2 * CFQ, or complete fairness queueing, disk scheduler.
4 * Based on ideas from a previously unfinished io
5 * scheduler (round robin per-process disk scheduling) and Andrea Arcangeli.
7 * Copyright (C) 2003 Jens Axboe <axboe@kernel.dk>
9 #include <linux/module.h>
10 #include <linux/blkdev.h>
11 #include <linux/elevator.h>
12 #include <linux/jiffies.h>
13 #include <linux/rbtree.h>
14 #include <linux/ioprio.h>
15 #include <linux/blktrace_api.h>
16 #include "blk-cgroup.h"
21 /* max queue in one round of service */
22 static const int cfq_quantum
= 8;
23 static const int cfq_fifo_expire
[2] = { HZ
/ 4, HZ
/ 8 };
24 /* maximum backwards seek, in KiB */
25 static const int cfq_back_max
= 16 * 1024;
26 /* penalty of a backwards seek */
27 static const int cfq_back_penalty
= 2;
28 static const int cfq_slice_sync
= HZ
/ 10;
29 static int cfq_slice_async
= HZ
/ 25;
30 static const int cfq_slice_async_rq
= 2;
31 static int cfq_slice_idle
= HZ
/ 125;
32 static const int cfq_target_latency
= HZ
* 3/10; /* 300 ms */
33 static const int cfq_hist_divisor
= 4;
36 * offset from end of service tree
38 #define CFQ_IDLE_DELAY (HZ / 5)
41 * below this threshold, we consider thinktime immediate
43 #define CFQ_MIN_TT (2)
45 #define CFQ_SLICE_SCALE (5)
46 #define CFQ_HW_QUEUE_MIN (5)
47 #define CFQ_SERVICE_SHIFT 12
49 #define CFQQ_SEEK_THR (sector_t)(8 * 100)
50 #define CFQQ_CLOSE_THR (sector_t)(8 * 1024)
51 #define CFQQ_SECT_THR_NONROT (sector_t)(2 * 32)
52 #define CFQQ_SEEKY(cfqq) (hweight32(cfqq->seek_history) > 32/8)
55 ((struct cfq_io_context *) (rq)->elevator_private)
56 #define RQ_CFQQ(rq) (struct cfq_queue *) ((rq)->elevator_private2)
58 static struct kmem_cache
*cfq_pool
;
59 static struct kmem_cache
*cfq_ioc_pool
;
61 static DEFINE_PER_CPU(unsigned long, cfq_ioc_count
);
62 static struct completion
*ioc_gone
;
63 static DEFINE_SPINLOCK(ioc_gone_lock
);
65 #define CFQ_PRIO_LISTS IOPRIO_BE_NR
66 #define cfq_class_idle(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
67 #define cfq_class_rt(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
69 #define sample_valid(samples) ((samples) > 80)
70 #define rb_entry_cfqg(node) rb_entry((node), struct cfq_group, rb_node)
73 * Most of our rbtree usage is for sorting with min extraction, so
74 * if we cache the leftmost node we don't have to walk down the tree
75 * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should
76 * move this into the elevator for the rq sorting as well.
82 unsigned total_weight
;
84 struct rb_node
*active
;
86 #define CFQ_RB_ROOT (struct cfq_rb_root) { .rb = RB_ROOT, .left = NULL, \
87 .count = 0, .min_vdisktime = 0, }
90 * Per process-grouping structure
95 /* various state flags, see below */
98 struct cfq_data
*cfqd
;
99 /* service_tree member */
100 struct rb_node rb_node
;
101 /* service_tree key */
102 unsigned long rb_key
;
103 /* prio tree member */
104 struct rb_node p_node
;
105 /* prio tree root we belong to, if any */
106 struct rb_root
*p_root
;
107 /* sorted list of pending requests */
108 struct rb_root sort_list
;
109 /* if fifo isn't expired, next request to serve */
110 struct request
*next_rq
;
111 /* requests queued in sort_list */
113 /* currently allocated requests */
115 /* fifo list of requests in sort_list */
116 struct list_head fifo
;
118 /* time when queue got scheduled in to dispatch first request. */
119 unsigned long dispatch_start
;
120 unsigned int allocated_slice
;
121 unsigned int slice_dispatch
;
122 /* time when first request from queue completed and slice started. */
123 unsigned long slice_start
;
124 unsigned long slice_end
;
127 /* pending metadata requests */
129 /* number of requests that are on the dispatch list or inside driver */
132 /* io prio of this group */
133 unsigned short ioprio
, org_ioprio
;
134 unsigned short ioprio_class
, org_ioprio_class
;
139 sector_t last_request_pos
;
141 struct cfq_rb_root
*service_tree
;
142 struct cfq_queue
*new_cfqq
;
143 struct cfq_group
*cfqg
;
144 struct cfq_group
*orig_cfqg
;
145 /* Sectors dispatched in current dispatch round */
146 unsigned long nr_sectors
;
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_DEBUG_CFQ_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
;
434 * We regard a request as SYNC, if it's either a read or has the SYNC bit
435 * set (in which case it could also be direct WRITE).
437 static inline bool cfq_bio_sync(struct bio
*bio
)
439 return bio_data_dir(bio
) == READ
|| bio_rw_flagged(bio
, BIO_RW_SYNCIO
);
443 * scheduler run of queue, if there are requests pending and no one in the
444 * driver that will restart queueing
446 static inline void cfq_schedule_dispatch(struct cfq_data
*cfqd
)
448 if (cfqd
->busy_queues
) {
449 cfq_log(cfqd
, "schedule dispatch");
450 kblockd_schedule_work(cfqd
->queue
, &cfqd
->unplug_work
);
454 static int cfq_queue_empty(struct request_queue
*q
)
456 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
458 return !cfqd
->rq_queued
;
462 * Scale schedule slice based on io priority. Use the sync time slice only
463 * if a queue is marked sync and has sync io queued. A sync queue with async
464 * io only, should not get full sync slice length.
466 static inline int cfq_prio_slice(struct cfq_data
*cfqd
, bool sync
,
469 const int base_slice
= cfqd
->cfq_slice
[sync
];
471 WARN_ON(prio
>= IOPRIO_BE_NR
);
473 return base_slice
+ (base_slice
/CFQ_SLICE_SCALE
* (4 - prio
));
477 cfq_prio_to_slice(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
479 return cfq_prio_slice(cfqd
, cfq_cfqq_sync(cfqq
), cfqq
->ioprio
);
482 static inline u64
cfq_scale_slice(unsigned long delta
, struct cfq_group
*cfqg
)
484 u64 d
= delta
<< CFQ_SERVICE_SHIFT
;
486 d
= d
* BLKIO_WEIGHT_DEFAULT
;
487 do_div(d
, cfqg
->weight
);
491 static inline u64
max_vdisktime(u64 min_vdisktime
, u64 vdisktime
)
493 s64 delta
= (s64
)(vdisktime
- min_vdisktime
);
495 min_vdisktime
= vdisktime
;
497 return min_vdisktime
;
500 static inline u64
min_vdisktime(u64 min_vdisktime
, u64 vdisktime
)
502 s64 delta
= (s64
)(vdisktime
- min_vdisktime
);
504 min_vdisktime
= vdisktime
;
506 return min_vdisktime
;
509 static void update_min_vdisktime(struct cfq_rb_root
*st
)
511 u64 vdisktime
= st
->min_vdisktime
;
512 struct cfq_group
*cfqg
;
515 cfqg
= rb_entry_cfqg(st
->active
);
516 vdisktime
= cfqg
->vdisktime
;
520 cfqg
= rb_entry_cfqg(st
->left
);
521 vdisktime
= min_vdisktime(vdisktime
, cfqg
->vdisktime
);
524 st
->min_vdisktime
= max_vdisktime(st
->min_vdisktime
, vdisktime
);
528 * get averaged number of queues of RT/BE priority.
529 * average is updated, with a formula that gives more weight to higher numbers,
530 * to quickly follows sudden increases and decrease slowly
533 static inline unsigned cfq_group_get_avg_queues(struct cfq_data
*cfqd
,
534 struct cfq_group
*cfqg
, bool rt
)
536 unsigned min_q
, max_q
;
537 unsigned mult
= cfq_hist_divisor
- 1;
538 unsigned round
= cfq_hist_divisor
/ 2;
539 unsigned busy
= cfq_group_busy_queues_wl(rt
, cfqd
, cfqg
);
541 min_q
= min(cfqg
->busy_queues_avg
[rt
], busy
);
542 max_q
= max(cfqg
->busy_queues_avg
[rt
], busy
);
543 cfqg
->busy_queues_avg
[rt
] = (mult
* max_q
+ min_q
+ round
) /
545 return cfqg
->busy_queues_avg
[rt
];
548 static inline unsigned
549 cfq_group_slice(struct cfq_data
*cfqd
, struct cfq_group
*cfqg
)
551 struct cfq_rb_root
*st
= &cfqd
->grp_service_tree
;
553 return cfq_target_latency
* cfqg
->weight
/ st
->total_weight
;
557 cfq_set_prio_slice(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
559 unsigned slice
= cfq_prio_to_slice(cfqd
, cfqq
);
560 if (cfqd
->cfq_latency
) {
562 * interested queues (we consider only the ones with the same
563 * priority class in the cfq group)
565 unsigned iq
= cfq_group_get_avg_queues(cfqd
, cfqq
->cfqg
,
567 unsigned sync_slice
= cfqd
->cfq_slice
[1];
568 unsigned expect_latency
= sync_slice
* iq
;
569 unsigned group_slice
= cfq_group_slice(cfqd
, cfqq
->cfqg
);
571 if (expect_latency
> group_slice
) {
572 unsigned base_low_slice
= 2 * cfqd
->cfq_slice_idle
;
573 /* scale low_slice according to IO priority
574 * and sync vs async */
576 min(slice
, base_low_slice
* slice
/ sync_slice
);
577 /* the adapted slice value is scaled to fit all iqs
578 * into the target latency */
579 slice
= max(slice
* group_slice
/ expect_latency
,
583 cfqq
->slice_start
= jiffies
;
584 cfqq
->slice_end
= jiffies
+ slice
;
585 cfqq
->allocated_slice
= slice
;
586 cfq_log_cfqq(cfqd
, cfqq
, "set_slice=%lu", cfqq
->slice_end
- jiffies
);
590 * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
591 * isn't valid until the first request from the dispatch is activated
592 * and the slice time set.
594 static inline bool cfq_slice_used(struct cfq_queue
*cfqq
)
596 if (cfq_cfqq_slice_new(cfqq
))
598 if (time_before(jiffies
, cfqq
->slice_end
))
605 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
606 * We choose the request that is closest to the head right now. Distance
607 * behind the head is penalized and only allowed to a certain extent.
609 static struct request
*
610 cfq_choose_req(struct cfq_data
*cfqd
, struct request
*rq1
, struct request
*rq2
, sector_t last
)
612 sector_t s1
, s2
, d1
= 0, d2
= 0;
613 unsigned long back_max
;
614 #define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */
615 #define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */
616 unsigned wrap
= 0; /* bit mask: requests behind the disk head? */
618 if (rq1
== NULL
|| rq1
== rq2
)
623 if (rq_is_sync(rq1
) && !rq_is_sync(rq2
))
625 else if (rq_is_sync(rq2
) && !rq_is_sync(rq1
))
627 if (rq_is_meta(rq1
) && !rq_is_meta(rq2
))
629 else if (rq_is_meta(rq2
) && !rq_is_meta(rq1
))
632 s1
= blk_rq_pos(rq1
);
633 s2
= blk_rq_pos(rq2
);
636 * by definition, 1KiB is 2 sectors
638 back_max
= cfqd
->cfq_back_max
* 2;
641 * Strict one way elevator _except_ in the case where we allow
642 * short backward seeks which are biased as twice the cost of a
643 * similar forward seek.
647 else if (s1
+ back_max
>= last
)
648 d1
= (last
- s1
) * cfqd
->cfq_back_penalty
;
650 wrap
|= CFQ_RQ1_WRAP
;
654 else if (s2
+ back_max
>= last
)
655 d2
= (last
- s2
) * cfqd
->cfq_back_penalty
;
657 wrap
|= CFQ_RQ2_WRAP
;
659 /* Found required data */
662 * By doing switch() on the bit mask "wrap" we avoid having to
663 * check two variables for all permutations: --> faster!
666 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
682 case (CFQ_RQ1_WRAP
|CFQ_RQ2_WRAP
): /* both rqs wrapped */
685 * Since both rqs are wrapped,
686 * start with the one that's further behind head
687 * (--> only *one* back seek required),
688 * since back seek takes more time than forward.
698 * The below is leftmost cache rbtree addon
700 static struct cfq_queue
*cfq_rb_first(struct cfq_rb_root
*root
)
702 /* Service tree is empty */
707 root
->left
= rb_first(&root
->rb
);
710 return rb_entry(root
->left
, struct cfq_queue
, rb_node
);
715 static struct cfq_group
*cfq_rb_first_group(struct cfq_rb_root
*root
)
718 root
->left
= rb_first(&root
->rb
);
721 return rb_entry_cfqg(root
->left
);
726 static void rb_erase_init(struct rb_node
*n
, struct rb_root
*root
)
732 static void cfq_rb_erase(struct rb_node
*n
, struct cfq_rb_root
*root
)
736 rb_erase_init(n
, &root
->rb
);
741 * would be nice to take fifo expire time into account as well
743 static struct request
*
744 cfq_find_next_rq(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
745 struct request
*last
)
747 struct rb_node
*rbnext
= rb_next(&last
->rb_node
);
748 struct rb_node
*rbprev
= rb_prev(&last
->rb_node
);
749 struct request
*next
= NULL
, *prev
= NULL
;
751 BUG_ON(RB_EMPTY_NODE(&last
->rb_node
));
754 prev
= rb_entry_rq(rbprev
);
757 next
= rb_entry_rq(rbnext
);
759 rbnext
= rb_first(&cfqq
->sort_list
);
760 if (rbnext
&& rbnext
!= &last
->rb_node
)
761 next
= rb_entry_rq(rbnext
);
764 return cfq_choose_req(cfqd
, next
, prev
, blk_rq_pos(last
));
767 static unsigned long cfq_slice_offset(struct cfq_data
*cfqd
,
768 struct cfq_queue
*cfqq
)
771 * just an approximation, should be ok.
773 return (cfqq
->cfqg
->nr_cfqq
- 1) * (cfq_prio_slice(cfqd
, 1, 0) -
774 cfq_prio_slice(cfqd
, cfq_cfqq_sync(cfqq
), cfqq
->ioprio
));
778 cfqg_key(struct cfq_rb_root
*st
, struct cfq_group
*cfqg
)
780 return cfqg
->vdisktime
- st
->min_vdisktime
;
784 __cfq_group_service_tree_add(struct cfq_rb_root
*st
, struct cfq_group
*cfqg
)
786 struct rb_node
**node
= &st
->rb
.rb_node
;
787 struct rb_node
*parent
= NULL
;
788 struct cfq_group
*__cfqg
;
789 s64 key
= cfqg_key(st
, cfqg
);
792 while (*node
!= NULL
) {
794 __cfqg
= rb_entry_cfqg(parent
);
796 if (key
< cfqg_key(st
, __cfqg
))
797 node
= &parent
->rb_left
;
799 node
= &parent
->rb_right
;
805 st
->left
= &cfqg
->rb_node
;
807 rb_link_node(&cfqg
->rb_node
, parent
, node
);
808 rb_insert_color(&cfqg
->rb_node
, &st
->rb
);
812 cfq_group_service_tree_add(struct cfq_data
*cfqd
, struct cfq_group
*cfqg
)
814 struct cfq_rb_root
*st
= &cfqd
->grp_service_tree
;
815 struct cfq_group
*__cfqg
;
823 * Currently put the group at the end. Later implement something
824 * so that groups get lesser vtime based on their weights, so that
825 * if group does not loose all if it was not continously backlogged.
827 n
= rb_last(&st
->rb
);
829 __cfqg
= rb_entry_cfqg(n
);
830 cfqg
->vdisktime
= __cfqg
->vdisktime
+ CFQ_IDLE_DELAY
;
832 cfqg
->vdisktime
= st
->min_vdisktime
;
834 __cfq_group_service_tree_add(st
, cfqg
);
836 st
->total_weight
+= cfqg
->weight
;
840 cfq_group_service_tree_del(struct cfq_data
*cfqd
, struct cfq_group
*cfqg
)
842 struct cfq_rb_root
*st
= &cfqd
->grp_service_tree
;
844 if (st
->active
== &cfqg
->rb_node
)
847 BUG_ON(cfqg
->nr_cfqq
< 1);
850 /* If there are other cfq queues under this group, don't delete it */
854 cfq_log_cfqg(cfqd
, cfqg
, "del_from_rr group");
856 st
->total_weight
-= cfqg
->weight
;
857 if (!RB_EMPTY_NODE(&cfqg
->rb_node
))
858 cfq_rb_erase(&cfqg
->rb_node
, st
);
859 cfqg
->saved_workload_slice
= 0;
860 blkiocg_update_blkio_group_dequeue_stats(&cfqg
->blkg
, 1);
863 static inline unsigned int cfq_cfqq_slice_usage(struct cfq_queue
*cfqq
)
865 unsigned int slice_used
;
868 * Queue got expired before even a single request completed or
869 * got expired immediately after first request completion.
871 if (!cfqq
->slice_start
|| cfqq
->slice_start
== jiffies
) {
873 * Also charge the seek time incurred to the group, otherwise
874 * if there are mutiple queues in the group, each can dispatch
875 * a single request on seeky media and cause lots of seek time
876 * and group will never know it.
878 slice_used
= max_t(unsigned, (jiffies
- cfqq
->dispatch_start
),
881 slice_used
= jiffies
- cfqq
->slice_start
;
882 if (slice_used
> cfqq
->allocated_slice
)
883 slice_used
= cfqq
->allocated_slice
;
886 cfq_log_cfqq(cfqq
->cfqd
, cfqq
, "sl_used=%u sect=%lu", slice_used
,
891 static void cfq_group_served(struct cfq_data
*cfqd
, struct cfq_group
*cfqg
,
892 struct cfq_queue
*cfqq
)
894 struct cfq_rb_root
*st
= &cfqd
->grp_service_tree
;
895 unsigned int used_sl
, charge_sl
;
896 int nr_sync
= cfqg
->nr_cfqq
- cfqg_busy_async_queues(cfqd
, cfqg
)
897 - cfqg
->service_tree_idle
.count
;
900 used_sl
= charge_sl
= cfq_cfqq_slice_usage(cfqq
);
902 if (!cfq_cfqq_sync(cfqq
) && !nr_sync
)
903 charge_sl
= cfqq
->allocated_slice
;
905 /* Can't update vdisktime while group is on service tree */
906 cfq_rb_erase(&cfqg
->rb_node
, st
);
907 cfqg
->vdisktime
+= cfq_scale_slice(charge_sl
, cfqg
);
908 __cfq_group_service_tree_add(st
, cfqg
);
910 /* This group is being expired. Save the context */
911 if (time_after(cfqd
->workload_expires
, jiffies
)) {
912 cfqg
->saved_workload_slice
= cfqd
->workload_expires
914 cfqg
->saved_workload
= cfqd
->serving_type
;
915 cfqg
->saved_serving_prio
= cfqd
->serving_prio
;
917 cfqg
->saved_workload_slice
= 0;
919 cfq_log_cfqg(cfqd
, cfqg
, "served: vt=%llu min_vt=%llu", cfqg
->vdisktime
,
921 blkiocg_update_blkio_group_stats(&cfqg
->blkg
, used_sl
,
925 #ifdef CONFIG_CFQ_GROUP_IOSCHED
926 static inline struct cfq_group
*cfqg_of_blkg(struct blkio_group
*blkg
)
929 return container_of(blkg
, struct cfq_group
, blkg
);
934 cfq_update_blkio_group_weight(struct blkio_group
*blkg
, unsigned int weight
)
936 cfqg_of_blkg(blkg
)->weight
= weight
;
939 static struct cfq_group
*
940 cfq_find_alloc_cfqg(struct cfq_data
*cfqd
, struct cgroup
*cgroup
, int create
)
942 struct blkio_cgroup
*blkcg
= cgroup_to_blkio_cgroup(cgroup
);
943 struct cfq_group
*cfqg
= NULL
;
946 struct cfq_rb_root
*st
;
947 struct backing_dev_info
*bdi
= &cfqd
->queue
->backing_dev_info
;
948 unsigned int major
, minor
;
950 cfqg
= cfqg_of_blkg(blkiocg_lookup_group(blkcg
, key
));
954 cfqg
= kzalloc_node(sizeof(*cfqg
), GFP_ATOMIC
, cfqd
->queue
->node
);
958 cfqg
->weight
= blkcg
->weight
;
959 for_each_cfqg_st(cfqg
, i
, j
, st
)
961 RB_CLEAR_NODE(&cfqg
->rb_node
);
964 * Take the initial reference that will be released on destroy
965 * This can be thought of a joint reference by cgroup and
966 * elevator which will be dropped by either elevator exit
967 * or cgroup deletion path depending on who is exiting first.
969 atomic_set(&cfqg
->ref
, 1);
971 /* Add group onto cgroup list */
972 sscanf(dev_name(bdi
->dev
), "%u:%u", &major
, &minor
);
973 blkiocg_add_blkio_group(blkcg
, &cfqg
->blkg
, (void *)cfqd
,
974 MKDEV(major
, minor
));
976 /* Add group on cfqd list */
977 hlist_add_head(&cfqg
->cfqd_node
, &cfqd
->cfqg_list
);
984 * Search for the cfq group current task belongs to. If create = 1, then also
985 * create the cfq group if it does not exist. request_queue lock must be held.
987 static struct cfq_group
*cfq_get_cfqg(struct cfq_data
*cfqd
, int create
)
989 struct cgroup
*cgroup
;
990 struct cfq_group
*cfqg
= NULL
;
993 cgroup
= task_cgroup(current
, blkio_subsys_id
);
994 cfqg
= cfq_find_alloc_cfqg(cfqd
, cgroup
, create
);
996 cfqg
= &cfqd
->root_group
;
1001 static void cfq_link_cfqq_cfqg(struct cfq_queue
*cfqq
, struct cfq_group
*cfqg
)
1003 /* Currently, all async queues are mapped to root group */
1004 if (!cfq_cfqq_sync(cfqq
))
1005 cfqg
= &cfqq
->cfqd
->root_group
;
1008 /* cfqq reference on cfqg */
1009 atomic_inc(&cfqq
->cfqg
->ref
);
1012 static void cfq_put_cfqg(struct cfq_group
*cfqg
)
1014 struct cfq_rb_root
*st
;
1017 BUG_ON(atomic_read(&cfqg
->ref
) <= 0);
1018 if (!atomic_dec_and_test(&cfqg
->ref
))
1020 for_each_cfqg_st(cfqg
, i
, j
, st
)
1021 BUG_ON(!RB_EMPTY_ROOT(&st
->rb
) || st
->active
!= NULL
);
1025 static void cfq_destroy_cfqg(struct cfq_data
*cfqd
, struct cfq_group
*cfqg
)
1027 /* Something wrong if we are trying to remove same group twice */
1028 BUG_ON(hlist_unhashed(&cfqg
->cfqd_node
));
1030 hlist_del_init(&cfqg
->cfqd_node
);
1033 * Put the reference taken at the time of creation so that when all
1034 * queues are gone, group can be destroyed.
1039 static void cfq_release_cfq_groups(struct cfq_data
*cfqd
)
1041 struct hlist_node
*pos
, *n
;
1042 struct cfq_group
*cfqg
;
1044 hlist_for_each_entry_safe(cfqg
, pos
, n
, &cfqd
->cfqg_list
, cfqd_node
) {
1046 * If cgroup removal path got to blk_group first and removed
1047 * it from cgroup list, then it will take care of destroying
1050 if (!blkiocg_del_blkio_group(&cfqg
->blkg
))
1051 cfq_destroy_cfqg(cfqd
, cfqg
);
1056 * Blk cgroup controller notification saying that blkio_group object is being
1057 * delinked as associated cgroup object is going away. That also means that
1058 * no new IO will come in this group. So get rid of this group as soon as
1059 * any pending IO in the group is finished.
1061 * This function is called under rcu_read_lock(). key is the rcu protected
1062 * pointer. That means "key" is a valid cfq_data pointer as long as we are rcu
1065 * "key" was fetched from blkio_group under blkio_cgroup->lock. That means
1066 * it should not be NULL as even if elevator was exiting, cgroup deltion
1067 * path got to it first.
1069 void cfq_unlink_blkio_group(void *key
, struct blkio_group
*blkg
)
1071 unsigned long flags
;
1072 struct cfq_data
*cfqd
= key
;
1074 spin_lock_irqsave(cfqd
->queue
->queue_lock
, flags
);
1075 cfq_destroy_cfqg(cfqd
, cfqg_of_blkg(blkg
));
1076 spin_unlock_irqrestore(cfqd
->queue
->queue_lock
, flags
);
1079 #else /* GROUP_IOSCHED */
1080 static struct cfq_group
*cfq_get_cfqg(struct cfq_data
*cfqd
, int create
)
1082 return &cfqd
->root_group
;
1085 cfq_link_cfqq_cfqg(struct cfq_queue
*cfqq
, struct cfq_group
*cfqg
) {
1089 static void cfq_release_cfq_groups(struct cfq_data
*cfqd
) {}
1090 static inline void cfq_put_cfqg(struct cfq_group
*cfqg
) {}
1092 #endif /* GROUP_IOSCHED */
1095 * The cfqd->service_trees holds all pending cfq_queue's that have
1096 * requests waiting to be processed. It is sorted in the order that
1097 * we will service the queues.
1099 static void cfq_service_tree_add(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
1102 struct rb_node
**p
, *parent
;
1103 struct cfq_queue
*__cfqq
;
1104 unsigned long rb_key
;
1105 struct cfq_rb_root
*service_tree
;
1108 int group_changed
= 0;
1110 #ifdef CONFIG_CFQ_GROUP_IOSCHED
1111 if (!cfqd
->cfq_group_isolation
1112 && cfqq_type(cfqq
) == SYNC_NOIDLE_WORKLOAD
1113 && cfqq
->cfqg
&& cfqq
->cfqg
!= &cfqd
->root_group
) {
1114 /* Move this cfq to root group */
1115 cfq_log_cfqq(cfqd
, cfqq
, "moving to root group");
1116 if (!RB_EMPTY_NODE(&cfqq
->rb_node
))
1117 cfq_group_service_tree_del(cfqd
, cfqq
->cfqg
);
1118 cfqq
->orig_cfqg
= cfqq
->cfqg
;
1119 cfqq
->cfqg
= &cfqd
->root_group
;
1120 atomic_inc(&cfqd
->root_group
.ref
);
1122 } else if (!cfqd
->cfq_group_isolation
1123 && cfqq_type(cfqq
) == SYNC_WORKLOAD
&& cfqq
->orig_cfqg
) {
1124 /* cfqq is sequential now needs to go to its original group */
1125 BUG_ON(cfqq
->cfqg
!= &cfqd
->root_group
);
1126 if (!RB_EMPTY_NODE(&cfqq
->rb_node
))
1127 cfq_group_service_tree_del(cfqd
, cfqq
->cfqg
);
1128 cfq_put_cfqg(cfqq
->cfqg
);
1129 cfqq
->cfqg
= cfqq
->orig_cfqg
;
1130 cfqq
->orig_cfqg
= NULL
;
1132 cfq_log_cfqq(cfqd
, cfqq
, "moved to origin group");
1136 service_tree
= service_tree_for(cfqq
->cfqg
, cfqq_prio(cfqq
),
1138 if (cfq_class_idle(cfqq
)) {
1139 rb_key
= CFQ_IDLE_DELAY
;
1140 parent
= rb_last(&service_tree
->rb
);
1141 if (parent
&& parent
!= &cfqq
->rb_node
) {
1142 __cfqq
= rb_entry(parent
, struct cfq_queue
, rb_node
);
1143 rb_key
+= __cfqq
->rb_key
;
1146 } else if (!add_front
) {
1148 * Get our rb key offset. Subtract any residual slice
1149 * value carried from last service. A negative resid
1150 * count indicates slice overrun, and this should position
1151 * the next service time further away in the tree.
1153 rb_key
= cfq_slice_offset(cfqd
, cfqq
) + jiffies
;
1154 rb_key
-= cfqq
->slice_resid
;
1155 cfqq
->slice_resid
= 0;
1158 __cfqq
= cfq_rb_first(service_tree
);
1159 rb_key
+= __cfqq
? __cfqq
->rb_key
: jiffies
;
1162 if (!RB_EMPTY_NODE(&cfqq
->rb_node
)) {
1165 * same position, nothing more to do
1167 if (rb_key
== cfqq
->rb_key
&&
1168 cfqq
->service_tree
== service_tree
)
1171 cfq_rb_erase(&cfqq
->rb_node
, cfqq
->service_tree
);
1172 cfqq
->service_tree
= NULL
;
1177 cfqq
->service_tree
= service_tree
;
1178 p
= &service_tree
->rb
.rb_node
;
1183 __cfqq
= rb_entry(parent
, struct cfq_queue
, rb_node
);
1186 * sort by key, that represents service time.
1188 if (time_before(rb_key
, __cfqq
->rb_key
))
1191 n
= &(*p
)->rb_right
;
1199 service_tree
->left
= &cfqq
->rb_node
;
1201 cfqq
->rb_key
= rb_key
;
1202 rb_link_node(&cfqq
->rb_node
, parent
, p
);
1203 rb_insert_color(&cfqq
->rb_node
, &service_tree
->rb
);
1204 service_tree
->count
++;
1205 if ((add_front
|| !new_cfqq
) && !group_changed
)
1207 cfq_group_service_tree_add(cfqd
, cfqq
->cfqg
);
1210 static struct cfq_queue
*
1211 cfq_prio_tree_lookup(struct cfq_data
*cfqd
, struct rb_root
*root
,
1212 sector_t sector
, struct rb_node
**ret_parent
,
1213 struct rb_node
***rb_link
)
1215 struct rb_node
**p
, *parent
;
1216 struct cfq_queue
*cfqq
= NULL
;
1224 cfqq
= rb_entry(parent
, struct cfq_queue
, p_node
);
1227 * Sort strictly based on sector. Smallest to the left,
1228 * largest to the right.
1230 if (sector
> blk_rq_pos(cfqq
->next_rq
))
1231 n
= &(*p
)->rb_right
;
1232 else if (sector
< blk_rq_pos(cfqq
->next_rq
))
1240 *ret_parent
= parent
;
1246 static void cfq_prio_tree_add(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
1248 struct rb_node
**p
, *parent
;
1249 struct cfq_queue
*__cfqq
;
1252 rb_erase(&cfqq
->p_node
, cfqq
->p_root
);
1253 cfqq
->p_root
= NULL
;
1256 if (cfq_class_idle(cfqq
))
1261 cfqq
->p_root
= &cfqd
->prio_trees
[cfqq
->org_ioprio
];
1262 __cfqq
= cfq_prio_tree_lookup(cfqd
, cfqq
->p_root
,
1263 blk_rq_pos(cfqq
->next_rq
), &parent
, &p
);
1265 rb_link_node(&cfqq
->p_node
, parent
, p
);
1266 rb_insert_color(&cfqq
->p_node
, cfqq
->p_root
);
1268 cfqq
->p_root
= NULL
;
1272 * Update cfqq's position in the service tree.
1274 static void cfq_resort_rr_list(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
1277 * Resorting requires the cfqq to be on the RR list already.
1279 if (cfq_cfqq_on_rr(cfqq
)) {
1280 cfq_service_tree_add(cfqd
, cfqq
, 0);
1281 cfq_prio_tree_add(cfqd
, cfqq
);
1286 * add to busy list of queues for service, trying to be fair in ordering
1287 * the pending list according to last request service
1289 static void cfq_add_cfqq_rr(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
1291 cfq_log_cfqq(cfqd
, cfqq
, "add_to_rr");
1292 BUG_ON(cfq_cfqq_on_rr(cfqq
));
1293 cfq_mark_cfqq_on_rr(cfqq
);
1294 cfqd
->busy_queues
++;
1296 cfq_resort_rr_list(cfqd
, cfqq
);
1300 * Called when the cfqq no longer has requests pending, remove it from
1303 static void cfq_del_cfqq_rr(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
1305 cfq_log_cfqq(cfqd
, cfqq
, "del_from_rr");
1306 BUG_ON(!cfq_cfqq_on_rr(cfqq
));
1307 cfq_clear_cfqq_on_rr(cfqq
);
1309 if (!RB_EMPTY_NODE(&cfqq
->rb_node
)) {
1310 cfq_rb_erase(&cfqq
->rb_node
, cfqq
->service_tree
);
1311 cfqq
->service_tree
= NULL
;
1314 rb_erase(&cfqq
->p_node
, cfqq
->p_root
);
1315 cfqq
->p_root
= NULL
;
1318 cfq_group_service_tree_del(cfqd
, cfqq
->cfqg
);
1319 BUG_ON(!cfqd
->busy_queues
);
1320 cfqd
->busy_queues
--;
1324 * rb tree support functions
1326 static void cfq_del_rq_rb(struct request
*rq
)
1328 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
1329 const int sync
= rq_is_sync(rq
);
1331 BUG_ON(!cfqq
->queued
[sync
]);
1332 cfqq
->queued
[sync
]--;
1334 elv_rb_del(&cfqq
->sort_list
, rq
);
1336 if (cfq_cfqq_on_rr(cfqq
) && RB_EMPTY_ROOT(&cfqq
->sort_list
)) {
1338 * Queue will be deleted from service tree when we actually
1339 * expire it later. Right now just remove it from prio tree
1343 rb_erase(&cfqq
->p_node
, cfqq
->p_root
);
1344 cfqq
->p_root
= NULL
;
1349 static void cfq_add_rq_rb(struct request
*rq
)
1351 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
1352 struct cfq_data
*cfqd
= cfqq
->cfqd
;
1353 struct request
*__alias
, *prev
;
1355 cfqq
->queued
[rq_is_sync(rq
)]++;
1358 * looks a little odd, but the first insert might return an alias.
1359 * if that happens, put the alias on the dispatch list
1361 while ((__alias
= elv_rb_add(&cfqq
->sort_list
, rq
)) != NULL
)
1362 cfq_dispatch_insert(cfqd
->queue
, __alias
);
1364 if (!cfq_cfqq_on_rr(cfqq
))
1365 cfq_add_cfqq_rr(cfqd
, cfqq
);
1368 * check if this request is a better next-serve candidate
1370 prev
= cfqq
->next_rq
;
1371 cfqq
->next_rq
= cfq_choose_req(cfqd
, cfqq
->next_rq
, rq
, cfqd
->last_position
);
1374 * adjust priority tree position, if ->next_rq changes
1376 if (prev
!= cfqq
->next_rq
)
1377 cfq_prio_tree_add(cfqd
, cfqq
);
1379 BUG_ON(!cfqq
->next_rq
);
1382 static void cfq_reposition_rq_rb(struct cfq_queue
*cfqq
, struct request
*rq
)
1384 elv_rb_del(&cfqq
->sort_list
, rq
);
1385 cfqq
->queued
[rq_is_sync(rq
)]--;
1389 static struct request
*
1390 cfq_find_rq_fmerge(struct cfq_data
*cfqd
, struct bio
*bio
)
1392 struct task_struct
*tsk
= current
;
1393 struct cfq_io_context
*cic
;
1394 struct cfq_queue
*cfqq
;
1396 cic
= cfq_cic_lookup(cfqd
, tsk
->io_context
);
1400 cfqq
= cic_to_cfqq(cic
, cfq_bio_sync(bio
));
1402 sector_t sector
= bio
->bi_sector
+ bio_sectors(bio
);
1404 return elv_rb_find(&cfqq
->sort_list
, sector
);
1410 static void cfq_activate_request(struct request_queue
*q
, struct request
*rq
)
1412 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
1414 cfqd
->rq_in_driver
++;
1415 cfq_log_cfqq(cfqd
, RQ_CFQQ(rq
), "activate rq, drv=%d",
1416 cfqd
->rq_in_driver
);
1418 cfqd
->last_position
= blk_rq_pos(rq
) + blk_rq_sectors(rq
);
1421 static void cfq_deactivate_request(struct request_queue
*q
, struct request
*rq
)
1423 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
1425 WARN_ON(!cfqd
->rq_in_driver
);
1426 cfqd
->rq_in_driver
--;
1427 cfq_log_cfqq(cfqd
, RQ_CFQQ(rq
), "deactivate rq, drv=%d",
1428 cfqd
->rq_in_driver
);
1431 static void cfq_remove_request(struct request
*rq
)
1433 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
1435 if (cfqq
->next_rq
== rq
)
1436 cfqq
->next_rq
= cfq_find_next_rq(cfqq
->cfqd
, cfqq
, rq
);
1438 list_del_init(&rq
->queuelist
);
1441 cfqq
->cfqd
->rq_queued
--;
1442 if (rq_is_meta(rq
)) {
1443 WARN_ON(!cfqq
->meta_pending
);
1444 cfqq
->meta_pending
--;
1448 static int cfq_merge(struct request_queue
*q
, struct request
**req
,
1451 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
1452 struct request
*__rq
;
1454 __rq
= cfq_find_rq_fmerge(cfqd
, bio
);
1455 if (__rq
&& elv_rq_merge_ok(__rq
, bio
)) {
1457 return ELEVATOR_FRONT_MERGE
;
1460 return ELEVATOR_NO_MERGE
;
1463 static void cfq_merged_request(struct request_queue
*q
, struct request
*req
,
1466 if (type
== ELEVATOR_FRONT_MERGE
) {
1467 struct cfq_queue
*cfqq
= RQ_CFQQ(req
);
1469 cfq_reposition_rq_rb(cfqq
, req
);
1474 cfq_merged_requests(struct request_queue
*q
, struct request
*rq
,
1475 struct request
*next
)
1477 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
1479 * reposition in fifo if next is older than rq
1481 if (!list_empty(&rq
->queuelist
) && !list_empty(&next
->queuelist
) &&
1482 time_before(rq_fifo_time(next
), rq_fifo_time(rq
))) {
1483 list_move(&rq
->queuelist
, &next
->queuelist
);
1484 rq_set_fifo_time(rq
, rq_fifo_time(next
));
1487 if (cfqq
->next_rq
== next
)
1489 cfq_remove_request(next
);
1492 static int cfq_allow_merge(struct request_queue
*q
, struct request
*rq
,
1495 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
1496 struct cfq_io_context
*cic
;
1497 struct cfq_queue
*cfqq
;
1500 * Disallow merge of a sync bio into an async request.
1502 if (cfq_bio_sync(bio
) && !rq_is_sync(rq
))
1506 * Lookup the cfqq that this bio will be queued with. Allow
1507 * merge only if rq is queued there.
1509 cic
= cfq_cic_lookup(cfqd
, current
->io_context
);
1513 cfqq
= cic_to_cfqq(cic
, cfq_bio_sync(bio
));
1514 return cfqq
== RQ_CFQQ(rq
);
1517 static void __cfq_set_active_queue(struct cfq_data
*cfqd
,
1518 struct cfq_queue
*cfqq
)
1521 cfq_log_cfqq(cfqd
, cfqq
, "set_active wl_prio:%d wl_type:%d",
1522 cfqd
->serving_prio
, cfqd
->serving_type
);
1523 cfqq
->slice_start
= 0;
1524 cfqq
->dispatch_start
= jiffies
;
1525 cfqq
->allocated_slice
= 0;
1526 cfqq
->slice_end
= 0;
1527 cfqq
->slice_dispatch
= 0;
1528 cfqq
->nr_sectors
= 0;
1530 cfq_clear_cfqq_wait_request(cfqq
);
1531 cfq_clear_cfqq_must_dispatch(cfqq
);
1532 cfq_clear_cfqq_must_alloc_slice(cfqq
);
1533 cfq_clear_cfqq_fifo_expire(cfqq
);
1534 cfq_mark_cfqq_slice_new(cfqq
);
1536 del_timer(&cfqd
->idle_slice_timer
);
1539 cfqd
->active_queue
= cfqq
;
1543 * current cfqq expired its slice (or was too idle), select new one
1546 __cfq_slice_expired(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
1549 cfq_log_cfqq(cfqd
, cfqq
, "slice expired t=%d", timed_out
);
1551 if (cfq_cfqq_wait_request(cfqq
))
1552 del_timer(&cfqd
->idle_slice_timer
);
1554 cfq_clear_cfqq_wait_request(cfqq
);
1555 cfq_clear_cfqq_wait_busy(cfqq
);
1558 * If this cfqq is shared between multiple processes, check to
1559 * make sure that those processes are still issuing I/Os within
1560 * the mean seek distance. If not, it may be time to break the
1561 * queues apart again.
1563 if (cfq_cfqq_coop(cfqq
) && CFQQ_SEEKY(cfqq
))
1564 cfq_mark_cfqq_split_coop(cfqq
);
1567 * store what was left of this slice, if the queue idled/timed out
1569 if (timed_out
&& !cfq_cfqq_slice_new(cfqq
)) {
1570 cfqq
->slice_resid
= cfqq
->slice_end
- jiffies
;
1571 cfq_log_cfqq(cfqd
, cfqq
, "resid=%ld", cfqq
->slice_resid
);
1574 cfq_group_served(cfqd
, cfqq
->cfqg
, cfqq
);
1576 if (cfq_cfqq_on_rr(cfqq
) && RB_EMPTY_ROOT(&cfqq
->sort_list
))
1577 cfq_del_cfqq_rr(cfqd
, cfqq
);
1579 cfq_resort_rr_list(cfqd
, cfqq
);
1581 if (cfqq
== cfqd
->active_queue
)
1582 cfqd
->active_queue
= NULL
;
1584 if (&cfqq
->cfqg
->rb_node
== cfqd
->grp_service_tree
.active
)
1585 cfqd
->grp_service_tree
.active
= NULL
;
1587 if (cfqd
->active_cic
) {
1588 put_io_context(cfqd
->active_cic
->ioc
);
1589 cfqd
->active_cic
= NULL
;
1593 static inline void cfq_slice_expired(struct cfq_data
*cfqd
, bool timed_out
)
1595 struct cfq_queue
*cfqq
= cfqd
->active_queue
;
1598 __cfq_slice_expired(cfqd
, cfqq
, timed_out
);
1602 * Get next queue for service. Unless we have a queue preemption,
1603 * we'll simply select the first cfqq in the service tree.
1605 static struct cfq_queue
*cfq_get_next_queue(struct cfq_data
*cfqd
)
1607 struct cfq_rb_root
*service_tree
=
1608 service_tree_for(cfqd
->serving_group
, cfqd
->serving_prio
,
1609 cfqd
->serving_type
);
1611 if (!cfqd
->rq_queued
)
1614 /* There is nothing to dispatch */
1617 if (RB_EMPTY_ROOT(&service_tree
->rb
))
1619 return cfq_rb_first(service_tree
);
1622 static struct cfq_queue
*cfq_get_next_queue_forced(struct cfq_data
*cfqd
)
1624 struct cfq_group
*cfqg
;
1625 struct cfq_queue
*cfqq
;
1627 struct cfq_rb_root
*st
;
1629 if (!cfqd
->rq_queued
)
1632 cfqg
= cfq_get_next_cfqg(cfqd
);
1636 for_each_cfqg_st(cfqg
, i
, j
, st
)
1637 if ((cfqq
= cfq_rb_first(st
)) != NULL
)
1643 * Get and set a new active queue for service.
1645 static struct cfq_queue
*cfq_set_active_queue(struct cfq_data
*cfqd
,
1646 struct cfq_queue
*cfqq
)
1649 cfqq
= cfq_get_next_queue(cfqd
);
1651 __cfq_set_active_queue(cfqd
, cfqq
);
1655 static inline sector_t
cfq_dist_from_last(struct cfq_data
*cfqd
,
1658 if (blk_rq_pos(rq
) >= cfqd
->last_position
)
1659 return blk_rq_pos(rq
) - cfqd
->last_position
;
1661 return cfqd
->last_position
- blk_rq_pos(rq
);
1664 static inline int cfq_rq_close(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
1667 return cfq_dist_from_last(cfqd
, rq
) <= CFQQ_CLOSE_THR
;
1670 static struct cfq_queue
*cfqq_close(struct cfq_data
*cfqd
,
1671 struct cfq_queue
*cur_cfqq
)
1673 struct rb_root
*root
= &cfqd
->prio_trees
[cur_cfqq
->org_ioprio
];
1674 struct rb_node
*parent
, *node
;
1675 struct cfq_queue
*__cfqq
;
1676 sector_t sector
= cfqd
->last_position
;
1678 if (RB_EMPTY_ROOT(root
))
1682 * First, if we find a request starting at the end of the last
1683 * request, choose it.
1685 __cfqq
= cfq_prio_tree_lookup(cfqd
, root
, sector
, &parent
, NULL
);
1690 * If the exact sector wasn't found, the parent of the NULL leaf
1691 * will contain the closest sector.
1693 __cfqq
= rb_entry(parent
, struct cfq_queue
, p_node
);
1694 if (cfq_rq_close(cfqd
, cur_cfqq
, __cfqq
->next_rq
))
1697 if (blk_rq_pos(__cfqq
->next_rq
) < sector
)
1698 node
= rb_next(&__cfqq
->p_node
);
1700 node
= rb_prev(&__cfqq
->p_node
);
1704 __cfqq
= rb_entry(node
, struct cfq_queue
, p_node
);
1705 if (cfq_rq_close(cfqd
, cur_cfqq
, __cfqq
->next_rq
))
1713 * cur_cfqq - passed in so that we don't decide that the current queue is
1714 * closely cooperating with itself.
1716 * So, basically we're assuming that that cur_cfqq has dispatched at least
1717 * one request, and that cfqd->last_position reflects a position on the disk
1718 * associated with the I/O issued by cur_cfqq. I'm not sure this is a valid
1721 static struct cfq_queue
*cfq_close_cooperator(struct cfq_data
*cfqd
,
1722 struct cfq_queue
*cur_cfqq
)
1724 struct cfq_queue
*cfqq
;
1726 if (!cfq_cfqq_sync(cur_cfqq
))
1728 if (CFQQ_SEEKY(cur_cfqq
))
1732 * Don't search priority tree if it's the only queue in the group.
1734 if (cur_cfqq
->cfqg
->nr_cfqq
== 1)
1738 * We should notice if some of the queues are cooperating, eg
1739 * working closely on the same area of the disk. In that case,
1740 * we can group them together and don't waste time idling.
1742 cfqq
= cfqq_close(cfqd
, cur_cfqq
);
1746 /* If new queue belongs to different cfq_group, don't choose it */
1747 if (cur_cfqq
->cfqg
!= cfqq
->cfqg
)
1751 * It only makes sense to merge sync queues.
1753 if (!cfq_cfqq_sync(cfqq
))
1755 if (CFQQ_SEEKY(cfqq
))
1759 * Do not merge queues of different priority classes
1761 if (cfq_class_rt(cfqq
) != cfq_class_rt(cur_cfqq
))
1768 * Determine whether we should enforce idle window for this queue.
1771 static bool cfq_should_idle(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
1773 enum wl_prio_t prio
= cfqq_prio(cfqq
);
1774 struct cfq_rb_root
*service_tree
= cfqq
->service_tree
;
1776 BUG_ON(!service_tree
);
1777 BUG_ON(!service_tree
->count
);
1779 /* We never do for idle class queues. */
1780 if (prio
== IDLE_WORKLOAD
)
1783 /* We do for queues that were marked with idle window flag. */
1784 if (cfq_cfqq_idle_window(cfqq
) &&
1785 !(blk_queue_nonrot(cfqd
->queue
) && cfqd
->hw_tag
))
1789 * Otherwise, we do only if they are the last ones
1790 * in their service tree.
1792 if (service_tree
->count
== 1 && cfq_cfqq_sync(cfqq
))
1794 cfq_log_cfqq(cfqd
, cfqq
, "Not idling. st->count:%d",
1795 service_tree
->count
);
1799 static void cfq_arm_slice_timer(struct cfq_data
*cfqd
)
1801 struct cfq_queue
*cfqq
= cfqd
->active_queue
;
1802 struct cfq_io_context
*cic
;
1806 * SSD device without seek penalty, disable idling. But only do so
1807 * for devices that support queuing, otherwise we still have a problem
1808 * with sync vs async workloads.
1810 if (blk_queue_nonrot(cfqd
->queue
) && cfqd
->hw_tag
)
1813 WARN_ON(!RB_EMPTY_ROOT(&cfqq
->sort_list
));
1814 WARN_ON(cfq_cfqq_slice_new(cfqq
));
1817 * idle is disabled, either manually or by past process history
1819 if (!cfqd
->cfq_slice_idle
|| !cfq_should_idle(cfqd
, cfqq
))
1823 * still active requests from this queue, don't idle
1825 if (cfqq
->dispatched
)
1829 * task has exited, don't wait
1831 cic
= cfqd
->active_cic
;
1832 if (!cic
|| !atomic_read(&cic
->ioc
->nr_tasks
))
1836 * If our average think time is larger than the remaining time
1837 * slice, then don't idle. This avoids overrunning the allotted
1840 if (sample_valid(cic
->ttime_samples
) &&
1841 (cfqq
->slice_end
- jiffies
< cic
->ttime_mean
)) {
1842 cfq_log_cfqq(cfqd
, cfqq
, "Not idling. think_time:%d",
1847 cfq_mark_cfqq_wait_request(cfqq
);
1849 sl
= cfqd
->cfq_slice_idle
;
1851 mod_timer(&cfqd
->idle_slice_timer
, jiffies
+ sl
);
1852 cfq_log_cfqq(cfqd
, cfqq
, "arm_idle: %lu", sl
);
1856 * Move request from internal lists to the request queue dispatch list.
1858 static void cfq_dispatch_insert(struct request_queue
*q
, struct request
*rq
)
1860 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
1861 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
1863 cfq_log_cfqq(cfqd
, cfqq
, "dispatch_insert");
1865 cfqq
->next_rq
= cfq_find_next_rq(cfqd
, cfqq
, rq
);
1866 cfq_remove_request(rq
);
1868 elv_dispatch_sort(q
, rq
);
1870 cfqd
->rq_in_flight
[cfq_cfqq_sync(cfqq
)]++;
1871 cfqq
->nr_sectors
+= blk_rq_sectors(rq
);
1875 * return expired entry, or NULL to just start from scratch in rbtree
1877 static struct request
*cfq_check_fifo(struct cfq_queue
*cfqq
)
1879 struct request
*rq
= NULL
;
1881 if (cfq_cfqq_fifo_expire(cfqq
))
1884 cfq_mark_cfqq_fifo_expire(cfqq
);
1886 if (list_empty(&cfqq
->fifo
))
1889 rq
= rq_entry_fifo(cfqq
->fifo
.next
);
1890 if (time_before(jiffies
, rq_fifo_time(rq
)))
1893 cfq_log_cfqq(cfqq
->cfqd
, cfqq
, "fifo=%p", rq
);
1898 cfq_prio_to_maxrq(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
1900 const int base_rq
= cfqd
->cfq_slice_async_rq
;
1902 WARN_ON(cfqq
->ioprio
>= IOPRIO_BE_NR
);
1904 return 2 * (base_rq
+ base_rq
* (CFQ_PRIO_LISTS
- 1 - cfqq
->ioprio
));
1908 * Must be called with the queue_lock held.
1910 static int cfqq_process_refs(struct cfq_queue
*cfqq
)
1912 int process_refs
, io_refs
;
1914 io_refs
= cfqq
->allocated
[READ
] + cfqq
->allocated
[WRITE
];
1915 process_refs
= atomic_read(&cfqq
->ref
) - io_refs
;
1916 BUG_ON(process_refs
< 0);
1917 return process_refs
;
1920 static void cfq_setup_merge(struct cfq_queue
*cfqq
, struct cfq_queue
*new_cfqq
)
1922 int process_refs
, new_process_refs
;
1923 struct cfq_queue
*__cfqq
;
1925 /* Avoid a circular list and skip interim queue merges */
1926 while ((__cfqq
= new_cfqq
->new_cfqq
)) {
1932 process_refs
= cfqq_process_refs(cfqq
);
1934 * If the process for the cfqq has gone away, there is no
1935 * sense in merging the queues.
1937 if (process_refs
== 0)
1941 * Merge in the direction of the lesser amount of work.
1943 new_process_refs
= cfqq_process_refs(new_cfqq
);
1944 if (new_process_refs
>= process_refs
) {
1945 cfqq
->new_cfqq
= new_cfqq
;
1946 atomic_add(process_refs
, &new_cfqq
->ref
);
1948 new_cfqq
->new_cfqq
= cfqq
;
1949 atomic_add(new_process_refs
, &cfqq
->ref
);
1953 static enum wl_type_t
cfq_choose_wl(struct cfq_data
*cfqd
,
1954 struct cfq_group
*cfqg
, enum wl_prio_t prio
)
1956 struct cfq_queue
*queue
;
1958 bool key_valid
= false;
1959 unsigned long lowest_key
= 0;
1960 enum wl_type_t cur_best
= SYNC_NOIDLE_WORKLOAD
;
1962 for (i
= 0; i
<= SYNC_WORKLOAD
; ++i
) {
1963 /* select the one with lowest rb_key */
1964 queue
= cfq_rb_first(service_tree_for(cfqg
, prio
, i
));
1966 (!key_valid
|| time_before(queue
->rb_key
, lowest_key
))) {
1967 lowest_key
= queue
->rb_key
;
1976 static void choose_service_tree(struct cfq_data
*cfqd
, struct cfq_group
*cfqg
)
1980 struct cfq_rb_root
*st
;
1981 unsigned group_slice
;
1984 cfqd
->serving_prio
= IDLE_WORKLOAD
;
1985 cfqd
->workload_expires
= jiffies
+ 1;
1989 /* Choose next priority. RT > BE > IDLE */
1990 if (cfq_group_busy_queues_wl(RT_WORKLOAD
, cfqd
, cfqg
))
1991 cfqd
->serving_prio
= RT_WORKLOAD
;
1992 else if (cfq_group_busy_queues_wl(BE_WORKLOAD
, cfqd
, cfqg
))
1993 cfqd
->serving_prio
= BE_WORKLOAD
;
1995 cfqd
->serving_prio
= IDLE_WORKLOAD
;
1996 cfqd
->workload_expires
= jiffies
+ 1;
2001 * For RT and BE, we have to choose also the type
2002 * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
2005 st
= service_tree_for(cfqg
, cfqd
->serving_prio
, cfqd
->serving_type
);
2009 * check workload expiration, and that we still have other queues ready
2011 if (count
&& !time_after(jiffies
, cfqd
->workload_expires
))
2014 /* otherwise select new workload type */
2015 cfqd
->serving_type
=
2016 cfq_choose_wl(cfqd
, cfqg
, cfqd
->serving_prio
);
2017 st
= service_tree_for(cfqg
, cfqd
->serving_prio
, cfqd
->serving_type
);
2021 * the workload slice is computed as a fraction of target latency
2022 * proportional to the number of queues in that workload, over
2023 * all the queues in the same priority class
2025 group_slice
= cfq_group_slice(cfqd
, cfqg
);
2027 slice
= group_slice
* count
/
2028 max_t(unsigned, cfqg
->busy_queues_avg
[cfqd
->serving_prio
],
2029 cfq_group_busy_queues_wl(cfqd
->serving_prio
, cfqd
, cfqg
));
2031 if (cfqd
->serving_type
== ASYNC_WORKLOAD
) {
2035 * Async queues are currently system wide. Just taking
2036 * proportion of queues with-in same group will lead to higher
2037 * async ratio system wide as generally root group is going
2038 * to have higher weight. A more accurate thing would be to
2039 * calculate system wide asnc/sync ratio.
2041 tmp
= cfq_target_latency
* cfqg_busy_async_queues(cfqd
, cfqg
);
2042 tmp
= tmp
/cfqd
->busy_queues
;
2043 slice
= min_t(unsigned, slice
, tmp
);
2045 /* async workload slice is scaled down according to
2046 * the sync/async slice ratio. */
2047 slice
= slice
* cfqd
->cfq_slice
[0] / cfqd
->cfq_slice
[1];
2049 /* sync workload slice is at least 2 * cfq_slice_idle */
2050 slice
= max(slice
, 2 * cfqd
->cfq_slice_idle
);
2052 slice
= max_t(unsigned, slice
, CFQ_MIN_TT
);
2053 cfq_log(cfqd
, "workload slice:%d", slice
);
2054 cfqd
->workload_expires
= jiffies
+ slice
;
2055 cfqd
->noidle_tree_requires_idle
= false;
2058 static struct cfq_group
*cfq_get_next_cfqg(struct cfq_data
*cfqd
)
2060 struct cfq_rb_root
*st
= &cfqd
->grp_service_tree
;
2061 struct cfq_group
*cfqg
;
2063 if (RB_EMPTY_ROOT(&st
->rb
))
2065 cfqg
= cfq_rb_first_group(st
);
2066 st
->active
= &cfqg
->rb_node
;
2067 update_min_vdisktime(st
);
2071 static void cfq_choose_cfqg(struct cfq_data
*cfqd
)
2073 struct cfq_group
*cfqg
= cfq_get_next_cfqg(cfqd
);
2075 cfqd
->serving_group
= cfqg
;
2077 /* Restore the workload type data */
2078 if (cfqg
->saved_workload_slice
) {
2079 cfqd
->workload_expires
= jiffies
+ cfqg
->saved_workload_slice
;
2080 cfqd
->serving_type
= cfqg
->saved_workload
;
2081 cfqd
->serving_prio
= cfqg
->saved_serving_prio
;
2083 cfqd
->workload_expires
= jiffies
- 1;
2085 choose_service_tree(cfqd
, cfqg
);
2089 * Select a queue for service. If we have a current active queue,
2090 * check whether to continue servicing it, or retrieve and set a new one.
2092 static struct cfq_queue
*cfq_select_queue(struct cfq_data
*cfqd
)
2094 struct cfq_queue
*cfqq
, *new_cfqq
= NULL
;
2096 cfqq
= cfqd
->active_queue
;
2100 if (!cfqd
->rq_queued
)
2104 * We were waiting for group to get backlogged. Expire the queue
2106 if (cfq_cfqq_wait_busy(cfqq
) && !RB_EMPTY_ROOT(&cfqq
->sort_list
))
2110 * The active queue has run out of time, expire it and select new.
2112 if (cfq_slice_used(cfqq
) && !cfq_cfqq_must_dispatch(cfqq
)) {
2114 * If slice had not expired at the completion of last request
2115 * we might not have turned on wait_busy flag. Don't expire
2116 * the queue yet. Allow the group to get backlogged.
2118 * The very fact that we have used the slice, that means we
2119 * have been idling all along on this queue and it should be
2120 * ok to wait for this request to complete.
2122 if (cfqq
->cfqg
->nr_cfqq
== 1 && RB_EMPTY_ROOT(&cfqq
->sort_list
)
2123 && cfqq
->dispatched
&& cfq_should_idle(cfqd
, cfqq
)) {
2131 * The active queue has requests and isn't expired, allow it to
2134 if (!RB_EMPTY_ROOT(&cfqq
->sort_list
))
2138 * If another queue has a request waiting within our mean seek
2139 * distance, let it run. The expire code will check for close
2140 * cooperators and put the close queue at the front of the service
2141 * tree. If possible, merge the expiring queue with the new cfqq.
2143 new_cfqq
= cfq_close_cooperator(cfqd
, cfqq
);
2145 if (!cfqq
->new_cfqq
)
2146 cfq_setup_merge(cfqq
, new_cfqq
);
2151 * No requests pending. If the active queue still has requests in
2152 * flight or is idling for a new request, allow either of these
2153 * conditions to happen (or time out) before selecting a new queue.
2155 if (timer_pending(&cfqd
->idle_slice_timer
) ||
2156 (cfqq
->dispatched
&& cfq_should_idle(cfqd
, cfqq
))) {
2162 cfq_slice_expired(cfqd
, 0);
2165 * Current queue expired. Check if we have to switch to a new
2169 cfq_choose_cfqg(cfqd
);
2171 cfqq
= cfq_set_active_queue(cfqd
, new_cfqq
);
2176 static int __cfq_forced_dispatch_cfqq(struct cfq_queue
*cfqq
)
2180 while (cfqq
->next_rq
) {
2181 cfq_dispatch_insert(cfqq
->cfqd
->queue
, cfqq
->next_rq
);
2185 BUG_ON(!list_empty(&cfqq
->fifo
));
2187 /* By default cfqq is not expired if it is empty. Do it explicitly */
2188 __cfq_slice_expired(cfqq
->cfqd
, cfqq
, 0);
2193 * Drain our current requests. Used for barriers and when switching
2194 * io schedulers on-the-fly.
2196 static int cfq_forced_dispatch(struct cfq_data
*cfqd
)
2198 struct cfq_queue
*cfqq
;
2201 while ((cfqq
= cfq_get_next_queue_forced(cfqd
)) != NULL
)
2202 dispatched
+= __cfq_forced_dispatch_cfqq(cfqq
);
2204 cfq_slice_expired(cfqd
, 0);
2205 BUG_ON(cfqd
->busy_queues
);
2207 cfq_log(cfqd
, "forced_dispatch=%d", dispatched
);
2211 static inline bool cfq_slice_used_soon(struct cfq_data
*cfqd
,
2212 struct cfq_queue
*cfqq
)
2214 /* the queue hasn't finished any request, can't estimate */
2215 if (cfq_cfqq_slice_new(cfqq
))
2217 if (time_after(jiffies
+ cfqd
->cfq_slice_idle
* cfqq
->dispatched
,
2224 static bool cfq_may_dispatch(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
2226 unsigned int max_dispatch
;
2229 * Drain async requests before we start sync IO
2231 if (cfq_should_idle(cfqd
, cfqq
) && cfqd
->rq_in_flight
[BLK_RW_ASYNC
])
2235 * If this is an async queue and we have sync IO in flight, let it wait
2237 if (cfqd
->rq_in_flight
[BLK_RW_SYNC
] && !cfq_cfqq_sync(cfqq
))
2240 max_dispatch
= max_t(unsigned int, cfqd
->cfq_quantum
/ 2, 1);
2241 if (cfq_class_idle(cfqq
))
2245 * Does this cfqq already have too much IO in flight?
2247 if (cfqq
->dispatched
>= max_dispatch
) {
2249 * idle queue must always only have a single IO in flight
2251 if (cfq_class_idle(cfqq
))
2255 * We have other queues, don't allow more IO from this one
2257 if (cfqd
->busy_queues
> 1 && cfq_slice_used_soon(cfqd
, cfqq
))
2261 * Sole queue user, no limit
2263 if (cfqd
->busy_queues
== 1)
2267 * Normally we start throttling cfqq when cfq_quantum/2
2268 * requests have been dispatched. But we can drive
2269 * deeper queue depths at the beginning of slice
2270 * subjected to upper limit of cfq_quantum.
2272 max_dispatch
= cfqd
->cfq_quantum
;
2276 * Async queues must wait a bit before being allowed dispatch.
2277 * We also ramp up the dispatch depth gradually for async IO,
2278 * based on the last sync IO we serviced
2280 if (!cfq_cfqq_sync(cfqq
) && cfqd
->cfq_latency
) {
2281 unsigned long last_sync
= jiffies
- cfqd
->last_delayed_sync
;
2284 depth
= last_sync
/ cfqd
->cfq_slice
[1];
2285 if (!depth
&& !cfqq
->dispatched
)
2287 if (depth
< max_dispatch
)
2288 max_dispatch
= depth
;
2292 * If we're below the current max, allow a dispatch
2294 return cfqq
->dispatched
< max_dispatch
;
2298 * Dispatch a request from cfqq, moving them to the request queue
2301 static bool cfq_dispatch_request(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
2305 BUG_ON(RB_EMPTY_ROOT(&cfqq
->sort_list
));
2307 if (!cfq_may_dispatch(cfqd
, cfqq
))
2311 * follow expired path, else get first next available
2313 rq
= cfq_check_fifo(cfqq
);
2318 * insert request into driver dispatch list
2320 cfq_dispatch_insert(cfqd
->queue
, rq
);
2322 if (!cfqd
->active_cic
) {
2323 struct cfq_io_context
*cic
= RQ_CIC(rq
);
2325 atomic_long_inc(&cic
->ioc
->refcount
);
2326 cfqd
->active_cic
= cic
;
2333 * Find the cfqq that we need to service and move a request from that to the
2336 static int cfq_dispatch_requests(struct request_queue
*q
, int force
)
2338 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
2339 struct cfq_queue
*cfqq
;
2341 if (!cfqd
->busy_queues
)
2344 if (unlikely(force
))
2345 return cfq_forced_dispatch(cfqd
);
2347 cfqq
= cfq_select_queue(cfqd
);
2352 * Dispatch a request from this cfqq, if it is allowed
2354 if (!cfq_dispatch_request(cfqd
, cfqq
))
2357 cfqq
->slice_dispatch
++;
2358 cfq_clear_cfqq_must_dispatch(cfqq
);
2361 * expire an async queue immediately if it has used up its slice. idle
2362 * queue always expire after 1 dispatch round.
2364 if (cfqd
->busy_queues
> 1 && ((!cfq_cfqq_sync(cfqq
) &&
2365 cfqq
->slice_dispatch
>= cfq_prio_to_maxrq(cfqd
, cfqq
)) ||
2366 cfq_class_idle(cfqq
))) {
2367 cfqq
->slice_end
= jiffies
+ 1;
2368 cfq_slice_expired(cfqd
, 0);
2371 cfq_log_cfqq(cfqd
, cfqq
, "dispatched a request");
2376 * task holds one reference to the queue, dropped when task exits. each rq
2377 * in-flight on this queue also holds a reference, dropped when rq is freed.
2379 * Each cfq queue took a reference on the parent group. Drop it now.
2380 * queue lock must be held here.
2382 static void cfq_put_queue(struct cfq_queue
*cfqq
)
2384 struct cfq_data
*cfqd
= cfqq
->cfqd
;
2385 struct cfq_group
*cfqg
, *orig_cfqg
;
2387 BUG_ON(atomic_read(&cfqq
->ref
) <= 0);
2389 if (!atomic_dec_and_test(&cfqq
->ref
))
2392 cfq_log_cfqq(cfqd
, cfqq
, "put_queue");
2393 BUG_ON(rb_first(&cfqq
->sort_list
));
2394 BUG_ON(cfqq
->allocated
[READ
] + cfqq
->allocated
[WRITE
]);
2396 orig_cfqg
= cfqq
->orig_cfqg
;
2398 if (unlikely(cfqd
->active_queue
== cfqq
)) {
2399 __cfq_slice_expired(cfqd
, cfqq
, 0);
2400 cfq_schedule_dispatch(cfqd
);
2403 BUG_ON(cfq_cfqq_on_rr(cfqq
));
2404 kmem_cache_free(cfq_pool
, cfqq
);
2407 cfq_put_cfqg(orig_cfqg
);
2411 * Must always be called with the rcu_read_lock() held
2414 __call_for_each_cic(struct io_context
*ioc
,
2415 void (*func
)(struct io_context
*, struct cfq_io_context
*))
2417 struct cfq_io_context
*cic
;
2418 struct hlist_node
*n
;
2420 hlist_for_each_entry_rcu(cic
, n
, &ioc
->cic_list
, cic_list
)
2425 * Call func for each cic attached to this ioc.
2428 call_for_each_cic(struct io_context
*ioc
,
2429 void (*func
)(struct io_context
*, struct cfq_io_context
*))
2432 __call_for_each_cic(ioc
, func
);
2436 static void cfq_cic_free_rcu(struct rcu_head
*head
)
2438 struct cfq_io_context
*cic
;
2440 cic
= container_of(head
, struct cfq_io_context
, rcu_head
);
2442 kmem_cache_free(cfq_ioc_pool
, cic
);
2443 elv_ioc_count_dec(cfq_ioc_count
);
2447 * CFQ scheduler is exiting, grab exit lock and check
2448 * the pending io context count. If it hits zero,
2449 * complete ioc_gone and set it back to NULL
2451 spin_lock(&ioc_gone_lock
);
2452 if (ioc_gone
&& !elv_ioc_count_read(cfq_ioc_count
)) {
2456 spin_unlock(&ioc_gone_lock
);
2460 static void cfq_cic_free(struct cfq_io_context
*cic
)
2462 call_rcu(&cic
->rcu_head
, cfq_cic_free_rcu
);
2465 static void cic_free_func(struct io_context
*ioc
, struct cfq_io_context
*cic
)
2467 unsigned long flags
;
2469 BUG_ON(!cic
->dead_key
);
2471 spin_lock_irqsave(&ioc
->lock
, flags
);
2472 radix_tree_delete(&ioc
->radix_root
, cic
->dead_key
);
2473 hlist_del_rcu(&cic
->cic_list
);
2474 spin_unlock_irqrestore(&ioc
->lock
, flags
);
2480 * Must be called with rcu_read_lock() held or preemption otherwise disabled.
2481 * Only two callers of this - ->dtor() which is called with the rcu_read_lock(),
2482 * and ->trim() which is called with the task lock held
2484 static void cfq_free_io_context(struct io_context
*ioc
)
2487 * ioc->refcount is zero here, or we are called from elv_unregister(),
2488 * so no more cic's are allowed to be linked into this ioc. So it
2489 * should be ok to iterate over the known list, we will see all cic's
2490 * since no new ones are added.
2492 __call_for_each_cic(ioc
, cic_free_func
);
2495 static void cfq_exit_cfqq(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
2497 struct cfq_queue
*__cfqq
, *next
;
2499 if (unlikely(cfqq
== cfqd
->active_queue
)) {
2500 __cfq_slice_expired(cfqd
, cfqq
, 0);
2501 cfq_schedule_dispatch(cfqd
);
2505 * If this queue was scheduled to merge with another queue, be
2506 * sure to drop the reference taken on that queue (and others in
2507 * the merge chain). See cfq_setup_merge and cfq_merge_cfqqs.
2509 __cfqq
= cfqq
->new_cfqq
;
2511 if (__cfqq
== cfqq
) {
2512 WARN(1, "cfqq->new_cfqq loop detected\n");
2515 next
= __cfqq
->new_cfqq
;
2516 cfq_put_queue(__cfqq
);
2520 cfq_put_queue(cfqq
);
2523 static void __cfq_exit_single_io_context(struct cfq_data
*cfqd
,
2524 struct cfq_io_context
*cic
)
2526 struct io_context
*ioc
= cic
->ioc
;
2528 list_del_init(&cic
->queue_list
);
2531 * Make sure key == NULL is seen for dead queues
2534 cic
->dead_key
= (unsigned long) cic
->key
;
2537 if (ioc
->ioc_data
== cic
)
2538 rcu_assign_pointer(ioc
->ioc_data
, NULL
);
2540 if (cic
->cfqq
[BLK_RW_ASYNC
]) {
2541 cfq_exit_cfqq(cfqd
, cic
->cfqq
[BLK_RW_ASYNC
]);
2542 cic
->cfqq
[BLK_RW_ASYNC
] = NULL
;
2545 if (cic
->cfqq
[BLK_RW_SYNC
]) {
2546 cfq_exit_cfqq(cfqd
, cic
->cfqq
[BLK_RW_SYNC
]);
2547 cic
->cfqq
[BLK_RW_SYNC
] = NULL
;
2551 static void cfq_exit_single_io_context(struct io_context
*ioc
,
2552 struct cfq_io_context
*cic
)
2554 struct cfq_data
*cfqd
= cic
->key
;
2557 struct request_queue
*q
= cfqd
->queue
;
2558 unsigned long flags
;
2560 spin_lock_irqsave(q
->queue_lock
, flags
);
2563 * Ensure we get a fresh copy of the ->key to prevent
2564 * race between exiting task and queue
2566 smp_read_barrier_depends();
2568 __cfq_exit_single_io_context(cfqd
, cic
);
2570 spin_unlock_irqrestore(q
->queue_lock
, flags
);
2575 * The process that ioc belongs to has exited, we need to clean up
2576 * and put the internal structures we have that belongs to that process.
2578 static void cfq_exit_io_context(struct io_context
*ioc
)
2580 call_for_each_cic(ioc
, cfq_exit_single_io_context
);
2583 static struct cfq_io_context
*
2584 cfq_alloc_io_context(struct cfq_data
*cfqd
, gfp_t gfp_mask
)
2586 struct cfq_io_context
*cic
;
2588 cic
= kmem_cache_alloc_node(cfq_ioc_pool
, gfp_mask
| __GFP_ZERO
,
2591 cic
->last_end_request
= jiffies
;
2592 INIT_LIST_HEAD(&cic
->queue_list
);
2593 INIT_HLIST_NODE(&cic
->cic_list
);
2594 cic
->dtor
= cfq_free_io_context
;
2595 cic
->exit
= cfq_exit_io_context
;
2596 elv_ioc_count_inc(cfq_ioc_count
);
2602 static void cfq_init_prio_data(struct cfq_queue
*cfqq
, struct io_context
*ioc
)
2604 struct task_struct
*tsk
= current
;
2607 if (!cfq_cfqq_prio_changed(cfqq
))
2610 ioprio_class
= IOPRIO_PRIO_CLASS(ioc
->ioprio
);
2611 switch (ioprio_class
) {
2613 printk(KERN_ERR
"cfq: bad prio %x\n", ioprio_class
);
2614 case IOPRIO_CLASS_NONE
:
2616 * no prio set, inherit CPU scheduling settings
2618 cfqq
->ioprio
= task_nice_ioprio(tsk
);
2619 cfqq
->ioprio_class
= task_nice_ioclass(tsk
);
2621 case IOPRIO_CLASS_RT
:
2622 cfqq
->ioprio
= task_ioprio(ioc
);
2623 cfqq
->ioprio_class
= IOPRIO_CLASS_RT
;
2625 case IOPRIO_CLASS_BE
:
2626 cfqq
->ioprio
= task_ioprio(ioc
);
2627 cfqq
->ioprio_class
= IOPRIO_CLASS_BE
;
2629 case IOPRIO_CLASS_IDLE
:
2630 cfqq
->ioprio_class
= IOPRIO_CLASS_IDLE
;
2632 cfq_clear_cfqq_idle_window(cfqq
);
2637 * keep track of original prio settings in case we have to temporarily
2638 * elevate the priority of this queue
2640 cfqq
->org_ioprio
= cfqq
->ioprio
;
2641 cfqq
->org_ioprio_class
= cfqq
->ioprio_class
;
2642 cfq_clear_cfqq_prio_changed(cfqq
);
2645 static void changed_ioprio(struct io_context
*ioc
, struct cfq_io_context
*cic
)
2647 struct cfq_data
*cfqd
= cic
->key
;
2648 struct cfq_queue
*cfqq
;
2649 unsigned long flags
;
2651 if (unlikely(!cfqd
))
2654 spin_lock_irqsave(cfqd
->queue
->queue_lock
, flags
);
2656 cfqq
= cic
->cfqq
[BLK_RW_ASYNC
];
2658 struct cfq_queue
*new_cfqq
;
2659 new_cfqq
= cfq_get_queue(cfqd
, BLK_RW_ASYNC
, cic
->ioc
,
2662 cic
->cfqq
[BLK_RW_ASYNC
] = new_cfqq
;
2663 cfq_put_queue(cfqq
);
2667 cfqq
= cic
->cfqq
[BLK_RW_SYNC
];
2669 cfq_mark_cfqq_prio_changed(cfqq
);
2671 spin_unlock_irqrestore(cfqd
->queue
->queue_lock
, flags
);
2674 static void cfq_ioc_set_ioprio(struct io_context
*ioc
)
2676 call_for_each_cic(ioc
, changed_ioprio
);
2677 ioc
->ioprio_changed
= 0;
2680 static void cfq_init_cfqq(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
2681 pid_t pid
, bool is_sync
)
2683 RB_CLEAR_NODE(&cfqq
->rb_node
);
2684 RB_CLEAR_NODE(&cfqq
->p_node
);
2685 INIT_LIST_HEAD(&cfqq
->fifo
);
2687 atomic_set(&cfqq
->ref
, 0);
2690 cfq_mark_cfqq_prio_changed(cfqq
);
2693 if (!cfq_class_idle(cfqq
))
2694 cfq_mark_cfqq_idle_window(cfqq
);
2695 cfq_mark_cfqq_sync(cfqq
);
2700 #ifdef CONFIG_CFQ_GROUP_IOSCHED
2701 static void changed_cgroup(struct io_context
*ioc
, struct cfq_io_context
*cic
)
2703 struct cfq_queue
*sync_cfqq
= cic_to_cfqq(cic
, 1);
2704 struct cfq_data
*cfqd
= cic
->key
;
2705 unsigned long flags
;
2706 struct request_queue
*q
;
2708 if (unlikely(!cfqd
))
2713 spin_lock_irqsave(q
->queue_lock
, flags
);
2717 * Drop reference to sync queue. A new sync queue will be
2718 * assigned in new group upon arrival of a fresh request.
2720 cfq_log_cfqq(cfqd
, sync_cfqq
, "changed cgroup");
2721 cic_set_cfqq(cic
, NULL
, 1);
2722 cfq_put_queue(sync_cfqq
);
2725 spin_unlock_irqrestore(q
->queue_lock
, flags
);
2728 static void cfq_ioc_set_cgroup(struct io_context
*ioc
)
2730 call_for_each_cic(ioc
, changed_cgroup
);
2731 ioc
->cgroup_changed
= 0;
2733 #endif /* CONFIG_CFQ_GROUP_IOSCHED */
2735 static struct cfq_queue
*
2736 cfq_find_alloc_queue(struct cfq_data
*cfqd
, bool is_sync
,
2737 struct io_context
*ioc
, gfp_t gfp_mask
)
2739 struct cfq_queue
*cfqq
, *new_cfqq
= NULL
;
2740 struct cfq_io_context
*cic
;
2741 struct cfq_group
*cfqg
;
2744 cfqg
= cfq_get_cfqg(cfqd
, 1);
2745 cic
= cfq_cic_lookup(cfqd
, ioc
);
2746 /* cic always exists here */
2747 cfqq
= cic_to_cfqq(cic
, is_sync
);
2750 * Always try a new alloc if we fell back to the OOM cfqq
2751 * originally, since it should just be a temporary situation.
2753 if (!cfqq
|| cfqq
== &cfqd
->oom_cfqq
) {
2758 } else if (gfp_mask
& __GFP_WAIT
) {
2759 spin_unlock_irq(cfqd
->queue
->queue_lock
);
2760 new_cfqq
= kmem_cache_alloc_node(cfq_pool
,
2761 gfp_mask
| __GFP_ZERO
,
2763 spin_lock_irq(cfqd
->queue
->queue_lock
);
2767 cfqq
= kmem_cache_alloc_node(cfq_pool
,
2768 gfp_mask
| __GFP_ZERO
,
2773 cfq_init_cfqq(cfqd
, cfqq
, current
->pid
, is_sync
);
2774 cfq_init_prio_data(cfqq
, ioc
);
2775 cfq_link_cfqq_cfqg(cfqq
, cfqg
);
2776 cfq_log_cfqq(cfqd
, cfqq
, "alloced");
2778 cfqq
= &cfqd
->oom_cfqq
;
2782 kmem_cache_free(cfq_pool
, new_cfqq
);
2787 static struct cfq_queue
**
2788 cfq_async_queue_prio(struct cfq_data
*cfqd
, int ioprio_class
, int ioprio
)
2790 switch (ioprio_class
) {
2791 case IOPRIO_CLASS_RT
:
2792 return &cfqd
->async_cfqq
[0][ioprio
];
2793 case IOPRIO_CLASS_BE
:
2794 return &cfqd
->async_cfqq
[1][ioprio
];
2795 case IOPRIO_CLASS_IDLE
:
2796 return &cfqd
->async_idle_cfqq
;
2802 static struct cfq_queue
*
2803 cfq_get_queue(struct cfq_data
*cfqd
, bool is_sync
, struct io_context
*ioc
,
2806 const int ioprio
= task_ioprio(ioc
);
2807 const int ioprio_class
= task_ioprio_class(ioc
);
2808 struct cfq_queue
**async_cfqq
= NULL
;
2809 struct cfq_queue
*cfqq
= NULL
;
2812 async_cfqq
= cfq_async_queue_prio(cfqd
, ioprio_class
, ioprio
);
2817 cfqq
= cfq_find_alloc_queue(cfqd
, is_sync
, ioc
, gfp_mask
);
2820 * pin the queue now that it's allocated, scheduler exit will prune it
2822 if (!is_sync
&& !(*async_cfqq
)) {
2823 atomic_inc(&cfqq
->ref
);
2827 atomic_inc(&cfqq
->ref
);
2832 * We drop cfq io contexts lazily, so we may find a dead one.
2835 cfq_drop_dead_cic(struct cfq_data
*cfqd
, struct io_context
*ioc
,
2836 struct cfq_io_context
*cic
)
2838 unsigned long flags
;
2840 WARN_ON(!list_empty(&cic
->queue_list
));
2842 spin_lock_irqsave(&ioc
->lock
, flags
);
2844 BUG_ON(ioc
->ioc_data
== cic
);
2846 radix_tree_delete(&ioc
->radix_root
, (unsigned long) cfqd
);
2847 hlist_del_rcu(&cic
->cic_list
);
2848 spin_unlock_irqrestore(&ioc
->lock
, flags
);
2853 static struct cfq_io_context
*
2854 cfq_cic_lookup(struct cfq_data
*cfqd
, struct io_context
*ioc
)
2856 struct cfq_io_context
*cic
;
2857 unsigned long flags
;
2866 * we maintain a last-hit cache, to avoid browsing over the tree
2868 cic
= rcu_dereference(ioc
->ioc_data
);
2869 if (cic
&& cic
->key
== cfqd
) {
2875 cic
= radix_tree_lookup(&ioc
->radix_root
, (unsigned long) cfqd
);
2879 /* ->key must be copied to avoid race with cfq_exit_queue() */
2882 cfq_drop_dead_cic(cfqd
, ioc
, cic
);
2887 spin_lock_irqsave(&ioc
->lock
, flags
);
2888 rcu_assign_pointer(ioc
->ioc_data
, cic
);
2889 spin_unlock_irqrestore(&ioc
->lock
, flags
);
2897 * Add cic into ioc, using cfqd as the search key. This enables us to lookup
2898 * the process specific cfq io context when entered from the block layer.
2899 * Also adds the cic to a per-cfqd list, used when this queue is removed.
2901 static int cfq_cic_link(struct cfq_data
*cfqd
, struct io_context
*ioc
,
2902 struct cfq_io_context
*cic
, gfp_t gfp_mask
)
2904 unsigned long flags
;
2907 ret
= radix_tree_preload(gfp_mask
);
2912 spin_lock_irqsave(&ioc
->lock
, flags
);
2913 ret
= radix_tree_insert(&ioc
->radix_root
,
2914 (unsigned long) cfqd
, cic
);
2916 hlist_add_head_rcu(&cic
->cic_list
, &ioc
->cic_list
);
2917 spin_unlock_irqrestore(&ioc
->lock
, flags
);
2919 radix_tree_preload_end();
2922 spin_lock_irqsave(cfqd
->queue
->queue_lock
, flags
);
2923 list_add(&cic
->queue_list
, &cfqd
->cic_list
);
2924 spin_unlock_irqrestore(cfqd
->queue
->queue_lock
, flags
);
2929 printk(KERN_ERR
"cfq: cic link failed!\n");
2935 * Setup general io context and cfq io context. There can be several cfq
2936 * io contexts per general io context, if this process is doing io to more
2937 * than one device managed by cfq.
2939 static struct cfq_io_context
*
2940 cfq_get_io_context(struct cfq_data
*cfqd
, gfp_t gfp_mask
)
2942 struct io_context
*ioc
= NULL
;
2943 struct cfq_io_context
*cic
;
2945 might_sleep_if(gfp_mask
& __GFP_WAIT
);
2947 ioc
= get_io_context(gfp_mask
, cfqd
->queue
->node
);
2951 cic
= cfq_cic_lookup(cfqd
, ioc
);
2955 cic
= cfq_alloc_io_context(cfqd
, gfp_mask
);
2959 if (cfq_cic_link(cfqd
, ioc
, cic
, gfp_mask
))
2963 smp_read_barrier_depends();
2964 if (unlikely(ioc
->ioprio_changed
))
2965 cfq_ioc_set_ioprio(ioc
);
2967 #ifdef CONFIG_CFQ_GROUP_IOSCHED
2968 if (unlikely(ioc
->cgroup_changed
))
2969 cfq_ioc_set_cgroup(ioc
);
2975 put_io_context(ioc
);
2980 cfq_update_io_thinktime(struct cfq_data
*cfqd
, struct cfq_io_context
*cic
)
2982 unsigned long elapsed
= jiffies
- cic
->last_end_request
;
2983 unsigned long ttime
= min(elapsed
, 2UL * cfqd
->cfq_slice_idle
);
2985 cic
->ttime_samples
= (7*cic
->ttime_samples
+ 256) / 8;
2986 cic
->ttime_total
= (7*cic
->ttime_total
+ 256*ttime
) / 8;
2987 cic
->ttime_mean
= (cic
->ttime_total
+ 128) / cic
->ttime_samples
;
2991 cfq_update_io_seektime(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
2995 sector_t n_sec
= blk_rq_sectors(rq
);
2996 if (cfqq
->last_request_pos
) {
2997 if (cfqq
->last_request_pos
< blk_rq_pos(rq
))
2998 sdist
= blk_rq_pos(rq
) - cfqq
->last_request_pos
;
3000 sdist
= cfqq
->last_request_pos
- blk_rq_pos(rq
);
3003 cfqq
->seek_history
<<= 1;
3004 if (blk_queue_nonrot(cfqd
->queue
))
3005 cfqq
->seek_history
|= (n_sec
< CFQQ_SECT_THR_NONROT
);
3007 cfqq
->seek_history
|= (sdist
> CFQQ_SEEK_THR
);
3011 * Disable idle window if the process thinks too long or seeks so much that
3015 cfq_update_idle_window(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
3016 struct cfq_io_context
*cic
)
3018 int old_idle
, enable_idle
;
3021 * Don't idle for async or idle io prio class
3023 if (!cfq_cfqq_sync(cfqq
) || cfq_class_idle(cfqq
))
3026 enable_idle
= old_idle
= cfq_cfqq_idle_window(cfqq
);
3028 if (cfqq
->queued
[0] + cfqq
->queued
[1] >= 4)
3029 cfq_mark_cfqq_deep(cfqq
);
3031 if (!atomic_read(&cic
->ioc
->nr_tasks
) || !cfqd
->cfq_slice_idle
||
3032 (!cfq_cfqq_deep(cfqq
) && CFQQ_SEEKY(cfqq
)))
3034 else if (sample_valid(cic
->ttime_samples
)) {
3035 if (cic
->ttime_mean
> cfqd
->cfq_slice_idle
)
3041 if (old_idle
!= enable_idle
) {
3042 cfq_log_cfqq(cfqd
, cfqq
, "idle=%d", enable_idle
);
3044 cfq_mark_cfqq_idle_window(cfqq
);
3046 cfq_clear_cfqq_idle_window(cfqq
);
3051 * Check if new_cfqq should preempt the currently active queue. Return 0 for
3052 * no or if we aren't sure, a 1 will cause a preempt.
3055 cfq_should_preempt(struct cfq_data
*cfqd
, struct cfq_queue
*new_cfqq
,
3058 struct cfq_queue
*cfqq
;
3060 cfqq
= cfqd
->active_queue
;
3064 if (cfq_class_idle(new_cfqq
))
3067 if (cfq_class_idle(cfqq
))
3071 * Don't allow a non-RT request to preempt an ongoing RT cfqq timeslice.
3073 if (cfq_class_rt(cfqq
) && !cfq_class_rt(new_cfqq
))
3077 * if the new request is sync, but the currently running queue is
3078 * not, let the sync request have priority.
3080 if (rq_is_sync(rq
) && !cfq_cfqq_sync(cfqq
))
3083 if (new_cfqq
->cfqg
!= cfqq
->cfqg
)
3086 if (cfq_slice_used(cfqq
))
3089 /* Allow preemption only if we are idling on sync-noidle tree */
3090 if (cfqd
->serving_type
== SYNC_NOIDLE_WORKLOAD
&&
3091 cfqq_type(new_cfqq
) == SYNC_NOIDLE_WORKLOAD
&&
3092 new_cfqq
->service_tree
->count
== 2 &&
3093 RB_EMPTY_ROOT(&cfqq
->sort_list
))
3097 * So both queues are sync. Let the new request get disk time if
3098 * it's a metadata request and the current queue is doing regular IO.
3100 if (rq_is_meta(rq
) && !cfqq
->meta_pending
)
3104 * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
3106 if (cfq_class_rt(new_cfqq
) && !cfq_class_rt(cfqq
))
3109 if (!cfqd
->active_cic
|| !cfq_cfqq_wait_request(cfqq
))
3113 * if this request is as-good as one we would expect from the
3114 * current cfqq, let it preempt
3116 if (cfq_rq_close(cfqd
, cfqq
, rq
))
3123 * cfqq preempts the active queue. if we allowed preempt with no slice left,
3124 * let it have half of its nominal slice.
3126 static void cfq_preempt_queue(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
3128 cfq_log_cfqq(cfqd
, cfqq
, "preempt");
3129 cfq_slice_expired(cfqd
, 1);
3132 * Put the new queue at the front of the of the current list,
3133 * so we know that it will be selected next.
3135 BUG_ON(!cfq_cfqq_on_rr(cfqq
));
3137 cfq_service_tree_add(cfqd
, cfqq
, 1);
3139 cfqq
->slice_end
= 0;
3140 cfq_mark_cfqq_slice_new(cfqq
);
3144 * Called when a new fs request (rq) is added (to cfqq). Check if there's
3145 * something we should do about it
3148 cfq_rq_enqueued(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
3151 struct cfq_io_context
*cic
= RQ_CIC(rq
);
3155 cfqq
->meta_pending
++;
3157 cfq_update_io_thinktime(cfqd
, cic
);
3158 cfq_update_io_seektime(cfqd
, cfqq
, rq
);
3159 cfq_update_idle_window(cfqd
, cfqq
, cic
);
3161 cfqq
->last_request_pos
= blk_rq_pos(rq
) + blk_rq_sectors(rq
);
3163 if (cfqq
== cfqd
->active_queue
) {
3165 * Remember that we saw a request from this process, but
3166 * don't start queuing just yet. Otherwise we risk seeing lots
3167 * of tiny requests, because we disrupt the normal plugging
3168 * and merging. If the request is already larger than a single
3169 * page, let it rip immediately. For that case we assume that
3170 * merging is already done. Ditto for a busy system that
3171 * has other work pending, don't risk delaying until the
3172 * idle timer unplug to continue working.
3174 if (cfq_cfqq_wait_request(cfqq
)) {
3175 if (blk_rq_bytes(rq
) > PAGE_CACHE_SIZE
||
3176 cfqd
->busy_queues
> 1) {
3177 del_timer(&cfqd
->idle_slice_timer
);
3178 cfq_clear_cfqq_wait_request(cfqq
);
3179 __blk_run_queue(cfqd
->queue
);
3181 cfq_mark_cfqq_must_dispatch(cfqq
);
3183 } else if (cfq_should_preempt(cfqd
, cfqq
, rq
)) {
3185 * not the active queue - expire current slice if it is
3186 * idle and has expired it's mean thinktime or this new queue
3187 * has some old slice time left and is of higher priority or
3188 * this new queue is RT and the current one is BE
3190 cfq_preempt_queue(cfqd
, cfqq
);
3191 __blk_run_queue(cfqd
->queue
);
3195 static void cfq_insert_request(struct request_queue
*q
, struct request
*rq
)
3197 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
3198 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
3200 cfq_log_cfqq(cfqd
, cfqq
, "insert_request");
3201 cfq_init_prio_data(cfqq
, RQ_CIC(rq
)->ioc
);
3203 rq_set_fifo_time(rq
, jiffies
+ cfqd
->cfq_fifo_expire
[rq_is_sync(rq
)]);
3204 list_add_tail(&rq
->queuelist
, &cfqq
->fifo
);
3207 cfq_rq_enqueued(cfqd
, cfqq
, rq
);
3211 * Update hw_tag based on peak queue depth over 50 samples under
3214 static void cfq_update_hw_tag(struct cfq_data
*cfqd
)
3216 struct cfq_queue
*cfqq
= cfqd
->active_queue
;
3218 if (cfqd
->rq_in_driver
> cfqd
->hw_tag_est_depth
)
3219 cfqd
->hw_tag_est_depth
= cfqd
->rq_in_driver
;
3221 if (cfqd
->hw_tag
== 1)
3224 if (cfqd
->rq_queued
<= CFQ_HW_QUEUE_MIN
&&
3225 cfqd
->rq_in_driver
<= CFQ_HW_QUEUE_MIN
)
3229 * If active queue hasn't enough requests and can idle, cfq might not
3230 * dispatch sufficient requests to hardware. Don't zero hw_tag in this
3233 if (cfqq
&& cfq_cfqq_idle_window(cfqq
) &&
3234 cfqq
->dispatched
+ cfqq
->queued
[0] + cfqq
->queued
[1] <
3235 CFQ_HW_QUEUE_MIN
&& cfqd
->rq_in_driver
< CFQ_HW_QUEUE_MIN
)
3238 if (cfqd
->hw_tag_samples
++ < 50)
3241 if (cfqd
->hw_tag_est_depth
>= CFQ_HW_QUEUE_MIN
)
3247 static bool cfq_should_wait_busy(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
3249 struct cfq_io_context
*cic
= cfqd
->active_cic
;
3251 /* If there are other queues in the group, don't wait */
3252 if (cfqq
->cfqg
->nr_cfqq
> 1)
3255 if (cfq_slice_used(cfqq
))
3258 /* if slice left is less than think time, wait busy */
3259 if (cic
&& sample_valid(cic
->ttime_samples
)
3260 && (cfqq
->slice_end
- jiffies
< cic
->ttime_mean
))
3264 * If think times is less than a jiffy than ttime_mean=0 and above
3265 * will not be true. It might happen that slice has not expired yet
3266 * but will expire soon (4-5 ns) during select_queue(). To cover the
3267 * case where think time is less than a jiffy, mark the queue wait
3268 * busy if only 1 jiffy is left in the slice.
3270 if (cfqq
->slice_end
- jiffies
== 1)
3276 static void cfq_completed_request(struct request_queue
*q
, struct request
*rq
)
3278 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
3279 struct cfq_data
*cfqd
= cfqq
->cfqd
;
3280 const int sync
= rq_is_sync(rq
);
3284 cfq_log_cfqq(cfqd
, cfqq
, "complete rqnoidle %d", !!rq_noidle(rq
));
3286 cfq_update_hw_tag(cfqd
);
3288 WARN_ON(!cfqd
->rq_in_driver
);
3289 WARN_ON(!cfqq
->dispatched
);
3290 cfqd
->rq_in_driver
--;
3293 cfqd
->rq_in_flight
[cfq_cfqq_sync(cfqq
)]--;
3296 RQ_CIC(rq
)->last_end_request
= now
;
3297 if (!time_after(rq
->start_time
+ cfqd
->cfq_fifo_expire
[1], now
))
3298 cfqd
->last_delayed_sync
= now
;
3302 * If this is the active queue, check if it needs to be expired,
3303 * or if we want to idle in case it has no pending requests.
3305 if (cfqd
->active_queue
== cfqq
) {
3306 const bool cfqq_empty
= RB_EMPTY_ROOT(&cfqq
->sort_list
);
3308 if (cfq_cfqq_slice_new(cfqq
)) {
3309 cfq_set_prio_slice(cfqd
, cfqq
);
3310 cfq_clear_cfqq_slice_new(cfqq
);
3314 * Should we wait for next request to come in before we expire
3317 if (cfq_should_wait_busy(cfqd
, cfqq
)) {
3318 cfqq
->slice_end
= jiffies
+ cfqd
->cfq_slice_idle
;
3319 cfq_mark_cfqq_wait_busy(cfqq
);
3320 cfq_log_cfqq(cfqd
, cfqq
, "will busy wait");
3324 * Idling is not enabled on:
3326 * - idle-priority queues
3328 * - queues with still some requests queued
3329 * - when there is a close cooperator
3331 if (cfq_slice_used(cfqq
) || cfq_class_idle(cfqq
))
3332 cfq_slice_expired(cfqd
, 1);
3333 else if (sync
&& cfqq_empty
&&
3334 !cfq_close_cooperator(cfqd
, cfqq
)) {
3335 cfqd
->noidle_tree_requires_idle
|= !rq_noidle(rq
);
3337 * Idling is enabled for SYNC_WORKLOAD.
3338 * SYNC_NOIDLE_WORKLOAD idles at the end of the tree
3339 * only if we processed at least one !rq_noidle request
3341 if (cfqd
->serving_type
== SYNC_WORKLOAD
3342 || cfqd
->noidle_tree_requires_idle
3343 || cfqq
->cfqg
->nr_cfqq
== 1)
3344 cfq_arm_slice_timer(cfqd
);
3348 if (!cfqd
->rq_in_driver
)
3349 cfq_schedule_dispatch(cfqd
);
3353 * we temporarily boost lower priority queues if they are holding fs exclusive
3354 * resources. they are boosted to normal prio (CLASS_BE/4)
3356 static void cfq_prio_boost(struct cfq_queue
*cfqq
)
3358 if (has_fs_excl()) {
3360 * boost idle prio on transactions that would lock out other
3361 * users of the filesystem
3363 if (cfq_class_idle(cfqq
))
3364 cfqq
->ioprio_class
= IOPRIO_CLASS_BE
;
3365 if (cfqq
->ioprio
> IOPRIO_NORM
)
3366 cfqq
->ioprio
= IOPRIO_NORM
;
3369 * unboost the queue (if needed)
3371 cfqq
->ioprio_class
= cfqq
->org_ioprio_class
;
3372 cfqq
->ioprio
= cfqq
->org_ioprio
;
3376 static inline int __cfq_may_queue(struct cfq_queue
*cfqq
)
3378 if (cfq_cfqq_wait_request(cfqq
) && !cfq_cfqq_must_alloc_slice(cfqq
)) {
3379 cfq_mark_cfqq_must_alloc_slice(cfqq
);
3380 return ELV_MQUEUE_MUST
;
3383 return ELV_MQUEUE_MAY
;
3386 static int cfq_may_queue(struct request_queue
*q
, int rw
)
3388 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
3389 struct task_struct
*tsk
= current
;
3390 struct cfq_io_context
*cic
;
3391 struct cfq_queue
*cfqq
;
3394 * don't force setup of a queue from here, as a call to may_queue
3395 * does not necessarily imply that a request actually will be queued.
3396 * so just lookup a possibly existing queue, or return 'may queue'
3399 cic
= cfq_cic_lookup(cfqd
, tsk
->io_context
);
3401 return ELV_MQUEUE_MAY
;
3403 cfqq
= cic_to_cfqq(cic
, rw_is_sync(rw
));
3405 cfq_init_prio_data(cfqq
, cic
->ioc
);
3406 cfq_prio_boost(cfqq
);
3408 return __cfq_may_queue(cfqq
);
3411 return ELV_MQUEUE_MAY
;
3415 * queue lock held here
3417 static void cfq_put_request(struct request
*rq
)
3419 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
3422 const int rw
= rq_data_dir(rq
);
3424 BUG_ON(!cfqq
->allocated
[rw
]);
3425 cfqq
->allocated
[rw
]--;
3427 put_io_context(RQ_CIC(rq
)->ioc
);
3429 rq
->elevator_private
= NULL
;
3430 rq
->elevator_private2
= NULL
;
3432 cfq_put_queue(cfqq
);
3436 static struct cfq_queue
*
3437 cfq_merge_cfqqs(struct cfq_data
*cfqd
, struct cfq_io_context
*cic
,
3438 struct cfq_queue
*cfqq
)
3440 cfq_log_cfqq(cfqd
, cfqq
, "merging with queue %p", cfqq
->new_cfqq
);
3441 cic_set_cfqq(cic
, cfqq
->new_cfqq
, 1);
3442 cfq_mark_cfqq_coop(cfqq
->new_cfqq
);
3443 cfq_put_queue(cfqq
);
3444 return cic_to_cfqq(cic
, 1);
3448 * Returns NULL if a new cfqq should be allocated, or the old cfqq if this
3449 * was the last process referring to said cfqq.
3451 static struct cfq_queue
*
3452 split_cfqq(struct cfq_io_context
*cic
, struct cfq_queue
*cfqq
)
3454 if (cfqq_process_refs(cfqq
) == 1) {
3455 cfqq
->pid
= current
->pid
;
3456 cfq_clear_cfqq_coop(cfqq
);
3457 cfq_clear_cfqq_split_coop(cfqq
);
3461 cic_set_cfqq(cic
, NULL
, 1);
3462 cfq_put_queue(cfqq
);
3466 * Allocate cfq data structures associated with this request.
3469 cfq_set_request(struct request_queue
*q
, struct request
*rq
, gfp_t gfp_mask
)
3471 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
3472 struct cfq_io_context
*cic
;
3473 const int rw
= rq_data_dir(rq
);
3474 const bool is_sync
= rq_is_sync(rq
);
3475 struct cfq_queue
*cfqq
;
3476 unsigned long flags
;
3478 might_sleep_if(gfp_mask
& __GFP_WAIT
);
3480 cic
= cfq_get_io_context(cfqd
, gfp_mask
);
3482 spin_lock_irqsave(q
->queue_lock
, flags
);
3488 cfqq
= cic_to_cfqq(cic
, is_sync
);
3489 if (!cfqq
|| cfqq
== &cfqd
->oom_cfqq
) {
3490 cfqq
= cfq_get_queue(cfqd
, is_sync
, cic
->ioc
, gfp_mask
);
3491 cic_set_cfqq(cic
, cfqq
, is_sync
);
3494 * If the queue was seeky for too long, break it apart.
3496 if (cfq_cfqq_coop(cfqq
) && cfq_cfqq_split_coop(cfqq
)) {
3497 cfq_log_cfqq(cfqd
, cfqq
, "breaking apart cfqq");
3498 cfqq
= split_cfqq(cic
, cfqq
);
3504 * Check to see if this queue is scheduled to merge with
3505 * another, closely cooperating queue. The merging of
3506 * queues happens here as it must be done in process context.
3507 * The reference on new_cfqq was taken in merge_cfqqs.
3510 cfqq
= cfq_merge_cfqqs(cfqd
, cic
, cfqq
);
3513 cfqq
->allocated
[rw
]++;
3514 atomic_inc(&cfqq
->ref
);
3516 spin_unlock_irqrestore(q
->queue_lock
, flags
);
3518 rq
->elevator_private
= cic
;
3519 rq
->elevator_private2
= cfqq
;
3524 put_io_context(cic
->ioc
);
3526 cfq_schedule_dispatch(cfqd
);
3527 spin_unlock_irqrestore(q
->queue_lock
, flags
);
3528 cfq_log(cfqd
, "set_request fail");
3532 static void cfq_kick_queue(struct work_struct
*work
)
3534 struct cfq_data
*cfqd
=
3535 container_of(work
, struct cfq_data
, unplug_work
);
3536 struct request_queue
*q
= cfqd
->queue
;
3538 spin_lock_irq(q
->queue_lock
);
3539 __blk_run_queue(cfqd
->queue
);
3540 spin_unlock_irq(q
->queue_lock
);
3544 * Timer running if the active_queue is currently idling inside its time slice
3546 static void cfq_idle_slice_timer(unsigned long data
)
3548 struct cfq_data
*cfqd
= (struct cfq_data
*) data
;
3549 struct cfq_queue
*cfqq
;
3550 unsigned long flags
;
3553 cfq_log(cfqd
, "idle timer fired");
3555 spin_lock_irqsave(cfqd
->queue
->queue_lock
, flags
);
3557 cfqq
= cfqd
->active_queue
;
3562 * We saw a request before the queue expired, let it through
3564 if (cfq_cfqq_must_dispatch(cfqq
))
3570 if (cfq_slice_used(cfqq
))
3574 * only expire and reinvoke request handler, if there are
3575 * other queues with pending requests
3577 if (!cfqd
->busy_queues
)
3581 * not expired and it has a request pending, let it dispatch
3583 if (!RB_EMPTY_ROOT(&cfqq
->sort_list
))
3587 * Queue depth flag is reset only when the idle didn't succeed
3589 cfq_clear_cfqq_deep(cfqq
);
3592 cfq_slice_expired(cfqd
, timed_out
);
3594 cfq_schedule_dispatch(cfqd
);
3596 spin_unlock_irqrestore(cfqd
->queue
->queue_lock
, flags
);
3599 static void cfq_shutdown_timer_wq(struct cfq_data
*cfqd
)
3601 del_timer_sync(&cfqd
->idle_slice_timer
);
3602 cancel_work_sync(&cfqd
->unplug_work
);
3605 static void cfq_put_async_queues(struct cfq_data
*cfqd
)
3609 for (i
= 0; i
< IOPRIO_BE_NR
; i
++) {
3610 if (cfqd
->async_cfqq
[0][i
])
3611 cfq_put_queue(cfqd
->async_cfqq
[0][i
]);
3612 if (cfqd
->async_cfqq
[1][i
])
3613 cfq_put_queue(cfqd
->async_cfqq
[1][i
]);
3616 if (cfqd
->async_idle_cfqq
)
3617 cfq_put_queue(cfqd
->async_idle_cfqq
);
3620 static void cfq_cfqd_free(struct rcu_head
*head
)
3622 kfree(container_of(head
, struct cfq_data
, rcu
));
3625 static void cfq_exit_queue(struct elevator_queue
*e
)
3627 struct cfq_data
*cfqd
= e
->elevator_data
;
3628 struct request_queue
*q
= cfqd
->queue
;
3630 cfq_shutdown_timer_wq(cfqd
);
3632 spin_lock_irq(q
->queue_lock
);
3634 if (cfqd
->active_queue
)
3635 __cfq_slice_expired(cfqd
, cfqd
->active_queue
, 0);
3637 while (!list_empty(&cfqd
->cic_list
)) {
3638 struct cfq_io_context
*cic
= list_entry(cfqd
->cic_list
.next
,
3639 struct cfq_io_context
,
3642 __cfq_exit_single_io_context(cfqd
, cic
);
3645 cfq_put_async_queues(cfqd
);
3646 cfq_release_cfq_groups(cfqd
);
3647 blkiocg_del_blkio_group(&cfqd
->root_group
.blkg
);
3649 spin_unlock_irq(q
->queue_lock
);
3651 cfq_shutdown_timer_wq(cfqd
);
3653 /* Wait for cfqg->blkg->key accessors to exit their grace periods. */
3654 call_rcu(&cfqd
->rcu
, cfq_cfqd_free
);
3657 static void *cfq_init_queue(struct request_queue
*q
)
3659 struct cfq_data
*cfqd
;
3661 struct cfq_group
*cfqg
;
3662 struct cfq_rb_root
*st
;
3664 cfqd
= kmalloc_node(sizeof(*cfqd
), GFP_KERNEL
| __GFP_ZERO
, q
->node
);
3668 /* Init root service tree */
3669 cfqd
->grp_service_tree
= CFQ_RB_ROOT
;
3671 /* Init root group */
3672 cfqg
= &cfqd
->root_group
;
3673 for_each_cfqg_st(cfqg
, i
, j
, st
)
3675 RB_CLEAR_NODE(&cfqg
->rb_node
);
3677 /* Give preference to root group over other groups */
3678 cfqg
->weight
= 2*BLKIO_WEIGHT_DEFAULT
;
3680 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3682 * Take a reference to root group which we never drop. This is just
3683 * to make sure that cfq_put_cfqg() does not try to kfree root group
3685 atomic_set(&cfqg
->ref
, 1);
3686 blkiocg_add_blkio_group(&blkio_root_cgroup
, &cfqg
->blkg
, (void *)cfqd
,
3690 * Not strictly needed (since RB_ROOT just clears the node and we
3691 * zeroed cfqd on alloc), but better be safe in case someone decides
3692 * to add magic to the rb code
3694 for (i
= 0; i
< CFQ_PRIO_LISTS
; i
++)
3695 cfqd
->prio_trees
[i
] = RB_ROOT
;
3698 * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
3699 * Grab a permanent reference to it, so that the normal code flow
3700 * will not attempt to free it.
3702 cfq_init_cfqq(cfqd
, &cfqd
->oom_cfqq
, 1, 0);
3703 atomic_inc(&cfqd
->oom_cfqq
.ref
);
3704 cfq_link_cfqq_cfqg(&cfqd
->oom_cfqq
, &cfqd
->root_group
);
3706 INIT_LIST_HEAD(&cfqd
->cic_list
);
3710 init_timer(&cfqd
->idle_slice_timer
);
3711 cfqd
->idle_slice_timer
.function
= cfq_idle_slice_timer
;
3712 cfqd
->idle_slice_timer
.data
= (unsigned long) cfqd
;
3714 INIT_WORK(&cfqd
->unplug_work
, cfq_kick_queue
);
3716 cfqd
->cfq_quantum
= cfq_quantum
;
3717 cfqd
->cfq_fifo_expire
[0] = cfq_fifo_expire
[0];
3718 cfqd
->cfq_fifo_expire
[1] = cfq_fifo_expire
[1];
3719 cfqd
->cfq_back_max
= cfq_back_max
;
3720 cfqd
->cfq_back_penalty
= cfq_back_penalty
;
3721 cfqd
->cfq_slice
[0] = cfq_slice_async
;
3722 cfqd
->cfq_slice
[1] = cfq_slice_sync
;
3723 cfqd
->cfq_slice_async_rq
= cfq_slice_async_rq
;
3724 cfqd
->cfq_slice_idle
= cfq_slice_idle
;
3725 cfqd
->cfq_latency
= 1;
3726 cfqd
->cfq_group_isolation
= 0;
3729 * we optimistically start assuming sync ops weren't delayed in last
3730 * second, in order to have larger depth for async operations.
3732 cfqd
->last_delayed_sync
= jiffies
- HZ
;
3733 INIT_RCU_HEAD(&cfqd
->rcu
);
3737 static void cfq_slab_kill(void)
3740 * Caller already ensured that pending RCU callbacks are completed,
3741 * so we should have no busy allocations at this point.
3744 kmem_cache_destroy(cfq_pool
);
3746 kmem_cache_destroy(cfq_ioc_pool
);
3749 static int __init
cfq_slab_setup(void)
3751 cfq_pool
= KMEM_CACHE(cfq_queue
, 0);
3755 cfq_ioc_pool
= KMEM_CACHE(cfq_io_context
, 0);
3766 * sysfs parts below -->
3769 cfq_var_show(unsigned int var
, char *page
)
3771 return sprintf(page
, "%d\n", var
);
3775 cfq_var_store(unsigned int *var
, const char *page
, size_t count
)
3777 char *p
= (char *) page
;
3779 *var
= simple_strtoul(p
, &p
, 10);
3783 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
3784 static ssize_t __FUNC(struct elevator_queue *e, char *page) \
3786 struct cfq_data *cfqd = e->elevator_data; \
3787 unsigned int __data = __VAR; \
3789 __data = jiffies_to_msecs(__data); \
3790 return cfq_var_show(__data, (page)); \
3792 SHOW_FUNCTION(cfq_quantum_show
, cfqd
->cfq_quantum
, 0);
3793 SHOW_FUNCTION(cfq_fifo_expire_sync_show
, cfqd
->cfq_fifo_expire
[1], 1);
3794 SHOW_FUNCTION(cfq_fifo_expire_async_show
, cfqd
->cfq_fifo_expire
[0], 1);
3795 SHOW_FUNCTION(cfq_back_seek_max_show
, cfqd
->cfq_back_max
, 0);
3796 SHOW_FUNCTION(cfq_back_seek_penalty_show
, cfqd
->cfq_back_penalty
, 0);
3797 SHOW_FUNCTION(cfq_slice_idle_show
, cfqd
->cfq_slice_idle
, 1);
3798 SHOW_FUNCTION(cfq_slice_sync_show
, cfqd
->cfq_slice
[1], 1);
3799 SHOW_FUNCTION(cfq_slice_async_show
, cfqd
->cfq_slice
[0], 1);
3800 SHOW_FUNCTION(cfq_slice_async_rq_show
, cfqd
->cfq_slice_async_rq
, 0);
3801 SHOW_FUNCTION(cfq_low_latency_show
, cfqd
->cfq_latency
, 0);
3802 SHOW_FUNCTION(cfq_group_isolation_show
, cfqd
->cfq_group_isolation
, 0);
3803 #undef SHOW_FUNCTION
3805 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
3806 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
3808 struct cfq_data *cfqd = e->elevator_data; \
3809 unsigned int __data; \
3810 int ret = cfq_var_store(&__data, (page), count); \
3811 if (__data < (MIN)) \
3813 else if (__data > (MAX)) \
3816 *(__PTR) = msecs_to_jiffies(__data); \
3818 *(__PTR) = __data; \
3821 STORE_FUNCTION(cfq_quantum_store
, &cfqd
->cfq_quantum
, 1, UINT_MAX
, 0);
3822 STORE_FUNCTION(cfq_fifo_expire_sync_store
, &cfqd
->cfq_fifo_expire
[1], 1,
3824 STORE_FUNCTION(cfq_fifo_expire_async_store
, &cfqd
->cfq_fifo_expire
[0], 1,
3826 STORE_FUNCTION(cfq_back_seek_max_store
, &cfqd
->cfq_back_max
, 0, UINT_MAX
, 0);
3827 STORE_FUNCTION(cfq_back_seek_penalty_store
, &cfqd
->cfq_back_penalty
, 1,
3829 STORE_FUNCTION(cfq_slice_idle_store
, &cfqd
->cfq_slice_idle
, 0, UINT_MAX
, 1);
3830 STORE_FUNCTION(cfq_slice_sync_store
, &cfqd
->cfq_slice
[1], 1, UINT_MAX
, 1);
3831 STORE_FUNCTION(cfq_slice_async_store
, &cfqd
->cfq_slice
[0], 1, UINT_MAX
, 1);
3832 STORE_FUNCTION(cfq_slice_async_rq_store
, &cfqd
->cfq_slice_async_rq
, 1,
3834 STORE_FUNCTION(cfq_low_latency_store
, &cfqd
->cfq_latency
, 0, 1, 0);
3835 STORE_FUNCTION(cfq_group_isolation_store
, &cfqd
->cfq_group_isolation
, 0, 1, 0);
3836 #undef STORE_FUNCTION
3838 #define CFQ_ATTR(name) \
3839 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
3841 static struct elv_fs_entry cfq_attrs
[] = {
3843 CFQ_ATTR(fifo_expire_sync
),
3844 CFQ_ATTR(fifo_expire_async
),
3845 CFQ_ATTR(back_seek_max
),
3846 CFQ_ATTR(back_seek_penalty
),
3847 CFQ_ATTR(slice_sync
),
3848 CFQ_ATTR(slice_async
),
3849 CFQ_ATTR(slice_async_rq
),
3850 CFQ_ATTR(slice_idle
),
3851 CFQ_ATTR(low_latency
),
3852 CFQ_ATTR(group_isolation
),
3856 static struct elevator_type iosched_cfq
= {
3858 .elevator_merge_fn
= cfq_merge
,
3859 .elevator_merged_fn
= cfq_merged_request
,
3860 .elevator_merge_req_fn
= cfq_merged_requests
,
3861 .elevator_allow_merge_fn
= cfq_allow_merge
,
3862 .elevator_dispatch_fn
= cfq_dispatch_requests
,
3863 .elevator_add_req_fn
= cfq_insert_request
,
3864 .elevator_activate_req_fn
= cfq_activate_request
,
3865 .elevator_deactivate_req_fn
= cfq_deactivate_request
,
3866 .elevator_queue_empty_fn
= cfq_queue_empty
,
3867 .elevator_completed_req_fn
= cfq_completed_request
,
3868 .elevator_former_req_fn
= elv_rb_former_request
,
3869 .elevator_latter_req_fn
= elv_rb_latter_request
,
3870 .elevator_set_req_fn
= cfq_set_request
,
3871 .elevator_put_req_fn
= cfq_put_request
,
3872 .elevator_may_queue_fn
= cfq_may_queue
,
3873 .elevator_init_fn
= cfq_init_queue
,
3874 .elevator_exit_fn
= cfq_exit_queue
,
3875 .trim
= cfq_free_io_context
,
3877 .elevator_attrs
= cfq_attrs
,
3878 .elevator_name
= "cfq",
3879 .elevator_owner
= THIS_MODULE
,
3882 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3883 static struct blkio_policy_type blkio_policy_cfq
= {
3885 .blkio_unlink_group_fn
= cfq_unlink_blkio_group
,
3886 .blkio_update_group_weight_fn
= cfq_update_blkio_group_weight
,
3890 static struct blkio_policy_type blkio_policy_cfq
;
3893 static int __init
cfq_init(void)
3896 * could be 0 on HZ < 1000 setups
3898 if (!cfq_slice_async
)
3899 cfq_slice_async
= 1;
3900 if (!cfq_slice_idle
)
3903 if (cfq_slab_setup())
3906 elv_register(&iosched_cfq
);
3907 blkio_policy_register(&blkio_policy_cfq
);
3912 static void __exit
cfq_exit(void)
3914 DECLARE_COMPLETION_ONSTACK(all_gone
);
3915 blkio_policy_unregister(&blkio_policy_cfq
);
3916 elv_unregister(&iosched_cfq
);
3917 ioc_gone
= &all_gone
;
3918 /* ioc_gone's update must be visible before reading ioc_count */
3922 * this also protects us from entering cfq_slab_kill() with
3923 * pending RCU callbacks
3925 if (elv_ioc_count_read(cfq_ioc_count
))
3926 wait_for_completion(&all_gone
);
3930 module_init(cfq_init
);
3931 module_exit(cfq_exit
);
3933 MODULE_AUTHOR("Jens Axboe");
3934 MODULE_LICENSE("GPL");
3935 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");