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>
23 /* max queue in one round of service */
24 static const int cfq_quantum
= 8;
25 static const int cfq_fifo_expire
[2] = { HZ
/ 4, HZ
/ 8 };
26 /* maximum backwards seek, in KiB */
27 static const int cfq_back_max
= 16 * 1024;
28 /* penalty of a backwards seek */
29 static const int cfq_back_penalty
= 2;
30 static const int cfq_slice_sync
= HZ
/ 10;
31 static int cfq_slice_async
= HZ
/ 25;
32 static const int cfq_slice_async_rq
= 2;
33 static int cfq_slice_idle
= HZ
/ 125;
34 static int cfq_group_idle
= HZ
/ 125;
35 static const int cfq_target_latency
= HZ
* 3/10; /* 300 ms */
36 static const int cfq_hist_divisor
= 4;
39 * offset from end of service tree
41 #define CFQ_IDLE_DELAY (HZ / 5)
44 * below this threshold, we consider thinktime immediate
46 #define CFQ_MIN_TT (2)
48 #define CFQ_SLICE_SCALE (5)
49 #define CFQ_HW_QUEUE_MIN (5)
50 #define CFQ_SERVICE_SHIFT 12
52 #define CFQQ_SEEK_THR (sector_t)(8 * 100)
53 #define CFQQ_CLOSE_THR (sector_t)(8 * 1024)
54 #define CFQQ_SECT_THR_NONROT (sector_t)(2 * 32)
55 #define CFQQ_SEEKY(cfqq) (hweight32(cfqq->seek_history) > 32/8)
58 ((struct cfq_io_context *) (rq)->elevator_private[0])
59 #define RQ_CFQQ(rq) (struct cfq_queue *) ((rq)->elevator_private[1])
60 #define RQ_CFQG(rq) (struct cfq_group *) ((rq)->elevator_private[2])
62 static struct kmem_cache
*cfq_pool
;
63 static struct kmem_cache
*cfq_ioc_pool
;
65 #define CFQ_PRIO_LISTS IOPRIO_BE_NR
66 #define cfq_class_idle(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
67 #define cfq_class_rt(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
69 #define sample_valid(samples) ((samples) > 80)
70 #define rb_entry_cfqg(node) rb_entry((node), struct cfq_group, rb_node)
73 * Most of our rbtree usage is for sorting with min extraction, so
74 * if we cache the leftmost node we don't have to walk down the tree
75 * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should
76 * move this into the elevator for the rq sorting as well.
82 unsigned total_weight
;
84 struct cfq_ttime ttime
;
86 #define CFQ_RB_ROOT (struct cfq_rb_root) { .rb = RB_ROOT, \
87 .ttime = {.last_end_request = jiffies,},}
90 * Per process-grouping structure
95 /* various state flags, see below */
98 struct cfq_data
*cfqd
;
99 /* service_tree member */
100 struct rb_node rb_node
;
101 /* service_tree key */
102 unsigned long rb_key
;
103 /* prio tree member */
104 struct rb_node p_node
;
105 /* prio tree root we belong to, if any */
106 struct rb_root
*p_root
;
107 /* sorted list of pending requests */
108 struct rb_root sort_list
;
109 /* if fifo isn't expired, next request to serve */
110 struct request
*next_rq
;
111 /* requests queued in sort_list */
113 /* currently allocated requests */
115 /* fifo list of requests in sort_list */
116 struct list_head fifo
;
118 /* time when queue got scheduled in to dispatch first request. */
119 unsigned long dispatch_start
;
120 unsigned int allocated_slice
;
121 unsigned int slice_dispatch
;
122 /* time when first request from queue completed and slice started. */
123 unsigned long slice_start
;
124 unsigned long slice_end
;
127 /* pending priority requests */
129 /* number of requests that are on the dispatch list or inside driver */
132 /* io prio of this group */
133 unsigned short ioprio
, org_ioprio
;
134 unsigned short ioprio_class
;
139 sector_t last_request_pos
;
141 struct cfq_rb_root
*service_tree
;
142 struct cfq_queue
*new_cfqq
;
143 struct cfq_group
*cfqg
;
144 /* Number of sectors dispatched from queue in single dispatch round */
145 unsigned long nr_sectors
;
149 * First index in the service_trees.
150 * IDLE is handled separately, so it has negative index
160 * Second index in the service_trees.
164 SYNC_NOIDLE_WORKLOAD
= 1,
168 /* This is per cgroup per device grouping structure */
170 /* group service_tree member */
171 struct rb_node rb_node
;
173 /* group service_tree key */
176 unsigned int new_weight
;
179 /* number of cfqq currently on this group */
183 * Per group busy queues average. Useful for workload slice calc. We
184 * create the array for each prio class but at run time it is used
185 * only for RT and BE class and slot for IDLE class remains unused.
186 * This is primarily done to avoid confusion and a gcc warning.
188 unsigned int busy_queues_avg
[CFQ_PRIO_NR
];
190 * rr lists of queues with requests. We maintain service trees for
191 * RT and BE classes. These trees are subdivided in subclasses
192 * of SYNC, SYNC_NOIDLE and ASYNC based on workload type. For IDLE
193 * class there is no subclassification and all the cfq queues go on
194 * a single tree service_tree_idle.
195 * Counts are embedded in the cfq_rb_root
197 struct cfq_rb_root service_trees
[2][3];
198 struct cfq_rb_root service_tree_idle
;
200 unsigned long saved_workload_slice
;
201 enum wl_type_t saved_workload
;
202 enum wl_prio_t saved_serving_prio
;
203 struct blkio_group blkg
;
204 #ifdef CONFIG_CFQ_GROUP_IOSCHED
205 struct hlist_node cfqd_node
;
208 /* number of requests that are on the dispatch list or inside driver */
210 struct cfq_ttime ttime
;
214 * Per block device queue structure
217 struct request_queue
*queue
;
218 /* Root service tree for cfq_groups */
219 struct cfq_rb_root grp_service_tree
;
220 struct cfq_group root_group
;
223 * The priority currently being served
225 enum wl_prio_t serving_prio
;
226 enum wl_type_t serving_type
;
227 unsigned long workload_expires
;
228 struct cfq_group
*serving_group
;
231 * Each priority tree is sorted by next_request position. These
232 * trees are used when determining if two or more queues are
233 * interleaving requests (see cfq_close_cooperator).
235 struct rb_root prio_trees
[CFQ_PRIO_LISTS
];
237 unsigned int busy_queues
;
238 unsigned int busy_sync_queues
;
244 * queue-depth detection
250 * -1 => indeterminate, (cfq will behave as if NCQ is present, to allow better detection)
251 * 1 => NCQ is present (hw_tag_est_depth is the estimated max depth)
254 int hw_tag_est_depth
;
255 unsigned int hw_tag_samples
;
258 * idle window management
260 struct timer_list idle_slice_timer
;
261 struct work_struct unplug_work
;
263 struct cfq_queue
*active_queue
;
264 struct cfq_io_context
*active_cic
;
267 * async queue for each priority case
269 struct cfq_queue
*async_cfqq
[2][IOPRIO_BE_NR
];
270 struct cfq_queue
*async_idle_cfqq
;
272 sector_t last_position
;
275 * tunables, see top of file
277 unsigned int cfq_quantum
;
278 unsigned int cfq_fifo_expire
[2];
279 unsigned int cfq_back_penalty
;
280 unsigned int cfq_back_max
;
281 unsigned int cfq_slice
[2];
282 unsigned int cfq_slice_async_rq
;
283 unsigned int cfq_slice_idle
;
284 unsigned int cfq_group_idle
;
285 unsigned int cfq_latency
;
287 struct list_head cic_list
;
290 * Fallback dummy cfqq for extreme OOM conditions
292 struct cfq_queue oom_cfqq
;
294 unsigned long last_delayed_sync
;
296 /* List of cfq groups being managed on this device*/
297 struct hlist_head cfqg_list
;
299 /* Number of groups which are on blkcg->blkg_list */
300 unsigned int nr_blkcg_linked_grps
;
303 static struct cfq_group
*cfq_get_next_cfqg(struct cfq_data
*cfqd
);
305 static struct cfq_rb_root
*service_tree_for(struct cfq_group
*cfqg
,
312 if (prio
== IDLE_WORKLOAD
)
313 return &cfqg
->service_tree_idle
;
315 return &cfqg
->service_trees
[prio
][type
];
318 enum cfqq_state_flags
{
319 CFQ_CFQQ_FLAG_on_rr
= 0, /* on round-robin busy list */
320 CFQ_CFQQ_FLAG_wait_request
, /* waiting for a request */
321 CFQ_CFQQ_FLAG_must_dispatch
, /* must be allowed a dispatch */
322 CFQ_CFQQ_FLAG_must_alloc_slice
, /* per-slice must_alloc flag */
323 CFQ_CFQQ_FLAG_fifo_expire
, /* FIFO checked in this slice */
324 CFQ_CFQQ_FLAG_idle_window
, /* slice idling enabled */
325 CFQ_CFQQ_FLAG_prio_changed
, /* task priority has changed */
326 CFQ_CFQQ_FLAG_slice_new
, /* no requests dispatched in slice */
327 CFQ_CFQQ_FLAG_sync
, /* synchronous queue */
328 CFQ_CFQQ_FLAG_coop
, /* cfqq is shared */
329 CFQ_CFQQ_FLAG_split_coop
, /* shared cfqq will be splitted */
330 CFQ_CFQQ_FLAG_deep
, /* sync cfqq experienced large depth */
331 CFQ_CFQQ_FLAG_wait_busy
, /* Waiting for next request */
334 #define CFQ_CFQQ_FNS(name) \
335 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \
337 (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name); \
339 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \
341 (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \
343 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \
345 return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \
349 CFQ_CFQQ_FNS(wait_request
);
350 CFQ_CFQQ_FNS(must_dispatch
);
351 CFQ_CFQQ_FNS(must_alloc_slice
);
352 CFQ_CFQQ_FNS(fifo_expire
);
353 CFQ_CFQQ_FNS(idle_window
);
354 CFQ_CFQQ_FNS(prio_changed
);
355 CFQ_CFQQ_FNS(slice_new
);
358 CFQ_CFQQ_FNS(split_coop
);
360 CFQ_CFQQ_FNS(wait_busy
);
363 #ifdef CONFIG_CFQ_GROUP_IOSCHED
364 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
365 blk_add_trace_msg((cfqd)->queue, "cfq%d%c %s " fmt, (cfqq)->pid, \
366 cfq_cfqq_sync((cfqq)) ? 'S' : 'A', \
367 blkg_path(&(cfqq)->cfqg->blkg), ##args)
369 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...) \
370 blk_add_trace_msg((cfqd)->queue, "%s " fmt, \
371 blkg_path(&(cfqg)->blkg), ##args) \
374 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
375 blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args)
376 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...) do {} while (0)
378 #define cfq_log(cfqd, fmt, args...) \
379 blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
381 /* Traverses through cfq group service trees */
382 #define for_each_cfqg_st(cfqg, i, j, st) \
383 for (i = 0; i <= IDLE_WORKLOAD; i++) \
384 for (j = 0, st = i < IDLE_WORKLOAD ? &cfqg->service_trees[i][j]\
385 : &cfqg->service_tree_idle; \
386 (i < IDLE_WORKLOAD && j <= SYNC_WORKLOAD) || \
387 (i == IDLE_WORKLOAD && j == 0); \
388 j++, st = i < IDLE_WORKLOAD ? \
389 &cfqg->service_trees[i][j]: NULL) \
391 static inline bool cfq_io_thinktime_big(struct cfq_data *cfqd,
392 struct cfq_ttime
*ttime
, bool group_idle
)
395 if (!sample_valid(ttime
->ttime_samples
))
398 slice
= cfqd
->cfq_group_idle
;
400 slice
= cfqd
->cfq_slice_idle
;
401 return ttime
->ttime_mean
> slice
;
404 static inline bool iops_mode(struct cfq_data
*cfqd
)
407 * If we are not idling on queues and it is a NCQ drive, parallel
408 * execution of requests is on and measuring time is not possible
409 * in most of the cases until and unless we drive shallower queue
410 * depths and that becomes a performance bottleneck. In such cases
411 * switch to start providing fairness in terms of number of IOs.
413 if (!cfqd
->cfq_slice_idle
&& cfqd
->hw_tag
)
419 static inline enum wl_prio_t
cfqq_prio(struct cfq_queue
*cfqq
)
421 if (cfq_class_idle(cfqq
))
422 return IDLE_WORKLOAD
;
423 if (cfq_class_rt(cfqq
))
429 static enum wl_type_t
cfqq_type(struct cfq_queue
*cfqq
)
431 if (!cfq_cfqq_sync(cfqq
))
432 return ASYNC_WORKLOAD
;
433 if (!cfq_cfqq_idle_window(cfqq
))
434 return SYNC_NOIDLE_WORKLOAD
;
435 return SYNC_WORKLOAD
;
438 static inline int cfq_group_busy_queues_wl(enum wl_prio_t wl
,
439 struct cfq_data
*cfqd
,
440 struct cfq_group
*cfqg
)
442 if (wl
== IDLE_WORKLOAD
)
443 return cfqg
->service_tree_idle
.count
;
445 return cfqg
->service_trees
[wl
][ASYNC_WORKLOAD
].count
446 + cfqg
->service_trees
[wl
][SYNC_NOIDLE_WORKLOAD
].count
447 + cfqg
->service_trees
[wl
][SYNC_WORKLOAD
].count
;
450 static inline int cfqg_busy_async_queues(struct cfq_data
*cfqd
,
451 struct cfq_group
*cfqg
)
453 return cfqg
->service_trees
[RT_WORKLOAD
][ASYNC_WORKLOAD
].count
454 + cfqg
->service_trees
[BE_WORKLOAD
][ASYNC_WORKLOAD
].count
;
457 static void cfq_dispatch_insert(struct request_queue
*, struct request
*);
458 static struct cfq_queue
*cfq_get_queue(struct cfq_data
*, bool,
459 struct io_context
*, gfp_t
);
460 static struct cfq_io_context
*cfq_cic_lookup(struct cfq_data
*,
461 struct io_context
*);
463 static inline struct cfq_queue
*cic_to_cfqq(struct cfq_io_context
*cic
,
466 return cic
->cfqq
[is_sync
];
469 static inline void cic_set_cfqq(struct cfq_io_context
*cic
,
470 struct cfq_queue
*cfqq
, bool is_sync
)
472 cic
->cfqq
[is_sync
] = cfqq
;
475 #define CIC_DEAD_KEY 1ul
476 #define CIC_DEAD_INDEX_SHIFT 1
478 static inline void *cfqd_dead_key(struct cfq_data
*cfqd
)
480 return (void *)(cfqd
->queue
->id
<< CIC_DEAD_INDEX_SHIFT
| CIC_DEAD_KEY
);
483 static inline struct cfq_data
*cic_to_cfqd(struct cfq_io_context
*cic
)
485 struct cfq_data
*cfqd
= cic
->key
;
487 if (unlikely((unsigned long) cfqd
& CIC_DEAD_KEY
))
494 * We regard a request as SYNC, if it's either a read or has the SYNC bit
495 * set (in which case it could also be direct WRITE).
497 static inline bool cfq_bio_sync(struct bio
*bio
)
499 return bio_data_dir(bio
) == READ
|| (bio
->bi_rw
& REQ_SYNC
);
503 * scheduler run of queue, if there are requests pending and no one in the
504 * driver that will restart queueing
506 static inline void cfq_schedule_dispatch(struct cfq_data
*cfqd
)
508 if (cfqd
->busy_queues
) {
509 cfq_log(cfqd
, "schedule dispatch");
510 kblockd_schedule_work(cfqd
->queue
, &cfqd
->unplug_work
);
515 * Scale schedule slice based on io priority. Use the sync time slice only
516 * if a queue is marked sync and has sync io queued. A sync queue with async
517 * io only, should not get full sync slice length.
519 static inline int cfq_prio_slice(struct cfq_data
*cfqd
, bool sync
,
522 const int base_slice
= cfqd
->cfq_slice
[sync
];
524 WARN_ON(prio
>= IOPRIO_BE_NR
);
526 return base_slice
+ (base_slice
/CFQ_SLICE_SCALE
* (4 - prio
));
530 cfq_prio_to_slice(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
532 return cfq_prio_slice(cfqd
, cfq_cfqq_sync(cfqq
), cfqq
->ioprio
);
535 static inline u64
cfq_scale_slice(unsigned long delta
, struct cfq_group
*cfqg
)
537 u64 d
= delta
<< CFQ_SERVICE_SHIFT
;
539 d
= d
* BLKIO_WEIGHT_DEFAULT
;
540 do_div(d
, cfqg
->weight
);
544 static inline u64
max_vdisktime(u64 min_vdisktime
, u64 vdisktime
)
546 s64 delta
= (s64
)(vdisktime
- min_vdisktime
);
548 min_vdisktime
= vdisktime
;
550 return min_vdisktime
;
553 static inline u64
min_vdisktime(u64 min_vdisktime
, u64 vdisktime
)
555 s64 delta
= (s64
)(vdisktime
- min_vdisktime
);
557 min_vdisktime
= vdisktime
;
559 return min_vdisktime
;
562 static void update_min_vdisktime(struct cfq_rb_root
*st
)
564 struct cfq_group
*cfqg
;
567 cfqg
= rb_entry_cfqg(st
->left
);
568 st
->min_vdisktime
= max_vdisktime(st
->min_vdisktime
,
574 * get averaged number of queues of RT/BE priority.
575 * average is updated, with a formula that gives more weight to higher numbers,
576 * to quickly follows sudden increases and decrease slowly
579 static inline unsigned cfq_group_get_avg_queues(struct cfq_data
*cfqd
,
580 struct cfq_group
*cfqg
, bool rt
)
582 unsigned min_q
, max_q
;
583 unsigned mult
= cfq_hist_divisor
- 1;
584 unsigned round
= cfq_hist_divisor
/ 2;
585 unsigned busy
= cfq_group_busy_queues_wl(rt
, cfqd
, cfqg
);
587 min_q
= min(cfqg
->busy_queues_avg
[rt
], busy
);
588 max_q
= max(cfqg
->busy_queues_avg
[rt
], busy
);
589 cfqg
->busy_queues_avg
[rt
] = (mult
* max_q
+ min_q
+ round
) /
591 return cfqg
->busy_queues_avg
[rt
];
594 static inline unsigned
595 cfq_group_slice(struct cfq_data
*cfqd
, struct cfq_group
*cfqg
)
597 struct cfq_rb_root
*st
= &cfqd
->grp_service_tree
;
599 return cfq_target_latency
* cfqg
->weight
/ st
->total_weight
;
602 static inline unsigned
603 cfq_scaled_cfqq_slice(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
605 unsigned slice
= cfq_prio_to_slice(cfqd
, cfqq
);
606 if (cfqd
->cfq_latency
) {
608 * interested queues (we consider only the ones with the same
609 * priority class in the cfq group)
611 unsigned iq
= cfq_group_get_avg_queues(cfqd
, cfqq
->cfqg
,
613 unsigned sync_slice
= cfqd
->cfq_slice
[1];
614 unsigned expect_latency
= sync_slice
* iq
;
615 unsigned group_slice
= cfq_group_slice(cfqd
, cfqq
->cfqg
);
617 if (expect_latency
> group_slice
) {
618 unsigned base_low_slice
= 2 * cfqd
->cfq_slice_idle
;
619 /* scale low_slice according to IO priority
620 * and sync vs async */
622 min(slice
, base_low_slice
* slice
/ sync_slice
);
623 /* the adapted slice value is scaled to fit all iqs
624 * into the target latency */
625 slice
= max(slice
* group_slice
/ expect_latency
,
633 cfq_set_prio_slice(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
635 unsigned slice
= cfq_scaled_cfqq_slice(cfqd
, cfqq
);
637 cfqq
->slice_start
= jiffies
;
638 cfqq
->slice_end
= jiffies
+ slice
;
639 cfqq
->allocated_slice
= slice
;
640 cfq_log_cfqq(cfqd
, cfqq
, "set_slice=%lu", cfqq
->slice_end
- jiffies
);
644 * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
645 * isn't valid until the first request from the dispatch is activated
646 * and the slice time set.
648 static inline bool cfq_slice_used(struct cfq_queue
*cfqq
)
650 if (cfq_cfqq_slice_new(cfqq
))
652 if (time_before(jiffies
, cfqq
->slice_end
))
659 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
660 * We choose the request that is closest to the head right now. Distance
661 * behind the head is penalized and only allowed to a certain extent.
663 static struct request
*
664 cfq_choose_req(struct cfq_data
*cfqd
, struct request
*rq1
, struct request
*rq2
, sector_t last
)
666 sector_t s1
, s2
, d1
= 0, d2
= 0;
667 unsigned long back_max
;
668 #define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */
669 #define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */
670 unsigned wrap
= 0; /* bit mask: requests behind the disk head? */
672 if (rq1
== NULL
|| rq1
== rq2
)
677 if (rq_is_sync(rq1
) != rq_is_sync(rq2
))
678 return rq_is_sync(rq1
) ? rq1
: rq2
;
680 if ((rq1
->cmd_flags
^ rq2
->cmd_flags
) & REQ_PRIO
)
681 return rq1
->cmd_flags
& REQ_PRIO
? rq1
: rq2
;
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_update_group_weight(struct cfq_group
*cfqg
)
865 BUG_ON(!RB_EMPTY_NODE(&cfqg
->rb_node
));
866 if (cfqg
->needs_update
) {
867 cfqg
->weight
= cfqg
->new_weight
;
868 cfqg
->needs_update
= false;
873 cfq_group_service_tree_add(struct cfq_rb_root
*st
, struct cfq_group
*cfqg
)
875 BUG_ON(!RB_EMPTY_NODE(&cfqg
->rb_node
));
877 cfq_update_group_weight(cfqg
);
878 __cfq_group_service_tree_add(st
, cfqg
);
879 st
->total_weight
+= cfqg
->weight
;
883 cfq_group_notify_queue_add(struct cfq_data
*cfqd
, struct cfq_group
*cfqg
)
885 struct cfq_rb_root
*st
= &cfqd
->grp_service_tree
;
886 struct cfq_group
*__cfqg
;
890 if (!RB_EMPTY_NODE(&cfqg
->rb_node
))
894 * Currently put the group at the end. Later implement something
895 * so that groups get lesser vtime based on their weights, so that
896 * if group does not loose all if it was not continuously backlogged.
898 n
= rb_last(&st
->rb
);
900 __cfqg
= rb_entry_cfqg(n
);
901 cfqg
->vdisktime
= __cfqg
->vdisktime
+ CFQ_IDLE_DELAY
;
903 cfqg
->vdisktime
= st
->min_vdisktime
;
904 cfq_group_service_tree_add(st
, cfqg
);
908 cfq_group_service_tree_del(struct cfq_rb_root
*st
, struct cfq_group
*cfqg
)
910 st
->total_weight
-= cfqg
->weight
;
911 if (!RB_EMPTY_NODE(&cfqg
->rb_node
))
912 cfq_rb_erase(&cfqg
->rb_node
, st
);
916 cfq_group_notify_queue_del(struct cfq_data
*cfqd
, struct cfq_group
*cfqg
)
918 struct cfq_rb_root
*st
= &cfqd
->grp_service_tree
;
920 BUG_ON(cfqg
->nr_cfqq
< 1);
923 /* If there are other cfq queues under this group, don't delete it */
927 cfq_log_cfqg(cfqd
, cfqg
, "del_from_rr group");
928 cfq_group_service_tree_del(st
, cfqg
);
929 cfqg
->saved_workload_slice
= 0;
930 cfq_blkiocg_update_dequeue_stats(&cfqg
->blkg
, 1);
933 static inline unsigned int cfq_cfqq_slice_usage(struct cfq_queue
*cfqq
,
934 unsigned int *unaccounted_time
)
936 unsigned int slice_used
;
939 * Queue got expired before even a single request completed or
940 * got expired immediately after first request completion.
942 if (!cfqq
->slice_start
|| cfqq
->slice_start
== jiffies
) {
944 * Also charge the seek time incurred to the group, otherwise
945 * if there are mutiple queues in the group, each can dispatch
946 * a single request on seeky media and cause lots of seek time
947 * and group will never know it.
949 slice_used
= max_t(unsigned, (jiffies
- cfqq
->dispatch_start
),
952 slice_used
= jiffies
- cfqq
->slice_start
;
953 if (slice_used
> cfqq
->allocated_slice
) {
954 *unaccounted_time
= slice_used
- cfqq
->allocated_slice
;
955 slice_used
= cfqq
->allocated_slice
;
957 if (time_after(cfqq
->slice_start
, cfqq
->dispatch_start
))
958 *unaccounted_time
+= cfqq
->slice_start
-
959 cfqq
->dispatch_start
;
965 static void cfq_group_served(struct cfq_data
*cfqd
, struct cfq_group
*cfqg
,
966 struct cfq_queue
*cfqq
)
968 struct cfq_rb_root
*st
= &cfqd
->grp_service_tree
;
969 unsigned int used_sl
, charge
, unaccounted_sl
= 0;
970 int nr_sync
= cfqg
->nr_cfqq
- cfqg_busy_async_queues(cfqd
, cfqg
)
971 - cfqg
->service_tree_idle
.count
;
974 used_sl
= charge
= cfq_cfqq_slice_usage(cfqq
, &unaccounted_sl
);
977 charge
= cfqq
->slice_dispatch
;
978 else if (!cfq_cfqq_sync(cfqq
) && !nr_sync
)
979 charge
= cfqq
->allocated_slice
;
981 /* Can't update vdisktime while group is on service tree */
982 cfq_group_service_tree_del(st
, cfqg
);
983 cfqg
->vdisktime
+= cfq_scale_slice(charge
, cfqg
);
984 /* If a new weight was requested, update now, off tree */
985 cfq_group_service_tree_add(st
, cfqg
);
987 /* This group is being expired. Save the context */
988 if (time_after(cfqd
->workload_expires
, jiffies
)) {
989 cfqg
->saved_workload_slice
= cfqd
->workload_expires
991 cfqg
->saved_workload
= cfqd
->serving_type
;
992 cfqg
->saved_serving_prio
= cfqd
->serving_prio
;
994 cfqg
->saved_workload_slice
= 0;
996 cfq_log_cfqg(cfqd
, cfqg
, "served: vt=%llu min_vt=%llu", cfqg
->vdisktime
,
998 cfq_log_cfqq(cfqq
->cfqd
, cfqq
,
999 "sl_used=%u disp=%u charge=%u iops=%u sect=%lu",
1000 used_sl
, cfqq
->slice_dispatch
, charge
,
1001 iops_mode(cfqd
), cfqq
->nr_sectors
);
1002 cfq_blkiocg_update_timeslice_used(&cfqg
->blkg
, used_sl
,
1004 cfq_blkiocg_set_start_empty_time(&cfqg
->blkg
);
1007 #ifdef CONFIG_CFQ_GROUP_IOSCHED
1008 static inline struct cfq_group
*cfqg_of_blkg(struct blkio_group
*blkg
)
1011 return container_of(blkg
, struct cfq_group
, blkg
);
1015 static void cfq_update_blkio_group_weight(void *key
, struct blkio_group
*blkg
,
1016 unsigned int weight
)
1018 struct cfq_group
*cfqg
= cfqg_of_blkg(blkg
);
1019 cfqg
->new_weight
= weight
;
1020 cfqg
->needs_update
= true;
1023 static void cfq_init_add_cfqg_lists(struct cfq_data
*cfqd
,
1024 struct cfq_group
*cfqg
, struct blkio_cgroup
*blkcg
)
1026 struct backing_dev_info
*bdi
= &cfqd
->queue
->backing_dev_info
;
1027 unsigned int major
, minor
;
1030 * Add group onto cgroup list. It might happen that bdi->dev is
1031 * not initialized yet. Initialize this new group without major
1032 * and minor info and this info will be filled in once a new thread
1036 sscanf(dev_name(bdi
->dev
), "%u:%u", &major
, &minor
);
1037 cfq_blkiocg_add_blkio_group(blkcg
, &cfqg
->blkg
,
1038 (void *)cfqd
, MKDEV(major
, minor
));
1040 cfq_blkiocg_add_blkio_group(blkcg
, &cfqg
->blkg
,
1043 cfqd
->nr_blkcg_linked_grps
++;
1044 cfqg
->weight
= blkcg_get_weight(blkcg
, cfqg
->blkg
.dev
);
1046 /* Add group on cfqd list */
1047 hlist_add_head(&cfqg
->cfqd_node
, &cfqd
->cfqg_list
);
1051 * Should be called from sleepable context. No request queue lock as per
1052 * cpu stats are allocated dynamically and alloc_percpu needs to be called
1053 * from sleepable context.
1055 static struct cfq_group
* cfq_alloc_cfqg(struct cfq_data
*cfqd
)
1057 struct cfq_group
*cfqg
= NULL
;
1059 struct cfq_rb_root
*st
;
1061 cfqg
= kzalloc_node(sizeof(*cfqg
), GFP_ATOMIC
, cfqd
->queue
->node
);
1065 for_each_cfqg_st(cfqg
, i
, j
, st
)
1067 RB_CLEAR_NODE(&cfqg
->rb_node
);
1069 cfqg
->ttime
.last_end_request
= jiffies
;
1072 * Take the initial reference that will be released on destroy
1073 * This can be thought of a joint reference by cgroup and
1074 * elevator which will be dropped by either elevator exit
1075 * or cgroup deletion path depending on who is exiting first.
1079 ret
= blkio_alloc_blkg_stats(&cfqg
->blkg
);
1088 static struct cfq_group
*
1089 cfq_find_cfqg(struct cfq_data
*cfqd
, struct blkio_cgroup
*blkcg
)
1091 struct cfq_group
*cfqg
= NULL
;
1093 struct backing_dev_info
*bdi
= &cfqd
->queue
->backing_dev_info
;
1094 unsigned int major
, minor
;
1097 * This is the common case when there are no blkio cgroups.
1098 * Avoid lookup in this case
1100 if (blkcg
== &blkio_root_cgroup
)
1101 cfqg
= &cfqd
->root_group
;
1103 cfqg
= cfqg_of_blkg(blkiocg_lookup_group(blkcg
, key
));
1105 if (cfqg
&& !cfqg
->blkg
.dev
&& bdi
->dev
&& dev_name(bdi
->dev
)) {
1106 sscanf(dev_name(bdi
->dev
), "%u:%u", &major
, &minor
);
1107 cfqg
->blkg
.dev
= MKDEV(major
, minor
);
1114 * Search for the cfq group current task belongs to. request_queue lock must
1117 static struct cfq_group
*cfq_get_cfqg(struct cfq_data
*cfqd
)
1119 struct blkio_cgroup
*blkcg
;
1120 struct cfq_group
*cfqg
= NULL
, *__cfqg
= NULL
;
1121 struct request_queue
*q
= cfqd
->queue
;
1124 blkcg
= task_blkio_cgroup(current
);
1125 cfqg
= cfq_find_cfqg(cfqd
, blkcg
);
1132 * Need to allocate a group. Allocation of group also needs allocation
1133 * of per cpu stats which in-turn takes a mutex() and can block. Hence
1134 * we need to drop rcu lock and queue_lock before we call alloc.
1136 * Not taking any queue reference here and assuming that queue is
1137 * around by the time we return. CFQ queue allocation code does
1138 * the same. It might be racy though.
1142 spin_unlock_irq(q
->queue_lock
);
1144 cfqg
= cfq_alloc_cfqg(cfqd
);
1146 spin_lock_irq(q
->queue_lock
);
1149 blkcg
= task_blkio_cgroup(current
);
1152 * If some other thread already allocated the group while we were
1153 * not holding queue lock, free up the group
1155 __cfqg
= cfq_find_cfqg(cfqd
, blkcg
);
1164 cfqg
= &cfqd
->root_group
;
1166 cfq_init_add_cfqg_lists(cfqd
, cfqg
, blkcg
);
1171 static inline struct cfq_group
*cfq_ref_get_cfqg(struct cfq_group
*cfqg
)
1177 static void cfq_link_cfqq_cfqg(struct cfq_queue
*cfqq
, struct cfq_group
*cfqg
)
1179 /* Currently, all async queues are mapped to root group */
1180 if (!cfq_cfqq_sync(cfqq
))
1181 cfqg
= &cfqq
->cfqd
->root_group
;
1184 /* cfqq reference on cfqg */
1188 static void cfq_put_cfqg(struct cfq_group
*cfqg
)
1190 struct cfq_rb_root
*st
;
1193 BUG_ON(cfqg
->ref
<= 0);
1197 for_each_cfqg_st(cfqg
, i
, j
, st
)
1198 BUG_ON(!RB_EMPTY_ROOT(&st
->rb
));
1199 free_percpu(cfqg
->blkg
.stats_cpu
);
1203 static void cfq_destroy_cfqg(struct cfq_data
*cfqd
, struct cfq_group
*cfqg
)
1205 /* Something wrong if we are trying to remove same group twice */
1206 BUG_ON(hlist_unhashed(&cfqg
->cfqd_node
));
1208 hlist_del_init(&cfqg
->cfqd_node
);
1210 BUG_ON(cfqd
->nr_blkcg_linked_grps
<= 0);
1211 cfqd
->nr_blkcg_linked_grps
--;
1214 * Put the reference taken at the time of creation so that when all
1215 * queues are gone, group can be destroyed.
1220 static void cfq_release_cfq_groups(struct cfq_data
*cfqd
)
1222 struct hlist_node
*pos
, *n
;
1223 struct cfq_group
*cfqg
;
1225 hlist_for_each_entry_safe(cfqg
, pos
, n
, &cfqd
->cfqg_list
, cfqd_node
) {
1227 * If cgroup removal path got to blk_group first and removed
1228 * it from cgroup list, then it will take care of destroying
1231 if (!cfq_blkiocg_del_blkio_group(&cfqg
->blkg
))
1232 cfq_destroy_cfqg(cfqd
, cfqg
);
1237 * Blk cgroup controller notification saying that blkio_group object is being
1238 * delinked as associated cgroup object is going away. That also means that
1239 * no new IO will come in this group. So get rid of this group as soon as
1240 * any pending IO in the group is finished.
1242 * This function is called under rcu_read_lock(). key is the rcu protected
1243 * pointer. That means "key" is a valid cfq_data pointer as long as we are rcu
1246 * "key" was fetched from blkio_group under blkio_cgroup->lock. That means
1247 * it should not be NULL as even if elevator was exiting, cgroup deltion
1248 * path got to it first.
1250 static void cfq_unlink_blkio_group(void *key
, struct blkio_group
*blkg
)
1252 unsigned long flags
;
1253 struct cfq_data
*cfqd
= key
;
1255 spin_lock_irqsave(cfqd
->queue
->queue_lock
, flags
);
1256 cfq_destroy_cfqg(cfqd
, cfqg_of_blkg(blkg
));
1257 spin_unlock_irqrestore(cfqd
->queue
->queue_lock
, flags
);
1260 #else /* GROUP_IOSCHED */
1261 static struct cfq_group
*cfq_get_cfqg(struct cfq_data
*cfqd
)
1263 return &cfqd
->root_group
;
1266 static inline struct cfq_group
*cfq_ref_get_cfqg(struct cfq_group
*cfqg
)
1272 cfq_link_cfqq_cfqg(struct cfq_queue
*cfqq
, struct cfq_group
*cfqg
) {
1276 static void cfq_release_cfq_groups(struct cfq_data
*cfqd
) {}
1277 static inline void cfq_put_cfqg(struct cfq_group
*cfqg
) {}
1279 #endif /* GROUP_IOSCHED */
1282 * The cfqd->service_trees holds all pending cfq_queue's that have
1283 * requests waiting to be processed. It is sorted in the order that
1284 * we will service the queues.
1286 static void cfq_service_tree_add(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
1289 struct rb_node
**p
, *parent
;
1290 struct cfq_queue
*__cfqq
;
1291 unsigned long rb_key
;
1292 struct cfq_rb_root
*service_tree
;
1296 service_tree
= service_tree_for(cfqq
->cfqg
, cfqq_prio(cfqq
),
1298 if (cfq_class_idle(cfqq
)) {
1299 rb_key
= CFQ_IDLE_DELAY
;
1300 parent
= rb_last(&service_tree
->rb
);
1301 if (parent
&& parent
!= &cfqq
->rb_node
) {
1302 __cfqq
= rb_entry(parent
, struct cfq_queue
, rb_node
);
1303 rb_key
+= __cfqq
->rb_key
;
1306 } else if (!add_front
) {
1308 * Get our rb key offset. Subtract any residual slice
1309 * value carried from last service. A negative resid
1310 * count indicates slice overrun, and this should position
1311 * the next service time further away in the tree.
1313 rb_key
= cfq_slice_offset(cfqd
, cfqq
) + jiffies
;
1314 rb_key
-= cfqq
->slice_resid
;
1315 cfqq
->slice_resid
= 0;
1318 __cfqq
= cfq_rb_first(service_tree
);
1319 rb_key
+= __cfqq
? __cfqq
->rb_key
: jiffies
;
1322 if (!RB_EMPTY_NODE(&cfqq
->rb_node
)) {
1325 * same position, nothing more to do
1327 if (rb_key
== cfqq
->rb_key
&&
1328 cfqq
->service_tree
== service_tree
)
1331 cfq_rb_erase(&cfqq
->rb_node
, cfqq
->service_tree
);
1332 cfqq
->service_tree
= NULL
;
1337 cfqq
->service_tree
= service_tree
;
1338 p
= &service_tree
->rb
.rb_node
;
1343 __cfqq
= rb_entry(parent
, struct cfq_queue
, rb_node
);
1346 * sort by key, that represents service time.
1348 if (time_before(rb_key
, __cfqq
->rb_key
))
1351 n
= &(*p
)->rb_right
;
1359 service_tree
->left
= &cfqq
->rb_node
;
1361 cfqq
->rb_key
= rb_key
;
1362 rb_link_node(&cfqq
->rb_node
, parent
, p
);
1363 rb_insert_color(&cfqq
->rb_node
, &service_tree
->rb
);
1364 service_tree
->count
++;
1365 if (add_front
|| !new_cfqq
)
1367 cfq_group_notify_queue_add(cfqd
, cfqq
->cfqg
);
1370 static struct cfq_queue
*
1371 cfq_prio_tree_lookup(struct cfq_data
*cfqd
, struct rb_root
*root
,
1372 sector_t sector
, struct rb_node
**ret_parent
,
1373 struct rb_node
***rb_link
)
1375 struct rb_node
**p
, *parent
;
1376 struct cfq_queue
*cfqq
= NULL
;
1384 cfqq
= rb_entry(parent
, struct cfq_queue
, p_node
);
1387 * Sort strictly based on sector. Smallest to the left,
1388 * largest to the right.
1390 if (sector
> blk_rq_pos(cfqq
->next_rq
))
1391 n
= &(*p
)->rb_right
;
1392 else if (sector
< blk_rq_pos(cfqq
->next_rq
))
1400 *ret_parent
= parent
;
1406 static void cfq_prio_tree_add(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
1408 struct rb_node
**p
, *parent
;
1409 struct cfq_queue
*__cfqq
;
1412 rb_erase(&cfqq
->p_node
, cfqq
->p_root
);
1413 cfqq
->p_root
= NULL
;
1416 if (cfq_class_idle(cfqq
))
1421 cfqq
->p_root
= &cfqd
->prio_trees
[cfqq
->org_ioprio
];
1422 __cfqq
= cfq_prio_tree_lookup(cfqd
, cfqq
->p_root
,
1423 blk_rq_pos(cfqq
->next_rq
), &parent
, &p
);
1425 rb_link_node(&cfqq
->p_node
, parent
, p
);
1426 rb_insert_color(&cfqq
->p_node
, cfqq
->p_root
);
1428 cfqq
->p_root
= NULL
;
1432 * Update cfqq's position in the service tree.
1434 static void cfq_resort_rr_list(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
1437 * Resorting requires the cfqq to be on the RR list already.
1439 if (cfq_cfqq_on_rr(cfqq
)) {
1440 cfq_service_tree_add(cfqd
, cfqq
, 0);
1441 cfq_prio_tree_add(cfqd
, cfqq
);
1446 * add to busy list of queues for service, trying to be fair in ordering
1447 * the pending list according to last request service
1449 static void cfq_add_cfqq_rr(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
1451 cfq_log_cfqq(cfqd
, cfqq
, "add_to_rr");
1452 BUG_ON(cfq_cfqq_on_rr(cfqq
));
1453 cfq_mark_cfqq_on_rr(cfqq
);
1454 cfqd
->busy_queues
++;
1455 if (cfq_cfqq_sync(cfqq
))
1456 cfqd
->busy_sync_queues
++;
1458 cfq_resort_rr_list(cfqd
, cfqq
);
1462 * Called when the cfqq no longer has requests pending, remove it from
1465 static void cfq_del_cfqq_rr(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
1467 cfq_log_cfqq(cfqd
, cfqq
, "del_from_rr");
1468 BUG_ON(!cfq_cfqq_on_rr(cfqq
));
1469 cfq_clear_cfqq_on_rr(cfqq
);
1471 if (!RB_EMPTY_NODE(&cfqq
->rb_node
)) {
1472 cfq_rb_erase(&cfqq
->rb_node
, cfqq
->service_tree
);
1473 cfqq
->service_tree
= NULL
;
1476 rb_erase(&cfqq
->p_node
, cfqq
->p_root
);
1477 cfqq
->p_root
= NULL
;
1480 cfq_group_notify_queue_del(cfqd
, cfqq
->cfqg
);
1481 BUG_ON(!cfqd
->busy_queues
);
1482 cfqd
->busy_queues
--;
1483 if (cfq_cfqq_sync(cfqq
))
1484 cfqd
->busy_sync_queues
--;
1488 * rb tree support functions
1490 static void cfq_del_rq_rb(struct request
*rq
)
1492 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
1493 const int sync
= rq_is_sync(rq
);
1495 BUG_ON(!cfqq
->queued
[sync
]);
1496 cfqq
->queued
[sync
]--;
1498 elv_rb_del(&cfqq
->sort_list
, rq
);
1500 if (cfq_cfqq_on_rr(cfqq
) && RB_EMPTY_ROOT(&cfqq
->sort_list
)) {
1502 * Queue will be deleted from service tree when we actually
1503 * expire it later. Right now just remove it from prio tree
1507 rb_erase(&cfqq
->p_node
, cfqq
->p_root
);
1508 cfqq
->p_root
= NULL
;
1513 static void cfq_add_rq_rb(struct request
*rq
)
1515 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
1516 struct cfq_data
*cfqd
= cfqq
->cfqd
;
1517 struct request
*prev
;
1519 cfqq
->queued
[rq_is_sync(rq
)]++;
1521 elv_rb_add(&cfqq
->sort_list
, rq
);
1523 if (!cfq_cfqq_on_rr(cfqq
))
1524 cfq_add_cfqq_rr(cfqd
, cfqq
);
1527 * check if this request is a better next-serve candidate
1529 prev
= cfqq
->next_rq
;
1530 cfqq
->next_rq
= cfq_choose_req(cfqd
, cfqq
->next_rq
, rq
, cfqd
->last_position
);
1533 * adjust priority tree position, if ->next_rq changes
1535 if (prev
!= cfqq
->next_rq
)
1536 cfq_prio_tree_add(cfqd
, cfqq
);
1538 BUG_ON(!cfqq
->next_rq
);
1541 static void cfq_reposition_rq_rb(struct cfq_queue
*cfqq
, struct request
*rq
)
1543 elv_rb_del(&cfqq
->sort_list
, rq
);
1544 cfqq
->queued
[rq_is_sync(rq
)]--;
1545 cfq_blkiocg_update_io_remove_stats(&(RQ_CFQG(rq
))->blkg
,
1546 rq_data_dir(rq
), rq_is_sync(rq
));
1548 cfq_blkiocg_update_io_add_stats(&(RQ_CFQG(rq
))->blkg
,
1549 &cfqq
->cfqd
->serving_group
->blkg
, rq_data_dir(rq
),
1553 static struct request
*
1554 cfq_find_rq_fmerge(struct cfq_data
*cfqd
, struct bio
*bio
)
1556 struct task_struct
*tsk
= current
;
1557 struct cfq_io_context
*cic
;
1558 struct cfq_queue
*cfqq
;
1560 cic
= cfq_cic_lookup(cfqd
, tsk
->io_context
);
1564 cfqq
= cic_to_cfqq(cic
, cfq_bio_sync(bio
));
1566 sector_t sector
= bio
->bi_sector
+ bio_sectors(bio
);
1568 return elv_rb_find(&cfqq
->sort_list
, sector
);
1574 static void cfq_activate_request(struct request_queue
*q
, struct request
*rq
)
1576 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
1578 cfqd
->rq_in_driver
++;
1579 cfq_log_cfqq(cfqd
, RQ_CFQQ(rq
), "activate rq, drv=%d",
1580 cfqd
->rq_in_driver
);
1582 cfqd
->last_position
= blk_rq_pos(rq
) + blk_rq_sectors(rq
);
1585 static void cfq_deactivate_request(struct request_queue
*q
, struct request
*rq
)
1587 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
1589 WARN_ON(!cfqd
->rq_in_driver
);
1590 cfqd
->rq_in_driver
--;
1591 cfq_log_cfqq(cfqd
, RQ_CFQQ(rq
), "deactivate rq, drv=%d",
1592 cfqd
->rq_in_driver
);
1595 static void cfq_remove_request(struct request
*rq
)
1597 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
1599 if (cfqq
->next_rq
== rq
)
1600 cfqq
->next_rq
= cfq_find_next_rq(cfqq
->cfqd
, cfqq
, rq
);
1602 list_del_init(&rq
->queuelist
);
1605 cfqq
->cfqd
->rq_queued
--;
1606 cfq_blkiocg_update_io_remove_stats(&(RQ_CFQG(rq
))->blkg
,
1607 rq_data_dir(rq
), rq_is_sync(rq
));
1608 if (rq
->cmd_flags
& REQ_PRIO
) {
1609 WARN_ON(!cfqq
->prio_pending
);
1610 cfqq
->prio_pending
--;
1614 static int cfq_merge(struct request_queue
*q
, struct request
**req
,
1617 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
1618 struct request
*__rq
;
1620 __rq
= cfq_find_rq_fmerge(cfqd
, bio
);
1621 if (__rq
&& elv_rq_merge_ok(__rq
, bio
)) {
1623 return ELEVATOR_FRONT_MERGE
;
1626 return ELEVATOR_NO_MERGE
;
1629 static void cfq_merged_request(struct request_queue
*q
, struct request
*req
,
1632 if (type
== ELEVATOR_FRONT_MERGE
) {
1633 struct cfq_queue
*cfqq
= RQ_CFQQ(req
);
1635 cfq_reposition_rq_rb(cfqq
, req
);
1639 static void cfq_bio_merged(struct request_queue
*q
, struct request
*req
,
1642 cfq_blkiocg_update_io_merged_stats(&(RQ_CFQG(req
))->blkg
,
1643 bio_data_dir(bio
), cfq_bio_sync(bio
));
1647 cfq_merged_requests(struct request_queue
*q
, struct request
*rq
,
1648 struct request
*next
)
1650 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
1652 * reposition in fifo if next is older than rq
1654 if (!list_empty(&rq
->queuelist
) && !list_empty(&next
->queuelist
) &&
1655 time_before(rq_fifo_time(next
), rq_fifo_time(rq
))) {
1656 list_move(&rq
->queuelist
, &next
->queuelist
);
1657 rq_set_fifo_time(rq
, rq_fifo_time(next
));
1660 if (cfqq
->next_rq
== next
)
1662 cfq_remove_request(next
);
1663 cfq_blkiocg_update_io_merged_stats(&(RQ_CFQG(rq
))->blkg
,
1664 rq_data_dir(next
), rq_is_sync(next
));
1667 static int cfq_allow_merge(struct request_queue
*q
, struct request
*rq
,
1670 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
1671 struct cfq_io_context
*cic
;
1672 struct cfq_queue
*cfqq
;
1675 * Disallow merge of a sync bio into an async request.
1677 if (cfq_bio_sync(bio
) && !rq_is_sync(rq
))
1681 * Lookup the cfqq that this bio will be queued with and allow
1682 * merge only if rq is queued there. This function can be called
1683 * from plug merge without queue_lock. In such cases, ioc of @rq
1684 * and %current are guaranteed to be equal. Avoid lookup which
1685 * requires queue_lock by using @rq's cic.
1687 if (current
->io_context
== RQ_CIC(rq
)->ioc
) {
1690 cic
= cfq_cic_lookup(cfqd
, current
->io_context
);
1695 cfqq
= cic_to_cfqq(cic
, cfq_bio_sync(bio
));
1696 return cfqq
== RQ_CFQQ(rq
);
1699 static inline void cfq_del_timer(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
1701 del_timer(&cfqd
->idle_slice_timer
);
1702 cfq_blkiocg_update_idle_time_stats(&cfqq
->cfqg
->blkg
);
1705 static void __cfq_set_active_queue(struct cfq_data
*cfqd
,
1706 struct cfq_queue
*cfqq
)
1709 cfq_log_cfqq(cfqd
, cfqq
, "set_active wl_prio:%d wl_type:%d",
1710 cfqd
->serving_prio
, cfqd
->serving_type
);
1711 cfq_blkiocg_update_avg_queue_size_stats(&cfqq
->cfqg
->blkg
);
1712 cfqq
->slice_start
= 0;
1713 cfqq
->dispatch_start
= jiffies
;
1714 cfqq
->allocated_slice
= 0;
1715 cfqq
->slice_end
= 0;
1716 cfqq
->slice_dispatch
= 0;
1717 cfqq
->nr_sectors
= 0;
1719 cfq_clear_cfqq_wait_request(cfqq
);
1720 cfq_clear_cfqq_must_dispatch(cfqq
);
1721 cfq_clear_cfqq_must_alloc_slice(cfqq
);
1722 cfq_clear_cfqq_fifo_expire(cfqq
);
1723 cfq_mark_cfqq_slice_new(cfqq
);
1725 cfq_del_timer(cfqd
, cfqq
);
1728 cfqd
->active_queue
= cfqq
;
1732 * current cfqq expired its slice (or was too idle), select new one
1735 __cfq_slice_expired(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
1738 cfq_log_cfqq(cfqd
, cfqq
, "slice expired t=%d", timed_out
);
1740 if (cfq_cfqq_wait_request(cfqq
))
1741 cfq_del_timer(cfqd
, cfqq
);
1743 cfq_clear_cfqq_wait_request(cfqq
);
1744 cfq_clear_cfqq_wait_busy(cfqq
);
1747 * If this cfqq is shared between multiple processes, check to
1748 * make sure that those processes are still issuing I/Os within
1749 * the mean seek distance. If not, it may be time to break the
1750 * queues apart again.
1752 if (cfq_cfqq_coop(cfqq
) && CFQQ_SEEKY(cfqq
))
1753 cfq_mark_cfqq_split_coop(cfqq
);
1756 * store what was left of this slice, if the queue idled/timed out
1759 if (cfq_cfqq_slice_new(cfqq
))
1760 cfqq
->slice_resid
= cfq_scaled_cfqq_slice(cfqd
, cfqq
);
1762 cfqq
->slice_resid
= cfqq
->slice_end
- jiffies
;
1763 cfq_log_cfqq(cfqd
, cfqq
, "resid=%ld", cfqq
->slice_resid
);
1766 cfq_group_served(cfqd
, cfqq
->cfqg
, cfqq
);
1768 if (cfq_cfqq_on_rr(cfqq
) && RB_EMPTY_ROOT(&cfqq
->sort_list
))
1769 cfq_del_cfqq_rr(cfqd
, cfqq
);
1771 cfq_resort_rr_list(cfqd
, cfqq
);
1773 if (cfqq
== cfqd
->active_queue
)
1774 cfqd
->active_queue
= NULL
;
1776 if (cfqd
->active_cic
) {
1777 put_io_context(cfqd
->active_cic
->ioc
, cfqd
->queue
);
1778 cfqd
->active_cic
= NULL
;
1782 static inline void cfq_slice_expired(struct cfq_data
*cfqd
, bool timed_out
)
1784 struct cfq_queue
*cfqq
= cfqd
->active_queue
;
1787 __cfq_slice_expired(cfqd
, cfqq
, timed_out
);
1791 * Get next queue for service. Unless we have a queue preemption,
1792 * we'll simply select the first cfqq in the service tree.
1794 static struct cfq_queue
*cfq_get_next_queue(struct cfq_data
*cfqd
)
1796 struct cfq_rb_root
*service_tree
=
1797 service_tree_for(cfqd
->serving_group
, cfqd
->serving_prio
,
1798 cfqd
->serving_type
);
1800 if (!cfqd
->rq_queued
)
1803 /* There is nothing to dispatch */
1806 if (RB_EMPTY_ROOT(&service_tree
->rb
))
1808 return cfq_rb_first(service_tree
);
1811 static struct cfq_queue
*cfq_get_next_queue_forced(struct cfq_data
*cfqd
)
1813 struct cfq_group
*cfqg
;
1814 struct cfq_queue
*cfqq
;
1816 struct cfq_rb_root
*st
;
1818 if (!cfqd
->rq_queued
)
1821 cfqg
= cfq_get_next_cfqg(cfqd
);
1825 for_each_cfqg_st(cfqg
, i
, j
, st
)
1826 if ((cfqq
= cfq_rb_first(st
)) != NULL
)
1832 * Get and set a new active queue for service.
1834 static struct cfq_queue
*cfq_set_active_queue(struct cfq_data
*cfqd
,
1835 struct cfq_queue
*cfqq
)
1838 cfqq
= cfq_get_next_queue(cfqd
);
1840 __cfq_set_active_queue(cfqd
, cfqq
);
1844 static inline sector_t
cfq_dist_from_last(struct cfq_data
*cfqd
,
1847 if (blk_rq_pos(rq
) >= cfqd
->last_position
)
1848 return blk_rq_pos(rq
) - cfqd
->last_position
;
1850 return cfqd
->last_position
- blk_rq_pos(rq
);
1853 static inline int cfq_rq_close(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
1856 return cfq_dist_from_last(cfqd
, rq
) <= CFQQ_CLOSE_THR
;
1859 static struct cfq_queue
*cfqq_close(struct cfq_data
*cfqd
,
1860 struct cfq_queue
*cur_cfqq
)
1862 struct rb_root
*root
= &cfqd
->prio_trees
[cur_cfqq
->org_ioprio
];
1863 struct rb_node
*parent
, *node
;
1864 struct cfq_queue
*__cfqq
;
1865 sector_t sector
= cfqd
->last_position
;
1867 if (RB_EMPTY_ROOT(root
))
1871 * First, if we find a request starting at the end of the last
1872 * request, choose it.
1874 __cfqq
= cfq_prio_tree_lookup(cfqd
, root
, sector
, &parent
, NULL
);
1879 * If the exact sector wasn't found, the parent of the NULL leaf
1880 * will contain the closest sector.
1882 __cfqq
= rb_entry(parent
, struct cfq_queue
, p_node
);
1883 if (cfq_rq_close(cfqd
, cur_cfqq
, __cfqq
->next_rq
))
1886 if (blk_rq_pos(__cfqq
->next_rq
) < sector
)
1887 node
= rb_next(&__cfqq
->p_node
);
1889 node
= rb_prev(&__cfqq
->p_node
);
1893 __cfqq
= rb_entry(node
, struct cfq_queue
, p_node
);
1894 if (cfq_rq_close(cfqd
, cur_cfqq
, __cfqq
->next_rq
))
1902 * cur_cfqq - passed in so that we don't decide that the current queue is
1903 * closely cooperating with itself.
1905 * So, basically we're assuming that that cur_cfqq has dispatched at least
1906 * one request, and that cfqd->last_position reflects a position on the disk
1907 * associated with the I/O issued by cur_cfqq. I'm not sure this is a valid
1910 static struct cfq_queue
*cfq_close_cooperator(struct cfq_data
*cfqd
,
1911 struct cfq_queue
*cur_cfqq
)
1913 struct cfq_queue
*cfqq
;
1915 if (cfq_class_idle(cur_cfqq
))
1917 if (!cfq_cfqq_sync(cur_cfqq
))
1919 if (CFQQ_SEEKY(cur_cfqq
))
1923 * Don't search priority tree if it's the only queue in the group.
1925 if (cur_cfqq
->cfqg
->nr_cfqq
== 1)
1929 * We should notice if some of the queues are cooperating, eg
1930 * working closely on the same area of the disk. In that case,
1931 * we can group them together and don't waste time idling.
1933 cfqq
= cfqq_close(cfqd
, cur_cfqq
);
1937 /* If new queue belongs to different cfq_group, don't choose it */
1938 if (cur_cfqq
->cfqg
!= cfqq
->cfqg
)
1942 * It only makes sense to merge sync queues.
1944 if (!cfq_cfqq_sync(cfqq
))
1946 if (CFQQ_SEEKY(cfqq
))
1950 * Do not merge queues of different priority classes
1952 if (cfq_class_rt(cfqq
) != cfq_class_rt(cur_cfqq
))
1959 * Determine whether we should enforce idle window for this queue.
1962 static bool cfq_should_idle(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
1964 enum wl_prio_t prio
= cfqq_prio(cfqq
);
1965 struct cfq_rb_root
*service_tree
= cfqq
->service_tree
;
1967 BUG_ON(!service_tree
);
1968 BUG_ON(!service_tree
->count
);
1970 if (!cfqd
->cfq_slice_idle
)
1973 /* We never do for idle class queues. */
1974 if (prio
== IDLE_WORKLOAD
)
1977 /* We do for queues that were marked with idle window flag. */
1978 if (cfq_cfqq_idle_window(cfqq
) &&
1979 !(blk_queue_nonrot(cfqd
->queue
) && cfqd
->hw_tag
))
1983 * Otherwise, we do only if they are the last ones
1984 * in their service tree.
1986 if (service_tree
->count
== 1 && cfq_cfqq_sync(cfqq
) &&
1987 !cfq_io_thinktime_big(cfqd
, &service_tree
->ttime
, false))
1989 cfq_log_cfqq(cfqd
, cfqq
, "Not idling. st->count:%d",
1990 service_tree
->count
);
1994 static void cfq_arm_slice_timer(struct cfq_data
*cfqd
)
1996 struct cfq_queue
*cfqq
= cfqd
->active_queue
;
1997 struct cfq_io_context
*cic
;
1998 unsigned long sl
, group_idle
= 0;
2001 * SSD device without seek penalty, disable idling. But only do so
2002 * for devices that support queuing, otherwise we still have a problem
2003 * with sync vs async workloads.
2005 if (blk_queue_nonrot(cfqd
->queue
) && cfqd
->hw_tag
)
2008 WARN_ON(!RB_EMPTY_ROOT(&cfqq
->sort_list
));
2009 WARN_ON(cfq_cfqq_slice_new(cfqq
));
2012 * idle is disabled, either manually or by past process history
2014 if (!cfq_should_idle(cfqd
, cfqq
)) {
2015 /* no queue idling. Check for group idling */
2016 if (cfqd
->cfq_group_idle
)
2017 group_idle
= cfqd
->cfq_group_idle
;
2023 * still active requests from this queue, don't idle
2025 if (cfqq
->dispatched
)
2029 * task has exited, don't wait
2031 cic
= cfqd
->active_cic
;
2032 if (!cic
|| !atomic_read(&cic
->ioc
->nr_tasks
))
2036 * If our average think time is larger than the remaining time
2037 * slice, then don't idle. This avoids overrunning the allotted
2040 if (sample_valid(cic
->ttime
.ttime_samples
) &&
2041 (cfqq
->slice_end
- jiffies
< cic
->ttime
.ttime_mean
)) {
2042 cfq_log_cfqq(cfqd
, cfqq
, "Not idling. think_time:%lu",
2043 cic
->ttime
.ttime_mean
);
2047 /* There are other queues in the group, don't do group idle */
2048 if (group_idle
&& cfqq
->cfqg
->nr_cfqq
> 1)
2051 cfq_mark_cfqq_wait_request(cfqq
);
2054 sl
= cfqd
->cfq_group_idle
;
2056 sl
= cfqd
->cfq_slice_idle
;
2058 mod_timer(&cfqd
->idle_slice_timer
, jiffies
+ sl
);
2059 cfq_blkiocg_update_set_idle_time_stats(&cfqq
->cfqg
->blkg
);
2060 cfq_log_cfqq(cfqd
, cfqq
, "arm_idle: %lu group_idle: %d", sl
,
2061 group_idle
? 1 : 0);
2065 * Move request from internal lists to the request queue dispatch list.
2067 static void cfq_dispatch_insert(struct request_queue
*q
, struct request
*rq
)
2069 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
2070 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
2072 cfq_log_cfqq(cfqd
, cfqq
, "dispatch_insert");
2074 cfqq
->next_rq
= cfq_find_next_rq(cfqd
, cfqq
, rq
);
2075 cfq_remove_request(rq
);
2077 (RQ_CFQG(rq
))->dispatched
++;
2078 elv_dispatch_sort(q
, rq
);
2080 cfqd
->rq_in_flight
[cfq_cfqq_sync(cfqq
)]++;
2081 cfqq
->nr_sectors
+= blk_rq_sectors(rq
);
2082 cfq_blkiocg_update_dispatch_stats(&cfqq
->cfqg
->blkg
, blk_rq_bytes(rq
),
2083 rq_data_dir(rq
), rq_is_sync(rq
));
2087 * return expired entry, or NULL to just start from scratch in rbtree
2089 static struct request
*cfq_check_fifo(struct cfq_queue
*cfqq
)
2091 struct request
*rq
= NULL
;
2093 if (cfq_cfqq_fifo_expire(cfqq
))
2096 cfq_mark_cfqq_fifo_expire(cfqq
);
2098 if (list_empty(&cfqq
->fifo
))
2101 rq
= rq_entry_fifo(cfqq
->fifo
.next
);
2102 if (time_before(jiffies
, rq_fifo_time(rq
)))
2105 cfq_log_cfqq(cfqq
->cfqd
, cfqq
, "fifo=%p", rq
);
2110 cfq_prio_to_maxrq(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
2112 const int base_rq
= cfqd
->cfq_slice_async_rq
;
2114 WARN_ON(cfqq
->ioprio
>= IOPRIO_BE_NR
);
2116 return 2 * base_rq
* (IOPRIO_BE_NR
- cfqq
->ioprio
);
2120 * Must be called with the queue_lock held.
2122 static int cfqq_process_refs(struct cfq_queue
*cfqq
)
2124 int process_refs
, io_refs
;
2126 io_refs
= cfqq
->allocated
[READ
] + cfqq
->allocated
[WRITE
];
2127 process_refs
= cfqq
->ref
- io_refs
;
2128 BUG_ON(process_refs
< 0);
2129 return process_refs
;
2132 static void cfq_setup_merge(struct cfq_queue
*cfqq
, struct cfq_queue
*new_cfqq
)
2134 int process_refs
, new_process_refs
;
2135 struct cfq_queue
*__cfqq
;
2138 * If there are no process references on the new_cfqq, then it is
2139 * unsafe to follow the ->new_cfqq chain as other cfqq's in the
2140 * chain may have dropped their last reference (not just their
2141 * last process reference).
2143 if (!cfqq_process_refs(new_cfqq
))
2146 /* Avoid a circular list and skip interim queue merges */
2147 while ((__cfqq
= new_cfqq
->new_cfqq
)) {
2153 process_refs
= cfqq_process_refs(cfqq
);
2154 new_process_refs
= cfqq_process_refs(new_cfqq
);
2156 * If the process for the cfqq has gone away, there is no
2157 * sense in merging the queues.
2159 if (process_refs
== 0 || new_process_refs
== 0)
2163 * Merge in the direction of the lesser amount of work.
2165 if (new_process_refs
>= process_refs
) {
2166 cfqq
->new_cfqq
= new_cfqq
;
2167 new_cfqq
->ref
+= process_refs
;
2169 new_cfqq
->new_cfqq
= cfqq
;
2170 cfqq
->ref
+= new_process_refs
;
2174 static enum wl_type_t
cfq_choose_wl(struct cfq_data
*cfqd
,
2175 struct cfq_group
*cfqg
, enum wl_prio_t prio
)
2177 struct cfq_queue
*queue
;
2179 bool key_valid
= false;
2180 unsigned long lowest_key
= 0;
2181 enum wl_type_t cur_best
= SYNC_NOIDLE_WORKLOAD
;
2183 for (i
= 0; i
<= SYNC_WORKLOAD
; ++i
) {
2184 /* select the one with lowest rb_key */
2185 queue
= cfq_rb_first(service_tree_for(cfqg
, prio
, i
));
2187 (!key_valid
|| time_before(queue
->rb_key
, lowest_key
))) {
2188 lowest_key
= queue
->rb_key
;
2197 static void choose_service_tree(struct cfq_data
*cfqd
, struct cfq_group
*cfqg
)
2201 struct cfq_rb_root
*st
;
2202 unsigned group_slice
;
2203 enum wl_prio_t original_prio
= cfqd
->serving_prio
;
2205 /* Choose next priority. RT > BE > IDLE */
2206 if (cfq_group_busy_queues_wl(RT_WORKLOAD
, cfqd
, cfqg
))
2207 cfqd
->serving_prio
= RT_WORKLOAD
;
2208 else if (cfq_group_busy_queues_wl(BE_WORKLOAD
, cfqd
, cfqg
))
2209 cfqd
->serving_prio
= BE_WORKLOAD
;
2211 cfqd
->serving_prio
= IDLE_WORKLOAD
;
2212 cfqd
->workload_expires
= jiffies
+ 1;
2216 if (original_prio
!= cfqd
->serving_prio
)
2220 * For RT and BE, we have to choose also the type
2221 * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
2224 st
= service_tree_for(cfqg
, cfqd
->serving_prio
, cfqd
->serving_type
);
2228 * check workload expiration, and that we still have other queues ready
2230 if (count
&& !time_after(jiffies
, cfqd
->workload_expires
))
2234 /* otherwise select new workload type */
2235 cfqd
->serving_type
=
2236 cfq_choose_wl(cfqd
, cfqg
, cfqd
->serving_prio
);
2237 st
= service_tree_for(cfqg
, cfqd
->serving_prio
, cfqd
->serving_type
);
2241 * the workload slice is computed as a fraction of target latency
2242 * proportional to the number of queues in that workload, over
2243 * all the queues in the same priority class
2245 group_slice
= cfq_group_slice(cfqd
, cfqg
);
2247 slice
= group_slice
* count
/
2248 max_t(unsigned, cfqg
->busy_queues_avg
[cfqd
->serving_prio
],
2249 cfq_group_busy_queues_wl(cfqd
->serving_prio
, cfqd
, cfqg
));
2251 if (cfqd
->serving_type
== ASYNC_WORKLOAD
) {
2255 * Async queues are currently system wide. Just taking
2256 * proportion of queues with-in same group will lead to higher
2257 * async ratio system wide as generally root group is going
2258 * to have higher weight. A more accurate thing would be to
2259 * calculate system wide asnc/sync ratio.
2261 tmp
= cfq_target_latency
* cfqg_busy_async_queues(cfqd
, cfqg
);
2262 tmp
= tmp
/cfqd
->busy_queues
;
2263 slice
= min_t(unsigned, slice
, tmp
);
2265 /* async workload slice is scaled down according to
2266 * the sync/async slice ratio. */
2267 slice
= slice
* cfqd
->cfq_slice
[0] / cfqd
->cfq_slice
[1];
2269 /* sync workload slice is at least 2 * cfq_slice_idle */
2270 slice
= max(slice
, 2 * cfqd
->cfq_slice_idle
);
2272 slice
= max_t(unsigned, slice
, CFQ_MIN_TT
);
2273 cfq_log(cfqd
, "workload slice:%d", slice
);
2274 cfqd
->workload_expires
= jiffies
+ slice
;
2277 static struct cfq_group
*cfq_get_next_cfqg(struct cfq_data
*cfqd
)
2279 struct cfq_rb_root
*st
= &cfqd
->grp_service_tree
;
2280 struct cfq_group
*cfqg
;
2282 if (RB_EMPTY_ROOT(&st
->rb
))
2284 cfqg
= cfq_rb_first_group(st
);
2285 update_min_vdisktime(st
);
2289 static void cfq_choose_cfqg(struct cfq_data
*cfqd
)
2291 struct cfq_group
*cfqg
= cfq_get_next_cfqg(cfqd
);
2293 cfqd
->serving_group
= cfqg
;
2295 /* Restore the workload type data */
2296 if (cfqg
->saved_workload_slice
) {
2297 cfqd
->workload_expires
= jiffies
+ cfqg
->saved_workload_slice
;
2298 cfqd
->serving_type
= cfqg
->saved_workload
;
2299 cfqd
->serving_prio
= cfqg
->saved_serving_prio
;
2301 cfqd
->workload_expires
= jiffies
- 1;
2303 choose_service_tree(cfqd
, cfqg
);
2307 * Select a queue for service. If we have a current active queue,
2308 * check whether to continue servicing it, or retrieve and set a new one.
2310 static struct cfq_queue
*cfq_select_queue(struct cfq_data
*cfqd
)
2312 struct cfq_queue
*cfqq
, *new_cfqq
= NULL
;
2314 cfqq
= cfqd
->active_queue
;
2318 if (!cfqd
->rq_queued
)
2322 * We were waiting for group to get backlogged. Expire the queue
2324 if (cfq_cfqq_wait_busy(cfqq
) && !RB_EMPTY_ROOT(&cfqq
->sort_list
))
2328 * The active queue has run out of time, expire it and select new.
2330 if (cfq_slice_used(cfqq
) && !cfq_cfqq_must_dispatch(cfqq
)) {
2332 * If slice had not expired at the completion of last request
2333 * we might not have turned on wait_busy flag. Don't expire
2334 * the queue yet. Allow the group to get backlogged.
2336 * The very fact that we have used the slice, that means we
2337 * have been idling all along on this queue and it should be
2338 * ok to wait for this request to complete.
2340 if (cfqq
->cfqg
->nr_cfqq
== 1 && RB_EMPTY_ROOT(&cfqq
->sort_list
)
2341 && cfqq
->dispatched
&& cfq_should_idle(cfqd
, cfqq
)) {
2345 goto check_group_idle
;
2349 * The active queue has requests and isn't expired, allow it to
2352 if (!RB_EMPTY_ROOT(&cfqq
->sort_list
))
2356 * If another queue has a request waiting within our mean seek
2357 * distance, let it run. The expire code will check for close
2358 * cooperators and put the close queue at the front of the service
2359 * tree. If possible, merge the expiring queue with the new cfqq.
2361 new_cfqq
= cfq_close_cooperator(cfqd
, cfqq
);
2363 if (!cfqq
->new_cfqq
)
2364 cfq_setup_merge(cfqq
, new_cfqq
);
2369 * No requests pending. If the active queue still has requests in
2370 * flight or is idling for a new request, allow either of these
2371 * conditions to happen (or time out) before selecting a new queue.
2373 if (timer_pending(&cfqd
->idle_slice_timer
)) {
2379 * This is a deep seek queue, but the device is much faster than
2380 * the queue can deliver, don't idle
2382 if (CFQQ_SEEKY(cfqq
) && cfq_cfqq_idle_window(cfqq
) &&
2383 (cfq_cfqq_slice_new(cfqq
) ||
2384 (cfqq
->slice_end
- jiffies
> jiffies
- cfqq
->slice_start
))) {
2385 cfq_clear_cfqq_deep(cfqq
);
2386 cfq_clear_cfqq_idle_window(cfqq
);
2389 if (cfqq
->dispatched
&& cfq_should_idle(cfqd
, cfqq
)) {
2395 * If group idle is enabled and there are requests dispatched from
2396 * this group, wait for requests to complete.
2399 if (cfqd
->cfq_group_idle
&& cfqq
->cfqg
->nr_cfqq
== 1 &&
2400 cfqq
->cfqg
->dispatched
&&
2401 !cfq_io_thinktime_big(cfqd
, &cfqq
->cfqg
->ttime
, true)) {
2407 cfq_slice_expired(cfqd
, 0);
2410 * Current queue expired. Check if we have to switch to a new
2414 cfq_choose_cfqg(cfqd
);
2416 cfqq
= cfq_set_active_queue(cfqd
, new_cfqq
);
2421 static int __cfq_forced_dispatch_cfqq(struct cfq_queue
*cfqq
)
2425 while (cfqq
->next_rq
) {
2426 cfq_dispatch_insert(cfqq
->cfqd
->queue
, cfqq
->next_rq
);
2430 BUG_ON(!list_empty(&cfqq
->fifo
));
2432 /* By default cfqq is not expired if it is empty. Do it explicitly */
2433 __cfq_slice_expired(cfqq
->cfqd
, cfqq
, 0);
2438 * Drain our current requests. Used for barriers and when switching
2439 * io schedulers on-the-fly.
2441 static int cfq_forced_dispatch(struct cfq_data
*cfqd
)
2443 struct cfq_queue
*cfqq
;
2446 /* Expire the timeslice of the current active queue first */
2447 cfq_slice_expired(cfqd
, 0);
2448 while ((cfqq
= cfq_get_next_queue_forced(cfqd
)) != NULL
) {
2449 __cfq_set_active_queue(cfqd
, cfqq
);
2450 dispatched
+= __cfq_forced_dispatch_cfqq(cfqq
);
2453 BUG_ON(cfqd
->busy_queues
);
2455 cfq_log(cfqd
, "forced_dispatch=%d", dispatched
);
2459 static inline bool cfq_slice_used_soon(struct cfq_data
*cfqd
,
2460 struct cfq_queue
*cfqq
)
2462 /* the queue hasn't finished any request, can't estimate */
2463 if (cfq_cfqq_slice_new(cfqq
))
2465 if (time_after(jiffies
+ cfqd
->cfq_slice_idle
* cfqq
->dispatched
,
2472 static bool cfq_may_dispatch(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
2474 unsigned int max_dispatch
;
2477 * Drain async requests before we start sync IO
2479 if (cfq_should_idle(cfqd
, cfqq
) && cfqd
->rq_in_flight
[BLK_RW_ASYNC
])
2483 * If this is an async queue and we have sync IO in flight, let it wait
2485 if (cfqd
->rq_in_flight
[BLK_RW_SYNC
] && !cfq_cfqq_sync(cfqq
))
2488 max_dispatch
= max_t(unsigned int, cfqd
->cfq_quantum
/ 2, 1);
2489 if (cfq_class_idle(cfqq
))
2493 * Does this cfqq already have too much IO in flight?
2495 if (cfqq
->dispatched
>= max_dispatch
) {
2496 bool promote_sync
= false;
2498 * idle queue must always only have a single IO in flight
2500 if (cfq_class_idle(cfqq
))
2504 * If there is only one sync queue
2505 * we can ignore async queue here and give the sync
2506 * queue no dispatch limit. The reason is a sync queue can
2507 * preempt async queue, limiting the sync queue doesn't make
2508 * sense. This is useful for aiostress test.
2510 if (cfq_cfqq_sync(cfqq
) && cfqd
->busy_sync_queues
== 1)
2511 promote_sync
= true;
2514 * We have other queues, don't allow more IO from this one
2516 if (cfqd
->busy_queues
> 1 && cfq_slice_used_soon(cfqd
, cfqq
) &&
2521 * Sole queue user, no limit
2523 if (cfqd
->busy_queues
== 1 || promote_sync
)
2527 * Normally we start throttling cfqq when cfq_quantum/2
2528 * requests have been dispatched. But we can drive
2529 * deeper queue depths at the beginning of slice
2530 * subjected to upper limit of cfq_quantum.
2532 max_dispatch
= cfqd
->cfq_quantum
;
2536 * Async queues must wait a bit before being allowed dispatch.
2537 * We also ramp up the dispatch depth gradually for async IO,
2538 * based on the last sync IO we serviced
2540 if (!cfq_cfqq_sync(cfqq
) && cfqd
->cfq_latency
) {
2541 unsigned long last_sync
= jiffies
- cfqd
->last_delayed_sync
;
2544 depth
= last_sync
/ cfqd
->cfq_slice
[1];
2545 if (!depth
&& !cfqq
->dispatched
)
2547 if (depth
< max_dispatch
)
2548 max_dispatch
= depth
;
2552 * If we're below the current max, allow a dispatch
2554 return cfqq
->dispatched
< max_dispatch
;
2558 * Dispatch a request from cfqq, moving them to the request queue
2561 static bool cfq_dispatch_request(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
2565 BUG_ON(RB_EMPTY_ROOT(&cfqq
->sort_list
));
2567 if (!cfq_may_dispatch(cfqd
, cfqq
))
2571 * follow expired path, else get first next available
2573 rq
= cfq_check_fifo(cfqq
);
2578 * insert request into driver dispatch list
2580 cfq_dispatch_insert(cfqd
->queue
, rq
);
2582 if (!cfqd
->active_cic
) {
2583 struct cfq_io_context
*cic
= RQ_CIC(rq
);
2585 atomic_long_inc(&cic
->ioc
->refcount
);
2586 cfqd
->active_cic
= cic
;
2593 * Find the cfqq that we need to service and move a request from that to the
2596 static int cfq_dispatch_requests(struct request_queue
*q
, int force
)
2598 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
2599 struct cfq_queue
*cfqq
;
2601 if (!cfqd
->busy_queues
)
2604 if (unlikely(force
))
2605 return cfq_forced_dispatch(cfqd
);
2607 cfqq
= cfq_select_queue(cfqd
);
2612 * Dispatch a request from this cfqq, if it is allowed
2614 if (!cfq_dispatch_request(cfqd
, cfqq
))
2617 cfqq
->slice_dispatch
++;
2618 cfq_clear_cfqq_must_dispatch(cfqq
);
2621 * expire an async queue immediately if it has used up its slice. idle
2622 * queue always expire after 1 dispatch round.
2624 if (cfqd
->busy_queues
> 1 && ((!cfq_cfqq_sync(cfqq
) &&
2625 cfqq
->slice_dispatch
>= cfq_prio_to_maxrq(cfqd
, cfqq
)) ||
2626 cfq_class_idle(cfqq
))) {
2627 cfqq
->slice_end
= jiffies
+ 1;
2628 cfq_slice_expired(cfqd
, 0);
2631 cfq_log_cfqq(cfqd
, cfqq
, "dispatched a request");
2636 * task holds one reference to the queue, dropped when task exits. each rq
2637 * in-flight on this queue also holds a reference, dropped when rq is freed.
2639 * Each cfq queue took a reference on the parent group. Drop it now.
2640 * queue lock must be held here.
2642 static void cfq_put_queue(struct cfq_queue
*cfqq
)
2644 struct cfq_data
*cfqd
= cfqq
->cfqd
;
2645 struct cfq_group
*cfqg
;
2647 BUG_ON(cfqq
->ref
<= 0);
2653 cfq_log_cfqq(cfqd
, cfqq
, "put_queue");
2654 BUG_ON(rb_first(&cfqq
->sort_list
));
2655 BUG_ON(cfqq
->allocated
[READ
] + cfqq
->allocated
[WRITE
]);
2658 if (unlikely(cfqd
->active_queue
== cfqq
)) {
2659 __cfq_slice_expired(cfqd
, cfqq
, 0);
2660 cfq_schedule_dispatch(cfqd
);
2663 BUG_ON(cfq_cfqq_on_rr(cfqq
));
2664 kmem_cache_free(cfq_pool
, cfqq
);
2668 static void cfq_cic_free_rcu(struct rcu_head
*head
)
2670 kmem_cache_free(cfq_ioc_pool
,
2671 container_of(head
, struct cfq_io_context
, rcu_head
));
2674 static void cfq_cic_free(struct cfq_io_context
*cic
)
2676 call_rcu(&cic
->rcu_head
, cfq_cic_free_rcu
);
2679 static void cfq_release_cic(struct cfq_io_context
*cic
)
2681 struct io_context
*ioc
= cic
->ioc
;
2682 unsigned long dead_key
= (unsigned long) cic
->key
;
2684 BUG_ON(!(dead_key
& CIC_DEAD_KEY
));
2685 radix_tree_delete(&ioc
->radix_root
, dead_key
>> CIC_DEAD_INDEX_SHIFT
);
2686 hlist_del(&cic
->cic_list
);
2690 static void cfq_put_cooperator(struct cfq_queue
*cfqq
)
2692 struct cfq_queue
*__cfqq
, *next
;
2695 * If this queue was scheduled to merge with another queue, be
2696 * sure to drop the reference taken on that queue (and others in
2697 * the merge chain). See cfq_setup_merge and cfq_merge_cfqqs.
2699 __cfqq
= cfqq
->new_cfqq
;
2701 if (__cfqq
== cfqq
) {
2702 WARN(1, "cfqq->new_cfqq loop detected\n");
2705 next
= __cfqq
->new_cfqq
;
2706 cfq_put_queue(__cfqq
);
2711 static void cfq_exit_cfqq(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
2713 if (unlikely(cfqq
== cfqd
->active_queue
)) {
2714 __cfq_slice_expired(cfqd
, cfqq
, 0);
2715 cfq_schedule_dispatch(cfqd
);
2718 cfq_put_cooperator(cfqq
);
2720 cfq_put_queue(cfqq
);
2723 static void cfq_exit_cic(struct cfq_io_context
*cic
)
2725 struct cfq_data
*cfqd
= cic_to_cfqd(cic
);
2726 struct io_context
*ioc
= cic
->ioc
;
2728 list_del_init(&cic
->queue_list
);
2729 cic
->key
= cfqd_dead_key(cfqd
);
2732 * Both setting lookup hint to and clearing it from @cic are done
2733 * under queue_lock. If it's not pointing to @cic now, it never
2734 * will. Hint assignment itself can race safely.
2736 if (rcu_dereference_raw(ioc
->ioc_data
) == cic
)
2737 rcu_assign_pointer(ioc
->ioc_data
, NULL
);
2739 if (cic
->cfqq
[BLK_RW_ASYNC
]) {
2740 cfq_exit_cfqq(cfqd
, cic
->cfqq
[BLK_RW_ASYNC
]);
2741 cic
->cfqq
[BLK_RW_ASYNC
] = NULL
;
2744 if (cic
->cfqq
[BLK_RW_SYNC
]) {
2745 cfq_exit_cfqq(cfqd
, cic
->cfqq
[BLK_RW_SYNC
]);
2746 cic
->cfqq
[BLK_RW_SYNC
] = NULL
;
2750 static struct cfq_io_context
*
2751 cfq_alloc_io_context(struct cfq_data
*cfqd
, gfp_t gfp_mask
)
2753 struct cfq_io_context
*cic
;
2755 cic
= kmem_cache_alloc_node(cfq_ioc_pool
, gfp_mask
| __GFP_ZERO
,
2758 cic
->ttime
.last_end_request
= jiffies
;
2759 INIT_LIST_HEAD(&cic
->queue_list
);
2760 INIT_HLIST_NODE(&cic
->cic_list
);
2761 cic
->exit
= cfq_exit_cic
;
2762 cic
->release
= cfq_release_cic
;
2768 static void cfq_init_prio_data(struct cfq_queue
*cfqq
, struct io_context
*ioc
)
2770 struct task_struct
*tsk
= current
;
2773 if (!cfq_cfqq_prio_changed(cfqq
))
2776 ioprio_class
= IOPRIO_PRIO_CLASS(ioc
->ioprio
);
2777 switch (ioprio_class
) {
2779 printk(KERN_ERR
"cfq: bad prio %x\n", ioprio_class
);
2780 case IOPRIO_CLASS_NONE
:
2782 * no prio set, inherit CPU scheduling settings
2784 cfqq
->ioprio
= task_nice_ioprio(tsk
);
2785 cfqq
->ioprio_class
= task_nice_ioclass(tsk
);
2787 case IOPRIO_CLASS_RT
:
2788 cfqq
->ioprio
= task_ioprio(ioc
);
2789 cfqq
->ioprio_class
= IOPRIO_CLASS_RT
;
2791 case IOPRIO_CLASS_BE
:
2792 cfqq
->ioprio
= task_ioprio(ioc
);
2793 cfqq
->ioprio_class
= IOPRIO_CLASS_BE
;
2795 case IOPRIO_CLASS_IDLE
:
2796 cfqq
->ioprio_class
= IOPRIO_CLASS_IDLE
;
2798 cfq_clear_cfqq_idle_window(cfqq
);
2803 * keep track of original prio settings in case we have to temporarily
2804 * elevate the priority of this queue
2806 cfqq
->org_ioprio
= cfqq
->ioprio
;
2807 cfq_clear_cfqq_prio_changed(cfqq
);
2810 static void changed_ioprio(struct cfq_io_context
*cic
)
2812 struct cfq_data
*cfqd
= cic_to_cfqd(cic
);
2813 struct cfq_queue
*cfqq
;
2815 if (unlikely(!cfqd
))
2818 cfqq
= cic
->cfqq
[BLK_RW_ASYNC
];
2820 struct cfq_queue
*new_cfqq
;
2821 new_cfqq
= cfq_get_queue(cfqd
, BLK_RW_ASYNC
, cic
->ioc
,
2824 cic
->cfqq
[BLK_RW_ASYNC
] = new_cfqq
;
2825 cfq_put_queue(cfqq
);
2829 cfqq
= cic
->cfqq
[BLK_RW_SYNC
];
2831 cfq_mark_cfqq_prio_changed(cfqq
);
2834 static void cfq_init_cfqq(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
2835 pid_t pid
, bool is_sync
)
2837 RB_CLEAR_NODE(&cfqq
->rb_node
);
2838 RB_CLEAR_NODE(&cfqq
->p_node
);
2839 INIT_LIST_HEAD(&cfqq
->fifo
);
2844 cfq_mark_cfqq_prio_changed(cfqq
);
2847 if (!cfq_class_idle(cfqq
))
2848 cfq_mark_cfqq_idle_window(cfqq
);
2849 cfq_mark_cfqq_sync(cfqq
);
2854 #ifdef CONFIG_CFQ_GROUP_IOSCHED
2855 static void changed_cgroup(struct cfq_io_context
*cic
)
2857 struct cfq_queue
*sync_cfqq
= cic_to_cfqq(cic
, 1);
2858 struct cfq_data
*cfqd
= cic_to_cfqd(cic
);
2859 struct request_queue
*q
;
2861 if (unlikely(!cfqd
))
2868 * Drop reference to sync queue. A new sync queue will be
2869 * assigned in new group upon arrival of a fresh request.
2871 cfq_log_cfqq(cfqd
, sync_cfqq
, "changed cgroup");
2872 cic_set_cfqq(cic
, NULL
, 1);
2873 cfq_put_queue(sync_cfqq
);
2876 #endif /* CONFIG_CFQ_GROUP_IOSCHED */
2878 static struct cfq_queue
*
2879 cfq_find_alloc_queue(struct cfq_data
*cfqd
, bool is_sync
,
2880 struct io_context
*ioc
, gfp_t gfp_mask
)
2882 struct cfq_queue
*cfqq
, *new_cfqq
= NULL
;
2883 struct cfq_io_context
*cic
;
2884 struct cfq_group
*cfqg
;
2887 cfqg
= cfq_get_cfqg(cfqd
);
2888 cic
= cfq_cic_lookup(cfqd
, ioc
);
2889 /* cic always exists here */
2890 cfqq
= cic_to_cfqq(cic
, is_sync
);
2893 * Always try a new alloc if we fell back to the OOM cfqq
2894 * originally, since it should just be a temporary situation.
2896 if (!cfqq
|| cfqq
== &cfqd
->oom_cfqq
) {
2901 } else if (gfp_mask
& __GFP_WAIT
) {
2902 spin_unlock_irq(cfqd
->queue
->queue_lock
);
2903 new_cfqq
= kmem_cache_alloc_node(cfq_pool
,
2904 gfp_mask
| __GFP_ZERO
,
2906 spin_lock_irq(cfqd
->queue
->queue_lock
);
2910 cfqq
= kmem_cache_alloc_node(cfq_pool
,
2911 gfp_mask
| __GFP_ZERO
,
2916 cfq_init_cfqq(cfqd
, cfqq
, current
->pid
, is_sync
);
2917 cfq_init_prio_data(cfqq
, ioc
);
2918 cfq_link_cfqq_cfqg(cfqq
, cfqg
);
2919 cfq_log_cfqq(cfqd
, cfqq
, "alloced");
2921 cfqq
= &cfqd
->oom_cfqq
;
2925 kmem_cache_free(cfq_pool
, new_cfqq
);
2930 static struct cfq_queue
**
2931 cfq_async_queue_prio(struct cfq_data
*cfqd
, int ioprio_class
, int ioprio
)
2933 switch (ioprio_class
) {
2934 case IOPRIO_CLASS_RT
:
2935 return &cfqd
->async_cfqq
[0][ioprio
];
2936 case IOPRIO_CLASS_BE
:
2937 return &cfqd
->async_cfqq
[1][ioprio
];
2938 case IOPRIO_CLASS_IDLE
:
2939 return &cfqd
->async_idle_cfqq
;
2945 static struct cfq_queue
*
2946 cfq_get_queue(struct cfq_data
*cfqd
, bool is_sync
, struct io_context
*ioc
,
2949 const int ioprio
= task_ioprio(ioc
);
2950 const int ioprio_class
= task_ioprio_class(ioc
);
2951 struct cfq_queue
**async_cfqq
= NULL
;
2952 struct cfq_queue
*cfqq
= NULL
;
2955 async_cfqq
= cfq_async_queue_prio(cfqd
, ioprio_class
, ioprio
);
2960 cfqq
= cfq_find_alloc_queue(cfqd
, is_sync
, ioc
, gfp_mask
);
2963 * pin the queue now that it's allocated, scheduler exit will prune it
2965 if (!is_sync
&& !(*async_cfqq
)) {
2975 * cfq_cic_lookup - lookup cfq_io_context
2976 * @cfqd: the associated cfq_data
2977 * @ioc: the associated io_context
2979 * Look up cfq_io_context associated with @cfqd - @ioc pair. Must be
2980 * called with queue_lock held.
2982 static struct cfq_io_context
*
2983 cfq_cic_lookup(struct cfq_data
*cfqd
, struct io_context
*ioc
)
2985 struct cfq_io_context
*cic
;
2987 lockdep_assert_held(cfqd
->queue
->queue_lock
);
2992 * cic's are indexed from @ioc using radix tree and hint pointer,
2993 * both of which are protected with RCU. All removals are done
2994 * holding both q and ioc locks, and we're holding q lock - if we
2995 * find a cic which points to us, it's guaranteed to be valid.
2998 cic
= rcu_dereference(ioc
->ioc_data
);
2999 if (cic
&& cic
->key
== cfqd
)
3002 cic
= radix_tree_lookup(&ioc
->radix_root
, cfqd
->queue
->id
);
3003 if (cic
&& cic
->key
== cfqd
)
3004 rcu_assign_pointer(ioc
->ioc_data
, cic
); /* allowed to race */
3013 * cfq_create_cic - create and link a cfq_io_context
3014 * @cfqd: cfqd of interest
3015 * @gfp_mask: allocation mask
3017 * Make sure cfq_io_context linking %current->io_context and @cfqd exists.
3018 * If ioc and/or cic doesn't exist, they will be created using @gfp_mask.
3020 static int cfq_create_cic(struct cfq_data
*cfqd
, gfp_t gfp_mask
)
3022 struct request_queue
*q
= cfqd
->queue
;
3023 struct cfq_io_context
*cic
= NULL
;
3024 struct io_context
*ioc
;
3027 might_sleep_if(gfp_mask
& __GFP_WAIT
);
3029 /* allocate stuff */
3030 ioc
= current_io_context(gfp_mask
, q
->node
);
3034 cic
= cfq_alloc_io_context(cfqd
, gfp_mask
);
3038 ret
= radix_tree_preload(gfp_mask
);
3044 cic
->q
= cfqd
->queue
;
3046 /* lock both q and ioc and try to link @cic */
3047 spin_lock_irq(q
->queue_lock
);
3048 spin_lock(&ioc
->lock
);
3050 ret
= radix_tree_insert(&ioc
->radix_root
, q
->id
, cic
);
3052 hlist_add_head(&cic
->cic_list
, &ioc
->cic_list
);
3053 list_add(&cic
->queue_list
, &cfqd
->cic_list
);
3055 } else if (ret
== -EEXIST
) {
3056 /* someone else already did it */
3060 spin_unlock(&ioc
->lock
);
3061 spin_unlock_irq(q
->queue_lock
);
3063 radix_tree_preload_end();
3066 printk(KERN_ERR
"cfq: cic link failed!\n");
3073 * cfq_get_io_context - acquire cfq_io_context and bump refcnt on io_context
3074 * @cfqd: cfqd to setup cic for
3075 * @gfp_mask: allocation mask
3077 * Return cfq_io_context associating @cfqd and %current->io_context and
3078 * bump refcnt on io_context. If ioc or cic doesn't exist, they're created
3081 * Must be called under queue_lock which may be released and re-acquired.
3082 * This function also may sleep depending on @gfp_mask.
3084 static struct cfq_io_context
*
3085 cfq_get_io_context(struct cfq_data
*cfqd
, gfp_t gfp_mask
)
3087 struct request_queue
*q
= cfqd
->queue
;
3088 struct cfq_io_context
*cic
= NULL
;
3089 struct io_context
*ioc
;
3092 lockdep_assert_held(q
->queue_lock
);
3096 ioc
= current
->io_context
;
3098 cic
= cfq_cic_lookup(cfqd
, ioc
);
3103 /* slow path - unlock, create missing ones and retry */
3104 spin_unlock_irq(q
->queue_lock
);
3105 err
= cfq_create_cic(cfqd
, gfp_mask
);
3106 spin_lock_irq(q
->queue_lock
);
3111 /* bump @ioc's refcnt and handle changed notifications */
3112 get_io_context(ioc
);
3114 if (unlikely(cic
->changed
)) {
3115 if (test_and_clear_bit(CIC_IOPRIO_CHANGED
, &cic
->changed
))
3116 changed_ioprio(cic
);
3117 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3118 if (test_and_clear_bit(CIC_CGROUP_CHANGED
, &cic
->changed
))
3119 changed_cgroup(cic
);
3127 __cfq_update_io_thinktime(struct cfq_ttime
*ttime
, unsigned long slice_idle
)
3129 unsigned long elapsed
= jiffies
- ttime
->last_end_request
;
3130 elapsed
= min(elapsed
, 2UL * slice_idle
);
3132 ttime
->ttime_samples
= (7*ttime
->ttime_samples
+ 256) / 8;
3133 ttime
->ttime_total
= (7*ttime
->ttime_total
+ 256*elapsed
) / 8;
3134 ttime
->ttime_mean
= (ttime
->ttime_total
+ 128) / ttime
->ttime_samples
;
3138 cfq_update_io_thinktime(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
3139 struct cfq_io_context
*cic
)
3141 if (cfq_cfqq_sync(cfqq
)) {
3142 __cfq_update_io_thinktime(&cic
->ttime
, cfqd
->cfq_slice_idle
);
3143 __cfq_update_io_thinktime(&cfqq
->service_tree
->ttime
,
3144 cfqd
->cfq_slice_idle
);
3146 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3147 __cfq_update_io_thinktime(&cfqq
->cfqg
->ttime
, cfqd
->cfq_group_idle
);
3152 cfq_update_io_seektime(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
3156 sector_t n_sec
= blk_rq_sectors(rq
);
3157 if (cfqq
->last_request_pos
) {
3158 if (cfqq
->last_request_pos
< blk_rq_pos(rq
))
3159 sdist
= blk_rq_pos(rq
) - cfqq
->last_request_pos
;
3161 sdist
= cfqq
->last_request_pos
- blk_rq_pos(rq
);
3164 cfqq
->seek_history
<<= 1;
3165 if (blk_queue_nonrot(cfqd
->queue
))
3166 cfqq
->seek_history
|= (n_sec
< CFQQ_SECT_THR_NONROT
);
3168 cfqq
->seek_history
|= (sdist
> CFQQ_SEEK_THR
);
3172 * Disable idle window if the process thinks too long or seeks so much that
3176 cfq_update_idle_window(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
3177 struct cfq_io_context
*cic
)
3179 int old_idle
, enable_idle
;
3182 * Don't idle for async or idle io prio class
3184 if (!cfq_cfqq_sync(cfqq
) || cfq_class_idle(cfqq
))
3187 enable_idle
= old_idle
= cfq_cfqq_idle_window(cfqq
);
3189 if (cfqq
->queued
[0] + cfqq
->queued
[1] >= 4)
3190 cfq_mark_cfqq_deep(cfqq
);
3192 if (cfqq
->next_rq
&& (cfqq
->next_rq
->cmd_flags
& REQ_NOIDLE
))
3194 else if (!atomic_read(&cic
->ioc
->nr_tasks
) || !cfqd
->cfq_slice_idle
||
3195 (!cfq_cfqq_deep(cfqq
) && CFQQ_SEEKY(cfqq
)))
3197 else if (sample_valid(cic
->ttime
.ttime_samples
)) {
3198 if (cic
->ttime
.ttime_mean
> cfqd
->cfq_slice_idle
)
3204 if (old_idle
!= enable_idle
) {
3205 cfq_log_cfqq(cfqd
, cfqq
, "idle=%d", enable_idle
);
3207 cfq_mark_cfqq_idle_window(cfqq
);
3209 cfq_clear_cfqq_idle_window(cfqq
);
3214 * Check if new_cfqq should preempt the currently active queue. Return 0 for
3215 * no or if we aren't sure, a 1 will cause a preempt.
3218 cfq_should_preempt(struct cfq_data
*cfqd
, struct cfq_queue
*new_cfqq
,
3221 struct cfq_queue
*cfqq
;
3223 cfqq
= cfqd
->active_queue
;
3227 if (cfq_class_idle(new_cfqq
))
3230 if (cfq_class_idle(cfqq
))
3234 * Don't allow a non-RT request to preempt an ongoing RT cfqq timeslice.
3236 if (cfq_class_rt(cfqq
) && !cfq_class_rt(new_cfqq
))
3240 * if the new request is sync, but the currently running queue is
3241 * not, let the sync request have priority.
3243 if (rq_is_sync(rq
) && !cfq_cfqq_sync(cfqq
))
3246 if (new_cfqq
->cfqg
!= cfqq
->cfqg
)
3249 if (cfq_slice_used(cfqq
))
3252 /* Allow preemption only if we are idling on sync-noidle tree */
3253 if (cfqd
->serving_type
== SYNC_NOIDLE_WORKLOAD
&&
3254 cfqq_type(new_cfqq
) == SYNC_NOIDLE_WORKLOAD
&&
3255 new_cfqq
->service_tree
->count
== 2 &&
3256 RB_EMPTY_ROOT(&cfqq
->sort_list
))
3260 * So both queues are sync. Let the new request get disk time if
3261 * it's a metadata request and the current queue is doing regular IO.
3263 if ((rq
->cmd_flags
& REQ_PRIO
) && !cfqq
->prio_pending
)
3267 * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
3269 if (cfq_class_rt(new_cfqq
) && !cfq_class_rt(cfqq
))
3272 /* An idle queue should not be idle now for some reason */
3273 if (RB_EMPTY_ROOT(&cfqq
->sort_list
) && !cfq_should_idle(cfqd
, cfqq
))
3276 if (!cfqd
->active_cic
|| !cfq_cfqq_wait_request(cfqq
))
3280 * if this request is as-good as one we would expect from the
3281 * current cfqq, let it preempt
3283 if (cfq_rq_close(cfqd
, cfqq
, rq
))
3290 * cfqq preempts the active queue. if we allowed preempt with no slice left,
3291 * let it have half of its nominal slice.
3293 static void cfq_preempt_queue(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
3295 struct cfq_queue
*old_cfqq
= cfqd
->active_queue
;
3297 cfq_log_cfqq(cfqd
, cfqq
, "preempt");
3298 cfq_slice_expired(cfqd
, 1);
3301 * workload type is changed, don't save slice, otherwise preempt
3304 if (cfqq_type(old_cfqq
) != cfqq_type(cfqq
))
3305 cfqq
->cfqg
->saved_workload_slice
= 0;
3308 * Put the new queue at the front of the of the current list,
3309 * so we know that it will be selected next.
3311 BUG_ON(!cfq_cfqq_on_rr(cfqq
));
3313 cfq_service_tree_add(cfqd
, cfqq
, 1);
3315 cfqq
->slice_end
= 0;
3316 cfq_mark_cfqq_slice_new(cfqq
);
3320 * Called when a new fs request (rq) is added (to cfqq). Check if there's
3321 * something we should do about it
3324 cfq_rq_enqueued(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
3327 struct cfq_io_context
*cic
= RQ_CIC(rq
);
3330 if (rq
->cmd_flags
& REQ_PRIO
)
3331 cfqq
->prio_pending
++;
3333 cfq_update_io_thinktime(cfqd
, cfqq
, cic
);
3334 cfq_update_io_seektime(cfqd
, cfqq
, rq
);
3335 cfq_update_idle_window(cfqd
, cfqq
, cic
);
3337 cfqq
->last_request_pos
= blk_rq_pos(rq
) + blk_rq_sectors(rq
);
3339 if (cfqq
== cfqd
->active_queue
) {
3341 * Remember that we saw a request from this process, but
3342 * don't start queuing just yet. Otherwise we risk seeing lots
3343 * of tiny requests, because we disrupt the normal plugging
3344 * and merging. If the request is already larger than a single
3345 * page, let it rip immediately. For that case we assume that
3346 * merging is already done. Ditto for a busy system that
3347 * has other work pending, don't risk delaying until the
3348 * idle timer unplug to continue working.
3350 if (cfq_cfqq_wait_request(cfqq
)) {
3351 if (blk_rq_bytes(rq
) > PAGE_CACHE_SIZE
||
3352 cfqd
->busy_queues
> 1) {
3353 cfq_del_timer(cfqd
, cfqq
);
3354 cfq_clear_cfqq_wait_request(cfqq
);
3355 __blk_run_queue(cfqd
->queue
);
3357 cfq_blkiocg_update_idle_time_stats(
3359 cfq_mark_cfqq_must_dispatch(cfqq
);
3362 } else if (cfq_should_preempt(cfqd
, cfqq
, rq
)) {
3364 * not the active queue - expire current slice if it is
3365 * idle and has expired it's mean thinktime or this new queue
3366 * has some old slice time left and is of higher priority or
3367 * this new queue is RT and the current one is BE
3369 cfq_preempt_queue(cfqd
, cfqq
);
3370 __blk_run_queue(cfqd
->queue
);
3374 static void cfq_insert_request(struct request_queue
*q
, struct request
*rq
)
3376 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
3377 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
3379 cfq_log_cfqq(cfqd
, cfqq
, "insert_request");
3380 cfq_init_prio_data(cfqq
, RQ_CIC(rq
)->ioc
);
3382 rq_set_fifo_time(rq
, jiffies
+ cfqd
->cfq_fifo_expire
[rq_is_sync(rq
)]);
3383 list_add_tail(&rq
->queuelist
, &cfqq
->fifo
);
3385 cfq_blkiocg_update_io_add_stats(&(RQ_CFQG(rq
))->blkg
,
3386 &cfqd
->serving_group
->blkg
, rq_data_dir(rq
),
3388 cfq_rq_enqueued(cfqd
, cfqq
, rq
);
3392 * Update hw_tag based on peak queue depth over 50 samples under
3395 static void cfq_update_hw_tag(struct cfq_data
*cfqd
)
3397 struct cfq_queue
*cfqq
= cfqd
->active_queue
;
3399 if (cfqd
->rq_in_driver
> cfqd
->hw_tag_est_depth
)
3400 cfqd
->hw_tag_est_depth
= cfqd
->rq_in_driver
;
3402 if (cfqd
->hw_tag
== 1)
3405 if (cfqd
->rq_queued
<= CFQ_HW_QUEUE_MIN
&&
3406 cfqd
->rq_in_driver
<= CFQ_HW_QUEUE_MIN
)
3410 * If active queue hasn't enough requests and can idle, cfq might not
3411 * dispatch sufficient requests to hardware. Don't zero hw_tag in this
3414 if (cfqq
&& cfq_cfqq_idle_window(cfqq
) &&
3415 cfqq
->dispatched
+ cfqq
->queued
[0] + cfqq
->queued
[1] <
3416 CFQ_HW_QUEUE_MIN
&& cfqd
->rq_in_driver
< CFQ_HW_QUEUE_MIN
)
3419 if (cfqd
->hw_tag_samples
++ < 50)
3422 if (cfqd
->hw_tag_est_depth
>= CFQ_HW_QUEUE_MIN
)
3428 static bool cfq_should_wait_busy(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
3430 struct cfq_io_context
*cic
= cfqd
->active_cic
;
3432 /* If the queue already has requests, don't wait */
3433 if (!RB_EMPTY_ROOT(&cfqq
->sort_list
))
3436 /* If there are other queues in the group, don't wait */
3437 if (cfqq
->cfqg
->nr_cfqq
> 1)
3440 /* the only queue in the group, but think time is big */
3441 if (cfq_io_thinktime_big(cfqd
, &cfqq
->cfqg
->ttime
, true))
3444 if (cfq_slice_used(cfqq
))
3447 /* if slice left is less than think time, wait busy */
3448 if (cic
&& sample_valid(cic
->ttime
.ttime_samples
)
3449 && (cfqq
->slice_end
- jiffies
< cic
->ttime
.ttime_mean
))
3453 * If think times is less than a jiffy than ttime_mean=0 and above
3454 * will not be true. It might happen that slice has not expired yet
3455 * but will expire soon (4-5 ns) during select_queue(). To cover the
3456 * case where think time is less than a jiffy, mark the queue wait
3457 * busy if only 1 jiffy is left in the slice.
3459 if (cfqq
->slice_end
- jiffies
== 1)
3465 static void cfq_completed_request(struct request_queue
*q
, struct request
*rq
)
3467 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
3468 struct cfq_data
*cfqd
= cfqq
->cfqd
;
3469 const int sync
= rq_is_sync(rq
);
3473 cfq_log_cfqq(cfqd
, cfqq
, "complete rqnoidle %d",
3474 !!(rq
->cmd_flags
& REQ_NOIDLE
));
3476 cfq_update_hw_tag(cfqd
);
3478 WARN_ON(!cfqd
->rq_in_driver
);
3479 WARN_ON(!cfqq
->dispatched
);
3480 cfqd
->rq_in_driver
--;
3482 (RQ_CFQG(rq
))->dispatched
--;
3483 cfq_blkiocg_update_completion_stats(&cfqq
->cfqg
->blkg
,
3484 rq_start_time_ns(rq
), rq_io_start_time_ns(rq
),
3485 rq_data_dir(rq
), rq_is_sync(rq
));
3487 cfqd
->rq_in_flight
[cfq_cfqq_sync(cfqq
)]--;
3490 struct cfq_rb_root
*service_tree
;
3492 RQ_CIC(rq
)->ttime
.last_end_request
= now
;
3494 if (cfq_cfqq_on_rr(cfqq
))
3495 service_tree
= cfqq
->service_tree
;
3497 service_tree
= service_tree_for(cfqq
->cfqg
,
3498 cfqq_prio(cfqq
), cfqq_type(cfqq
));
3499 service_tree
->ttime
.last_end_request
= now
;
3500 if (!time_after(rq
->start_time
+ cfqd
->cfq_fifo_expire
[1], now
))
3501 cfqd
->last_delayed_sync
= now
;
3504 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3505 cfqq
->cfqg
->ttime
.last_end_request
= now
;
3509 * If this is the active queue, check if it needs to be expired,
3510 * or if we want to idle in case it has no pending requests.
3512 if (cfqd
->active_queue
== cfqq
) {
3513 const bool cfqq_empty
= RB_EMPTY_ROOT(&cfqq
->sort_list
);
3515 if (cfq_cfqq_slice_new(cfqq
)) {
3516 cfq_set_prio_slice(cfqd
, cfqq
);
3517 cfq_clear_cfqq_slice_new(cfqq
);
3521 * Should we wait for next request to come in before we expire
3524 if (cfq_should_wait_busy(cfqd
, cfqq
)) {
3525 unsigned long extend_sl
= cfqd
->cfq_slice_idle
;
3526 if (!cfqd
->cfq_slice_idle
)
3527 extend_sl
= cfqd
->cfq_group_idle
;
3528 cfqq
->slice_end
= jiffies
+ extend_sl
;
3529 cfq_mark_cfqq_wait_busy(cfqq
);
3530 cfq_log_cfqq(cfqd
, cfqq
, "will busy wait");
3534 * Idling is not enabled on:
3536 * - idle-priority queues
3538 * - queues with still some requests queued
3539 * - when there is a close cooperator
3541 if (cfq_slice_used(cfqq
) || cfq_class_idle(cfqq
))
3542 cfq_slice_expired(cfqd
, 1);
3543 else if (sync
&& cfqq_empty
&&
3544 !cfq_close_cooperator(cfqd
, cfqq
)) {
3545 cfq_arm_slice_timer(cfqd
);
3549 if (!cfqd
->rq_in_driver
)
3550 cfq_schedule_dispatch(cfqd
);
3553 static inline int __cfq_may_queue(struct cfq_queue
*cfqq
)
3555 if (cfq_cfqq_wait_request(cfqq
) && !cfq_cfqq_must_alloc_slice(cfqq
)) {
3556 cfq_mark_cfqq_must_alloc_slice(cfqq
);
3557 return ELV_MQUEUE_MUST
;
3560 return ELV_MQUEUE_MAY
;
3563 static int cfq_may_queue(struct request_queue
*q
, int rw
)
3565 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
3566 struct task_struct
*tsk
= current
;
3567 struct cfq_io_context
*cic
;
3568 struct cfq_queue
*cfqq
;
3571 * don't force setup of a queue from here, as a call to may_queue
3572 * does not necessarily imply that a request actually will be queued.
3573 * so just lookup a possibly existing queue, or return 'may queue'
3576 cic
= cfq_cic_lookup(cfqd
, tsk
->io_context
);
3578 return ELV_MQUEUE_MAY
;
3580 cfqq
= cic_to_cfqq(cic
, rw_is_sync(rw
));
3582 cfq_init_prio_data(cfqq
, cic
->ioc
);
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
, cfqq
->cfqd
->queue
);
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
;
3660 might_sleep_if(gfp_mask
& __GFP_WAIT
);
3662 spin_lock_irq(q
->queue_lock
);
3663 cic
= cfq_get_io_context(cfqd
, gfp_mask
);
3668 cfqq
= cic_to_cfqq(cic
, is_sync
);
3669 if (!cfqq
|| cfqq
== &cfqd
->oom_cfqq
) {
3670 cfqq
= cfq_get_queue(cfqd
, is_sync
, cic
->ioc
, gfp_mask
);
3671 cic_set_cfqq(cic
, cfqq
, is_sync
);
3674 * If the queue was seeky for too long, break it apart.
3676 if (cfq_cfqq_coop(cfqq
) && cfq_cfqq_split_coop(cfqq
)) {
3677 cfq_log_cfqq(cfqd
, cfqq
, "breaking apart cfqq");
3678 cfqq
= split_cfqq(cic
, cfqq
);
3684 * Check to see if this queue is scheduled to merge with
3685 * another, closely cooperating queue. The merging of
3686 * queues happens here as it must be done in process context.
3687 * The reference on new_cfqq was taken in merge_cfqqs.
3690 cfqq
= cfq_merge_cfqqs(cfqd
, cic
, cfqq
);
3693 cfqq
->allocated
[rw
]++;
3696 rq
->elevator_private
[0] = cic
;
3697 rq
->elevator_private
[1] = cfqq
;
3698 rq
->elevator_private
[2] = cfq_ref_get_cfqg(cfqq
->cfqg
);
3699 spin_unlock_irq(q
->queue_lock
);
3703 cfq_schedule_dispatch(cfqd
);
3704 spin_unlock_irq(q
->queue_lock
);
3705 cfq_log(cfqd
, "set_request fail");
3709 static void cfq_kick_queue(struct work_struct
*work
)
3711 struct cfq_data
*cfqd
=
3712 container_of(work
, struct cfq_data
, unplug_work
);
3713 struct request_queue
*q
= cfqd
->queue
;
3715 spin_lock_irq(q
->queue_lock
);
3716 __blk_run_queue(cfqd
->queue
);
3717 spin_unlock_irq(q
->queue_lock
);
3721 * Timer running if the active_queue is currently idling inside its time slice
3723 static void cfq_idle_slice_timer(unsigned long data
)
3725 struct cfq_data
*cfqd
= (struct cfq_data
*) data
;
3726 struct cfq_queue
*cfqq
;
3727 unsigned long flags
;
3730 cfq_log(cfqd
, "idle timer fired");
3732 spin_lock_irqsave(cfqd
->queue
->queue_lock
, flags
);
3734 cfqq
= cfqd
->active_queue
;
3739 * We saw a request before the queue expired, let it through
3741 if (cfq_cfqq_must_dispatch(cfqq
))
3747 if (cfq_slice_used(cfqq
))
3751 * only expire and reinvoke request handler, if there are
3752 * other queues with pending requests
3754 if (!cfqd
->busy_queues
)
3758 * not expired and it has a request pending, let it dispatch
3760 if (!RB_EMPTY_ROOT(&cfqq
->sort_list
))
3764 * Queue depth flag is reset only when the idle didn't succeed
3766 cfq_clear_cfqq_deep(cfqq
);
3769 cfq_slice_expired(cfqd
, timed_out
);
3771 cfq_schedule_dispatch(cfqd
);
3773 spin_unlock_irqrestore(cfqd
->queue
->queue_lock
, flags
);
3776 static void cfq_shutdown_timer_wq(struct cfq_data
*cfqd
)
3778 del_timer_sync(&cfqd
->idle_slice_timer
);
3779 cancel_work_sync(&cfqd
->unplug_work
);
3782 static void cfq_put_async_queues(struct cfq_data
*cfqd
)
3786 for (i
= 0; i
< IOPRIO_BE_NR
; i
++) {
3787 if (cfqd
->async_cfqq
[0][i
])
3788 cfq_put_queue(cfqd
->async_cfqq
[0][i
]);
3789 if (cfqd
->async_cfqq
[1][i
])
3790 cfq_put_queue(cfqd
->async_cfqq
[1][i
]);
3793 if (cfqd
->async_idle_cfqq
)
3794 cfq_put_queue(cfqd
->async_idle_cfqq
);
3797 static void cfq_exit_queue(struct elevator_queue
*e
)
3799 struct cfq_data
*cfqd
= e
->elevator_data
;
3800 struct request_queue
*q
= cfqd
->queue
;
3803 cfq_shutdown_timer_wq(cfqd
);
3805 spin_lock_irq(q
->queue_lock
);
3807 if (cfqd
->active_queue
)
3808 __cfq_slice_expired(cfqd
, cfqd
->active_queue
, 0);
3810 while (!list_empty(&cfqd
->cic_list
)) {
3811 struct cfq_io_context
*cic
= list_entry(cfqd
->cic_list
.next
,
3812 struct cfq_io_context
,
3814 struct io_context
*ioc
= cic
->ioc
;
3816 spin_lock(&ioc
->lock
);
3818 cfq_release_cic(cic
);
3819 spin_unlock(&ioc
->lock
);
3822 cfq_put_async_queues(cfqd
);
3823 cfq_release_cfq_groups(cfqd
);
3826 * If there are groups which we could not unlink from blkcg list,
3827 * wait for a rcu period for them to be freed.
3829 if (cfqd
->nr_blkcg_linked_grps
)
3832 spin_unlock_irq(q
->queue_lock
);
3834 cfq_shutdown_timer_wq(cfqd
);
3837 * Wait for cfqg->blkg->key accessors to exit their grace periods.
3838 * Do this wait only if there are other unlinked groups out
3839 * there. This can happen if cgroup deletion path claimed the
3840 * responsibility of cleaning up a group before queue cleanup code
3843 * Do not call synchronize_rcu() unconditionally as there are drivers
3844 * which create/delete request queue hundreds of times during scan/boot
3845 * and synchronize_rcu() can take significant time and slow down boot.
3850 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3851 /* Free up per cpu stats for root group */
3852 free_percpu(cfqd
->root_group
.blkg
.stats_cpu
);
3857 static void *cfq_init_queue(struct request_queue
*q
)
3859 struct cfq_data
*cfqd
;
3861 struct cfq_group
*cfqg
;
3862 struct cfq_rb_root
*st
;
3864 cfqd
= kmalloc_node(sizeof(*cfqd
), GFP_KERNEL
| __GFP_ZERO
, q
->node
);
3868 /* Init root service tree */
3869 cfqd
->grp_service_tree
= CFQ_RB_ROOT
;
3871 /* Init root group */
3872 cfqg
= &cfqd
->root_group
;
3873 for_each_cfqg_st(cfqg
, i
, j
, st
)
3875 RB_CLEAR_NODE(&cfqg
->rb_node
);
3877 /* Give preference to root group over other groups */
3878 cfqg
->weight
= 2*BLKIO_WEIGHT_DEFAULT
;
3880 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3882 * Set root group reference to 2. One reference will be dropped when
3883 * all groups on cfqd->cfqg_list are being deleted during queue exit.
3884 * Other reference will remain there as we don't want to delete this
3885 * group as it is statically allocated and gets destroyed when
3886 * throtl_data goes away.
3890 if (blkio_alloc_blkg_stats(&cfqg
->blkg
)) {
3898 cfq_blkiocg_add_blkio_group(&blkio_root_cgroup
, &cfqg
->blkg
,
3901 cfqd
->nr_blkcg_linked_grps
++;
3903 /* Add group on cfqd->cfqg_list */
3904 hlist_add_head(&cfqg
->cfqd_node
, &cfqd
->cfqg_list
);
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_completed_req_fn
= cfq_completed_request
,
4084 .elevator_former_req_fn
= elv_rb_former_request
,
4085 .elevator_latter_req_fn
= elv_rb_latter_request
,
4086 .elevator_set_req_fn
= cfq_set_request
,
4087 .elevator_put_req_fn
= cfq_put_request
,
4088 .elevator_may_queue_fn
= cfq_may_queue
,
4089 .elevator_init_fn
= cfq_init_queue
,
4090 .elevator_exit_fn
= cfq_exit_queue
,
4092 .elevator_attrs
= cfq_attrs
,
4093 .elevator_name
= "cfq",
4094 .elevator_owner
= THIS_MODULE
,
4097 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4098 static struct blkio_policy_type blkio_policy_cfq
= {
4100 .blkio_unlink_group_fn
= cfq_unlink_blkio_group
,
4101 .blkio_update_group_weight_fn
= cfq_update_blkio_group_weight
,
4103 .plid
= BLKIO_POLICY_PROP
,
4106 static struct blkio_policy_type blkio_policy_cfq
;
4109 static int __init
cfq_init(void)
4112 * could be 0 on HZ < 1000 setups
4114 if (!cfq_slice_async
)
4115 cfq_slice_async
= 1;
4116 if (!cfq_slice_idle
)
4119 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4120 if (!cfq_group_idle
)
4125 if (cfq_slab_setup())
4128 elv_register(&iosched_cfq
);
4129 blkio_policy_register(&blkio_policy_cfq
);
4134 static void __exit
cfq_exit(void)
4136 blkio_policy_unregister(&blkio_policy_cfq
);
4137 elv_unregister(&iosched_cfq
);
4138 rcu_barrier(); /* make sure all cic RCU frees are complete */
4142 module_init(cfq_init
);
4143 module_exit(cfq_exit
);
4145 MODULE_AUTHOR("Jens Axboe");
4146 MODULE_LICENSE("GPL");
4147 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");