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)
57 #define RQ_CIC(rq) icq_to_cic((rq)->elv.icq)
58 #define RQ_CFQQ(rq) (struct cfq_queue *) ((rq)->elv.priv[0])
59 #define RQ_CFQG(rq) (struct cfq_group *) ((rq)->elv.priv[1])
61 static struct kmem_cache
*cfq_pool
;
62 static struct kmem_cache
*cfq_icq_pool
;
64 #define CFQ_PRIO_LISTS IOPRIO_BE_NR
65 #define cfq_class_idle(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
66 #define cfq_class_rt(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
68 #define sample_valid(samples) ((samples) > 80)
69 #define rb_entry_cfqg(node) rb_entry((node), struct cfq_group, rb_node)
72 unsigned long last_end_request
;
74 unsigned long ttime_total
;
75 unsigned long ttime_samples
;
76 unsigned long ttime_mean
;
80 * Most of our rbtree usage is for sorting with min extraction, so
81 * if we cache the leftmost node we don't have to walk down the tree
82 * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should
83 * move this into the elevator for the rq sorting as well.
89 unsigned total_weight
;
91 struct cfq_ttime ttime
;
93 #define CFQ_RB_ROOT (struct cfq_rb_root) { .rb = RB_ROOT, \
94 .ttime = {.last_end_request = jiffies,},}
97 * Per process-grouping structure
100 /* reference count */
102 /* various state flags, see below */
104 /* parent cfq_data */
105 struct cfq_data
*cfqd
;
106 /* service_tree member */
107 struct rb_node rb_node
;
108 /* service_tree key */
109 unsigned long rb_key
;
110 /* prio tree member */
111 struct rb_node p_node
;
112 /* prio tree root we belong to, if any */
113 struct rb_root
*p_root
;
114 /* sorted list of pending requests */
115 struct rb_root sort_list
;
116 /* if fifo isn't expired, next request to serve */
117 struct request
*next_rq
;
118 /* requests queued in sort_list */
120 /* currently allocated requests */
122 /* fifo list of requests in sort_list */
123 struct list_head fifo
;
125 /* time when queue got scheduled in to dispatch first request. */
126 unsigned long dispatch_start
;
127 unsigned int allocated_slice
;
128 unsigned int slice_dispatch
;
129 /* time when first request from queue completed and slice started. */
130 unsigned long slice_start
;
131 unsigned long slice_end
;
134 /* pending priority requests */
136 /* number of requests that are on the dispatch list or inside driver */
139 /* io prio of this group */
140 unsigned short ioprio
, org_ioprio
;
141 unsigned short ioprio_class
;
146 sector_t last_request_pos
;
148 struct cfq_rb_root
*service_tree
;
149 struct cfq_queue
*new_cfqq
;
150 struct cfq_group
*cfqg
;
151 /* Number of sectors dispatched from queue in single dispatch round */
152 unsigned long nr_sectors
;
156 * First index in the service_trees.
157 * IDLE is handled separately, so it has negative index
167 * Second index in the service_trees.
171 SYNC_NOIDLE_WORKLOAD
= 1,
175 /* This is per cgroup per device grouping structure */
177 /* group service_tree member */
178 struct rb_node rb_node
;
180 /* group service_tree key */
183 unsigned int new_weight
;
186 /* number of cfqq currently on this group */
190 * Per group busy queues average. Useful for workload slice calc. We
191 * create the array for each prio class but at run time it is used
192 * only for RT and BE class and slot for IDLE class remains unused.
193 * This is primarily done to avoid confusion and a gcc warning.
195 unsigned int busy_queues_avg
[CFQ_PRIO_NR
];
197 * rr lists of queues with requests. We maintain service trees for
198 * RT and BE classes. These trees are subdivided in subclasses
199 * of SYNC, SYNC_NOIDLE and ASYNC based on workload type. For IDLE
200 * class there is no subclassification and all the cfq queues go on
201 * a single tree service_tree_idle.
202 * Counts are embedded in the cfq_rb_root
204 struct cfq_rb_root service_trees
[2][3];
205 struct cfq_rb_root service_tree_idle
;
207 unsigned long saved_workload_slice
;
208 enum wl_type_t saved_workload
;
209 enum wl_prio_t saved_serving_prio
;
210 struct blkio_group blkg
;
211 #ifdef CONFIG_CFQ_GROUP_IOSCHED
212 struct hlist_node cfqd_node
;
215 /* number of requests that are on the dispatch list or inside driver */
217 struct cfq_ttime ttime
;
221 struct io_cq icq
; /* must be the first member */
222 struct cfq_queue
*cfqq
[2];
223 struct cfq_ttime ttime
;
227 * Per block device queue structure
230 struct request_queue
*queue
;
231 /* Root service tree for cfq_groups */
232 struct cfq_rb_root grp_service_tree
;
233 struct cfq_group root_group
;
236 * The priority currently being served
238 enum wl_prio_t serving_prio
;
239 enum wl_type_t serving_type
;
240 unsigned long workload_expires
;
241 struct cfq_group
*serving_group
;
244 * Each priority tree is sorted by next_request position. These
245 * trees are used when determining if two or more queues are
246 * interleaving requests (see cfq_close_cooperator).
248 struct rb_root prio_trees
[CFQ_PRIO_LISTS
];
250 unsigned int busy_queues
;
251 unsigned int busy_sync_queues
;
257 * queue-depth detection
263 * -1 => indeterminate, (cfq will behave as if NCQ is present, to allow better detection)
264 * 1 => NCQ is present (hw_tag_est_depth is the estimated max depth)
267 int hw_tag_est_depth
;
268 unsigned int hw_tag_samples
;
271 * idle window management
273 struct timer_list idle_slice_timer
;
274 struct work_struct unplug_work
;
276 struct cfq_queue
*active_queue
;
277 struct cfq_io_cq
*active_cic
;
280 * async queue for each priority case
282 struct cfq_queue
*async_cfqq
[2][IOPRIO_BE_NR
];
283 struct cfq_queue
*async_idle_cfqq
;
285 sector_t last_position
;
288 * tunables, see top of file
290 unsigned int cfq_quantum
;
291 unsigned int cfq_fifo_expire
[2];
292 unsigned int cfq_back_penalty
;
293 unsigned int cfq_back_max
;
294 unsigned int cfq_slice
[2];
295 unsigned int cfq_slice_async_rq
;
296 unsigned int cfq_slice_idle
;
297 unsigned int cfq_group_idle
;
298 unsigned int cfq_latency
;
301 * Fallback dummy cfqq for extreme OOM conditions
303 struct cfq_queue oom_cfqq
;
305 unsigned long last_delayed_sync
;
307 /* List of cfq groups being managed on this device*/
308 struct hlist_head cfqg_list
;
310 /* Number of groups which are on blkcg->blkg_list */
311 unsigned int nr_blkcg_linked_grps
;
314 static struct cfq_group
*cfq_get_next_cfqg(struct cfq_data
*cfqd
);
316 static struct cfq_rb_root
*service_tree_for(struct cfq_group
*cfqg
,
323 if (prio
== IDLE_WORKLOAD
)
324 return &cfqg
->service_tree_idle
;
326 return &cfqg
->service_trees
[prio
][type
];
329 enum cfqq_state_flags
{
330 CFQ_CFQQ_FLAG_on_rr
= 0, /* on round-robin busy list */
331 CFQ_CFQQ_FLAG_wait_request
, /* waiting for a request */
332 CFQ_CFQQ_FLAG_must_dispatch
, /* must be allowed a dispatch */
333 CFQ_CFQQ_FLAG_must_alloc_slice
, /* per-slice must_alloc flag */
334 CFQ_CFQQ_FLAG_fifo_expire
, /* FIFO checked in this slice */
335 CFQ_CFQQ_FLAG_idle_window
, /* slice idling enabled */
336 CFQ_CFQQ_FLAG_prio_changed
, /* task priority has changed */
337 CFQ_CFQQ_FLAG_slice_new
, /* no requests dispatched in slice */
338 CFQ_CFQQ_FLAG_sync
, /* synchronous queue */
339 CFQ_CFQQ_FLAG_coop
, /* cfqq is shared */
340 CFQ_CFQQ_FLAG_split_coop
, /* shared cfqq will be splitted */
341 CFQ_CFQQ_FLAG_deep
, /* sync cfqq experienced large depth */
342 CFQ_CFQQ_FLAG_wait_busy
, /* Waiting for next request */
345 #define CFQ_CFQQ_FNS(name) \
346 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \
348 (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name); \
350 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \
352 (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \
354 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \
356 return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \
360 CFQ_CFQQ_FNS(wait_request
);
361 CFQ_CFQQ_FNS(must_dispatch
);
362 CFQ_CFQQ_FNS(must_alloc_slice
);
363 CFQ_CFQQ_FNS(fifo_expire
);
364 CFQ_CFQQ_FNS(idle_window
);
365 CFQ_CFQQ_FNS(prio_changed
);
366 CFQ_CFQQ_FNS(slice_new
);
369 CFQ_CFQQ_FNS(split_coop
);
371 CFQ_CFQQ_FNS(wait_busy
);
374 #ifdef CONFIG_CFQ_GROUP_IOSCHED
375 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
376 blk_add_trace_msg((cfqd)->queue, "cfq%d%c %s " fmt, (cfqq)->pid, \
377 cfq_cfqq_sync((cfqq)) ? 'S' : 'A', \
378 blkg_path(&(cfqq)->cfqg->blkg), ##args)
380 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...) \
381 blk_add_trace_msg((cfqd)->queue, "%s " fmt, \
382 blkg_path(&(cfqg)->blkg), ##args) \
385 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
386 blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args)
387 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...) do {} while (0)
389 #define cfq_log(cfqd, fmt, args...) \
390 blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
392 /* Traverses through cfq group service trees */
393 #define for_each_cfqg_st(cfqg, i, j, st) \
394 for (i = 0; i <= IDLE_WORKLOAD; i++) \
395 for (j = 0, st = i < IDLE_WORKLOAD ? &cfqg->service_trees[i][j]\
396 : &cfqg->service_tree_idle; \
397 (i < IDLE_WORKLOAD && j <= SYNC_WORKLOAD) || \
398 (i == IDLE_WORKLOAD && j == 0); \
399 j++, st = i < IDLE_WORKLOAD ? \
400 &cfqg->service_trees[i][j]: NULL) \
402 static inline bool cfq_io_thinktime_big(struct cfq_data *cfqd,
403 struct cfq_ttime
*ttime
, bool group_idle
)
406 if (!sample_valid(ttime
->ttime_samples
))
409 slice
= cfqd
->cfq_group_idle
;
411 slice
= cfqd
->cfq_slice_idle
;
412 return ttime
->ttime_mean
> slice
;
415 static inline bool iops_mode(struct cfq_data
*cfqd
)
418 * If we are not idling on queues and it is a NCQ drive, parallel
419 * execution of requests is on and measuring time is not possible
420 * in most of the cases until and unless we drive shallower queue
421 * depths and that becomes a performance bottleneck. In such cases
422 * switch to start providing fairness in terms of number of IOs.
424 if (!cfqd
->cfq_slice_idle
&& cfqd
->hw_tag
)
430 static inline enum wl_prio_t
cfqq_prio(struct cfq_queue
*cfqq
)
432 if (cfq_class_idle(cfqq
))
433 return IDLE_WORKLOAD
;
434 if (cfq_class_rt(cfqq
))
440 static enum wl_type_t
cfqq_type(struct cfq_queue
*cfqq
)
442 if (!cfq_cfqq_sync(cfqq
))
443 return ASYNC_WORKLOAD
;
444 if (!cfq_cfqq_idle_window(cfqq
))
445 return SYNC_NOIDLE_WORKLOAD
;
446 return SYNC_WORKLOAD
;
449 static inline int cfq_group_busy_queues_wl(enum wl_prio_t wl
,
450 struct cfq_data
*cfqd
,
451 struct cfq_group
*cfqg
)
453 if (wl
== IDLE_WORKLOAD
)
454 return cfqg
->service_tree_idle
.count
;
456 return cfqg
->service_trees
[wl
][ASYNC_WORKLOAD
].count
457 + cfqg
->service_trees
[wl
][SYNC_NOIDLE_WORKLOAD
].count
458 + cfqg
->service_trees
[wl
][SYNC_WORKLOAD
].count
;
461 static inline int cfqg_busy_async_queues(struct cfq_data
*cfqd
,
462 struct cfq_group
*cfqg
)
464 return cfqg
->service_trees
[RT_WORKLOAD
][ASYNC_WORKLOAD
].count
465 + cfqg
->service_trees
[BE_WORKLOAD
][ASYNC_WORKLOAD
].count
;
468 static void cfq_dispatch_insert(struct request_queue
*, struct request
*);
469 static struct cfq_queue
*cfq_get_queue(struct cfq_data
*, bool,
470 struct io_context
*, gfp_t
);
471 static struct cfq_io_cq
*cfq_cic_lookup(struct cfq_data
*, struct io_context
*);
473 static inline struct cfq_io_cq
*icq_to_cic(struct io_cq
*icq
)
475 /* cic->icq is the first member, %NULL will convert to %NULL */
476 return container_of(icq
, struct cfq_io_cq
, icq
);
479 static inline struct cfq_queue
*cic_to_cfqq(struct cfq_io_cq
*cic
, bool is_sync
)
481 return cic
->cfqq
[is_sync
];
484 static inline void cic_set_cfqq(struct cfq_io_cq
*cic
, struct cfq_queue
*cfqq
,
487 cic
->cfqq
[is_sync
] = cfqq
;
490 static inline struct cfq_data
*cic_to_cfqd(struct cfq_io_cq
*cic
)
492 return cic
->icq
.q
->elevator
->elevator_data
;
496 * We regard a request as SYNC, if it's either a read or has the SYNC bit
497 * set (in which case it could also be direct WRITE).
499 static inline bool cfq_bio_sync(struct bio
*bio
)
501 return bio_data_dir(bio
) == READ
|| (bio
->bi_rw
& REQ_SYNC
);
505 * scheduler run of queue, if there are requests pending and no one in the
506 * driver that will restart queueing
508 static inline void cfq_schedule_dispatch(struct cfq_data
*cfqd
)
510 if (cfqd
->busy_queues
) {
511 cfq_log(cfqd
, "schedule dispatch");
512 kblockd_schedule_work(cfqd
->queue
, &cfqd
->unplug_work
);
517 * Scale schedule slice based on io priority. Use the sync time slice only
518 * if a queue is marked sync and has sync io queued. A sync queue with async
519 * io only, should not get full sync slice length.
521 static inline int cfq_prio_slice(struct cfq_data
*cfqd
, bool sync
,
524 const int base_slice
= cfqd
->cfq_slice
[sync
];
526 WARN_ON(prio
>= IOPRIO_BE_NR
);
528 return base_slice
+ (base_slice
/CFQ_SLICE_SCALE
* (4 - prio
));
532 cfq_prio_to_slice(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
534 return cfq_prio_slice(cfqd
, cfq_cfqq_sync(cfqq
), cfqq
->ioprio
);
537 static inline u64
cfq_scale_slice(unsigned long delta
, struct cfq_group
*cfqg
)
539 u64 d
= delta
<< CFQ_SERVICE_SHIFT
;
541 d
= d
* BLKIO_WEIGHT_DEFAULT
;
542 do_div(d
, cfqg
->weight
);
546 static inline u64
max_vdisktime(u64 min_vdisktime
, u64 vdisktime
)
548 s64 delta
= (s64
)(vdisktime
- min_vdisktime
);
550 min_vdisktime
= vdisktime
;
552 return min_vdisktime
;
555 static inline u64
min_vdisktime(u64 min_vdisktime
, u64 vdisktime
)
557 s64 delta
= (s64
)(vdisktime
- min_vdisktime
);
559 min_vdisktime
= vdisktime
;
561 return min_vdisktime
;
564 static void update_min_vdisktime(struct cfq_rb_root
*st
)
566 struct cfq_group
*cfqg
;
569 cfqg
= rb_entry_cfqg(st
->left
);
570 st
->min_vdisktime
= max_vdisktime(st
->min_vdisktime
,
576 * get averaged number of queues of RT/BE priority.
577 * average is updated, with a formula that gives more weight to higher numbers,
578 * to quickly follows sudden increases and decrease slowly
581 static inline unsigned cfq_group_get_avg_queues(struct cfq_data
*cfqd
,
582 struct cfq_group
*cfqg
, bool rt
)
584 unsigned min_q
, max_q
;
585 unsigned mult
= cfq_hist_divisor
- 1;
586 unsigned round
= cfq_hist_divisor
/ 2;
587 unsigned busy
= cfq_group_busy_queues_wl(rt
, cfqd
, cfqg
);
589 min_q
= min(cfqg
->busy_queues_avg
[rt
], busy
);
590 max_q
= max(cfqg
->busy_queues_avg
[rt
], busy
);
591 cfqg
->busy_queues_avg
[rt
] = (mult
* max_q
+ min_q
+ round
) /
593 return cfqg
->busy_queues_avg
[rt
];
596 static inline unsigned
597 cfq_group_slice(struct cfq_data
*cfqd
, struct cfq_group
*cfqg
)
599 struct cfq_rb_root
*st
= &cfqd
->grp_service_tree
;
601 return cfq_target_latency
* cfqg
->weight
/ st
->total_weight
;
604 static inline unsigned
605 cfq_scaled_cfqq_slice(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
607 unsigned slice
= cfq_prio_to_slice(cfqd
, cfqq
);
608 if (cfqd
->cfq_latency
) {
610 * interested queues (we consider only the ones with the same
611 * priority class in the cfq group)
613 unsigned iq
= cfq_group_get_avg_queues(cfqd
, cfqq
->cfqg
,
615 unsigned sync_slice
= cfqd
->cfq_slice
[1];
616 unsigned expect_latency
= sync_slice
* iq
;
617 unsigned group_slice
= cfq_group_slice(cfqd
, cfqq
->cfqg
);
619 if (expect_latency
> group_slice
) {
620 unsigned base_low_slice
= 2 * cfqd
->cfq_slice_idle
;
621 /* scale low_slice according to IO priority
622 * and sync vs async */
624 min(slice
, base_low_slice
* slice
/ sync_slice
);
625 /* the adapted slice value is scaled to fit all iqs
626 * into the target latency */
627 slice
= max(slice
* group_slice
/ expect_latency
,
635 cfq_set_prio_slice(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
637 unsigned slice
= cfq_scaled_cfqq_slice(cfqd
, cfqq
);
639 cfqq
->slice_start
= jiffies
;
640 cfqq
->slice_end
= jiffies
+ slice
;
641 cfqq
->allocated_slice
= slice
;
642 cfq_log_cfqq(cfqd
, cfqq
, "set_slice=%lu", cfqq
->slice_end
- jiffies
);
646 * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
647 * isn't valid until the first request from the dispatch is activated
648 * and the slice time set.
650 static inline bool cfq_slice_used(struct cfq_queue
*cfqq
)
652 if (cfq_cfqq_slice_new(cfqq
))
654 if (time_before(jiffies
, cfqq
->slice_end
))
661 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
662 * We choose the request that is closest to the head right now. Distance
663 * behind the head is penalized and only allowed to a certain extent.
665 static struct request
*
666 cfq_choose_req(struct cfq_data
*cfqd
, struct request
*rq1
, struct request
*rq2
, sector_t last
)
668 sector_t s1
, s2
, d1
= 0, d2
= 0;
669 unsigned long back_max
;
670 #define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */
671 #define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */
672 unsigned wrap
= 0; /* bit mask: requests behind the disk head? */
674 if (rq1
== NULL
|| rq1
== rq2
)
679 if (rq_is_sync(rq1
) != rq_is_sync(rq2
))
680 return rq_is_sync(rq1
) ? rq1
: rq2
;
682 if ((rq1
->cmd_flags
^ rq2
->cmd_flags
) & REQ_PRIO
)
683 return rq1
->cmd_flags
& REQ_PRIO
? rq1
: rq2
;
685 s1
= blk_rq_pos(rq1
);
686 s2
= blk_rq_pos(rq2
);
689 * by definition, 1KiB is 2 sectors
691 back_max
= cfqd
->cfq_back_max
* 2;
694 * Strict one way elevator _except_ in the case where we allow
695 * short backward seeks which are biased as twice the cost of a
696 * similar forward seek.
700 else if (s1
+ back_max
>= last
)
701 d1
= (last
- s1
) * cfqd
->cfq_back_penalty
;
703 wrap
|= CFQ_RQ1_WRAP
;
707 else if (s2
+ back_max
>= last
)
708 d2
= (last
- s2
) * cfqd
->cfq_back_penalty
;
710 wrap
|= CFQ_RQ2_WRAP
;
712 /* Found required data */
715 * By doing switch() on the bit mask "wrap" we avoid having to
716 * check two variables for all permutations: --> faster!
719 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
735 case (CFQ_RQ1_WRAP
|CFQ_RQ2_WRAP
): /* both rqs wrapped */
738 * Since both rqs are wrapped,
739 * start with the one that's further behind head
740 * (--> only *one* back seek required),
741 * since back seek takes more time than forward.
751 * The below is leftmost cache rbtree addon
753 static struct cfq_queue
*cfq_rb_first(struct cfq_rb_root
*root
)
755 /* Service tree is empty */
760 root
->left
= rb_first(&root
->rb
);
763 return rb_entry(root
->left
, struct cfq_queue
, rb_node
);
768 static struct cfq_group
*cfq_rb_first_group(struct cfq_rb_root
*root
)
771 root
->left
= rb_first(&root
->rb
);
774 return rb_entry_cfqg(root
->left
);
779 static void rb_erase_init(struct rb_node
*n
, struct rb_root
*root
)
785 static void cfq_rb_erase(struct rb_node
*n
, struct cfq_rb_root
*root
)
789 rb_erase_init(n
, &root
->rb
);
794 * would be nice to take fifo expire time into account as well
796 static struct request
*
797 cfq_find_next_rq(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
798 struct request
*last
)
800 struct rb_node
*rbnext
= rb_next(&last
->rb_node
);
801 struct rb_node
*rbprev
= rb_prev(&last
->rb_node
);
802 struct request
*next
= NULL
, *prev
= NULL
;
804 BUG_ON(RB_EMPTY_NODE(&last
->rb_node
));
807 prev
= rb_entry_rq(rbprev
);
810 next
= rb_entry_rq(rbnext
);
812 rbnext
= rb_first(&cfqq
->sort_list
);
813 if (rbnext
&& rbnext
!= &last
->rb_node
)
814 next
= rb_entry_rq(rbnext
);
817 return cfq_choose_req(cfqd
, next
, prev
, blk_rq_pos(last
));
820 static unsigned long cfq_slice_offset(struct cfq_data
*cfqd
,
821 struct cfq_queue
*cfqq
)
824 * just an approximation, should be ok.
826 return (cfqq
->cfqg
->nr_cfqq
- 1) * (cfq_prio_slice(cfqd
, 1, 0) -
827 cfq_prio_slice(cfqd
, cfq_cfqq_sync(cfqq
), cfqq
->ioprio
));
831 cfqg_key(struct cfq_rb_root
*st
, struct cfq_group
*cfqg
)
833 return cfqg
->vdisktime
- st
->min_vdisktime
;
837 __cfq_group_service_tree_add(struct cfq_rb_root
*st
, struct cfq_group
*cfqg
)
839 struct rb_node
**node
= &st
->rb
.rb_node
;
840 struct rb_node
*parent
= NULL
;
841 struct cfq_group
*__cfqg
;
842 s64 key
= cfqg_key(st
, cfqg
);
845 while (*node
!= NULL
) {
847 __cfqg
= rb_entry_cfqg(parent
);
849 if (key
< cfqg_key(st
, __cfqg
))
850 node
= &parent
->rb_left
;
852 node
= &parent
->rb_right
;
858 st
->left
= &cfqg
->rb_node
;
860 rb_link_node(&cfqg
->rb_node
, parent
, node
);
861 rb_insert_color(&cfqg
->rb_node
, &st
->rb
);
865 cfq_update_group_weight(struct cfq_group
*cfqg
)
867 BUG_ON(!RB_EMPTY_NODE(&cfqg
->rb_node
));
868 if (cfqg
->needs_update
) {
869 cfqg
->weight
= cfqg
->new_weight
;
870 cfqg
->needs_update
= false;
875 cfq_group_service_tree_add(struct cfq_rb_root
*st
, struct cfq_group
*cfqg
)
877 BUG_ON(!RB_EMPTY_NODE(&cfqg
->rb_node
));
879 cfq_update_group_weight(cfqg
);
880 __cfq_group_service_tree_add(st
, cfqg
);
881 st
->total_weight
+= cfqg
->weight
;
885 cfq_group_notify_queue_add(struct cfq_data
*cfqd
, struct cfq_group
*cfqg
)
887 struct cfq_rb_root
*st
= &cfqd
->grp_service_tree
;
888 struct cfq_group
*__cfqg
;
892 if (!RB_EMPTY_NODE(&cfqg
->rb_node
))
896 * Currently put the group at the end. Later implement something
897 * so that groups get lesser vtime based on their weights, so that
898 * if group does not loose all if it was not continuously backlogged.
900 n
= rb_last(&st
->rb
);
902 __cfqg
= rb_entry_cfqg(n
);
903 cfqg
->vdisktime
= __cfqg
->vdisktime
+ CFQ_IDLE_DELAY
;
905 cfqg
->vdisktime
= st
->min_vdisktime
;
906 cfq_group_service_tree_add(st
, cfqg
);
910 cfq_group_service_tree_del(struct cfq_rb_root
*st
, struct cfq_group
*cfqg
)
912 st
->total_weight
-= cfqg
->weight
;
913 if (!RB_EMPTY_NODE(&cfqg
->rb_node
))
914 cfq_rb_erase(&cfqg
->rb_node
, st
);
918 cfq_group_notify_queue_del(struct cfq_data
*cfqd
, struct cfq_group
*cfqg
)
920 struct cfq_rb_root
*st
= &cfqd
->grp_service_tree
;
922 BUG_ON(cfqg
->nr_cfqq
< 1);
925 /* If there are other cfq queues under this group, don't delete it */
929 cfq_log_cfqg(cfqd
, cfqg
, "del_from_rr group");
930 cfq_group_service_tree_del(st
, cfqg
);
931 cfqg
->saved_workload_slice
= 0;
932 cfq_blkiocg_update_dequeue_stats(&cfqg
->blkg
, 1);
935 static inline unsigned int cfq_cfqq_slice_usage(struct cfq_queue
*cfqq
,
936 unsigned int *unaccounted_time
)
938 unsigned int slice_used
;
941 * Queue got expired before even a single request completed or
942 * got expired immediately after first request completion.
944 if (!cfqq
->slice_start
|| cfqq
->slice_start
== jiffies
) {
946 * Also charge the seek time incurred to the group, otherwise
947 * if there are mutiple queues in the group, each can dispatch
948 * a single request on seeky media and cause lots of seek time
949 * and group will never know it.
951 slice_used
= max_t(unsigned, (jiffies
- cfqq
->dispatch_start
),
954 slice_used
= jiffies
- cfqq
->slice_start
;
955 if (slice_used
> cfqq
->allocated_slice
) {
956 *unaccounted_time
= slice_used
- cfqq
->allocated_slice
;
957 slice_used
= cfqq
->allocated_slice
;
959 if (time_after(cfqq
->slice_start
, cfqq
->dispatch_start
))
960 *unaccounted_time
+= cfqq
->slice_start
-
961 cfqq
->dispatch_start
;
967 static void cfq_group_served(struct cfq_data
*cfqd
, struct cfq_group
*cfqg
,
968 struct cfq_queue
*cfqq
)
970 struct cfq_rb_root
*st
= &cfqd
->grp_service_tree
;
971 unsigned int used_sl
, charge
, unaccounted_sl
= 0;
972 int nr_sync
= cfqg
->nr_cfqq
- cfqg_busy_async_queues(cfqd
, cfqg
)
973 - cfqg
->service_tree_idle
.count
;
976 used_sl
= charge
= cfq_cfqq_slice_usage(cfqq
, &unaccounted_sl
);
979 charge
= cfqq
->slice_dispatch
;
980 else if (!cfq_cfqq_sync(cfqq
) && !nr_sync
)
981 charge
= cfqq
->allocated_slice
;
983 /* Can't update vdisktime while group is on service tree */
984 cfq_group_service_tree_del(st
, cfqg
);
985 cfqg
->vdisktime
+= cfq_scale_slice(charge
, cfqg
);
986 /* If a new weight was requested, update now, off tree */
987 cfq_group_service_tree_add(st
, cfqg
);
989 /* This group is being expired. Save the context */
990 if (time_after(cfqd
->workload_expires
, jiffies
)) {
991 cfqg
->saved_workload_slice
= cfqd
->workload_expires
993 cfqg
->saved_workload
= cfqd
->serving_type
;
994 cfqg
->saved_serving_prio
= cfqd
->serving_prio
;
996 cfqg
->saved_workload_slice
= 0;
998 cfq_log_cfqg(cfqd
, cfqg
, "served: vt=%llu min_vt=%llu", cfqg
->vdisktime
,
1000 cfq_log_cfqq(cfqq
->cfqd
, cfqq
,
1001 "sl_used=%u disp=%u charge=%u iops=%u sect=%lu",
1002 used_sl
, cfqq
->slice_dispatch
, charge
,
1003 iops_mode(cfqd
), cfqq
->nr_sectors
);
1004 cfq_blkiocg_update_timeslice_used(&cfqg
->blkg
, used_sl
,
1006 cfq_blkiocg_set_start_empty_time(&cfqg
->blkg
);
1009 #ifdef CONFIG_CFQ_GROUP_IOSCHED
1010 static inline struct cfq_group
*cfqg_of_blkg(struct blkio_group
*blkg
)
1013 return container_of(blkg
, struct cfq_group
, blkg
);
1017 static void cfq_update_blkio_group_weight(void *key
, struct blkio_group
*blkg
,
1018 unsigned int weight
)
1020 struct cfq_group
*cfqg
= cfqg_of_blkg(blkg
);
1021 cfqg
->new_weight
= weight
;
1022 cfqg
->needs_update
= true;
1025 static void cfq_init_add_cfqg_lists(struct cfq_data
*cfqd
,
1026 struct cfq_group
*cfqg
, struct blkio_cgroup
*blkcg
)
1028 struct backing_dev_info
*bdi
= &cfqd
->queue
->backing_dev_info
;
1029 unsigned int major
, minor
;
1032 * Add group onto cgroup list. It might happen that bdi->dev is
1033 * not initialized yet. Initialize this new group without major
1034 * and minor info and this info will be filled in once a new thread
1038 sscanf(dev_name(bdi
->dev
), "%u:%u", &major
, &minor
);
1039 cfq_blkiocg_add_blkio_group(blkcg
, &cfqg
->blkg
,
1040 (void *)cfqd
, MKDEV(major
, minor
));
1042 cfq_blkiocg_add_blkio_group(blkcg
, &cfqg
->blkg
,
1045 cfqd
->nr_blkcg_linked_grps
++;
1046 cfqg
->weight
= blkcg_get_weight(blkcg
, cfqg
->blkg
.dev
);
1048 /* Add group on cfqd list */
1049 hlist_add_head(&cfqg
->cfqd_node
, &cfqd
->cfqg_list
);
1053 * Should be called from sleepable context. No request queue lock as per
1054 * cpu stats are allocated dynamically and alloc_percpu needs to be called
1055 * from sleepable context.
1057 static struct cfq_group
* cfq_alloc_cfqg(struct cfq_data
*cfqd
)
1059 struct cfq_group
*cfqg
= NULL
;
1061 struct cfq_rb_root
*st
;
1063 cfqg
= kzalloc_node(sizeof(*cfqg
), GFP_ATOMIC
, cfqd
->queue
->node
);
1067 for_each_cfqg_st(cfqg
, i
, j
, st
)
1069 RB_CLEAR_NODE(&cfqg
->rb_node
);
1071 cfqg
->ttime
.last_end_request
= jiffies
;
1074 * Take the initial reference that will be released on destroy
1075 * This can be thought of a joint reference by cgroup and
1076 * elevator which will be dropped by either elevator exit
1077 * or cgroup deletion path depending on who is exiting first.
1081 ret
= blkio_alloc_blkg_stats(&cfqg
->blkg
);
1090 static struct cfq_group
*
1091 cfq_find_cfqg(struct cfq_data
*cfqd
, struct blkio_cgroup
*blkcg
)
1093 struct cfq_group
*cfqg
= NULL
;
1095 struct backing_dev_info
*bdi
= &cfqd
->queue
->backing_dev_info
;
1096 unsigned int major
, minor
;
1099 * This is the common case when there are no blkio cgroups.
1100 * Avoid lookup in this case
1102 if (blkcg
== &blkio_root_cgroup
)
1103 cfqg
= &cfqd
->root_group
;
1105 cfqg
= cfqg_of_blkg(blkiocg_lookup_group(blkcg
, key
));
1107 if (cfqg
&& !cfqg
->blkg
.dev
&& bdi
->dev
&& dev_name(bdi
->dev
)) {
1108 sscanf(dev_name(bdi
->dev
), "%u:%u", &major
, &minor
);
1109 cfqg
->blkg
.dev
= MKDEV(major
, minor
);
1116 * Search for the cfq group current task belongs to. request_queue lock must
1119 static struct cfq_group
*cfq_get_cfqg(struct cfq_data
*cfqd
)
1121 struct blkio_cgroup
*blkcg
;
1122 struct cfq_group
*cfqg
= NULL
, *__cfqg
= NULL
;
1123 struct request_queue
*q
= cfqd
->queue
;
1126 blkcg
= task_blkio_cgroup(current
);
1127 cfqg
= cfq_find_cfqg(cfqd
, blkcg
);
1134 * Need to allocate a group. Allocation of group also needs allocation
1135 * of per cpu stats which in-turn takes a mutex() and can block. Hence
1136 * we need to drop rcu lock and queue_lock before we call alloc.
1138 * Not taking any queue reference here and assuming that queue is
1139 * around by the time we return. CFQ queue allocation code does
1140 * the same. It might be racy though.
1144 spin_unlock_irq(q
->queue_lock
);
1146 cfqg
= cfq_alloc_cfqg(cfqd
);
1148 spin_lock_irq(q
->queue_lock
);
1151 blkcg
= task_blkio_cgroup(current
);
1154 * If some other thread already allocated the group while we were
1155 * not holding queue lock, free up the group
1157 __cfqg
= cfq_find_cfqg(cfqd
, blkcg
);
1166 cfqg
= &cfqd
->root_group
;
1168 cfq_init_add_cfqg_lists(cfqd
, cfqg
, blkcg
);
1173 static inline struct cfq_group
*cfq_ref_get_cfqg(struct cfq_group
*cfqg
)
1179 static void cfq_link_cfqq_cfqg(struct cfq_queue
*cfqq
, struct cfq_group
*cfqg
)
1181 /* Currently, all async queues are mapped to root group */
1182 if (!cfq_cfqq_sync(cfqq
))
1183 cfqg
= &cfqq
->cfqd
->root_group
;
1186 /* cfqq reference on cfqg */
1190 static void cfq_put_cfqg(struct cfq_group
*cfqg
)
1192 struct cfq_rb_root
*st
;
1195 BUG_ON(cfqg
->ref
<= 0);
1199 for_each_cfqg_st(cfqg
, i
, j
, st
)
1200 BUG_ON(!RB_EMPTY_ROOT(&st
->rb
));
1201 free_percpu(cfqg
->blkg
.stats_cpu
);
1205 static void cfq_destroy_cfqg(struct cfq_data
*cfqd
, struct cfq_group
*cfqg
)
1207 /* Something wrong if we are trying to remove same group twice */
1208 BUG_ON(hlist_unhashed(&cfqg
->cfqd_node
));
1210 hlist_del_init(&cfqg
->cfqd_node
);
1212 BUG_ON(cfqd
->nr_blkcg_linked_grps
<= 0);
1213 cfqd
->nr_blkcg_linked_grps
--;
1216 * Put the reference taken at the time of creation so that when all
1217 * queues are gone, group can be destroyed.
1222 static void cfq_release_cfq_groups(struct cfq_data
*cfqd
)
1224 struct hlist_node
*pos
, *n
;
1225 struct cfq_group
*cfqg
;
1227 hlist_for_each_entry_safe(cfqg
, pos
, n
, &cfqd
->cfqg_list
, cfqd_node
) {
1229 * If cgroup removal path got to blk_group first and removed
1230 * it from cgroup list, then it will take care of destroying
1233 if (!cfq_blkiocg_del_blkio_group(&cfqg
->blkg
))
1234 cfq_destroy_cfqg(cfqd
, cfqg
);
1239 * Blk cgroup controller notification saying that blkio_group object is being
1240 * delinked as associated cgroup object is going away. That also means that
1241 * no new IO will come in this group. So get rid of this group as soon as
1242 * any pending IO in the group is finished.
1244 * This function is called under rcu_read_lock(). key is the rcu protected
1245 * pointer. That means "key" is a valid cfq_data pointer as long as we are rcu
1248 * "key" was fetched from blkio_group under blkio_cgroup->lock. That means
1249 * it should not be NULL as even if elevator was exiting, cgroup deltion
1250 * path got to it first.
1252 static void cfq_unlink_blkio_group(void *key
, struct blkio_group
*blkg
)
1254 unsigned long flags
;
1255 struct cfq_data
*cfqd
= key
;
1257 spin_lock_irqsave(cfqd
->queue
->queue_lock
, flags
);
1258 cfq_destroy_cfqg(cfqd
, cfqg_of_blkg(blkg
));
1259 spin_unlock_irqrestore(cfqd
->queue
->queue_lock
, flags
);
1262 #else /* GROUP_IOSCHED */
1263 static struct cfq_group
*cfq_get_cfqg(struct cfq_data
*cfqd
)
1265 return &cfqd
->root_group
;
1268 static inline struct cfq_group
*cfq_ref_get_cfqg(struct cfq_group
*cfqg
)
1274 cfq_link_cfqq_cfqg(struct cfq_queue
*cfqq
, struct cfq_group
*cfqg
) {
1278 static void cfq_release_cfq_groups(struct cfq_data
*cfqd
) {}
1279 static inline void cfq_put_cfqg(struct cfq_group
*cfqg
) {}
1281 #endif /* GROUP_IOSCHED */
1284 * The cfqd->service_trees holds all pending cfq_queue's that have
1285 * requests waiting to be processed. It is sorted in the order that
1286 * we will service the queues.
1288 static void cfq_service_tree_add(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
1291 struct rb_node
**p
, *parent
;
1292 struct cfq_queue
*__cfqq
;
1293 unsigned long rb_key
;
1294 struct cfq_rb_root
*service_tree
;
1298 service_tree
= service_tree_for(cfqq
->cfqg
, cfqq_prio(cfqq
),
1300 if (cfq_class_idle(cfqq
)) {
1301 rb_key
= CFQ_IDLE_DELAY
;
1302 parent
= rb_last(&service_tree
->rb
);
1303 if (parent
&& parent
!= &cfqq
->rb_node
) {
1304 __cfqq
= rb_entry(parent
, struct cfq_queue
, rb_node
);
1305 rb_key
+= __cfqq
->rb_key
;
1308 } else if (!add_front
) {
1310 * Get our rb key offset. Subtract any residual slice
1311 * value carried from last service. A negative resid
1312 * count indicates slice overrun, and this should position
1313 * the next service time further away in the tree.
1315 rb_key
= cfq_slice_offset(cfqd
, cfqq
) + jiffies
;
1316 rb_key
-= cfqq
->slice_resid
;
1317 cfqq
->slice_resid
= 0;
1320 __cfqq
= cfq_rb_first(service_tree
);
1321 rb_key
+= __cfqq
? __cfqq
->rb_key
: jiffies
;
1324 if (!RB_EMPTY_NODE(&cfqq
->rb_node
)) {
1327 * same position, nothing more to do
1329 if (rb_key
== cfqq
->rb_key
&&
1330 cfqq
->service_tree
== service_tree
)
1333 cfq_rb_erase(&cfqq
->rb_node
, cfqq
->service_tree
);
1334 cfqq
->service_tree
= NULL
;
1339 cfqq
->service_tree
= service_tree
;
1340 p
= &service_tree
->rb
.rb_node
;
1345 __cfqq
= rb_entry(parent
, struct cfq_queue
, rb_node
);
1348 * sort by key, that represents service time.
1350 if (time_before(rb_key
, __cfqq
->rb_key
))
1353 n
= &(*p
)->rb_right
;
1361 service_tree
->left
= &cfqq
->rb_node
;
1363 cfqq
->rb_key
= rb_key
;
1364 rb_link_node(&cfqq
->rb_node
, parent
, p
);
1365 rb_insert_color(&cfqq
->rb_node
, &service_tree
->rb
);
1366 service_tree
->count
++;
1367 if (add_front
|| !new_cfqq
)
1369 cfq_group_notify_queue_add(cfqd
, cfqq
->cfqg
);
1372 static struct cfq_queue
*
1373 cfq_prio_tree_lookup(struct cfq_data
*cfqd
, struct rb_root
*root
,
1374 sector_t sector
, struct rb_node
**ret_parent
,
1375 struct rb_node
***rb_link
)
1377 struct rb_node
**p
, *parent
;
1378 struct cfq_queue
*cfqq
= NULL
;
1386 cfqq
= rb_entry(parent
, struct cfq_queue
, p_node
);
1389 * Sort strictly based on sector. Smallest to the left,
1390 * largest to the right.
1392 if (sector
> blk_rq_pos(cfqq
->next_rq
))
1393 n
= &(*p
)->rb_right
;
1394 else if (sector
< blk_rq_pos(cfqq
->next_rq
))
1402 *ret_parent
= parent
;
1408 static void cfq_prio_tree_add(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
1410 struct rb_node
**p
, *parent
;
1411 struct cfq_queue
*__cfqq
;
1414 rb_erase(&cfqq
->p_node
, cfqq
->p_root
);
1415 cfqq
->p_root
= NULL
;
1418 if (cfq_class_idle(cfqq
))
1423 cfqq
->p_root
= &cfqd
->prio_trees
[cfqq
->org_ioprio
];
1424 __cfqq
= cfq_prio_tree_lookup(cfqd
, cfqq
->p_root
,
1425 blk_rq_pos(cfqq
->next_rq
), &parent
, &p
);
1427 rb_link_node(&cfqq
->p_node
, parent
, p
);
1428 rb_insert_color(&cfqq
->p_node
, cfqq
->p_root
);
1430 cfqq
->p_root
= NULL
;
1434 * Update cfqq's position in the service tree.
1436 static void cfq_resort_rr_list(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
1439 * Resorting requires the cfqq to be on the RR list already.
1441 if (cfq_cfqq_on_rr(cfqq
)) {
1442 cfq_service_tree_add(cfqd
, cfqq
, 0);
1443 cfq_prio_tree_add(cfqd
, cfqq
);
1448 * add to busy list of queues for service, trying to be fair in ordering
1449 * the pending list according to last request service
1451 static void cfq_add_cfqq_rr(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
1453 cfq_log_cfqq(cfqd
, cfqq
, "add_to_rr");
1454 BUG_ON(cfq_cfqq_on_rr(cfqq
));
1455 cfq_mark_cfqq_on_rr(cfqq
);
1456 cfqd
->busy_queues
++;
1457 if (cfq_cfqq_sync(cfqq
))
1458 cfqd
->busy_sync_queues
++;
1460 cfq_resort_rr_list(cfqd
, cfqq
);
1464 * Called when the cfqq no longer has requests pending, remove it from
1467 static void cfq_del_cfqq_rr(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
1469 cfq_log_cfqq(cfqd
, cfqq
, "del_from_rr");
1470 BUG_ON(!cfq_cfqq_on_rr(cfqq
));
1471 cfq_clear_cfqq_on_rr(cfqq
);
1473 if (!RB_EMPTY_NODE(&cfqq
->rb_node
)) {
1474 cfq_rb_erase(&cfqq
->rb_node
, cfqq
->service_tree
);
1475 cfqq
->service_tree
= NULL
;
1478 rb_erase(&cfqq
->p_node
, cfqq
->p_root
);
1479 cfqq
->p_root
= NULL
;
1482 cfq_group_notify_queue_del(cfqd
, cfqq
->cfqg
);
1483 BUG_ON(!cfqd
->busy_queues
);
1484 cfqd
->busy_queues
--;
1485 if (cfq_cfqq_sync(cfqq
))
1486 cfqd
->busy_sync_queues
--;
1490 * rb tree support functions
1492 static void cfq_del_rq_rb(struct request
*rq
)
1494 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
1495 const int sync
= rq_is_sync(rq
);
1497 BUG_ON(!cfqq
->queued
[sync
]);
1498 cfqq
->queued
[sync
]--;
1500 elv_rb_del(&cfqq
->sort_list
, rq
);
1502 if (cfq_cfqq_on_rr(cfqq
) && RB_EMPTY_ROOT(&cfqq
->sort_list
)) {
1504 * Queue will be deleted from service tree when we actually
1505 * expire it later. Right now just remove it from prio tree
1509 rb_erase(&cfqq
->p_node
, cfqq
->p_root
);
1510 cfqq
->p_root
= NULL
;
1515 static void cfq_add_rq_rb(struct request
*rq
)
1517 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
1518 struct cfq_data
*cfqd
= cfqq
->cfqd
;
1519 struct request
*prev
;
1521 cfqq
->queued
[rq_is_sync(rq
)]++;
1523 elv_rb_add(&cfqq
->sort_list
, rq
);
1525 if (!cfq_cfqq_on_rr(cfqq
))
1526 cfq_add_cfqq_rr(cfqd
, cfqq
);
1529 * check if this request is a better next-serve candidate
1531 prev
= cfqq
->next_rq
;
1532 cfqq
->next_rq
= cfq_choose_req(cfqd
, cfqq
->next_rq
, rq
, cfqd
->last_position
);
1535 * adjust priority tree position, if ->next_rq changes
1537 if (prev
!= cfqq
->next_rq
)
1538 cfq_prio_tree_add(cfqd
, cfqq
);
1540 BUG_ON(!cfqq
->next_rq
);
1543 static void cfq_reposition_rq_rb(struct cfq_queue
*cfqq
, struct request
*rq
)
1545 elv_rb_del(&cfqq
->sort_list
, rq
);
1546 cfqq
->queued
[rq_is_sync(rq
)]--;
1547 cfq_blkiocg_update_io_remove_stats(&(RQ_CFQG(rq
))->blkg
,
1548 rq_data_dir(rq
), rq_is_sync(rq
));
1550 cfq_blkiocg_update_io_add_stats(&(RQ_CFQG(rq
))->blkg
,
1551 &cfqq
->cfqd
->serving_group
->blkg
, rq_data_dir(rq
),
1555 static struct request
*
1556 cfq_find_rq_fmerge(struct cfq_data
*cfqd
, struct bio
*bio
)
1558 struct task_struct
*tsk
= current
;
1559 struct cfq_io_cq
*cic
;
1560 struct cfq_queue
*cfqq
;
1562 cic
= cfq_cic_lookup(cfqd
, tsk
->io_context
);
1566 cfqq
= cic_to_cfqq(cic
, cfq_bio_sync(bio
));
1568 sector_t sector
= bio
->bi_sector
+ bio_sectors(bio
);
1570 return elv_rb_find(&cfqq
->sort_list
, sector
);
1576 static void cfq_activate_request(struct request_queue
*q
, struct request
*rq
)
1578 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
1580 cfqd
->rq_in_driver
++;
1581 cfq_log_cfqq(cfqd
, RQ_CFQQ(rq
), "activate rq, drv=%d",
1582 cfqd
->rq_in_driver
);
1584 cfqd
->last_position
= blk_rq_pos(rq
) + blk_rq_sectors(rq
);
1587 static void cfq_deactivate_request(struct request_queue
*q
, struct request
*rq
)
1589 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
1591 WARN_ON(!cfqd
->rq_in_driver
);
1592 cfqd
->rq_in_driver
--;
1593 cfq_log_cfqq(cfqd
, RQ_CFQQ(rq
), "deactivate rq, drv=%d",
1594 cfqd
->rq_in_driver
);
1597 static void cfq_remove_request(struct request
*rq
)
1599 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
1601 if (cfqq
->next_rq
== rq
)
1602 cfqq
->next_rq
= cfq_find_next_rq(cfqq
->cfqd
, cfqq
, rq
);
1604 list_del_init(&rq
->queuelist
);
1607 cfqq
->cfqd
->rq_queued
--;
1608 cfq_blkiocg_update_io_remove_stats(&(RQ_CFQG(rq
))->blkg
,
1609 rq_data_dir(rq
), rq_is_sync(rq
));
1610 if (rq
->cmd_flags
& REQ_PRIO
) {
1611 WARN_ON(!cfqq
->prio_pending
);
1612 cfqq
->prio_pending
--;
1616 static int cfq_merge(struct request_queue
*q
, struct request
**req
,
1619 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
1620 struct request
*__rq
;
1622 __rq
= cfq_find_rq_fmerge(cfqd
, bio
);
1623 if (__rq
&& elv_rq_merge_ok(__rq
, bio
)) {
1625 return ELEVATOR_FRONT_MERGE
;
1628 return ELEVATOR_NO_MERGE
;
1631 static void cfq_merged_request(struct request_queue
*q
, struct request
*req
,
1634 if (type
== ELEVATOR_FRONT_MERGE
) {
1635 struct cfq_queue
*cfqq
= RQ_CFQQ(req
);
1637 cfq_reposition_rq_rb(cfqq
, req
);
1641 static void cfq_bio_merged(struct request_queue
*q
, struct request
*req
,
1644 cfq_blkiocg_update_io_merged_stats(&(RQ_CFQG(req
))->blkg
,
1645 bio_data_dir(bio
), cfq_bio_sync(bio
));
1649 cfq_merged_requests(struct request_queue
*q
, struct request
*rq
,
1650 struct request
*next
)
1652 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
1654 * reposition in fifo if next is older than rq
1656 if (!list_empty(&rq
->queuelist
) && !list_empty(&next
->queuelist
) &&
1657 time_before(rq_fifo_time(next
), rq_fifo_time(rq
))) {
1658 list_move(&rq
->queuelist
, &next
->queuelist
);
1659 rq_set_fifo_time(rq
, rq_fifo_time(next
));
1662 if (cfqq
->next_rq
== next
)
1664 cfq_remove_request(next
);
1665 cfq_blkiocg_update_io_merged_stats(&(RQ_CFQG(rq
))->blkg
,
1666 rq_data_dir(next
), rq_is_sync(next
));
1669 static int cfq_allow_merge(struct request_queue
*q
, struct request
*rq
,
1672 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
1673 struct cfq_io_cq
*cic
;
1674 struct cfq_queue
*cfqq
;
1677 * Disallow merge of a sync bio into an async request.
1679 if (cfq_bio_sync(bio
) && !rq_is_sync(rq
))
1683 * Lookup the cfqq that this bio will be queued with and allow
1684 * merge only if rq is queued there. This function can be called
1685 * from plug merge without queue_lock. In such cases, ioc of @rq
1686 * and %current are guaranteed to be equal. Avoid lookup which
1687 * requires queue_lock by using @rq's cic.
1689 if (current
->io_context
== RQ_CIC(rq
)->icq
.ioc
) {
1692 cic
= cfq_cic_lookup(cfqd
, current
->io_context
);
1697 cfqq
= cic_to_cfqq(cic
, cfq_bio_sync(bio
));
1698 return cfqq
== RQ_CFQQ(rq
);
1701 static inline void cfq_del_timer(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
1703 del_timer(&cfqd
->idle_slice_timer
);
1704 cfq_blkiocg_update_idle_time_stats(&cfqq
->cfqg
->blkg
);
1707 static void __cfq_set_active_queue(struct cfq_data
*cfqd
,
1708 struct cfq_queue
*cfqq
)
1711 cfq_log_cfqq(cfqd
, cfqq
, "set_active wl_prio:%d wl_type:%d",
1712 cfqd
->serving_prio
, cfqd
->serving_type
);
1713 cfq_blkiocg_update_avg_queue_size_stats(&cfqq
->cfqg
->blkg
);
1714 cfqq
->slice_start
= 0;
1715 cfqq
->dispatch_start
= jiffies
;
1716 cfqq
->allocated_slice
= 0;
1717 cfqq
->slice_end
= 0;
1718 cfqq
->slice_dispatch
= 0;
1719 cfqq
->nr_sectors
= 0;
1721 cfq_clear_cfqq_wait_request(cfqq
);
1722 cfq_clear_cfqq_must_dispatch(cfqq
);
1723 cfq_clear_cfqq_must_alloc_slice(cfqq
);
1724 cfq_clear_cfqq_fifo_expire(cfqq
);
1725 cfq_mark_cfqq_slice_new(cfqq
);
1727 cfq_del_timer(cfqd
, cfqq
);
1730 cfqd
->active_queue
= cfqq
;
1734 * current cfqq expired its slice (or was too idle), select new one
1737 __cfq_slice_expired(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
1740 cfq_log_cfqq(cfqd
, cfqq
, "slice expired t=%d", timed_out
);
1742 if (cfq_cfqq_wait_request(cfqq
))
1743 cfq_del_timer(cfqd
, cfqq
);
1745 cfq_clear_cfqq_wait_request(cfqq
);
1746 cfq_clear_cfqq_wait_busy(cfqq
);
1749 * If this cfqq is shared between multiple processes, check to
1750 * make sure that those processes are still issuing I/Os within
1751 * the mean seek distance. If not, it may be time to break the
1752 * queues apart again.
1754 if (cfq_cfqq_coop(cfqq
) && CFQQ_SEEKY(cfqq
))
1755 cfq_mark_cfqq_split_coop(cfqq
);
1758 * store what was left of this slice, if the queue idled/timed out
1761 if (cfq_cfqq_slice_new(cfqq
))
1762 cfqq
->slice_resid
= cfq_scaled_cfqq_slice(cfqd
, cfqq
);
1764 cfqq
->slice_resid
= cfqq
->slice_end
- jiffies
;
1765 cfq_log_cfqq(cfqd
, cfqq
, "resid=%ld", cfqq
->slice_resid
);
1768 cfq_group_served(cfqd
, cfqq
->cfqg
, cfqq
);
1770 if (cfq_cfqq_on_rr(cfqq
) && RB_EMPTY_ROOT(&cfqq
->sort_list
))
1771 cfq_del_cfqq_rr(cfqd
, cfqq
);
1773 cfq_resort_rr_list(cfqd
, cfqq
);
1775 if (cfqq
== cfqd
->active_queue
)
1776 cfqd
->active_queue
= NULL
;
1778 if (cfqd
->active_cic
) {
1779 put_io_context(cfqd
->active_cic
->icq
.ioc
, cfqd
->queue
);
1780 cfqd
->active_cic
= NULL
;
1784 static inline void cfq_slice_expired(struct cfq_data
*cfqd
, bool timed_out
)
1786 struct cfq_queue
*cfqq
= cfqd
->active_queue
;
1789 __cfq_slice_expired(cfqd
, cfqq
, timed_out
);
1793 * Get next queue for service. Unless we have a queue preemption,
1794 * we'll simply select the first cfqq in the service tree.
1796 static struct cfq_queue
*cfq_get_next_queue(struct cfq_data
*cfqd
)
1798 struct cfq_rb_root
*service_tree
=
1799 service_tree_for(cfqd
->serving_group
, cfqd
->serving_prio
,
1800 cfqd
->serving_type
);
1802 if (!cfqd
->rq_queued
)
1805 /* There is nothing to dispatch */
1808 if (RB_EMPTY_ROOT(&service_tree
->rb
))
1810 return cfq_rb_first(service_tree
);
1813 static struct cfq_queue
*cfq_get_next_queue_forced(struct cfq_data
*cfqd
)
1815 struct cfq_group
*cfqg
;
1816 struct cfq_queue
*cfqq
;
1818 struct cfq_rb_root
*st
;
1820 if (!cfqd
->rq_queued
)
1823 cfqg
= cfq_get_next_cfqg(cfqd
);
1827 for_each_cfqg_st(cfqg
, i
, j
, st
)
1828 if ((cfqq
= cfq_rb_first(st
)) != NULL
)
1834 * Get and set a new active queue for service.
1836 static struct cfq_queue
*cfq_set_active_queue(struct cfq_data
*cfqd
,
1837 struct cfq_queue
*cfqq
)
1840 cfqq
= cfq_get_next_queue(cfqd
);
1842 __cfq_set_active_queue(cfqd
, cfqq
);
1846 static inline sector_t
cfq_dist_from_last(struct cfq_data
*cfqd
,
1849 if (blk_rq_pos(rq
) >= cfqd
->last_position
)
1850 return blk_rq_pos(rq
) - cfqd
->last_position
;
1852 return cfqd
->last_position
- blk_rq_pos(rq
);
1855 static inline int cfq_rq_close(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
1858 return cfq_dist_from_last(cfqd
, rq
) <= CFQQ_CLOSE_THR
;
1861 static struct cfq_queue
*cfqq_close(struct cfq_data
*cfqd
,
1862 struct cfq_queue
*cur_cfqq
)
1864 struct rb_root
*root
= &cfqd
->prio_trees
[cur_cfqq
->org_ioprio
];
1865 struct rb_node
*parent
, *node
;
1866 struct cfq_queue
*__cfqq
;
1867 sector_t sector
= cfqd
->last_position
;
1869 if (RB_EMPTY_ROOT(root
))
1873 * First, if we find a request starting at the end of the last
1874 * request, choose it.
1876 __cfqq
= cfq_prio_tree_lookup(cfqd
, root
, sector
, &parent
, NULL
);
1881 * If the exact sector wasn't found, the parent of the NULL leaf
1882 * will contain the closest sector.
1884 __cfqq
= rb_entry(parent
, struct cfq_queue
, p_node
);
1885 if (cfq_rq_close(cfqd
, cur_cfqq
, __cfqq
->next_rq
))
1888 if (blk_rq_pos(__cfqq
->next_rq
) < sector
)
1889 node
= rb_next(&__cfqq
->p_node
);
1891 node
= rb_prev(&__cfqq
->p_node
);
1895 __cfqq
= rb_entry(node
, struct cfq_queue
, p_node
);
1896 if (cfq_rq_close(cfqd
, cur_cfqq
, __cfqq
->next_rq
))
1904 * cur_cfqq - passed in so that we don't decide that the current queue is
1905 * closely cooperating with itself.
1907 * So, basically we're assuming that that cur_cfqq has dispatched at least
1908 * one request, and that cfqd->last_position reflects a position on the disk
1909 * associated with the I/O issued by cur_cfqq. I'm not sure this is a valid
1912 static struct cfq_queue
*cfq_close_cooperator(struct cfq_data
*cfqd
,
1913 struct cfq_queue
*cur_cfqq
)
1915 struct cfq_queue
*cfqq
;
1917 if (cfq_class_idle(cur_cfqq
))
1919 if (!cfq_cfqq_sync(cur_cfqq
))
1921 if (CFQQ_SEEKY(cur_cfqq
))
1925 * Don't search priority tree if it's the only queue in the group.
1927 if (cur_cfqq
->cfqg
->nr_cfqq
== 1)
1931 * We should notice if some of the queues are cooperating, eg
1932 * working closely on the same area of the disk. In that case,
1933 * we can group them together and don't waste time idling.
1935 cfqq
= cfqq_close(cfqd
, cur_cfqq
);
1939 /* If new queue belongs to different cfq_group, don't choose it */
1940 if (cur_cfqq
->cfqg
!= cfqq
->cfqg
)
1944 * It only makes sense to merge sync queues.
1946 if (!cfq_cfqq_sync(cfqq
))
1948 if (CFQQ_SEEKY(cfqq
))
1952 * Do not merge queues of different priority classes
1954 if (cfq_class_rt(cfqq
) != cfq_class_rt(cur_cfqq
))
1961 * Determine whether we should enforce idle window for this queue.
1964 static bool cfq_should_idle(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
1966 enum wl_prio_t prio
= cfqq_prio(cfqq
);
1967 struct cfq_rb_root
*service_tree
= cfqq
->service_tree
;
1969 BUG_ON(!service_tree
);
1970 BUG_ON(!service_tree
->count
);
1972 if (!cfqd
->cfq_slice_idle
)
1975 /* We never do for idle class queues. */
1976 if (prio
== IDLE_WORKLOAD
)
1979 /* We do for queues that were marked with idle window flag. */
1980 if (cfq_cfqq_idle_window(cfqq
) &&
1981 !(blk_queue_nonrot(cfqd
->queue
) && cfqd
->hw_tag
))
1985 * Otherwise, we do only if they are the last ones
1986 * in their service tree.
1988 if (service_tree
->count
== 1 && cfq_cfqq_sync(cfqq
) &&
1989 !cfq_io_thinktime_big(cfqd
, &service_tree
->ttime
, false))
1991 cfq_log_cfqq(cfqd
, cfqq
, "Not idling. st->count:%d",
1992 service_tree
->count
);
1996 static void cfq_arm_slice_timer(struct cfq_data
*cfqd
)
1998 struct cfq_queue
*cfqq
= cfqd
->active_queue
;
1999 struct cfq_io_cq
*cic
;
2000 unsigned long sl
, group_idle
= 0;
2003 * SSD device without seek penalty, disable idling. But only do so
2004 * for devices that support queuing, otherwise we still have a problem
2005 * with sync vs async workloads.
2007 if (blk_queue_nonrot(cfqd
->queue
) && cfqd
->hw_tag
)
2010 WARN_ON(!RB_EMPTY_ROOT(&cfqq
->sort_list
));
2011 WARN_ON(cfq_cfqq_slice_new(cfqq
));
2014 * idle is disabled, either manually or by past process history
2016 if (!cfq_should_idle(cfqd
, cfqq
)) {
2017 /* no queue idling. Check for group idling */
2018 if (cfqd
->cfq_group_idle
)
2019 group_idle
= cfqd
->cfq_group_idle
;
2025 * still active requests from this queue, don't idle
2027 if (cfqq
->dispatched
)
2031 * task has exited, don't wait
2033 cic
= cfqd
->active_cic
;
2034 if (!cic
|| !atomic_read(&cic
->icq
.ioc
->nr_tasks
))
2038 * If our average think time is larger than the remaining time
2039 * slice, then don't idle. This avoids overrunning the allotted
2042 if (sample_valid(cic
->ttime
.ttime_samples
) &&
2043 (cfqq
->slice_end
- jiffies
< cic
->ttime
.ttime_mean
)) {
2044 cfq_log_cfqq(cfqd
, cfqq
, "Not idling. think_time:%lu",
2045 cic
->ttime
.ttime_mean
);
2049 /* There are other queues in the group, don't do group idle */
2050 if (group_idle
&& cfqq
->cfqg
->nr_cfqq
> 1)
2053 cfq_mark_cfqq_wait_request(cfqq
);
2056 sl
= cfqd
->cfq_group_idle
;
2058 sl
= cfqd
->cfq_slice_idle
;
2060 mod_timer(&cfqd
->idle_slice_timer
, jiffies
+ sl
);
2061 cfq_blkiocg_update_set_idle_time_stats(&cfqq
->cfqg
->blkg
);
2062 cfq_log_cfqq(cfqd
, cfqq
, "arm_idle: %lu group_idle: %d", sl
,
2063 group_idle
? 1 : 0);
2067 * Move request from internal lists to the request queue dispatch list.
2069 static void cfq_dispatch_insert(struct request_queue
*q
, struct request
*rq
)
2071 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
2072 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
2074 cfq_log_cfqq(cfqd
, cfqq
, "dispatch_insert");
2076 cfqq
->next_rq
= cfq_find_next_rq(cfqd
, cfqq
, rq
);
2077 cfq_remove_request(rq
);
2079 (RQ_CFQG(rq
))->dispatched
++;
2080 elv_dispatch_sort(q
, rq
);
2082 cfqd
->rq_in_flight
[cfq_cfqq_sync(cfqq
)]++;
2083 cfqq
->nr_sectors
+= blk_rq_sectors(rq
);
2084 cfq_blkiocg_update_dispatch_stats(&cfqq
->cfqg
->blkg
, blk_rq_bytes(rq
),
2085 rq_data_dir(rq
), rq_is_sync(rq
));
2089 * return expired entry, or NULL to just start from scratch in rbtree
2091 static struct request
*cfq_check_fifo(struct cfq_queue
*cfqq
)
2093 struct request
*rq
= NULL
;
2095 if (cfq_cfqq_fifo_expire(cfqq
))
2098 cfq_mark_cfqq_fifo_expire(cfqq
);
2100 if (list_empty(&cfqq
->fifo
))
2103 rq
= rq_entry_fifo(cfqq
->fifo
.next
);
2104 if (time_before(jiffies
, rq_fifo_time(rq
)))
2107 cfq_log_cfqq(cfqq
->cfqd
, cfqq
, "fifo=%p", rq
);
2112 cfq_prio_to_maxrq(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
2114 const int base_rq
= cfqd
->cfq_slice_async_rq
;
2116 WARN_ON(cfqq
->ioprio
>= IOPRIO_BE_NR
);
2118 return 2 * base_rq
* (IOPRIO_BE_NR
- cfqq
->ioprio
);
2122 * Must be called with the queue_lock held.
2124 static int cfqq_process_refs(struct cfq_queue
*cfqq
)
2126 int process_refs
, io_refs
;
2128 io_refs
= cfqq
->allocated
[READ
] + cfqq
->allocated
[WRITE
];
2129 process_refs
= cfqq
->ref
- io_refs
;
2130 BUG_ON(process_refs
< 0);
2131 return process_refs
;
2134 static void cfq_setup_merge(struct cfq_queue
*cfqq
, struct cfq_queue
*new_cfqq
)
2136 int process_refs
, new_process_refs
;
2137 struct cfq_queue
*__cfqq
;
2140 * If there are no process references on the new_cfqq, then it is
2141 * unsafe to follow the ->new_cfqq chain as other cfqq's in the
2142 * chain may have dropped their last reference (not just their
2143 * last process reference).
2145 if (!cfqq_process_refs(new_cfqq
))
2148 /* Avoid a circular list and skip interim queue merges */
2149 while ((__cfqq
= new_cfqq
->new_cfqq
)) {
2155 process_refs
= cfqq_process_refs(cfqq
);
2156 new_process_refs
= cfqq_process_refs(new_cfqq
);
2158 * If the process for the cfqq has gone away, there is no
2159 * sense in merging the queues.
2161 if (process_refs
== 0 || new_process_refs
== 0)
2165 * Merge in the direction of the lesser amount of work.
2167 if (new_process_refs
>= process_refs
) {
2168 cfqq
->new_cfqq
= new_cfqq
;
2169 new_cfqq
->ref
+= process_refs
;
2171 new_cfqq
->new_cfqq
= cfqq
;
2172 cfqq
->ref
+= new_process_refs
;
2176 static enum wl_type_t
cfq_choose_wl(struct cfq_data
*cfqd
,
2177 struct cfq_group
*cfqg
, enum wl_prio_t prio
)
2179 struct cfq_queue
*queue
;
2181 bool key_valid
= false;
2182 unsigned long lowest_key
= 0;
2183 enum wl_type_t cur_best
= SYNC_NOIDLE_WORKLOAD
;
2185 for (i
= 0; i
<= SYNC_WORKLOAD
; ++i
) {
2186 /* select the one with lowest rb_key */
2187 queue
= cfq_rb_first(service_tree_for(cfqg
, prio
, i
));
2189 (!key_valid
|| time_before(queue
->rb_key
, lowest_key
))) {
2190 lowest_key
= queue
->rb_key
;
2199 static void choose_service_tree(struct cfq_data
*cfqd
, struct cfq_group
*cfqg
)
2203 struct cfq_rb_root
*st
;
2204 unsigned group_slice
;
2205 enum wl_prio_t original_prio
= cfqd
->serving_prio
;
2207 /* Choose next priority. RT > BE > IDLE */
2208 if (cfq_group_busy_queues_wl(RT_WORKLOAD
, cfqd
, cfqg
))
2209 cfqd
->serving_prio
= RT_WORKLOAD
;
2210 else if (cfq_group_busy_queues_wl(BE_WORKLOAD
, cfqd
, cfqg
))
2211 cfqd
->serving_prio
= BE_WORKLOAD
;
2213 cfqd
->serving_prio
= IDLE_WORKLOAD
;
2214 cfqd
->workload_expires
= jiffies
+ 1;
2218 if (original_prio
!= cfqd
->serving_prio
)
2222 * For RT and BE, we have to choose also the type
2223 * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
2226 st
= service_tree_for(cfqg
, cfqd
->serving_prio
, cfqd
->serving_type
);
2230 * check workload expiration, and that we still have other queues ready
2232 if (count
&& !time_after(jiffies
, cfqd
->workload_expires
))
2236 /* otherwise select new workload type */
2237 cfqd
->serving_type
=
2238 cfq_choose_wl(cfqd
, cfqg
, cfqd
->serving_prio
);
2239 st
= service_tree_for(cfqg
, cfqd
->serving_prio
, cfqd
->serving_type
);
2243 * the workload slice is computed as a fraction of target latency
2244 * proportional to the number of queues in that workload, over
2245 * all the queues in the same priority class
2247 group_slice
= cfq_group_slice(cfqd
, cfqg
);
2249 slice
= group_slice
* count
/
2250 max_t(unsigned, cfqg
->busy_queues_avg
[cfqd
->serving_prio
],
2251 cfq_group_busy_queues_wl(cfqd
->serving_prio
, cfqd
, cfqg
));
2253 if (cfqd
->serving_type
== ASYNC_WORKLOAD
) {
2257 * Async queues are currently system wide. Just taking
2258 * proportion of queues with-in same group will lead to higher
2259 * async ratio system wide as generally root group is going
2260 * to have higher weight. A more accurate thing would be to
2261 * calculate system wide asnc/sync ratio.
2263 tmp
= cfq_target_latency
* cfqg_busy_async_queues(cfqd
, cfqg
);
2264 tmp
= tmp
/cfqd
->busy_queues
;
2265 slice
= min_t(unsigned, slice
, tmp
);
2267 /* async workload slice is scaled down according to
2268 * the sync/async slice ratio. */
2269 slice
= slice
* cfqd
->cfq_slice
[0] / cfqd
->cfq_slice
[1];
2271 /* sync workload slice is at least 2 * cfq_slice_idle */
2272 slice
= max(slice
, 2 * cfqd
->cfq_slice_idle
);
2274 slice
= max_t(unsigned, slice
, CFQ_MIN_TT
);
2275 cfq_log(cfqd
, "workload slice:%d", slice
);
2276 cfqd
->workload_expires
= jiffies
+ slice
;
2279 static struct cfq_group
*cfq_get_next_cfqg(struct cfq_data
*cfqd
)
2281 struct cfq_rb_root
*st
= &cfqd
->grp_service_tree
;
2282 struct cfq_group
*cfqg
;
2284 if (RB_EMPTY_ROOT(&st
->rb
))
2286 cfqg
= cfq_rb_first_group(st
);
2287 update_min_vdisktime(st
);
2291 static void cfq_choose_cfqg(struct cfq_data
*cfqd
)
2293 struct cfq_group
*cfqg
= cfq_get_next_cfqg(cfqd
);
2295 cfqd
->serving_group
= cfqg
;
2297 /* Restore the workload type data */
2298 if (cfqg
->saved_workload_slice
) {
2299 cfqd
->workload_expires
= jiffies
+ cfqg
->saved_workload_slice
;
2300 cfqd
->serving_type
= cfqg
->saved_workload
;
2301 cfqd
->serving_prio
= cfqg
->saved_serving_prio
;
2303 cfqd
->workload_expires
= jiffies
- 1;
2305 choose_service_tree(cfqd
, cfqg
);
2309 * Select a queue for service. If we have a current active queue,
2310 * check whether to continue servicing it, or retrieve and set a new one.
2312 static struct cfq_queue
*cfq_select_queue(struct cfq_data
*cfqd
)
2314 struct cfq_queue
*cfqq
, *new_cfqq
= NULL
;
2316 cfqq
= cfqd
->active_queue
;
2320 if (!cfqd
->rq_queued
)
2324 * We were waiting for group to get backlogged. Expire the queue
2326 if (cfq_cfqq_wait_busy(cfqq
) && !RB_EMPTY_ROOT(&cfqq
->sort_list
))
2330 * The active queue has run out of time, expire it and select new.
2332 if (cfq_slice_used(cfqq
) && !cfq_cfqq_must_dispatch(cfqq
)) {
2334 * If slice had not expired at the completion of last request
2335 * we might not have turned on wait_busy flag. Don't expire
2336 * the queue yet. Allow the group to get backlogged.
2338 * The very fact that we have used the slice, that means we
2339 * have been idling all along on this queue and it should be
2340 * ok to wait for this request to complete.
2342 if (cfqq
->cfqg
->nr_cfqq
== 1 && RB_EMPTY_ROOT(&cfqq
->sort_list
)
2343 && cfqq
->dispatched
&& cfq_should_idle(cfqd
, cfqq
)) {
2347 goto check_group_idle
;
2351 * The active queue has requests and isn't expired, allow it to
2354 if (!RB_EMPTY_ROOT(&cfqq
->sort_list
))
2358 * If another queue has a request waiting within our mean seek
2359 * distance, let it run. The expire code will check for close
2360 * cooperators and put the close queue at the front of the service
2361 * tree. If possible, merge the expiring queue with the new cfqq.
2363 new_cfqq
= cfq_close_cooperator(cfqd
, cfqq
);
2365 if (!cfqq
->new_cfqq
)
2366 cfq_setup_merge(cfqq
, new_cfqq
);
2371 * No requests pending. If the active queue still has requests in
2372 * flight or is idling for a new request, allow either of these
2373 * conditions to happen (or time out) before selecting a new queue.
2375 if (timer_pending(&cfqd
->idle_slice_timer
)) {
2381 * This is a deep seek queue, but the device is much faster than
2382 * the queue can deliver, don't idle
2384 if (CFQQ_SEEKY(cfqq
) && cfq_cfqq_idle_window(cfqq
) &&
2385 (cfq_cfqq_slice_new(cfqq
) ||
2386 (cfqq
->slice_end
- jiffies
> jiffies
- cfqq
->slice_start
))) {
2387 cfq_clear_cfqq_deep(cfqq
);
2388 cfq_clear_cfqq_idle_window(cfqq
);
2391 if (cfqq
->dispatched
&& cfq_should_idle(cfqd
, cfqq
)) {
2397 * If group idle is enabled and there are requests dispatched from
2398 * this group, wait for requests to complete.
2401 if (cfqd
->cfq_group_idle
&& cfqq
->cfqg
->nr_cfqq
== 1 &&
2402 cfqq
->cfqg
->dispatched
&&
2403 !cfq_io_thinktime_big(cfqd
, &cfqq
->cfqg
->ttime
, true)) {
2409 cfq_slice_expired(cfqd
, 0);
2412 * Current queue expired. Check if we have to switch to a new
2416 cfq_choose_cfqg(cfqd
);
2418 cfqq
= cfq_set_active_queue(cfqd
, new_cfqq
);
2423 static int __cfq_forced_dispatch_cfqq(struct cfq_queue
*cfqq
)
2427 while (cfqq
->next_rq
) {
2428 cfq_dispatch_insert(cfqq
->cfqd
->queue
, cfqq
->next_rq
);
2432 BUG_ON(!list_empty(&cfqq
->fifo
));
2434 /* By default cfqq is not expired if it is empty. Do it explicitly */
2435 __cfq_slice_expired(cfqq
->cfqd
, cfqq
, 0);
2440 * Drain our current requests. Used for barriers and when switching
2441 * io schedulers on-the-fly.
2443 static int cfq_forced_dispatch(struct cfq_data
*cfqd
)
2445 struct cfq_queue
*cfqq
;
2448 /* Expire the timeslice of the current active queue first */
2449 cfq_slice_expired(cfqd
, 0);
2450 while ((cfqq
= cfq_get_next_queue_forced(cfqd
)) != NULL
) {
2451 __cfq_set_active_queue(cfqd
, cfqq
);
2452 dispatched
+= __cfq_forced_dispatch_cfqq(cfqq
);
2455 BUG_ON(cfqd
->busy_queues
);
2457 cfq_log(cfqd
, "forced_dispatch=%d", dispatched
);
2461 static inline bool cfq_slice_used_soon(struct cfq_data
*cfqd
,
2462 struct cfq_queue
*cfqq
)
2464 /* the queue hasn't finished any request, can't estimate */
2465 if (cfq_cfqq_slice_new(cfqq
))
2467 if (time_after(jiffies
+ cfqd
->cfq_slice_idle
* cfqq
->dispatched
,
2474 static bool cfq_may_dispatch(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
2476 unsigned int max_dispatch
;
2479 * Drain async requests before we start sync IO
2481 if (cfq_should_idle(cfqd
, cfqq
) && cfqd
->rq_in_flight
[BLK_RW_ASYNC
])
2485 * If this is an async queue and we have sync IO in flight, let it wait
2487 if (cfqd
->rq_in_flight
[BLK_RW_SYNC
] && !cfq_cfqq_sync(cfqq
))
2490 max_dispatch
= max_t(unsigned int, cfqd
->cfq_quantum
/ 2, 1);
2491 if (cfq_class_idle(cfqq
))
2495 * Does this cfqq already have too much IO in flight?
2497 if (cfqq
->dispatched
>= max_dispatch
) {
2498 bool promote_sync
= false;
2500 * idle queue must always only have a single IO in flight
2502 if (cfq_class_idle(cfqq
))
2506 * If there is only one sync queue
2507 * we can ignore async queue here and give the sync
2508 * queue no dispatch limit. The reason is a sync queue can
2509 * preempt async queue, limiting the sync queue doesn't make
2510 * sense. This is useful for aiostress test.
2512 if (cfq_cfqq_sync(cfqq
) && cfqd
->busy_sync_queues
== 1)
2513 promote_sync
= true;
2516 * We have other queues, don't allow more IO from this one
2518 if (cfqd
->busy_queues
> 1 && cfq_slice_used_soon(cfqd
, cfqq
) &&
2523 * Sole queue user, no limit
2525 if (cfqd
->busy_queues
== 1 || promote_sync
)
2529 * Normally we start throttling cfqq when cfq_quantum/2
2530 * requests have been dispatched. But we can drive
2531 * deeper queue depths at the beginning of slice
2532 * subjected to upper limit of cfq_quantum.
2534 max_dispatch
= cfqd
->cfq_quantum
;
2538 * Async queues must wait a bit before being allowed dispatch.
2539 * We also ramp up the dispatch depth gradually for async IO,
2540 * based on the last sync IO we serviced
2542 if (!cfq_cfqq_sync(cfqq
) && cfqd
->cfq_latency
) {
2543 unsigned long last_sync
= jiffies
- cfqd
->last_delayed_sync
;
2546 depth
= last_sync
/ cfqd
->cfq_slice
[1];
2547 if (!depth
&& !cfqq
->dispatched
)
2549 if (depth
< max_dispatch
)
2550 max_dispatch
= depth
;
2554 * If we're below the current max, allow a dispatch
2556 return cfqq
->dispatched
< max_dispatch
;
2560 * Dispatch a request from cfqq, moving them to the request queue
2563 static bool cfq_dispatch_request(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
2567 BUG_ON(RB_EMPTY_ROOT(&cfqq
->sort_list
));
2569 if (!cfq_may_dispatch(cfqd
, cfqq
))
2573 * follow expired path, else get first next available
2575 rq
= cfq_check_fifo(cfqq
);
2580 * insert request into driver dispatch list
2582 cfq_dispatch_insert(cfqd
->queue
, rq
);
2584 if (!cfqd
->active_cic
) {
2585 struct cfq_io_cq
*cic
= RQ_CIC(rq
);
2587 atomic_long_inc(&cic
->icq
.ioc
->refcount
);
2588 cfqd
->active_cic
= cic
;
2595 * Find the cfqq that we need to service and move a request from that to the
2598 static int cfq_dispatch_requests(struct request_queue
*q
, int force
)
2600 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
2601 struct cfq_queue
*cfqq
;
2603 if (!cfqd
->busy_queues
)
2606 if (unlikely(force
))
2607 return cfq_forced_dispatch(cfqd
);
2609 cfqq
= cfq_select_queue(cfqd
);
2614 * Dispatch a request from this cfqq, if it is allowed
2616 if (!cfq_dispatch_request(cfqd
, cfqq
))
2619 cfqq
->slice_dispatch
++;
2620 cfq_clear_cfqq_must_dispatch(cfqq
);
2623 * expire an async queue immediately if it has used up its slice. idle
2624 * queue always expire after 1 dispatch round.
2626 if (cfqd
->busy_queues
> 1 && ((!cfq_cfqq_sync(cfqq
) &&
2627 cfqq
->slice_dispatch
>= cfq_prio_to_maxrq(cfqd
, cfqq
)) ||
2628 cfq_class_idle(cfqq
))) {
2629 cfqq
->slice_end
= jiffies
+ 1;
2630 cfq_slice_expired(cfqd
, 0);
2633 cfq_log_cfqq(cfqd
, cfqq
, "dispatched a request");
2638 * task holds one reference to the queue, dropped when task exits. each rq
2639 * in-flight on this queue also holds a reference, dropped when rq is freed.
2641 * Each cfq queue took a reference on the parent group. Drop it now.
2642 * queue lock must be held here.
2644 static void cfq_put_queue(struct cfq_queue
*cfqq
)
2646 struct cfq_data
*cfqd
= cfqq
->cfqd
;
2647 struct cfq_group
*cfqg
;
2649 BUG_ON(cfqq
->ref
<= 0);
2655 cfq_log_cfqq(cfqd
, cfqq
, "put_queue");
2656 BUG_ON(rb_first(&cfqq
->sort_list
));
2657 BUG_ON(cfqq
->allocated
[READ
] + cfqq
->allocated
[WRITE
]);
2660 if (unlikely(cfqd
->active_queue
== cfqq
)) {
2661 __cfq_slice_expired(cfqd
, cfqq
, 0);
2662 cfq_schedule_dispatch(cfqd
);
2665 BUG_ON(cfq_cfqq_on_rr(cfqq
));
2666 kmem_cache_free(cfq_pool
, cfqq
);
2670 static void cfq_icq_free_rcu(struct rcu_head
*head
)
2672 kmem_cache_free(cfq_icq_pool
,
2673 icq_to_cic(container_of(head
, struct io_cq
, rcu_head
)));
2676 static void cfq_icq_free(struct io_cq
*icq
)
2678 call_rcu(&icq
->rcu_head
, cfq_icq_free_rcu
);
2681 static void cfq_release_icq(struct io_cq
*icq
)
2683 struct io_context
*ioc
= icq
->ioc
;
2685 radix_tree_delete(&ioc
->icq_tree
, icq
->q
->id
);
2686 hlist_del(&icq
->ioc_node
);
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_icq(struct io_cq
*icq
)
2725 struct cfq_io_cq
*cic
= icq_to_cic(icq
);
2726 struct cfq_data
*cfqd
= cic_to_cfqd(cic
);
2727 struct io_context
*ioc
= icq
->ioc
;
2729 list_del_init(&icq
->q_node
);
2732 * Both setting lookup hint to and clearing it from @icq are done
2733 * under queue_lock. If it's not pointing to @icq now, it never
2734 * will. Hint assignment itself can race safely.
2736 if (rcu_dereference_raw(ioc
->icq_hint
) == icq
)
2737 rcu_assign_pointer(ioc
->icq_hint
, 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_cq
*cfq_alloc_cic(struct cfq_data
*cfqd
, gfp_t gfp_mask
)
2752 struct cfq_io_cq
*cic
;
2754 cic
= kmem_cache_alloc_node(cfq_icq_pool
, gfp_mask
| __GFP_ZERO
,
2757 cic
->ttime
.last_end_request
= jiffies
;
2758 INIT_LIST_HEAD(&cic
->icq
.q_node
);
2759 INIT_HLIST_NODE(&cic
->icq
.ioc_node
);
2760 cic
->icq
.exit
= cfq_exit_icq
;
2761 cic
->icq
.release
= cfq_release_icq
;
2767 static void cfq_init_prio_data(struct cfq_queue
*cfqq
, struct io_context
*ioc
)
2769 struct task_struct
*tsk
= current
;
2772 if (!cfq_cfqq_prio_changed(cfqq
))
2775 ioprio_class
= IOPRIO_PRIO_CLASS(ioc
->ioprio
);
2776 switch (ioprio_class
) {
2778 printk(KERN_ERR
"cfq: bad prio %x\n", ioprio_class
);
2779 case IOPRIO_CLASS_NONE
:
2781 * no prio set, inherit CPU scheduling settings
2783 cfqq
->ioprio
= task_nice_ioprio(tsk
);
2784 cfqq
->ioprio_class
= task_nice_ioclass(tsk
);
2786 case IOPRIO_CLASS_RT
:
2787 cfqq
->ioprio
= task_ioprio(ioc
);
2788 cfqq
->ioprio_class
= IOPRIO_CLASS_RT
;
2790 case IOPRIO_CLASS_BE
:
2791 cfqq
->ioprio
= task_ioprio(ioc
);
2792 cfqq
->ioprio_class
= IOPRIO_CLASS_BE
;
2794 case IOPRIO_CLASS_IDLE
:
2795 cfqq
->ioprio_class
= IOPRIO_CLASS_IDLE
;
2797 cfq_clear_cfqq_idle_window(cfqq
);
2802 * keep track of original prio settings in case we have to temporarily
2803 * elevate the priority of this queue
2805 cfqq
->org_ioprio
= cfqq
->ioprio
;
2806 cfq_clear_cfqq_prio_changed(cfqq
);
2809 static void changed_ioprio(struct cfq_io_cq
*cic
)
2811 struct cfq_data
*cfqd
= cic_to_cfqd(cic
);
2812 struct cfq_queue
*cfqq
;
2814 if (unlikely(!cfqd
))
2817 cfqq
= cic
->cfqq
[BLK_RW_ASYNC
];
2819 struct cfq_queue
*new_cfqq
;
2820 new_cfqq
= cfq_get_queue(cfqd
, BLK_RW_ASYNC
, cic
->icq
.ioc
,
2823 cic
->cfqq
[BLK_RW_ASYNC
] = new_cfqq
;
2824 cfq_put_queue(cfqq
);
2828 cfqq
= cic
->cfqq
[BLK_RW_SYNC
];
2830 cfq_mark_cfqq_prio_changed(cfqq
);
2833 static void cfq_init_cfqq(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
2834 pid_t pid
, bool is_sync
)
2836 RB_CLEAR_NODE(&cfqq
->rb_node
);
2837 RB_CLEAR_NODE(&cfqq
->p_node
);
2838 INIT_LIST_HEAD(&cfqq
->fifo
);
2843 cfq_mark_cfqq_prio_changed(cfqq
);
2846 if (!cfq_class_idle(cfqq
))
2847 cfq_mark_cfqq_idle_window(cfqq
);
2848 cfq_mark_cfqq_sync(cfqq
);
2853 #ifdef CONFIG_CFQ_GROUP_IOSCHED
2854 static void changed_cgroup(struct cfq_io_cq
*cic
)
2856 struct cfq_queue
*sync_cfqq
= cic_to_cfqq(cic
, 1);
2857 struct cfq_data
*cfqd
= cic_to_cfqd(cic
);
2858 struct request_queue
*q
;
2860 if (unlikely(!cfqd
))
2867 * Drop reference to sync queue. A new sync queue will be
2868 * assigned in new group upon arrival of a fresh request.
2870 cfq_log_cfqq(cfqd
, sync_cfqq
, "changed cgroup");
2871 cic_set_cfqq(cic
, NULL
, 1);
2872 cfq_put_queue(sync_cfqq
);
2875 #endif /* CONFIG_CFQ_GROUP_IOSCHED */
2877 static struct cfq_queue
*
2878 cfq_find_alloc_queue(struct cfq_data
*cfqd
, bool is_sync
,
2879 struct io_context
*ioc
, gfp_t gfp_mask
)
2881 struct cfq_queue
*cfqq
, *new_cfqq
= NULL
;
2882 struct cfq_io_cq
*cic
;
2883 struct cfq_group
*cfqg
;
2886 cfqg
= cfq_get_cfqg(cfqd
);
2887 cic
= cfq_cic_lookup(cfqd
, ioc
);
2888 /* cic always exists here */
2889 cfqq
= cic_to_cfqq(cic
, is_sync
);
2892 * Always try a new alloc if we fell back to the OOM cfqq
2893 * originally, since it should just be a temporary situation.
2895 if (!cfqq
|| cfqq
== &cfqd
->oom_cfqq
) {
2900 } else if (gfp_mask
& __GFP_WAIT
) {
2901 spin_unlock_irq(cfqd
->queue
->queue_lock
);
2902 new_cfqq
= kmem_cache_alloc_node(cfq_pool
,
2903 gfp_mask
| __GFP_ZERO
,
2905 spin_lock_irq(cfqd
->queue
->queue_lock
);
2909 cfqq
= kmem_cache_alloc_node(cfq_pool
,
2910 gfp_mask
| __GFP_ZERO
,
2915 cfq_init_cfqq(cfqd
, cfqq
, current
->pid
, is_sync
);
2916 cfq_init_prio_data(cfqq
, ioc
);
2917 cfq_link_cfqq_cfqg(cfqq
, cfqg
);
2918 cfq_log_cfqq(cfqd
, cfqq
, "alloced");
2920 cfqq
= &cfqd
->oom_cfqq
;
2924 kmem_cache_free(cfq_pool
, new_cfqq
);
2929 static struct cfq_queue
**
2930 cfq_async_queue_prio(struct cfq_data
*cfqd
, int ioprio_class
, int ioprio
)
2932 switch (ioprio_class
) {
2933 case IOPRIO_CLASS_RT
:
2934 return &cfqd
->async_cfqq
[0][ioprio
];
2935 case IOPRIO_CLASS_BE
:
2936 return &cfqd
->async_cfqq
[1][ioprio
];
2937 case IOPRIO_CLASS_IDLE
:
2938 return &cfqd
->async_idle_cfqq
;
2944 static struct cfq_queue
*
2945 cfq_get_queue(struct cfq_data
*cfqd
, bool is_sync
, struct io_context
*ioc
,
2948 const int ioprio
= task_ioprio(ioc
);
2949 const int ioprio_class
= task_ioprio_class(ioc
);
2950 struct cfq_queue
**async_cfqq
= NULL
;
2951 struct cfq_queue
*cfqq
= NULL
;
2954 async_cfqq
= cfq_async_queue_prio(cfqd
, ioprio_class
, ioprio
);
2959 cfqq
= cfq_find_alloc_queue(cfqd
, is_sync
, ioc
, gfp_mask
);
2962 * pin the queue now that it's allocated, scheduler exit will prune it
2964 if (!is_sync
&& !(*async_cfqq
)) {
2974 * cfq_cic_lookup - lookup cfq_io_cq
2975 * @cfqd: the associated cfq_data
2976 * @ioc: the associated io_context
2978 * Look up cfq_io_cq associated with @cfqd - @ioc pair. Must be called
2979 * with queue_lock held.
2981 static struct cfq_io_cq
*
2982 cfq_cic_lookup(struct cfq_data
*cfqd
, struct io_context
*ioc
)
2984 struct request_queue
*q
= cfqd
->queue
;
2987 lockdep_assert_held(cfqd
->queue
->queue_lock
);
2992 * icq'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 icq which points to us, it's guaranteed to be valid.
2998 icq
= rcu_dereference(ioc
->icq_hint
);
2999 if (icq
&& icq
->q
== q
)
3002 icq
= radix_tree_lookup(&ioc
->icq_tree
, cfqd
->queue
->id
);
3003 if (icq
&& icq
->q
== q
)
3004 rcu_assign_pointer(ioc
->icq_hint
, icq
); /* allowed to race */
3009 return icq_to_cic(icq
);
3013 * cfq_create_cic - create and link a cfq_io_cq
3014 * @cfqd: cfqd of interest
3015 * @gfp_mask: allocation mask
3017 * Make sure cfq_io_cq linking %current->io_context and @cfqd exists. If
3018 * 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 io_cq
*icq
= NULL
;
3024 struct cfq_io_cq
*cic
;
3025 struct io_context
*ioc
;
3028 might_sleep_if(gfp_mask
& __GFP_WAIT
);
3030 /* allocate stuff */
3031 ioc
= create_io_context(current
, gfp_mask
, q
->node
);
3035 cic
= cfq_alloc_cic(cfqd
, gfp_mask
);
3040 ret
= radix_tree_preload(gfp_mask
);
3045 icq
->q
= cfqd
->queue
;
3047 /* lock both q and ioc and try to link @icq */
3048 spin_lock_irq(q
->queue_lock
);
3049 spin_lock(&ioc
->lock
);
3051 ret
= radix_tree_insert(&ioc
->icq_tree
, q
->id
, icq
);
3053 hlist_add_head(&icq
->ioc_node
, &ioc
->icq_list
);
3054 list_add(&icq
->q_node
, &q
->icq_list
);
3056 } else if (ret
== -EEXIST
) {
3057 /* someone else already did it */
3061 spin_unlock(&ioc
->lock
);
3062 spin_unlock_irq(q
->queue_lock
);
3064 radix_tree_preload_end();
3067 printk(KERN_ERR
"cfq: icq link failed!\n");
3074 * cfq_get_cic - acquire cfq_io_cq and bump refcnt on io_context
3075 * @cfqd: cfqd to setup cic for
3076 * @gfp_mask: allocation mask
3078 * Return cfq_io_cq associating @cfqd and %current->io_context and
3079 * bump refcnt on io_context. If ioc or cic doesn't exist, they're created
3082 * Must be called under queue_lock which may be released and re-acquired.
3083 * This function also may sleep depending on @gfp_mask.
3085 static struct cfq_io_cq
*cfq_get_cic(struct cfq_data
*cfqd
, gfp_t gfp_mask
)
3087 struct request_queue
*q
= cfqd
->queue
;
3088 struct cfq_io_cq
*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
->icq
.changed
)) {
3115 if (test_and_clear_bit(ICQ_IOPRIO_CHANGED
, &cic
->icq
.changed
))
3116 changed_ioprio(cic
);
3117 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3118 if (test_and_clear_bit(ICQ_CGROUP_CHANGED
, &cic
->icq
.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_cq
*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_cq
*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
->icq
.ioc
->nr_tasks
) ||
3195 !cfqd
->cfq_slice_idle
||
3196 (!cfq_cfqq_deep(cfqq
) && CFQQ_SEEKY(cfqq
)))
3198 else if (sample_valid(cic
->ttime
.ttime_samples
)) {
3199 if (cic
->ttime
.ttime_mean
> cfqd
->cfq_slice_idle
)
3205 if (old_idle
!= enable_idle
) {
3206 cfq_log_cfqq(cfqd
, cfqq
, "idle=%d", enable_idle
);
3208 cfq_mark_cfqq_idle_window(cfqq
);
3210 cfq_clear_cfqq_idle_window(cfqq
);
3215 * Check if new_cfqq should preempt the currently active queue. Return 0 for
3216 * no or if we aren't sure, a 1 will cause a preempt.
3219 cfq_should_preempt(struct cfq_data
*cfqd
, struct cfq_queue
*new_cfqq
,
3222 struct cfq_queue
*cfqq
;
3224 cfqq
= cfqd
->active_queue
;
3228 if (cfq_class_idle(new_cfqq
))
3231 if (cfq_class_idle(cfqq
))
3235 * Don't allow a non-RT request to preempt an ongoing RT cfqq timeslice.
3237 if (cfq_class_rt(cfqq
) && !cfq_class_rt(new_cfqq
))
3241 * if the new request is sync, but the currently running queue is
3242 * not, let the sync request have priority.
3244 if (rq_is_sync(rq
) && !cfq_cfqq_sync(cfqq
))
3247 if (new_cfqq
->cfqg
!= cfqq
->cfqg
)
3250 if (cfq_slice_used(cfqq
))
3253 /* Allow preemption only if we are idling on sync-noidle tree */
3254 if (cfqd
->serving_type
== SYNC_NOIDLE_WORKLOAD
&&
3255 cfqq_type(new_cfqq
) == SYNC_NOIDLE_WORKLOAD
&&
3256 new_cfqq
->service_tree
->count
== 2 &&
3257 RB_EMPTY_ROOT(&cfqq
->sort_list
))
3261 * So both queues are sync. Let the new request get disk time if
3262 * it's a metadata request and the current queue is doing regular IO.
3264 if ((rq
->cmd_flags
& REQ_PRIO
) && !cfqq
->prio_pending
)
3268 * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
3270 if (cfq_class_rt(new_cfqq
) && !cfq_class_rt(cfqq
))
3273 /* An idle queue should not be idle now for some reason */
3274 if (RB_EMPTY_ROOT(&cfqq
->sort_list
) && !cfq_should_idle(cfqd
, cfqq
))
3277 if (!cfqd
->active_cic
|| !cfq_cfqq_wait_request(cfqq
))
3281 * if this request is as-good as one we would expect from the
3282 * current cfqq, let it preempt
3284 if (cfq_rq_close(cfqd
, cfqq
, rq
))
3291 * cfqq preempts the active queue. if we allowed preempt with no slice left,
3292 * let it have half of its nominal slice.
3294 static void cfq_preempt_queue(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
3296 struct cfq_queue
*old_cfqq
= cfqd
->active_queue
;
3298 cfq_log_cfqq(cfqd
, cfqq
, "preempt");
3299 cfq_slice_expired(cfqd
, 1);
3302 * workload type is changed, don't save slice, otherwise preempt
3305 if (cfqq_type(old_cfqq
) != cfqq_type(cfqq
))
3306 cfqq
->cfqg
->saved_workload_slice
= 0;
3309 * Put the new queue at the front of the of the current list,
3310 * so we know that it will be selected next.
3312 BUG_ON(!cfq_cfqq_on_rr(cfqq
));
3314 cfq_service_tree_add(cfqd
, cfqq
, 1);
3316 cfqq
->slice_end
= 0;
3317 cfq_mark_cfqq_slice_new(cfqq
);
3321 * Called when a new fs request (rq) is added (to cfqq). Check if there's
3322 * something we should do about it
3325 cfq_rq_enqueued(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
3328 struct cfq_io_cq
*cic
= RQ_CIC(rq
);
3331 if (rq
->cmd_flags
& REQ_PRIO
)
3332 cfqq
->prio_pending
++;
3334 cfq_update_io_thinktime(cfqd
, cfqq
, cic
);
3335 cfq_update_io_seektime(cfqd
, cfqq
, rq
);
3336 cfq_update_idle_window(cfqd
, cfqq
, cic
);
3338 cfqq
->last_request_pos
= blk_rq_pos(rq
) + blk_rq_sectors(rq
);
3340 if (cfqq
== cfqd
->active_queue
) {
3342 * Remember that we saw a request from this process, but
3343 * don't start queuing just yet. Otherwise we risk seeing lots
3344 * of tiny requests, because we disrupt the normal plugging
3345 * and merging. If the request is already larger than a single
3346 * page, let it rip immediately. For that case we assume that
3347 * merging is already done. Ditto for a busy system that
3348 * has other work pending, don't risk delaying until the
3349 * idle timer unplug to continue working.
3351 if (cfq_cfqq_wait_request(cfqq
)) {
3352 if (blk_rq_bytes(rq
) > PAGE_CACHE_SIZE
||
3353 cfqd
->busy_queues
> 1) {
3354 cfq_del_timer(cfqd
, cfqq
);
3355 cfq_clear_cfqq_wait_request(cfqq
);
3356 __blk_run_queue(cfqd
->queue
);
3358 cfq_blkiocg_update_idle_time_stats(
3360 cfq_mark_cfqq_must_dispatch(cfqq
);
3363 } else if (cfq_should_preempt(cfqd
, cfqq
, rq
)) {
3365 * not the active queue - expire current slice if it is
3366 * idle and has expired it's mean thinktime or this new queue
3367 * has some old slice time left and is of higher priority or
3368 * this new queue is RT and the current one is BE
3370 cfq_preempt_queue(cfqd
, cfqq
);
3371 __blk_run_queue(cfqd
->queue
);
3375 static void cfq_insert_request(struct request_queue
*q
, struct request
*rq
)
3377 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
3378 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
3380 cfq_log_cfqq(cfqd
, cfqq
, "insert_request");
3381 cfq_init_prio_data(cfqq
, RQ_CIC(rq
)->icq
.ioc
);
3383 rq_set_fifo_time(rq
, jiffies
+ cfqd
->cfq_fifo_expire
[rq_is_sync(rq
)]);
3384 list_add_tail(&rq
->queuelist
, &cfqq
->fifo
);
3386 cfq_blkiocg_update_io_add_stats(&(RQ_CFQG(rq
))->blkg
,
3387 &cfqd
->serving_group
->blkg
, rq_data_dir(rq
),
3389 cfq_rq_enqueued(cfqd
, cfqq
, rq
);
3393 * Update hw_tag based on peak queue depth over 50 samples under
3396 static void cfq_update_hw_tag(struct cfq_data
*cfqd
)
3398 struct cfq_queue
*cfqq
= cfqd
->active_queue
;
3400 if (cfqd
->rq_in_driver
> cfqd
->hw_tag_est_depth
)
3401 cfqd
->hw_tag_est_depth
= cfqd
->rq_in_driver
;
3403 if (cfqd
->hw_tag
== 1)
3406 if (cfqd
->rq_queued
<= CFQ_HW_QUEUE_MIN
&&
3407 cfqd
->rq_in_driver
<= CFQ_HW_QUEUE_MIN
)
3411 * If active queue hasn't enough requests and can idle, cfq might not
3412 * dispatch sufficient requests to hardware. Don't zero hw_tag in this
3415 if (cfqq
&& cfq_cfqq_idle_window(cfqq
) &&
3416 cfqq
->dispatched
+ cfqq
->queued
[0] + cfqq
->queued
[1] <
3417 CFQ_HW_QUEUE_MIN
&& cfqd
->rq_in_driver
< CFQ_HW_QUEUE_MIN
)
3420 if (cfqd
->hw_tag_samples
++ < 50)
3423 if (cfqd
->hw_tag_est_depth
>= CFQ_HW_QUEUE_MIN
)
3429 static bool cfq_should_wait_busy(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
3431 struct cfq_io_cq
*cic
= cfqd
->active_cic
;
3433 /* If the queue already has requests, don't wait */
3434 if (!RB_EMPTY_ROOT(&cfqq
->sort_list
))
3437 /* If there are other queues in the group, don't wait */
3438 if (cfqq
->cfqg
->nr_cfqq
> 1)
3441 /* the only queue in the group, but think time is big */
3442 if (cfq_io_thinktime_big(cfqd
, &cfqq
->cfqg
->ttime
, true))
3445 if (cfq_slice_used(cfqq
))
3448 /* if slice left is less than think time, wait busy */
3449 if (cic
&& sample_valid(cic
->ttime
.ttime_samples
)
3450 && (cfqq
->slice_end
- jiffies
< cic
->ttime
.ttime_mean
))
3454 * If think times is less than a jiffy than ttime_mean=0 and above
3455 * will not be true. It might happen that slice has not expired yet
3456 * but will expire soon (4-5 ns) during select_queue(). To cover the
3457 * case where think time is less than a jiffy, mark the queue wait
3458 * busy if only 1 jiffy is left in the slice.
3460 if (cfqq
->slice_end
- jiffies
== 1)
3466 static void cfq_completed_request(struct request_queue
*q
, struct request
*rq
)
3468 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
3469 struct cfq_data
*cfqd
= cfqq
->cfqd
;
3470 const int sync
= rq_is_sync(rq
);
3474 cfq_log_cfqq(cfqd
, cfqq
, "complete rqnoidle %d",
3475 !!(rq
->cmd_flags
& REQ_NOIDLE
));
3477 cfq_update_hw_tag(cfqd
);
3479 WARN_ON(!cfqd
->rq_in_driver
);
3480 WARN_ON(!cfqq
->dispatched
);
3481 cfqd
->rq_in_driver
--;
3483 (RQ_CFQG(rq
))->dispatched
--;
3484 cfq_blkiocg_update_completion_stats(&cfqq
->cfqg
->blkg
,
3485 rq_start_time_ns(rq
), rq_io_start_time_ns(rq
),
3486 rq_data_dir(rq
), rq_is_sync(rq
));
3488 cfqd
->rq_in_flight
[cfq_cfqq_sync(cfqq
)]--;
3491 struct cfq_rb_root
*service_tree
;
3493 RQ_CIC(rq
)->ttime
.last_end_request
= now
;
3495 if (cfq_cfqq_on_rr(cfqq
))
3496 service_tree
= cfqq
->service_tree
;
3498 service_tree
= service_tree_for(cfqq
->cfqg
,
3499 cfqq_prio(cfqq
), cfqq_type(cfqq
));
3500 service_tree
->ttime
.last_end_request
= now
;
3501 if (!time_after(rq
->start_time
+ cfqd
->cfq_fifo_expire
[1], now
))
3502 cfqd
->last_delayed_sync
= now
;
3505 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3506 cfqq
->cfqg
->ttime
.last_end_request
= now
;
3510 * If this is the active queue, check if it needs to be expired,
3511 * or if we want to idle in case it has no pending requests.
3513 if (cfqd
->active_queue
== cfqq
) {
3514 const bool cfqq_empty
= RB_EMPTY_ROOT(&cfqq
->sort_list
);
3516 if (cfq_cfqq_slice_new(cfqq
)) {
3517 cfq_set_prio_slice(cfqd
, cfqq
);
3518 cfq_clear_cfqq_slice_new(cfqq
);
3522 * Should we wait for next request to come in before we expire
3525 if (cfq_should_wait_busy(cfqd
, cfqq
)) {
3526 unsigned long extend_sl
= cfqd
->cfq_slice_idle
;
3527 if (!cfqd
->cfq_slice_idle
)
3528 extend_sl
= cfqd
->cfq_group_idle
;
3529 cfqq
->slice_end
= jiffies
+ extend_sl
;
3530 cfq_mark_cfqq_wait_busy(cfqq
);
3531 cfq_log_cfqq(cfqd
, cfqq
, "will busy wait");
3535 * Idling is not enabled on:
3537 * - idle-priority queues
3539 * - queues with still some requests queued
3540 * - when there is a close cooperator
3542 if (cfq_slice_used(cfqq
) || cfq_class_idle(cfqq
))
3543 cfq_slice_expired(cfqd
, 1);
3544 else if (sync
&& cfqq_empty
&&
3545 !cfq_close_cooperator(cfqd
, cfqq
)) {
3546 cfq_arm_slice_timer(cfqd
);
3550 if (!cfqd
->rq_in_driver
)
3551 cfq_schedule_dispatch(cfqd
);
3554 static inline int __cfq_may_queue(struct cfq_queue
*cfqq
)
3556 if (cfq_cfqq_wait_request(cfqq
) && !cfq_cfqq_must_alloc_slice(cfqq
)) {
3557 cfq_mark_cfqq_must_alloc_slice(cfqq
);
3558 return ELV_MQUEUE_MUST
;
3561 return ELV_MQUEUE_MAY
;
3564 static int cfq_may_queue(struct request_queue
*q
, int rw
)
3566 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
3567 struct task_struct
*tsk
= current
;
3568 struct cfq_io_cq
*cic
;
3569 struct cfq_queue
*cfqq
;
3572 * don't force setup of a queue from here, as a call to may_queue
3573 * does not necessarily imply that a request actually will be queued.
3574 * so just lookup a possibly existing queue, or return 'may queue'
3577 cic
= cfq_cic_lookup(cfqd
, tsk
->io_context
);
3579 return ELV_MQUEUE_MAY
;
3581 cfqq
= cic_to_cfqq(cic
, rw_is_sync(rw
));
3583 cfq_init_prio_data(cfqq
, cic
->icq
.ioc
);
3585 return __cfq_may_queue(cfqq
);
3588 return ELV_MQUEUE_MAY
;
3592 * queue lock held here
3594 static void cfq_put_request(struct request
*rq
)
3596 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
3599 const int rw
= rq_data_dir(rq
);
3601 BUG_ON(!cfqq
->allocated
[rw
]);
3602 cfqq
->allocated
[rw
]--;
3604 put_io_context(RQ_CIC(rq
)->icq
.ioc
, cfqq
->cfqd
->queue
);
3606 /* Put down rq reference on cfqg */
3607 cfq_put_cfqg(RQ_CFQG(rq
));
3608 rq
->elv
.priv
[0] = NULL
;
3609 rq
->elv
.priv
[1] = NULL
;
3611 cfq_put_queue(cfqq
);
3615 static struct cfq_queue
*
3616 cfq_merge_cfqqs(struct cfq_data
*cfqd
, struct cfq_io_cq
*cic
,
3617 struct cfq_queue
*cfqq
)
3619 cfq_log_cfqq(cfqd
, cfqq
, "merging with queue %p", cfqq
->new_cfqq
);
3620 cic_set_cfqq(cic
, cfqq
->new_cfqq
, 1);
3621 cfq_mark_cfqq_coop(cfqq
->new_cfqq
);
3622 cfq_put_queue(cfqq
);
3623 return cic_to_cfqq(cic
, 1);
3627 * Returns NULL if a new cfqq should be allocated, or the old cfqq if this
3628 * was the last process referring to said cfqq.
3630 static struct cfq_queue
*
3631 split_cfqq(struct cfq_io_cq
*cic
, struct cfq_queue
*cfqq
)
3633 if (cfqq_process_refs(cfqq
) == 1) {
3634 cfqq
->pid
= current
->pid
;
3635 cfq_clear_cfqq_coop(cfqq
);
3636 cfq_clear_cfqq_split_coop(cfqq
);
3640 cic_set_cfqq(cic
, NULL
, 1);
3642 cfq_put_cooperator(cfqq
);
3644 cfq_put_queue(cfqq
);
3648 * Allocate cfq data structures associated with this request.
3651 cfq_set_request(struct request_queue
*q
, struct request
*rq
, gfp_t gfp_mask
)
3653 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
3654 struct cfq_io_cq
*cic
;
3655 const int rw
= rq_data_dir(rq
);
3656 const bool is_sync
= rq_is_sync(rq
);
3657 struct cfq_queue
*cfqq
;
3659 might_sleep_if(gfp_mask
& __GFP_WAIT
);
3661 spin_lock_irq(q
->queue_lock
);
3662 cic
= cfq_get_cic(cfqd
, gfp_mask
);
3667 cfqq
= cic_to_cfqq(cic
, is_sync
);
3668 if (!cfqq
|| cfqq
== &cfqd
->oom_cfqq
) {
3669 cfqq
= cfq_get_queue(cfqd
, is_sync
, cic
->icq
.ioc
, gfp_mask
);
3670 cic_set_cfqq(cic
, cfqq
, is_sync
);
3673 * If the queue was seeky for too long, break it apart.
3675 if (cfq_cfqq_coop(cfqq
) && cfq_cfqq_split_coop(cfqq
)) {
3676 cfq_log_cfqq(cfqd
, cfqq
, "breaking apart cfqq");
3677 cfqq
= split_cfqq(cic
, cfqq
);
3683 * Check to see if this queue is scheduled to merge with
3684 * another, closely cooperating queue. The merging of
3685 * queues happens here as it must be done in process context.
3686 * The reference on new_cfqq was taken in merge_cfqqs.
3689 cfqq
= cfq_merge_cfqqs(cfqd
, cic
, cfqq
);
3692 cfqq
->allocated
[rw
]++;
3695 rq
->elv
.icq
= &cic
->icq
;
3696 rq
->elv
.priv
[0] = cfqq
;
3697 rq
->elv
.priv
[1] = cfq_ref_get_cfqg(cfqq
->cfqg
);
3698 spin_unlock_irq(q
->queue_lock
);
3702 cfq_schedule_dispatch(cfqd
);
3703 spin_unlock_irq(q
->queue_lock
);
3704 cfq_log(cfqd
, "set_request fail");
3708 static void cfq_kick_queue(struct work_struct
*work
)
3710 struct cfq_data
*cfqd
=
3711 container_of(work
, struct cfq_data
, unplug_work
);
3712 struct request_queue
*q
= cfqd
->queue
;
3714 spin_lock_irq(q
->queue_lock
);
3715 __blk_run_queue(cfqd
->queue
);
3716 spin_unlock_irq(q
->queue_lock
);
3720 * Timer running if the active_queue is currently idling inside its time slice
3722 static void cfq_idle_slice_timer(unsigned long data
)
3724 struct cfq_data
*cfqd
= (struct cfq_data
*) data
;
3725 struct cfq_queue
*cfqq
;
3726 unsigned long flags
;
3729 cfq_log(cfqd
, "idle timer fired");
3731 spin_lock_irqsave(cfqd
->queue
->queue_lock
, flags
);
3733 cfqq
= cfqd
->active_queue
;
3738 * We saw a request before the queue expired, let it through
3740 if (cfq_cfqq_must_dispatch(cfqq
))
3746 if (cfq_slice_used(cfqq
))
3750 * only expire and reinvoke request handler, if there are
3751 * other queues with pending requests
3753 if (!cfqd
->busy_queues
)
3757 * not expired and it has a request pending, let it dispatch
3759 if (!RB_EMPTY_ROOT(&cfqq
->sort_list
))
3763 * Queue depth flag is reset only when the idle didn't succeed
3765 cfq_clear_cfqq_deep(cfqq
);
3768 cfq_slice_expired(cfqd
, timed_out
);
3770 cfq_schedule_dispatch(cfqd
);
3772 spin_unlock_irqrestore(cfqd
->queue
->queue_lock
, flags
);
3775 static void cfq_shutdown_timer_wq(struct cfq_data
*cfqd
)
3777 del_timer_sync(&cfqd
->idle_slice_timer
);
3778 cancel_work_sync(&cfqd
->unplug_work
);
3781 static void cfq_put_async_queues(struct cfq_data
*cfqd
)
3785 for (i
= 0; i
< IOPRIO_BE_NR
; i
++) {
3786 if (cfqd
->async_cfqq
[0][i
])
3787 cfq_put_queue(cfqd
->async_cfqq
[0][i
]);
3788 if (cfqd
->async_cfqq
[1][i
])
3789 cfq_put_queue(cfqd
->async_cfqq
[1][i
]);
3792 if (cfqd
->async_idle_cfqq
)
3793 cfq_put_queue(cfqd
->async_idle_cfqq
);
3796 static void cfq_exit_queue(struct elevator_queue
*e
)
3798 struct cfq_data
*cfqd
= e
->elevator_data
;
3799 struct request_queue
*q
= cfqd
->queue
;
3802 cfq_shutdown_timer_wq(cfqd
);
3804 spin_lock_irq(q
->queue_lock
);
3806 if (cfqd
->active_queue
)
3807 __cfq_slice_expired(cfqd
, cfqd
->active_queue
, 0);
3809 while (!list_empty(&q
->icq_list
)) {
3810 struct io_cq
*icq
= list_entry(q
->icq_list
.next
,
3811 struct io_cq
, q_node
);
3812 struct io_context
*ioc
= icq
->ioc
;
3814 spin_lock(&ioc
->lock
);
3816 cfq_release_icq(icq
);
3817 spin_unlock(&ioc
->lock
);
3820 cfq_put_async_queues(cfqd
);
3821 cfq_release_cfq_groups(cfqd
);
3824 * If there are groups which we could not unlink from blkcg list,
3825 * wait for a rcu period for them to be freed.
3827 if (cfqd
->nr_blkcg_linked_grps
)
3830 spin_unlock_irq(q
->queue_lock
);
3832 cfq_shutdown_timer_wq(cfqd
);
3835 * Wait for cfqg->blkg->key accessors to exit their grace periods.
3836 * Do this wait only if there are other unlinked groups out
3837 * there. This can happen if cgroup deletion path claimed the
3838 * responsibility of cleaning up a group before queue cleanup code
3841 * Do not call synchronize_rcu() unconditionally as there are drivers
3842 * which create/delete request queue hundreds of times during scan/boot
3843 * and synchronize_rcu() can take significant time and slow down boot.
3848 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3849 /* Free up per cpu stats for root group */
3850 free_percpu(cfqd
->root_group
.blkg
.stats_cpu
);
3855 static void *cfq_init_queue(struct request_queue
*q
)
3857 struct cfq_data
*cfqd
;
3859 struct cfq_group
*cfqg
;
3860 struct cfq_rb_root
*st
;
3862 cfqd
= kmalloc_node(sizeof(*cfqd
), GFP_KERNEL
| __GFP_ZERO
, q
->node
);
3866 /* Init root service tree */
3867 cfqd
->grp_service_tree
= CFQ_RB_ROOT
;
3869 /* Init root group */
3870 cfqg
= &cfqd
->root_group
;
3871 for_each_cfqg_st(cfqg
, i
, j
, st
)
3873 RB_CLEAR_NODE(&cfqg
->rb_node
);
3875 /* Give preference to root group over other groups */
3876 cfqg
->weight
= 2*BLKIO_WEIGHT_DEFAULT
;
3878 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3880 * Set root group reference to 2. One reference will be dropped when
3881 * all groups on cfqd->cfqg_list are being deleted during queue exit.
3882 * Other reference will remain there as we don't want to delete this
3883 * group as it is statically allocated and gets destroyed when
3884 * throtl_data goes away.
3888 if (blkio_alloc_blkg_stats(&cfqg
->blkg
)) {
3896 cfq_blkiocg_add_blkio_group(&blkio_root_cgroup
, &cfqg
->blkg
,
3899 cfqd
->nr_blkcg_linked_grps
++;
3901 /* Add group on cfqd->cfqg_list */
3902 hlist_add_head(&cfqg
->cfqd_node
, &cfqd
->cfqg_list
);
3905 * Not strictly needed (since RB_ROOT just clears the node and we
3906 * zeroed cfqd on alloc), but better be safe in case someone decides
3907 * to add magic to the rb code
3909 for (i
= 0; i
< CFQ_PRIO_LISTS
; i
++)
3910 cfqd
->prio_trees
[i
] = RB_ROOT
;
3913 * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
3914 * Grab a permanent reference to it, so that the normal code flow
3915 * will not attempt to free it.
3917 cfq_init_cfqq(cfqd
, &cfqd
->oom_cfqq
, 1, 0);
3918 cfqd
->oom_cfqq
.ref
++;
3919 cfq_link_cfqq_cfqg(&cfqd
->oom_cfqq
, &cfqd
->root_group
);
3923 init_timer(&cfqd
->idle_slice_timer
);
3924 cfqd
->idle_slice_timer
.function
= cfq_idle_slice_timer
;
3925 cfqd
->idle_slice_timer
.data
= (unsigned long) cfqd
;
3927 INIT_WORK(&cfqd
->unplug_work
, cfq_kick_queue
);
3929 cfqd
->cfq_quantum
= cfq_quantum
;
3930 cfqd
->cfq_fifo_expire
[0] = cfq_fifo_expire
[0];
3931 cfqd
->cfq_fifo_expire
[1] = cfq_fifo_expire
[1];
3932 cfqd
->cfq_back_max
= cfq_back_max
;
3933 cfqd
->cfq_back_penalty
= cfq_back_penalty
;
3934 cfqd
->cfq_slice
[0] = cfq_slice_async
;
3935 cfqd
->cfq_slice
[1] = cfq_slice_sync
;
3936 cfqd
->cfq_slice_async_rq
= cfq_slice_async_rq
;
3937 cfqd
->cfq_slice_idle
= cfq_slice_idle
;
3938 cfqd
->cfq_group_idle
= cfq_group_idle
;
3939 cfqd
->cfq_latency
= 1;
3942 * we optimistically start assuming sync ops weren't delayed in last
3943 * second, in order to have larger depth for async operations.
3945 cfqd
->last_delayed_sync
= jiffies
- HZ
;
3949 static void cfq_slab_kill(void)
3952 * Caller already ensured that pending RCU callbacks are completed,
3953 * so we should have no busy allocations at this point.
3956 kmem_cache_destroy(cfq_pool
);
3958 kmem_cache_destroy(cfq_icq_pool
);
3961 static int __init
cfq_slab_setup(void)
3963 cfq_pool
= KMEM_CACHE(cfq_queue
, 0);
3967 cfq_icq_pool
= KMEM_CACHE(cfq_io_cq
, 0);
3978 * sysfs parts below -->
3981 cfq_var_show(unsigned int var
, char *page
)
3983 return sprintf(page
, "%d\n", var
);
3987 cfq_var_store(unsigned int *var
, const char *page
, size_t count
)
3989 char *p
= (char *) page
;
3991 *var
= simple_strtoul(p
, &p
, 10);
3995 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
3996 static ssize_t __FUNC(struct elevator_queue *e, char *page) \
3998 struct cfq_data *cfqd = e->elevator_data; \
3999 unsigned int __data = __VAR; \
4001 __data = jiffies_to_msecs(__data); \
4002 return cfq_var_show(__data, (page)); \
4004 SHOW_FUNCTION(cfq_quantum_show
, cfqd
->cfq_quantum
, 0);
4005 SHOW_FUNCTION(cfq_fifo_expire_sync_show
, cfqd
->cfq_fifo_expire
[1], 1);
4006 SHOW_FUNCTION(cfq_fifo_expire_async_show
, cfqd
->cfq_fifo_expire
[0], 1);
4007 SHOW_FUNCTION(cfq_back_seek_max_show
, cfqd
->cfq_back_max
, 0);
4008 SHOW_FUNCTION(cfq_back_seek_penalty_show
, cfqd
->cfq_back_penalty
, 0);
4009 SHOW_FUNCTION(cfq_slice_idle_show
, cfqd
->cfq_slice_idle
, 1);
4010 SHOW_FUNCTION(cfq_group_idle_show
, cfqd
->cfq_group_idle
, 1);
4011 SHOW_FUNCTION(cfq_slice_sync_show
, cfqd
->cfq_slice
[1], 1);
4012 SHOW_FUNCTION(cfq_slice_async_show
, cfqd
->cfq_slice
[0], 1);
4013 SHOW_FUNCTION(cfq_slice_async_rq_show
, cfqd
->cfq_slice_async_rq
, 0);
4014 SHOW_FUNCTION(cfq_low_latency_show
, cfqd
->cfq_latency
, 0);
4015 #undef SHOW_FUNCTION
4017 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
4018 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
4020 struct cfq_data *cfqd = e->elevator_data; \
4021 unsigned int __data; \
4022 int ret = cfq_var_store(&__data, (page), count); \
4023 if (__data < (MIN)) \
4025 else if (__data > (MAX)) \
4028 *(__PTR) = msecs_to_jiffies(__data); \
4030 *(__PTR) = __data; \
4033 STORE_FUNCTION(cfq_quantum_store
, &cfqd
->cfq_quantum
, 1, UINT_MAX
, 0);
4034 STORE_FUNCTION(cfq_fifo_expire_sync_store
, &cfqd
->cfq_fifo_expire
[1], 1,
4036 STORE_FUNCTION(cfq_fifo_expire_async_store
, &cfqd
->cfq_fifo_expire
[0], 1,
4038 STORE_FUNCTION(cfq_back_seek_max_store
, &cfqd
->cfq_back_max
, 0, UINT_MAX
, 0);
4039 STORE_FUNCTION(cfq_back_seek_penalty_store
, &cfqd
->cfq_back_penalty
, 1,
4041 STORE_FUNCTION(cfq_slice_idle_store
, &cfqd
->cfq_slice_idle
, 0, UINT_MAX
, 1);
4042 STORE_FUNCTION(cfq_group_idle_store
, &cfqd
->cfq_group_idle
, 0, UINT_MAX
, 1);
4043 STORE_FUNCTION(cfq_slice_sync_store
, &cfqd
->cfq_slice
[1], 1, UINT_MAX
, 1);
4044 STORE_FUNCTION(cfq_slice_async_store
, &cfqd
->cfq_slice
[0], 1, UINT_MAX
, 1);
4045 STORE_FUNCTION(cfq_slice_async_rq_store
, &cfqd
->cfq_slice_async_rq
, 1,
4047 STORE_FUNCTION(cfq_low_latency_store
, &cfqd
->cfq_latency
, 0, 1, 0);
4048 #undef STORE_FUNCTION
4050 #define CFQ_ATTR(name) \
4051 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
4053 static struct elv_fs_entry cfq_attrs
[] = {
4055 CFQ_ATTR(fifo_expire_sync
),
4056 CFQ_ATTR(fifo_expire_async
),
4057 CFQ_ATTR(back_seek_max
),
4058 CFQ_ATTR(back_seek_penalty
),
4059 CFQ_ATTR(slice_sync
),
4060 CFQ_ATTR(slice_async
),
4061 CFQ_ATTR(slice_async_rq
),
4062 CFQ_ATTR(slice_idle
),
4063 CFQ_ATTR(group_idle
),
4064 CFQ_ATTR(low_latency
),
4068 static struct elevator_type iosched_cfq
= {
4070 .elevator_merge_fn
= cfq_merge
,
4071 .elevator_merged_fn
= cfq_merged_request
,
4072 .elevator_merge_req_fn
= cfq_merged_requests
,
4073 .elevator_allow_merge_fn
= cfq_allow_merge
,
4074 .elevator_bio_merged_fn
= cfq_bio_merged
,
4075 .elevator_dispatch_fn
= cfq_dispatch_requests
,
4076 .elevator_add_req_fn
= cfq_insert_request
,
4077 .elevator_activate_req_fn
= cfq_activate_request
,
4078 .elevator_deactivate_req_fn
= cfq_deactivate_request
,
4079 .elevator_completed_req_fn
= cfq_completed_request
,
4080 .elevator_former_req_fn
= elv_rb_former_request
,
4081 .elevator_latter_req_fn
= elv_rb_latter_request
,
4082 .elevator_set_req_fn
= cfq_set_request
,
4083 .elevator_put_req_fn
= cfq_put_request
,
4084 .elevator_may_queue_fn
= cfq_may_queue
,
4085 .elevator_init_fn
= cfq_init_queue
,
4086 .elevator_exit_fn
= cfq_exit_queue
,
4088 .elevator_attrs
= cfq_attrs
,
4089 .elevator_name
= "cfq",
4090 .elevator_owner
= THIS_MODULE
,
4093 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4094 static struct blkio_policy_type blkio_policy_cfq
= {
4096 .blkio_unlink_group_fn
= cfq_unlink_blkio_group
,
4097 .blkio_update_group_weight_fn
= cfq_update_blkio_group_weight
,
4099 .plid
= BLKIO_POLICY_PROP
,
4102 static struct blkio_policy_type blkio_policy_cfq
;
4105 static int __init
cfq_init(void)
4108 * could be 0 on HZ < 1000 setups
4110 if (!cfq_slice_async
)
4111 cfq_slice_async
= 1;
4112 if (!cfq_slice_idle
)
4115 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4116 if (!cfq_group_idle
)
4121 if (cfq_slab_setup())
4124 elv_register(&iosched_cfq
);
4125 blkio_policy_register(&blkio_policy_cfq
);
4130 static void __exit
cfq_exit(void)
4132 blkio_policy_unregister(&blkio_policy_cfq
);
4133 elv_unregister(&iosched_cfq
);
4134 rcu_barrier(); /* make sure all cic RCU frees are complete */
4138 module_init(cfq_init
);
4139 module_exit(cfq_exit
);
4141 MODULE_AUTHOR("Jens Axboe");
4142 MODULE_LICENSE("GPL");
4143 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");