2 * CFQ, or complete fairness queueing, disk scheduler.
4 * Based on ideas from a previously unfinished io
5 * scheduler (round robin per-process disk scheduling) and Andrea Arcangeli.
7 * Copyright (C) 2003 Jens Axboe <axboe@kernel.dk>
9 #include <linux/module.h>
10 #include <linux/slab.h>
11 #include <linux/blkdev.h>
12 #include <linux/elevator.h>
13 #include <linux/jiffies.h>
14 #include <linux/rbtree.h>
15 #include <linux/ioprio.h>
16 #include <linux/blktrace_api.h>
23 /* max queue in one round of service */
24 static const int cfq_quantum
= 8;
25 static const int cfq_fifo_expire
[2] = { HZ
/ 4, HZ
/ 8 };
26 /* maximum backwards seek, in KiB */
27 static const int cfq_back_max
= 16 * 1024;
28 /* penalty of a backwards seek */
29 static const int cfq_back_penalty
= 2;
30 static const int cfq_slice_sync
= HZ
/ 10;
31 static int cfq_slice_async
= HZ
/ 25;
32 static const int cfq_slice_async_rq
= 2;
33 static int cfq_slice_idle
= HZ
/ 125;
34 static int cfq_group_idle
= HZ
/ 125;
35 static const int cfq_target_latency
= HZ
* 3/10; /* 300 ms */
36 static const int cfq_hist_divisor
= 4;
39 * offset from end of service tree
41 #define CFQ_IDLE_DELAY (HZ / 5)
44 * below this threshold, we consider thinktime immediate
46 #define CFQ_MIN_TT (2)
48 #define CFQ_SLICE_SCALE (5)
49 #define CFQ_HW_QUEUE_MIN (5)
50 #define CFQ_SERVICE_SHIFT 12
52 #define CFQQ_SEEK_THR (sector_t)(8 * 100)
53 #define CFQQ_CLOSE_THR (sector_t)(8 * 1024)
54 #define CFQQ_SECT_THR_NONROT (sector_t)(2 * 32)
55 #define CFQQ_SEEKY(cfqq) (hweight32(cfqq->seek_history) > 32/8)
58 ((struct cfq_io_context *) (rq)->elevator_private[0])
59 #define RQ_CFQQ(rq) (struct cfq_queue *) ((rq)->elevator_private[1])
60 #define RQ_CFQG(rq) (struct cfq_group *) ((rq)->elevator_private[2])
62 static struct kmem_cache
*cfq_pool
;
63 static struct kmem_cache
*cfq_ioc_pool
;
65 static DEFINE_PER_CPU(unsigned long, cfq_ioc_count
);
66 static struct completion
*ioc_gone
;
67 static DEFINE_SPINLOCK(ioc_gone_lock
);
69 #define CFQ_PRIO_LISTS IOPRIO_BE_NR
70 #define cfq_class_idle(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
71 #define cfq_class_rt(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
73 #define sample_valid(samples) ((samples) > 80)
74 #define rb_entry_cfqg(node) rb_entry((node), struct cfq_group, rb_node)
77 * Most of our rbtree usage is for sorting with min extraction, so
78 * if we cache the leftmost node we don't have to walk down the tree
79 * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should
80 * move this into the elevator for the rq sorting as well.
86 unsigned total_weight
;
88 struct cfq_ttime ttime
;
90 #define CFQ_RB_ROOT (struct cfq_rb_root) { .rb = RB_ROOT, \
91 .ttime = {.last_end_request = jiffies,},}
94 * Per process-grouping structure
99 /* various state flags, see below */
101 /* parent cfq_data */
102 struct cfq_data
*cfqd
;
103 /* service_tree member */
104 struct rb_node rb_node
;
105 /* service_tree key */
106 unsigned long rb_key
;
107 /* prio tree member */
108 struct rb_node p_node
;
109 /* prio tree root we belong to, if any */
110 struct rb_root
*p_root
;
111 /* sorted list of pending requests */
112 struct rb_root sort_list
;
113 /* if fifo isn't expired, next request to serve */
114 struct request
*next_rq
;
115 /* requests queued in sort_list */
117 /* currently allocated requests */
119 /* fifo list of requests in sort_list */
120 struct list_head fifo
;
122 /* time when queue got scheduled in to dispatch first request. */
123 unsigned long dispatch_start
;
124 unsigned int allocated_slice
;
125 unsigned int slice_dispatch
;
126 /* time when first request from queue completed and slice started. */
127 unsigned long slice_start
;
128 unsigned long slice_end
;
131 /* pending priority requests */
133 /* number of requests that are on the dispatch list or inside driver */
136 /* io prio of this group */
137 unsigned short ioprio
, org_ioprio
;
138 unsigned short ioprio_class
;
143 sector_t last_request_pos
;
145 struct cfq_rb_root
*service_tree
;
146 struct cfq_queue
*new_cfqq
;
147 struct cfq_group
*cfqg
;
148 /* Number of sectors dispatched from queue in single dispatch round */
149 unsigned long nr_sectors
;
153 * First index in the service_trees.
154 * IDLE is handled separately, so it has negative index
164 * Second index in the service_trees.
168 SYNC_NOIDLE_WORKLOAD
= 1,
172 /* This is per cgroup per device grouping structure */
174 /* group service_tree member */
175 struct rb_node rb_node
;
177 /* group service_tree key */
180 unsigned int new_weight
;
183 /* number of cfqq currently on this group */
187 * Per group busy queues average. Useful for workload slice calc. We
188 * create the array for each prio class but at run time it is used
189 * only for RT and BE class and slot for IDLE class remains unused.
190 * This is primarily done to avoid confusion and a gcc warning.
192 unsigned int busy_queues_avg
[CFQ_PRIO_NR
];
194 * rr lists of queues with requests. We maintain service trees for
195 * RT and BE classes. These trees are subdivided in subclasses
196 * of SYNC, SYNC_NOIDLE and ASYNC based on workload type. For IDLE
197 * class there is no subclassification and all the cfq queues go on
198 * a single tree service_tree_idle.
199 * Counts are embedded in the cfq_rb_root
201 struct cfq_rb_root service_trees
[2][3];
202 struct cfq_rb_root service_tree_idle
;
204 unsigned long saved_workload_slice
;
205 enum wl_type_t saved_workload
;
206 enum wl_prio_t saved_serving_prio
;
207 struct blkio_group blkg
;
208 #ifdef CONFIG_CFQ_GROUP_IOSCHED
209 struct hlist_node cfqd_node
;
212 /* number of requests that are on the dispatch list or inside driver */
214 struct cfq_ttime ttime
;
218 * Per block device queue structure
221 struct request_queue
*queue
;
222 /* Root service tree for cfq_groups */
223 struct cfq_rb_root grp_service_tree
;
224 struct cfq_group root_group
;
227 * The priority currently being served
229 enum wl_prio_t serving_prio
;
230 enum wl_type_t serving_type
;
231 unsigned long workload_expires
;
232 struct cfq_group
*serving_group
;
235 * Each priority tree is sorted by next_request position. These
236 * trees are used when determining if two or more queues are
237 * interleaving requests (see cfq_close_cooperator).
239 struct rb_root prio_trees
[CFQ_PRIO_LISTS
];
241 unsigned int busy_queues
;
242 unsigned int busy_sync_queues
;
248 * queue-depth detection
254 * -1 => indeterminate, (cfq will behave as if NCQ is present, to allow better detection)
255 * 1 => NCQ is present (hw_tag_est_depth is the estimated max depth)
258 int hw_tag_est_depth
;
259 unsigned int hw_tag_samples
;
262 * idle window management
264 struct timer_list idle_slice_timer
;
265 struct work_struct unplug_work
;
267 struct cfq_queue
*active_queue
;
268 struct cfq_io_context
*active_cic
;
271 * async queue for each priority case
273 struct cfq_queue
*async_cfqq
[2][IOPRIO_BE_NR
];
274 struct cfq_queue
*async_idle_cfqq
;
276 sector_t last_position
;
279 * tunables, see top of file
281 unsigned int cfq_quantum
;
282 unsigned int cfq_fifo_expire
[2];
283 unsigned int cfq_back_penalty
;
284 unsigned int cfq_back_max
;
285 unsigned int cfq_slice
[2];
286 unsigned int cfq_slice_async_rq
;
287 unsigned int cfq_slice_idle
;
288 unsigned int cfq_group_idle
;
289 unsigned int cfq_latency
;
291 struct list_head cic_list
;
294 * Fallback dummy cfqq for extreme OOM conditions
296 struct cfq_queue oom_cfqq
;
298 unsigned long last_delayed_sync
;
300 /* List of cfq groups being managed on this device*/
301 struct hlist_head cfqg_list
;
303 /* Number of groups which are on blkcg->blkg_list */
304 unsigned int nr_blkcg_linked_grps
;
307 static struct cfq_group
*cfq_get_next_cfqg(struct cfq_data
*cfqd
);
309 static struct cfq_rb_root
*service_tree_for(struct cfq_group
*cfqg
,
316 if (prio
== IDLE_WORKLOAD
)
317 return &cfqg
->service_tree_idle
;
319 return &cfqg
->service_trees
[prio
][type
];
322 enum cfqq_state_flags
{
323 CFQ_CFQQ_FLAG_on_rr
= 0, /* on round-robin busy list */
324 CFQ_CFQQ_FLAG_wait_request
, /* waiting for a request */
325 CFQ_CFQQ_FLAG_must_dispatch
, /* must be allowed a dispatch */
326 CFQ_CFQQ_FLAG_must_alloc_slice
, /* per-slice must_alloc flag */
327 CFQ_CFQQ_FLAG_fifo_expire
, /* FIFO checked in this slice */
328 CFQ_CFQQ_FLAG_idle_window
, /* slice idling enabled */
329 CFQ_CFQQ_FLAG_prio_changed
, /* task priority has changed */
330 CFQ_CFQQ_FLAG_slice_new
, /* no requests dispatched in slice */
331 CFQ_CFQQ_FLAG_sync
, /* synchronous queue */
332 CFQ_CFQQ_FLAG_coop
, /* cfqq is shared */
333 CFQ_CFQQ_FLAG_split_coop
, /* shared cfqq will be splitted */
334 CFQ_CFQQ_FLAG_deep
, /* sync cfqq experienced large depth */
335 CFQ_CFQQ_FLAG_wait_busy
, /* Waiting for next request */
338 #define CFQ_CFQQ_FNS(name) \
339 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \
341 (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name); \
343 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \
345 (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \
347 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \
349 return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \
353 CFQ_CFQQ_FNS(wait_request
);
354 CFQ_CFQQ_FNS(must_dispatch
);
355 CFQ_CFQQ_FNS(must_alloc_slice
);
356 CFQ_CFQQ_FNS(fifo_expire
);
357 CFQ_CFQQ_FNS(idle_window
);
358 CFQ_CFQQ_FNS(prio_changed
);
359 CFQ_CFQQ_FNS(slice_new
);
362 CFQ_CFQQ_FNS(split_coop
);
364 CFQ_CFQQ_FNS(wait_busy
);
367 #ifdef CONFIG_CFQ_GROUP_IOSCHED
368 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
369 blk_add_trace_msg((cfqd)->queue, "cfq%d%c %s " fmt, (cfqq)->pid, \
370 cfq_cfqq_sync((cfqq)) ? 'S' : 'A', \
371 blkg_path(&(cfqq)->cfqg->blkg), ##args)
373 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...) \
374 blk_add_trace_msg((cfqd)->queue, "%s " fmt, \
375 blkg_path(&(cfqg)->blkg), ##args) \
378 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
379 blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args)
380 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...) do {} while (0)
382 #define cfq_log(cfqd, fmt, args...) \
383 blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
385 /* Traverses through cfq group service trees */
386 #define for_each_cfqg_st(cfqg, i, j, st) \
387 for (i = 0; i <= IDLE_WORKLOAD; i++) \
388 for (j = 0, st = i < IDLE_WORKLOAD ? &cfqg->service_trees[i][j]\
389 : &cfqg->service_tree_idle; \
390 (i < IDLE_WORKLOAD && j <= SYNC_WORKLOAD) || \
391 (i == IDLE_WORKLOAD && j == 0); \
392 j++, st = i < IDLE_WORKLOAD ? \
393 &cfqg->service_trees[i][j]: NULL) \
395 static inline bool cfq_io_thinktime_big(struct cfq_data *cfqd,
396 struct cfq_ttime
*ttime
, bool group_idle
)
399 if (!sample_valid(ttime
->ttime_samples
))
402 slice
= cfqd
->cfq_group_idle
;
404 slice
= cfqd
->cfq_slice_idle
;
405 return ttime
->ttime_mean
> slice
;
408 static inline bool iops_mode(struct cfq_data
*cfqd
)
411 * If we are not idling on queues and it is a NCQ drive, parallel
412 * execution of requests is on and measuring time is not possible
413 * in most of the cases until and unless we drive shallower queue
414 * depths and that becomes a performance bottleneck. In such cases
415 * switch to start providing fairness in terms of number of IOs.
417 if (!cfqd
->cfq_slice_idle
&& cfqd
->hw_tag
)
423 static inline enum wl_prio_t
cfqq_prio(struct cfq_queue
*cfqq
)
425 if (cfq_class_idle(cfqq
))
426 return IDLE_WORKLOAD
;
427 if (cfq_class_rt(cfqq
))
433 static enum wl_type_t
cfqq_type(struct cfq_queue
*cfqq
)
435 if (!cfq_cfqq_sync(cfqq
))
436 return ASYNC_WORKLOAD
;
437 if (!cfq_cfqq_idle_window(cfqq
))
438 return SYNC_NOIDLE_WORKLOAD
;
439 return SYNC_WORKLOAD
;
442 static inline int cfq_group_busy_queues_wl(enum wl_prio_t wl
,
443 struct cfq_data
*cfqd
,
444 struct cfq_group
*cfqg
)
446 if (wl
== IDLE_WORKLOAD
)
447 return cfqg
->service_tree_idle
.count
;
449 return cfqg
->service_trees
[wl
][ASYNC_WORKLOAD
].count
450 + cfqg
->service_trees
[wl
][SYNC_NOIDLE_WORKLOAD
].count
451 + cfqg
->service_trees
[wl
][SYNC_WORKLOAD
].count
;
454 static inline int cfqg_busy_async_queues(struct cfq_data
*cfqd
,
455 struct cfq_group
*cfqg
)
457 return cfqg
->service_trees
[RT_WORKLOAD
][ASYNC_WORKLOAD
].count
458 + cfqg
->service_trees
[BE_WORKLOAD
][ASYNC_WORKLOAD
].count
;
461 static void cfq_dispatch_insert(struct request_queue
*, struct request
*);
462 static struct cfq_queue
*cfq_get_queue(struct cfq_data
*, bool,
463 struct io_context
*, gfp_t
);
464 static struct cfq_io_context
*cfq_cic_lookup(struct cfq_data
*,
465 struct io_context
*);
467 static inline struct cfq_queue
*cic_to_cfqq(struct cfq_io_context
*cic
,
470 return cic
->cfqq
[is_sync
];
473 static inline void cic_set_cfqq(struct cfq_io_context
*cic
,
474 struct cfq_queue
*cfqq
, bool is_sync
)
476 cic
->cfqq
[is_sync
] = cfqq
;
479 #define CIC_DEAD_KEY 1ul
480 #define CIC_DEAD_INDEX_SHIFT 1
482 static inline void *cfqd_dead_key(struct cfq_data
*cfqd
)
484 return (void *)(cfqd
->queue
->id
<< CIC_DEAD_INDEX_SHIFT
| CIC_DEAD_KEY
);
487 static inline struct cfq_data
*cic_to_cfqd(struct cfq_io_context
*cic
)
489 struct cfq_data
*cfqd
= cic
->key
;
491 if (unlikely((unsigned long) cfqd
& CIC_DEAD_KEY
))
498 * We regard a request as SYNC, if it's either a read or has the SYNC bit
499 * set (in which case it could also be direct WRITE).
501 static inline bool cfq_bio_sync(struct bio
*bio
)
503 return bio_data_dir(bio
) == READ
|| (bio
->bi_rw
& REQ_SYNC
);
507 * scheduler run of queue, if there are requests pending and no one in the
508 * driver that will restart queueing
510 static inline void cfq_schedule_dispatch(struct cfq_data
*cfqd
)
512 if (cfqd
->busy_queues
) {
513 cfq_log(cfqd
, "schedule dispatch");
514 kblockd_schedule_work(cfqd
->queue
, &cfqd
->unplug_work
);
519 * Scale schedule slice based on io priority. Use the sync time slice only
520 * if a queue is marked sync and has sync io queued. A sync queue with async
521 * io only, should not get full sync slice length.
523 static inline int cfq_prio_slice(struct cfq_data
*cfqd
, bool sync
,
526 const int base_slice
= cfqd
->cfq_slice
[sync
];
528 WARN_ON(prio
>= IOPRIO_BE_NR
);
530 return base_slice
+ (base_slice
/CFQ_SLICE_SCALE
* (4 - prio
));
534 cfq_prio_to_slice(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
536 return cfq_prio_slice(cfqd
, cfq_cfqq_sync(cfqq
), cfqq
->ioprio
);
539 static inline u64
cfq_scale_slice(unsigned long delta
, struct cfq_group
*cfqg
)
541 u64 d
= delta
<< CFQ_SERVICE_SHIFT
;
543 d
= d
* BLKIO_WEIGHT_DEFAULT
;
544 do_div(d
, cfqg
->weight
);
548 static inline u64
max_vdisktime(u64 min_vdisktime
, u64 vdisktime
)
550 s64 delta
= (s64
)(vdisktime
- min_vdisktime
);
552 min_vdisktime
= vdisktime
;
554 return min_vdisktime
;
557 static inline u64
min_vdisktime(u64 min_vdisktime
, u64 vdisktime
)
559 s64 delta
= (s64
)(vdisktime
- min_vdisktime
);
561 min_vdisktime
= vdisktime
;
563 return min_vdisktime
;
566 static void update_min_vdisktime(struct cfq_rb_root
*st
)
568 struct cfq_group
*cfqg
;
571 cfqg
= rb_entry_cfqg(st
->left
);
572 st
->min_vdisktime
= max_vdisktime(st
->min_vdisktime
,
578 * get averaged number of queues of RT/BE priority.
579 * average is updated, with a formula that gives more weight to higher numbers,
580 * to quickly follows sudden increases and decrease slowly
583 static inline unsigned cfq_group_get_avg_queues(struct cfq_data
*cfqd
,
584 struct cfq_group
*cfqg
, bool rt
)
586 unsigned min_q
, max_q
;
587 unsigned mult
= cfq_hist_divisor
- 1;
588 unsigned round
= cfq_hist_divisor
/ 2;
589 unsigned busy
= cfq_group_busy_queues_wl(rt
, cfqd
, cfqg
);
591 min_q
= min(cfqg
->busy_queues_avg
[rt
], busy
);
592 max_q
= max(cfqg
->busy_queues_avg
[rt
], busy
);
593 cfqg
->busy_queues_avg
[rt
] = (mult
* max_q
+ min_q
+ round
) /
595 return cfqg
->busy_queues_avg
[rt
];
598 static inline unsigned
599 cfq_group_slice(struct cfq_data
*cfqd
, struct cfq_group
*cfqg
)
601 struct cfq_rb_root
*st
= &cfqd
->grp_service_tree
;
603 return cfq_target_latency
* cfqg
->weight
/ st
->total_weight
;
606 static inline unsigned
607 cfq_scaled_cfqq_slice(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
609 unsigned slice
= cfq_prio_to_slice(cfqd
, cfqq
);
610 if (cfqd
->cfq_latency
) {
612 * interested queues (we consider only the ones with the same
613 * priority class in the cfq group)
615 unsigned iq
= cfq_group_get_avg_queues(cfqd
, cfqq
->cfqg
,
617 unsigned sync_slice
= cfqd
->cfq_slice
[1];
618 unsigned expect_latency
= sync_slice
* iq
;
619 unsigned group_slice
= cfq_group_slice(cfqd
, cfqq
->cfqg
);
621 if (expect_latency
> group_slice
) {
622 unsigned base_low_slice
= 2 * cfqd
->cfq_slice_idle
;
623 /* scale low_slice according to IO priority
624 * and sync vs async */
626 min(slice
, base_low_slice
* slice
/ sync_slice
);
627 /* the adapted slice value is scaled to fit all iqs
628 * into the target latency */
629 slice
= max(slice
* group_slice
/ expect_latency
,
637 cfq_set_prio_slice(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
639 unsigned slice
= cfq_scaled_cfqq_slice(cfqd
, cfqq
);
641 cfqq
->slice_start
= jiffies
;
642 cfqq
->slice_end
= jiffies
+ slice
;
643 cfqq
->allocated_slice
= slice
;
644 cfq_log_cfqq(cfqd
, cfqq
, "set_slice=%lu", cfqq
->slice_end
- jiffies
);
648 * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
649 * isn't valid until the first request from the dispatch is activated
650 * and the slice time set.
652 static inline bool cfq_slice_used(struct cfq_queue
*cfqq
)
654 if (cfq_cfqq_slice_new(cfqq
))
656 if (time_before(jiffies
, cfqq
->slice_end
))
663 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
664 * We choose the request that is closest to the head right now. Distance
665 * behind the head is penalized and only allowed to a certain extent.
667 static struct request
*
668 cfq_choose_req(struct cfq_data
*cfqd
, struct request
*rq1
, struct request
*rq2
, sector_t last
)
670 sector_t s1
, s2
, d1
= 0, d2
= 0;
671 unsigned long back_max
;
672 #define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */
673 #define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */
674 unsigned wrap
= 0; /* bit mask: requests behind the disk head? */
676 if (rq1
== NULL
|| rq1
== rq2
)
681 if (rq_is_sync(rq1
) != rq_is_sync(rq2
))
682 return rq_is_sync(rq1
) ? rq1
: rq2
;
684 if ((rq1
->cmd_flags
^ rq2
->cmd_flags
) & REQ_PRIO
)
685 return rq1
->cmd_flags
& REQ_PRIO
? rq1
: rq2
;
687 s1
= blk_rq_pos(rq1
);
688 s2
= blk_rq_pos(rq2
);
691 * by definition, 1KiB is 2 sectors
693 back_max
= cfqd
->cfq_back_max
* 2;
696 * Strict one way elevator _except_ in the case where we allow
697 * short backward seeks which are biased as twice the cost of a
698 * similar forward seek.
702 else if (s1
+ back_max
>= last
)
703 d1
= (last
- s1
) * cfqd
->cfq_back_penalty
;
705 wrap
|= CFQ_RQ1_WRAP
;
709 else if (s2
+ back_max
>= last
)
710 d2
= (last
- s2
) * cfqd
->cfq_back_penalty
;
712 wrap
|= CFQ_RQ2_WRAP
;
714 /* Found required data */
717 * By doing switch() on the bit mask "wrap" we avoid having to
718 * check two variables for all permutations: --> faster!
721 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
737 case (CFQ_RQ1_WRAP
|CFQ_RQ2_WRAP
): /* both rqs wrapped */
740 * Since both rqs are wrapped,
741 * start with the one that's further behind head
742 * (--> only *one* back seek required),
743 * since back seek takes more time than forward.
753 * The below is leftmost cache rbtree addon
755 static struct cfq_queue
*cfq_rb_first(struct cfq_rb_root
*root
)
757 /* Service tree is empty */
762 root
->left
= rb_first(&root
->rb
);
765 return rb_entry(root
->left
, struct cfq_queue
, rb_node
);
770 static struct cfq_group
*cfq_rb_first_group(struct cfq_rb_root
*root
)
773 root
->left
= rb_first(&root
->rb
);
776 return rb_entry_cfqg(root
->left
);
781 static void rb_erase_init(struct rb_node
*n
, struct rb_root
*root
)
787 static void cfq_rb_erase(struct rb_node
*n
, struct cfq_rb_root
*root
)
791 rb_erase_init(n
, &root
->rb
);
796 * would be nice to take fifo expire time into account as well
798 static struct request
*
799 cfq_find_next_rq(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
800 struct request
*last
)
802 struct rb_node
*rbnext
= rb_next(&last
->rb_node
);
803 struct rb_node
*rbprev
= rb_prev(&last
->rb_node
);
804 struct request
*next
= NULL
, *prev
= NULL
;
806 BUG_ON(RB_EMPTY_NODE(&last
->rb_node
));
809 prev
= rb_entry_rq(rbprev
);
812 next
= rb_entry_rq(rbnext
);
814 rbnext
= rb_first(&cfqq
->sort_list
);
815 if (rbnext
&& rbnext
!= &last
->rb_node
)
816 next
= rb_entry_rq(rbnext
);
819 return cfq_choose_req(cfqd
, next
, prev
, blk_rq_pos(last
));
822 static unsigned long cfq_slice_offset(struct cfq_data
*cfqd
,
823 struct cfq_queue
*cfqq
)
826 * just an approximation, should be ok.
828 return (cfqq
->cfqg
->nr_cfqq
- 1) * (cfq_prio_slice(cfqd
, 1, 0) -
829 cfq_prio_slice(cfqd
, cfq_cfqq_sync(cfqq
), cfqq
->ioprio
));
833 cfqg_key(struct cfq_rb_root
*st
, struct cfq_group
*cfqg
)
835 return cfqg
->vdisktime
- st
->min_vdisktime
;
839 __cfq_group_service_tree_add(struct cfq_rb_root
*st
, struct cfq_group
*cfqg
)
841 struct rb_node
**node
= &st
->rb
.rb_node
;
842 struct rb_node
*parent
= NULL
;
843 struct cfq_group
*__cfqg
;
844 s64 key
= cfqg_key(st
, cfqg
);
847 while (*node
!= NULL
) {
849 __cfqg
= rb_entry_cfqg(parent
);
851 if (key
< cfqg_key(st
, __cfqg
))
852 node
= &parent
->rb_left
;
854 node
= &parent
->rb_right
;
860 st
->left
= &cfqg
->rb_node
;
862 rb_link_node(&cfqg
->rb_node
, parent
, node
);
863 rb_insert_color(&cfqg
->rb_node
, &st
->rb
);
867 cfq_update_group_weight(struct cfq_group
*cfqg
)
869 BUG_ON(!RB_EMPTY_NODE(&cfqg
->rb_node
));
870 if (cfqg
->needs_update
) {
871 cfqg
->weight
= cfqg
->new_weight
;
872 cfqg
->needs_update
= false;
877 cfq_group_service_tree_add(struct cfq_rb_root
*st
, struct cfq_group
*cfqg
)
879 BUG_ON(!RB_EMPTY_NODE(&cfqg
->rb_node
));
881 cfq_update_group_weight(cfqg
);
882 __cfq_group_service_tree_add(st
, cfqg
);
883 st
->total_weight
+= cfqg
->weight
;
887 cfq_group_notify_queue_add(struct cfq_data
*cfqd
, struct cfq_group
*cfqg
)
889 struct cfq_rb_root
*st
= &cfqd
->grp_service_tree
;
890 struct cfq_group
*__cfqg
;
894 if (!RB_EMPTY_NODE(&cfqg
->rb_node
))
898 * Currently put the group at the end. Later implement something
899 * so that groups get lesser vtime based on their weights, so that
900 * if group does not loose all if it was not continuously backlogged.
902 n
= rb_last(&st
->rb
);
904 __cfqg
= rb_entry_cfqg(n
);
905 cfqg
->vdisktime
= __cfqg
->vdisktime
+ CFQ_IDLE_DELAY
;
907 cfqg
->vdisktime
= st
->min_vdisktime
;
908 cfq_group_service_tree_add(st
, cfqg
);
912 cfq_group_service_tree_del(struct cfq_rb_root
*st
, struct cfq_group
*cfqg
)
914 st
->total_weight
-= cfqg
->weight
;
915 if (!RB_EMPTY_NODE(&cfqg
->rb_node
))
916 cfq_rb_erase(&cfqg
->rb_node
, st
);
920 cfq_group_notify_queue_del(struct cfq_data
*cfqd
, struct cfq_group
*cfqg
)
922 struct cfq_rb_root
*st
= &cfqd
->grp_service_tree
;
924 BUG_ON(cfqg
->nr_cfqq
< 1);
927 /* If there are other cfq queues under this group, don't delete it */
931 cfq_log_cfqg(cfqd
, cfqg
, "del_from_rr group");
932 cfq_group_service_tree_del(st
, cfqg
);
933 cfqg
->saved_workload_slice
= 0;
934 cfq_blkiocg_update_dequeue_stats(&cfqg
->blkg
, 1);
937 static inline unsigned int cfq_cfqq_slice_usage(struct cfq_queue
*cfqq
,
938 unsigned int *unaccounted_time
)
940 unsigned int slice_used
;
943 * Queue got expired before even a single request completed or
944 * got expired immediately after first request completion.
946 if (!cfqq
->slice_start
|| cfqq
->slice_start
== jiffies
) {
948 * Also charge the seek time incurred to the group, otherwise
949 * if there are mutiple queues in the group, each can dispatch
950 * a single request on seeky media and cause lots of seek time
951 * and group will never know it.
953 slice_used
= max_t(unsigned, (jiffies
- cfqq
->dispatch_start
),
956 slice_used
= jiffies
- cfqq
->slice_start
;
957 if (slice_used
> cfqq
->allocated_slice
) {
958 *unaccounted_time
= slice_used
- cfqq
->allocated_slice
;
959 slice_used
= cfqq
->allocated_slice
;
961 if (time_after(cfqq
->slice_start
, cfqq
->dispatch_start
))
962 *unaccounted_time
+= cfqq
->slice_start
-
963 cfqq
->dispatch_start
;
969 static void cfq_group_served(struct cfq_data
*cfqd
, struct cfq_group
*cfqg
,
970 struct cfq_queue
*cfqq
)
972 struct cfq_rb_root
*st
= &cfqd
->grp_service_tree
;
973 unsigned int used_sl
, charge
, unaccounted_sl
= 0;
974 int nr_sync
= cfqg
->nr_cfqq
- cfqg_busy_async_queues(cfqd
, cfqg
)
975 - cfqg
->service_tree_idle
.count
;
978 used_sl
= charge
= cfq_cfqq_slice_usage(cfqq
, &unaccounted_sl
);
981 charge
= cfqq
->slice_dispatch
;
982 else if (!cfq_cfqq_sync(cfqq
) && !nr_sync
)
983 charge
= cfqq
->allocated_slice
;
985 /* Can't update vdisktime while group is on service tree */
986 cfq_group_service_tree_del(st
, cfqg
);
987 cfqg
->vdisktime
+= cfq_scale_slice(charge
, cfqg
);
988 /* If a new weight was requested, update now, off tree */
989 cfq_group_service_tree_add(st
, cfqg
);
991 /* This group is being expired. Save the context */
992 if (time_after(cfqd
->workload_expires
, jiffies
)) {
993 cfqg
->saved_workload_slice
= cfqd
->workload_expires
995 cfqg
->saved_workload
= cfqd
->serving_type
;
996 cfqg
->saved_serving_prio
= cfqd
->serving_prio
;
998 cfqg
->saved_workload_slice
= 0;
1000 cfq_log_cfqg(cfqd
, cfqg
, "served: vt=%llu min_vt=%llu", cfqg
->vdisktime
,
1002 cfq_log_cfqq(cfqq
->cfqd
, cfqq
,
1003 "sl_used=%u disp=%u charge=%u iops=%u sect=%lu",
1004 used_sl
, cfqq
->slice_dispatch
, charge
,
1005 iops_mode(cfqd
), cfqq
->nr_sectors
);
1006 cfq_blkiocg_update_timeslice_used(&cfqg
->blkg
, used_sl
,
1008 cfq_blkiocg_set_start_empty_time(&cfqg
->blkg
);
1011 #ifdef CONFIG_CFQ_GROUP_IOSCHED
1012 static inline struct cfq_group
*cfqg_of_blkg(struct blkio_group
*blkg
)
1015 return container_of(blkg
, struct cfq_group
, blkg
);
1019 static void cfq_update_blkio_group_weight(void *key
, struct blkio_group
*blkg
,
1020 unsigned int weight
)
1022 struct cfq_group
*cfqg
= cfqg_of_blkg(blkg
);
1023 cfqg
->new_weight
= weight
;
1024 cfqg
->needs_update
= true;
1027 static void cfq_init_add_cfqg_lists(struct cfq_data
*cfqd
,
1028 struct cfq_group
*cfqg
, struct blkio_cgroup
*blkcg
)
1030 struct backing_dev_info
*bdi
= &cfqd
->queue
->backing_dev_info
;
1031 unsigned int major
, minor
;
1034 * Add group onto cgroup list. It might happen that bdi->dev is
1035 * not initialized yet. Initialize this new group without major
1036 * and minor info and this info will be filled in once a new thread
1040 sscanf(dev_name(bdi
->dev
), "%u:%u", &major
, &minor
);
1041 cfq_blkiocg_add_blkio_group(blkcg
, &cfqg
->blkg
,
1042 (void *)cfqd
, MKDEV(major
, minor
));
1044 cfq_blkiocg_add_blkio_group(blkcg
, &cfqg
->blkg
,
1047 cfqd
->nr_blkcg_linked_grps
++;
1048 cfqg
->weight
= blkcg_get_weight(blkcg
, cfqg
->blkg
.dev
);
1050 /* Add group on cfqd list */
1051 hlist_add_head(&cfqg
->cfqd_node
, &cfqd
->cfqg_list
);
1055 * Should be called from sleepable context. No request queue lock as per
1056 * cpu stats are allocated dynamically and alloc_percpu needs to be called
1057 * from sleepable context.
1059 static struct cfq_group
* cfq_alloc_cfqg(struct cfq_data
*cfqd
)
1061 struct cfq_group
*cfqg
= NULL
;
1063 struct cfq_rb_root
*st
;
1065 cfqg
= kzalloc_node(sizeof(*cfqg
), GFP_ATOMIC
, cfqd
->queue
->node
);
1069 for_each_cfqg_st(cfqg
, i
, j
, st
)
1071 RB_CLEAR_NODE(&cfqg
->rb_node
);
1073 cfqg
->ttime
.last_end_request
= jiffies
;
1076 * Take the initial reference that will be released on destroy
1077 * This can be thought of a joint reference by cgroup and
1078 * elevator which will be dropped by either elevator exit
1079 * or cgroup deletion path depending on who is exiting first.
1083 ret
= blkio_alloc_blkg_stats(&cfqg
->blkg
);
1092 static struct cfq_group
*
1093 cfq_find_cfqg(struct cfq_data
*cfqd
, struct blkio_cgroup
*blkcg
)
1095 struct cfq_group
*cfqg
= NULL
;
1097 struct backing_dev_info
*bdi
= &cfqd
->queue
->backing_dev_info
;
1098 unsigned int major
, minor
;
1101 * This is the common case when there are no blkio cgroups.
1102 * Avoid lookup in this case
1104 if (blkcg
== &blkio_root_cgroup
)
1105 cfqg
= &cfqd
->root_group
;
1107 cfqg
= cfqg_of_blkg(blkiocg_lookup_group(blkcg
, key
));
1109 if (cfqg
&& !cfqg
->blkg
.dev
&& bdi
->dev
&& dev_name(bdi
->dev
)) {
1110 sscanf(dev_name(bdi
->dev
), "%u:%u", &major
, &minor
);
1111 cfqg
->blkg
.dev
= MKDEV(major
, minor
);
1118 * Search for the cfq group current task belongs to. request_queue lock must
1121 static struct cfq_group
*cfq_get_cfqg(struct cfq_data
*cfqd
)
1123 struct blkio_cgroup
*blkcg
;
1124 struct cfq_group
*cfqg
= NULL
, *__cfqg
= NULL
;
1125 struct request_queue
*q
= cfqd
->queue
;
1128 blkcg
= task_blkio_cgroup(current
);
1129 cfqg
= cfq_find_cfqg(cfqd
, blkcg
);
1136 * Need to allocate a group. Allocation of group also needs allocation
1137 * of per cpu stats which in-turn takes a mutex() and can block. Hence
1138 * we need to drop rcu lock and queue_lock before we call alloc.
1140 * Not taking any queue reference here and assuming that queue is
1141 * around by the time we return. CFQ queue allocation code does
1142 * the same. It might be racy though.
1146 spin_unlock_irq(q
->queue_lock
);
1148 cfqg
= cfq_alloc_cfqg(cfqd
);
1150 spin_lock_irq(q
->queue_lock
);
1153 blkcg
= task_blkio_cgroup(current
);
1156 * If some other thread already allocated the group while we were
1157 * not holding queue lock, free up the group
1159 __cfqg
= cfq_find_cfqg(cfqd
, blkcg
);
1168 cfqg
= &cfqd
->root_group
;
1170 cfq_init_add_cfqg_lists(cfqd
, cfqg
, blkcg
);
1175 static inline struct cfq_group
*cfq_ref_get_cfqg(struct cfq_group
*cfqg
)
1181 static void cfq_link_cfqq_cfqg(struct cfq_queue
*cfqq
, struct cfq_group
*cfqg
)
1183 /* Currently, all async queues are mapped to root group */
1184 if (!cfq_cfqq_sync(cfqq
))
1185 cfqg
= &cfqq
->cfqd
->root_group
;
1188 /* cfqq reference on cfqg */
1192 static void cfq_put_cfqg(struct cfq_group
*cfqg
)
1194 struct cfq_rb_root
*st
;
1197 BUG_ON(cfqg
->ref
<= 0);
1201 for_each_cfqg_st(cfqg
, i
, j
, st
)
1202 BUG_ON(!RB_EMPTY_ROOT(&st
->rb
));
1203 free_percpu(cfqg
->blkg
.stats_cpu
);
1207 static void cfq_destroy_cfqg(struct cfq_data
*cfqd
, struct cfq_group
*cfqg
)
1209 /* Something wrong if we are trying to remove same group twice */
1210 BUG_ON(hlist_unhashed(&cfqg
->cfqd_node
));
1212 hlist_del_init(&cfqg
->cfqd_node
);
1214 BUG_ON(cfqd
->nr_blkcg_linked_grps
<= 0);
1215 cfqd
->nr_blkcg_linked_grps
--;
1218 * Put the reference taken at the time of creation so that when all
1219 * queues are gone, group can be destroyed.
1224 static void cfq_release_cfq_groups(struct cfq_data
*cfqd
)
1226 struct hlist_node
*pos
, *n
;
1227 struct cfq_group
*cfqg
;
1229 hlist_for_each_entry_safe(cfqg
, pos
, n
, &cfqd
->cfqg_list
, cfqd_node
) {
1231 * If cgroup removal path got to blk_group first and removed
1232 * it from cgroup list, then it will take care of destroying
1235 if (!cfq_blkiocg_del_blkio_group(&cfqg
->blkg
))
1236 cfq_destroy_cfqg(cfqd
, cfqg
);
1241 * Blk cgroup controller notification saying that blkio_group object is being
1242 * delinked as associated cgroup object is going away. That also means that
1243 * no new IO will come in this group. So get rid of this group as soon as
1244 * any pending IO in the group is finished.
1246 * This function is called under rcu_read_lock(). key is the rcu protected
1247 * pointer. That means "key" is a valid cfq_data pointer as long as we are rcu
1250 * "key" was fetched from blkio_group under blkio_cgroup->lock. That means
1251 * it should not be NULL as even if elevator was exiting, cgroup deltion
1252 * path got to it first.
1254 static void cfq_unlink_blkio_group(void *key
, struct blkio_group
*blkg
)
1256 unsigned long flags
;
1257 struct cfq_data
*cfqd
= key
;
1259 spin_lock_irqsave(cfqd
->queue
->queue_lock
, flags
);
1260 cfq_destroy_cfqg(cfqd
, cfqg_of_blkg(blkg
));
1261 spin_unlock_irqrestore(cfqd
->queue
->queue_lock
, flags
);
1264 #else /* GROUP_IOSCHED */
1265 static struct cfq_group
*cfq_get_cfqg(struct cfq_data
*cfqd
)
1267 return &cfqd
->root_group
;
1270 static inline struct cfq_group
*cfq_ref_get_cfqg(struct cfq_group
*cfqg
)
1276 cfq_link_cfqq_cfqg(struct cfq_queue
*cfqq
, struct cfq_group
*cfqg
) {
1280 static void cfq_release_cfq_groups(struct cfq_data
*cfqd
) {}
1281 static inline void cfq_put_cfqg(struct cfq_group
*cfqg
) {}
1283 #endif /* GROUP_IOSCHED */
1286 * The cfqd->service_trees holds all pending cfq_queue's that have
1287 * requests waiting to be processed. It is sorted in the order that
1288 * we will service the queues.
1290 static void cfq_service_tree_add(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
1293 struct rb_node
**p
, *parent
;
1294 struct cfq_queue
*__cfqq
;
1295 unsigned long rb_key
;
1296 struct cfq_rb_root
*service_tree
;
1300 service_tree
= service_tree_for(cfqq
->cfqg
, cfqq_prio(cfqq
),
1302 if (cfq_class_idle(cfqq
)) {
1303 rb_key
= CFQ_IDLE_DELAY
;
1304 parent
= rb_last(&service_tree
->rb
);
1305 if (parent
&& parent
!= &cfqq
->rb_node
) {
1306 __cfqq
= rb_entry(parent
, struct cfq_queue
, rb_node
);
1307 rb_key
+= __cfqq
->rb_key
;
1310 } else if (!add_front
) {
1312 * Get our rb key offset. Subtract any residual slice
1313 * value carried from last service. A negative resid
1314 * count indicates slice overrun, and this should position
1315 * the next service time further away in the tree.
1317 rb_key
= cfq_slice_offset(cfqd
, cfqq
) + jiffies
;
1318 rb_key
-= cfqq
->slice_resid
;
1319 cfqq
->slice_resid
= 0;
1322 __cfqq
= cfq_rb_first(service_tree
);
1323 rb_key
+= __cfqq
? __cfqq
->rb_key
: jiffies
;
1326 if (!RB_EMPTY_NODE(&cfqq
->rb_node
)) {
1329 * same position, nothing more to do
1331 if (rb_key
== cfqq
->rb_key
&&
1332 cfqq
->service_tree
== service_tree
)
1335 cfq_rb_erase(&cfqq
->rb_node
, cfqq
->service_tree
);
1336 cfqq
->service_tree
= NULL
;
1341 cfqq
->service_tree
= service_tree
;
1342 p
= &service_tree
->rb
.rb_node
;
1347 __cfqq
= rb_entry(parent
, struct cfq_queue
, rb_node
);
1350 * sort by key, that represents service time.
1352 if (time_before(rb_key
, __cfqq
->rb_key
))
1355 n
= &(*p
)->rb_right
;
1363 service_tree
->left
= &cfqq
->rb_node
;
1365 cfqq
->rb_key
= rb_key
;
1366 rb_link_node(&cfqq
->rb_node
, parent
, p
);
1367 rb_insert_color(&cfqq
->rb_node
, &service_tree
->rb
);
1368 service_tree
->count
++;
1369 if (add_front
|| !new_cfqq
)
1371 cfq_group_notify_queue_add(cfqd
, cfqq
->cfqg
);
1374 static struct cfq_queue
*
1375 cfq_prio_tree_lookup(struct cfq_data
*cfqd
, struct rb_root
*root
,
1376 sector_t sector
, struct rb_node
**ret_parent
,
1377 struct rb_node
***rb_link
)
1379 struct rb_node
**p
, *parent
;
1380 struct cfq_queue
*cfqq
= NULL
;
1388 cfqq
= rb_entry(parent
, struct cfq_queue
, p_node
);
1391 * Sort strictly based on sector. Smallest to the left,
1392 * largest to the right.
1394 if (sector
> blk_rq_pos(cfqq
->next_rq
))
1395 n
= &(*p
)->rb_right
;
1396 else if (sector
< blk_rq_pos(cfqq
->next_rq
))
1404 *ret_parent
= parent
;
1410 static void cfq_prio_tree_add(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
1412 struct rb_node
**p
, *parent
;
1413 struct cfq_queue
*__cfqq
;
1416 rb_erase(&cfqq
->p_node
, cfqq
->p_root
);
1417 cfqq
->p_root
= NULL
;
1420 if (cfq_class_idle(cfqq
))
1425 cfqq
->p_root
= &cfqd
->prio_trees
[cfqq
->org_ioprio
];
1426 __cfqq
= cfq_prio_tree_lookup(cfqd
, cfqq
->p_root
,
1427 blk_rq_pos(cfqq
->next_rq
), &parent
, &p
);
1429 rb_link_node(&cfqq
->p_node
, parent
, p
);
1430 rb_insert_color(&cfqq
->p_node
, cfqq
->p_root
);
1432 cfqq
->p_root
= NULL
;
1436 * Update cfqq's position in the service tree.
1438 static void cfq_resort_rr_list(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
1441 * Resorting requires the cfqq to be on the RR list already.
1443 if (cfq_cfqq_on_rr(cfqq
)) {
1444 cfq_service_tree_add(cfqd
, cfqq
, 0);
1445 cfq_prio_tree_add(cfqd
, cfqq
);
1450 * add to busy list of queues for service, trying to be fair in ordering
1451 * the pending list according to last request service
1453 static void cfq_add_cfqq_rr(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
1455 cfq_log_cfqq(cfqd
, cfqq
, "add_to_rr");
1456 BUG_ON(cfq_cfqq_on_rr(cfqq
));
1457 cfq_mark_cfqq_on_rr(cfqq
);
1458 cfqd
->busy_queues
++;
1459 if (cfq_cfqq_sync(cfqq
))
1460 cfqd
->busy_sync_queues
++;
1462 cfq_resort_rr_list(cfqd
, cfqq
);
1466 * Called when the cfqq no longer has requests pending, remove it from
1469 static void cfq_del_cfqq_rr(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
1471 cfq_log_cfqq(cfqd
, cfqq
, "del_from_rr");
1472 BUG_ON(!cfq_cfqq_on_rr(cfqq
));
1473 cfq_clear_cfqq_on_rr(cfqq
);
1475 if (!RB_EMPTY_NODE(&cfqq
->rb_node
)) {
1476 cfq_rb_erase(&cfqq
->rb_node
, cfqq
->service_tree
);
1477 cfqq
->service_tree
= NULL
;
1480 rb_erase(&cfqq
->p_node
, cfqq
->p_root
);
1481 cfqq
->p_root
= NULL
;
1484 cfq_group_notify_queue_del(cfqd
, cfqq
->cfqg
);
1485 BUG_ON(!cfqd
->busy_queues
);
1486 cfqd
->busy_queues
--;
1487 if (cfq_cfqq_sync(cfqq
))
1488 cfqd
->busy_sync_queues
--;
1492 * rb tree support functions
1494 static void cfq_del_rq_rb(struct request
*rq
)
1496 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
1497 const int sync
= rq_is_sync(rq
);
1499 BUG_ON(!cfqq
->queued
[sync
]);
1500 cfqq
->queued
[sync
]--;
1502 elv_rb_del(&cfqq
->sort_list
, rq
);
1504 if (cfq_cfqq_on_rr(cfqq
) && RB_EMPTY_ROOT(&cfqq
->sort_list
)) {
1506 * Queue will be deleted from service tree when we actually
1507 * expire it later. Right now just remove it from prio tree
1511 rb_erase(&cfqq
->p_node
, cfqq
->p_root
);
1512 cfqq
->p_root
= NULL
;
1517 static void cfq_add_rq_rb(struct request
*rq
)
1519 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
1520 struct cfq_data
*cfqd
= cfqq
->cfqd
;
1521 struct request
*prev
;
1523 cfqq
->queued
[rq_is_sync(rq
)]++;
1525 elv_rb_add(&cfqq
->sort_list
, rq
);
1527 if (!cfq_cfqq_on_rr(cfqq
))
1528 cfq_add_cfqq_rr(cfqd
, cfqq
);
1531 * check if this request is a better next-serve candidate
1533 prev
= cfqq
->next_rq
;
1534 cfqq
->next_rq
= cfq_choose_req(cfqd
, cfqq
->next_rq
, rq
, cfqd
->last_position
);
1537 * adjust priority tree position, if ->next_rq changes
1539 if (prev
!= cfqq
->next_rq
)
1540 cfq_prio_tree_add(cfqd
, cfqq
);
1542 BUG_ON(!cfqq
->next_rq
);
1545 static void cfq_reposition_rq_rb(struct cfq_queue
*cfqq
, struct request
*rq
)
1547 elv_rb_del(&cfqq
->sort_list
, rq
);
1548 cfqq
->queued
[rq_is_sync(rq
)]--;
1549 cfq_blkiocg_update_io_remove_stats(&(RQ_CFQG(rq
))->blkg
,
1550 rq_data_dir(rq
), rq_is_sync(rq
));
1552 cfq_blkiocg_update_io_add_stats(&(RQ_CFQG(rq
))->blkg
,
1553 &cfqq
->cfqd
->serving_group
->blkg
, rq_data_dir(rq
),
1557 static struct request
*
1558 cfq_find_rq_fmerge(struct cfq_data
*cfqd
, struct bio
*bio
)
1560 struct task_struct
*tsk
= current
;
1561 struct cfq_io_context
*cic
;
1562 struct cfq_queue
*cfqq
;
1564 cic
= cfq_cic_lookup(cfqd
, tsk
->io_context
);
1568 cfqq
= cic_to_cfqq(cic
, cfq_bio_sync(bio
));
1570 sector_t sector
= bio
->bi_sector
+ bio_sectors(bio
);
1572 return elv_rb_find(&cfqq
->sort_list
, sector
);
1578 static void cfq_activate_request(struct request_queue
*q
, struct request
*rq
)
1580 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
1582 cfqd
->rq_in_driver
++;
1583 cfq_log_cfqq(cfqd
, RQ_CFQQ(rq
), "activate rq, drv=%d",
1584 cfqd
->rq_in_driver
);
1586 cfqd
->last_position
= blk_rq_pos(rq
) + blk_rq_sectors(rq
);
1589 static void cfq_deactivate_request(struct request_queue
*q
, struct request
*rq
)
1591 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
1593 WARN_ON(!cfqd
->rq_in_driver
);
1594 cfqd
->rq_in_driver
--;
1595 cfq_log_cfqq(cfqd
, RQ_CFQQ(rq
), "deactivate rq, drv=%d",
1596 cfqd
->rq_in_driver
);
1599 static void cfq_remove_request(struct request
*rq
)
1601 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
1603 if (cfqq
->next_rq
== rq
)
1604 cfqq
->next_rq
= cfq_find_next_rq(cfqq
->cfqd
, cfqq
, rq
);
1606 list_del_init(&rq
->queuelist
);
1609 cfqq
->cfqd
->rq_queued
--;
1610 cfq_blkiocg_update_io_remove_stats(&(RQ_CFQG(rq
))->blkg
,
1611 rq_data_dir(rq
), rq_is_sync(rq
));
1612 if (rq
->cmd_flags
& REQ_PRIO
) {
1613 WARN_ON(!cfqq
->prio_pending
);
1614 cfqq
->prio_pending
--;
1618 static int cfq_merge(struct request_queue
*q
, struct request
**req
,
1621 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
1622 struct request
*__rq
;
1624 __rq
= cfq_find_rq_fmerge(cfqd
, bio
);
1625 if (__rq
&& elv_rq_merge_ok(__rq
, bio
)) {
1627 return ELEVATOR_FRONT_MERGE
;
1630 return ELEVATOR_NO_MERGE
;
1633 static void cfq_merged_request(struct request_queue
*q
, struct request
*req
,
1636 if (type
== ELEVATOR_FRONT_MERGE
) {
1637 struct cfq_queue
*cfqq
= RQ_CFQQ(req
);
1639 cfq_reposition_rq_rb(cfqq
, req
);
1643 static void cfq_bio_merged(struct request_queue
*q
, struct request
*req
,
1646 cfq_blkiocg_update_io_merged_stats(&(RQ_CFQG(req
))->blkg
,
1647 bio_data_dir(bio
), cfq_bio_sync(bio
));
1651 cfq_merged_requests(struct request_queue
*q
, struct request
*rq
,
1652 struct request
*next
)
1654 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
1656 * reposition in fifo if next is older than rq
1658 if (!list_empty(&rq
->queuelist
) && !list_empty(&next
->queuelist
) &&
1659 time_before(rq_fifo_time(next
), rq_fifo_time(rq
))) {
1660 list_move(&rq
->queuelist
, &next
->queuelist
);
1661 rq_set_fifo_time(rq
, rq_fifo_time(next
));
1664 if (cfqq
->next_rq
== next
)
1666 cfq_remove_request(next
);
1667 cfq_blkiocg_update_io_merged_stats(&(RQ_CFQG(rq
))->blkg
,
1668 rq_data_dir(next
), rq_is_sync(next
));
1671 static int cfq_allow_merge(struct request_queue
*q
, struct request
*rq
,
1674 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
1675 struct cfq_io_context
*cic
;
1676 struct cfq_queue
*cfqq
;
1679 * Disallow merge of a sync bio into an async request.
1681 if (cfq_bio_sync(bio
) && !rq_is_sync(rq
))
1685 * Lookup the cfqq that this bio will be queued with. Allow
1686 * merge only if rq is queued there.
1688 cic
= cfq_cic_lookup(cfqd
, current
->io_context
);
1692 cfqq
= cic_to_cfqq(cic
, cfq_bio_sync(bio
));
1693 return cfqq
== RQ_CFQQ(rq
);
1696 static inline void cfq_del_timer(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
1698 del_timer(&cfqd
->idle_slice_timer
);
1699 cfq_blkiocg_update_idle_time_stats(&cfqq
->cfqg
->blkg
);
1702 static void __cfq_set_active_queue(struct cfq_data
*cfqd
,
1703 struct cfq_queue
*cfqq
)
1706 cfq_log_cfqq(cfqd
, cfqq
, "set_active wl_prio:%d wl_type:%d",
1707 cfqd
->serving_prio
, cfqd
->serving_type
);
1708 cfq_blkiocg_update_avg_queue_size_stats(&cfqq
->cfqg
->blkg
);
1709 cfqq
->slice_start
= 0;
1710 cfqq
->dispatch_start
= jiffies
;
1711 cfqq
->allocated_slice
= 0;
1712 cfqq
->slice_end
= 0;
1713 cfqq
->slice_dispatch
= 0;
1714 cfqq
->nr_sectors
= 0;
1716 cfq_clear_cfqq_wait_request(cfqq
);
1717 cfq_clear_cfqq_must_dispatch(cfqq
);
1718 cfq_clear_cfqq_must_alloc_slice(cfqq
);
1719 cfq_clear_cfqq_fifo_expire(cfqq
);
1720 cfq_mark_cfqq_slice_new(cfqq
);
1722 cfq_del_timer(cfqd
, cfqq
);
1725 cfqd
->active_queue
= cfqq
;
1729 * current cfqq expired its slice (or was too idle), select new one
1732 __cfq_slice_expired(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
1735 cfq_log_cfqq(cfqd
, cfqq
, "slice expired t=%d", timed_out
);
1737 if (cfq_cfqq_wait_request(cfqq
))
1738 cfq_del_timer(cfqd
, cfqq
);
1740 cfq_clear_cfqq_wait_request(cfqq
);
1741 cfq_clear_cfqq_wait_busy(cfqq
);
1744 * If this cfqq is shared between multiple processes, check to
1745 * make sure that those processes are still issuing I/Os within
1746 * the mean seek distance. If not, it may be time to break the
1747 * queues apart again.
1749 if (cfq_cfqq_coop(cfqq
) && CFQQ_SEEKY(cfqq
))
1750 cfq_mark_cfqq_split_coop(cfqq
);
1753 * store what was left of this slice, if the queue idled/timed out
1756 if (cfq_cfqq_slice_new(cfqq
))
1757 cfqq
->slice_resid
= cfq_scaled_cfqq_slice(cfqd
, cfqq
);
1759 cfqq
->slice_resid
= cfqq
->slice_end
- jiffies
;
1760 cfq_log_cfqq(cfqd
, cfqq
, "resid=%ld", cfqq
->slice_resid
);
1763 cfq_group_served(cfqd
, cfqq
->cfqg
, cfqq
);
1765 if (cfq_cfqq_on_rr(cfqq
) && RB_EMPTY_ROOT(&cfqq
->sort_list
))
1766 cfq_del_cfqq_rr(cfqd
, cfqq
);
1768 cfq_resort_rr_list(cfqd
, cfqq
);
1770 if (cfqq
== cfqd
->active_queue
)
1771 cfqd
->active_queue
= NULL
;
1773 if (cfqd
->active_cic
) {
1774 put_io_context(cfqd
->active_cic
->ioc
);
1775 cfqd
->active_cic
= NULL
;
1779 static inline void cfq_slice_expired(struct cfq_data
*cfqd
, bool timed_out
)
1781 struct cfq_queue
*cfqq
= cfqd
->active_queue
;
1784 __cfq_slice_expired(cfqd
, cfqq
, timed_out
);
1788 * Get next queue for service. Unless we have a queue preemption,
1789 * we'll simply select the first cfqq in the service tree.
1791 static struct cfq_queue
*cfq_get_next_queue(struct cfq_data
*cfqd
)
1793 struct cfq_rb_root
*service_tree
=
1794 service_tree_for(cfqd
->serving_group
, cfqd
->serving_prio
,
1795 cfqd
->serving_type
);
1797 if (!cfqd
->rq_queued
)
1800 /* There is nothing to dispatch */
1803 if (RB_EMPTY_ROOT(&service_tree
->rb
))
1805 return cfq_rb_first(service_tree
);
1808 static struct cfq_queue
*cfq_get_next_queue_forced(struct cfq_data
*cfqd
)
1810 struct cfq_group
*cfqg
;
1811 struct cfq_queue
*cfqq
;
1813 struct cfq_rb_root
*st
;
1815 if (!cfqd
->rq_queued
)
1818 cfqg
= cfq_get_next_cfqg(cfqd
);
1822 for_each_cfqg_st(cfqg
, i
, j
, st
)
1823 if ((cfqq
= cfq_rb_first(st
)) != NULL
)
1829 * Get and set a new active queue for service.
1831 static struct cfq_queue
*cfq_set_active_queue(struct cfq_data
*cfqd
,
1832 struct cfq_queue
*cfqq
)
1835 cfqq
= cfq_get_next_queue(cfqd
);
1837 __cfq_set_active_queue(cfqd
, cfqq
);
1841 static inline sector_t
cfq_dist_from_last(struct cfq_data
*cfqd
,
1844 if (blk_rq_pos(rq
) >= cfqd
->last_position
)
1845 return blk_rq_pos(rq
) - cfqd
->last_position
;
1847 return cfqd
->last_position
- blk_rq_pos(rq
);
1850 static inline int cfq_rq_close(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
1853 return cfq_dist_from_last(cfqd
, rq
) <= CFQQ_CLOSE_THR
;
1856 static struct cfq_queue
*cfqq_close(struct cfq_data
*cfqd
,
1857 struct cfq_queue
*cur_cfqq
)
1859 struct rb_root
*root
= &cfqd
->prio_trees
[cur_cfqq
->org_ioprio
];
1860 struct rb_node
*parent
, *node
;
1861 struct cfq_queue
*__cfqq
;
1862 sector_t sector
= cfqd
->last_position
;
1864 if (RB_EMPTY_ROOT(root
))
1868 * First, if we find a request starting at the end of the last
1869 * request, choose it.
1871 __cfqq
= cfq_prio_tree_lookup(cfqd
, root
, sector
, &parent
, NULL
);
1876 * If the exact sector wasn't found, the parent of the NULL leaf
1877 * will contain the closest sector.
1879 __cfqq
= rb_entry(parent
, struct cfq_queue
, p_node
);
1880 if (cfq_rq_close(cfqd
, cur_cfqq
, __cfqq
->next_rq
))
1883 if (blk_rq_pos(__cfqq
->next_rq
) < sector
)
1884 node
= rb_next(&__cfqq
->p_node
);
1886 node
= rb_prev(&__cfqq
->p_node
);
1890 __cfqq
= rb_entry(node
, struct cfq_queue
, p_node
);
1891 if (cfq_rq_close(cfqd
, cur_cfqq
, __cfqq
->next_rq
))
1899 * cur_cfqq - passed in so that we don't decide that the current queue is
1900 * closely cooperating with itself.
1902 * So, basically we're assuming that that cur_cfqq has dispatched at least
1903 * one request, and that cfqd->last_position reflects a position on the disk
1904 * associated with the I/O issued by cur_cfqq. I'm not sure this is a valid
1907 static struct cfq_queue
*cfq_close_cooperator(struct cfq_data
*cfqd
,
1908 struct cfq_queue
*cur_cfqq
)
1910 struct cfq_queue
*cfqq
;
1912 if (cfq_class_idle(cur_cfqq
))
1914 if (!cfq_cfqq_sync(cur_cfqq
))
1916 if (CFQQ_SEEKY(cur_cfqq
))
1920 * Don't search priority tree if it's the only queue in the group.
1922 if (cur_cfqq
->cfqg
->nr_cfqq
== 1)
1926 * We should notice if some of the queues are cooperating, eg
1927 * working closely on the same area of the disk. In that case,
1928 * we can group them together and don't waste time idling.
1930 cfqq
= cfqq_close(cfqd
, cur_cfqq
);
1934 /* If new queue belongs to different cfq_group, don't choose it */
1935 if (cur_cfqq
->cfqg
!= cfqq
->cfqg
)
1939 * It only makes sense to merge sync queues.
1941 if (!cfq_cfqq_sync(cfqq
))
1943 if (CFQQ_SEEKY(cfqq
))
1947 * Do not merge queues of different priority classes
1949 if (cfq_class_rt(cfqq
) != cfq_class_rt(cur_cfqq
))
1956 * Determine whether we should enforce idle window for this queue.
1959 static bool cfq_should_idle(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
1961 enum wl_prio_t prio
= cfqq_prio(cfqq
);
1962 struct cfq_rb_root
*service_tree
= cfqq
->service_tree
;
1964 BUG_ON(!service_tree
);
1965 BUG_ON(!service_tree
->count
);
1967 if (!cfqd
->cfq_slice_idle
)
1970 /* We never do for idle class queues. */
1971 if (prio
== IDLE_WORKLOAD
)
1974 /* We do for queues that were marked with idle window flag. */
1975 if (cfq_cfqq_idle_window(cfqq
) &&
1976 !(blk_queue_nonrot(cfqd
->queue
) && cfqd
->hw_tag
))
1980 * Otherwise, we do only if they are the last ones
1981 * in their service tree.
1983 if (service_tree
->count
== 1 && cfq_cfqq_sync(cfqq
) &&
1984 !cfq_io_thinktime_big(cfqd
, &service_tree
->ttime
, false))
1986 cfq_log_cfqq(cfqd
, cfqq
, "Not idling. st->count:%d",
1987 service_tree
->count
);
1991 static void cfq_arm_slice_timer(struct cfq_data
*cfqd
)
1993 struct cfq_queue
*cfqq
= cfqd
->active_queue
;
1994 struct cfq_io_context
*cic
;
1995 unsigned long sl
, group_idle
= 0;
1998 * SSD device without seek penalty, disable idling. But only do so
1999 * for devices that support queuing, otherwise we still have a problem
2000 * with sync vs async workloads.
2002 if (blk_queue_nonrot(cfqd
->queue
) && cfqd
->hw_tag
)
2005 WARN_ON(!RB_EMPTY_ROOT(&cfqq
->sort_list
));
2006 WARN_ON(cfq_cfqq_slice_new(cfqq
));
2009 * idle is disabled, either manually or by past process history
2011 if (!cfq_should_idle(cfqd
, cfqq
)) {
2012 /* no queue idling. Check for group idling */
2013 if (cfqd
->cfq_group_idle
)
2014 group_idle
= cfqd
->cfq_group_idle
;
2020 * still active requests from this queue, don't idle
2022 if (cfqq
->dispatched
)
2026 * task has exited, don't wait
2028 cic
= cfqd
->active_cic
;
2029 if (!cic
|| !atomic_read(&cic
->ioc
->nr_tasks
))
2033 * If our average think time is larger than the remaining time
2034 * slice, then don't idle. This avoids overrunning the allotted
2037 if (sample_valid(cic
->ttime
.ttime_samples
) &&
2038 (cfqq
->slice_end
- jiffies
< cic
->ttime
.ttime_mean
)) {
2039 cfq_log_cfqq(cfqd
, cfqq
, "Not idling. think_time:%lu",
2040 cic
->ttime
.ttime_mean
);
2044 /* There are other queues in the group, don't do group idle */
2045 if (group_idle
&& cfqq
->cfqg
->nr_cfqq
> 1)
2048 cfq_mark_cfqq_wait_request(cfqq
);
2051 sl
= cfqd
->cfq_group_idle
;
2053 sl
= cfqd
->cfq_slice_idle
;
2055 mod_timer(&cfqd
->idle_slice_timer
, jiffies
+ sl
);
2056 cfq_blkiocg_update_set_idle_time_stats(&cfqq
->cfqg
->blkg
);
2057 cfq_log_cfqq(cfqd
, cfqq
, "arm_idle: %lu group_idle: %d", sl
,
2058 group_idle
? 1 : 0);
2062 * Move request from internal lists to the request queue dispatch list.
2064 static void cfq_dispatch_insert(struct request_queue
*q
, struct request
*rq
)
2066 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
2067 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
2069 cfq_log_cfqq(cfqd
, cfqq
, "dispatch_insert");
2071 cfqq
->next_rq
= cfq_find_next_rq(cfqd
, cfqq
, rq
);
2072 cfq_remove_request(rq
);
2074 (RQ_CFQG(rq
))->dispatched
++;
2075 elv_dispatch_sort(q
, rq
);
2077 cfqd
->rq_in_flight
[cfq_cfqq_sync(cfqq
)]++;
2078 cfqq
->nr_sectors
+= blk_rq_sectors(rq
);
2079 cfq_blkiocg_update_dispatch_stats(&cfqq
->cfqg
->blkg
, blk_rq_bytes(rq
),
2080 rq_data_dir(rq
), rq_is_sync(rq
));
2084 * return expired entry, or NULL to just start from scratch in rbtree
2086 static struct request
*cfq_check_fifo(struct cfq_queue
*cfqq
)
2088 struct request
*rq
= NULL
;
2090 if (cfq_cfqq_fifo_expire(cfqq
))
2093 cfq_mark_cfqq_fifo_expire(cfqq
);
2095 if (list_empty(&cfqq
->fifo
))
2098 rq
= rq_entry_fifo(cfqq
->fifo
.next
);
2099 if (time_before(jiffies
, rq_fifo_time(rq
)))
2102 cfq_log_cfqq(cfqq
->cfqd
, cfqq
, "fifo=%p", rq
);
2107 cfq_prio_to_maxrq(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
2109 const int base_rq
= cfqd
->cfq_slice_async_rq
;
2111 WARN_ON(cfqq
->ioprio
>= IOPRIO_BE_NR
);
2113 return 2 * base_rq
* (IOPRIO_BE_NR
- cfqq
->ioprio
);
2117 * Must be called with the queue_lock held.
2119 static int cfqq_process_refs(struct cfq_queue
*cfqq
)
2121 int process_refs
, io_refs
;
2123 io_refs
= cfqq
->allocated
[READ
] + cfqq
->allocated
[WRITE
];
2124 process_refs
= cfqq
->ref
- io_refs
;
2125 BUG_ON(process_refs
< 0);
2126 return process_refs
;
2129 static void cfq_setup_merge(struct cfq_queue
*cfqq
, struct cfq_queue
*new_cfqq
)
2131 int process_refs
, new_process_refs
;
2132 struct cfq_queue
*__cfqq
;
2135 * If there are no process references on the new_cfqq, then it is
2136 * unsafe to follow the ->new_cfqq chain as other cfqq's in the
2137 * chain may have dropped their last reference (not just their
2138 * last process reference).
2140 if (!cfqq_process_refs(new_cfqq
))
2143 /* Avoid a circular list and skip interim queue merges */
2144 while ((__cfqq
= new_cfqq
->new_cfqq
)) {
2150 process_refs
= cfqq_process_refs(cfqq
);
2151 new_process_refs
= cfqq_process_refs(new_cfqq
);
2153 * If the process for the cfqq has gone away, there is no
2154 * sense in merging the queues.
2156 if (process_refs
== 0 || new_process_refs
== 0)
2160 * Merge in the direction of the lesser amount of work.
2162 if (new_process_refs
>= process_refs
) {
2163 cfqq
->new_cfqq
= new_cfqq
;
2164 new_cfqq
->ref
+= process_refs
;
2166 new_cfqq
->new_cfqq
= cfqq
;
2167 cfqq
->ref
+= new_process_refs
;
2171 static enum wl_type_t
cfq_choose_wl(struct cfq_data
*cfqd
,
2172 struct cfq_group
*cfqg
, enum wl_prio_t prio
)
2174 struct cfq_queue
*queue
;
2176 bool key_valid
= false;
2177 unsigned long lowest_key
= 0;
2178 enum wl_type_t cur_best
= SYNC_NOIDLE_WORKLOAD
;
2180 for (i
= 0; i
<= SYNC_WORKLOAD
; ++i
) {
2181 /* select the one with lowest rb_key */
2182 queue
= cfq_rb_first(service_tree_for(cfqg
, prio
, i
));
2184 (!key_valid
|| time_before(queue
->rb_key
, lowest_key
))) {
2185 lowest_key
= queue
->rb_key
;
2194 static void choose_service_tree(struct cfq_data
*cfqd
, struct cfq_group
*cfqg
)
2198 struct cfq_rb_root
*st
;
2199 unsigned group_slice
;
2200 enum wl_prio_t original_prio
= cfqd
->serving_prio
;
2202 /* Choose next priority. RT > BE > IDLE */
2203 if (cfq_group_busy_queues_wl(RT_WORKLOAD
, cfqd
, cfqg
))
2204 cfqd
->serving_prio
= RT_WORKLOAD
;
2205 else if (cfq_group_busy_queues_wl(BE_WORKLOAD
, cfqd
, cfqg
))
2206 cfqd
->serving_prio
= BE_WORKLOAD
;
2208 cfqd
->serving_prio
= IDLE_WORKLOAD
;
2209 cfqd
->workload_expires
= jiffies
+ 1;
2213 if (original_prio
!= cfqd
->serving_prio
)
2217 * For RT and BE, we have to choose also the type
2218 * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
2221 st
= service_tree_for(cfqg
, cfqd
->serving_prio
, cfqd
->serving_type
);
2225 * check workload expiration, and that we still have other queues ready
2227 if (count
&& !time_after(jiffies
, cfqd
->workload_expires
))
2231 /* otherwise select new workload type */
2232 cfqd
->serving_type
=
2233 cfq_choose_wl(cfqd
, cfqg
, cfqd
->serving_prio
);
2234 st
= service_tree_for(cfqg
, cfqd
->serving_prio
, cfqd
->serving_type
);
2238 * the workload slice is computed as a fraction of target latency
2239 * proportional to the number of queues in that workload, over
2240 * all the queues in the same priority class
2242 group_slice
= cfq_group_slice(cfqd
, cfqg
);
2244 slice
= group_slice
* count
/
2245 max_t(unsigned, cfqg
->busy_queues_avg
[cfqd
->serving_prio
],
2246 cfq_group_busy_queues_wl(cfqd
->serving_prio
, cfqd
, cfqg
));
2248 if (cfqd
->serving_type
== ASYNC_WORKLOAD
) {
2252 * Async queues are currently system wide. Just taking
2253 * proportion of queues with-in same group will lead to higher
2254 * async ratio system wide as generally root group is going
2255 * to have higher weight. A more accurate thing would be to
2256 * calculate system wide asnc/sync ratio.
2258 tmp
= cfq_target_latency
* cfqg_busy_async_queues(cfqd
, cfqg
);
2259 tmp
= tmp
/cfqd
->busy_queues
;
2260 slice
= min_t(unsigned, slice
, tmp
);
2262 /* async workload slice is scaled down according to
2263 * the sync/async slice ratio. */
2264 slice
= slice
* cfqd
->cfq_slice
[0] / cfqd
->cfq_slice
[1];
2266 /* sync workload slice is at least 2 * cfq_slice_idle */
2267 slice
= max(slice
, 2 * cfqd
->cfq_slice_idle
);
2269 slice
= max_t(unsigned, slice
, CFQ_MIN_TT
);
2270 cfq_log(cfqd
, "workload slice:%d", slice
);
2271 cfqd
->workload_expires
= jiffies
+ slice
;
2274 static struct cfq_group
*cfq_get_next_cfqg(struct cfq_data
*cfqd
)
2276 struct cfq_rb_root
*st
= &cfqd
->grp_service_tree
;
2277 struct cfq_group
*cfqg
;
2279 if (RB_EMPTY_ROOT(&st
->rb
))
2281 cfqg
= cfq_rb_first_group(st
);
2282 update_min_vdisktime(st
);
2286 static void cfq_choose_cfqg(struct cfq_data
*cfqd
)
2288 struct cfq_group
*cfqg
= cfq_get_next_cfqg(cfqd
);
2290 cfqd
->serving_group
= cfqg
;
2292 /* Restore the workload type data */
2293 if (cfqg
->saved_workload_slice
) {
2294 cfqd
->workload_expires
= jiffies
+ cfqg
->saved_workload_slice
;
2295 cfqd
->serving_type
= cfqg
->saved_workload
;
2296 cfqd
->serving_prio
= cfqg
->saved_serving_prio
;
2298 cfqd
->workload_expires
= jiffies
- 1;
2300 choose_service_tree(cfqd
, cfqg
);
2304 * Select a queue for service. If we have a current active queue,
2305 * check whether to continue servicing it, or retrieve and set a new one.
2307 static struct cfq_queue
*cfq_select_queue(struct cfq_data
*cfqd
)
2309 struct cfq_queue
*cfqq
, *new_cfqq
= NULL
;
2311 cfqq
= cfqd
->active_queue
;
2315 if (!cfqd
->rq_queued
)
2319 * We were waiting for group to get backlogged. Expire the queue
2321 if (cfq_cfqq_wait_busy(cfqq
) && !RB_EMPTY_ROOT(&cfqq
->sort_list
))
2325 * The active queue has run out of time, expire it and select new.
2327 if (cfq_slice_used(cfqq
) && !cfq_cfqq_must_dispatch(cfqq
)) {
2329 * If slice had not expired at the completion of last request
2330 * we might not have turned on wait_busy flag. Don't expire
2331 * the queue yet. Allow the group to get backlogged.
2333 * The very fact that we have used the slice, that means we
2334 * have been idling all along on this queue and it should be
2335 * ok to wait for this request to complete.
2337 if (cfqq
->cfqg
->nr_cfqq
== 1 && RB_EMPTY_ROOT(&cfqq
->sort_list
)
2338 && cfqq
->dispatched
&& cfq_should_idle(cfqd
, cfqq
)) {
2342 goto check_group_idle
;
2346 * The active queue has requests and isn't expired, allow it to
2349 if (!RB_EMPTY_ROOT(&cfqq
->sort_list
))
2353 * If another queue has a request waiting within our mean seek
2354 * distance, let it run. The expire code will check for close
2355 * cooperators and put the close queue at the front of the service
2356 * tree. If possible, merge the expiring queue with the new cfqq.
2358 new_cfqq
= cfq_close_cooperator(cfqd
, cfqq
);
2360 if (!cfqq
->new_cfqq
)
2361 cfq_setup_merge(cfqq
, new_cfqq
);
2366 * No requests pending. If the active queue still has requests in
2367 * flight or is idling for a new request, allow either of these
2368 * conditions to happen (or time out) before selecting a new queue.
2370 if (timer_pending(&cfqd
->idle_slice_timer
)) {
2376 * This is a deep seek queue, but the device is much faster than
2377 * the queue can deliver, don't idle
2379 if (CFQQ_SEEKY(cfqq
) && cfq_cfqq_idle_window(cfqq
) &&
2380 (cfq_cfqq_slice_new(cfqq
) ||
2381 (cfqq
->slice_end
- jiffies
> jiffies
- cfqq
->slice_start
))) {
2382 cfq_clear_cfqq_deep(cfqq
);
2383 cfq_clear_cfqq_idle_window(cfqq
);
2386 if (cfqq
->dispatched
&& cfq_should_idle(cfqd
, cfqq
)) {
2392 * If group idle is enabled and there are requests dispatched from
2393 * this group, wait for requests to complete.
2396 if (cfqd
->cfq_group_idle
&& cfqq
->cfqg
->nr_cfqq
== 1 &&
2397 cfqq
->cfqg
->dispatched
&&
2398 !cfq_io_thinktime_big(cfqd
, &cfqq
->cfqg
->ttime
, true)) {
2404 cfq_slice_expired(cfqd
, 0);
2407 * Current queue expired. Check if we have to switch to a new
2411 cfq_choose_cfqg(cfqd
);
2413 cfqq
= cfq_set_active_queue(cfqd
, new_cfqq
);
2418 static int __cfq_forced_dispatch_cfqq(struct cfq_queue
*cfqq
)
2422 while (cfqq
->next_rq
) {
2423 cfq_dispatch_insert(cfqq
->cfqd
->queue
, cfqq
->next_rq
);
2427 BUG_ON(!list_empty(&cfqq
->fifo
));
2429 /* By default cfqq is not expired if it is empty. Do it explicitly */
2430 __cfq_slice_expired(cfqq
->cfqd
, cfqq
, 0);
2435 * Drain our current requests. Used for barriers and when switching
2436 * io schedulers on-the-fly.
2438 static int cfq_forced_dispatch(struct cfq_data
*cfqd
)
2440 struct cfq_queue
*cfqq
;
2443 /* Expire the timeslice of the current active queue first */
2444 cfq_slice_expired(cfqd
, 0);
2445 while ((cfqq
= cfq_get_next_queue_forced(cfqd
)) != NULL
) {
2446 __cfq_set_active_queue(cfqd
, cfqq
);
2447 dispatched
+= __cfq_forced_dispatch_cfqq(cfqq
);
2450 BUG_ON(cfqd
->busy_queues
);
2452 cfq_log(cfqd
, "forced_dispatch=%d", dispatched
);
2456 static inline bool cfq_slice_used_soon(struct cfq_data
*cfqd
,
2457 struct cfq_queue
*cfqq
)
2459 /* the queue hasn't finished any request, can't estimate */
2460 if (cfq_cfqq_slice_new(cfqq
))
2462 if (time_after(jiffies
+ cfqd
->cfq_slice_idle
* cfqq
->dispatched
,
2469 static bool cfq_may_dispatch(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
2471 unsigned int max_dispatch
;
2474 * Drain async requests before we start sync IO
2476 if (cfq_should_idle(cfqd
, cfqq
) && cfqd
->rq_in_flight
[BLK_RW_ASYNC
])
2480 * If this is an async queue and we have sync IO in flight, let it wait
2482 if (cfqd
->rq_in_flight
[BLK_RW_SYNC
] && !cfq_cfqq_sync(cfqq
))
2485 max_dispatch
= max_t(unsigned int, cfqd
->cfq_quantum
/ 2, 1);
2486 if (cfq_class_idle(cfqq
))
2490 * Does this cfqq already have too much IO in flight?
2492 if (cfqq
->dispatched
>= max_dispatch
) {
2493 bool promote_sync
= false;
2495 * idle queue must always only have a single IO in flight
2497 if (cfq_class_idle(cfqq
))
2501 * If there is only one sync queue
2502 * we can ignore async queue here and give the sync
2503 * queue no dispatch limit. The reason is a sync queue can
2504 * preempt async queue, limiting the sync queue doesn't make
2505 * sense. This is useful for aiostress test.
2507 if (cfq_cfqq_sync(cfqq
) && cfqd
->busy_sync_queues
== 1)
2508 promote_sync
= true;
2511 * We have other queues, don't allow more IO from this one
2513 if (cfqd
->busy_queues
> 1 && cfq_slice_used_soon(cfqd
, cfqq
) &&
2518 * Sole queue user, no limit
2520 if (cfqd
->busy_queues
== 1 || promote_sync
)
2524 * Normally we start throttling cfqq when cfq_quantum/2
2525 * requests have been dispatched. But we can drive
2526 * deeper queue depths at the beginning of slice
2527 * subjected to upper limit of cfq_quantum.
2529 max_dispatch
= cfqd
->cfq_quantum
;
2533 * Async queues must wait a bit before being allowed dispatch.
2534 * We also ramp up the dispatch depth gradually for async IO,
2535 * based on the last sync IO we serviced
2537 if (!cfq_cfqq_sync(cfqq
) && cfqd
->cfq_latency
) {
2538 unsigned long last_sync
= jiffies
- cfqd
->last_delayed_sync
;
2541 depth
= last_sync
/ cfqd
->cfq_slice
[1];
2542 if (!depth
&& !cfqq
->dispatched
)
2544 if (depth
< max_dispatch
)
2545 max_dispatch
= depth
;
2549 * If we're below the current max, allow a dispatch
2551 return cfqq
->dispatched
< max_dispatch
;
2555 * Dispatch a request from cfqq, moving them to the request queue
2558 static bool cfq_dispatch_request(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
2562 BUG_ON(RB_EMPTY_ROOT(&cfqq
->sort_list
));
2564 if (!cfq_may_dispatch(cfqd
, cfqq
))
2568 * follow expired path, else get first next available
2570 rq
= cfq_check_fifo(cfqq
);
2575 * insert request into driver dispatch list
2577 cfq_dispatch_insert(cfqd
->queue
, rq
);
2579 if (!cfqd
->active_cic
) {
2580 struct cfq_io_context
*cic
= RQ_CIC(rq
);
2582 atomic_long_inc(&cic
->ioc
->refcount
);
2583 cfqd
->active_cic
= cic
;
2590 * Find the cfqq that we need to service and move a request from that to the
2593 static int cfq_dispatch_requests(struct request_queue
*q
, int force
)
2595 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
2596 struct cfq_queue
*cfqq
;
2598 if (!cfqd
->busy_queues
)
2601 if (unlikely(force
))
2602 return cfq_forced_dispatch(cfqd
);
2604 cfqq
= cfq_select_queue(cfqd
);
2609 * Dispatch a request from this cfqq, if it is allowed
2611 if (!cfq_dispatch_request(cfqd
, cfqq
))
2614 cfqq
->slice_dispatch
++;
2615 cfq_clear_cfqq_must_dispatch(cfqq
);
2618 * expire an async queue immediately if it has used up its slice. idle
2619 * queue always expire after 1 dispatch round.
2621 if (cfqd
->busy_queues
> 1 && ((!cfq_cfqq_sync(cfqq
) &&
2622 cfqq
->slice_dispatch
>= cfq_prio_to_maxrq(cfqd
, cfqq
)) ||
2623 cfq_class_idle(cfqq
))) {
2624 cfqq
->slice_end
= jiffies
+ 1;
2625 cfq_slice_expired(cfqd
, 0);
2628 cfq_log_cfqq(cfqd
, cfqq
, "dispatched a request");
2633 * task holds one reference to the queue, dropped when task exits. each rq
2634 * in-flight on this queue also holds a reference, dropped when rq is freed.
2636 * Each cfq queue took a reference on the parent group. Drop it now.
2637 * queue lock must be held here.
2639 static void cfq_put_queue(struct cfq_queue
*cfqq
)
2641 struct cfq_data
*cfqd
= cfqq
->cfqd
;
2642 struct cfq_group
*cfqg
;
2644 BUG_ON(cfqq
->ref
<= 0);
2650 cfq_log_cfqq(cfqd
, cfqq
, "put_queue");
2651 BUG_ON(rb_first(&cfqq
->sort_list
));
2652 BUG_ON(cfqq
->allocated
[READ
] + cfqq
->allocated
[WRITE
]);
2655 if (unlikely(cfqd
->active_queue
== cfqq
)) {
2656 __cfq_slice_expired(cfqd
, cfqq
, 0);
2657 cfq_schedule_dispatch(cfqd
);
2660 BUG_ON(cfq_cfqq_on_rr(cfqq
));
2661 kmem_cache_free(cfq_pool
, cfqq
);
2666 * Call func for each cic attached to this ioc.
2669 call_for_each_cic(struct io_context
*ioc
,
2670 void (*func
)(struct io_context
*, struct cfq_io_context
*))
2672 struct cfq_io_context
*cic
;
2673 struct hlist_node
*n
;
2677 hlist_for_each_entry_rcu(cic
, n
, &ioc
->cic_list
, cic_list
)
2683 static void cfq_cic_free_rcu(struct rcu_head
*head
)
2685 struct cfq_io_context
*cic
;
2687 cic
= container_of(head
, struct cfq_io_context
, rcu_head
);
2689 kmem_cache_free(cfq_ioc_pool
, cic
);
2690 elv_ioc_count_dec(cfq_ioc_count
);
2694 * CFQ scheduler is exiting, grab exit lock and check
2695 * the pending io context count. If it hits zero,
2696 * complete ioc_gone and set it back to NULL
2698 spin_lock(&ioc_gone_lock
);
2699 if (ioc_gone
&& !elv_ioc_count_read(cfq_ioc_count
)) {
2703 spin_unlock(&ioc_gone_lock
);
2707 static void cfq_cic_free(struct cfq_io_context
*cic
)
2709 call_rcu(&cic
->rcu_head
, cfq_cic_free_rcu
);
2712 static void cfq_release_cic(struct cfq_io_context
*cic
)
2714 struct io_context
*ioc
= cic
->ioc
;
2715 unsigned long dead_key
= (unsigned long) cic
->key
;
2717 BUG_ON(!(dead_key
& CIC_DEAD_KEY
));
2718 radix_tree_delete(&ioc
->radix_root
, dead_key
>> CIC_DEAD_INDEX_SHIFT
);
2719 hlist_del_rcu(&cic
->cic_list
);
2723 static void cic_free_func(struct io_context
*ioc
, struct cfq_io_context
*cic
)
2725 unsigned long flags
;
2727 spin_lock_irqsave(&ioc
->lock
, flags
);
2728 cfq_release_cic(cic
);
2729 spin_unlock_irqrestore(&ioc
->lock
, flags
);
2733 * Must be called with rcu_read_lock() held or preemption otherwise disabled.
2734 * Only two callers of this - ->dtor() which is called with the rcu_read_lock(),
2735 * and ->trim() which is called with the task lock held
2737 static void cfq_free_io_context(struct io_context
*ioc
)
2740 * ioc->refcount is zero here, or we are called from elv_unregister(),
2741 * so no more cic's are allowed to be linked into this ioc. So it
2742 * should be ok to iterate over the known list, we will see all cic's
2743 * since no new ones are added.
2745 call_for_each_cic(ioc
, cic_free_func
);
2748 static void cfq_put_cooperator(struct cfq_queue
*cfqq
)
2750 struct cfq_queue
*__cfqq
, *next
;
2753 * If this queue was scheduled to merge with another queue, be
2754 * sure to drop the reference taken on that queue (and others in
2755 * the merge chain). See cfq_setup_merge and cfq_merge_cfqqs.
2757 __cfqq
= cfqq
->new_cfqq
;
2759 if (__cfqq
== cfqq
) {
2760 WARN(1, "cfqq->new_cfqq loop detected\n");
2763 next
= __cfqq
->new_cfqq
;
2764 cfq_put_queue(__cfqq
);
2769 static void cfq_exit_cfqq(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
2771 if (unlikely(cfqq
== cfqd
->active_queue
)) {
2772 __cfq_slice_expired(cfqd
, cfqq
, 0);
2773 cfq_schedule_dispatch(cfqd
);
2776 cfq_put_cooperator(cfqq
);
2778 cfq_put_queue(cfqq
);
2781 static void cfq_exit_cic(struct cfq_io_context
*cic
)
2783 struct cfq_data
*cfqd
= cic_to_cfqd(cic
);
2784 struct io_context
*ioc
= cic
->ioc
;
2786 list_del_init(&cic
->queue_list
);
2789 * Make sure dead mark is seen for dead queues
2792 cic
->key
= cfqd_dead_key(cfqd
);
2795 if (rcu_dereference(ioc
->ioc_data
) == cic
) {
2797 spin_lock(&ioc
->lock
);
2798 rcu_assign_pointer(ioc
->ioc_data
, NULL
);
2799 spin_unlock(&ioc
->lock
);
2803 if (cic
->cfqq
[BLK_RW_ASYNC
]) {
2804 cfq_exit_cfqq(cfqd
, cic
->cfqq
[BLK_RW_ASYNC
]);
2805 cic
->cfqq
[BLK_RW_ASYNC
] = NULL
;
2808 if (cic
->cfqq
[BLK_RW_SYNC
]) {
2809 cfq_exit_cfqq(cfqd
, cic
->cfqq
[BLK_RW_SYNC
]);
2810 cic
->cfqq
[BLK_RW_SYNC
] = NULL
;
2814 static void cfq_exit_single_io_context(struct io_context
*ioc
,
2815 struct cfq_io_context
*cic
)
2817 struct cfq_data
*cfqd
= cic_to_cfqd(cic
);
2820 struct request_queue
*q
= cfqd
->queue
;
2821 unsigned long flags
;
2823 spin_lock_irqsave(q
->queue_lock
, flags
);
2826 * Ensure we get a fresh copy of the ->key to prevent
2827 * race between exiting task and queue
2829 smp_read_barrier_depends();
2830 if (cic
->key
== cfqd
)
2833 spin_unlock_irqrestore(q
->queue_lock
, flags
);
2838 * The process that ioc belongs to has exited, we need to clean up
2839 * and put the internal structures we have that belongs to that process.
2841 static void cfq_exit_io_context(struct io_context
*ioc
)
2843 call_for_each_cic(ioc
, cfq_exit_single_io_context
);
2846 static struct cfq_io_context
*
2847 cfq_alloc_io_context(struct cfq_data
*cfqd
, gfp_t gfp_mask
)
2849 struct cfq_io_context
*cic
;
2851 cic
= kmem_cache_alloc_node(cfq_ioc_pool
, gfp_mask
| __GFP_ZERO
,
2854 cic
->ttime
.last_end_request
= jiffies
;
2855 INIT_LIST_HEAD(&cic
->queue_list
);
2856 INIT_HLIST_NODE(&cic
->cic_list
);
2857 cic
->dtor
= cfq_free_io_context
;
2858 cic
->exit
= cfq_exit_io_context
;
2859 elv_ioc_count_inc(cfq_ioc_count
);
2865 static void cfq_init_prio_data(struct cfq_queue
*cfqq
, struct io_context
*ioc
)
2867 struct task_struct
*tsk
= current
;
2870 if (!cfq_cfqq_prio_changed(cfqq
))
2873 ioprio_class
= IOPRIO_PRIO_CLASS(ioc
->ioprio
);
2874 switch (ioprio_class
) {
2876 printk(KERN_ERR
"cfq: bad prio %x\n", ioprio_class
);
2877 case IOPRIO_CLASS_NONE
:
2879 * no prio set, inherit CPU scheduling settings
2881 cfqq
->ioprio
= task_nice_ioprio(tsk
);
2882 cfqq
->ioprio_class
= task_nice_ioclass(tsk
);
2884 case IOPRIO_CLASS_RT
:
2885 cfqq
->ioprio
= task_ioprio(ioc
);
2886 cfqq
->ioprio_class
= IOPRIO_CLASS_RT
;
2888 case IOPRIO_CLASS_BE
:
2889 cfqq
->ioprio
= task_ioprio(ioc
);
2890 cfqq
->ioprio_class
= IOPRIO_CLASS_BE
;
2892 case IOPRIO_CLASS_IDLE
:
2893 cfqq
->ioprio_class
= IOPRIO_CLASS_IDLE
;
2895 cfq_clear_cfqq_idle_window(cfqq
);
2900 * keep track of original prio settings in case we have to temporarily
2901 * elevate the priority of this queue
2903 cfqq
->org_ioprio
= cfqq
->ioprio
;
2904 cfq_clear_cfqq_prio_changed(cfqq
);
2907 static void changed_ioprio(struct cfq_io_context
*cic
)
2909 struct cfq_data
*cfqd
= cic_to_cfqd(cic
);
2910 struct cfq_queue
*cfqq
;
2911 unsigned long flags
;
2913 if (unlikely(!cfqd
))
2916 spin_lock_irqsave(cfqd
->queue
->queue_lock
, flags
);
2918 cfqq
= cic
->cfqq
[BLK_RW_ASYNC
];
2920 struct cfq_queue
*new_cfqq
;
2921 new_cfqq
= cfq_get_queue(cfqd
, BLK_RW_ASYNC
, cic
->ioc
,
2924 cic
->cfqq
[BLK_RW_ASYNC
] = new_cfqq
;
2925 cfq_put_queue(cfqq
);
2929 cfqq
= cic
->cfqq
[BLK_RW_SYNC
];
2931 cfq_mark_cfqq_prio_changed(cfqq
);
2933 spin_unlock_irqrestore(cfqd
->queue
->queue_lock
, flags
);
2936 static void cfq_init_cfqq(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
2937 pid_t pid
, bool is_sync
)
2939 RB_CLEAR_NODE(&cfqq
->rb_node
);
2940 RB_CLEAR_NODE(&cfqq
->p_node
);
2941 INIT_LIST_HEAD(&cfqq
->fifo
);
2946 cfq_mark_cfqq_prio_changed(cfqq
);
2949 if (!cfq_class_idle(cfqq
))
2950 cfq_mark_cfqq_idle_window(cfqq
);
2951 cfq_mark_cfqq_sync(cfqq
);
2956 #ifdef CONFIG_CFQ_GROUP_IOSCHED
2957 static void changed_cgroup(struct cfq_io_context
*cic
)
2959 struct cfq_queue
*sync_cfqq
= cic_to_cfqq(cic
, 1);
2960 struct cfq_data
*cfqd
= cic_to_cfqd(cic
);
2961 unsigned long flags
;
2962 struct request_queue
*q
;
2964 if (unlikely(!cfqd
))
2969 spin_lock_irqsave(q
->queue_lock
, flags
);
2973 * Drop reference to sync queue. A new sync queue will be
2974 * assigned in new group upon arrival of a fresh request.
2976 cfq_log_cfqq(cfqd
, sync_cfqq
, "changed cgroup");
2977 cic_set_cfqq(cic
, NULL
, 1);
2978 cfq_put_queue(sync_cfqq
);
2981 spin_unlock_irqrestore(q
->queue_lock
, flags
);
2983 #endif /* CONFIG_CFQ_GROUP_IOSCHED */
2985 static struct cfq_queue
*
2986 cfq_find_alloc_queue(struct cfq_data
*cfqd
, bool is_sync
,
2987 struct io_context
*ioc
, gfp_t gfp_mask
)
2989 struct cfq_queue
*cfqq
, *new_cfqq
= NULL
;
2990 struct cfq_io_context
*cic
;
2991 struct cfq_group
*cfqg
;
2994 cfqg
= cfq_get_cfqg(cfqd
);
2995 cic
= cfq_cic_lookup(cfqd
, ioc
);
2996 /* cic always exists here */
2997 cfqq
= cic_to_cfqq(cic
, is_sync
);
3000 * Always try a new alloc if we fell back to the OOM cfqq
3001 * originally, since it should just be a temporary situation.
3003 if (!cfqq
|| cfqq
== &cfqd
->oom_cfqq
) {
3008 } else if (gfp_mask
& __GFP_WAIT
) {
3009 spin_unlock_irq(cfqd
->queue
->queue_lock
);
3010 new_cfqq
= kmem_cache_alloc_node(cfq_pool
,
3011 gfp_mask
| __GFP_ZERO
,
3013 spin_lock_irq(cfqd
->queue
->queue_lock
);
3017 cfqq
= kmem_cache_alloc_node(cfq_pool
,
3018 gfp_mask
| __GFP_ZERO
,
3023 cfq_init_cfqq(cfqd
, cfqq
, current
->pid
, is_sync
);
3024 cfq_init_prio_data(cfqq
, ioc
);
3025 cfq_link_cfqq_cfqg(cfqq
, cfqg
);
3026 cfq_log_cfqq(cfqd
, cfqq
, "alloced");
3028 cfqq
= &cfqd
->oom_cfqq
;
3032 kmem_cache_free(cfq_pool
, new_cfqq
);
3037 static struct cfq_queue
**
3038 cfq_async_queue_prio(struct cfq_data
*cfqd
, int ioprio_class
, int ioprio
)
3040 switch (ioprio_class
) {
3041 case IOPRIO_CLASS_RT
:
3042 return &cfqd
->async_cfqq
[0][ioprio
];
3043 case IOPRIO_CLASS_BE
:
3044 return &cfqd
->async_cfqq
[1][ioprio
];
3045 case IOPRIO_CLASS_IDLE
:
3046 return &cfqd
->async_idle_cfqq
;
3052 static struct cfq_queue
*
3053 cfq_get_queue(struct cfq_data
*cfqd
, bool is_sync
, struct io_context
*ioc
,
3056 const int ioprio
= task_ioprio(ioc
);
3057 const int ioprio_class
= task_ioprio_class(ioc
);
3058 struct cfq_queue
**async_cfqq
= NULL
;
3059 struct cfq_queue
*cfqq
= NULL
;
3062 async_cfqq
= cfq_async_queue_prio(cfqd
, ioprio_class
, ioprio
);
3067 cfqq
= cfq_find_alloc_queue(cfqd
, is_sync
, ioc
, gfp_mask
);
3070 * pin the queue now that it's allocated, scheduler exit will prune it
3072 if (!is_sync
&& !(*async_cfqq
)) {
3082 * We drop cfq io contexts lazily, so we may find a dead one.
3085 cfq_drop_dead_cic(struct cfq_data
*cfqd
, struct io_context
*ioc
,
3086 struct cfq_io_context
*cic
)
3088 unsigned long flags
;
3090 WARN_ON(!list_empty(&cic
->queue_list
));
3091 BUG_ON(cic
->key
!= cfqd_dead_key(cfqd
));
3093 spin_lock_irqsave(&ioc
->lock
, flags
);
3095 BUG_ON(rcu_dereference_check(ioc
->ioc_data
,
3096 lockdep_is_held(&ioc
->lock
)) == cic
);
3098 radix_tree_delete(&ioc
->radix_root
, cfqd
->queue
->id
);
3099 hlist_del_rcu(&cic
->cic_list
);
3100 spin_unlock_irqrestore(&ioc
->lock
, flags
);
3105 static struct cfq_io_context
*
3106 cfq_cic_lookup(struct cfq_data
*cfqd
, struct io_context
*ioc
)
3108 struct cfq_io_context
*cic
;
3109 unsigned long flags
;
3117 * we maintain a last-hit cache, to avoid browsing over the tree
3119 cic
= rcu_dereference(ioc
->ioc_data
);
3120 if (cic
&& cic
->key
== cfqd
) {
3126 cic
= radix_tree_lookup(&ioc
->radix_root
, cfqd
->queue
->id
);
3130 if (unlikely(cic
->key
!= cfqd
)) {
3131 cfq_drop_dead_cic(cfqd
, ioc
, cic
);
3136 spin_lock_irqsave(&ioc
->lock
, flags
);
3137 rcu_assign_pointer(ioc
->ioc_data
, cic
);
3138 spin_unlock_irqrestore(&ioc
->lock
, flags
);
3146 * Add cic into ioc, using cfqd as the search key. This enables us to lookup
3147 * the process specific cfq io context when entered from the block layer.
3148 * Also adds the cic to a per-cfqd list, used when this queue is removed.
3150 static int cfq_cic_link(struct cfq_data
*cfqd
, struct io_context
*ioc
,
3151 struct cfq_io_context
*cic
, gfp_t gfp_mask
)
3153 unsigned long flags
;
3156 ret
= radix_tree_preload(gfp_mask
);
3162 cic
->q
= cfqd
->queue
;
3164 spin_lock_irqsave(&ioc
->lock
, flags
);
3165 ret
= radix_tree_insert(&ioc
->radix_root
, cfqd
->queue
->id
, cic
);
3167 hlist_add_head_rcu(&cic
->cic_list
, &ioc
->cic_list
);
3168 spin_unlock_irqrestore(&ioc
->lock
, flags
);
3170 radix_tree_preload_end();
3173 spin_lock_irqsave(cfqd
->queue
->queue_lock
, flags
);
3174 list_add(&cic
->queue_list
, &cfqd
->cic_list
);
3175 spin_unlock_irqrestore(cfqd
->queue
->queue_lock
, flags
);
3179 printk(KERN_ERR
"cfq: cic link failed!\n");
3184 * Setup general io context and cfq io context. There can be several cfq
3185 * io contexts per general io context, if this process is doing io to more
3186 * than one device managed by cfq.
3188 static struct cfq_io_context
*
3189 cfq_get_io_context(struct cfq_data
*cfqd
, gfp_t gfp_mask
)
3191 struct io_context
*ioc
= NULL
;
3192 struct cfq_io_context
*cic
= NULL
;
3194 might_sleep_if(gfp_mask
& __GFP_WAIT
);
3196 ioc
= current_io_context(gfp_mask
, cfqd
->queue
->node
);
3200 cic
= cfq_cic_lookup(cfqd
, ioc
);
3204 cic
= cfq_alloc_io_context(cfqd
, gfp_mask
);
3208 if (cfq_cic_link(cfqd
, ioc
, cic
, gfp_mask
))
3211 get_io_context(ioc
);
3213 if (unlikely(cic
->changed
)) {
3214 if (test_and_clear_bit(CIC_IOPRIO_CHANGED
, &cic
->changed
))
3215 changed_ioprio(cic
);
3216 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3217 if (test_and_clear_bit(CIC_CGROUP_CHANGED
, &cic
->changed
))
3218 changed_cgroup(cic
);
3230 __cfq_update_io_thinktime(struct cfq_ttime
*ttime
, unsigned long slice_idle
)
3232 unsigned long elapsed
= jiffies
- ttime
->last_end_request
;
3233 elapsed
= min(elapsed
, 2UL * slice_idle
);
3235 ttime
->ttime_samples
= (7*ttime
->ttime_samples
+ 256) / 8;
3236 ttime
->ttime_total
= (7*ttime
->ttime_total
+ 256*elapsed
) / 8;
3237 ttime
->ttime_mean
= (ttime
->ttime_total
+ 128) / ttime
->ttime_samples
;
3241 cfq_update_io_thinktime(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
3242 struct cfq_io_context
*cic
)
3244 if (cfq_cfqq_sync(cfqq
)) {
3245 __cfq_update_io_thinktime(&cic
->ttime
, cfqd
->cfq_slice_idle
);
3246 __cfq_update_io_thinktime(&cfqq
->service_tree
->ttime
,
3247 cfqd
->cfq_slice_idle
);
3249 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3250 __cfq_update_io_thinktime(&cfqq
->cfqg
->ttime
, cfqd
->cfq_group_idle
);
3255 cfq_update_io_seektime(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
3259 sector_t n_sec
= blk_rq_sectors(rq
);
3260 if (cfqq
->last_request_pos
) {
3261 if (cfqq
->last_request_pos
< blk_rq_pos(rq
))
3262 sdist
= blk_rq_pos(rq
) - cfqq
->last_request_pos
;
3264 sdist
= cfqq
->last_request_pos
- blk_rq_pos(rq
);
3267 cfqq
->seek_history
<<= 1;
3268 if (blk_queue_nonrot(cfqd
->queue
))
3269 cfqq
->seek_history
|= (n_sec
< CFQQ_SECT_THR_NONROT
);
3271 cfqq
->seek_history
|= (sdist
> CFQQ_SEEK_THR
);
3275 * Disable idle window if the process thinks too long or seeks so much that
3279 cfq_update_idle_window(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
3280 struct cfq_io_context
*cic
)
3282 int old_idle
, enable_idle
;
3285 * Don't idle for async or idle io prio class
3287 if (!cfq_cfqq_sync(cfqq
) || cfq_class_idle(cfqq
))
3290 enable_idle
= old_idle
= cfq_cfqq_idle_window(cfqq
);
3292 if (cfqq
->queued
[0] + cfqq
->queued
[1] >= 4)
3293 cfq_mark_cfqq_deep(cfqq
);
3295 if (cfqq
->next_rq
&& (cfqq
->next_rq
->cmd_flags
& REQ_NOIDLE
))
3297 else if (!atomic_read(&cic
->ioc
->nr_tasks
) || !cfqd
->cfq_slice_idle
||
3298 (!cfq_cfqq_deep(cfqq
) && CFQQ_SEEKY(cfqq
)))
3300 else if (sample_valid(cic
->ttime
.ttime_samples
)) {
3301 if (cic
->ttime
.ttime_mean
> cfqd
->cfq_slice_idle
)
3307 if (old_idle
!= enable_idle
) {
3308 cfq_log_cfqq(cfqd
, cfqq
, "idle=%d", enable_idle
);
3310 cfq_mark_cfqq_idle_window(cfqq
);
3312 cfq_clear_cfqq_idle_window(cfqq
);
3317 * Check if new_cfqq should preempt the currently active queue. Return 0 for
3318 * no or if we aren't sure, a 1 will cause a preempt.
3321 cfq_should_preempt(struct cfq_data
*cfqd
, struct cfq_queue
*new_cfqq
,
3324 struct cfq_queue
*cfqq
;
3326 cfqq
= cfqd
->active_queue
;
3330 if (cfq_class_idle(new_cfqq
))
3333 if (cfq_class_idle(cfqq
))
3337 * Don't allow a non-RT request to preempt an ongoing RT cfqq timeslice.
3339 if (cfq_class_rt(cfqq
) && !cfq_class_rt(new_cfqq
))
3343 * if the new request is sync, but the currently running queue is
3344 * not, let the sync request have priority.
3346 if (rq_is_sync(rq
) && !cfq_cfqq_sync(cfqq
))
3349 if (new_cfqq
->cfqg
!= cfqq
->cfqg
)
3352 if (cfq_slice_used(cfqq
))
3355 /* Allow preemption only if we are idling on sync-noidle tree */
3356 if (cfqd
->serving_type
== SYNC_NOIDLE_WORKLOAD
&&
3357 cfqq_type(new_cfqq
) == SYNC_NOIDLE_WORKLOAD
&&
3358 new_cfqq
->service_tree
->count
== 2 &&
3359 RB_EMPTY_ROOT(&cfqq
->sort_list
))
3363 * So both queues are sync. Let the new request get disk time if
3364 * it's a metadata request and the current queue is doing regular IO.
3366 if ((rq
->cmd_flags
& REQ_PRIO
) && !cfqq
->prio_pending
)
3370 * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
3372 if (cfq_class_rt(new_cfqq
) && !cfq_class_rt(cfqq
))
3375 /* An idle queue should not be idle now for some reason */
3376 if (RB_EMPTY_ROOT(&cfqq
->sort_list
) && !cfq_should_idle(cfqd
, cfqq
))
3379 if (!cfqd
->active_cic
|| !cfq_cfqq_wait_request(cfqq
))
3383 * if this request is as-good as one we would expect from the
3384 * current cfqq, let it preempt
3386 if (cfq_rq_close(cfqd
, cfqq
, rq
))
3393 * cfqq preempts the active queue. if we allowed preempt with no slice left,
3394 * let it have half of its nominal slice.
3396 static void cfq_preempt_queue(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
3398 struct cfq_queue
*old_cfqq
= cfqd
->active_queue
;
3400 cfq_log_cfqq(cfqd
, cfqq
, "preempt");
3401 cfq_slice_expired(cfqd
, 1);
3404 * workload type is changed, don't save slice, otherwise preempt
3407 if (cfqq_type(old_cfqq
) != cfqq_type(cfqq
))
3408 cfqq
->cfqg
->saved_workload_slice
= 0;
3411 * Put the new queue at the front of the of the current list,
3412 * so we know that it will be selected next.
3414 BUG_ON(!cfq_cfqq_on_rr(cfqq
));
3416 cfq_service_tree_add(cfqd
, cfqq
, 1);
3418 cfqq
->slice_end
= 0;
3419 cfq_mark_cfqq_slice_new(cfqq
);
3423 * Called when a new fs request (rq) is added (to cfqq). Check if there's
3424 * something we should do about it
3427 cfq_rq_enqueued(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
3430 struct cfq_io_context
*cic
= RQ_CIC(rq
);
3433 if (rq
->cmd_flags
& REQ_PRIO
)
3434 cfqq
->prio_pending
++;
3436 cfq_update_io_thinktime(cfqd
, cfqq
, cic
);
3437 cfq_update_io_seektime(cfqd
, cfqq
, rq
);
3438 cfq_update_idle_window(cfqd
, cfqq
, cic
);
3440 cfqq
->last_request_pos
= blk_rq_pos(rq
) + blk_rq_sectors(rq
);
3442 if (cfqq
== cfqd
->active_queue
) {
3444 * Remember that we saw a request from this process, but
3445 * don't start queuing just yet. Otherwise we risk seeing lots
3446 * of tiny requests, because we disrupt the normal plugging
3447 * and merging. If the request is already larger than a single
3448 * page, let it rip immediately. For that case we assume that
3449 * merging is already done. Ditto for a busy system that
3450 * has other work pending, don't risk delaying until the
3451 * idle timer unplug to continue working.
3453 if (cfq_cfqq_wait_request(cfqq
)) {
3454 if (blk_rq_bytes(rq
) > PAGE_CACHE_SIZE
||
3455 cfqd
->busy_queues
> 1) {
3456 cfq_del_timer(cfqd
, cfqq
);
3457 cfq_clear_cfqq_wait_request(cfqq
);
3458 __blk_run_queue(cfqd
->queue
);
3460 cfq_blkiocg_update_idle_time_stats(
3462 cfq_mark_cfqq_must_dispatch(cfqq
);
3465 } else if (cfq_should_preempt(cfqd
, cfqq
, rq
)) {
3467 * not the active queue - expire current slice if it is
3468 * idle and has expired it's mean thinktime or this new queue
3469 * has some old slice time left and is of higher priority or
3470 * this new queue is RT and the current one is BE
3472 cfq_preempt_queue(cfqd
, cfqq
);
3473 __blk_run_queue(cfqd
->queue
);
3477 static void cfq_insert_request(struct request_queue
*q
, struct request
*rq
)
3479 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
3480 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
3482 cfq_log_cfqq(cfqd
, cfqq
, "insert_request");
3483 cfq_init_prio_data(cfqq
, RQ_CIC(rq
)->ioc
);
3485 rq_set_fifo_time(rq
, jiffies
+ cfqd
->cfq_fifo_expire
[rq_is_sync(rq
)]);
3486 list_add_tail(&rq
->queuelist
, &cfqq
->fifo
);
3488 cfq_blkiocg_update_io_add_stats(&(RQ_CFQG(rq
))->blkg
,
3489 &cfqd
->serving_group
->blkg
, rq_data_dir(rq
),
3491 cfq_rq_enqueued(cfqd
, cfqq
, rq
);
3495 * Update hw_tag based on peak queue depth over 50 samples under
3498 static void cfq_update_hw_tag(struct cfq_data
*cfqd
)
3500 struct cfq_queue
*cfqq
= cfqd
->active_queue
;
3502 if (cfqd
->rq_in_driver
> cfqd
->hw_tag_est_depth
)
3503 cfqd
->hw_tag_est_depth
= cfqd
->rq_in_driver
;
3505 if (cfqd
->hw_tag
== 1)
3508 if (cfqd
->rq_queued
<= CFQ_HW_QUEUE_MIN
&&
3509 cfqd
->rq_in_driver
<= CFQ_HW_QUEUE_MIN
)
3513 * If active queue hasn't enough requests and can idle, cfq might not
3514 * dispatch sufficient requests to hardware. Don't zero hw_tag in this
3517 if (cfqq
&& cfq_cfqq_idle_window(cfqq
) &&
3518 cfqq
->dispatched
+ cfqq
->queued
[0] + cfqq
->queued
[1] <
3519 CFQ_HW_QUEUE_MIN
&& cfqd
->rq_in_driver
< CFQ_HW_QUEUE_MIN
)
3522 if (cfqd
->hw_tag_samples
++ < 50)
3525 if (cfqd
->hw_tag_est_depth
>= CFQ_HW_QUEUE_MIN
)
3531 static bool cfq_should_wait_busy(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
3533 struct cfq_io_context
*cic
= cfqd
->active_cic
;
3535 /* If the queue already has requests, don't wait */
3536 if (!RB_EMPTY_ROOT(&cfqq
->sort_list
))
3539 /* If there are other queues in the group, don't wait */
3540 if (cfqq
->cfqg
->nr_cfqq
> 1)
3543 /* the only queue in the group, but think time is big */
3544 if (cfq_io_thinktime_big(cfqd
, &cfqq
->cfqg
->ttime
, true))
3547 if (cfq_slice_used(cfqq
))
3550 /* if slice left is less than think time, wait busy */
3551 if (cic
&& sample_valid(cic
->ttime
.ttime_samples
)
3552 && (cfqq
->slice_end
- jiffies
< cic
->ttime
.ttime_mean
))
3556 * If think times is less than a jiffy than ttime_mean=0 and above
3557 * will not be true. It might happen that slice has not expired yet
3558 * but will expire soon (4-5 ns) during select_queue(). To cover the
3559 * case where think time is less than a jiffy, mark the queue wait
3560 * busy if only 1 jiffy is left in the slice.
3562 if (cfqq
->slice_end
- jiffies
== 1)
3568 static void cfq_completed_request(struct request_queue
*q
, struct request
*rq
)
3570 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
3571 struct cfq_data
*cfqd
= cfqq
->cfqd
;
3572 const int sync
= rq_is_sync(rq
);
3576 cfq_log_cfqq(cfqd
, cfqq
, "complete rqnoidle %d",
3577 !!(rq
->cmd_flags
& REQ_NOIDLE
));
3579 cfq_update_hw_tag(cfqd
);
3581 WARN_ON(!cfqd
->rq_in_driver
);
3582 WARN_ON(!cfqq
->dispatched
);
3583 cfqd
->rq_in_driver
--;
3585 (RQ_CFQG(rq
))->dispatched
--;
3586 cfq_blkiocg_update_completion_stats(&cfqq
->cfqg
->blkg
,
3587 rq_start_time_ns(rq
), rq_io_start_time_ns(rq
),
3588 rq_data_dir(rq
), rq_is_sync(rq
));
3590 cfqd
->rq_in_flight
[cfq_cfqq_sync(cfqq
)]--;
3593 struct cfq_rb_root
*service_tree
;
3595 RQ_CIC(rq
)->ttime
.last_end_request
= now
;
3597 if (cfq_cfqq_on_rr(cfqq
))
3598 service_tree
= cfqq
->service_tree
;
3600 service_tree
= service_tree_for(cfqq
->cfqg
,
3601 cfqq_prio(cfqq
), cfqq_type(cfqq
));
3602 service_tree
->ttime
.last_end_request
= now
;
3603 if (!time_after(rq
->start_time
+ cfqd
->cfq_fifo_expire
[1], now
))
3604 cfqd
->last_delayed_sync
= now
;
3607 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3608 cfqq
->cfqg
->ttime
.last_end_request
= now
;
3612 * If this is the active queue, check if it needs to be expired,
3613 * or if we want to idle in case it has no pending requests.
3615 if (cfqd
->active_queue
== cfqq
) {
3616 const bool cfqq_empty
= RB_EMPTY_ROOT(&cfqq
->sort_list
);
3618 if (cfq_cfqq_slice_new(cfqq
)) {
3619 cfq_set_prio_slice(cfqd
, cfqq
);
3620 cfq_clear_cfqq_slice_new(cfqq
);
3624 * Should we wait for next request to come in before we expire
3627 if (cfq_should_wait_busy(cfqd
, cfqq
)) {
3628 unsigned long extend_sl
= cfqd
->cfq_slice_idle
;
3629 if (!cfqd
->cfq_slice_idle
)
3630 extend_sl
= cfqd
->cfq_group_idle
;
3631 cfqq
->slice_end
= jiffies
+ extend_sl
;
3632 cfq_mark_cfqq_wait_busy(cfqq
);
3633 cfq_log_cfqq(cfqd
, cfqq
, "will busy wait");
3637 * Idling is not enabled on:
3639 * - idle-priority queues
3641 * - queues with still some requests queued
3642 * - when there is a close cooperator
3644 if (cfq_slice_used(cfqq
) || cfq_class_idle(cfqq
))
3645 cfq_slice_expired(cfqd
, 1);
3646 else if (sync
&& cfqq_empty
&&
3647 !cfq_close_cooperator(cfqd
, cfqq
)) {
3648 cfq_arm_slice_timer(cfqd
);
3652 if (!cfqd
->rq_in_driver
)
3653 cfq_schedule_dispatch(cfqd
);
3656 static inline int __cfq_may_queue(struct cfq_queue
*cfqq
)
3658 if (cfq_cfqq_wait_request(cfqq
) && !cfq_cfqq_must_alloc_slice(cfqq
)) {
3659 cfq_mark_cfqq_must_alloc_slice(cfqq
);
3660 return ELV_MQUEUE_MUST
;
3663 return ELV_MQUEUE_MAY
;
3666 static int cfq_may_queue(struct request_queue
*q
, int rw
)
3668 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
3669 struct task_struct
*tsk
= current
;
3670 struct cfq_io_context
*cic
;
3671 struct cfq_queue
*cfqq
;
3674 * don't force setup of a queue from here, as a call to may_queue
3675 * does not necessarily imply that a request actually will be queued.
3676 * so just lookup a possibly existing queue, or return 'may queue'
3679 cic
= cfq_cic_lookup(cfqd
, tsk
->io_context
);
3681 return ELV_MQUEUE_MAY
;
3683 cfqq
= cic_to_cfqq(cic
, rw_is_sync(rw
));
3685 cfq_init_prio_data(cfqq
, cic
->ioc
);
3687 return __cfq_may_queue(cfqq
);
3690 return ELV_MQUEUE_MAY
;
3694 * queue lock held here
3696 static void cfq_put_request(struct request
*rq
)
3698 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
3701 const int rw
= rq_data_dir(rq
);
3703 BUG_ON(!cfqq
->allocated
[rw
]);
3704 cfqq
->allocated
[rw
]--;
3706 put_io_context(RQ_CIC(rq
)->ioc
);
3708 rq
->elevator_private
[0] = NULL
;
3709 rq
->elevator_private
[1] = NULL
;
3711 /* Put down rq reference on cfqg */
3712 cfq_put_cfqg(RQ_CFQG(rq
));
3713 rq
->elevator_private
[2] = NULL
;
3715 cfq_put_queue(cfqq
);
3719 static struct cfq_queue
*
3720 cfq_merge_cfqqs(struct cfq_data
*cfqd
, struct cfq_io_context
*cic
,
3721 struct cfq_queue
*cfqq
)
3723 cfq_log_cfqq(cfqd
, cfqq
, "merging with queue %p", cfqq
->new_cfqq
);
3724 cic_set_cfqq(cic
, cfqq
->new_cfqq
, 1);
3725 cfq_mark_cfqq_coop(cfqq
->new_cfqq
);
3726 cfq_put_queue(cfqq
);
3727 return cic_to_cfqq(cic
, 1);
3731 * Returns NULL if a new cfqq should be allocated, or the old cfqq if this
3732 * was the last process referring to said cfqq.
3734 static struct cfq_queue
*
3735 split_cfqq(struct cfq_io_context
*cic
, struct cfq_queue
*cfqq
)
3737 if (cfqq_process_refs(cfqq
) == 1) {
3738 cfqq
->pid
= current
->pid
;
3739 cfq_clear_cfqq_coop(cfqq
);
3740 cfq_clear_cfqq_split_coop(cfqq
);
3744 cic_set_cfqq(cic
, NULL
, 1);
3746 cfq_put_cooperator(cfqq
);
3748 cfq_put_queue(cfqq
);
3752 * Allocate cfq data structures associated with this request.
3755 cfq_set_request(struct request_queue
*q
, struct request
*rq
, gfp_t gfp_mask
)
3757 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
3758 struct cfq_io_context
*cic
;
3759 const int rw
= rq_data_dir(rq
);
3760 const bool is_sync
= rq_is_sync(rq
);
3761 struct cfq_queue
*cfqq
;
3762 unsigned long flags
;
3764 might_sleep_if(gfp_mask
& __GFP_WAIT
);
3766 cic
= cfq_get_io_context(cfqd
, gfp_mask
);
3768 spin_lock_irqsave(q
->queue_lock
, flags
);
3774 cfqq
= cic_to_cfqq(cic
, is_sync
);
3775 if (!cfqq
|| cfqq
== &cfqd
->oom_cfqq
) {
3776 cfqq
= cfq_get_queue(cfqd
, is_sync
, cic
->ioc
, gfp_mask
);
3777 cic_set_cfqq(cic
, cfqq
, is_sync
);
3780 * If the queue was seeky for too long, break it apart.
3782 if (cfq_cfqq_coop(cfqq
) && cfq_cfqq_split_coop(cfqq
)) {
3783 cfq_log_cfqq(cfqd
, cfqq
, "breaking apart cfqq");
3784 cfqq
= split_cfqq(cic
, cfqq
);
3790 * Check to see if this queue is scheduled to merge with
3791 * another, closely cooperating queue. The merging of
3792 * queues happens here as it must be done in process context.
3793 * The reference on new_cfqq was taken in merge_cfqqs.
3796 cfqq
= cfq_merge_cfqqs(cfqd
, cic
, cfqq
);
3799 cfqq
->allocated
[rw
]++;
3802 rq
->elevator_private
[0] = cic
;
3803 rq
->elevator_private
[1] = cfqq
;
3804 rq
->elevator_private
[2] = cfq_ref_get_cfqg(cfqq
->cfqg
);
3805 spin_unlock_irqrestore(q
->queue_lock
, flags
);
3809 cfq_schedule_dispatch(cfqd
);
3810 spin_unlock_irqrestore(q
->queue_lock
, flags
);
3811 cfq_log(cfqd
, "set_request fail");
3815 static void cfq_kick_queue(struct work_struct
*work
)
3817 struct cfq_data
*cfqd
=
3818 container_of(work
, struct cfq_data
, unplug_work
);
3819 struct request_queue
*q
= cfqd
->queue
;
3821 spin_lock_irq(q
->queue_lock
);
3822 __blk_run_queue(cfqd
->queue
);
3823 spin_unlock_irq(q
->queue_lock
);
3827 * Timer running if the active_queue is currently idling inside its time slice
3829 static void cfq_idle_slice_timer(unsigned long data
)
3831 struct cfq_data
*cfqd
= (struct cfq_data
*) data
;
3832 struct cfq_queue
*cfqq
;
3833 unsigned long flags
;
3836 cfq_log(cfqd
, "idle timer fired");
3838 spin_lock_irqsave(cfqd
->queue
->queue_lock
, flags
);
3840 cfqq
= cfqd
->active_queue
;
3845 * We saw a request before the queue expired, let it through
3847 if (cfq_cfqq_must_dispatch(cfqq
))
3853 if (cfq_slice_used(cfqq
))
3857 * only expire and reinvoke request handler, if there are
3858 * other queues with pending requests
3860 if (!cfqd
->busy_queues
)
3864 * not expired and it has a request pending, let it dispatch
3866 if (!RB_EMPTY_ROOT(&cfqq
->sort_list
))
3870 * Queue depth flag is reset only when the idle didn't succeed
3872 cfq_clear_cfqq_deep(cfqq
);
3875 cfq_slice_expired(cfqd
, timed_out
);
3877 cfq_schedule_dispatch(cfqd
);
3879 spin_unlock_irqrestore(cfqd
->queue
->queue_lock
, flags
);
3882 static void cfq_shutdown_timer_wq(struct cfq_data
*cfqd
)
3884 del_timer_sync(&cfqd
->idle_slice_timer
);
3885 cancel_work_sync(&cfqd
->unplug_work
);
3888 static void cfq_put_async_queues(struct cfq_data
*cfqd
)
3892 for (i
= 0; i
< IOPRIO_BE_NR
; i
++) {
3893 if (cfqd
->async_cfqq
[0][i
])
3894 cfq_put_queue(cfqd
->async_cfqq
[0][i
]);
3895 if (cfqd
->async_cfqq
[1][i
])
3896 cfq_put_queue(cfqd
->async_cfqq
[1][i
]);
3899 if (cfqd
->async_idle_cfqq
)
3900 cfq_put_queue(cfqd
->async_idle_cfqq
);
3903 static void cfq_exit_queue(struct elevator_queue
*e
)
3905 struct cfq_data
*cfqd
= e
->elevator_data
;
3906 struct request_queue
*q
= cfqd
->queue
;
3909 cfq_shutdown_timer_wq(cfqd
);
3911 spin_lock_irq(q
->queue_lock
);
3913 if (cfqd
->active_queue
)
3914 __cfq_slice_expired(cfqd
, cfqd
->active_queue
, 0);
3916 while (!list_empty(&cfqd
->cic_list
)) {
3917 struct cfq_io_context
*cic
= list_entry(cfqd
->cic_list
.next
,
3918 struct cfq_io_context
,
3924 cfq_put_async_queues(cfqd
);
3925 cfq_release_cfq_groups(cfqd
);
3928 * If there are groups which we could not unlink from blkcg list,
3929 * wait for a rcu period for them to be freed.
3931 if (cfqd
->nr_blkcg_linked_grps
)
3934 spin_unlock_irq(q
->queue_lock
);
3936 cfq_shutdown_timer_wq(cfqd
);
3939 * Wait for cfqg->blkg->key accessors to exit their grace periods.
3940 * Do this wait only if there are other unlinked groups out
3941 * there. This can happen if cgroup deletion path claimed the
3942 * responsibility of cleaning up a group before queue cleanup code
3945 * Do not call synchronize_rcu() unconditionally as there are drivers
3946 * which create/delete request queue hundreds of times during scan/boot
3947 * and synchronize_rcu() can take significant time and slow down boot.
3952 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3953 /* Free up per cpu stats for root group */
3954 free_percpu(cfqd
->root_group
.blkg
.stats_cpu
);
3959 static void *cfq_init_queue(struct request_queue
*q
)
3961 struct cfq_data
*cfqd
;
3963 struct cfq_group
*cfqg
;
3964 struct cfq_rb_root
*st
;
3966 cfqd
= kmalloc_node(sizeof(*cfqd
), GFP_KERNEL
| __GFP_ZERO
, q
->node
);
3970 /* Init root service tree */
3971 cfqd
->grp_service_tree
= CFQ_RB_ROOT
;
3973 /* Init root group */
3974 cfqg
= &cfqd
->root_group
;
3975 for_each_cfqg_st(cfqg
, i
, j
, st
)
3977 RB_CLEAR_NODE(&cfqg
->rb_node
);
3979 /* Give preference to root group over other groups */
3980 cfqg
->weight
= 2*BLKIO_WEIGHT_DEFAULT
;
3982 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3984 * Set root group reference to 2. One reference will be dropped when
3985 * all groups on cfqd->cfqg_list are being deleted during queue exit.
3986 * Other reference will remain there as we don't want to delete this
3987 * group as it is statically allocated and gets destroyed when
3988 * throtl_data goes away.
3992 if (blkio_alloc_blkg_stats(&cfqg
->blkg
)) {
4000 cfq_blkiocg_add_blkio_group(&blkio_root_cgroup
, &cfqg
->blkg
,
4003 cfqd
->nr_blkcg_linked_grps
++;
4005 /* Add group on cfqd->cfqg_list */
4006 hlist_add_head(&cfqg
->cfqd_node
, &cfqd
->cfqg_list
);
4009 * Not strictly needed (since RB_ROOT just clears the node and we
4010 * zeroed cfqd on alloc), but better be safe in case someone decides
4011 * to add magic to the rb code
4013 for (i
= 0; i
< CFQ_PRIO_LISTS
; i
++)
4014 cfqd
->prio_trees
[i
] = RB_ROOT
;
4017 * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
4018 * Grab a permanent reference to it, so that the normal code flow
4019 * will not attempt to free it.
4021 cfq_init_cfqq(cfqd
, &cfqd
->oom_cfqq
, 1, 0);
4022 cfqd
->oom_cfqq
.ref
++;
4023 cfq_link_cfqq_cfqg(&cfqd
->oom_cfqq
, &cfqd
->root_group
);
4025 INIT_LIST_HEAD(&cfqd
->cic_list
);
4029 init_timer(&cfqd
->idle_slice_timer
);
4030 cfqd
->idle_slice_timer
.function
= cfq_idle_slice_timer
;
4031 cfqd
->idle_slice_timer
.data
= (unsigned long) cfqd
;
4033 INIT_WORK(&cfqd
->unplug_work
, cfq_kick_queue
);
4035 cfqd
->cfq_quantum
= cfq_quantum
;
4036 cfqd
->cfq_fifo_expire
[0] = cfq_fifo_expire
[0];
4037 cfqd
->cfq_fifo_expire
[1] = cfq_fifo_expire
[1];
4038 cfqd
->cfq_back_max
= cfq_back_max
;
4039 cfqd
->cfq_back_penalty
= cfq_back_penalty
;
4040 cfqd
->cfq_slice
[0] = cfq_slice_async
;
4041 cfqd
->cfq_slice
[1] = cfq_slice_sync
;
4042 cfqd
->cfq_slice_async_rq
= cfq_slice_async_rq
;
4043 cfqd
->cfq_slice_idle
= cfq_slice_idle
;
4044 cfqd
->cfq_group_idle
= cfq_group_idle
;
4045 cfqd
->cfq_latency
= 1;
4048 * we optimistically start assuming sync ops weren't delayed in last
4049 * second, in order to have larger depth for async operations.
4051 cfqd
->last_delayed_sync
= jiffies
- HZ
;
4055 static void cfq_slab_kill(void)
4058 * Caller already ensured that pending RCU callbacks are completed,
4059 * so we should have no busy allocations at this point.
4062 kmem_cache_destroy(cfq_pool
);
4064 kmem_cache_destroy(cfq_ioc_pool
);
4067 static int __init
cfq_slab_setup(void)
4069 cfq_pool
= KMEM_CACHE(cfq_queue
, 0);
4073 cfq_ioc_pool
= KMEM_CACHE(cfq_io_context
, 0);
4084 * sysfs parts below -->
4087 cfq_var_show(unsigned int var
, char *page
)
4089 return sprintf(page
, "%d\n", var
);
4093 cfq_var_store(unsigned int *var
, const char *page
, size_t count
)
4095 char *p
= (char *) page
;
4097 *var
= simple_strtoul(p
, &p
, 10);
4101 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
4102 static ssize_t __FUNC(struct elevator_queue *e, char *page) \
4104 struct cfq_data *cfqd = e->elevator_data; \
4105 unsigned int __data = __VAR; \
4107 __data = jiffies_to_msecs(__data); \
4108 return cfq_var_show(__data, (page)); \
4110 SHOW_FUNCTION(cfq_quantum_show
, cfqd
->cfq_quantum
, 0);
4111 SHOW_FUNCTION(cfq_fifo_expire_sync_show
, cfqd
->cfq_fifo_expire
[1], 1);
4112 SHOW_FUNCTION(cfq_fifo_expire_async_show
, cfqd
->cfq_fifo_expire
[0], 1);
4113 SHOW_FUNCTION(cfq_back_seek_max_show
, cfqd
->cfq_back_max
, 0);
4114 SHOW_FUNCTION(cfq_back_seek_penalty_show
, cfqd
->cfq_back_penalty
, 0);
4115 SHOW_FUNCTION(cfq_slice_idle_show
, cfqd
->cfq_slice_idle
, 1);
4116 SHOW_FUNCTION(cfq_group_idle_show
, cfqd
->cfq_group_idle
, 1);
4117 SHOW_FUNCTION(cfq_slice_sync_show
, cfqd
->cfq_slice
[1], 1);
4118 SHOW_FUNCTION(cfq_slice_async_show
, cfqd
->cfq_slice
[0], 1);
4119 SHOW_FUNCTION(cfq_slice_async_rq_show
, cfqd
->cfq_slice_async_rq
, 0);
4120 SHOW_FUNCTION(cfq_low_latency_show
, cfqd
->cfq_latency
, 0);
4121 #undef SHOW_FUNCTION
4123 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
4124 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
4126 struct cfq_data *cfqd = e->elevator_data; \
4127 unsigned int __data; \
4128 int ret = cfq_var_store(&__data, (page), count); \
4129 if (__data < (MIN)) \
4131 else if (__data > (MAX)) \
4134 *(__PTR) = msecs_to_jiffies(__data); \
4136 *(__PTR) = __data; \
4139 STORE_FUNCTION(cfq_quantum_store
, &cfqd
->cfq_quantum
, 1, UINT_MAX
, 0);
4140 STORE_FUNCTION(cfq_fifo_expire_sync_store
, &cfqd
->cfq_fifo_expire
[1], 1,
4142 STORE_FUNCTION(cfq_fifo_expire_async_store
, &cfqd
->cfq_fifo_expire
[0], 1,
4144 STORE_FUNCTION(cfq_back_seek_max_store
, &cfqd
->cfq_back_max
, 0, UINT_MAX
, 0);
4145 STORE_FUNCTION(cfq_back_seek_penalty_store
, &cfqd
->cfq_back_penalty
, 1,
4147 STORE_FUNCTION(cfq_slice_idle_store
, &cfqd
->cfq_slice_idle
, 0, UINT_MAX
, 1);
4148 STORE_FUNCTION(cfq_group_idle_store
, &cfqd
->cfq_group_idle
, 0, UINT_MAX
, 1);
4149 STORE_FUNCTION(cfq_slice_sync_store
, &cfqd
->cfq_slice
[1], 1, UINT_MAX
, 1);
4150 STORE_FUNCTION(cfq_slice_async_store
, &cfqd
->cfq_slice
[0], 1, UINT_MAX
, 1);
4151 STORE_FUNCTION(cfq_slice_async_rq_store
, &cfqd
->cfq_slice_async_rq
, 1,
4153 STORE_FUNCTION(cfq_low_latency_store
, &cfqd
->cfq_latency
, 0, 1, 0);
4154 #undef STORE_FUNCTION
4156 #define CFQ_ATTR(name) \
4157 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
4159 static struct elv_fs_entry cfq_attrs
[] = {
4161 CFQ_ATTR(fifo_expire_sync
),
4162 CFQ_ATTR(fifo_expire_async
),
4163 CFQ_ATTR(back_seek_max
),
4164 CFQ_ATTR(back_seek_penalty
),
4165 CFQ_ATTR(slice_sync
),
4166 CFQ_ATTR(slice_async
),
4167 CFQ_ATTR(slice_async_rq
),
4168 CFQ_ATTR(slice_idle
),
4169 CFQ_ATTR(group_idle
),
4170 CFQ_ATTR(low_latency
),
4174 static struct elevator_type iosched_cfq
= {
4176 .elevator_merge_fn
= cfq_merge
,
4177 .elevator_merged_fn
= cfq_merged_request
,
4178 .elevator_merge_req_fn
= cfq_merged_requests
,
4179 .elevator_allow_merge_fn
= cfq_allow_merge
,
4180 .elevator_bio_merged_fn
= cfq_bio_merged
,
4181 .elevator_dispatch_fn
= cfq_dispatch_requests
,
4182 .elevator_add_req_fn
= cfq_insert_request
,
4183 .elevator_activate_req_fn
= cfq_activate_request
,
4184 .elevator_deactivate_req_fn
= cfq_deactivate_request
,
4185 .elevator_completed_req_fn
= cfq_completed_request
,
4186 .elevator_former_req_fn
= elv_rb_former_request
,
4187 .elevator_latter_req_fn
= elv_rb_latter_request
,
4188 .elevator_set_req_fn
= cfq_set_request
,
4189 .elevator_put_req_fn
= cfq_put_request
,
4190 .elevator_may_queue_fn
= cfq_may_queue
,
4191 .elevator_init_fn
= cfq_init_queue
,
4192 .elevator_exit_fn
= cfq_exit_queue
,
4193 .trim
= cfq_free_io_context
,
4195 .elevator_attrs
= cfq_attrs
,
4196 .elevator_name
= "cfq",
4197 .elevator_owner
= THIS_MODULE
,
4200 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4201 static struct blkio_policy_type blkio_policy_cfq
= {
4203 .blkio_unlink_group_fn
= cfq_unlink_blkio_group
,
4204 .blkio_update_group_weight_fn
= cfq_update_blkio_group_weight
,
4206 .plid
= BLKIO_POLICY_PROP
,
4209 static struct blkio_policy_type blkio_policy_cfq
;
4212 static int __init
cfq_init(void)
4215 * could be 0 on HZ < 1000 setups
4217 if (!cfq_slice_async
)
4218 cfq_slice_async
= 1;
4219 if (!cfq_slice_idle
)
4222 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4223 if (!cfq_group_idle
)
4228 if (cfq_slab_setup())
4231 elv_register(&iosched_cfq
);
4232 blkio_policy_register(&blkio_policy_cfq
);
4237 static void __exit
cfq_exit(void)
4239 DECLARE_COMPLETION_ONSTACK(all_gone
);
4240 blkio_policy_unregister(&blkio_policy_cfq
);
4241 elv_unregister(&iosched_cfq
);
4242 ioc_gone
= &all_gone
;
4243 /* ioc_gone's update must be visible before reading ioc_count */
4247 * this also protects us from entering cfq_slab_kill() with
4248 * pending RCU callbacks
4250 if (elv_ioc_count_read(cfq_ioc_count
))
4251 wait_for_completion(&all_gone
);
4255 module_init(cfq_init
);
4256 module_exit(cfq_exit
);
4258 MODULE_AUTHOR("Jens Axboe");
4259 MODULE_LICENSE("GPL");
4260 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");