2 * CFQ, or complete fairness queueing, disk scheduler.
4 * Based on ideas from a previously unfinished io
5 * scheduler (round robin per-process disk scheduling) and Andrea Arcangeli.
7 * Copyright (C) 2003 Jens Axboe <axboe@kernel.dk>
9 #include <linux/module.h>
10 #include <linux/slab.h>
11 #include <linux/blkdev.h>
12 #include <linux/elevator.h>
13 #include <linux/jiffies.h>
14 #include <linux/rbtree.h>
15 #include <linux/ioprio.h>
16 #include <linux/blktrace_api.h>
22 /* max queue in one round of service */
23 static const int cfq_quantum
= 8;
24 static const int cfq_fifo_expire
[2] = { HZ
/ 4, HZ
/ 8 };
25 /* maximum backwards seek, in KiB */
26 static const int cfq_back_max
= 16 * 1024;
27 /* penalty of a backwards seek */
28 static const int cfq_back_penalty
= 2;
29 static const int cfq_slice_sync
= HZ
/ 10;
30 static int cfq_slice_async
= HZ
/ 25;
31 static const int cfq_slice_async_rq
= 2;
32 static int cfq_slice_idle
= HZ
/ 125;
33 static int cfq_group_idle
= HZ
/ 125;
34 static const int cfq_target_latency
= HZ
* 3/10; /* 300 ms */
35 static const int cfq_hist_divisor
= 4;
38 * offset from end of service tree
40 #define CFQ_IDLE_DELAY (HZ / 5)
43 * below this threshold, we consider thinktime immediate
45 #define CFQ_MIN_TT (2)
47 #define CFQ_SLICE_SCALE (5)
48 #define CFQ_HW_QUEUE_MIN (5)
49 #define CFQ_SERVICE_SHIFT 12
51 #define CFQQ_SEEK_THR (sector_t)(8 * 100)
52 #define CFQQ_CLOSE_THR (sector_t)(8 * 1024)
53 #define CFQQ_SECT_THR_NONROT (sector_t)(2 * 32)
54 #define CFQQ_SEEKY(cfqq) (hweight32(cfqq->seek_history) > 32/8)
57 ((struct cfq_io_context *) (rq)->elevator_private[0])
58 #define RQ_CFQQ(rq) (struct cfq_queue *) ((rq)->elevator_private[1])
59 #define RQ_CFQG(rq) (struct cfq_group *) ((rq)->elevator_private[2])
61 static struct kmem_cache
*cfq_pool
;
62 static struct kmem_cache
*cfq_ioc_pool
;
64 static DEFINE_PER_CPU(unsigned long, cfq_ioc_count
);
65 static struct completion
*ioc_gone
;
66 static DEFINE_SPINLOCK(ioc_gone_lock
);
68 static DEFINE_SPINLOCK(cic_index_lock
);
69 static DEFINE_IDA(cic_index_ida
);
71 #define CFQ_PRIO_LISTS IOPRIO_BE_NR
72 #define cfq_class_idle(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
73 #define cfq_class_rt(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
75 #define sample_valid(samples) ((samples) > 80)
76 #define rb_entry_cfqg(node) rb_entry((node), struct cfq_group, rb_node)
79 * Most of our rbtree usage is for sorting with min extraction, so
80 * if we cache the leftmost node we don't have to walk down the tree
81 * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should
82 * move this into the elevator for the rq sorting as well.
88 unsigned total_weight
;
91 #define CFQ_RB_ROOT (struct cfq_rb_root) { .rb = RB_ROOT, .left = NULL, \
92 .count = 0, .min_vdisktime = 0, }
95 * Per process-grouping structure
100 /* various state flags, see below */
102 /* parent cfq_data */
103 struct cfq_data
*cfqd
;
104 /* service_tree member */
105 struct rb_node rb_node
;
106 /* service_tree key */
107 unsigned long rb_key
;
108 /* prio tree member */
109 struct rb_node p_node
;
110 /* prio tree root we belong to, if any */
111 struct rb_root
*p_root
;
112 /* sorted list of pending requests */
113 struct rb_root sort_list
;
114 /* if fifo isn't expired, next request to serve */
115 struct request
*next_rq
;
116 /* requests queued in sort_list */
118 /* currently allocated requests */
120 /* fifo list of requests in sort_list */
121 struct list_head fifo
;
123 /* time when queue got scheduled in to dispatch first request. */
124 unsigned long dispatch_start
;
125 unsigned int allocated_slice
;
126 unsigned int slice_dispatch
;
127 /* time when first request from queue completed and slice started. */
128 unsigned long slice_start
;
129 unsigned long slice_end
;
132 /* pending metadata requests */
134 /* number of requests that are on the dispatch list or inside driver */
137 /* io prio of this group */
138 unsigned short ioprio
, org_ioprio
;
139 unsigned short ioprio_class
, org_ioprio_class
;
144 sector_t last_request_pos
;
146 struct cfq_rb_root
*service_tree
;
147 struct cfq_queue
*new_cfqq
;
148 struct cfq_group
*cfqg
;
149 /* Number of sectors dispatched from queue in single dispatch round */
150 unsigned long nr_sectors
;
154 * First index in the service_trees.
155 * IDLE is handled separately, so it has negative index
165 * Second index in the service_trees.
169 SYNC_NOIDLE_WORKLOAD
= 1,
173 /* This is per cgroup per device grouping structure */
175 /* group service_tree member */
176 struct rb_node rb_node
;
178 /* group service_tree key */
182 /* number of cfqq currently on this group */
186 * Per group busy queus average. Useful for workload slice calc. We
187 * create the array for each prio class but at run time it is used
188 * only for RT and BE class and slot for IDLE class remains unused.
189 * This is primarily done to avoid confusion and a gcc warning.
191 unsigned int busy_queues_avg
[CFQ_PRIO_NR
];
193 * rr lists of queues with requests. We maintain service trees for
194 * RT and BE classes. These trees are subdivided in subclasses
195 * of SYNC, SYNC_NOIDLE and ASYNC based on workload type. For IDLE
196 * class there is no subclassification and all the cfq queues go on
197 * a single tree service_tree_idle.
198 * Counts are embedded in the cfq_rb_root
200 struct cfq_rb_root service_trees
[2][3];
201 struct cfq_rb_root service_tree_idle
;
203 unsigned long saved_workload_slice
;
204 enum wl_type_t saved_workload
;
205 enum wl_prio_t saved_serving_prio
;
206 struct blkio_group blkg
;
207 #ifdef CONFIG_CFQ_GROUP_IOSCHED
208 struct hlist_node cfqd_node
;
211 /* number of requests that are on the dispatch list or inside driver */
216 * Per block device queue structure
219 struct request_queue
*queue
;
220 /* Root service tree for cfq_groups */
221 struct cfq_rb_root grp_service_tree
;
222 struct cfq_group root_group
;
225 * The priority currently being served
227 enum wl_prio_t serving_prio
;
228 enum wl_type_t serving_type
;
229 unsigned long workload_expires
;
230 struct cfq_group
*serving_group
;
233 * Each priority tree is sorted by next_request position. These
234 * trees are used when determining if two or more queues are
235 * interleaving requests (see cfq_close_cooperator).
237 struct rb_root prio_trees
[CFQ_PRIO_LISTS
];
239 unsigned int busy_queues
;
240 unsigned int busy_sync_queues
;
246 * queue-depth detection
252 * -1 => indeterminate, (cfq will behave as if NCQ is present, to allow better detection)
253 * 1 => NCQ is present (hw_tag_est_depth is the estimated max depth)
256 int hw_tag_est_depth
;
257 unsigned int hw_tag_samples
;
260 * idle window management
262 struct timer_list idle_slice_timer
;
263 struct work_struct unplug_work
;
265 struct cfq_queue
*active_queue
;
266 struct cfq_io_context
*active_cic
;
269 * async queue for each priority case
271 struct cfq_queue
*async_cfqq
[2][IOPRIO_BE_NR
];
272 struct cfq_queue
*async_idle_cfqq
;
274 sector_t last_position
;
277 * tunables, see top of file
279 unsigned int cfq_quantum
;
280 unsigned int cfq_fifo_expire
[2];
281 unsigned int cfq_back_penalty
;
282 unsigned int cfq_back_max
;
283 unsigned int cfq_slice
[2];
284 unsigned int cfq_slice_async_rq
;
285 unsigned int cfq_slice_idle
;
286 unsigned int cfq_group_idle
;
287 unsigned int cfq_latency
;
289 unsigned int cic_index
;
290 struct list_head cic_list
;
293 * Fallback dummy cfqq for extreme OOM conditions
295 struct cfq_queue oom_cfqq
;
297 unsigned long last_delayed_sync
;
299 /* List of cfq groups being managed on this device*/
300 struct hlist_head cfqg_list
;
304 static struct cfq_group
*cfq_get_next_cfqg(struct cfq_data
*cfqd
);
306 static struct cfq_rb_root
*service_tree_for(struct cfq_group
*cfqg
,
313 if (prio
== IDLE_WORKLOAD
)
314 return &cfqg
->service_tree_idle
;
316 return &cfqg
->service_trees
[prio
][type
];
319 enum cfqq_state_flags
{
320 CFQ_CFQQ_FLAG_on_rr
= 0, /* on round-robin busy list */
321 CFQ_CFQQ_FLAG_wait_request
, /* waiting for a request */
322 CFQ_CFQQ_FLAG_must_dispatch
, /* must be allowed a dispatch */
323 CFQ_CFQQ_FLAG_must_alloc_slice
, /* per-slice must_alloc flag */
324 CFQ_CFQQ_FLAG_fifo_expire
, /* FIFO checked in this slice */
325 CFQ_CFQQ_FLAG_idle_window
, /* slice idling enabled */
326 CFQ_CFQQ_FLAG_prio_changed
, /* task priority has changed */
327 CFQ_CFQQ_FLAG_slice_new
, /* no requests dispatched in slice */
328 CFQ_CFQQ_FLAG_sync
, /* synchronous queue */
329 CFQ_CFQQ_FLAG_coop
, /* cfqq is shared */
330 CFQ_CFQQ_FLAG_split_coop
, /* shared cfqq will be splitted */
331 CFQ_CFQQ_FLAG_deep
, /* sync cfqq experienced large depth */
332 CFQ_CFQQ_FLAG_wait_busy
, /* Waiting for next request */
335 #define CFQ_CFQQ_FNS(name) \
336 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \
338 (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name); \
340 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \
342 (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \
344 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \
346 return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \
350 CFQ_CFQQ_FNS(wait_request
);
351 CFQ_CFQQ_FNS(must_dispatch
);
352 CFQ_CFQQ_FNS(must_alloc_slice
);
353 CFQ_CFQQ_FNS(fifo_expire
);
354 CFQ_CFQQ_FNS(idle_window
);
355 CFQ_CFQQ_FNS(prio_changed
);
356 CFQ_CFQQ_FNS(slice_new
);
359 CFQ_CFQQ_FNS(split_coop
);
361 CFQ_CFQQ_FNS(wait_busy
);
364 #ifdef CONFIG_CFQ_GROUP_IOSCHED
365 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
366 blk_add_trace_msg((cfqd)->queue, "cfq%d%c %s " fmt, (cfqq)->pid, \
367 cfq_cfqq_sync((cfqq)) ? 'S' : 'A', \
368 blkg_path(&(cfqq)->cfqg->blkg), ##args);
370 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...) \
371 blk_add_trace_msg((cfqd)->queue, "%s " fmt, \
372 blkg_path(&(cfqg)->blkg), ##args); \
375 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
376 blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args)
377 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...) do {} while (0);
379 #define cfq_log(cfqd, fmt, args...) \
380 blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
382 /* Traverses through cfq group service trees */
383 #define for_each_cfqg_st(cfqg, i, j, st) \
384 for (i = 0; i <= IDLE_WORKLOAD; i++) \
385 for (j = 0, st = i < IDLE_WORKLOAD ? &cfqg->service_trees[i][j]\
386 : &cfqg->service_tree_idle; \
387 (i < IDLE_WORKLOAD && j <= SYNC_WORKLOAD) || \
388 (i == IDLE_WORKLOAD && j == 0); \
389 j++, st = i < IDLE_WORKLOAD ? \
390 &cfqg->service_trees[i][j]: NULL) \
393 static inline bool iops_mode(struct cfq_data *cfqd)
396 * If we are not idling on queues and it is a NCQ drive, parallel
397 * execution of requests is on and measuring time is not possible
398 * in most of the cases until and unless we drive shallower queue
399 * depths and that becomes a performance bottleneck. In such cases
400 * switch to start providing fairness in terms of number of IOs.
402 if (!cfqd
->cfq_slice_idle
&& cfqd
->hw_tag
)
408 static inline enum wl_prio_t
cfqq_prio(struct cfq_queue
*cfqq
)
410 if (cfq_class_idle(cfqq
))
411 return IDLE_WORKLOAD
;
412 if (cfq_class_rt(cfqq
))
418 static enum wl_type_t
cfqq_type(struct cfq_queue
*cfqq
)
420 if (!cfq_cfqq_sync(cfqq
))
421 return ASYNC_WORKLOAD
;
422 if (!cfq_cfqq_idle_window(cfqq
))
423 return SYNC_NOIDLE_WORKLOAD
;
424 return SYNC_WORKLOAD
;
427 static inline int cfq_group_busy_queues_wl(enum wl_prio_t wl
,
428 struct cfq_data
*cfqd
,
429 struct cfq_group
*cfqg
)
431 if (wl
== IDLE_WORKLOAD
)
432 return cfqg
->service_tree_idle
.count
;
434 return cfqg
->service_trees
[wl
][ASYNC_WORKLOAD
].count
435 + cfqg
->service_trees
[wl
][SYNC_NOIDLE_WORKLOAD
].count
436 + cfqg
->service_trees
[wl
][SYNC_WORKLOAD
].count
;
439 static inline int cfqg_busy_async_queues(struct cfq_data
*cfqd
,
440 struct cfq_group
*cfqg
)
442 return cfqg
->service_trees
[RT_WORKLOAD
][ASYNC_WORKLOAD
].count
443 + cfqg
->service_trees
[BE_WORKLOAD
][ASYNC_WORKLOAD
].count
;
446 static void cfq_dispatch_insert(struct request_queue
*, struct request
*);
447 static struct cfq_queue
*cfq_get_queue(struct cfq_data
*, bool,
448 struct io_context
*, gfp_t
);
449 static struct cfq_io_context
*cfq_cic_lookup(struct cfq_data
*,
450 struct io_context
*);
452 static inline struct cfq_queue
*cic_to_cfqq(struct cfq_io_context
*cic
,
455 return cic
->cfqq
[is_sync
];
458 static inline void cic_set_cfqq(struct cfq_io_context
*cic
,
459 struct cfq_queue
*cfqq
, bool is_sync
)
461 cic
->cfqq
[is_sync
] = cfqq
;
464 #define CIC_DEAD_KEY 1ul
465 #define CIC_DEAD_INDEX_SHIFT 1
467 static inline void *cfqd_dead_key(struct cfq_data
*cfqd
)
469 return (void *)(cfqd
->cic_index
<< CIC_DEAD_INDEX_SHIFT
| CIC_DEAD_KEY
);
472 static inline struct cfq_data
*cic_to_cfqd(struct cfq_io_context
*cic
)
474 struct cfq_data
*cfqd
= cic
->key
;
476 if (unlikely((unsigned long) cfqd
& CIC_DEAD_KEY
))
483 * We regard a request as SYNC, if it's either a read or has the SYNC bit
484 * set (in which case it could also be direct WRITE).
486 static inline bool cfq_bio_sync(struct bio
*bio
)
488 return bio_data_dir(bio
) == READ
|| (bio
->bi_rw
& REQ_SYNC
);
492 * scheduler run of queue, if there are requests pending and no one in the
493 * driver that will restart queueing
495 static inline void cfq_schedule_dispatch(struct cfq_data
*cfqd
)
497 if (cfqd
->busy_queues
) {
498 cfq_log(cfqd
, "schedule dispatch");
499 kblockd_schedule_work(cfqd
->queue
, &cfqd
->unplug_work
);
503 static int cfq_queue_empty(struct request_queue
*q
)
505 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
507 return !cfqd
->rq_queued
;
511 * Scale schedule slice based on io priority. Use the sync time slice only
512 * if a queue is marked sync and has sync io queued. A sync queue with async
513 * io only, should not get full sync slice length.
515 static inline int cfq_prio_slice(struct cfq_data
*cfqd
, bool sync
,
518 const int base_slice
= cfqd
->cfq_slice
[sync
];
520 WARN_ON(prio
>= IOPRIO_BE_NR
);
522 return base_slice
+ (base_slice
/CFQ_SLICE_SCALE
* (4 - prio
));
526 cfq_prio_to_slice(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
528 return cfq_prio_slice(cfqd
, cfq_cfqq_sync(cfqq
), cfqq
->ioprio
);
531 static inline u64
cfq_scale_slice(unsigned long delta
, struct cfq_group
*cfqg
)
533 u64 d
= delta
<< CFQ_SERVICE_SHIFT
;
535 d
= d
* BLKIO_WEIGHT_DEFAULT
;
536 do_div(d
, cfqg
->weight
);
540 static inline u64
max_vdisktime(u64 min_vdisktime
, u64 vdisktime
)
542 s64 delta
= (s64
)(vdisktime
- min_vdisktime
);
544 min_vdisktime
= vdisktime
;
546 return min_vdisktime
;
549 static inline u64
min_vdisktime(u64 min_vdisktime
, u64 vdisktime
)
551 s64 delta
= (s64
)(vdisktime
- min_vdisktime
);
553 min_vdisktime
= vdisktime
;
555 return min_vdisktime
;
558 static void update_min_vdisktime(struct cfq_rb_root
*st
)
560 struct cfq_group
*cfqg
;
563 cfqg
= rb_entry_cfqg(st
->left
);
564 st
->min_vdisktime
= max_vdisktime(st
->min_vdisktime
,
570 * get averaged number of queues of RT/BE priority.
571 * average is updated, with a formula that gives more weight to higher numbers,
572 * to quickly follows sudden increases and decrease slowly
575 static inline unsigned cfq_group_get_avg_queues(struct cfq_data
*cfqd
,
576 struct cfq_group
*cfqg
, bool rt
)
578 unsigned min_q
, max_q
;
579 unsigned mult
= cfq_hist_divisor
- 1;
580 unsigned round
= cfq_hist_divisor
/ 2;
581 unsigned busy
= cfq_group_busy_queues_wl(rt
, cfqd
, cfqg
);
583 min_q
= min(cfqg
->busy_queues_avg
[rt
], busy
);
584 max_q
= max(cfqg
->busy_queues_avg
[rt
], busy
);
585 cfqg
->busy_queues_avg
[rt
] = (mult
* max_q
+ min_q
+ round
) /
587 return cfqg
->busy_queues_avg
[rt
];
590 static inline unsigned
591 cfq_group_slice(struct cfq_data
*cfqd
, struct cfq_group
*cfqg
)
593 struct cfq_rb_root
*st
= &cfqd
->grp_service_tree
;
595 return cfq_target_latency
* cfqg
->weight
/ st
->total_weight
;
598 static inline unsigned
599 cfq_scaled_cfqq_slice(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
601 unsigned slice
= cfq_prio_to_slice(cfqd
, cfqq
);
602 if (cfqd
->cfq_latency
) {
604 * interested queues (we consider only the ones with the same
605 * priority class in the cfq group)
607 unsigned iq
= cfq_group_get_avg_queues(cfqd
, cfqq
->cfqg
,
609 unsigned sync_slice
= cfqd
->cfq_slice
[1];
610 unsigned expect_latency
= sync_slice
* iq
;
611 unsigned group_slice
= cfq_group_slice(cfqd
, cfqq
->cfqg
);
613 if (expect_latency
> group_slice
) {
614 unsigned base_low_slice
= 2 * cfqd
->cfq_slice_idle
;
615 /* scale low_slice according to IO priority
616 * and sync vs async */
618 min(slice
, base_low_slice
* slice
/ sync_slice
);
619 /* the adapted slice value is scaled to fit all iqs
620 * into the target latency */
621 slice
= max(slice
* group_slice
/ expect_latency
,
629 cfq_set_prio_slice(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
631 unsigned slice
= cfq_scaled_cfqq_slice(cfqd
, cfqq
);
633 cfqq
->slice_start
= jiffies
;
634 cfqq
->slice_end
= jiffies
+ slice
;
635 cfqq
->allocated_slice
= slice
;
636 cfq_log_cfqq(cfqd
, cfqq
, "set_slice=%lu", cfqq
->slice_end
- jiffies
);
640 * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
641 * isn't valid until the first request from the dispatch is activated
642 * and the slice time set.
644 static inline bool cfq_slice_used(struct cfq_queue
*cfqq
)
646 if (cfq_cfqq_slice_new(cfqq
))
648 if (time_before(jiffies
, cfqq
->slice_end
))
655 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
656 * We choose the request that is closest to the head right now. Distance
657 * behind the head is penalized and only allowed to a certain extent.
659 static struct request
*
660 cfq_choose_req(struct cfq_data
*cfqd
, struct request
*rq1
, struct request
*rq2
, sector_t last
)
662 sector_t s1
, s2
, d1
= 0, d2
= 0;
663 unsigned long back_max
;
664 #define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */
665 #define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */
666 unsigned wrap
= 0; /* bit mask: requests behind the disk head? */
668 if (rq1
== NULL
|| rq1
== rq2
)
673 if (rq_is_sync(rq1
) && !rq_is_sync(rq2
))
675 else if (rq_is_sync(rq2
) && !rq_is_sync(rq1
))
677 if ((rq1
->cmd_flags
& REQ_META
) && !(rq2
->cmd_flags
& REQ_META
))
679 else if ((rq2
->cmd_flags
& REQ_META
) &&
680 !(rq1
->cmd_flags
& REQ_META
))
683 s1
= blk_rq_pos(rq1
);
684 s2
= blk_rq_pos(rq2
);
687 * by definition, 1KiB is 2 sectors
689 back_max
= cfqd
->cfq_back_max
* 2;
692 * Strict one way elevator _except_ in the case where we allow
693 * short backward seeks which are biased as twice the cost of a
694 * similar forward seek.
698 else if (s1
+ back_max
>= last
)
699 d1
= (last
- s1
) * cfqd
->cfq_back_penalty
;
701 wrap
|= CFQ_RQ1_WRAP
;
705 else if (s2
+ back_max
>= last
)
706 d2
= (last
- s2
) * cfqd
->cfq_back_penalty
;
708 wrap
|= CFQ_RQ2_WRAP
;
710 /* Found required data */
713 * By doing switch() on the bit mask "wrap" we avoid having to
714 * check two variables for all permutations: --> faster!
717 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
733 case (CFQ_RQ1_WRAP
|CFQ_RQ2_WRAP
): /* both rqs wrapped */
736 * Since both rqs are wrapped,
737 * start with the one that's further behind head
738 * (--> only *one* back seek required),
739 * since back seek takes more time than forward.
749 * The below is leftmost cache rbtree addon
751 static struct cfq_queue
*cfq_rb_first(struct cfq_rb_root
*root
)
753 /* Service tree is empty */
758 root
->left
= rb_first(&root
->rb
);
761 return rb_entry(root
->left
, struct cfq_queue
, rb_node
);
766 static struct cfq_group
*cfq_rb_first_group(struct cfq_rb_root
*root
)
769 root
->left
= rb_first(&root
->rb
);
772 return rb_entry_cfqg(root
->left
);
777 static void rb_erase_init(struct rb_node
*n
, struct rb_root
*root
)
783 static void cfq_rb_erase(struct rb_node
*n
, struct cfq_rb_root
*root
)
787 rb_erase_init(n
, &root
->rb
);
792 * would be nice to take fifo expire time into account as well
794 static struct request
*
795 cfq_find_next_rq(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
796 struct request
*last
)
798 struct rb_node
*rbnext
= rb_next(&last
->rb_node
);
799 struct rb_node
*rbprev
= rb_prev(&last
->rb_node
);
800 struct request
*next
= NULL
, *prev
= NULL
;
802 BUG_ON(RB_EMPTY_NODE(&last
->rb_node
));
805 prev
= rb_entry_rq(rbprev
);
808 next
= rb_entry_rq(rbnext
);
810 rbnext
= rb_first(&cfqq
->sort_list
);
811 if (rbnext
&& rbnext
!= &last
->rb_node
)
812 next
= rb_entry_rq(rbnext
);
815 return cfq_choose_req(cfqd
, next
, prev
, blk_rq_pos(last
));
818 static unsigned long cfq_slice_offset(struct cfq_data
*cfqd
,
819 struct cfq_queue
*cfqq
)
822 * just an approximation, should be ok.
824 return (cfqq
->cfqg
->nr_cfqq
- 1) * (cfq_prio_slice(cfqd
, 1, 0) -
825 cfq_prio_slice(cfqd
, cfq_cfqq_sync(cfqq
), cfqq
->ioprio
));
829 cfqg_key(struct cfq_rb_root
*st
, struct cfq_group
*cfqg
)
831 return cfqg
->vdisktime
- st
->min_vdisktime
;
835 __cfq_group_service_tree_add(struct cfq_rb_root
*st
, struct cfq_group
*cfqg
)
837 struct rb_node
**node
= &st
->rb
.rb_node
;
838 struct rb_node
*parent
= NULL
;
839 struct cfq_group
*__cfqg
;
840 s64 key
= cfqg_key(st
, cfqg
);
843 while (*node
!= NULL
) {
845 __cfqg
= rb_entry_cfqg(parent
);
847 if (key
< cfqg_key(st
, __cfqg
))
848 node
= &parent
->rb_left
;
850 node
= &parent
->rb_right
;
856 st
->left
= &cfqg
->rb_node
;
858 rb_link_node(&cfqg
->rb_node
, parent
, node
);
859 rb_insert_color(&cfqg
->rb_node
, &st
->rb
);
863 cfq_group_service_tree_add(struct cfq_data
*cfqd
, struct cfq_group
*cfqg
)
865 struct cfq_rb_root
*st
= &cfqd
->grp_service_tree
;
866 struct cfq_group
*__cfqg
;
870 if (!RB_EMPTY_NODE(&cfqg
->rb_node
))
874 * Currently put the group at the end. Later implement something
875 * so that groups get lesser vtime based on their weights, so that
876 * if group does not loose all if it was not continously backlogged.
878 n
= rb_last(&st
->rb
);
880 __cfqg
= rb_entry_cfqg(n
);
881 cfqg
->vdisktime
= __cfqg
->vdisktime
+ CFQ_IDLE_DELAY
;
883 cfqg
->vdisktime
= st
->min_vdisktime
;
885 __cfq_group_service_tree_add(st
, cfqg
);
886 st
->total_weight
+= cfqg
->weight
;
890 cfq_group_service_tree_del(struct cfq_data
*cfqd
, struct cfq_group
*cfqg
)
892 struct cfq_rb_root
*st
= &cfqd
->grp_service_tree
;
894 BUG_ON(cfqg
->nr_cfqq
< 1);
897 /* If there are other cfq queues under this group, don't delete it */
901 cfq_log_cfqg(cfqd
, cfqg
, "del_from_rr group");
902 st
->total_weight
-= cfqg
->weight
;
903 if (!RB_EMPTY_NODE(&cfqg
->rb_node
))
904 cfq_rb_erase(&cfqg
->rb_node
, st
);
905 cfqg
->saved_workload_slice
= 0;
906 cfq_blkiocg_update_dequeue_stats(&cfqg
->blkg
, 1);
909 static inline unsigned int cfq_cfqq_slice_usage(struct cfq_queue
*cfqq
)
911 unsigned int slice_used
;
914 * Queue got expired before even a single request completed or
915 * got expired immediately after first request completion.
917 if (!cfqq
->slice_start
|| cfqq
->slice_start
== jiffies
) {
919 * Also charge the seek time incurred to the group, otherwise
920 * if there are mutiple queues in the group, each can dispatch
921 * a single request on seeky media and cause lots of seek time
922 * and group will never know it.
924 slice_used
= max_t(unsigned, (jiffies
- cfqq
->dispatch_start
),
927 slice_used
= jiffies
- cfqq
->slice_start
;
928 if (slice_used
> cfqq
->allocated_slice
)
929 slice_used
= cfqq
->allocated_slice
;
935 static void cfq_group_served(struct cfq_data
*cfqd
, struct cfq_group
*cfqg
,
936 struct cfq_queue
*cfqq
)
938 struct cfq_rb_root
*st
= &cfqd
->grp_service_tree
;
939 unsigned int used_sl
, charge
;
940 int nr_sync
= cfqg
->nr_cfqq
- cfqg_busy_async_queues(cfqd
, cfqg
)
941 - cfqg
->service_tree_idle
.count
;
944 used_sl
= charge
= cfq_cfqq_slice_usage(cfqq
);
947 charge
= cfqq
->slice_dispatch
;
948 else if (!cfq_cfqq_sync(cfqq
) && !nr_sync
)
949 charge
= cfqq
->allocated_slice
;
951 /* Can't update vdisktime while group is on service tree */
952 cfq_rb_erase(&cfqg
->rb_node
, st
);
953 cfqg
->vdisktime
+= cfq_scale_slice(charge
, cfqg
);
954 __cfq_group_service_tree_add(st
, cfqg
);
956 /* This group is being expired. Save the context */
957 if (time_after(cfqd
->workload_expires
, jiffies
)) {
958 cfqg
->saved_workload_slice
= cfqd
->workload_expires
960 cfqg
->saved_workload
= cfqd
->serving_type
;
961 cfqg
->saved_serving_prio
= cfqd
->serving_prio
;
963 cfqg
->saved_workload_slice
= 0;
965 cfq_log_cfqg(cfqd
, cfqg
, "served: vt=%llu min_vt=%llu", cfqg
->vdisktime
,
967 cfq_log_cfqq(cfqq
->cfqd
, cfqq
, "sl_used=%u disp=%u charge=%u iops=%u"
968 " sect=%u", used_sl
, cfqq
->slice_dispatch
, charge
,
969 iops_mode(cfqd
), cfqq
->nr_sectors
);
970 cfq_blkiocg_update_timeslice_used(&cfqg
->blkg
, used_sl
);
971 cfq_blkiocg_set_start_empty_time(&cfqg
->blkg
);
974 #ifdef CONFIG_CFQ_GROUP_IOSCHED
975 static inline struct cfq_group
*cfqg_of_blkg(struct blkio_group
*blkg
)
978 return container_of(blkg
, struct cfq_group
, blkg
);
982 void cfq_update_blkio_group_weight(void *key
, struct blkio_group
*blkg
,
985 cfqg_of_blkg(blkg
)->weight
= weight
;
988 static struct cfq_group
*
989 cfq_find_alloc_cfqg(struct cfq_data
*cfqd
, struct cgroup
*cgroup
, int create
)
991 struct blkio_cgroup
*blkcg
= cgroup_to_blkio_cgroup(cgroup
);
992 struct cfq_group
*cfqg
= NULL
;
995 struct cfq_rb_root
*st
;
996 struct backing_dev_info
*bdi
= &cfqd
->queue
->backing_dev_info
;
997 unsigned int major
, minor
;
999 cfqg
= cfqg_of_blkg(blkiocg_lookup_group(blkcg
, key
));
1000 if (cfqg
&& !cfqg
->blkg
.dev
&& bdi
->dev
&& dev_name(bdi
->dev
)) {
1001 sscanf(dev_name(bdi
->dev
), "%u:%u", &major
, &minor
);
1002 cfqg
->blkg
.dev
= MKDEV(major
, minor
);
1005 if (cfqg
|| !create
)
1008 cfqg
= kzalloc_node(sizeof(*cfqg
), GFP_ATOMIC
, cfqd
->queue
->node
);
1012 for_each_cfqg_st(cfqg
, i
, j
, st
)
1014 RB_CLEAR_NODE(&cfqg
->rb_node
);
1017 * Take the initial reference that will be released on destroy
1018 * This can be thought of a joint reference by cgroup and
1019 * elevator which will be dropped by either elevator exit
1020 * or cgroup deletion path depending on who is exiting first.
1025 * Add group onto cgroup list. It might happen that bdi->dev is
1026 * not initialized yet. Initialize this new group without major
1027 * and minor info and this info will be filled in once a new thread
1028 * comes for IO. See code above.
1031 sscanf(dev_name(bdi
->dev
), "%u:%u", &major
, &minor
);
1032 cfq_blkiocg_add_blkio_group(blkcg
, &cfqg
->blkg
, (void *)cfqd
,
1033 MKDEV(major
, minor
));
1035 cfq_blkiocg_add_blkio_group(blkcg
, &cfqg
->blkg
, (void *)cfqd
,
1038 cfqg
->weight
= blkcg_get_weight(blkcg
, cfqg
->blkg
.dev
);
1040 /* Add group on cfqd list */
1041 hlist_add_head(&cfqg
->cfqd_node
, &cfqd
->cfqg_list
);
1048 * Search for the cfq group current task belongs to. If create = 1, then also
1049 * create the cfq group if it does not exist. request_queue lock must be held.
1051 static struct cfq_group
*cfq_get_cfqg(struct cfq_data
*cfqd
, int create
)
1053 struct cgroup
*cgroup
;
1054 struct cfq_group
*cfqg
= NULL
;
1057 cgroup
= task_cgroup(current
, blkio_subsys_id
);
1058 cfqg
= cfq_find_alloc_cfqg(cfqd
, cgroup
, create
);
1059 if (!cfqg
&& create
)
1060 cfqg
= &cfqd
->root_group
;
1065 static inline struct cfq_group
*cfq_ref_get_cfqg(struct cfq_group
*cfqg
)
1071 static void cfq_link_cfqq_cfqg(struct cfq_queue
*cfqq
, struct cfq_group
*cfqg
)
1073 /* Currently, all async queues are mapped to root group */
1074 if (!cfq_cfqq_sync(cfqq
))
1075 cfqg
= &cfqq
->cfqd
->root_group
;
1078 /* cfqq reference on cfqg */
1082 static void cfq_put_cfqg(struct cfq_group
*cfqg
)
1084 struct cfq_rb_root
*st
;
1087 BUG_ON(cfqg
->ref
<= 0);
1091 for_each_cfqg_st(cfqg
, i
, j
, st
)
1092 BUG_ON(!RB_EMPTY_ROOT(&st
->rb
));
1096 static void cfq_destroy_cfqg(struct cfq_data
*cfqd
, struct cfq_group
*cfqg
)
1098 /* Something wrong if we are trying to remove same group twice */
1099 BUG_ON(hlist_unhashed(&cfqg
->cfqd_node
));
1101 hlist_del_init(&cfqg
->cfqd_node
);
1104 * Put the reference taken at the time of creation so that when all
1105 * queues are gone, group can be destroyed.
1110 static void cfq_release_cfq_groups(struct cfq_data
*cfqd
)
1112 struct hlist_node
*pos
, *n
;
1113 struct cfq_group
*cfqg
;
1115 hlist_for_each_entry_safe(cfqg
, pos
, n
, &cfqd
->cfqg_list
, cfqd_node
) {
1117 * If cgroup removal path got to blk_group first and removed
1118 * it from cgroup list, then it will take care of destroying
1121 if (!cfq_blkiocg_del_blkio_group(&cfqg
->blkg
))
1122 cfq_destroy_cfqg(cfqd
, cfqg
);
1127 * Blk cgroup controller notification saying that blkio_group object is being
1128 * delinked as associated cgroup object is going away. That also means that
1129 * no new IO will come in this group. So get rid of this group as soon as
1130 * any pending IO in the group is finished.
1132 * This function is called under rcu_read_lock(). key is the rcu protected
1133 * pointer. That means "key" is a valid cfq_data pointer as long as we are rcu
1136 * "key" was fetched from blkio_group under blkio_cgroup->lock. That means
1137 * it should not be NULL as even if elevator was exiting, cgroup deltion
1138 * path got to it first.
1140 void cfq_unlink_blkio_group(void *key
, struct blkio_group
*blkg
)
1142 unsigned long flags
;
1143 struct cfq_data
*cfqd
= key
;
1145 spin_lock_irqsave(cfqd
->queue
->queue_lock
, flags
);
1146 cfq_destroy_cfqg(cfqd
, cfqg_of_blkg(blkg
));
1147 spin_unlock_irqrestore(cfqd
->queue
->queue_lock
, flags
);
1150 #else /* GROUP_IOSCHED */
1151 static struct cfq_group
*cfq_get_cfqg(struct cfq_data
*cfqd
, int create
)
1153 return &cfqd
->root_group
;
1156 static inline struct cfq_group
*cfq_ref_get_cfqg(struct cfq_group
*cfqg
)
1162 cfq_link_cfqq_cfqg(struct cfq_queue
*cfqq
, struct cfq_group
*cfqg
) {
1166 static void cfq_release_cfq_groups(struct cfq_data
*cfqd
) {}
1167 static inline void cfq_put_cfqg(struct cfq_group
*cfqg
) {}
1169 #endif /* GROUP_IOSCHED */
1172 * The cfqd->service_trees holds all pending cfq_queue's that have
1173 * requests waiting to be processed. It is sorted in the order that
1174 * we will service the queues.
1176 static void cfq_service_tree_add(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
1179 struct rb_node
**p
, *parent
;
1180 struct cfq_queue
*__cfqq
;
1181 unsigned long rb_key
;
1182 struct cfq_rb_root
*service_tree
;
1185 int group_changed
= 0;
1187 service_tree
= service_tree_for(cfqq
->cfqg
, cfqq_prio(cfqq
),
1189 if (cfq_class_idle(cfqq
)) {
1190 rb_key
= CFQ_IDLE_DELAY
;
1191 parent
= rb_last(&service_tree
->rb
);
1192 if (parent
&& parent
!= &cfqq
->rb_node
) {
1193 __cfqq
= rb_entry(parent
, struct cfq_queue
, rb_node
);
1194 rb_key
+= __cfqq
->rb_key
;
1197 } else if (!add_front
) {
1199 * Get our rb key offset. Subtract any residual slice
1200 * value carried from last service. A negative resid
1201 * count indicates slice overrun, and this should position
1202 * the next service time further away in the tree.
1204 rb_key
= cfq_slice_offset(cfqd
, cfqq
) + jiffies
;
1205 rb_key
-= cfqq
->slice_resid
;
1206 cfqq
->slice_resid
= 0;
1209 __cfqq
= cfq_rb_first(service_tree
);
1210 rb_key
+= __cfqq
? __cfqq
->rb_key
: jiffies
;
1213 if (!RB_EMPTY_NODE(&cfqq
->rb_node
)) {
1216 * same position, nothing more to do
1218 if (rb_key
== cfqq
->rb_key
&&
1219 cfqq
->service_tree
== service_tree
)
1222 cfq_rb_erase(&cfqq
->rb_node
, cfqq
->service_tree
);
1223 cfqq
->service_tree
= NULL
;
1228 cfqq
->service_tree
= service_tree
;
1229 p
= &service_tree
->rb
.rb_node
;
1234 __cfqq
= rb_entry(parent
, struct cfq_queue
, rb_node
);
1237 * sort by key, that represents service time.
1239 if (time_before(rb_key
, __cfqq
->rb_key
))
1242 n
= &(*p
)->rb_right
;
1250 service_tree
->left
= &cfqq
->rb_node
;
1252 cfqq
->rb_key
= rb_key
;
1253 rb_link_node(&cfqq
->rb_node
, parent
, p
);
1254 rb_insert_color(&cfqq
->rb_node
, &service_tree
->rb
);
1255 service_tree
->count
++;
1256 if ((add_front
|| !new_cfqq
) && !group_changed
)
1258 cfq_group_service_tree_add(cfqd
, cfqq
->cfqg
);
1261 static struct cfq_queue
*
1262 cfq_prio_tree_lookup(struct cfq_data
*cfqd
, struct rb_root
*root
,
1263 sector_t sector
, struct rb_node
**ret_parent
,
1264 struct rb_node
***rb_link
)
1266 struct rb_node
**p
, *parent
;
1267 struct cfq_queue
*cfqq
= NULL
;
1275 cfqq
= rb_entry(parent
, struct cfq_queue
, p_node
);
1278 * Sort strictly based on sector. Smallest to the left,
1279 * largest to the right.
1281 if (sector
> blk_rq_pos(cfqq
->next_rq
))
1282 n
= &(*p
)->rb_right
;
1283 else if (sector
< blk_rq_pos(cfqq
->next_rq
))
1291 *ret_parent
= parent
;
1297 static void cfq_prio_tree_add(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
1299 struct rb_node
**p
, *parent
;
1300 struct cfq_queue
*__cfqq
;
1303 rb_erase(&cfqq
->p_node
, cfqq
->p_root
);
1304 cfqq
->p_root
= NULL
;
1307 if (cfq_class_idle(cfqq
))
1312 cfqq
->p_root
= &cfqd
->prio_trees
[cfqq
->org_ioprio
];
1313 __cfqq
= cfq_prio_tree_lookup(cfqd
, cfqq
->p_root
,
1314 blk_rq_pos(cfqq
->next_rq
), &parent
, &p
);
1316 rb_link_node(&cfqq
->p_node
, parent
, p
);
1317 rb_insert_color(&cfqq
->p_node
, cfqq
->p_root
);
1319 cfqq
->p_root
= NULL
;
1323 * Update cfqq's position in the service tree.
1325 static void cfq_resort_rr_list(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
1328 * Resorting requires the cfqq to be on the RR list already.
1330 if (cfq_cfqq_on_rr(cfqq
)) {
1331 cfq_service_tree_add(cfqd
, cfqq
, 0);
1332 cfq_prio_tree_add(cfqd
, cfqq
);
1337 * add to busy list of queues for service, trying to be fair in ordering
1338 * the pending list according to last request service
1340 static void cfq_add_cfqq_rr(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
1342 cfq_log_cfqq(cfqd
, cfqq
, "add_to_rr");
1343 BUG_ON(cfq_cfqq_on_rr(cfqq
));
1344 cfq_mark_cfqq_on_rr(cfqq
);
1345 cfqd
->busy_queues
++;
1346 if (cfq_cfqq_sync(cfqq
))
1347 cfqd
->busy_sync_queues
++;
1349 cfq_resort_rr_list(cfqd
, cfqq
);
1353 * Called when the cfqq no longer has requests pending, remove it from
1356 static void cfq_del_cfqq_rr(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
1358 cfq_log_cfqq(cfqd
, cfqq
, "del_from_rr");
1359 BUG_ON(!cfq_cfqq_on_rr(cfqq
));
1360 cfq_clear_cfqq_on_rr(cfqq
);
1362 if (!RB_EMPTY_NODE(&cfqq
->rb_node
)) {
1363 cfq_rb_erase(&cfqq
->rb_node
, cfqq
->service_tree
);
1364 cfqq
->service_tree
= NULL
;
1367 rb_erase(&cfqq
->p_node
, cfqq
->p_root
);
1368 cfqq
->p_root
= NULL
;
1371 cfq_group_service_tree_del(cfqd
, cfqq
->cfqg
);
1372 BUG_ON(!cfqd
->busy_queues
);
1373 cfqd
->busy_queues
--;
1374 if (cfq_cfqq_sync(cfqq
))
1375 cfqd
->busy_sync_queues
--;
1379 * rb tree support functions
1381 static void cfq_del_rq_rb(struct request
*rq
)
1383 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
1384 const int sync
= rq_is_sync(rq
);
1386 BUG_ON(!cfqq
->queued
[sync
]);
1387 cfqq
->queued
[sync
]--;
1389 elv_rb_del(&cfqq
->sort_list
, rq
);
1391 if (cfq_cfqq_on_rr(cfqq
) && RB_EMPTY_ROOT(&cfqq
->sort_list
)) {
1393 * Queue will be deleted from service tree when we actually
1394 * expire it later. Right now just remove it from prio tree
1398 rb_erase(&cfqq
->p_node
, cfqq
->p_root
);
1399 cfqq
->p_root
= NULL
;
1404 static void cfq_add_rq_rb(struct request
*rq
)
1406 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
1407 struct cfq_data
*cfqd
= cfqq
->cfqd
;
1408 struct request
*__alias
, *prev
;
1410 cfqq
->queued
[rq_is_sync(rq
)]++;
1413 * looks a little odd, but the first insert might return an alias.
1414 * if that happens, put the alias on the dispatch list
1416 while ((__alias
= elv_rb_add(&cfqq
->sort_list
, rq
)) != NULL
)
1417 cfq_dispatch_insert(cfqd
->queue
, __alias
);
1419 if (!cfq_cfqq_on_rr(cfqq
))
1420 cfq_add_cfqq_rr(cfqd
, cfqq
);
1423 * check if this request is a better next-serve candidate
1425 prev
= cfqq
->next_rq
;
1426 cfqq
->next_rq
= cfq_choose_req(cfqd
, cfqq
->next_rq
, rq
, cfqd
->last_position
);
1429 * adjust priority tree position, if ->next_rq changes
1431 if (prev
!= cfqq
->next_rq
)
1432 cfq_prio_tree_add(cfqd
, cfqq
);
1434 BUG_ON(!cfqq
->next_rq
);
1437 static void cfq_reposition_rq_rb(struct cfq_queue
*cfqq
, struct request
*rq
)
1439 elv_rb_del(&cfqq
->sort_list
, rq
);
1440 cfqq
->queued
[rq_is_sync(rq
)]--;
1441 cfq_blkiocg_update_io_remove_stats(&(RQ_CFQG(rq
))->blkg
,
1442 rq_data_dir(rq
), rq_is_sync(rq
));
1444 cfq_blkiocg_update_io_add_stats(&(RQ_CFQG(rq
))->blkg
,
1445 &cfqq
->cfqd
->serving_group
->blkg
, rq_data_dir(rq
),
1449 static struct request
*
1450 cfq_find_rq_fmerge(struct cfq_data
*cfqd
, struct bio
*bio
)
1452 struct task_struct
*tsk
= current
;
1453 struct cfq_io_context
*cic
;
1454 struct cfq_queue
*cfqq
;
1456 cic
= cfq_cic_lookup(cfqd
, tsk
->io_context
);
1460 cfqq
= cic_to_cfqq(cic
, cfq_bio_sync(bio
));
1462 sector_t sector
= bio
->bi_sector
+ bio_sectors(bio
);
1464 return elv_rb_find(&cfqq
->sort_list
, sector
);
1470 static void cfq_activate_request(struct request_queue
*q
, struct request
*rq
)
1472 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
1474 cfqd
->rq_in_driver
++;
1475 cfq_log_cfqq(cfqd
, RQ_CFQQ(rq
), "activate rq, drv=%d",
1476 cfqd
->rq_in_driver
);
1478 cfqd
->last_position
= blk_rq_pos(rq
) + blk_rq_sectors(rq
);
1481 static void cfq_deactivate_request(struct request_queue
*q
, struct request
*rq
)
1483 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
1485 WARN_ON(!cfqd
->rq_in_driver
);
1486 cfqd
->rq_in_driver
--;
1487 cfq_log_cfqq(cfqd
, RQ_CFQQ(rq
), "deactivate rq, drv=%d",
1488 cfqd
->rq_in_driver
);
1491 static void cfq_remove_request(struct request
*rq
)
1493 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
1495 if (cfqq
->next_rq
== rq
)
1496 cfqq
->next_rq
= cfq_find_next_rq(cfqq
->cfqd
, cfqq
, rq
);
1498 list_del_init(&rq
->queuelist
);
1501 cfqq
->cfqd
->rq_queued
--;
1502 cfq_blkiocg_update_io_remove_stats(&(RQ_CFQG(rq
))->blkg
,
1503 rq_data_dir(rq
), rq_is_sync(rq
));
1504 if (rq
->cmd_flags
& REQ_META
) {
1505 WARN_ON(!cfqq
->meta_pending
);
1506 cfqq
->meta_pending
--;
1510 static int cfq_merge(struct request_queue
*q
, struct request
**req
,
1513 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
1514 struct request
*__rq
;
1516 __rq
= cfq_find_rq_fmerge(cfqd
, bio
);
1517 if (__rq
&& elv_rq_merge_ok(__rq
, bio
)) {
1519 return ELEVATOR_FRONT_MERGE
;
1522 return ELEVATOR_NO_MERGE
;
1525 static void cfq_merged_request(struct request_queue
*q
, struct request
*req
,
1528 if (type
== ELEVATOR_FRONT_MERGE
) {
1529 struct cfq_queue
*cfqq
= RQ_CFQQ(req
);
1531 cfq_reposition_rq_rb(cfqq
, req
);
1535 static void cfq_bio_merged(struct request_queue
*q
, struct request
*req
,
1538 cfq_blkiocg_update_io_merged_stats(&(RQ_CFQG(req
))->blkg
,
1539 bio_data_dir(bio
), cfq_bio_sync(bio
));
1543 cfq_merged_requests(struct request_queue
*q
, struct request
*rq
,
1544 struct request
*next
)
1546 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
1548 * reposition in fifo if next is older than rq
1550 if (!list_empty(&rq
->queuelist
) && !list_empty(&next
->queuelist
) &&
1551 time_before(rq_fifo_time(next
), rq_fifo_time(rq
))) {
1552 list_move(&rq
->queuelist
, &next
->queuelist
);
1553 rq_set_fifo_time(rq
, rq_fifo_time(next
));
1556 if (cfqq
->next_rq
== next
)
1558 cfq_remove_request(next
);
1559 cfq_blkiocg_update_io_merged_stats(&(RQ_CFQG(rq
))->blkg
,
1560 rq_data_dir(next
), rq_is_sync(next
));
1563 static int cfq_allow_merge(struct request_queue
*q
, struct request
*rq
,
1566 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
1567 struct cfq_io_context
*cic
;
1568 struct cfq_queue
*cfqq
;
1571 * Disallow merge of a sync bio into an async request.
1573 if (cfq_bio_sync(bio
) && !rq_is_sync(rq
))
1577 * Lookup the cfqq that this bio will be queued with. Allow
1578 * merge only if rq is queued there.
1580 cic
= cfq_cic_lookup(cfqd
, current
->io_context
);
1584 cfqq
= cic_to_cfqq(cic
, cfq_bio_sync(bio
));
1585 return cfqq
== RQ_CFQQ(rq
);
1588 static inline void cfq_del_timer(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
1590 del_timer(&cfqd
->idle_slice_timer
);
1591 cfq_blkiocg_update_idle_time_stats(&cfqq
->cfqg
->blkg
);
1594 static void __cfq_set_active_queue(struct cfq_data
*cfqd
,
1595 struct cfq_queue
*cfqq
)
1598 cfq_log_cfqq(cfqd
, cfqq
, "set_active wl_prio:%d wl_type:%d",
1599 cfqd
->serving_prio
, cfqd
->serving_type
);
1600 cfq_blkiocg_update_avg_queue_size_stats(&cfqq
->cfqg
->blkg
);
1601 cfqq
->slice_start
= 0;
1602 cfqq
->dispatch_start
= jiffies
;
1603 cfqq
->allocated_slice
= 0;
1604 cfqq
->slice_end
= 0;
1605 cfqq
->slice_dispatch
= 0;
1606 cfqq
->nr_sectors
= 0;
1608 cfq_clear_cfqq_wait_request(cfqq
);
1609 cfq_clear_cfqq_must_dispatch(cfqq
);
1610 cfq_clear_cfqq_must_alloc_slice(cfqq
);
1611 cfq_clear_cfqq_fifo_expire(cfqq
);
1612 cfq_mark_cfqq_slice_new(cfqq
);
1614 cfq_del_timer(cfqd
, cfqq
);
1617 cfqd
->active_queue
= cfqq
;
1621 * current cfqq expired its slice (or was too idle), select new one
1624 __cfq_slice_expired(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
1627 cfq_log_cfqq(cfqd
, cfqq
, "slice expired t=%d", timed_out
);
1629 if (cfq_cfqq_wait_request(cfqq
))
1630 cfq_del_timer(cfqd
, cfqq
);
1632 cfq_clear_cfqq_wait_request(cfqq
);
1633 cfq_clear_cfqq_wait_busy(cfqq
);
1636 * If this cfqq is shared between multiple processes, check to
1637 * make sure that those processes are still issuing I/Os within
1638 * the mean seek distance. If not, it may be time to break the
1639 * queues apart again.
1641 if (cfq_cfqq_coop(cfqq
) && CFQQ_SEEKY(cfqq
))
1642 cfq_mark_cfqq_split_coop(cfqq
);
1645 * store what was left of this slice, if the queue idled/timed out
1648 if (cfq_cfqq_slice_new(cfqq
))
1649 cfqq
->slice_resid
= cfq_scaled_cfqq_slice(cfqd
, cfqq
);
1651 cfqq
->slice_resid
= cfqq
->slice_end
- jiffies
;
1652 cfq_log_cfqq(cfqd
, cfqq
, "resid=%ld", cfqq
->slice_resid
);
1655 cfq_group_served(cfqd
, cfqq
->cfqg
, cfqq
);
1657 if (cfq_cfqq_on_rr(cfqq
) && RB_EMPTY_ROOT(&cfqq
->sort_list
))
1658 cfq_del_cfqq_rr(cfqd
, cfqq
);
1660 cfq_resort_rr_list(cfqd
, cfqq
);
1662 if (cfqq
== cfqd
->active_queue
)
1663 cfqd
->active_queue
= NULL
;
1665 if (cfqd
->active_cic
) {
1666 put_io_context(cfqd
->active_cic
->ioc
);
1667 cfqd
->active_cic
= NULL
;
1671 static inline void cfq_slice_expired(struct cfq_data
*cfqd
, bool timed_out
)
1673 struct cfq_queue
*cfqq
= cfqd
->active_queue
;
1676 __cfq_slice_expired(cfqd
, cfqq
, timed_out
);
1680 * Get next queue for service. Unless we have a queue preemption,
1681 * we'll simply select the first cfqq in the service tree.
1683 static struct cfq_queue
*cfq_get_next_queue(struct cfq_data
*cfqd
)
1685 struct cfq_rb_root
*service_tree
=
1686 service_tree_for(cfqd
->serving_group
, cfqd
->serving_prio
,
1687 cfqd
->serving_type
);
1689 if (!cfqd
->rq_queued
)
1692 /* There is nothing to dispatch */
1695 if (RB_EMPTY_ROOT(&service_tree
->rb
))
1697 return cfq_rb_first(service_tree
);
1700 static struct cfq_queue
*cfq_get_next_queue_forced(struct cfq_data
*cfqd
)
1702 struct cfq_group
*cfqg
;
1703 struct cfq_queue
*cfqq
;
1705 struct cfq_rb_root
*st
;
1707 if (!cfqd
->rq_queued
)
1710 cfqg
= cfq_get_next_cfqg(cfqd
);
1714 for_each_cfqg_st(cfqg
, i
, j
, st
)
1715 if ((cfqq
= cfq_rb_first(st
)) != NULL
)
1721 * Get and set a new active queue for service.
1723 static struct cfq_queue
*cfq_set_active_queue(struct cfq_data
*cfqd
,
1724 struct cfq_queue
*cfqq
)
1727 cfqq
= cfq_get_next_queue(cfqd
);
1729 __cfq_set_active_queue(cfqd
, cfqq
);
1733 static inline sector_t
cfq_dist_from_last(struct cfq_data
*cfqd
,
1736 if (blk_rq_pos(rq
) >= cfqd
->last_position
)
1737 return blk_rq_pos(rq
) - cfqd
->last_position
;
1739 return cfqd
->last_position
- blk_rq_pos(rq
);
1742 static inline int cfq_rq_close(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
1745 return cfq_dist_from_last(cfqd
, rq
) <= CFQQ_CLOSE_THR
;
1748 static struct cfq_queue
*cfqq_close(struct cfq_data
*cfqd
,
1749 struct cfq_queue
*cur_cfqq
)
1751 struct rb_root
*root
= &cfqd
->prio_trees
[cur_cfqq
->org_ioprio
];
1752 struct rb_node
*parent
, *node
;
1753 struct cfq_queue
*__cfqq
;
1754 sector_t sector
= cfqd
->last_position
;
1756 if (RB_EMPTY_ROOT(root
))
1760 * First, if we find a request starting at the end of the last
1761 * request, choose it.
1763 __cfqq
= cfq_prio_tree_lookup(cfqd
, root
, sector
, &parent
, NULL
);
1768 * If the exact sector wasn't found, the parent of the NULL leaf
1769 * will contain the closest sector.
1771 __cfqq
= rb_entry(parent
, struct cfq_queue
, p_node
);
1772 if (cfq_rq_close(cfqd
, cur_cfqq
, __cfqq
->next_rq
))
1775 if (blk_rq_pos(__cfqq
->next_rq
) < sector
)
1776 node
= rb_next(&__cfqq
->p_node
);
1778 node
= rb_prev(&__cfqq
->p_node
);
1782 __cfqq
= rb_entry(node
, struct cfq_queue
, p_node
);
1783 if (cfq_rq_close(cfqd
, cur_cfqq
, __cfqq
->next_rq
))
1791 * cur_cfqq - passed in so that we don't decide that the current queue is
1792 * closely cooperating with itself.
1794 * So, basically we're assuming that that cur_cfqq has dispatched at least
1795 * one request, and that cfqd->last_position reflects a position on the disk
1796 * associated with the I/O issued by cur_cfqq. I'm not sure this is a valid
1799 static struct cfq_queue
*cfq_close_cooperator(struct cfq_data
*cfqd
,
1800 struct cfq_queue
*cur_cfqq
)
1802 struct cfq_queue
*cfqq
;
1804 if (cfq_class_idle(cur_cfqq
))
1806 if (!cfq_cfqq_sync(cur_cfqq
))
1808 if (CFQQ_SEEKY(cur_cfqq
))
1812 * Don't search priority tree if it's the only queue in the group.
1814 if (cur_cfqq
->cfqg
->nr_cfqq
== 1)
1818 * We should notice if some of the queues are cooperating, eg
1819 * working closely on the same area of the disk. In that case,
1820 * we can group them together and don't waste time idling.
1822 cfqq
= cfqq_close(cfqd
, cur_cfqq
);
1826 /* If new queue belongs to different cfq_group, don't choose it */
1827 if (cur_cfqq
->cfqg
!= cfqq
->cfqg
)
1831 * It only makes sense to merge sync queues.
1833 if (!cfq_cfqq_sync(cfqq
))
1835 if (CFQQ_SEEKY(cfqq
))
1839 * Do not merge queues of different priority classes
1841 if (cfq_class_rt(cfqq
) != cfq_class_rt(cur_cfqq
))
1848 * Determine whether we should enforce idle window for this queue.
1851 static bool cfq_should_idle(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
1853 enum wl_prio_t prio
= cfqq_prio(cfqq
);
1854 struct cfq_rb_root
*service_tree
= cfqq
->service_tree
;
1856 BUG_ON(!service_tree
);
1857 BUG_ON(!service_tree
->count
);
1859 if (!cfqd
->cfq_slice_idle
)
1862 /* We never do for idle class queues. */
1863 if (prio
== IDLE_WORKLOAD
)
1866 /* We do for queues that were marked with idle window flag. */
1867 if (cfq_cfqq_idle_window(cfqq
) &&
1868 !(blk_queue_nonrot(cfqd
->queue
) && cfqd
->hw_tag
))
1872 * Otherwise, we do only if they are the last ones
1873 * in their service tree.
1875 if (service_tree
->count
== 1 && cfq_cfqq_sync(cfqq
))
1877 cfq_log_cfqq(cfqd
, cfqq
, "Not idling. st->count:%d",
1878 service_tree
->count
);
1882 static void cfq_arm_slice_timer(struct cfq_data
*cfqd
)
1884 struct cfq_queue
*cfqq
= cfqd
->active_queue
;
1885 struct cfq_io_context
*cic
;
1886 unsigned long sl
, group_idle
= 0;
1889 * SSD device without seek penalty, disable idling. But only do so
1890 * for devices that support queuing, otherwise we still have a problem
1891 * with sync vs async workloads.
1893 if (blk_queue_nonrot(cfqd
->queue
) && cfqd
->hw_tag
)
1896 WARN_ON(!RB_EMPTY_ROOT(&cfqq
->sort_list
));
1897 WARN_ON(cfq_cfqq_slice_new(cfqq
));
1900 * idle is disabled, either manually or by past process history
1902 if (!cfq_should_idle(cfqd
, cfqq
)) {
1903 /* no queue idling. Check for group idling */
1904 if (cfqd
->cfq_group_idle
)
1905 group_idle
= cfqd
->cfq_group_idle
;
1911 * still active requests from this queue, don't idle
1913 if (cfqq
->dispatched
)
1917 * task has exited, don't wait
1919 cic
= cfqd
->active_cic
;
1920 if (!cic
|| !atomic_read(&cic
->ioc
->nr_tasks
))
1924 * If our average think time is larger than the remaining time
1925 * slice, then don't idle. This avoids overrunning the allotted
1928 if (sample_valid(cic
->ttime_samples
) &&
1929 (cfqq
->slice_end
- jiffies
< cic
->ttime_mean
)) {
1930 cfq_log_cfqq(cfqd
, cfqq
, "Not idling. think_time:%d",
1935 /* There are other queues in the group, don't do group idle */
1936 if (group_idle
&& cfqq
->cfqg
->nr_cfqq
> 1)
1939 cfq_mark_cfqq_wait_request(cfqq
);
1942 sl
= cfqd
->cfq_group_idle
;
1944 sl
= cfqd
->cfq_slice_idle
;
1946 mod_timer(&cfqd
->idle_slice_timer
, jiffies
+ sl
);
1947 cfq_blkiocg_update_set_idle_time_stats(&cfqq
->cfqg
->blkg
);
1948 cfq_log_cfqq(cfqd
, cfqq
, "arm_idle: %lu group_idle: %d", sl
,
1949 group_idle
? 1 : 0);
1953 * Move request from internal lists to the request queue dispatch list.
1955 static void cfq_dispatch_insert(struct request_queue
*q
, struct request
*rq
)
1957 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
1958 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
1960 cfq_log_cfqq(cfqd
, cfqq
, "dispatch_insert");
1962 cfqq
->next_rq
= cfq_find_next_rq(cfqd
, cfqq
, rq
);
1963 cfq_remove_request(rq
);
1965 (RQ_CFQG(rq
))->dispatched
++;
1966 elv_dispatch_sort(q
, rq
);
1968 cfqd
->rq_in_flight
[cfq_cfqq_sync(cfqq
)]++;
1969 cfqq
->nr_sectors
+= blk_rq_sectors(rq
);
1970 cfq_blkiocg_update_dispatch_stats(&cfqq
->cfqg
->blkg
, blk_rq_bytes(rq
),
1971 rq_data_dir(rq
), rq_is_sync(rq
));
1975 * return expired entry, or NULL to just start from scratch in rbtree
1977 static struct request
*cfq_check_fifo(struct cfq_queue
*cfqq
)
1979 struct request
*rq
= NULL
;
1981 if (cfq_cfqq_fifo_expire(cfqq
))
1984 cfq_mark_cfqq_fifo_expire(cfqq
);
1986 if (list_empty(&cfqq
->fifo
))
1989 rq
= rq_entry_fifo(cfqq
->fifo
.next
);
1990 if (time_before(jiffies
, rq_fifo_time(rq
)))
1993 cfq_log_cfqq(cfqq
->cfqd
, cfqq
, "fifo=%p", rq
);
1998 cfq_prio_to_maxrq(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
2000 const int base_rq
= cfqd
->cfq_slice_async_rq
;
2002 WARN_ON(cfqq
->ioprio
>= IOPRIO_BE_NR
);
2004 return 2 * (base_rq
+ base_rq
* (CFQ_PRIO_LISTS
- 1 - cfqq
->ioprio
));
2008 * Must be called with the queue_lock held.
2010 static int cfqq_process_refs(struct cfq_queue
*cfqq
)
2012 int process_refs
, io_refs
;
2014 io_refs
= cfqq
->allocated
[READ
] + cfqq
->allocated
[WRITE
];
2015 process_refs
= cfqq
->ref
- io_refs
;
2016 BUG_ON(process_refs
< 0);
2017 return process_refs
;
2020 static void cfq_setup_merge(struct cfq_queue
*cfqq
, struct cfq_queue
*new_cfqq
)
2022 int process_refs
, new_process_refs
;
2023 struct cfq_queue
*__cfqq
;
2026 * If there are no process references on the new_cfqq, then it is
2027 * unsafe to follow the ->new_cfqq chain as other cfqq's in the
2028 * chain may have dropped their last reference (not just their
2029 * last process reference).
2031 if (!cfqq_process_refs(new_cfqq
))
2034 /* Avoid a circular list and skip interim queue merges */
2035 while ((__cfqq
= new_cfqq
->new_cfqq
)) {
2041 process_refs
= cfqq_process_refs(cfqq
);
2042 new_process_refs
= cfqq_process_refs(new_cfqq
);
2044 * If the process for the cfqq has gone away, there is no
2045 * sense in merging the queues.
2047 if (process_refs
== 0 || new_process_refs
== 0)
2051 * Merge in the direction of the lesser amount of work.
2053 if (new_process_refs
>= process_refs
) {
2054 cfqq
->new_cfqq
= new_cfqq
;
2055 new_cfqq
->ref
+= process_refs
;
2057 new_cfqq
->new_cfqq
= cfqq
;
2058 cfqq
->ref
+= new_process_refs
;
2062 static enum wl_type_t
cfq_choose_wl(struct cfq_data
*cfqd
,
2063 struct cfq_group
*cfqg
, enum wl_prio_t prio
)
2065 struct cfq_queue
*queue
;
2067 bool key_valid
= false;
2068 unsigned long lowest_key
= 0;
2069 enum wl_type_t cur_best
= SYNC_NOIDLE_WORKLOAD
;
2071 for (i
= 0; i
<= SYNC_WORKLOAD
; ++i
) {
2072 /* select the one with lowest rb_key */
2073 queue
= cfq_rb_first(service_tree_for(cfqg
, prio
, i
));
2075 (!key_valid
|| time_before(queue
->rb_key
, lowest_key
))) {
2076 lowest_key
= queue
->rb_key
;
2085 static void choose_service_tree(struct cfq_data
*cfqd
, struct cfq_group
*cfqg
)
2089 struct cfq_rb_root
*st
;
2090 unsigned group_slice
;
2091 enum wl_prio_t original_prio
= cfqd
->serving_prio
;
2093 /* Choose next priority. RT > BE > IDLE */
2094 if (cfq_group_busy_queues_wl(RT_WORKLOAD
, cfqd
, cfqg
))
2095 cfqd
->serving_prio
= RT_WORKLOAD
;
2096 else if (cfq_group_busy_queues_wl(BE_WORKLOAD
, cfqd
, cfqg
))
2097 cfqd
->serving_prio
= BE_WORKLOAD
;
2099 cfqd
->serving_prio
= IDLE_WORKLOAD
;
2100 cfqd
->workload_expires
= jiffies
+ 1;
2104 if (original_prio
!= cfqd
->serving_prio
)
2108 * For RT and BE, we have to choose also the type
2109 * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
2112 st
= service_tree_for(cfqg
, cfqd
->serving_prio
, cfqd
->serving_type
);
2116 * check workload expiration, and that we still have other queues ready
2118 if (count
&& !time_after(jiffies
, cfqd
->workload_expires
))
2122 /* otherwise select new workload type */
2123 cfqd
->serving_type
=
2124 cfq_choose_wl(cfqd
, cfqg
, cfqd
->serving_prio
);
2125 st
= service_tree_for(cfqg
, cfqd
->serving_prio
, cfqd
->serving_type
);
2129 * the workload slice is computed as a fraction of target latency
2130 * proportional to the number of queues in that workload, over
2131 * all the queues in the same priority class
2133 group_slice
= cfq_group_slice(cfqd
, cfqg
);
2135 slice
= group_slice
* count
/
2136 max_t(unsigned, cfqg
->busy_queues_avg
[cfqd
->serving_prio
],
2137 cfq_group_busy_queues_wl(cfqd
->serving_prio
, cfqd
, cfqg
));
2139 if (cfqd
->serving_type
== ASYNC_WORKLOAD
) {
2143 * Async queues are currently system wide. Just taking
2144 * proportion of queues with-in same group will lead to higher
2145 * async ratio system wide as generally root group is going
2146 * to have higher weight. A more accurate thing would be to
2147 * calculate system wide asnc/sync ratio.
2149 tmp
= cfq_target_latency
* cfqg_busy_async_queues(cfqd
, cfqg
);
2150 tmp
= tmp
/cfqd
->busy_queues
;
2151 slice
= min_t(unsigned, slice
, tmp
);
2153 /* async workload slice is scaled down according to
2154 * the sync/async slice ratio. */
2155 slice
= slice
* cfqd
->cfq_slice
[0] / cfqd
->cfq_slice
[1];
2157 /* sync workload slice is at least 2 * cfq_slice_idle */
2158 slice
= max(slice
, 2 * cfqd
->cfq_slice_idle
);
2160 slice
= max_t(unsigned, slice
, CFQ_MIN_TT
);
2161 cfq_log(cfqd
, "workload slice:%d", slice
);
2162 cfqd
->workload_expires
= jiffies
+ slice
;
2165 static struct cfq_group
*cfq_get_next_cfqg(struct cfq_data
*cfqd
)
2167 struct cfq_rb_root
*st
= &cfqd
->grp_service_tree
;
2168 struct cfq_group
*cfqg
;
2170 if (RB_EMPTY_ROOT(&st
->rb
))
2172 cfqg
= cfq_rb_first_group(st
);
2173 update_min_vdisktime(st
);
2177 static void cfq_choose_cfqg(struct cfq_data
*cfqd
)
2179 struct cfq_group
*cfqg
= cfq_get_next_cfqg(cfqd
);
2181 cfqd
->serving_group
= cfqg
;
2183 /* Restore the workload type data */
2184 if (cfqg
->saved_workload_slice
) {
2185 cfqd
->workload_expires
= jiffies
+ cfqg
->saved_workload_slice
;
2186 cfqd
->serving_type
= cfqg
->saved_workload
;
2187 cfqd
->serving_prio
= cfqg
->saved_serving_prio
;
2189 cfqd
->workload_expires
= jiffies
- 1;
2191 choose_service_tree(cfqd
, cfqg
);
2195 * Select a queue for service. If we have a current active queue,
2196 * check whether to continue servicing it, or retrieve and set a new one.
2198 static struct cfq_queue
*cfq_select_queue(struct cfq_data
*cfqd
)
2200 struct cfq_queue
*cfqq
, *new_cfqq
= NULL
;
2202 cfqq
= cfqd
->active_queue
;
2206 if (!cfqd
->rq_queued
)
2210 * We were waiting for group to get backlogged. Expire the queue
2212 if (cfq_cfqq_wait_busy(cfqq
) && !RB_EMPTY_ROOT(&cfqq
->sort_list
))
2216 * The active queue has run out of time, expire it and select new.
2218 if (cfq_slice_used(cfqq
) && !cfq_cfqq_must_dispatch(cfqq
)) {
2220 * If slice had not expired at the completion of last request
2221 * we might not have turned on wait_busy flag. Don't expire
2222 * the queue yet. Allow the group to get backlogged.
2224 * The very fact that we have used the slice, that means we
2225 * have been idling all along on this queue and it should be
2226 * ok to wait for this request to complete.
2228 if (cfqq
->cfqg
->nr_cfqq
== 1 && RB_EMPTY_ROOT(&cfqq
->sort_list
)
2229 && cfqq
->dispatched
&& cfq_should_idle(cfqd
, cfqq
)) {
2233 goto check_group_idle
;
2237 * The active queue has requests and isn't expired, allow it to
2240 if (!RB_EMPTY_ROOT(&cfqq
->sort_list
))
2244 * If another queue has a request waiting within our mean seek
2245 * distance, let it run. The expire code will check for close
2246 * cooperators and put the close queue at the front of the service
2247 * tree. If possible, merge the expiring queue with the new cfqq.
2249 new_cfqq
= cfq_close_cooperator(cfqd
, cfqq
);
2251 if (!cfqq
->new_cfqq
)
2252 cfq_setup_merge(cfqq
, new_cfqq
);
2257 * No requests pending. If the active queue still has requests in
2258 * flight or is idling for a new request, allow either of these
2259 * conditions to happen (or time out) before selecting a new queue.
2261 if (timer_pending(&cfqd
->idle_slice_timer
)) {
2267 * This is a deep seek queue, but the device is much faster than
2268 * the queue can deliver, don't idle
2270 if (CFQQ_SEEKY(cfqq
) && cfq_cfqq_idle_window(cfqq
) &&
2271 (cfq_cfqq_slice_new(cfqq
) ||
2272 (cfqq
->slice_end
- jiffies
> jiffies
- cfqq
->slice_start
))) {
2273 cfq_clear_cfqq_deep(cfqq
);
2274 cfq_clear_cfqq_idle_window(cfqq
);
2277 if (cfqq
->dispatched
&& cfq_should_idle(cfqd
, cfqq
)) {
2283 * If group idle is enabled and there are requests dispatched from
2284 * this group, wait for requests to complete.
2287 if (cfqd
->cfq_group_idle
&& cfqq
->cfqg
->nr_cfqq
== 1
2288 && cfqq
->cfqg
->dispatched
) {
2294 cfq_slice_expired(cfqd
, 0);
2297 * Current queue expired. Check if we have to switch to a new
2301 cfq_choose_cfqg(cfqd
);
2303 cfqq
= cfq_set_active_queue(cfqd
, new_cfqq
);
2308 static int __cfq_forced_dispatch_cfqq(struct cfq_queue
*cfqq
)
2312 while (cfqq
->next_rq
) {
2313 cfq_dispatch_insert(cfqq
->cfqd
->queue
, cfqq
->next_rq
);
2317 BUG_ON(!list_empty(&cfqq
->fifo
));
2319 /* By default cfqq is not expired if it is empty. Do it explicitly */
2320 __cfq_slice_expired(cfqq
->cfqd
, cfqq
, 0);
2325 * Drain our current requests. Used for barriers and when switching
2326 * io schedulers on-the-fly.
2328 static int cfq_forced_dispatch(struct cfq_data
*cfqd
)
2330 struct cfq_queue
*cfqq
;
2333 /* Expire the timeslice of the current active queue first */
2334 cfq_slice_expired(cfqd
, 0);
2335 while ((cfqq
= cfq_get_next_queue_forced(cfqd
)) != NULL
) {
2336 __cfq_set_active_queue(cfqd
, cfqq
);
2337 dispatched
+= __cfq_forced_dispatch_cfqq(cfqq
);
2340 BUG_ON(cfqd
->busy_queues
);
2342 cfq_log(cfqd
, "forced_dispatch=%d", dispatched
);
2346 static inline bool cfq_slice_used_soon(struct cfq_data
*cfqd
,
2347 struct cfq_queue
*cfqq
)
2349 /* the queue hasn't finished any request, can't estimate */
2350 if (cfq_cfqq_slice_new(cfqq
))
2352 if (time_after(jiffies
+ cfqd
->cfq_slice_idle
* cfqq
->dispatched
,
2359 static bool cfq_may_dispatch(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
2361 unsigned int max_dispatch
;
2364 * Drain async requests before we start sync IO
2366 if (cfq_should_idle(cfqd
, cfqq
) && cfqd
->rq_in_flight
[BLK_RW_ASYNC
])
2370 * If this is an async queue and we have sync IO in flight, let it wait
2372 if (cfqd
->rq_in_flight
[BLK_RW_SYNC
] && !cfq_cfqq_sync(cfqq
))
2375 max_dispatch
= max_t(unsigned int, cfqd
->cfq_quantum
/ 2, 1);
2376 if (cfq_class_idle(cfqq
))
2380 * Does this cfqq already have too much IO in flight?
2382 if (cfqq
->dispatched
>= max_dispatch
) {
2383 bool promote_sync
= false;
2385 * idle queue must always only have a single IO in flight
2387 if (cfq_class_idle(cfqq
))
2391 * If there is only one sync queue, and its think time is
2392 * small, we can ignore async queue here and give the sync
2393 * queue no dispatch limit. The reason is a sync queue can
2394 * preempt async queue, limiting the sync queue doesn't make
2395 * sense. This is useful for aiostress test.
2397 if (cfq_cfqq_sync(cfqq
) && cfqd
->busy_sync_queues
== 1) {
2398 struct cfq_io_context
*cic
= RQ_CIC(cfqq
->next_rq
);
2400 if (sample_valid(cic
->ttime_samples
) &&
2401 cic
->ttime_mean
< cfqd
->cfq_slice_idle
)
2402 promote_sync
= true;
2406 * We have other queues, don't allow more IO from this one
2408 if (cfqd
->busy_queues
> 1 && cfq_slice_used_soon(cfqd
, cfqq
) &&
2413 * Sole queue user, no limit
2415 if (cfqd
->busy_queues
== 1 || promote_sync
)
2419 * Normally we start throttling cfqq when cfq_quantum/2
2420 * requests have been dispatched. But we can drive
2421 * deeper queue depths at the beginning of slice
2422 * subjected to upper limit of cfq_quantum.
2424 max_dispatch
= cfqd
->cfq_quantum
;
2428 * Async queues must wait a bit before being allowed dispatch.
2429 * We also ramp up the dispatch depth gradually for async IO,
2430 * based on the last sync IO we serviced
2432 if (!cfq_cfqq_sync(cfqq
) && cfqd
->cfq_latency
) {
2433 unsigned long last_sync
= jiffies
- cfqd
->last_delayed_sync
;
2436 depth
= last_sync
/ cfqd
->cfq_slice
[1];
2437 if (!depth
&& !cfqq
->dispatched
)
2439 if (depth
< max_dispatch
)
2440 max_dispatch
= depth
;
2444 * If we're below the current max, allow a dispatch
2446 return cfqq
->dispatched
< max_dispatch
;
2450 * Dispatch a request from cfqq, moving them to the request queue
2453 static bool cfq_dispatch_request(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
2457 BUG_ON(RB_EMPTY_ROOT(&cfqq
->sort_list
));
2459 if (!cfq_may_dispatch(cfqd
, cfqq
))
2463 * follow expired path, else get first next available
2465 rq
= cfq_check_fifo(cfqq
);
2470 * insert request into driver dispatch list
2472 cfq_dispatch_insert(cfqd
->queue
, rq
);
2474 if (!cfqd
->active_cic
) {
2475 struct cfq_io_context
*cic
= RQ_CIC(rq
);
2477 atomic_long_inc(&cic
->ioc
->refcount
);
2478 cfqd
->active_cic
= cic
;
2485 * Find the cfqq that we need to service and move a request from that to the
2488 static int cfq_dispatch_requests(struct request_queue
*q
, int force
)
2490 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
2491 struct cfq_queue
*cfqq
;
2493 if (!cfqd
->busy_queues
)
2496 if (unlikely(force
))
2497 return cfq_forced_dispatch(cfqd
);
2499 cfqq
= cfq_select_queue(cfqd
);
2504 * Dispatch a request from this cfqq, if it is allowed
2506 if (!cfq_dispatch_request(cfqd
, cfqq
))
2509 cfqq
->slice_dispatch
++;
2510 cfq_clear_cfqq_must_dispatch(cfqq
);
2513 * expire an async queue immediately if it has used up its slice. idle
2514 * queue always expire after 1 dispatch round.
2516 if (cfqd
->busy_queues
> 1 && ((!cfq_cfqq_sync(cfqq
) &&
2517 cfqq
->slice_dispatch
>= cfq_prio_to_maxrq(cfqd
, cfqq
)) ||
2518 cfq_class_idle(cfqq
))) {
2519 cfqq
->slice_end
= jiffies
+ 1;
2520 cfq_slice_expired(cfqd
, 0);
2523 cfq_log_cfqq(cfqd
, cfqq
, "dispatched a request");
2528 * task holds one reference to the queue, dropped when task exits. each rq
2529 * in-flight on this queue also holds a reference, dropped when rq is freed.
2531 * Each cfq queue took a reference on the parent group. Drop it now.
2532 * queue lock must be held here.
2534 static void cfq_put_queue(struct cfq_queue
*cfqq
)
2536 struct cfq_data
*cfqd
= cfqq
->cfqd
;
2537 struct cfq_group
*cfqg
;
2539 BUG_ON(cfqq
->ref
<= 0);
2545 cfq_log_cfqq(cfqd
, cfqq
, "put_queue");
2546 BUG_ON(rb_first(&cfqq
->sort_list
));
2547 BUG_ON(cfqq
->allocated
[READ
] + cfqq
->allocated
[WRITE
]);
2550 if (unlikely(cfqd
->active_queue
== cfqq
)) {
2551 __cfq_slice_expired(cfqd
, cfqq
, 0);
2552 cfq_schedule_dispatch(cfqd
);
2555 BUG_ON(cfq_cfqq_on_rr(cfqq
));
2556 kmem_cache_free(cfq_pool
, cfqq
);
2561 * Must always be called with the rcu_read_lock() held
2564 __call_for_each_cic(struct io_context
*ioc
,
2565 void (*func
)(struct io_context
*, struct cfq_io_context
*))
2567 struct cfq_io_context
*cic
;
2568 struct hlist_node
*n
;
2570 hlist_for_each_entry_rcu(cic
, n
, &ioc
->cic_list
, cic_list
)
2575 * Call func for each cic attached to this ioc.
2578 call_for_each_cic(struct io_context
*ioc
,
2579 void (*func
)(struct io_context
*, struct cfq_io_context
*))
2582 __call_for_each_cic(ioc
, func
);
2586 static void cfq_cic_free_rcu(struct rcu_head
*head
)
2588 struct cfq_io_context
*cic
;
2590 cic
= container_of(head
, struct cfq_io_context
, rcu_head
);
2592 kmem_cache_free(cfq_ioc_pool
, cic
);
2593 elv_ioc_count_dec(cfq_ioc_count
);
2597 * CFQ scheduler is exiting, grab exit lock and check
2598 * the pending io context count. If it hits zero,
2599 * complete ioc_gone and set it back to NULL
2601 spin_lock(&ioc_gone_lock
);
2602 if (ioc_gone
&& !elv_ioc_count_read(cfq_ioc_count
)) {
2606 spin_unlock(&ioc_gone_lock
);
2610 static void cfq_cic_free(struct cfq_io_context
*cic
)
2612 call_rcu(&cic
->rcu_head
, cfq_cic_free_rcu
);
2615 static void cic_free_func(struct io_context
*ioc
, struct cfq_io_context
*cic
)
2617 unsigned long flags
;
2618 unsigned long dead_key
= (unsigned long) cic
->key
;
2620 BUG_ON(!(dead_key
& CIC_DEAD_KEY
));
2622 spin_lock_irqsave(&ioc
->lock
, flags
);
2623 radix_tree_delete(&ioc
->radix_root
, dead_key
>> CIC_DEAD_INDEX_SHIFT
);
2624 hlist_del_rcu(&cic
->cic_list
);
2625 spin_unlock_irqrestore(&ioc
->lock
, flags
);
2631 * Must be called with rcu_read_lock() held or preemption otherwise disabled.
2632 * Only two callers of this - ->dtor() which is called with the rcu_read_lock(),
2633 * and ->trim() which is called with the task lock held
2635 static void cfq_free_io_context(struct io_context
*ioc
)
2638 * ioc->refcount is zero here, or we are called from elv_unregister(),
2639 * so no more cic's are allowed to be linked into this ioc. So it
2640 * should be ok to iterate over the known list, we will see all cic's
2641 * since no new ones are added.
2643 __call_for_each_cic(ioc
, cic_free_func
);
2646 static void cfq_put_cooperator(struct cfq_queue
*cfqq
)
2648 struct cfq_queue
*__cfqq
, *next
;
2651 * If this queue was scheduled to merge with another queue, be
2652 * sure to drop the reference taken on that queue (and others in
2653 * the merge chain). See cfq_setup_merge and cfq_merge_cfqqs.
2655 __cfqq
= cfqq
->new_cfqq
;
2657 if (__cfqq
== cfqq
) {
2658 WARN(1, "cfqq->new_cfqq loop detected\n");
2661 next
= __cfqq
->new_cfqq
;
2662 cfq_put_queue(__cfqq
);
2667 static void cfq_exit_cfqq(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
2669 if (unlikely(cfqq
== cfqd
->active_queue
)) {
2670 __cfq_slice_expired(cfqd
, cfqq
, 0);
2671 cfq_schedule_dispatch(cfqd
);
2674 cfq_put_cooperator(cfqq
);
2676 cfq_put_queue(cfqq
);
2679 static void __cfq_exit_single_io_context(struct cfq_data
*cfqd
,
2680 struct cfq_io_context
*cic
)
2682 struct io_context
*ioc
= cic
->ioc
;
2684 list_del_init(&cic
->queue_list
);
2687 * Make sure dead mark is seen for dead queues
2690 cic
->key
= cfqd_dead_key(cfqd
);
2692 if (ioc
->ioc_data
== cic
)
2693 rcu_assign_pointer(ioc
->ioc_data
, NULL
);
2695 if (cic
->cfqq
[BLK_RW_ASYNC
]) {
2696 cfq_exit_cfqq(cfqd
, cic
->cfqq
[BLK_RW_ASYNC
]);
2697 cic
->cfqq
[BLK_RW_ASYNC
] = NULL
;
2700 if (cic
->cfqq
[BLK_RW_SYNC
]) {
2701 cfq_exit_cfqq(cfqd
, cic
->cfqq
[BLK_RW_SYNC
]);
2702 cic
->cfqq
[BLK_RW_SYNC
] = NULL
;
2706 static void cfq_exit_single_io_context(struct io_context
*ioc
,
2707 struct cfq_io_context
*cic
)
2709 struct cfq_data
*cfqd
= cic_to_cfqd(cic
);
2712 struct request_queue
*q
= cfqd
->queue
;
2713 unsigned long flags
;
2715 spin_lock_irqsave(q
->queue_lock
, flags
);
2718 * Ensure we get a fresh copy of the ->key to prevent
2719 * race between exiting task and queue
2721 smp_read_barrier_depends();
2722 if (cic
->key
== cfqd
)
2723 __cfq_exit_single_io_context(cfqd
, cic
);
2725 spin_unlock_irqrestore(q
->queue_lock
, flags
);
2730 * The process that ioc belongs to has exited, we need to clean up
2731 * and put the internal structures we have that belongs to that process.
2733 static void cfq_exit_io_context(struct io_context
*ioc
)
2735 call_for_each_cic(ioc
, cfq_exit_single_io_context
);
2738 static struct cfq_io_context
*
2739 cfq_alloc_io_context(struct cfq_data
*cfqd
, gfp_t gfp_mask
)
2741 struct cfq_io_context
*cic
;
2743 cic
= kmem_cache_alloc_node(cfq_ioc_pool
, gfp_mask
| __GFP_ZERO
,
2746 cic
->last_end_request
= jiffies
;
2747 INIT_LIST_HEAD(&cic
->queue_list
);
2748 INIT_HLIST_NODE(&cic
->cic_list
);
2749 cic
->dtor
= cfq_free_io_context
;
2750 cic
->exit
= cfq_exit_io_context
;
2751 elv_ioc_count_inc(cfq_ioc_count
);
2757 static void cfq_init_prio_data(struct cfq_queue
*cfqq
, struct io_context
*ioc
)
2759 struct task_struct
*tsk
= current
;
2762 if (!cfq_cfqq_prio_changed(cfqq
))
2765 ioprio_class
= IOPRIO_PRIO_CLASS(ioc
->ioprio
);
2766 switch (ioprio_class
) {
2768 printk(KERN_ERR
"cfq: bad prio %x\n", ioprio_class
);
2769 case IOPRIO_CLASS_NONE
:
2771 * no prio set, inherit CPU scheduling settings
2773 cfqq
->ioprio
= task_nice_ioprio(tsk
);
2774 cfqq
->ioprio_class
= task_nice_ioclass(tsk
);
2776 case IOPRIO_CLASS_RT
:
2777 cfqq
->ioprio
= task_ioprio(ioc
);
2778 cfqq
->ioprio_class
= IOPRIO_CLASS_RT
;
2780 case IOPRIO_CLASS_BE
:
2781 cfqq
->ioprio
= task_ioprio(ioc
);
2782 cfqq
->ioprio_class
= IOPRIO_CLASS_BE
;
2784 case IOPRIO_CLASS_IDLE
:
2785 cfqq
->ioprio_class
= IOPRIO_CLASS_IDLE
;
2787 cfq_clear_cfqq_idle_window(cfqq
);
2792 * keep track of original prio settings in case we have to temporarily
2793 * elevate the priority of this queue
2795 cfqq
->org_ioprio
= cfqq
->ioprio
;
2796 cfqq
->org_ioprio_class
= cfqq
->ioprio_class
;
2797 cfq_clear_cfqq_prio_changed(cfqq
);
2800 static void changed_ioprio(struct io_context
*ioc
, struct cfq_io_context
*cic
)
2802 struct cfq_data
*cfqd
= cic_to_cfqd(cic
);
2803 struct cfq_queue
*cfqq
;
2804 unsigned long flags
;
2806 if (unlikely(!cfqd
))
2809 spin_lock_irqsave(cfqd
->queue
->queue_lock
, flags
);
2811 cfqq
= cic
->cfqq
[BLK_RW_ASYNC
];
2813 struct cfq_queue
*new_cfqq
;
2814 new_cfqq
= cfq_get_queue(cfqd
, BLK_RW_ASYNC
, cic
->ioc
,
2817 cic
->cfqq
[BLK_RW_ASYNC
] = new_cfqq
;
2818 cfq_put_queue(cfqq
);
2822 cfqq
= cic
->cfqq
[BLK_RW_SYNC
];
2824 cfq_mark_cfqq_prio_changed(cfqq
);
2826 spin_unlock_irqrestore(cfqd
->queue
->queue_lock
, flags
);
2829 static void cfq_ioc_set_ioprio(struct io_context
*ioc
)
2831 call_for_each_cic(ioc
, changed_ioprio
);
2832 ioc
->ioprio_changed
= 0;
2835 static void cfq_init_cfqq(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
2836 pid_t pid
, bool is_sync
)
2838 RB_CLEAR_NODE(&cfqq
->rb_node
);
2839 RB_CLEAR_NODE(&cfqq
->p_node
);
2840 INIT_LIST_HEAD(&cfqq
->fifo
);
2845 cfq_mark_cfqq_prio_changed(cfqq
);
2848 if (!cfq_class_idle(cfqq
))
2849 cfq_mark_cfqq_idle_window(cfqq
);
2850 cfq_mark_cfqq_sync(cfqq
);
2855 #ifdef CONFIG_CFQ_GROUP_IOSCHED
2856 static void changed_cgroup(struct io_context
*ioc
, struct cfq_io_context
*cic
)
2858 struct cfq_queue
*sync_cfqq
= cic_to_cfqq(cic
, 1);
2859 struct cfq_data
*cfqd
= cic_to_cfqd(cic
);
2860 unsigned long flags
;
2861 struct request_queue
*q
;
2863 if (unlikely(!cfqd
))
2868 spin_lock_irqsave(q
->queue_lock
, flags
);
2872 * Drop reference to sync queue. A new sync queue will be
2873 * assigned in new group upon arrival of a fresh request.
2875 cfq_log_cfqq(cfqd
, sync_cfqq
, "changed cgroup");
2876 cic_set_cfqq(cic
, NULL
, 1);
2877 cfq_put_queue(sync_cfqq
);
2880 spin_unlock_irqrestore(q
->queue_lock
, flags
);
2883 static void cfq_ioc_set_cgroup(struct io_context
*ioc
)
2885 call_for_each_cic(ioc
, changed_cgroup
);
2886 ioc
->cgroup_changed
= 0;
2888 #endif /* CONFIG_CFQ_GROUP_IOSCHED */
2890 static struct cfq_queue
*
2891 cfq_find_alloc_queue(struct cfq_data
*cfqd
, bool is_sync
,
2892 struct io_context
*ioc
, gfp_t gfp_mask
)
2894 struct cfq_queue
*cfqq
, *new_cfqq
= NULL
;
2895 struct cfq_io_context
*cic
;
2896 struct cfq_group
*cfqg
;
2899 cfqg
= cfq_get_cfqg(cfqd
, 1);
2900 cic
= cfq_cic_lookup(cfqd
, ioc
);
2901 /* cic always exists here */
2902 cfqq
= cic_to_cfqq(cic
, is_sync
);
2905 * Always try a new alloc if we fell back to the OOM cfqq
2906 * originally, since it should just be a temporary situation.
2908 if (!cfqq
|| cfqq
== &cfqd
->oom_cfqq
) {
2913 } else if (gfp_mask
& __GFP_WAIT
) {
2914 spin_unlock_irq(cfqd
->queue
->queue_lock
);
2915 new_cfqq
= kmem_cache_alloc_node(cfq_pool
,
2916 gfp_mask
| __GFP_ZERO
,
2918 spin_lock_irq(cfqd
->queue
->queue_lock
);
2922 cfqq
= kmem_cache_alloc_node(cfq_pool
,
2923 gfp_mask
| __GFP_ZERO
,
2928 cfq_init_cfqq(cfqd
, cfqq
, current
->pid
, is_sync
);
2929 cfq_init_prio_data(cfqq
, ioc
);
2930 cfq_link_cfqq_cfqg(cfqq
, cfqg
);
2931 cfq_log_cfqq(cfqd
, cfqq
, "alloced");
2933 cfqq
= &cfqd
->oom_cfqq
;
2937 kmem_cache_free(cfq_pool
, new_cfqq
);
2942 static struct cfq_queue
**
2943 cfq_async_queue_prio(struct cfq_data
*cfqd
, int ioprio_class
, int ioprio
)
2945 switch (ioprio_class
) {
2946 case IOPRIO_CLASS_RT
:
2947 return &cfqd
->async_cfqq
[0][ioprio
];
2948 case IOPRIO_CLASS_BE
:
2949 return &cfqd
->async_cfqq
[1][ioprio
];
2950 case IOPRIO_CLASS_IDLE
:
2951 return &cfqd
->async_idle_cfqq
;
2957 static struct cfq_queue
*
2958 cfq_get_queue(struct cfq_data
*cfqd
, bool is_sync
, struct io_context
*ioc
,
2961 const int ioprio
= task_ioprio(ioc
);
2962 const int ioprio_class
= task_ioprio_class(ioc
);
2963 struct cfq_queue
**async_cfqq
= NULL
;
2964 struct cfq_queue
*cfqq
= NULL
;
2967 async_cfqq
= cfq_async_queue_prio(cfqd
, ioprio_class
, ioprio
);
2972 cfqq
= cfq_find_alloc_queue(cfqd
, is_sync
, ioc
, gfp_mask
);
2975 * pin the queue now that it's allocated, scheduler exit will prune it
2977 if (!is_sync
&& !(*async_cfqq
)) {
2987 * We drop cfq io contexts lazily, so we may find a dead one.
2990 cfq_drop_dead_cic(struct cfq_data
*cfqd
, struct io_context
*ioc
,
2991 struct cfq_io_context
*cic
)
2993 unsigned long flags
;
2995 WARN_ON(!list_empty(&cic
->queue_list
));
2996 BUG_ON(cic
->key
!= cfqd_dead_key(cfqd
));
2998 spin_lock_irqsave(&ioc
->lock
, flags
);
3000 BUG_ON(ioc
->ioc_data
== cic
);
3002 radix_tree_delete(&ioc
->radix_root
, cfqd
->cic_index
);
3003 hlist_del_rcu(&cic
->cic_list
);
3004 spin_unlock_irqrestore(&ioc
->lock
, flags
);
3009 static struct cfq_io_context
*
3010 cfq_cic_lookup(struct cfq_data
*cfqd
, struct io_context
*ioc
)
3012 struct cfq_io_context
*cic
;
3013 unsigned long flags
;
3021 * we maintain a last-hit cache, to avoid browsing over the tree
3023 cic
= rcu_dereference(ioc
->ioc_data
);
3024 if (cic
&& cic
->key
== cfqd
) {
3030 cic
= radix_tree_lookup(&ioc
->radix_root
, cfqd
->cic_index
);
3034 if (unlikely(cic
->key
!= cfqd
)) {
3035 cfq_drop_dead_cic(cfqd
, ioc
, cic
);
3040 spin_lock_irqsave(&ioc
->lock
, flags
);
3041 rcu_assign_pointer(ioc
->ioc_data
, cic
);
3042 spin_unlock_irqrestore(&ioc
->lock
, flags
);
3050 * Add cic into ioc, using cfqd as the search key. This enables us to lookup
3051 * the process specific cfq io context when entered from the block layer.
3052 * Also adds the cic to a per-cfqd list, used when this queue is removed.
3054 static int cfq_cic_link(struct cfq_data
*cfqd
, struct io_context
*ioc
,
3055 struct cfq_io_context
*cic
, gfp_t gfp_mask
)
3057 unsigned long flags
;
3060 ret
= radix_tree_preload(gfp_mask
);
3065 spin_lock_irqsave(&ioc
->lock
, flags
);
3066 ret
= radix_tree_insert(&ioc
->radix_root
,
3067 cfqd
->cic_index
, cic
);
3069 hlist_add_head_rcu(&cic
->cic_list
, &ioc
->cic_list
);
3070 spin_unlock_irqrestore(&ioc
->lock
, flags
);
3072 radix_tree_preload_end();
3075 spin_lock_irqsave(cfqd
->queue
->queue_lock
, flags
);
3076 list_add(&cic
->queue_list
, &cfqd
->cic_list
);
3077 spin_unlock_irqrestore(cfqd
->queue
->queue_lock
, flags
);
3082 printk(KERN_ERR
"cfq: cic link failed!\n");
3088 * Setup general io context and cfq io context. There can be several cfq
3089 * io contexts per general io context, if this process is doing io to more
3090 * than one device managed by cfq.
3092 static struct cfq_io_context
*
3093 cfq_get_io_context(struct cfq_data
*cfqd
, gfp_t gfp_mask
)
3095 struct io_context
*ioc
= NULL
;
3096 struct cfq_io_context
*cic
;
3098 might_sleep_if(gfp_mask
& __GFP_WAIT
);
3100 ioc
= get_io_context(gfp_mask
, cfqd
->queue
->node
);
3104 cic
= cfq_cic_lookup(cfqd
, ioc
);
3108 cic
= cfq_alloc_io_context(cfqd
, gfp_mask
);
3112 if (cfq_cic_link(cfqd
, ioc
, cic
, gfp_mask
))
3116 smp_read_barrier_depends();
3117 if (unlikely(ioc
->ioprio_changed
))
3118 cfq_ioc_set_ioprio(ioc
);
3120 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3121 if (unlikely(ioc
->cgroup_changed
))
3122 cfq_ioc_set_cgroup(ioc
);
3128 put_io_context(ioc
);
3133 cfq_update_io_thinktime(struct cfq_data
*cfqd
, struct cfq_io_context
*cic
)
3135 unsigned long elapsed
= jiffies
- cic
->last_end_request
;
3136 unsigned long ttime
= min(elapsed
, 2UL * cfqd
->cfq_slice_idle
);
3138 cic
->ttime_samples
= (7*cic
->ttime_samples
+ 256) / 8;
3139 cic
->ttime_total
= (7*cic
->ttime_total
+ 256*ttime
) / 8;
3140 cic
->ttime_mean
= (cic
->ttime_total
+ 128) / cic
->ttime_samples
;
3144 cfq_update_io_seektime(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
3148 sector_t n_sec
= blk_rq_sectors(rq
);
3149 if (cfqq
->last_request_pos
) {
3150 if (cfqq
->last_request_pos
< blk_rq_pos(rq
))
3151 sdist
= blk_rq_pos(rq
) - cfqq
->last_request_pos
;
3153 sdist
= cfqq
->last_request_pos
- blk_rq_pos(rq
);
3156 cfqq
->seek_history
<<= 1;
3157 if (blk_queue_nonrot(cfqd
->queue
))
3158 cfqq
->seek_history
|= (n_sec
< CFQQ_SECT_THR_NONROT
);
3160 cfqq
->seek_history
|= (sdist
> CFQQ_SEEK_THR
);
3164 * Disable idle window if the process thinks too long or seeks so much that
3168 cfq_update_idle_window(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
3169 struct cfq_io_context
*cic
)
3171 int old_idle
, enable_idle
;
3174 * Don't idle for async or idle io prio class
3176 if (!cfq_cfqq_sync(cfqq
) || cfq_class_idle(cfqq
))
3179 enable_idle
= old_idle
= cfq_cfqq_idle_window(cfqq
);
3181 if (cfqq
->queued
[0] + cfqq
->queued
[1] >= 4)
3182 cfq_mark_cfqq_deep(cfqq
);
3184 if (cfqq
->next_rq
&& (cfqq
->next_rq
->cmd_flags
& REQ_NOIDLE
))
3186 else if (!atomic_read(&cic
->ioc
->nr_tasks
) || !cfqd
->cfq_slice_idle
||
3187 (!cfq_cfqq_deep(cfqq
) && CFQQ_SEEKY(cfqq
)))
3189 else if (sample_valid(cic
->ttime_samples
)) {
3190 if (cic
->ttime_mean
> cfqd
->cfq_slice_idle
)
3196 if (old_idle
!= enable_idle
) {
3197 cfq_log_cfqq(cfqd
, cfqq
, "idle=%d", enable_idle
);
3199 cfq_mark_cfqq_idle_window(cfqq
);
3201 cfq_clear_cfqq_idle_window(cfqq
);
3206 * Check if new_cfqq should preempt the currently active queue. Return 0 for
3207 * no or if we aren't sure, a 1 will cause a preempt.
3210 cfq_should_preempt(struct cfq_data
*cfqd
, struct cfq_queue
*new_cfqq
,
3213 struct cfq_queue
*cfqq
;
3215 cfqq
= cfqd
->active_queue
;
3219 if (cfq_class_idle(new_cfqq
))
3222 if (cfq_class_idle(cfqq
))
3226 * Don't allow a non-RT request to preempt an ongoing RT cfqq timeslice.
3228 if (cfq_class_rt(cfqq
) && !cfq_class_rt(new_cfqq
))
3232 * if the new request is sync, but the currently running queue is
3233 * not, let the sync request have priority.
3235 if (rq_is_sync(rq
) && !cfq_cfqq_sync(cfqq
))
3238 if (new_cfqq
->cfqg
!= cfqq
->cfqg
)
3241 if (cfq_slice_used(cfqq
))
3244 /* Allow preemption only if we are idling on sync-noidle tree */
3245 if (cfqd
->serving_type
== SYNC_NOIDLE_WORKLOAD
&&
3246 cfqq_type(new_cfqq
) == SYNC_NOIDLE_WORKLOAD
&&
3247 new_cfqq
->service_tree
->count
== 2 &&
3248 RB_EMPTY_ROOT(&cfqq
->sort_list
))
3252 * So both queues are sync. Let the new request get disk time if
3253 * it's a metadata request and the current queue is doing regular IO.
3255 if ((rq
->cmd_flags
& REQ_META
) && !cfqq
->meta_pending
)
3259 * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
3261 if (cfq_class_rt(new_cfqq
) && !cfq_class_rt(cfqq
))
3264 /* An idle queue should not be idle now for some reason */
3265 if (RB_EMPTY_ROOT(&cfqq
->sort_list
) && !cfq_should_idle(cfqd
, cfqq
))
3268 if (!cfqd
->active_cic
|| !cfq_cfqq_wait_request(cfqq
))
3272 * if this request is as-good as one we would expect from the
3273 * current cfqq, let it preempt
3275 if (cfq_rq_close(cfqd
, cfqq
, rq
))
3282 * cfqq preempts the active queue. if we allowed preempt with no slice left,
3283 * let it have half of its nominal slice.
3285 static void cfq_preempt_queue(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
3287 struct cfq_queue
*old_cfqq
= cfqd
->active_queue
;
3289 cfq_log_cfqq(cfqd
, cfqq
, "preempt");
3290 cfq_slice_expired(cfqd
, 1);
3293 * workload type is changed, don't save slice, otherwise preempt
3296 if (cfqq_type(old_cfqq
) != cfqq_type(cfqq
))
3297 cfqq
->cfqg
->saved_workload_slice
= 0;
3300 * Put the new queue at the front of the of the current list,
3301 * so we know that it will be selected next.
3303 BUG_ON(!cfq_cfqq_on_rr(cfqq
));
3305 cfq_service_tree_add(cfqd
, cfqq
, 1);
3307 cfqq
->slice_end
= 0;
3308 cfq_mark_cfqq_slice_new(cfqq
);
3312 * Called when a new fs request (rq) is added (to cfqq). Check if there's
3313 * something we should do about it
3316 cfq_rq_enqueued(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
3319 struct cfq_io_context
*cic
= RQ_CIC(rq
);
3322 if (rq
->cmd_flags
& REQ_META
)
3323 cfqq
->meta_pending
++;
3325 cfq_update_io_thinktime(cfqd
, cic
);
3326 cfq_update_io_seektime(cfqd
, cfqq
, rq
);
3327 cfq_update_idle_window(cfqd
, cfqq
, cic
);
3329 cfqq
->last_request_pos
= blk_rq_pos(rq
) + blk_rq_sectors(rq
);
3331 if (cfqq
== cfqd
->active_queue
) {
3333 * Remember that we saw a request from this process, but
3334 * don't start queuing just yet. Otherwise we risk seeing lots
3335 * of tiny requests, because we disrupt the normal plugging
3336 * and merging. If the request is already larger than a single
3337 * page, let it rip immediately. For that case we assume that
3338 * merging is already done. Ditto for a busy system that
3339 * has other work pending, don't risk delaying until the
3340 * idle timer unplug to continue working.
3342 if (cfq_cfqq_wait_request(cfqq
)) {
3343 if (blk_rq_bytes(rq
) > PAGE_CACHE_SIZE
||
3344 cfqd
->busy_queues
> 1) {
3345 cfq_del_timer(cfqd
, cfqq
);
3346 cfq_clear_cfqq_wait_request(cfqq
);
3347 __blk_run_queue(cfqd
->queue
);
3349 cfq_blkiocg_update_idle_time_stats(
3351 cfq_mark_cfqq_must_dispatch(cfqq
);
3354 } else if (cfq_should_preempt(cfqd
, cfqq
, rq
)) {
3356 * not the active queue - expire current slice if it is
3357 * idle and has expired it's mean thinktime or this new queue
3358 * has some old slice time left and is of higher priority or
3359 * this new queue is RT and the current one is BE
3361 cfq_preempt_queue(cfqd
, cfqq
);
3362 __blk_run_queue(cfqd
->queue
);
3366 static void cfq_insert_request(struct request_queue
*q
, struct request
*rq
)
3368 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
3369 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
3371 cfq_log_cfqq(cfqd
, cfqq
, "insert_request");
3372 cfq_init_prio_data(cfqq
, RQ_CIC(rq
)->ioc
);
3374 rq_set_fifo_time(rq
, jiffies
+ cfqd
->cfq_fifo_expire
[rq_is_sync(rq
)]);
3375 list_add_tail(&rq
->queuelist
, &cfqq
->fifo
);
3377 cfq_blkiocg_update_io_add_stats(&(RQ_CFQG(rq
))->blkg
,
3378 &cfqd
->serving_group
->blkg
, rq_data_dir(rq
),
3380 cfq_rq_enqueued(cfqd
, cfqq
, rq
);
3384 * Update hw_tag based on peak queue depth over 50 samples under
3387 static void cfq_update_hw_tag(struct cfq_data
*cfqd
)
3389 struct cfq_queue
*cfqq
= cfqd
->active_queue
;
3391 if (cfqd
->rq_in_driver
> cfqd
->hw_tag_est_depth
)
3392 cfqd
->hw_tag_est_depth
= cfqd
->rq_in_driver
;
3394 if (cfqd
->hw_tag
== 1)
3397 if (cfqd
->rq_queued
<= CFQ_HW_QUEUE_MIN
&&
3398 cfqd
->rq_in_driver
<= CFQ_HW_QUEUE_MIN
)
3402 * If active queue hasn't enough requests and can idle, cfq might not
3403 * dispatch sufficient requests to hardware. Don't zero hw_tag in this
3406 if (cfqq
&& cfq_cfqq_idle_window(cfqq
) &&
3407 cfqq
->dispatched
+ cfqq
->queued
[0] + cfqq
->queued
[1] <
3408 CFQ_HW_QUEUE_MIN
&& cfqd
->rq_in_driver
< CFQ_HW_QUEUE_MIN
)
3411 if (cfqd
->hw_tag_samples
++ < 50)
3414 if (cfqd
->hw_tag_est_depth
>= CFQ_HW_QUEUE_MIN
)
3420 static bool cfq_should_wait_busy(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
3422 struct cfq_io_context
*cic
= cfqd
->active_cic
;
3424 /* If the queue already has requests, don't wait */
3425 if (!RB_EMPTY_ROOT(&cfqq
->sort_list
))
3428 /* If there are other queues in the group, don't wait */
3429 if (cfqq
->cfqg
->nr_cfqq
> 1)
3432 if (cfq_slice_used(cfqq
))
3435 /* if slice left is less than think time, wait busy */
3436 if (cic
&& sample_valid(cic
->ttime_samples
)
3437 && (cfqq
->slice_end
- jiffies
< cic
->ttime_mean
))
3441 * If think times is less than a jiffy than ttime_mean=0 and above
3442 * will not be true. It might happen that slice has not expired yet
3443 * but will expire soon (4-5 ns) during select_queue(). To cover the
3444 * case where think time is less than a jiffy, mark the queue wait
3445 * busy if only 1 jiffy is left in the slice.
3447 if (cfqq
->slice_end
- jiffies
== 1)
3453 static void cfq_completed_request(struct request_queue
*q
, struct request
*rq
)
3455 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
3456 struct cfq_data
*cfqd
= cfqq
->cfqd
;
3457 const int sync
= rq_is_sync(rq
);
3461 cfq_log_cfqq(cfqd
, cfqq
, "complete rqnoidle %d",
3462 !!(rq
->cmd_flags
& REQ_NOIDLE
));
3464 cfq_update_hw_tag(cfqd
);
3466 WARN_ON(!cfqd
->rq_in_driver
);
3467 WARN_ON(!cfqq
->dispatched
);
3468 cfqd
->rq_in_driver
--;
3470 (RQ_CFQG(rq
))->dispatched
--;
3471 cfq_blkiocg_update_completion_stats(&cfqq
->cfqg
->blkg
,
3472 rq_start_time_ns(rq
), rq_io_start_time_ns(rq
),
3473 rq_data_dir(rq
), rq_is_sync(rq
));
3475 cfqd
->rq_in_flight
[cfq_cfqq_sync(cfqq
)]--;
3478 RQ_CIC(rq
)->last_end_request
= now
;
3479 if (!time_after(rq
->start_time
+ cfqd
->cfq_fifo_expire
[1], now
))
3480 cfqd
->last_delayed_sync
= now
;
3484 * If this is the active queue, check if it needs to be expired,
3485 * or if we want to idle in case it has no pending requests.
3487 if (cfqd
->active_queue
== cfqq
) {
3488 const bool cfqq_empty
= RB_EMPTY_ROOT(&cfqq
->sort_list
);
3490 if (cfq_cfqq_slice_new(cfqq
)) {
3491 cfq_set_prio_slice(cfqd
, cfqq
);
3492 cfq_clear_cfqq_slice_new(cfqq
);
3496 * Should we wait for next request to come in before we expire
3499 if (cfq_should_wait_busy(cfqd
, cfqq
)) {
3500 unsigned long extend_sl
= cfqd
->cfq_slice_idle
;
3501 if (!cfqd
->cfq_slice_idle
)
3502 extend_sl
= cfqd
->cfq_group_idle
;
3503 cfqq
->slice_end
= jiffies
+ extend_sl
;
3504 cfq_mark_cfqq_wait_busy(cfqq
);
3505 cfq_log_cfqq(cfqd
, cfqq
, "will busy wait");
3509 * Idling is not enabled on:
3511 * - idle-priority queues
3513 * - queues with still some requests queued
3514 * - when there is a close cooperator
3516 if (cfq_slice_used(cfqq
) || cfq_class_idle(cfqq
))
3517 cfq_slice_expired(cfqd
, 1);
3518 else if (sync
&& cfqq_empty
&&
3519 !cfq_close_cooperator(cfqd
, cfqq
)) {
3520 cfq_arm_slice_timer(cfqd
);
3524 if (!cfqd
->rq_in_driver
)
3525 cfq_schedule_dispatch(cfqd
);
3529 * we temporarily boost lower priority queues if they are holding fs exclusive
3530 * resources. they are boosted to normal prio (CLASS_BE/4)
3532 static void cfq_prio_boost(struct cfq_queue
*cfqq
)
3534 if (has_fs_excl()) {
3536 * boost idle prio on transactions that would lock out other
3537 * users of the filesystem
3539 if (cfq_class_idle(cfqq
))
3540 cfqq
->ioprio_class
= IOPRIO_CLASS_BE
;
3541 if (cfqq
->ioprio
> IOPRIO_NORM
)
3542 cfqq
->ioprio
= IOPRIO_NORM
;
3545 * unboost the queue (if needed)
3547 cfqq
->ioprio_class
= cfqq
->org_ioprio_class
;
3548 cfqq
->ioprio
= cfqq
->org_ioprio
;
3552 static inline int __cfq_may_queue(struct cfq_queue
*cfqq
)
3554 if (cfq_cfqq_wait_request(cfqq
) && !cfq_cfqq_must_alloc_slice(cfqq
)) {
3555 cfq_mark_cfqq_must_alloc_slice(cfqq
);
3556 return ELV_MQUEUE_MUST
;
3559 return ELV_MQUEUE_MAY
;
3562 static int cfq_may_queue(struct request_queue
*q
, int rw
)
3564 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
3565 struct task_struct
*tsk
= current
;
3566 struct cfq_io_context
*cic
;
3567 struct cfq_queue
*cfqq
;
3570 * don't force setup of a queue from here, as a call to may_queue
3571 * does not necessarily imply that a request actually will be queued.
3572 * so just lookup a possibly existing queue, or return 'may queue'
3575 cic
= cfq_cic_lookup(cfqd
, tsk
->io_context
);
3577 return ELV_MQUEUE_MAY
;
3579 cfqq
= cic_to_cfqq(cic
, rw_is_sync(rw
));
3581 cfq_init_prio_data(cfqq
, cic
->ioc
);
3582 cfq_prio_boost(cfqq
);
3584 return __cfq_may_queue(cfqq
);
3587 return ELV_MQUEUE_MAY
;
3591 * queue lock held here
3593 static void cfq_put_request(struct request
*rq
)
3595 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
3598 const int rw
= rq_data_dir(rq
);
3600 BUG_ON(!cfqq
->allocated
[rw
]);
3601 cfqq
->allocated
[rw
]--;
3603 put_io_context(RQ_CIC(rq
)->ioc
);
3605 rq
->elevator_private
[0] = NULL
;
3606 rq
->elevator_private
[1] = NULL
;
3608 /* Put down rq reference on cfqg */
3609 cfq_put_cfqg(RQ_CFQG(rq
));
3610 rq
->elevator_private
[2] = NULL
;
3612 cfq_put_queue(cfqq
);
3616 static struct cfq_queue
*
3617 cfq_merge_cfqqs(struct cfq_data
*cfqd
, struct cfq_io_context
*cic
,
3618 struct cfq_queue
*cfqq
)
3620 cfq_log_cfqq(cfqd
, cfqq
, "merging with queue %p", cfqq
->new_cfqq
);
3621 cic_set_cfqq(cic
, cfqq
->new_cfqq
, 1);
3622 cfq_mark_cfqq_coop(cfqq
->new_cfqq
);
3623 cfq_put_queue(cfqq
);
3624 return cic_to_cfqq(cic
, 1);
3628 * Returns NULL if a new cfqq should be allocated, or the old cfqq if this
3629 * was the last process referring to said cfqq.
3631 static struct cfq_queue
*
3632 split_cfqq(struct cfq_io_context
*cic
, struct cfq_queue
*cfqq
)
3634 if (cfqq_process_refs(cfqq
) == 1) {
3635 cfqq
->pid
= current
->pid
;
3636 cfq_clear_cfqq_coop(cfqq
);
3637 cfq_clear_cfqq_split_coop(cfqq
);
3641 cic_set_cfqq(cic
, NULL
, 1);
3643 cfq_put_cooperator(cfqq
);
3645 cfq_put_queue(cfqq
);
3649 * Allocate cfq data structures associated with this request.
3652 cfq_set_request(struct request_queue
*q
, struct request
*rq
, gfp_t gfp_mask
)
3654 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
3655 struct cfq_io_context
*cic
;
3656 const int rw
= rq_data_dir(rq
);
3657 const bool is_sync
= rq_is_sync(rq
);
3658 struct cfq_queue
*cfqq
;
3659 unsigned long flags
;
3661 might_sleep_if(gfp_mask
& __GFP_WAIT
);
3663 cic
= cfq_get_io_context(cfqd
, gfp_mask
);
3665 spin_lock_irqsave(q
->queue_lock
, flags
);
3671 cfqq
= cic_to_cfqq(cic
, is_sync
);
3672 if (!cfqq
|| cfqq
== &cfqd
->oom_cfqq
) {
3673 cfqq
= cfq_get_queue(cfqd
, is_sync
, cic
->ioc
, gfp_mask
);
3674 cic_set_cfqq(cic
, cfqq
, is_sync
);
3677 * If the queue was seeky for too long, break it apart.
3679 if (cfq_cfqq_coop(cfqq
) && cfq_cfqq_split_coop(cfqq
)) {
3680 cfq_log_cfqq(cfqd
, cfqq
, "breaking apart cfqq");
3681 cfqq
= split_cfqq(cic
, cfqq
);
3687 * Check to see if this queue is scheduled to merge with
3688 * another, closely cooperating queue. The merging of
3689 * queues happens here as it must be done in process context.
3690 * The reference on new_cfqq was taken in merge_cfqqs.
3693 cfqq
= cfq_merge_cfqqs(cfqd
, cic
, cfqq
);
3696 cfqq
->allocated
[rw
]++;
3699 rq
->elevator_private
[0] = cic
;
3700 rq
->elevator_private
[1] = cfqq
;
3701 rq
->elevator_private
[2] = cfq_ref_get_cfqg(cfqq
->cfqg
);
3702 spin_unlock_irqrestore(q
->queue_lock
, flags
);
3707 put_io_context(cic
->ioc
);
3709 cfq_schedule_dispatch(cfqd
);
3710 spin_unlock_irqrestore(q
->queue_lock
, flags
);
3711 cfq_log(cfqd
, "set_request fail");
3715 static void cfq_kick_queue(struct work_struct
*work
)
3717 struct cfq_data
*cfqd
=
3718 container_of(work
, struct cfq_data
, unplug_work
);
3719 struct request_queue
*q
= cfqd
->queue
;
3721 spin_lock_irq(q
->queue_lock
);
3722 __blk_run_queue(cfqd
->queue
);
3723 spin_unlock_irq(q
->queue_lock
);
3727 * Timer running if the active_queue is currently idling inside its time slice
3729 static void cfq_idle_slice_timer(unsigned long data
)
3731 struct cfq_data
*cfqd
= (struct cfq_data
*) data
;
3732 struct cfq_queue
*cfqq
;
3733 unsigned long flags
;
3736 cfq_log(cfqd
, "idle timer fired");
3738 spin_lock_irqsave(cfqd
->queue
->queue_lock
, flags
);
3740 cfqq
= cfqd
->active_queue
;
3745 * We saw a request before the queue expired, let it through
3747 if (cfq_cfqq_must_dispatch(cfqq
))
3753 if (cfq_slice_used(cfqq
))
3757 * only expire and reinvoke request handler, if there are
3758 * other queues with pending requests
3760 if (!cfqd
->busy_queues
)
3764 * not expired and it has a request pending, let it dispatch
3766 if (!RB_EMPTY_ROOT(&cfqq
->sort_list
))
3770 * Queue depth flag is reset only when the idle didn't succeed
3772 cfq_clear_cfqq_deep(cfqq
);
3775 cfq_slice_expired(cfqd
, timed_out
);
3777 cfq_schedule_dispatch(cfqd
);
3779 spin_unlock_irqrestore(cfqd
->queue
->queue_lock
, flags
);
3782 static void cfq_shutdown_timer_wq(struct cfq_data
*cfqd
)
3784 del_timer_sync(&cfqd
->idle_slice_timer
);
3785 cancel_work_sync(&cfqd
->unplug_work
);
3788 static void cfq_put_async_queues(struct cfq_data
*cfqd
)
3792 for (i
= 0; i
< IOPRIO_BE_NR
; i
++) {
3793 if (cfqd
->async_cfqq
[0][i
])
3794 cfq_put_queue(cfqd
->async_cfqq
[0][i
]);
3795 if (cfqd
->async_cfqq
[1][i
])
3796 cfq_put_queue(cfqd
->async_cfqq
[1][i
]);
3799 if (cfqd
->async_idle_cfqq
)
3800 cfq_put_queue(cfqd
->async_idle_cfqq
);
3803 static void cfq_cfqd_free(struct rcu_head
*head
)
3805 kfree(container_of(head
, struct cfq_data
, rcu
));
3808 static void cfq_exit_queue(struct elevator_queue
*e
)
3810 struct cfq_data
*cfqd
= e
->elevator_data
;
3811 struct request_queue
*q
= cfqd
->queue
;
3813 cfq_shutdown_timer_wq(cfqd
);
3815 spin_lock_irq(q
->queue_lock
);
3817 if (cfqd
->active_queue
)
3818 __cfq_slice_expired(cfqd
, cfqd
->active_queue
, 0);
3820 while (!list_empty(&cfqd
->cic_list
)) {
3821 struct cfq_io_context
*cic
= list_entry(cfqd
->cic_list
.next
,
3822 struct cfq_io_context
,
3825 __cfq_exit_single_io_context(cfqd
, cic
);
3828 cfq_put_async_queues(cfqd
);
3829 cfq_release_cfq_groups(cfqd
);
3830 cfq_blkiocg_del_blkio_group(&cfqd
->root_group
.blkg
);
3832 spin_unlock_irq(q
->queue_lock
);
3834 cfq_shutdown_timer_wq(cfqd
);
3836 spin_lock(&cic_index_lock
);
3837 ida_remove(&cic_index_ida
, cfqd
->cic_index
);
3838 spin_unlock(&cic_index_lock
);
3840 /* Wait for cfqg->blkg->key accessors to exit their grace periods. */
3841 call_rcu(&cfqd
->rcu
, cfq_cfqd_free
);
3844 static int cfq_alloc_cic_index(void)
3849 if (!ida_pre_get(&cic_index_ida
, GFP_KERNEL
))
3852 spin_lock(&cic_index_lock
);
3853 error
= ida_get_new(&cic_index_ida
, &index
);
3854 spin_unlock(&cic_index_lock
);
3855 if (error
&& error
!= -EAGAIN
)
3862 static void *cfq_init_queue(struct request_queue
*q
)
3864 struct cfq_data
*cfqd
;
3866 struct cfq_group
*cfqg
;
3867 struct cfq_rb_root
*st
;
3869 i
= cfq_alloc_cic_index();
3873 cfqd
= kmalloc_node(sizeof(*cfqd
), GFP_KERNEL
| __GFP_ZERO
, q
->node
);
3878 * Don't need take queue_lock in the routine, since we are
3879 * initializing the ioscheduler, and nobody is using cfqd
3881 cfqd
->cic_index
= i
;
3883 /* Init root service tree */
3884 cfqd
->grp_service_tree
= CFQ_RB_ROOT
;
3886 /* Init root group */
3887 cfqg
= &cfqd
->root_group
;
3888 for_each_cfqg_st(cfqg
, i
, j
, st
)
3890 RB_CLEAR_NODE(&cfqg
->rb_node
);
3892 /* Give preference to root group over other groups */
3893 cfqg
->weight
= 2*BLKIO_WEIGHT_DEFAULT
;
3895 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3897 * Take a reference to root group which we never drop. This is just
3898 * to make sure that cfq_put_cfqg() does not try to kfree root group
3902 cfq_blkiocg_add_blkio_group(&blkio_root_cgroup
, &cfqg
->blkg
,
3907 * Not strictly needed (since RB_ROOT just clears the node and we
3908 * zeroed cfqd on alloc), but better be safe in case someone decides
3909 * to add magic to the rb code
3911 for (i
= 0; i
< CFQ_PRIO_LISTS
; i
++)
3912 cfqd
->prio_trees
[i
] = RB_ROOT
;
3915 * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
3916 * Grab a permanent reference to it, so that the normal code flow
3917 * will not attempt to free it.
3919 cfq_init_cfqq(cfqd
, &cfqd
->oom_cfqq
, 1, 0);
3920 cfqd
->oom_cfqq
.ref
++;
3921 cfq_link_cfqq_cfqg(&cfqd
->oom_cfqq
, &cfqd
->root_group
);
3923 INIT_LIST_HEAD(&cfqd
->cic_list
);
3927 init_timer(&cfqd
->idle_slice_timer
);
3928 cfqd
->idle_slice_timer
.function
= cfq_idle_slice_timer
;
3929 cfqd
->idle_slice_timer
.data
= (unsigned long) cfqd
;
3931 INIT_WORK(&cfqd
->unplug_work
, cfq_kick_queue
);
3933 cfqd
->cfq_quantum
= cfq_quantum
;
3934 cfqd
->cfq_fifo_expire
[0] = cfq_fifo_expire
[0];
3935 cfqd
->cfq_fifo_expire
[1] = cfq_fifo_expire
[1];
3936 cfqd
->cfq_back_max
= cfq_back_max
;
3937 cfqd
->cfq_back_penalty
= cfq_back_penalty
;
3938 cfqd
->cfq_slice
[0] = cfq_slice_async
;
3939 cfqd
->cfq_slice
[1] = cfq_slice_sync
;
3940 cfqd
->cfq_slice_async_rq
= cfq_slice_async_rq
;
3941 cfqd
->cfq_slice_idle
= cfq_slice_idle
;
3942 cfqd
->cfq_group_idle
= cfq_group_idle
;
3943 cfqd
->cfq_latency
= 1;
3946 * we optimistically start assuming sync ops weren't delayed in last
3947 * second, in order to have larger depth for async operations.
3949 cfqd
->last_delayed_sync
= jiffies
- HZ
;
3953 static void cfq_slab_kill(void)
3956 * Caller already ensured that pending RCU callbacks are completed,
3957 * so we should have no busy allocations at this point.
3960 kmem_cache_destroy(cfq_pool
);
3962 kmem_cache_destroy(cfq_ioc_pool
);
3965 static int __init
cfq_slab_setup(void)
3967 cfq_pool
= KMEM_CACHE(cfq_queue
, 0);
3971 cfq_ioc_pool
= KMEM_CACHE(cfq_io_context
, 0);
3982 * sysfs parts below -->
3985 cfq_var_show(unsigned int var
, char *page
)
3987 return sprintf(page
, "%d\n", var
);
3991 cfq_var_store(unsigned int *var
, const char *page
, size_t count
)
3993 char *p
= (char *) page
;
3995 *var
= simple_strtoul(p
, &p
, 10);
3999 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
4000 static ssize_t __FUNC(struct elevator_queue *e, char *page) \
4002 struct cfq_data *cfqd = e->elevator_data; \
4003 unsigned int __data = __VAR; \
4005 __data = jiffies_to_msecs(__data); \
4006 return cfq_var_show(__data, (page)); \
4008 SHOW_FUNCTION(cfq_quantum_show
, cfqd
->cfq_quantum
, 0);
4009 SHOW_FUNCTION(cfq_fifo_expire_sync_show
, cfqd
->cfq_fifo_expire
[1], 1);
4010 SHOW_FUNCTION(cfq_fifo_expire_async_show
, cfqd
->cfq_fifo_expire
[0], 1);
4011 SHOW_FUNCTION(cfq_back_seek_max_show
, cfqd
->cfq_back_max
, 0);
4012 SHOW_FUNCTION(cfq_back_seek_penalty_show
, cfqd
->cfq_back_penalty
, 0);
4013 SHOW_FUNCTION(cfq_slice_idle_show
, cfqd
->cfq_slice_idle
, 1);
4014 SHOW_FUNCTION(cfq_group_idle_show
, cfqd
->cfq_group_idle
, 1);
4015 SHOW_FUNCTION(cfq_slice_sync_show
, cfqd
->cfq_slice
[1], 1);
4016 SHOW_FUNCTION(cfq_slice_async_show
, cfqd
->cfq_slice
[0], 1);
4017 SHOW_FUNCTION(cfq_slice_async_rq_show
, cfqd
->cfq_slice_async_rq
, 0);
4018 SHOW_FUNCTION(cfq_low_latency_show
, cfqd
->cfq_latency
, 0);
4019 #undef SHOW_FUNCTION
4021 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
4022 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
4024 struct cfq_data *cfqd = e->elevator_data; \
4025 unsigned int __data; \
4026 int ret = cfq_var_store(&__data, (page), count); \
4027 if (__data < (MIN)) \
4029 else if (__data > (MAX)) \
4032 *(__PTR) = msecs_to_jiffies(__data); \
4034 *(__PTR) = __data; \
4037 STORE_FUNCTION(cfq_quantum_store
, &cfqd
->cfq_quantum
, 1, UINT_MAX
, 0);
4038 STORE_FUNCTION(cfq_fifo_expire_sync_store
, &cfqd
->cfq_fifo_expire
[1], 1,
4040 STORE_FUNCTION(cfq_fifo_expire_async_store
, &cfqd
->cfq_fifo_expire
[0], 1,
4042 STORE_FUNCTION(cfq_back_seek_max_store
, &cfqd
->cfq_back_max
, 0, UINT_MAX
, 0);
4043 STORE_FUNCTION(cfq_back_seek_penalty_store
, &cfqd
->cfq_back_penalty
, 1,
4045 STORE_FUNCTION(cfq_slice_idle_store
, &cfqd
->cfq_slice_idle
, 0, UINT_MAX
, 1);
4046 STORE_FUNCTION(cfq_group_idle_store
, &cfqd
->cfq_group_idle
, 0, UINT_MAX
, 1);
4047 STORE_FUNCTION(cfq_slice_sync_store
, &cfqd
->cfq_slice
[1], 1, UINT_MAX
, 1);
4048 STORE_FUNCTION(cfq_slice_async_store
, &cfqd
->cfq_slice
[0], 1, UINT_MAX
, 1);
4049 STORE_FUNCTION(cfq_slice_async_rq_store
, &cfqd
->cfq_slice_async_rq
, 1,
4051 STORE_FUNCTION(cfq_low_latency_store
, &cfqd
->cfq_latency
, 0, 1, 0);
4052 #undef STORE_FUNCTION
4054 #define CFQ_ATTR(name) \
4055 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
4057 static struct elv_fs_entry cfq_attrs
[] = {
4059 CFQ_ATTR(fifo_expire_sync
),
4060 CFQ_ATTR(fifo_expire_async
),
4061 CFQ_ATTR(back_seek_max
),
4062 CFQ_ATTR(back_seek_penalty
),
4063 CFQ_ATTR(slice_sync
),
4064 CFQ_ATTR(slice_async
),
4065 CFQ_ATTR(slice_async_rq
),
4066 CFQ_ATTR(slice_idle
),
4067 CFQ_ATTR(group_idle
),
4068 CFQ_ATTR(low_latency
),
4072 static struct elevator_type iosched_cfq
= {
4074 .elevator_merge_fn
= cfq_merge
,
4075 .elevator_merged_fn
= cfq_merged_request
,
4076 .elevator_merge_req_fn
= cfq_merged_requests
,
4077 .elevator_allow_merge_fn
= cfq_allow_merge
,
4078 .elevator_bio_merged_fn
= cfq_bio_merged
,
4079 .elevator_dispatch_fn
= cfq_dispatch_requests
,
4080 .elevator_add_req_fn
= cfq_insert_request
,
4081 .elevator_activate_req_fn
= cfq_activate_request
,
4082 .elevator_deactivate_req_fn
= cfq_deactivate_request
,
4083 .elevator_queue_empty_fn
= cfq_queue_empty
,
4084 .elevator_completed_req_fn
= cfq_completed_request
,
4085 .elevator_former_req_fn
= elv_rb_former_request
,
4086 .elevator_latter_req_fn
= elv_rb_latter_request
,
4087 .elevator_set_req_fn
= cfq_set_request
,
4088 .elevator_put_req_fn
= cfq_put_request
,
4089 .elevator_may_queue_fn
= cfq_may_queue
,
4090 .elevator_init_fn
= cfq_init_queue
,
4091 .elevator_exit_fn
= cfq_exit_queue
,
4092 .trim
= cfq_free_io_context
,
4094 .elevator_attrs
= cfq_attrs
,
4095 .elevator_name
= "cfq",
4096 .elevator_owner
= THIS_MODULE
,
4099 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4100 static struct blkio_policy_type blkio_policy_cfq
= {
4102 .blkio_unlink_group_fn
= cfq_unlink_blkio_group
,
4103 .blkio_update_group_weight_fn
= cfq_update_blkio_group_weight
,
4105 .plid
= BLKIO_POLICY_PROP
,
4108 static struct blkio_policy_type blkio_policy_cfq
;
4111 static int __init
cfq_init(void)
4114 * could be 0 on HZ < 1000 setups
4116 if (!cfq_slice_async
)
4117 cfq_slice_async
= 1;
4118 if (!cfq_slice_idle
)
4121 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4122 if (!cfq_group_idle
)
4127 if (cfq_slab_setup())
4130 elv_register(&iosched_cfq
);
4131 blkio_policy_register(&blkio_policy_cfq
);
4136 static void __exit
cfq_exit(void)
4138 DECLARE_COMPLETION_ONSTACK(all_gone
);
4139 blkio_policy_unregister(&blkio_policy_cfq
);
4140 elv_unregister(&iosched_cfq
);
4141 ioc_gone
= &all_gone
;
4142 /* ioc_gone's update must be visible before reading ioc_count */
4146 * this also protects us from entering cfq_slab_kill() with
4147 * pending RCU callbacks
4149 if (elv_ioc_count_read(cfq_ioc_count
))
4150 wait_for_completion(&all_gone
);
4151 ida_destroy(&cic_index_ida
);
4155 module_init(cfq_init
);
4156 module_exit(cfq_exit
);
4158 MODULE_AUTHOR("Jens Axboe");
4159 MODULE_LICENSE("GPL");
4160 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");