq_log_cfqq(cfqd
, cfqq
, "preempt");
3189 cfq_slice_expired(cfqd
, 1);
3192 * Put the new queue at the front of the of the current list,
3193 * so we know that it will be selected next.
3195 BUG_ON(!cfq_cfqq_on_rr(cfqq
));
3197 cfq_service_tree_add(cfqd
, cfqq
, 1);
3199 cfqq
->slice_end
= 0;
3200 cfq_mark_cfqq_slice_new(cfqq
);
3204 * Called when a new fs request (rq) is added (to cfqq). Check if there's
3205 * something we should do about it
3208 cfq_rq_enqueued(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
3211 struct cfq_io_context
*cic
= RQ_CIC(rq
);
3215 cfqq
->meta_pending
++;
3217 cfq_update_io_thinktime(cfqd
, cic
);
3218 cfq_update_io_seektime(cfqd
, cfqq
, rq
);
3219 cfq_update_idle_window(cfqd
, cfqq
, cic
);
3221 cfqq
->last_request_pos
= blk_rq_pos(rq
) + blk_rq_sectors(rq
);
3223 if (cfqq
== cfqd
->active_queue
) {
3225 * Remember that we saw a request from this process, but
3226 * don't start queuing just yet. Otherwise we risk seeing lots
3227 * of tiny requests, because we disrupt the normal plugging
3228 * and merging. If the request is already larger than a single
3229 * page, let it rip immediately. For that case we assume that
3230 * merging is already done. Ditto for a busy system that
3231 * has other work pending, don't risk delaying until the
3232 * idle timer unplug to continue working.
3234 if (cfq_cfqq_wait_request(cfqq
)) {
3235 if (blk_rq_bytes(rq
) > PAGE_CACHE_SIZE
||
3236 cfqd
->busy_queues
> 1) {
3237 cfq_del_timer(cfqd
, cfqq
);
3238 cfq_clear_cfqq_wait_request(cfqq
);
3239 __blk_run_queue(cfqd
->queue
);
3241 blkiocg_update_idle_time_stats(
3243 cfq_mark_cfqq_must_dispatch(cfqq
);
3246 } else if (cfq_should_preempt(cfqd
, cfqq
, rq
)) {
3248 * not the active queue - expire current slice if it is
3249 * idle and has expired it's mean thinktime or this new queue
3250 * has some old slice time left and is of higher priority or
3251 * this new queue is RT and the current one is BE
3253 cfq_preempt_queue(cfqd
, cfqq
);
3254 __blk_run_queue(cfqd
->queue
);
3258 static void cfq_insert_request(struct request_queue
*q
, struct request
*rq
)
3260 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
3261 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
3263 cfq_log_cfqq(cfqd
, cfqq
, "insert_request");
3264 cfq_init_prio_data(cfqq
, RQ_CIC(rq
)->ioc
);
3266 rq_set_fifo_time(rq
, jiffies
+ cfqd
->cfq_fifo_expire
[rq_is_sync(rq
)]);
3267 list_add_tail(&rq
->queuelist
, &cfqq
->fifo
);
3269 blkiocg_update_io_add_stats(&(RQ_CFQG(rq
))->blkg
,
3270 &cfqd
->serving_group
->blkg
, rq_data_dir(rq
),
3272 cfq_rq_enqueued(cfqd
, cfqq
, rq
);
3276 * Update hw_tag based on peak queue depth over 50 samples under
3279 static void cfq_update_hw_tag(struct cfq_data
*cfqd
)
3281 struct cfq_queue
*cfqq
= cfqd
->active_queue
;
3283 if (cfqd
->rq_in_driver
> cfqd
->hw_tag_est_depth
)
3284 cfqd
->hw_tag_est_depth
= cfqd
->rq_in_driver
;
3286 if (cfqd
->hw_tag
== 1)
3289 if (cfqd
->rq_queued
<= CFQ_HW_QUEUE_MIN
&&
3290 cfqd
->rq_in_driver
<= CFQ_HW_QUEUE_MIN
)
3294 * If active queue hasn't enough requests and can idle, cfq might not
3295 * dispatch sufficient requests to hardware. Don't zero hw_tag in this
3298 if (cfqq
&& cfq_cfqq_idle_window(cfqq
) &&
3299 cfqq
->dispatched
+ cfqq
->queued
[0] + cfqq
->queued
[1] <
3300 CFQ_HW_QUEUE_MIN
&& cfqd
->rq_in_driver
< CFQ_HW_QUEUE_MIN
)
3303 if (cfqd
->hw_tag_samples
++ < 50)
3306 if (cfqd
->hw_tag_est_depth
>= CFQ_HW_QUEUE_MIN
)
3312 static bool cfq_should_wait_busy(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
3314 struct cfq_io_context
*cic
= cfqd
->active_cic
;
3316 /* If there are other queues in the group, don't wait */
3317 if (cfqq
->cfqg
->nr_cfqq
> 1)
3320 if (cfq_slice_used(cfqq
))
3323 /* if slice left is less than think time, wait busy */
3324 if (cic
&& sample_valid(cic
->ttime_samples
)
3325 && (cfqq
->slice_end
- jiffies
< cic
->ttime_mean
))
3329 * If think times is less than a jiffy than ttime_mean=0 and above
3330 * will not be true. It might happen that slice has not expired yet
3331 * but will expire soon (4-5 ns) during select_queue(). To cover the
3332 * case where think time is less than a jiffy, mark the queue wait
3333 * busy if only 1 jiffy is left in the slice.
3335 if (cfqq
->slice_end
- jiffies
== 1)
3341 static void cfq_completed_request(struct request_queue
*q
, struct request
*rq
)
3343 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
3344 struct cfq_data
*cfqd
= cfqq
->cfqd
;
3345 const int sync
= rq_is_sync(rq
);
3349 cfq_log_cfqq(cfqd
, cfqq
, "complete rqnoidle %d", !!rq_noidle(rq
));
3351 cfq_update_hw_tag(cfqd
);
3353 WARN_ON(!cfqd
->rq_in_driver
);
3354 WARN_ON(!cfqq
->dispatched
);
3355 cfqd
->rq_in_driver
--;
3357 blkiocg_update_completion_stats(&cfqq
->cfqg
->blkg
, rq_start_time_ns(rq
),
3358 rq_io_start_time_ns(rq
), rq_data_dir(rq
),
3361 cfqd
->rq_in_flight
[cfq_cfqq_sync(cfqq
)]--;
3364 RQ_CIC(rq
)->last_end_request
= now
;
3365 if (!time_after(rq
->start_time
+ cfqd
->cfq_fifo_expire
[1], now
))
3366 cfqd
->last_delayed_sync
= now
;
3370 * If this is the active queue, check if it needs to be expired,
3371 * or if we want to idle in case it has no pending requests.
3373 if (cfqd
->active_queue
== cfqq
) {
3374 const bool cfqq_empty
= RB_EMPTY_ROOT(&cfqq
->sort_list
);
3376 if (cfq_cfqq_slice_new(cfqq
)) {
3377 cfq_set_prio_slice(cfqd
, cfqq
);
3378 cfq_clear_cfqq_slice_new(cfqq
);
3382 * Should we wait for next request to come in before we expire
3385 if (cfq_should_wait_busy(cfqd
, cfqq
)) {
3386 cfqq
->slice_end
= jiffies
+ cfqd
->cfq_slice_idle
;
3387 cfq_mark_cfqq_wait_busy(cfqq
);
3388 cfq_log_cfqq(cfqd
, cfqq
, "will busy wait");
3392 * Idling is not enabled on:
3394 * - idle-priority queues
3396 * - queues with still some requests queued
3397 * - when there is a close cooperator
3399 if (cfq_slice_used(cfqq
) || cfq_class_idle(cfqq
))
3400 cfq_slice_expired(cfqd
, 1);
3401 else if (sync
&& cfqq_empty
&&
3402 !cfq_close_cooperator(cfqd
, cfqq
)) {
3403 cfqd
->noidle_tree_requires_idle
|= !rq_noidle(rq
);
3405 * Idling is enabled for SYNC_WORKLOAD.
3406 * SYNC_NOIDLE_WORKLOAD idles at the end of the tree
3407 * only if we processed at least one !rq_noidle request
3409 if (cfqd
->serving_type
== SYNC_WORKLOAD
3410 || cfqd
->noidle_tree_requires_idle
3411 || cfqq
->cfqg
->nr_cfqq
== 1)
3412 cfq_arm_slice_timer(cfqd
);
3416 if (!cfqd
->rq_in_driver
)
3417 cfq_schedule_dispatch(cfqd
);
3421 * we temporarily boost lower priority queues if they are holding fs exclusive
3422 * resources. they are boosted to normal prio (CLASS_BE/4)
3424 static void cfq_prio_boost(struct cfq_queue
*cfqq
)
3426 if (has_fs_excl()) {
3428 * boost idle prio on transactions that would lock out other
3429 * users of the filesystem
3431 if (cfq_class_idle(cfqq
))
3432 cfqq
->ioprio_class
= IOPRIO_CLASS_BE
;
3433 if (cfqq
->ioprio
> IOPRIO_NORM
)
3434 cfqq
->ioprio
= IOPRIO_NORM
;
3437 * unboost the queue (if needed)
3439 cfqq
->ioprio_class
= cfqq
->org_ioprio_class
;
3440 cfqq
->ioprio
= cfqq
->org_ioprio
;
3444 static inline int __cfq_may_queue(struct cfq_queue
*cfqq
)
3446 if (cfq_cfqq_wait_request(cfqq
) && !cfq_cfqq_must_alloc_slice(cfqq
)) {
3447 cfq_mark_cfqq_must_alloc_slice(cfqq
);
3448 return ELV_MQUEUE_MUST
;
3451 return ELV_MQUEUE_MAY
;
3454 static int cfq_may_queue(struct request_queue
*q
, int rw
)
3456 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
3457 struct task_struct
*tsk
= current
;
3458 struct cfq_io_context
*cic
;
3459 struct cfq_queue
*cfqq
;
3462 * don't force setup of a queue from here, as a call to may_queue
3463 * does not necessarily imply that a request actually will be queued.
3464 * so just lookup a possibly existing queue, or return 'may queue'
3467 cic
= cfq_cic_lookup(cfqd
, tsk
->io_context
);
3469 return ELV_MQUEUE_MAY
;
3471 cfqq
= cic_to_cfqq(cic
, rw_is_sync(rw
));
3473 cfq_init_prio_data(cfqq
, cic
->ioc
);
3474 cfq_prio_boost(cfqq
);
3476 return __cfq_may_queue(cfqq
);
3479 return ELV_MQUEUE_MAY
;
3483 * queue lock held here
3485 static void cfq_put_request(struct request
*rq
)
3487 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
3490 const int rw
= rq_data_dir(rq
);
3492 BUG_ON(!cfqq
->allocated
[rw
]);
3493 cfqq
->allocated
[rw
]--;
3495 put_io_context(RQ_CIC(rq
)->ioc
);
3497 rq
->elevator_private
= NULL
;
3498 rq
->elevator_private2
= NULL
;
3500 /* Put down rq reference on cfqg */
3501 cfq_put_cfqg(RQ_CFQG(rq
));
3502 rq
->elevator_private3
= NULL
;
3504 cfq_put_queue(cfqq
);
3508 static struct cfq_queue
*
3509 cfq_merge_cfqqs(struct cfq_data
*cfqd
, struct cfq_io_context
*cic
,
3510 struct cfq_queue
*cfqq
)
3512 cfq_log_cfqq(cfqd
, cfqq
, "merging with queue %p", cfqq
->new_cfqq
);
3513 cic_set_cfqq(cic
, cfqq
->new_cfqq
, 1);
3514 cfq_mark_cfqq_coop(cfqq
->new_cfqq
);
3515 cfq_put_queue(cfqq
);
3516 return cic_to_cfqq(cic
, 1);
3520 * Returns NULL if a new cfqq should be allocated, or the old cfqq if this
3521 * was the last process referring to said cfqq.
3523 static struct cfq_queue
*
3524 split_cfqq(struct cfq_io_context
*cic
, struct cfq_queue
*cfqq
)
3526 if (cfqq_process_refs(cfqq
) == 1) {
3527 cfqq
->pid
= current
->pid
;
3528 cfq_clear_cfqq_coop(cfqq
);
3529 cfq_clear_cfqq_split_coop(cfqq
);
3533 cic_set_cfqq(cic
, NULL
, 1);
3534 cfq_put_queue(cfqq
);
3538 * Allocate cfq data structures associated with this request.
3541 cfq_set_request(struct request_queue
*q
, struct request
*rq
, gfp_t gfp_mask
)
3543 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
3544 struct cfq_io_context
*cic
;
3545 const int rw
= rq_data_dir(rq
);
3546 const bool is_sync
= rq_is_sync(rq
);
3547 struct cfq_queue
*cfqq
;
3548 unsigned long flags
;
3550 might_sleep_if(gfp_mask
& __GFP_WAIT
);
3552 cic
= cfq_get_io_context(cfqd
, gfp_mask
);
3554 spin_lock_irqsave(q
->queue_lock
, flags
);
3560 cfqq
= cic_to_cfqq(cic
, is_sync
);
3561 if (!cfqq
|| cfqq
== &cfqd
->oom_cfqq
) {
3562 cfqq
= cfq_get_queue(cfqd
, is_sync
, cic
->ioc
, gfp_mask
);
3563 cic_set_cfqq(cic
, cfqq
, is_sync
);
3566 * If the queue was seeky for too long, break it apart.
3568 if (cfq_cfqq_coop(cfqq
) && cfq_cfqq_split_coop(cfqq
)) {
3569 cfq_log_cfqq(cfqd
, cfqq
, "breaking apart cfqq");
3570 cfqq
= split_cfqq(cic
, cfqq
);
3576 * Check to see if this queue is scheduled to merge with
3577 * another, closely cooperating queue. The merging of
3578 * queues happens here as it must be done in process context.
3579 * The reference on new_cfqq was taken in merge_cfqqs.
3582 cfqq
= cfq_merge_cfqqs(cfqd
, cic
, cfqq
);
3585 cfqq
->allocated
[rw
]++;
3586 atomic_inc(&cfqq
->ref
);
3588 spin_unlock_irqrestore(q
->queue_lock
, flags
);
3590 rq
->elevator_private
= cic
;
3591 rq
->elevator_private2
= cfqq
;
3592 rq
->elevator_private3
= cfq_ref_get_cfqg(cfqq
->cfqg
);
3597 put_io_context(cic
->ioc
);
3599 cfq_schedule_dispatch(cfqd
);
3600 spin_unlock_irqrestore(q
->queue_lock
, flags
);
3601 cfq_log(cfqd
, "set_request fail");
3605 static void cfq_kick_queue(struct work_struct
*work
)
3607 struct cfq_data
*cfqd
=
3608 container_of(work
, struct cfq_data
, unplug_work
);
3609 struct request_queue
*q
= cfqd
->queue
;
3611 spin_lock_irq(q
->queue_lock
);
3612 __blk_run_queue(cfqd
->queue
);
3613 spin_unlock_irq(q
->queue_lock
);
3617 * Timer running if the active_queue is currently idling inside its time slice
3619 static void cfq_idle_slice_timer(unsigned long data
)
3621 struct cfq_data
*cfqd
= (struct cfq_data
*) data
;
3622 struct cfq_queue
*cfqq
;
3623 unsigned long flags
;
3626 cfq_log(cfqd
, "idle timer fired");
3628 spin_lock_irqsave(cfqd
->queue
->queue_lock
, flags
);
3630 cfqq
= cfqd
->active_queue
;
3635 * We saw a request before the queue expired, let it through
3637 if (cfq_cfqq_must_dispatch(cfqq
))
3643 if (cfq_slice_used(cfqq
))
3647 * only expire and reinvoke request handler, if there are
3648 * other queues with pending requests
3650 if (!cfqd
->busy_queues
)
3654 * not expired and it has a request pending, let it dispatch
3656 if (!RB_EMPTY_ROOT(&cfqq
->sort_list
))
3660 * Queue depth flag is reset only when the idle didn't succeed
3662 cfq_clear_cfqq_deep(cfqq
);
3665 cfq_slice_expired(cfqd
, timed_out
);
3667 cfq_schedule_dispatch(cfqd
);
3669 spin_unlock_irqrestore(cfqd
->queue
->queue_lock
, flags
);
3672 static void cfq_shutdown_timer_wq(struct cfq_data
*cfqd
)
3674 del_timer_sync(&cfqd
->idle_slice_timer
);
3675 cancel_work_sync(&cfqd
->unplug_work
);
3678 static void cfq_put_async_queues(struct cfq_data
*cfqd
)
3682 for (i
= 0; i
< IOPRIO_BE_NR
; i
++) {
3683 if (cfqd
->async_cfqq
[0][i
])
3684 cfq_put_queue(cfqd
->async_cfqq
[0][i
]);
3685 if (cfqd
->async_cfqq
[1][i
])
3686 cfq_put_queue(cfqd
->async_cfqq
[1][i
]);
3689 if (cfqd
->async_idle_cfqq
)
3690 cfq_put_queue(cfqd
->async_idle_cfqq
);
3693 static void cfq_cfqd_free(struct rcu_head
*head
)
3695 kfree(container_of(head
, struct cfq_data
, rcu
));
3698 static void cfq_exit_queue(struct elevator_queue
*e
)
3700 struct cfq_data
*cfqd
= e
->elevator_data
;
3701 struct request_queue
*q
= cfqd
->queue
;
3703 cfq_shutdown_timer_wq(cfqd
);
3705 spin_lock_irq(q
->queue_lock
);
3707 if (cfqd
->active_queue
)
3708 __cfq_slice_expired(cfqd
, cfqd
->active_queue
, 0);
3710 while (!list_empty(&cfqd
->cic_list
)) {
3711 struct cfq_io_context
*cic
= list_entry(cfqd
->cic_list
.next
,
3712 struct cfq_io_context
,
3715 __cfq_exit_single_io_context(cfqd
, cic
);
3718 cfq_put_async_queues(cfqd
);
3719 cfq_release_cfq_groups(cfqd
);
3720 blkiocg_del_blkio_group(&cfqd
->root_group
.blkg
);
3722 spin_unlock_irq(q
->queue_lock
);
3724 cfq_shutdown_timer_wq(cfqd
);
3726 /* Wait for cfqg->blkg->key accessors to exit their grace periods. */
3727 call_rcu(&cfqd
->rcu
, cfq_cfqd_free
);
3730 static void *cfq_init_queue(struct request_queue
*q
)
3732 struct cfq_data
*cfqd
;
3734 struct cfq_group
*cfqg
;
3735 struct cfq_rb_root
*st
;
3737 cfqd
= kmalloc_node(sizeof(*cfqd
), GFP_KERNEL
| __GFP_ZERO
, q
->node
);
3741 /* Init root service tree */
3742 cfqd
->grp_service_tree
= CFQ_RB_ROOT
;
3744 /* Init root group */
3745 cfqg
= &cfqd
->root_group
;
3746 for_each_cfqg_st(cfqg
, i
, j
, st
)
3748 RB_CLEAR_NODE(&cfqg
->rb_node
);
3750 /* Give preference to root group over other groups */
3751 cfqg
->weight
= 2*BLKIO_WEIGHT_DEFAULT
;
3753 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3755 * Take a reference to root group which we never drop. This is just
3756 * to make sure that cfq_put_cfqg() does not try to kfree root group
3758 atomic_set(&cfqg
->ref
, 1);
3760 blkiocg_add_blkio_group(&blkio_root_cgroup
, &cfqg
->blkg
, (void *)cfqd
,
3765 * Not strictly needed (since RB_ROOT just clears the node and we
3766 * zeroed cfqd on alloc), but better be safe in case someone decides
3767 * to add magic to the rb code
3769 for (i
= 0; i
< CFQ_PRIO_LISTS
; i
++)
3770 cfqd
->prio_trees
[i
] = RB_ROOT
;
3773 * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
3774 * Grab a permanent reference to it, so that the normal code flow
3775 * will not attempt to free it.
3777 cfq_init_cfqq(cfqd
, &cfqd
->oom_cfqq
, 1, 0);
3778 atomic_inc(&cfqd
->oom_cfqq
.ref
);
3779 cfq_link_cfqq_cfqg(&cfqd
->oom_cfqq
, &cfqd
->root_group
);
3781 INIT_LIST_HEAD(&cfqd
->cic_list
);
3785 init_timer(&cfqd
->idle_slice_timer
);
3786 cfqd
->idle_slice_timer
.function
= cfq_idle_slice_timer
;
3787 cfqd
->idle_slice_timer
.data
= (unsigned long) cfqd
;
3789 INIT_WORK(&cfqd
->unplug_work
, cfq_kick_queue
);
3791 cfqd
->cfq_quantum
= cfq_quantum
;
3792 cfqd
->cfq_fifo_expire
[0] = cfq_fifo_expire
[0];
3793 cfqd
->cfq_fifo_expire
[1] = cfq_fifo_expire
[1];
3794 cfqd
->cfq_back_max
= cfq_back_max
;
3795 cfqd
->cfq_back_penalty
= cfq_back_penalty
;
3796 cfqd
->cfq_slice
[0] = cfq_slice_async
;
3797 cfqd
->cfq_slice
[1] = cfq_slice_sync
;
3798 cfqd
->cfq_slice_async_rq
= cfq_slice_async_rq
;
3799 cfqd
->cfq_slice_idle
= cfq_slice_idle
;
3800 cfqd
->cfq_latency
= 1;
3801 cfqd
->cfq_group_isolation
= 0;
3804 * we optimistically start assuming sync ops weren't delayed in last
3805 * second, in order to have larger depth for async operations.
3807 cfqd
->last_delayed_sync
= jiffies
- HZ
;
3811 static void cfq_slab_kill(void)
3814 * Caller already ensured that pending RCU callbacks are completed,
3815 * so we should have no busy allocations at this point.
3818 kmem_cache_destroy(cfq_pool
);
3820 kmem_cache_destroy(cfq_ioc_pool
);
3823 static int __init
cfq_slab_setup(void)
3825 cfq_pool
= KMEM_CACHE(cfq_queue
, 0);
3829 cfq_ioc_pool
= KMEM_CACHE(cfq_io_context
, 0);
3840 * sysfs parts below -->
3843 cfq_var_show(unsigned int var
, char *page
)
3845 return sprintf(page
, "%d\n", var
);
3849 cfq_var_store(unsigned int *var
, const char *page
, size_t count
)
3851 char *p
= (char *) page
;
3853 *var
= simple_strtoul(p
, &p
, 10);
3857 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
3858 static ssize_t __FUNC(struct elevator_queue *e, char *page) \
3860 struct cfq_data *cfqd = e->elevator_data; \
3861 unsigned int __data = __VAR; \
3863 __data = jiffies_to_msecs(__data); \
3864 return cfq_var_show(__data, (page)); \
3866 SHOW_FUNCTION(cfq_quantum_show
, cfqd
->cfq_quantum
, 0);
3867 SHOW_FUNCTION(cfq_fifo_expire_sync_show
, cfqd
->cfq_fifo_expire
[1], 1);
3868 SHOW_FUNCTION(cfq_fifo_expire_async_show
, cfqd
->cfq_fifo_expire
[0], 1);
3869 SHOW_FUNCTION(cfq_back_seek_max_show
, cfqd
->cfq_back_max
, 0);
3870 SHOW_FUNCTION(cfq_back_seek_penalty_show
, cfqd
->cfq_back_penalty
, 0);
3871 SHOW_FUNCTION(cfq_slice_idle_show
, cfqd
->cfq_slice_idle
, 1);
3872 SHOW_FUNCTION(cfq_slice_sync_show
, cfqd
->cfq_slice
[1], 1);
3873 SHOW_FUNCTION(cfq_slice_async_show
, cfqd
->cfq_slice
[0], 1);
3874 SHOW_FUNCTION(cfq_slice_async_rq_show
, cfqd
->cfq_slice_async_rq
, 0);
3875 SHOW_FUNCTION(cfq_low_latency_show
, cfqd
->cfq_latency
, 0);
3876 SHOW_FUNCTION(cfq_group_isolation_show
, cfqd
->cfq_group_isolation
, 0);
3877 #undef SHOW_FUNCTION
3879 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
3880 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
3882 struct cfq_data *cfqd = e->elevator_data; \
3883 unsigned int __data; \
3884 int ret = cfq_var_store(&__data, (page), count); \
3885 if (__data < (MIN)) \
3887 else if (__data > (MAX)) \
3890 *(__PTR) = msecs_to_jiffies(__data); \
3892 *(__PTR) = __data; \
3895 STORE_FUNCTION(cfq_quantum_store
, &cfqd
->cfq_quantum
, 1, UINT_MAX
, 0);
3896 STORE_FUNCTION(cfq_fifo_expire_sync_store
, &cfqd
->cfq_fifo_expire
[1], 1,
3898 STORE_FUNCTION(cfq_fifo_expire_async_store
, &cfqd
->cfq_fifo_expire
[0], 1,
3900 STORE_FUNCTION(cfq_back_seek_max_store
, &cfqd
->cfq_back_max
, 0, UINT_MAX
, 0);
3901 STORE_FUNCTION(cfq_back_seek_penalty_store
, &cfqd
->cfq_back_penalty
, 1,
3903 STORE_FUNCTION(cfq_slice_idle_store
, &cfqd
->cfq_slice_idle
, 0, UINT_MAX
, 1);
3904 STORE_FUNCTION(cfq_slice_sync_store
, &cfqd
->cfq_slice
[1], 1, UINT_MAX
, 1);
3905 STORE_FUNCTION(cfq_slice_async_store
, &cfqd
->cfq_slice
[0], 1, UINT_MAX
, 1);
3906 STORE_FUNCTION(cfq_slice_async_rq_store
, &cfqd
->cfq_slice_async_rq
, 1,
3908 STORE_FUNCTION(cfq_low_latency_store
, &cfqd
->cfq_latency
, 0, 1, 0);
3909 STORE_FUNCTION(cfq_group_isolation_store
, &cfqd
->cfq_group_isolation
, 0, 1, 0);
3910 #undef STORE_FUNCTION
3912 #define CFQ_ATTR(name) \
3913 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
3915 static struct elv_fs_entry cfq_attrs
[] = {
3917 CFQ_ATTR(fifo_expire_sync
),
3918 CFQ_ATTR(fifo_expire_async
),
3919 CFQ_ATTR(back_seek_max
),
3920 CFQ_ATTR(back_seek_penalty
),
3921 CFQ_ATTR(slice_sync
),
3922 CFQ_ATTR(slice_async
),
3923 CFQ_ATTR(slice_async_rq
),
3924 CFQ_ATTR(slice_idle
),
3925 CFQ_ATTR(low_latency
),
3926 CFQ_ATTR(group_isolation
),
3930 static struct elevator_type iosched_cfq
= {
3932 .elevator_merge_fn
= cfq_merge
,
3933 .elevator_merged_fn
= cfq_merged_request
,
3934 .elevator_merge_req_fn
= cfq_merged_requests
,
3935 .elevator_allow_merge_fn
= cfq_allow_merge
,
3936 .elevator_bio_merged_fn
= cfq_bio_merged
,
3937 .elevator_dispatch_fn
= cfq_dispatch_requests
,
3938 .elevator_add_req_fn
= cfq_insert_request
,
3939 .elevator_activate_req_fn
= cfq_activate_request
,
3940 .elevator_deactivate_req_fn
= cfq_deactivate_request
,
3941 .elevator_queue_empty_fn
= cfq_queue_empty
,
3942 .elevator_completed_req_fn
= cfq_completed_request
,
3943 .elevator_former_req_fn
= elv_rb_former_request
,
3944 .elevator_latter_req_fn
= elv_rb_latter_request
,
3945 .elevator_set_req_fn
= cfq_set_request
,
3946 .elevator_put_req_fn
= cfq_put_request
,
3947 .elevator_may_queue_fn
= cfq_may_queue
,
3948 .elevator_init_fn
= cfq_init_queue
,
3949 .elevator_exit_fn
= cfq_exit_queue
,
3950 .trim
= cfq_free_io_context
,
3952 .elevator_attrs
= cfq_attrs
,
3953 .elevator_name
= "cfq",
3954 .elevator_owner
= THIS_MODULE
,
3957 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3958 static struct blkio_policy_type blkio_policy_cfq
= {
3960 .blkio_unlink_group_fn
= cfq_unlink_blkio_group
,
3961 .blkio_update_group_weight_fn
= cfq_update_blkio_group_weight
,
3965 static struct blkio_policy_type blkio_policy_cfq
;
3968 static int __init
cfq_init(void)
3971 * could be 0 on HZ < 1000 setups
3973 if (!cfq_slice_async
)
3974 cfq_slice_async
= 1;
3975 if (!cfq_slice_idle
)
3978 if (cfq_slab_setup())
3981 elv_register(&iosched_cfq
);
3982 blkio_policy_register(&blkio_policy_cfq
);
3987 static void __exit
cfq_exit(void)
3989 DECLARE_COMPLETION_ONSTACK(all_gone
);
3990 blkio_policy_unregister(&blkio_policy_cfq
);
3991 elv_unregister(&iosched_cfq
);
3992 ioc_gone
= &all_gone
;
3993 /* ioc_gone's update must be visible before reading ioc_count */
3997 * this also protects us from entering cfq_slab_kill() with
3998 * pending RCU callbacks
4000 if (elv_ioc_count_read(cfq_ioc_count
))
4001 wait_for_completion(&all_gone
);
4005 module_init(cfq_init
);
4006 module_exit(cfq_exit
);
4008 MODULE_AUTHOR("Jens Axboe");
4009 MODULE_LICENSE("GPL");
4010 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");