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>
20 static struct blkio_policy_type blkio_policy_cfq
;
25 /* max queue in one round of service */
26 static const int cfq_quantum
= 8;
27 static const int cfq_fifo_expire
[2] = { HZ
/ 4, HZ
/ 8 };
28 /* maximum backwards seek, in KiB */
29 static const int cfq_back_max
= 16 * 1024;
30 /* penalty of a backwards seek */
31 static const int cfq_back_penalty
= 2;
32 static const int cfq_slice_sync
= HZ
/ 10;
33 static int cfq_slice_async
= HZ
/ 25;
34 static const int cfq_slice_async_rq
= 2;
35 static int cfq_slice_idle
= HZ
/ 125;
36 static int cfq_group_idle
= HZ
/ 125;
37 static const int cfq_target_latency
= HZ
* 3/10; /* 300 ms */
38 static const int cfq_hist_divisor
= 4;
41 * offset from end of service tree
43 #define CFQ_IDLE_DELAY (HZ / 5)
46 * below this threshold, we consider thinktime immediate
48 #define CFQ_MIN_TT (2)
50 #define CFQ_SLICE_SCALE (5)
51 #define CFQ_HW_QUEUE_MIN (5)
52 #define CFQ_SERVICE_SHIFT 12
54 #define CFQQ_SEEK_THR (sector_t)(8 * 100)
55 #define CFQQ_CLOSE_THR (sector_t)(8 * 1024)
56 #define CFQQ_SECT_THR_NONROT (sector_t)(2 * 32)
57 #define CFQQ_SEEKY(cfqq) (hweight32(cfqq->seek_history) > 32/8)
59 #define RQ_CIC(rq) icq_to_cic((rq)->elv.icq)
60 #define RQ_CFQQ(rq) (struct cfq_queue *) ((rq)->elv.priv[0])
61 #define RQ_CFQG(rq) (struct cfq_group *) ((rq)->elv.priv[1])
63 static struct kmem_cache
*cfq_pool
;
65 #define CFQ_PRIO_LISTS IOPRIO_BE_NR
66 #define cfq_class_idle(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
67 #define cfq_class_rt(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
69 #define sample_valid(samples) ((samples) > 80)
70 #define rb_entry_cfqg(node) rb_entry((node), struct cfq_group, rb_node)
73 unsigned long last_end_request
;
75 unsigned long ttime_total
;
76 unsigned long ttime_samples
;
77 unsigned long ttime_mean
;
81 * Most of our rbtree usage is for sorting with min extraction, so
82 * if we cache the leftmost node we don't have to walk down the tree
83 * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should
84 * move this into the elevator for the rq sorting as well.
90 unsigned total_weight
;
92 struct cfq_ttime ttime
;
94 #define CFQ_RB_ROOT (struct cfq_rb_root) { .rb = RB_ROOT, \
95 .ttime = {.last_end_request = jiffies,},}
98 * Per process-grouping structure
101 /* reference count */
103 /* various state flags, see below */
105 /* parent cfq_data */
106 struct cfq_data
*cfqd
;
107 /* service_tree member */
108 struct rb_node rb_node
;
109 /* service_tree key */
110 unsigned long rb_key
;
111 /* prio tree member */
112 struct rb_node p_node
;
113 /* prio tree root we belong to, if any */
114 struct rb_root
*p_root
;
115 /* sorted list of pending requests */
116 struct rb_root sort_list
;
117 /* if fifo isn't expired, next request to serve */
118 struct request
*next_rq
;
119 /* requests queued in sort_list */
121 /* currently allocated requests */
123 /* fifo list of requests in sort_list */
124 struct list_head fifo
;
126 /* time when queue got scheduled in to dispatch first request. */
127 unsigned long dispatch_start
;
128 unsigned int allocated_slice
;
129 unsigned int slice_dispatch
;
130 /* time when first request from queue completed and slice started. */
131 unsigned long slice_start
;
132 unsigned long slice_end
;
135 /* pending priority requests */
137 /* number of requests that are on the dispatch list or inside driver */
140 /* io prio of this group */
141 unsigned short ioprio
, org_ioprio
;
142 unsigned short ioprio_class
;
147 sector_t last_request_pos
;
149 struct cfq_rb_root
*service_tree
;
150 struct cfq_queue
*new_cfqq
;
151 struct cfq_group
*cfqg
;
152 /* Number of sectors dispatched from queue in single dispatch round */
153 unsigned long nr_sectors
;
157 * First index in the service_trees.
158 * IDLE is handled separately, so it has negative index
168 * Second index in the service_trees.
172 SYNC_NOIDLE_WORKLOAD
= 1,
176 /* This is per cgroup per device grouping structure */
178 /* group service_tree member */
179 struct rb_node rb_node
;
181 /* group service_tree key */
184 unsigned int new_weight
;
187 /* number of cfqq currently on this group */
191 * Per group busy queues average. Useful for workload slice calc. We
192 * create the array for each prio class but at run time it is used
193 * only for RT and BE class and slot for IDLE class remains unused.
194 * This is primarily done to avoid confusion and a gcc warning.
196 unsigned int busy_queues_avg
[CFQ_PRIO_NR
];
198 * rr lists of queues with requests. We maintain service trees for
199 * RT and BE classes. These trees are subdivided in subclasses
200 * of SYNC, SYNC_NOIDLE and ASYNC based on workload type. For IDLE
201 * class there is no subclassification and all the cfq queues go on
202 * a single tree service_tree_idle.
203 * Counts are embedded in the cfq_rb_root
205 struct cfq_rb_root service_trees
[2][3];
206 struct cfq_rb_root service_tree_idle
;
208 unsigned long saved_workload_slice
;
209 enum wl_type_t saved_workload
;
210 enum wl_prio_t saved_serving_prio
;
212 /* number of requests that are on the dispatch list or inside driver */
214 struct cfq_ttime ttime
;
218 struct io_cq icq
; /* must be the first member */
219 struct cfq_queue
*cfqq
[2];
220 struct cfq_ttime ttime
;
221 int ioprio
; /* the current ioprio */
222 #ifdef CONFIG_CFQ_GROUP_IOSCHED
223 uint64_t blkcg_id
; /* the current blkcg ID */
228 * Per block device queue structure
231 struct request_queue
*queue
;
232 /* Root service tree for cfq_groups */
233 struct cfq_rb_root grp_service_tree
;
234 struct cfq_group
*root_group
;
237 * The priority currently being served
239 enum wl_prio_t serving_prio
;
240 enum wl_type_t serving_type
;
241 unsigned long workload_expires
;
242 struct cfq_group
*serving_group
;
245 * Each priority tree is sorted by next_request position. These
246 * trees are used when determining if two or more queues are
247 * interleaving requests (see cfq_close_cooperator).
249 struct rb_root prio_trees
[CFQ_PRIO_LISTS
];
251 unsigned int busy_queues
;
252 unsigned int busy_sync_queues
;
258 * queue-depth detection
264 * -1 => indeterminate, (cfq will behave as if NCQ is present, to allow better detection)
265 * 1 => NCQ is present (hw_tag_est_depth is the estimated max depth)
268 int hw_tag_est_depth
;
269 unsigned int hw_tag_samples
;
272 * idle window management
274 struct timer_list idle_slice_timer
;
275 struct work_struct unplug_work
;
277 struct cfq_queue
*active_queue
;
278 struct cfq_io_cq
*active_cic
;
281 * async queue for each priority case
283 struct cfq_queue
*async_cfqq
[2][IOPRIO_BE_NR
];
284 struct cfq_queue
*async_idle_cfqq
;
286 sector_t last_position
;
289 * tunables, see top of file
291 unsigned int cfq_quantum
;
292 unsigned int cfq_fifo_expire
[2];
293 unsigned int cfq_back_penalty
;
294 unsigned int cfq_back_max
;
295 unsigned int cfq_slice
[2];
296 unsigned int cfq_slice_async_rq
;
297 unsigned int cfq_slice_idle
;
298 unsigned int cfq_group_idle
;
299 unsigned int cfq_latency
;
302 * Fallback dummy cfqq for extreme OOM conditions
304 struct cfq_queue oom_cfqq
;
306 unsigned long last_delayed_sync
;
309 static struct cfq_group
*cfq_get_next_cfqg(struct cfq_data
*cfqd
);
311 static struct cfq_rb_root
*service_tree_for(struct cfq_group
*cfqg
,
318 if (prio
== IDLE_WORKLOAD
)
319 return &cfqg
->service_tree_idle
;
321 return &cfqg
->service_trees
[prio
][type
];
324 enum cfqq_state_flags
{
325 CFQ_CFQQ_FLAG_on_rr
= 0, /* on round-robin busy list */
326 CFQ_CFQQ_FLAG_wait_request
, /* waiting for a request */
327 CFQ_CFQQ_FLAG_must_dispatch
, /* must be allowed a dispatch */
328 CFQ_CFQQ_FLAG_must_alloc_slice
, /* per-slice must_alloc flag */
329 CFQ_CFQQ_FLAG_fifo_expire
, /* FIFO checked in this slice */
330 CFQ_CFQQ_FLAG_idle_window
, /* slice idling enabled */
331 CFQ_CFQQ_FLAG_prio_changed
, /* task priority has changed */
332 CFQ_CFQQ_FLAG_slice_new
, /* no requests dispatched in slice */
333 CFQ_CFQQ_FLAG_sync
, /* synchronous queue */
334 CFQ_CFQQ_FLAG_coop
, /* cfqq is shared */
335 CFQ_CFQQ_FLAG_split_coop
, /* shared cfqq will be splitted */
336 CFQ_CFQQ_FLAG_deep
, /* sync cfqq experienced large depth */
337 CFQ_CFQQ_FLAG_wait_busy
, /* Waiting for next request */
340 #define CFQ_CFQQ_FNS(name) \
341 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \
343 (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name); \
345 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \
347 (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \
349 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \
351 return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \
355 CFQ_CFQQ_FNS(wait_request
);
356 CFQ_CFQQ_FNS(must_dispatch
);
357 CFQ_CFQQ_FNS(must_alloc_slice
);
358 CFQ_CFQQ_FNS(fifo_expire
);
359 CFQ_CFQQ_FNS(idle_window
);
360 CFQ_CFQQ_FNS(prio_changed
);
361 CFQ_CFQQ_FNS(slice_new
);
364 CFQ_CFQQ_FNS(split_coop
);
366 CFQ_CFQQ_FNS(wait_busy
);
369 #ifdef CONFIG_CFQ_GROUP_IOSCHED
370 static inline struct cfq_group
*blkg_to_cfqg(struct blkio_group
*blkg
)
372 return blkg_to_pdata(blkg
, &blkio_policy_cfq
);
375 static inline struct blkio_group
*cfqg_to_blkg(struct cfq_group
*cfqg
)
377 return pdata_to_blkg(cfqg
);
380 static inline void cfqg_get(struct cfq_group
*cfqg
)
382 return blkg_get(cfqg_to_blkg(cfqg
));
385 static inline void cfqg_put(struct cfq_group
*cfqg
)
387 return blkg_put(cfqg_to_blkg(cfqg
));
390 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
391 blk_add_trace_msg((cfqd)->queue, "cfq%d%c %s " fmt, (cfqq)->pid, \
392 cfq_cfqq_sync((cfqq)) ? 'S' : 'A', \
393 blkg_path(cfqg_to_blkg((cfqq)->cfqg)), ##args)
395 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...) \
396 blk_add_trace_msg((cfqd)->queue, "%s " fmt, \
397 blkg_path(cfqg_to_blkg((cfqg))), ##args) \
399 #else /* CONFIG_CFQ_GROUP_IOSCHED */
401 static inline struct cfq_group
*blkg_to_cfqg(struct blkio_group
*blkg
) { return NULL
; }
402 static inline struct blkio_group
*cfqg_to_blkg(struct cfq_group
*cfqg
) { return NULL
; }
403 static inline void cfqg_get(struct cfq_group
*cfqg
) { }
404 static inline void cfqg_put(struct cfq_group
*cfqg
) { }
406 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
407 blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args)
408 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...) do {} while (0)
410 #endif /* CONFIG_CFQ_GROUP_IOSCHED */
412 #define cfq_log(cfqd, fmt, args...) \
413 blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
415 /* Traverses through cfq group service trees */
416 #define for_each_cfqg_st(cfqg, i, j, st) \
417 for (i = 0; i <= IDLE_WORKLOAD; i++) \
418 for (j = 0, st = i < IDLE_WORKLOAD ? &cfqg->service_trees[i][j]\
419 : &cfqg->service_tree_idle; \
420 (i < IDLE_WORKLOAD && j <= SYNC_WORKLOAD) || \
421 (i == IDLE_WORKLOAD && j == 0); \
422 j++, st = i < IDLE_WORKLOAD ? \
423 &cfqg->service_trees[i][j]: NULL) \
425 static inline bool cfq_io_thinktime_big(struct cfq_data *cfqd,
426 struct cfq_ttime
*ttime
, bool group_idle
)
429 if (!sample_valid(ttime
->ttime_samples
))
432 slice
= cfqd
->cfq_group_idle
;
434 slice
= cfqd
->cfq_slice_idle
;
435 return ttime
->ttime_mean
> slice
;
438 static inline bool iops_mode(struct cfq_data
*cfqd
)
441 * If we are not idling on queues and it is a NCQ drive, parallel
442 * execution of requests is on and measuring time is not possible
443 * in most of the cases until and unless we drive shallower queue
444 * depths and that becomes a performance bottleneck. In such cases
445 * switch to start providing fairness in terms of number of IOs.
447 if (!cfqd
->cfq_slice_idle
&& cfqd
->hw_tag
)
453 static inline enum wl_prio_t
cfqq_prio(struct cfq_queue
*cfqq
)
455 if (cfq_class_idle(cfqq
))
456 return IDLE_WORKLOAD
;
457 if (cfq_class_rt(cfqq
))
463 static enum wl_type_t
cfqq_type(struct cfq_queue
*cfqq
)
465 if (!cfq_cfqq_sync(cfqq
))
466 return ASYNC_WORKLOAD
;
467 if (!cfq_cfqq_idle_window(cfqq
))
468 return SYNC_NOIDLE_WORKLOAD
;
469 return SYNC_WORKLOAD
;
472 static inline int cfq_group_busy_queues_wl(enum wl_prio_t wl
,
473 struct cfq_data
*cfqd
,
474 struct cfq_group
*cfqg
)
476 if (wl
== IDLE_WORKLOAD
)
477 return cfqg
->service_tree_idle
.count
;
479 return cfqg
->service_trees
[wl
][ASYNC_WORKLOAD
].count
480 + cfqg
->service_trees
[wl
][SYNC_NOIDLE_WORKLOAD
].count
481 + cfqg
->service_trees
[wl
][SYNC_WORKLOAD
].count
;
484 static inline int cfqg_busy_async_queues(struct cfq_data
*cfqd
,
485 struct cfq_group
*cfqg
)
487 return cfqg
->service_trees
[RT_WORKLOAD
][ASYNC_WORKLOAD
].count
488 + cfqg
->service_trees
[BE_WORKLOAD
][ASYNC_WORKLOAD
].count
;
491 static void cfq_dispatch_insert(struct request_queue
*, struct request
*);
492 static struct cfq_queue
*cfq_get_queue(struct cfq_data
*cfqd
, bool is_sync
,
493 struct cfq_io_cq
*cic
, struct bio
*bio
,
496 static inline struct cfq_io_cq
*icq_to_cic(struct io_cq
*icq
)
498 /* cic->icq is the first member, %NULL will convert to %NULL */
499 return container_of(icq
, struct cfq_io_cq
, icq
);
502 static inline struct cfq_io_cq
*cfq_cic_lookup(struct cfq_data
*cfqd
,
503 struct io_context
*ioc
)
506 return icq_to_cic(ioc_lookup_icq(ioc
, cfqd
->queue
));
510 static inline struct cfq_queue
*cic_to_cfqq(struct cfq_io_cq
*cic
, bool is_sync
)
512 return cic
->cfqq
[is_sync
];
515 static inline void cic_set_cfqq(struct cfq_io_cq
*cic
, struct cfq_queue
*cfqq
,
518 cic
->cfqq
[is_sync
] = cfqq
;
521 static inline struct cfq_data
*cic_to_cfqd(struct cfq_io_cq
*cic
)
523 return cic
->icq
.q
->elevator
->elevator_data
;
527 * We regard a request as SYNC, if it's either a read or has the SYNC bit
528 * set (in which case it could also be direct WRITE).
530 static inline bool cfq_bio_sync(struct bio
*bio
)
532 return bio_data_dir(bio
) == READ
|| (bio
->bi_rw
& REQ_SYNC
);
536 * scheduler run of queue, if there are requests pending and no one in the
537 * driver that will restart queueing
539 static inline void cfq_schedule_dispatch(struct cfq_data
*cfqd
)
541 if (cfqd
->busy_queues
) {
542 cfq_log(cfqd
, "schedule dispatch");
543 kblockd_schedule_work(cfqd
->queue
, &cfqd
->unplug_work
);
548 * Scale schedule slice based on io priority. Use the sync time slice only
549 * if a queue is marked sync and has sync io queued. A sync queue with async
550 * io only, should not get full sync slice length.
552 static inline int cfq_prio_slice(struct cfq_data
*cfqd
, bool sync
,
555 const int base_slice
= cfqd
->cfq_slice
[sync
];
557 WARN_ON(prio
>= IOPRIO_BE_NR
);
559 return base_slice
+ (base_slice
/CFQ_SLICE_SCALE
* (4 - prio
));
563 cfq_prio_to_slice(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
565 return cfq_prio_slice(cfqd
, cfq_cfqq_sync(cfqq
), cfqq
->ioprio
);
568 static inline u64
cfq_scale_slice(unsigned long delta
, struct cfq_group
*cfqg
)
570 u64 d
= delta
<< CFQ_SERVICE_SHIFT
;
572 d
= d
* BLKIO_WEIGHT_DEFAULT
;
573 do_div(d
, cfqg
->weight
);
577 static inline u64
max_vdisktime(u64 min_vdisktime
, u64 vdisktime
)
579 s64 delta
= (s64
)(vdisktime
- min_vdisktime
);
581 min_vdisktime
= vdisktime
;
583 return min_vdisktime
;
586 static inline u64
min_vdisktime(u64 min_vdisktime
, u64 vdisktime
)
588 s64 delta
= (s64
)(vdisktime
- min_vdisktime
);
590 min_vdisktime
= vdisktime
;
592 return min_vdisktime
;
595 static void update_min_vdisktime(struct cfq_rb_root
*st
)
597 struct cfq_group
*cfqg
;
600 cfqg
= rb_entry_cfqg(st
->left
);
601 st
->min_vdisktime
= max_vdisktime(st
->min_vdisktime
,
607 * get averaged number of queues of RT/BE priority.
608 * average is updated, with a formula that gives more weight to higher numbers,
609 * to quickly follows sudden increases and decrease slowly
612 static inline unsigned cfq_group_get_avg_queues(struct cfq_data
*cfqd
,
613 struct cfq_group
*cfqg
, bool rt
)
615 unsigned min_q
, max_q
;
616 unsigned mult
= cfq_hist_divisor
- 1;
617 unsigned round
= cfq_hist_divisor
/ 2;
618 unsigned busy
= cfq_group_busy_queues_wl(rt
, cfqd
, cfqg
);
620 min_q
= min(cfqg
->busy_queues_avg
[rt
], busy
);
621 max_q
= max(cfqg
->busy_queues_avg
[rt
], busy
);
622 cfqg
->busy_queues_avg
[rt
] = (mult
* max_q
+ min_q
+ round
) /
624 return cfqg
->busy_queues_avg
[rt
];
627 static inline unsigned
628 cfq_group_slice(struct cfq_data
*cfqd
, struct cfq_group
*cfqg
)
630 struct cfq_rb_root
*st
= &cfqd
->grp_service_tree
;
632 return cfq_target_latency
* cfqg
->weight
/ st
->total_weight
;
635 static inline unsigned
636 cfq_scaled_cfqq_slice(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
638 unsigned slice
= cfq_prio_to_slice(cfqd
, cfqq
);
639 if (cfqd
->cfq_latency
) {
641 * interested queues (we consider only the ones with the same
642 * priority class in the cfq group)
644 unsigned iq
= cfq_group_get_avg_queues(cfqd
, cfqq
->cfqg
,
646 unsigned sync_slice
= cfqd
->cfq_slice
[1];
647 unsigned expect_latency
= sync_slice
* iq
;
648 unsigned group_slice
= cfq_group_slice(cfqd
, cfqq
->cfqg
);
650 if (expect_latency
> group_slice
) {
651 unsigned base_low_slice
= 2 * cfqd
->cfq_slice_idle
;
652 /* scale low_slice according to IO priority
653 * and sync vs async */
655 min(slice
, base_low_slice
* slice
/ sync_slice
);
656 /* the adapted slice value is scaled to fit all iqs
657 * into the target latency */
658 slice
= max(slice
* group_slice
/ expect_latency
,
666 cfq_set_prio_slice(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
668 unsigned slice
= cfq_scaled_cfqq_slice(cfqd
, cfqq
);
670 cfqq
->slice_start
= jiffies
;
671 cfqq
->slice_end
= jiffies
+ slice
;
672 cfqq
->allocated_slice
= slice
;
673 cfq_log_cfqq(cfqd
, cfqq
, "set_slice=%lu", cfqq
->slice_end
- jiffies
);
677 * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
678 * isn't valid until the first request from the dispatch is activated
679 * and the slice time set.
681 static inline bool cfq_slice_used(struct cfq_queue
*cfqq
)
683 if (cfq_cfqq_slice_new(cfqq
))
685 if (time_before(jiffies
, cfqq
->slice_end
))
692 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
693 * We choose the request that is closest to the head right now. Distance
694 * behind the head is penalized and only allowed to a certain extent.
696 static struct request
*
697 cfq_choose_req(struct cfq_data
*cfqd
, struct request
*rq1
, struct request
*rq2
, sector_t last
)
699 sector_t s1
, s2
, d1
= 0, d2
= 0;
700 unsigned long back_max
;
701 #define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */
702 #define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */
703 unsigned wrap
= 0; /* bit mask: requests behind the disk head? */
705 if (rq1
== NULL
|| rq1
== rq2
)
710 if (rq_is_sync(rq1
) != rq_is_sync(rq2
))
711 return rq_is_sync(rq1
) ? rq1
: rq2
;
713 if ((rq1
->cmd_flags
^ rq2
->cmd_flags
) & REQ_PRIO
)
714 return rq1
->cmd_flags
& REQ_PRIO
? rq1
: rq2
;
716 s1
= blk_rq_pos(rq1
);
717 s2
= blk_rq_pos(rq2
);
720 * by definition, 1KiB is 2 sectors
722 back_max
= cfqd
->cfq_back_max
* 2;
725 * Strict one way elevator _except_ in the case where we allow
726 * short backward seeks which are biased as twice the cost of a
727 * similar forward seek.
731 else if (s1
+ back_max
>= last
)
732 d1
= (last
- s1
) * cfqd
->cfq_back_penalty
;
734 wrap
|= CFQ_RQ1_WRAP
;
738 else if (s2
+ back_max
>= last
)
739 d2
= (last
- s2
) * cfqd
->cfq_back_penalty
;
741 wrap
|= CFQ_RQ2_WRAP
;
743 /* Found required data */
746 * By doing switch() on the bit mask "wrap" we avoid having to
747 * check two variables for all permutations: --> faster!
750 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
766 case (CFQ_RQ1_WRAP
|CFQ_RQ2_WRAP
): /* both rqs wrapped */
769 * Since both rqs are wrapped,
770 * start with the one that's further behind head
771 * (--> only *one* back seek required),
772 * since back seek takes more time than forward.
782 * The below is leftmost cache rbtree addon
784 static struct cfq_queue
*cfq_rb_first(struct cfq_rb_root
*root
)
786 /* Service tree is empty */
791 root
->left
= rb_first(&root
->rb
);
794 return rb_entry(root
->left
, struct cfq_queue
, rb_node
);
799 static struct cfq_group
*cfq_rb_first_group(struct cfq_rb_root
*root
)
802 root
->left
= rb_first(&root
->rb
);
805 return rb_entry_cfqg(root
->left
);
810 static void rb_erase_init(struct rb_node
*n
, struct rb_root
*root
)
816 static void cfq_rb_erase(struct rb_node
*n
, struct cfq_rb_root
*root
)
820 rb_erase_init(n
, &root
->rb
);
825 * would be nice to take fifo expire time into account as well
827 static struct request
*
828 cfq_find_next_rq(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
829 struct request
*last
)
831 struct rb_node
*rbnext
= rb_next(&last
->rb_node
);
832 struct rb_node
*rbprev
= rb_prev(&last
->rb_node
);
833 struct request
*next
= NULL
, *prev
= NULL
;
835 BUG_ON(RB_EMPTY_NODE(&last
->rb_node
));
838 prev
= rb_entry_rq(rbprev
);
841 next
= rb_entry_rq(rbnext
);
843 rbnext
= rb_first(&cfqq
->sort_list
);
844 if (rbnext
&& rbnext
!= &last
->rb_node
)
845 next
= rb_entry_rq(rbnext
);
848 return cfq_choose_req(cfqd
, next
, prev
, blk_rq_pos(last
));
851 static unsigned long cfq_slice_offset(struct cfq_data
*cfqd
,
852 struct cfq_queue
*cfqq
)
855 * just an approximation, should be ok.
857 return (cfqq
->cfqg
->nr_cfqq
- 1) * (cfq_prio_slice(cfqd
, 1, 0) -
858 cfq_prio_slice(cfqd
, cfq_cfqq_sync(cfqq
), cfqq
->ioprio
));
862 cfqg_key(struct cfq_rb_root
*st
, struct cfq_group
*cfqg
)
864 return cfqg
->vdisktime
- st
->min_vdisktime
;
868 __cfq_group_service_tree_add(struct cfq_rb_root
*st
, struct cfq_group
*cfqg
)
870 struct rb_node
**node
= &st
->rb
.rb_node
;
871 struct rb_node
*parent
= NULL
;
872 struct cfq_group
*__cfqg
;
873 s64 key
= cfqg_key(st
, cfqg
);
876 while (*node
!= NULL
) {
878 __cfqg
= rb_entry_cfqg(parent
);
880 if (key
< cfqg_key(st
, __cfqg
))
881 node
= &parent
->rb_left
;
883 node
= &parent
->rb_right
;
889 st
->left
= &cfqg
->rb_node
;
891 rb_link_node(&cfqg
->rb_node
, parent
, node
);
892 rb_insert_color(&cfqg
->rb_node
, &st
->rb
);
896 cfq_update_group_weight(struct cfq_group
*cfqg
)
898 BUG_ON(!RB_EMPTY_NODE(&cfqg
->rb_node
));
899 if (cfqg
->needs_update
) {
900 cfqg
->weight
= cfqg
->new_weight
;
901 cfqg
->needs_update
= false;
906 cfq_group_service_tree_add(struct cfq_rb_root
*st
, struct cfq_group
*cfqg
)
908 BUG_ON(!RB_EMPTY_NODE(&cfqg
->rb_node
));
910 cfq_update_group_weight(cfqg
);
911 __cfq_group_service_tree_add(st
, cfqg
);
912 st
->total_weight
+= cfqg
->weight
;
916 cfq_group_notify_queue_add(struct cfq_data
*cfqd
, struct cfq_group
*cfqg
)
918 struct cfq_rb_root
*st
= &cfqd
->grp_service_tree
;
919 struct cfq_group
*__cfqg
;
923 if (!RB_EMPTY_NODE(&cfqg
->rb_node
))
927 * Currently put the group at the end. Later implement something
928 * so that groups get lesser vtime based on their weights, so that
929 * if group does not loose all if it was not continuously backlogged.
931 n
= rb_last(&st
->rb
);
933 __cfqg
= rb_entry_cfqg(n
);
934 cfqg
->vdisktime
= __cfqg
->vdisktime
+ CFQ_IDLE_DELAY
;
936 cfqg
->vdisktime
= st
->min_vdisktime
;
937 cfq_group_service_tree_add(st
, cfqg
);
941 cfq_group_service_tree_del(struct cfq_rb_root
*st
, struct cfq_group
*cfqg
)
943 st
->total_weight
-= cfqg
->weight
;
944 if (!RB_EMPTY_NODE(&cfqg
->rb_node
))
945 cfq_rb_erase(&cfqg
->rb_node
, st
);
949 cfq_group_notify_queue_del(struct cfq_data
*cfqd
, struct cfq_group
*cfqg
)
951 struct cfq_rb_root
*st
= &cfqd
->grp_service_tree
;
953 BUG_ON(cfqg
->nr_cfqq
< 1);
956 /* If there are other cfq queues under this group, don't delete it */
960 cfq_log_cfqg(cfqd
, cfqg
, "del_from_rr group");
961 cfq_group_service_tree_del(st
, cfqg
);
962 cfqg
->saved_workload_slice
= 0;
963 cfq_blkiocg_update_dequeue_stats(cfqg_to_blkg(cfqg
),
964 &blkio_policy_cfq
, 1);
967 static inline unsigned int cfq_cfqq_slice_usage(struct cfq_queue
*cfqq
,
968 unsigned int *unaccounted_time
)
970 unsigned int slice_used
;
973 * Queue got expired before even a single request completed or
974 * got expired immediately after first request completion.
976 if (!cfqq
->slice_start
|| cfqq
->slice_start
== jiffies
) {
978 * Also charge the seek time incurred to the group, otherwise
979 * if there are mutiple queues in the group, each can dispatch
980 * a single request on seeky media and cause lots of seek time
981 * and group will never know it.
983 slice_used
= max_t(unsigned, (jiffies
- cfqq
->dispatch_start
),
986 slice_used
= jiffies
- cfqq
->slice_start
;
987 if (slice_used
> cfqq
->allocated_slice
) {
988 *unaccounted_time
= slice_used
- cfqq
->allocated_slice
;
989 slice_used
= cfqq
->allocated_slice
;
991 if (time_after(cfqq
->slice_start
, cfqq
->dispatch_start
))
992 *unaccounted_time
+= cfqq
->slice_start
-
993 cfqq
->dispatch_start
;
999 static void cfq_group_served(struct cfq_data
*cfqd
, struct cfq_group
*cfqg
,
1000 struct cfq_queue
*cfqq
)
1002 struct cfq_rb_root
*st
= &cfqd
->grp_service_tree
;
1003 unsigned int used_sl
, charge
, unaccounted_sl
= 0;
1004 int nr_sync
= cfqg
->nr_cfqq
- cfqg_busy_async_queues(cfqd
, cfqg
)
1005 - cfqg
->service_tree_idle
.count
;
1007 BUG_ON(nr_sync
< 0);
1008 used_sl
= charge
= cfq_cfqq_slice_usage(cfqq
, &unaccounted_sl
);
1010 if (iops_mode(cfqd
))
1011 charge
= cfqq
->slice_dispatch
;
1012 else if (!cfq_cfqq_sync(cfqq
) && !nr_sync
)
1013 charge
= cfqq
->allocated_slice
;
1015 /* Can't update vdisktime while group is on service tree */
1016 cfq_group_service_tree_del(st
, cfqg
);
1017 cfqg
->vdisktime
+= cfq_scale_slice(charge
, cfqg
);
1018 /* If a new weight was requested, update now, off tree */
1019 cfq_group_service_tree_add(st
, cfqg
);
1021 /* This group is being expired. Save the context */
1022 if (time_after(cfqd
->workload_expires
, jiffies
)) {
1023 cfqg
->saved_workload_slice
= cfqd
->workload_expires
1025 cfqg
->saved_workload
= cfqd
->serving_type
;
1026 cfqg
->saved_serving_prio
= cfqd
->serving_prio
;
1028 cfqg
->saved_workload_slice
= 0;
1030 cfq_log_cfqg(cfqd
, cfqg
, "served: vt=%llu min_vt=%llu", cfqg
->vdisktime
,
1032 cfq_log_cfqq(cfqq
->cfqd
, cfqq
,
1033 "sl_used=%u disp=%u charge=%u iops=%u sect=%lu",
1034 used_sl
, cfqq
->slice_dispatch
, charge
,
1035 iops_mode(cfqd
), cfqq
->nr_sectors
);
1036 cfq_blkiocg_update_timeslice_used(cfqg_to_blkg(cfqg
), &blkio_policy_cfq
,
1037 used_sl
, unaccounted_sl
);
1038 cfq_blkiocg_set_start_empty_time(cfqg_to_blkg(cfqg
), &blkio_policy_cfq
);
1042 * cfq_init_cfqg_base - initialize base part of a cfq_group
1043 * @cfqg: cfq_group to initialize
1045 * Initialize the base part which is used whether %CONFIG_CFQ_GROUP_IOSCHED
1046 * is enabled or not.
1048 static void cfq_init_cfqg_base(struct cfq_group
*cfqg
)
1050 struct cfq_rb_root
*st
;
1053 for_each_cfqg_st(cfqg
, i
, j
, st
)
1055 RB_CLEAR_NODE(&cfqg
->rb_node
);
1057 cfqg
->ttime
.last_end_request
= jiffies
;
1060 #ifdef CONFIG_CFQ_GROUP_IOSCHED
1061 static void cfq_update_blkio_group_weight(struct request_queue
*q
,
1062 struct blkio_group
*blkg
,
1063 unsigned int weight
)
1065 struct cfq_group
*cfqg
= blkg_to_cfqg(blkg
);
1067 cfqg
->new_weight
= weight
;
1068 cfqg
->needs_update
= true;
1071 static void cfq_init_blkio_group(struct blkio_group
*blkg
)
1073 struct cfq_group
*cfqg
= blkg_to_cfqg(blkg
);
1075 cfq_init_cfqg_base(cfqg
);
1076 cfqg
->weight
= blkg
->blkcg
->weight
;
1080 * Search for the cfq group current task belongs to. request_queue lock must
1083 static struct cfq_group
*cfq_lookup_create_cfqg(struct cfq_data
*cfqd
,
1084 struct blkio_cgroup
*blkcg
)
1086 struct request_queue
*q
= cfqd
->queue
;
1087 struct cfq_group
*cfqg
= NULL
;
1089 /* avoid lookup for the common case where there's no blkio cgroup */
1090 if (blkcg
== &blkio_root_cgroup
) {
1091 cfqg
= cfqd
->root_group
;
1093 struct blkio_group
*blkg
;
1095 blkg
= blkg_lookup_create(blkcg
, q
, false);
1097 cfqg
= blkg_to_cfqg(blkg
);
1103 static void cfq_link_cfqq_cfqg(struct cfq_queue
*cfqq
, struct cfq_group
*cfqg
)
1105 /* Currently, all async queues are mapped to root group */
1106 if (!cfq_cfqq_sync(cfqq
))
1107 cfqg
= cfqq
->cfqd
->root_group
;
1110 /* cfqq reference on cfqg */
1114 #else /* GROUP_IOSCHED */
1115 static struct cfq_group
*cfq_lookup_create_cfqg(struct cfq_data
*cfqd
,
1116 struct blkio_cgroup
*blkcg
)
1118 return cfqd
->root_group
;
1122 cfq_link_cfqq_cfqg(struct cfq_queue
*cfqq
, struct cfq_group
*cfqg
) {
1126 #endif /* GROUP_IOSCHED */
1129 * The cfqd->service_trees holds all pending cfq_queue's that have
1130 * requests waiting to be processed. It is sorted in the order that
1131 * we will service the queues.
1133 static void cfq_service_tree_add(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
1136 struct rb_node
**p
, *parent
;
1137 struct cfq_queue
*__cfqq
;
1138 unsigned long rb_key
;
1139 struct cfq_rb_root
*service_tree
;
1143 service_tree
= service_tree_for(cfqq
->cfqg
, cfqq_prio(cfqq
),
1145 if (cfq_class_idle(cfqq
)) {
1146 rb_key
= CFQ_IDLE_DELAY
;
1147 parent
= rb_last(&service_tree
->rb
);
1148 if (parent
&& parent
!= &cfqq
->rb_node
) {
1149 __cfqq
= rb_entry(parent
, struct cfq_queue
, rb_node
);
1150 rb_key
+= __cfqq
->rb_key
;
1153 } else if (!add_front
) {
1155 * Get our rb key offset. Subtract any residual slice
1156 * value carried from last service. A negative resid
1157 * count indicates slice overrun, and this should position
1158 * the next service time further away in the tree.
1160 rb_key
= cfq_slice_offset(cfqd
, cfqq
) + jiffies
;
1161 rb_key
-= cfqq
->slice_resid
;
1162 cfqq
->slice_resid
= 0;
1165 __cfqq
= cfq_rb_first(service_tree
);
1166 rb_key
+= __cfqq
? __cfqq
->rb_key
: jiffies
;
1169 if (!RB_EMPTY_NODE(&cfqq
->rb_node
)) {
1172 * same position, nothing more to do
1174 if (rb_key
== cfqq
->rb_key
&&
1175 cfqq
->service_tree
== service_tree
)
1178 cfq_rb_erase(&cfqq
->rb_node
, cfqq
->service_tree
);
1179 cfqq
->service_tree
= NULL
;
1184 cfqq
->service_tree
= service_tree
;
1185 p
= &service_tree
->rb
.rb_node
;
1190 __cfqq
= rb_entry(parent
, struct cfq_queue
, rb_node
);
1193 * sort by key, that represents service time.
1195 if (time_before(rb_key
, __cfqq
->rb_key
))
1198 n
= &(*p
)->rb_right
;
1206 service_tree
->left
= &cfqq
->rb_node
;
1208 cfqq
->rb_key
= rb_key
;
1209 rb_link_node(&cfqq
->rb_node
, parent
, p
);
1210 rb_insert_color(&cfqq
->rb_node
, &service_tree
->rb
);
1211 service_tree
->count
++;
1212 if (add_front
|| !new_cfqq
)
1214 cfq_group_notify_queue_add(cfqd
, cfqq
->cfqg
);
1217 static struct cfq_queue
*
1218 cfq_prio_tree_lookup(struct cfq_data
*cfqd
, struct rb_root
*root
,
1219 sector_t sector
, struct rb_node
**ret_parent
,
1220 struct rb_node
***rb_link
)
1222 struct rb_node
**p
, *parent
;
1223 struct cfq_queue
*cfqq
= NULL
;
1231 cfqq
= rb_entry(parent
, struct cfq_queue
, p_node
);
1234 * Sort strictly based on sector. Smallest to the left,
1235 * largest to the right.
1237 if (sector
> blk_rq_pos(cfqq
->next_rq
))
1238 n
= &(*p
)->rb_right
;
1239 else if (sector
< blk_rq_pos(cfqq
->next_rq
))
1247 *ret_parent
= parent
;
1253 static void cfq_prio_tree_add(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
1255 struct rb_node
**p
, *parent
;
1256 struct cfq_queue
*__cfqq
;
1259 rb_erase(&cfqq
->p_node
, cfqq
->p_root
);
1260 cfqq
->p_root
= NULL
;
1263 if (cfq_class_idle(cfqq
))
1268 cfqq
->p_root
= &cfqd
->prio_trees
[cfqq
->org_ioprio
];
1269 __cfqq
= cfq_prio_tree_lookup(cfqd
, cfqq
->p_root
,
1270 blk_rq_pos(cfqq
->next_rq
), &parent
, &p
);
1272 rb_link_node(&cfqq
->p_node
, parent
, p
);
1273 rb_insert_color(&cfqq
->p_node
, cfqq
->p_root
);
1275 cfqq
->p_root
= NULL
;
1279 * Update cfqq's position in the service tree.
1281 static void cfq_resort_rr_list(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
1284 * Resorting requires the cfqq to be on the RR list already.
1286 if (cfq_cfqq_on_rr(cfqq
)) {
1287 cfq_service_tree_add(cfqd
, cfqq
, 0);
1288 cfq_prio_tree_add(cfqd
, cfqq
);
1293 * add to busy list of queues for service, trying to be fair in ordering
1294 * the pending list according to last request service
1296 static void cfq_add_cfqq_rr(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
1298 cfq_log_cfqq(cfqd
, cfqq
, "add_to_rr");
1299 BUG_ON(cfq_cfqq_on_rr(cfqq
));
1300 cfq_mark_cfqq_on_rr(cfqq
);
1301 cfqd
->busy_queues
++;
1302 if (cfq_cfqq_sync(cfqq
))
1303 cfqd
->busy_sync_queues
++;
1305 cfq_resort_rr_list(cfqd
, cfqq
);
1309 * Called when the cfqq no longer has requests pending, remove it from
1312 static void cfq_del_cfqq_rr(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
1314 cfq_log_cfqq(cfqd
, cfqq
, "del_from_rr");
1315 BUG_ON(!cfq_cfqq_on_rr(cfqq
));
1316 cfq_clear_cfqq_on_rr(cfqq
);
1318 if (!RB_EMPTY_NODE(&cfqq
->rb_node
)) {
1319 cfq_rb_erase(&cfqq
->rb_node
, cfqq
->service_tree
);
1320 cfqq
->service_tree
= NULL
;
1323 rb_erase(&cfqq
->p_node
, cfqq
->p_root
);
1324 cfqq
->p_root
= NULL
;
1327 cfq_group_notify_queue_del(cfqd
, cfqq
->cfqg
);
1328 BUG_ON(!cfqd
->busy_queues
);
1329 cfqd
->busy_queues
--;
1330 if (cfq_cfqq_sync(cfqq
))
1331 cfqd
->busy_sync_queues
--;
1335 * rb tree support functions
1337 static void cfq_del_rq_rb(struct request
*rq
)
1339 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
1340 const int sync
= rq_is_sync(rq
);
1342 BUG_ON(!cfqq
->queued
[sync
]);
1343 cfqq
->queued
[sync
]--;
1345 elv_rb_del(&cfqq
->sort_list
, rq
);
1347 if (cfq_cfqq_on_rr(cfqq
) && RB_EMPTY_ROOT(&cfqq
->sort_list
)) {
1349 * Queue will be deleted from service tree when we actually
1350 * expire it later. Right now just remove it from prio tree
1354 rb_erase(&cfqq
->p_node
, cfqq
->p_root
);
1355 cfqq
->p_root
= NULL
;
1360 static void cfq_add_rq_rb(struct request
*rq
)
1362 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
1363 struct cfq_data
*cfqd
= cfqq
->cfqd
;
1364 struct request
*prev
;
1366 cfqq
->queued
[rq_is_sync(rq
)]++;
1368 elv_rb_add(&cfqq
->sort_list
, rq
);
1370 if (!cfq_cfqq_on_rr(cfqq
))
1371 cfq_add_cfqq_rr(cfqd
, cfqq
);
1374 * check if this request is a better next-serve candidate
1376 prev
= cfqq
->next_rq
;
1377 cfqq
->next_rq
= cfq_choose_req(cfqd
, cfqq
->next_rq
, rq
, cfqd
->last_position
);
1380 * adjust priority tree position, if ->next_rq changes
1382 if (prev
!= cfqq
->next_rq
)
1383 cfq_prio_tree_add(cfqd
, cfqq
);
1385 BUG_ON(!cfqq
->next_rq
);
1388 static void cfq_reposition_rq_rb(struct cfq_queue
*cfqq
, struct request
*rq
)
1390 elv_rb_del(&cfqq
->sort_list
, rq
);
1391 cfqq
->queued
[rq_is_sync(rq
)]--;
1392 cfq_blkiocg_update_io_remove_stats(cfqg_to_blkg(RQ_CFQG(rq
)),
1393 &blkio_policy_cfq
, rq_data_dir(rq
),
1396 cfq_blkiocg_update_io_add_stats(cfqg_to_blkg(RQ_CFQG(rq
)),
1398 cfqg_to_blkg(cfqq
->cfqd
->serving_group
),
1399 rq_data_dir(rq
), rq_is_sync(rq
));
1402 static struct request
*
1403 cfq_find_rq_fmerge(struct cfq_data
*cfqd
, struct bio
*bio
)
1405 struct task_struct
*tsk
= current
;
1406 struct cfq_io_cq
*cic
;
1407 struct cfq_queue
*cfqq
;
1409 cic
= cfq_cic_lookup(cfqd
, tsk
->io_context
);
1413 cfqq
= cic_to_cfqq(cic
, cfq_bio_sync(bio
));
1415 sector_t sector
= bio
->bi_sector
+ bio_sectors(bio
);
1417 return elv_rb_find(&cfqq
->sort_list
, sector
);
1423 static void cfq_activate_request(struct request_queue
*q
, struct request
*rq
)
1425 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
1427 cfqd
->rq_in_driver
++;
1428 cfq_log_cfqq(cfqd
, RQ_CFQQ(rq
), "activate rq, drv=%d",
1429 cfqd
->rq_in_driver
);
1431 cfqd
->last_position
= blk_rq_pos(rq
) + blk_rq_sectors(rq
);
1434 static void cfq_deactivate_request(struct request_queue
*q
, struct request
*rq
)
1436 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
1438 WARN_ON(!cfqd
->rq_in_driver
);
1439 cfqd
->rq_in_driver
--;
1440 cfq_log_cfqq(cfqd
, RQ_CFQQ(rq
), "deactivate rq, drv=%d",
1441 cfqd
->rq_in_driver
);
1444 static void cfq_remove_request(struct request
*rq
)
1446 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
1448 if (cfqq
->next_rq
== rq
)
1449 cfqq
->next_rq
= cfq_find_next_rq(cfqq
->cfqd
, cfqq
, rq
);
1451 list_del_init(&rq
->queuelist
);
1454 cfqq
->cfqd
->rq_queued
--;
1455 cfq_blkiocg_update_io_remove_stats(cfqg_to_blkg(RQ_CFQG(rq
)),
1456 &blkio_policy_cfq
, rq_data_dir(rq
),
1458 if (rq
->cmd_flags
& REQ_PRIO
) {
1459 WARN_ON(!cfqq
->prio_pending
);
1460 cfqq
->prio_pending
--;
1464 static int cfq_merge(struct request_queue
*q
, struct request
**req
,
1467 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
1468 struct request
*__rq
;
1470 __rq
= cfq_find_rq_fmerge(cfqd
, bio
);
1471 if (__rq
&& elv_rq_merge_ok(__rq
, bio
)) {
1473 return ELEVATOR_FRONT_MERGE
;
1476 return ELEVATOR_NO_MERGE
;
1479 static void cfq_merged_request(struct request_queue
*q
, struct request
*req
,
1482 if (type
== ELEVATOR_FRONT_MERGE
) {
1483 struct cfq_queue
*cfqq
= RQ_CFQQ(req
);
1485 cfq_reposition_rq_rb(cfqq
, req
);
1489 static void cfq_bio_merged(struct request_queue
*q
, struct request
*req
,
1492 cfq_blkiocg_update_io_merged_stats(cfqg_to_blkg(RQ_CFQG(req
)),
1493 &blkio_policy_cfq
, bio_data_dir(bio
),
1498 cfq_merged_requests(struct request_queue
*q
, struct request
*rq
,
1499 struct request
*next
)
1501 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
1502 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
1505 * reposition in fifo if next is older than rq
1507 if (!list_empty(&rq
->queuelist
) && !list_empty(&next
->queuelist
) &&
1508 time_before(rq_fifo_time(next
), rq_fifo_time(rq
))) {
1509 list_move(&rq
->queuelist
, &next
->queuelist
);
1510 rq_set_fifo_time(rq
, rq_fifo_time(next
));
1513 if (cfqq
->next_rq
== next
)
1515 cfq_remove_request(next
);
1516 cfq_blkiocg_update_io_merged_stats(cfqg_to_blkg(RQ_CFQG(rq
)),
1517 &blkio_policy_cfq
, rq_data_dir(next
),
1520 cfqq
= RQ_CFQQ(next
);
1522 * all requests of this queue are merged to other queues, delete it
1523 * from the service tree. If it's the active_queue,
1524 * cfq_dispatch_requests() will choose to expire it or do idle
1526 if (cfq_cfqq_on_rr(cfqq
) && RB_EMPTY_ROOT(&cfqq
->sort_list
) &&
1527 cfqq
!= cfqd
->active_queue
)
1528 cfq_del_cfqq_rr(cfqd
, cfqq
);
1531 static int cfq_allow_merge(struct request_queue
*q
, struct request
*rq
,
1534 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
1535 struct cfq_io_cq
*cic
;
1536 struct cfq_queue
*cfqq
;
1539 * Disallow merge of a sync bio into an async request.
1541 if (cfq_bio_sync(bio
) && !rq_is_sync(rq
))
1545 * Lookup the cfqq that this bio will be queued with and allow
1546 * merge only if rq is queued there.
1548 cic
= cfq_cic_lookup(cfqd
, current
->io_context
);
1552 cfqq
= cic_to_cfqq(cic
, cfq_bio_sync(bio
));
1553 return cfqq
== RQ_CFQQ(rq
);
1556 static inline void cfq_del_timer(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
1558 del_timer(&cfqd
->idle_slice_timer
);
1559 cfq_blkiocg_update_idle_time_stats(cfqg_to_blkg(cfqq
->cfqg
),
1563 static void __cfq_set_active_queue(struct cfq_data
*cfqd
,
1564 struct cfq_queue
*cfqq
)
1567 cfq_log_cfqq(cfqd
, cfqq
, "set_active wl_prio:%d wl_type:%d",
1568 cfqd
->serving_prio
, cfqd
->serving_type
);
1569 cfq_blkiocg_update_avg_queue_size_stats(cfqg_to_blkg(cfqq
->cfqg
),
1571 cfqq
->slice_start
= 0;
1572 cfqq
->dispatch_start
= jiffies
;
1573 cfqq
->allocated_slice
= 0;
1574 cfqq
->slice_end
= 0;
1575 cfqq
->slice_dispatch
= 0;
1576 cfqq
->nr_sectors
= 0;
1578 cfq_clear_cfqq_wait_request(cfqq
);
1579 cfq_clear_cfqq_must_dispatch(cfqq
);
1580 cfq_clear_cfqq_must_alloc_slice(cfqq
);
1581 cfq_clear_cfqq_fifo_expire(cfqq
);
1582 cfq_mark_cfqq_slice_new(cfqq
);
1584 cfq_del_timer(cfqd
, cfqq
);
1587 cfqd
->active_queue
= cfqq
;
1591 * current cfqq expired its slice (or was too idle), select new one
1594 __cfq_slice_expired(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
1597 cfq_log_cfqq(cfqd
, cfqq
, "slice expired t=%d", timed_out
);
1599 if (cfq_cfqq_wait_request(cfqq
))
1600 cfq_del_timer(cfqd
, cfqq
);
1602 cfq_clear_cfqq_wait_request(cfqq
);
1603 cfq_clear_cfqq_wait_busy(cfqq
);
1606 * If this cfqq is shared between multiple processes, check to
1607 * make sure that those processes are still issuing I/Os within
1608 * the mean seek distance. If not, it may be time to break the
1609 * queues apart again.
1611 if (cfq_cfqq_coop(cfqq
) && CFQQ_SEEKY(cfqq
))
1612 cfq_mark_cfqq_split_coop(cfqq
);
1615 * store what was left of this slice, if the queue idled/timed out
1618 if (cfq_cfqq_slice_new(cfqq
))
1619 cfqq
->slice_resid
= cfq_scaled_cfqq_slice(cfqd
, cfqq
);
1621 cfqq
->slice_resid
= cfqq
->slice_end
- jiffies
;
1622 cfq_log_cfqq(cfqd
, cfqq
, "resid=%ld", cfqq
->slice_resid
);
1625 cfq_group_served(cfqd
, cfqq
->cfqg
, cfqq
);
1627 if (cfq_cfqq_on_rr(cfqq
) && RB_EMPTY_ROOT(&cfqq
->sort_list
))
1628 cfq_del_cfqq_rr(cfqd
, cfqq
);
1630 cfq_resort_rr_list(cfqd
, cfqq
);
1632 if (cfqq
== cfqd
->active_queue
)
1633 cfqd
->active_queue
= NULL
;
1635 if (cfqd
->active_cic
) {
1636 put_io_context(cfqd
->active_cic
->icq
.ioc
);
1637 cfqd
->active_cic
= NULL
;
1641 static inline void cfq_slice_expired(struct cfq_data
*cfqd
, bool timed_out
)
1643 struct cfq_queue
*cfqq
= cfqd
->active_queue
;
1646 __cfq_slice_expired(cfqd
, cfqq
, timed_out
);
1650 * Get next queue for service. Unless we have a queue preemption,
1651 * we'll simply select the first cfqq in the service tree.
1653 static struct cfq_queue
*cfq_get_next_queue(struct cfq_data
*cfqd
)
1655 struct cfq_rb_root
*service_tree
=
1656 service_tree_for(cfqd
->serving_group
, cfqd
->serving_prio
,
1657 cfqd
->serving_type
);
1659 if (!cfqd
->rq_queued
)
1662 /* There is nothing to dispatch */
1665 if (RB_EMPTY_ROOT(&service_tree
->rb
))
1667 return cfq_rb_first(service_tree
);
1670 static struct cfq_queue
*cfq_get_next_queue_forced(struct cfq_data
*cfqd
)
1672 struct cfq_group
*cfqg
;
1673 struct cfq_queue
*cfqq
;
1675 struct cfq_rb_root
*st
;
1677 if (!cfqd
->rq_queued
)
1680 cfqg
= cfq_get_next_cfqg(cfqd
);
1684 for_each_cfqg_st(cfqg
, i
, j
, st
)
1685 if ((cfqq
= cfq_rb_first(st
)) != NULL
)
1691 * Get and set a new active queue for service.
1693 static struct cfq_queue
*cfq_set_active_queue(struct cfq_data
*cfqd
,
1694 struct cfq_queue
*cfqq
)
1697 cfqq
= cfq_get_next_queue(cfqd
);
1699 __cfq_set_active_queue(cfqd
, cfqq
);
1703 static inline sector_t
cfq_dist_from_last(struct cfq_data
*cfqd
,
1706 if (blk_rq_pos(rq
) >= cfqd
->last_position
)
1707 return blk_rq_pos(rq
) - cfqd
->last_position
;
1709 return cfqd
->last_position
- blk_rq_pos(rq
);
1712 static inline int cfq_rq_close(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
1715 return cfq_dist_from_last(cfqd
, rq
) <= CFQQ_CLOSE_THR
;
1718 static struct cfq_queue
*cfqq_close(struct cfq_data
*cfqd
,
1719 struct cfq_queue
*cur_cfqq
)
1721 struct rb_root
*root
= &cfqd
->prio_trees
[cur_cfqq
->org_ioprio
];
1722 struct rb_node
*parent
, *node
;
1723 struct cfq_queue
*__cfqq
;
1724 sector_t sector
= cfqd
->last_position
;
1726 if (RB_EMPTY_ROOT(root
))
1730 * First, if we find a request starting at the end of the last
1731 * request, choose it.
1733 __cfqq
= cfq_prio_tree_lookup(cfqd
, root
, sector
, &parent
, NULL
);
1738 * If the exact sector wasn't found, the parent of the NULL leaf
1739 * will contain the closest sector.
1741 __cfqq
= rb_entry(parent
, struct cfq_queue
, p_node
);
1742 if (cfq_rq_close(cfqd
, cur_cfqq
, __cfqq
->next_rq
))
1745 if (blk_rq_pos(__cfqq
->next_rq
) < sector
)
1746 node
= rb_next(&__cfqq
->p_node
);
1748 node
= rb_prev(&__cfqq
->p_node
);
1752 __cfqq
= rb_entry(node
, struct cfq_queue
, p_node
);
1753 if (cfq_rq_close(cfqd
, cur_cfqq
, __cfqq
->next_rq
))
1761 * cur_cfqq - passed in so that we don't decide that the current queue is
1762 * closely cooperating with itself.
1764 * So, basically we're assuming that that cur_cfqq has dispatched at least
1765 * one request, and that cfqd->last_position reflects a position on the disk
1766 * associated with the I/O issued by cur_cfqq. I'm not sure this is a valid
1769 static struct cfq_queue
*cfq_close_cooperator(struct cfq_data
*cfqd
,
1770 struct cfq_queue
*cur_cfqq
)
1772 struct cfq_queue
*cfqq
;
1774 if (cfq_class_idle(cur_cfqq
))
1776 if (!cfq_cfqq_sync(cur_cfqq
))
1778 if (CFQQ_SEEKY(cur_cfqq
))
1782 * Don't search priority tree if it's the only queue in the group.
1784 if (cur_cfqq
->cfqg
->nr_cfqq
== 1)
1788 * We should notice if some of the queues are cooperating, eg
1789 * working closely on the same area of the disk. In that case,
1790 * we can group them together and don't waste time idling.
1792 cfqq
= cfqq_close(cfqd
, cur_cfqq
);
1796 /* If new queue belongs to different cfq_group, don't choose it */
1797 if (cur_cfqq
->cfqg
!= cfqq
->cfqg
)
1801 * It only makes sense to merge sync queues.
1803 if (!cfq_cfqq_sync(cfqq
))
1805 if (CFQQ_SEEKY(cfqq
))
1809 * Do not merge queues of different priority classes
1811 if (cfq_class_rt(cfqq
) != cfq_class_rt(cur_cfqq
))
1818 * Determine whether we should enforce idle window for this queue.
1821 static bool cfq_should_idle(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
1823 enum wl_prio_t prio
= cfqq_prio(cfqq
);
1824 struct cfq_rb_root
*service_tree
= cfqq
->service_tree
;
1826 BUG_ON(!service_tree
);
1827 BUG_ON(!service_tree
->count
);
1829 if (!cfqd
->cfq_slice_idle
)
1832 /* We never do for idle class queues. */
1833 if (prio
== IDLE_WORKLOAD
)
1836 /* We do for queues that were marked with idle window flag. */
1837 if (cfq_cfqq_idle_window(cfqq
) &&
1838 !(blk_queue_nonrot(cfqd
->queue
) && cfqd
->hw_tag
))
1842 * Otherwise, we do only if they are the last ones
1843 * in their service tree.
1845 if (service_tree
->count
== 1 && cfq_cfqq_sync(cfqq
) &&
1846 !cfq_io_thinktime_big(cfqd
, &service_tree
->ttime
, false))
1848 cfq_log_cfqq(cfqd
, cfqq
, "Not idling. st->count:%d",
1849 service_tree
->count
);
1853 static void cfq_arm_slice_timer(struct cfq_data
*cfqd
)
1855 struct cfq_queue
*cfqq
= cfqd
->active_queue
;
1856 struct cfq_io_cq
*cic
;
1857 unsigned long sl
, group_idle
= 0;
1860 * SSD device without seek penalty, disable idling. But only do so
1861 * for devices that support queuing, otherwise we still have a problem
1862 * with sync vs async workloads.
1864 if (blk_queue_nonrot(cfqd
->queue
) && cfqd
->hw_tag
)
1867 WARN_ON(!RB_EMPTY_ROOT(&cfqq
->sort_list
));
1868 WARN_ON(cfq_cfqq_slice_new(cfqq
));
1871 * idle is disabled, either manually or by past process history
1873 if (!cfq_should_idle(cfqd
, cfqq
)) {
1874 /* no queue idling. Check for group idling */
1875 if (cfqd
->cfq_group_idle
)
1876 group_idle
= cfqd
->cfq_group_idle
;
1882 * still active requests from this queue, don't idle
1884 if (cfqq
->dispatched
)
1888 * task has exited, don't wait
1890 cic
= cfqd
->active_cic
;
1891 if (!cic
|| !atomic_read(&cic
->icq
.ioc
->active_ref
))
1895 * If our average think time is larger than the remaining time
1896 * slice, then don't idle. This avoids overrunning the allotted
1899 if (sample_valid(cic
->ttime
.ttime_samples
) &&
1900 (cfqq
->slice_end
- jiffies
< cic
->ttime
.ttime_mean
)) {
1901 cfq_log_cfqq(cfqd
, cfqq
, "Not idling. think_time:%lu",
1902 cic
->ttime
.ttime_mean
);
1906 /* There are other queues in the group, don't do group idle */
1907 if (group_idle
&& cfqq
->cfqg
->nr_cfqq
> 1)
1910 cfq_mark_cfqq_wait_request(cfqq
);
1913 sl
= cfqd
->cfq_group_idle
;
1915 sl
= cfqd
->cfq_slice_idle
;
1917 mod_timer(&cfqd
->idle_slice_timer
, jiffies
+ sl
);
1918 cfq_blkiocg_update_set_idle_time_stats(cfqg_to_blkg(cfqq
->cfqg
),
1920 cfq_log_cfqq(cfqd
, cfqq
, "arm_idle: %lu group_idle: %d", sl
,
1921 group_idle
? 1 : 0);
1925 * Move request from internal lists to the request queue dispatch list.
1927 static void cfq_dispatch_insert(struct request_queue
*q
, struct request
*rq
)
1929 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
1930 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
1932 cfq_log_cfqq(cfqd
, cfqq
, "dispatch_insert");
1934 cfqq
->next_rq
= cfq_find_next_rq(cfqd
, cfqq
, rq
);
1935 cfq_remove_request(rq
);
1937 (RQ_CFQG(rq
))->dispatched
++;
1938 elv_dispatch_sort(q
, rq
);
1940 cfqd
->rq_in_flight
[cfq_cfqq_sync(cfqq
)]++;
1941 cfqq
->nr_sectors
+= blk_rq_sectors(rq
);
1942 cfq_blkiocg_update_dispatch_stats(cfqg_to_blkg(cfqq
->cfqg
),
1943 &blkio_policy_cfq
, blk_rq_bytes(rq
),
1944 rq_data_dir(rq
), rq_is_sync(rq
));
1948 * return expired entry, or NULL to just start from scratch in rbtree
1950 static struct request
*cfq_check_fifo(struct cfq_queue
*cfqq
)
1952 struct request
*rq
= NULL
;
1954 if (cfq_cfqq_fifo_expire(cfqq
))
1957 cfq_mark_cfqq_fifo_expire(cfqq
);
1959 if (list_empty(&cfqq
->fifo
))
1962 rq
= rq_entry_fifo(cfqq
->fifo
.next
);
1963 if (time_before(jiffies
, rq_fifo_time(rq
)))
1966 cfq_log_cfqq(cfqq
->cfqd
, cfqq
, "fifo=%p", rq
);
1971 cfq_prio_to_maxrq(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
1973 const int base_rq
= cfqd
->cfq_slice_async_rq
;
1975 WARN_ON(cfqq
->ioprio
>= IOPRIO_BE_NR
);
1977 return 2 * base_rq
* (IOPRIO_BE_NR
- cfqq
->ioprio
);
1981 * Must be called with the queue_lock held.
1983 static int cfqq_process_refs(struct cfq_queue
*cfqq
)
1985 int process_refs
, io_refs
;
1987 io_refs
= cfqq
->allocated
[READ
] + cfqq
->allocated
[WRITE
];
1988 process_refs
= cfqq
->ref
- io_refs
;
1989 BUG_ON(process_refs
< 0);
1990 return process_refs
;
1993 static void cfq_setup_merge(struct cfq_queue
*cfqq
, struct cfq_queue
*new_cfqq
)
1995 int process_refs
, new_process_refs
;
1996 struct cfq_queue
*__cfqq
;
1999 * If there are no process references on the new_cfqq, then it is
2000 * unsafe to follow the ->new_cfqq chain as other cfqq's in the
2001 * chain may have dropped their last reference (not just their
2002 * last process reference).
2004 if (!cfqq_process_refs(new_cfqq
))
2007 /* Avoid a circular list and skip interim queue merges */
2008 while ((__cfqq
= new_cfqq
->new_cfqq
)) {
2014 process_refs
= cfqq_process_refs(cfqq
);
2015 new_process_refs
= cfqq_process_refs(new_cfqq
);
2017 * If the process for the cfqq has gone away, there is no
2018 * sense in merging the queues.
2020 if (process_refs
== 0 || new_process_refs
== 0)
2024 * Merge in the direction of the lesser amount of work.
2026 if (new_process_refs
>= process_refs
) {
2027 cfqq
->new_cfqq
= new_cfqq
;
2028 new_cfqq
->ref
+= process_refs
;
2030 new_cfqq
->new_cfqq
= cfqq
;
2031 cfqq
->ref
+= new_process_refs
;
2035 static enum wl_type_t
cfq_choose_wl(struct cfq_data
*cfqd
,
2036 struct cfq_group
*cfqg
, enum wl_prio_t prio
)
2038 struct cfq_queue
*queue
;
2040 bool key_valid
= false;
2041 unsigned long lowest_key
= 0;
2042 enum wl_type_t cur_best
= SYNC_NOIDLE_WORKLOAD
;
2044 for (i
= 0; i
<= SYNC_WORKLOAD
; ++i
) {
2045 /* select the one with lowest rb_key */
2046 queue
= cfq_rb_first(service_tree_for(cfqg
, prio
, i
));
2048 (!key_valid
|| time_before(queue
->rb_key
, lowest_key
))) {
2049 lowest_key
= queue
->rb_key
;
2058 static void choose_service_tree(struct cfq_data
*cfqd
, struct cfq_group
*cfqg
)
2062 struct cfq_rb_root
*st
;
2063 unsigned group_slice
;
2064 enum wl_prio_t original_prio
= cfqd
->serving_prio
;
2066 /* Choose next priority. RT > BE > IDLE */
2067 if (cfq_group_busy_queues_wl(RT_WORKLOAD
, cfqd
, cfqg
))
2068 cfqd
->serving_prio
= RT_WORKLOAD
;
2069 else if (cfq_group_busy_queues_wl(BE_WORKLOAD
, cfqd
, cfqg
))
2070 cfqd
->serving_prio
= BE_WORKLOAD
;
2072 cfqd
->serving_prio
= IDLE_WORKLOAD
;
2073 cfqd
->workload_expires
= jiffies
+ 1;
2077 if (original_prio
!= cfqd
->serving_prio
)
2081 * For RT and BE, we have to choose also the type
2082 * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
2085 st
= service_tree_for(cfqg
, cfqd
->serving_prio
, cfqd
->serving_type
);
2089 * check workload expiration, and that we still have other queues ready
2091 if (count
&& !time_after(jiffies
, cfqd
->workload_expires
))
2095 /* otherwise select new workload type */
2096 cfqd
->serving_type
=
2097 cfq_choose_wl(cfqd
, cfqg
, cfqd
->serving_prio
);
2098 st
= service_tree_for(cfqg
, cfqd
->serving_prio
, cfqd
->serving_type
);
2102 * the workload slice is computed as a fraction of target latency
2103 * proportional to the number of queues in that workload, over
2104 * all the queues in the same priority class
2106 group_slice
= cfq_group_slice(cfqd
, cfqg
);
2108 slice
= group_slice
* count
/
2109 max_t(unsigned, cfqg
->busy_queues_avg
[cfqd
->serving_prio
],
2110 cfq_group_busy_queues_wl(cfqd
->serving_prio
, cfqd
, cfqg
));
2112 if (cfqd
->serving_type
== ASYNC_WORKLOAD
) {
2116 * Async queues are currently system wide. Just taking
2117 * proportion of queues with-in same group will lead to higher
2118 * async ratio system wide as generally root group is going
2119 * to have higher weight. A more accurate thing would be to
2120 * calculate system wide asnc/sync ratio.
2122 tmp
= cfq_target_latency
* cfqg_busy_async_queues(cfqd
, cfqg
);
2123 tmp
= tmp
/cfqd
->busy_queues
;
2124 slice
= min_t(unsigned, slice
, tmp
);
2126 /* async workload slice is scaled down according to
2127 * the sync/async slice ratio. */
2128 slice
= slice
* cfqd
->cfq_slice
[0] / cfqd
->cfq_slice
[1];
2130 /* sync workload slice is at least 2 * cfq_slice_idle */
2131 slice
= max(slice
, 2 * cfqd
->cfq_slice_idle
);
2133 slice
= max_t(unsigned, slice
, CFQ_MIN_TT
);
2134 cfq_log(cfqd
, "workload slice:%d", slice
);
2135 cfqd
->workload_expires
= jiffies
+ slice
;
2138 static struct cfq_group
*cfq_get_next_cfqg(struct cfq_data
*cfqd
)
2140 struct cfq_rb_root
*st
= &cfqd
->grp_service_tree
;
2141 struct cfq_group
*cfqg
;
2143 if (RB_EMPTY_ROOT(&st
->rb
))
2145 cfqg
= cfq_rb_first_group(st
);
2146 update_min_vdisktime(st
);
2150 static void cfq_choose_cfqg(struct cfq_data
*cfqd
)
2152 struct cfq_group
*cfqg
= cfq_get_next_cfqg(cfqd
);
2154 cfqd
->serving_group
= cfqg
;
2156 /* Restore the workload type data */
2157 if (cfqg
->saved_workload_slice
) {
2158 cfqd
->workload_expires
= jiffies
+ cfqg
->saved_workload_slice
;
2159 cfqd
->serving_type
= cfqg
->saved_workload
;
2160 cfqd
->serving_prio
= cfqg
->saved_serving_prio
;
2162 cfqd
->workload_expires
= jiffies
- 1;
2164 choose_service_tree(cfqd
, cfqg
);
2168 * Select a queue for service. If we have a current active queue,
2169 * check whether to continue servicing it, or retrieve and set a new one.
2171 static struct cfq_queue
*cfq_select_queue(struct cfq_data
*cfqd
)
2173 struct cfq_queue
*cfqq
, *new_cfqq
= NULL
;
2175 cfqq
= cfqd
->active_queue
;
2179 if (!cfqd
->rq_queued
)
2183 * We were waiting for group to get backlogged. Expire the queue
2185 if (cfq_cfqq_wait_busy(cfqq
) && !RB_EMPTY_ROOT(&cfqq
->sort_list
))
2189 * The active queue has run out of time, expire it and select new.
2191 if (cfq_slice_used(cfqq
) && !cfq_cfqq_must_dispatch(cfqq
)) {
2193 * If slice had not expired at the completion of last request
2194 * we might not have turned on wait_busy flag. Don't expire
2195 * the queue yet. Allow the group to get backlogged.
2197 * The very fact that we have used the slice, that means we
2198 * have been idling all along on this queue and it should be
2199 * ok to wait for this request to complete.
2201 if (cfqq
->cfqg
->nr_cfqq
== 1 && RB_EMPTY_ROOT(&cfqq
->sort_list
)
2202 && cfqq
->dispatched
&& cfq_should_idle(cfqd
, cfqq
)) {
2206 goto check_group_idle
;
2210 * The active queue has requests and isn't expired, allow it to
2213 if (!RB_EMPTY_ROOT(&cfqq
->sort_list
))
2217 * If another queue has a request waiting within our mean seek
2218 * distance, let it run. The expire code will check for close
2219 * cooperators and put the close queue at the front of the service
2220 * tree. If possible, merge the expiring queue with the new cfqq.
2222 new_cfqq
= cfq_close_cooperator(cfqd
, cfqq
);
2224 if (!cfqq
->new_cfqq
)
2225 cfq_setup_merge(cfqq
, new_cfqq
);
2230 * No requests pending. If the active queue still has requests in
2231 * flight or is idling for a new request, allow either of these
2232 * conditions to happen (or time out) before selecting a new queue.
2234 if (timer_pending(&cfqd
->idle_slice_timer
)) {
2240 * This is a deep seek queue, but the device is much faster than
2241 * the queue can deliver, don't idle
2243 if (CFQQ_SEEKY(cfqq
) && cfq_cfqq_idle_window(cfqq
) &&
2244 (cfq_cfqq_slice_new(cfqq
) ||
2245 (cfqq
->slice_end
- jiffies
> jiffies
- cfqq
->slice_start
))) {
2246 cfq_clear_cfqq_deep(cfqq
);
2247 cfq_clear_cfqq_idle_window(cfqq
);
2250 if (cfqq
->dispatched
&& cfq_should_idle(cfqd
, cfqq
)) {
2256 * If group idle is enabled and there are requests dispatched from
2257 * this group, wait for requests to complete.
2260 if (cfqd
->cfq_group_idle
&& cfqq
->cfqg
->nr_cfqq
== 1 &&
2261 cfqq
->cfqg
->dispatched
&&
2262 !cfq_io_thinktime_big(cfqd
, &cfqq
->cfqg
->ttime
, true)) {
2268 cfq_slice_expired(cfqd
, 0);
2271 * Current queue expired. Check if we have to switch to a new
2275 cfq_choose_cfqg(cfqd
);
2277 cfqq
= cfq_set_active_queue(cfqd
, new_cfqq
);
2282 static int __cfq_forced_dispatch_cfqq(struct cfq_queue
*cfqq
)
2286 while (cfqq
->next_rq
) {
2287 cfq_dispatch_insert(cfqq
->cfqd
->queue
, cfqq
->next_rq
);
2291 BUG_ON(!list_empty(&cfqq
->fifo
));
2293 /* By default cfqq is not expired if it is empty. Do it explicitly */
2294 __cfq_slice_expired(cfqq
->cfqd
, cfqq
, 0);
2299 * Drain our current requests. Used for barriers and when switching
2300 * io schedulers on-the-fly.
2302 static int cfq_forced_dispatch(struct cfq_data
*cfqd
)
2304 struct cfq_queue
*cfqq
;
2307 /* Expire the timeslice of the current active queue first */
2308 cfq_slice_expired(cfqd
, 0);
2309 while ((cfqq
= cfq_get_next_queue_forced(cfqd
)) != NULL
) {
2310 __cfq_set_active_queue(cfqd
, cfqq
);
2311 dispatched
+= __cfq_forced_dispatch_cfqq(cfqq
);
2314 BUG_ON(cfqd
->busy_queues
);
2316 cfq_log(cfqd
, "forced_dispatch=%d", dispatched
);
2320 static inline bool cfq_slice_used_soon(struct cfq_data
*cfqd
,
2321 struct cfq_queue
*cfqq
)
2323 /* the queue hasn't finished any request, can't estimate */
2324 if (cfq_cfqq_slice_new(cfqq
))
2326 if (time_after(jiffies
+ cfqd
->cfq_slice_idle
* cfqq
->dispatched
,
2333 static bool cfq_may_dispatch(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
2335 unsigned int max_dispatch
;
2338 * Drain async requests before we start sync IO
2340 if (cfq_should_idle(cfqd
, cfqq
) && cfqd
->rq_in_flight
[BLK_RW_ASYNC
])
2344 * If this is an async queue and we have sync IO in flight, let it wait
2346 if (cfqd
->rq_in_flight
[BLK_RW_SYNC
] && !cfq_cfqq_sync(cfqq
))
2349 max_dispatch
= max_t(unsigned int, cfqd
->cfq_quantum
/ 2, 1);
2350 if (cfq_class_idle(cfqq
))
2354 * Does this cfqq already have too much IO in flight?
2356 if (cfqq
->dispatched
>= max_dispatch
) {
2357 bool promote_sync
= false;
2359 * idle queue must always only have a single IO in flight
2361 if (cfq_class_idle(cfqq
))
2365 * If there is only one sync queue
2366 * we can ignore async queue here and give the sync
2367 * queue no dispatch limit. The reason is a sync queue can
2368 * preempt async queue, limiting the sync queue doesn't make
2369 * sense. This is useful for aiostress test.
2371 if (cfq_cfqq_sync(cfqq
) && cfqd
->busy_sync_queues
== 1)
2372 promote_sync
= true;
2375 * We have other queues, don't allow more IO from this one
2377 if (cfqd
->busy_queues
> 1 && cfq_slice_used_soon(cfqd
, cfqq
) &&
2382 * Sole queue user, no limit
2384 if (cfqd
->busy_queues
== 1 || promote_sync
)
2388 * Normally we start throttling cfqq when cfq_quantum/2
2389 * requests have been dispatched. But we can drive
2390 * deeper queue depths at the beginning of slice
2391 * subjected to upper limit of cfq_quantum.
2393 max_dispatch
= cfqd
->cfq_quantum
;
2397 * Async queues must wait a bit before being allowed dispatch.
2398 * We also ramp up the dispatch depth gradually for async IO,
2399 * based on the last sync IO we serviced
2401 if (!cfq_cfqq_sync(cfqq
) && cfqd
->cfq_latency
) {
2402 unsigned long last_sync
= jiffies
- cfqd
->last_delayed_sync
;
2405 depth
= last_sync
/ cfqd
->cfq_slice
[1];
2406 if (!depth
&& !cfqq
->dispatched
)
2408 if (depth
< max_dispatch
)
2409 max_dispatch
= depth
;
2413 * If we're below the current max, allow a dispatch
2415 return cfqq
->dispatched
< max_dispatch
;
2419 * Dispatch a request from cfqq, moving them to the request queue
2422 static bool cfq_dispatch_request(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
2426 BUG_ON(RB_EMPTY_ROOT(&cfqq
->sort_list
));
2428 if (!cfq_may_dispatch(cfqd
, cfqq
))
2432 * follow expired path, else get first next available
2434 rq
= cfq_check_fifo(cfqq
);
2439 * insert request into driver dispatch list
2441 cfq_dispatch_insert(cfqd
->queue
, rq
);
2443 if (!cfqd
->active_cic
) {
2444 struct cfq_io_cq
*cic
= RQ_CIC(rq
);
2446 atomic_long_inc(&cic
->icq
.ioc
->refcount
);
2447 cfqd
->active_cic
= cic
;
2454 * Find the cfqq that we need to service and move a request from that to the
2457 static int cfq_dispatch_requests(struct request_queue
*q
, int force
)
2459 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
2460 struct cfq_queue
*cfqq
;
2462 if (!cfqd
->busy_queues
)
2465 if (unlikely(force
))
2466 return cfq_forced_dispatch(cfqd
);
2468 cfqq
= cfq_select_queue(cfqd
);
2473 * Dispatch a request from this cfqq, if it is allowed
2475 if (!cfq_dispatch_request(cfqd
, cfqq
))
2478 cfqq
->slice_dispatch
++;
2479 cfq_clear_cfqq_must_dispatch(cfqq
);
2482 * expire an async queue immediately if it has used up its slice. idle
2483 * queue always expire after 1 dispatch round.
2485 if (cfqd
->busy_queues
> 1 && ((!cfq_cfqq_sync(cfqq
) &&
2486 cfqq
->slice_dispatch
>= cfq_prio_to_maxrq(cfqd
, cfqq
)) ||
2487 cfq_class_idle(cfqq
))) {
2488 cfqq
->slice_end
= jiffies
+ 1;
2489 cfq_slice_expired(cfqd
, 0);
2492 cfq_log_cfqq(cfqd
, cfqq
, "dispatched a request");
2497 * task holds one reference to the queue, dropped when task exits. each rq
2498 * in-flight on this queue also holds a reference, dropped when rq is freed.
2500 * Each cfq queue took a reference on the parent group. Drop it now.
2501 * queue lock must be held here.
2503 static void cfq_put_queue(struct cfq_queue
*cfqq
)
2505 struct cfq_data
*cfqd
= cfqq
->cfqd
;
2506 struct cfq_group
*cfqg
;
2508 BUG_ON(cfqq
->ref
<= 0);
2514 cfq_log_cfqq(cfqd
, cfqq
, "put_queue");
2515 BUG_ON(rb_first(&cfqq
->sort_list
));
2516 BUG_ON(cfqq
->allocated
[READ
] + cfqq
->allocated
[WRITE
]);
2519 if (unlikely(cfqd
->active_queue
== cfqq
)) {
2520 __cfq_slice_expired(cfqd
, cfqq
, 0);
2521 cfq_schedule_dispatch(cfqd
);
2524 BUG_ON(cfq_cfqq_on_rr(cfqq
));
2525 kmem_cache_free(cfq_pool
, cfqq
);
2529 static void cfq_put_cooperator(struct cfq_queue
*cfqq
)
2531 struct cfq_queue
*__cfqq
, *next
;
2534 * If this queue was scheduled to merge with another queue, be
2535 * sure to drop the reference taken on that queue (and others in
2536 * the merge chain). See cfq_setup_merge and cfq_merge_cfqqs.
2538 __cfqq
= cfqq
->new_cfqq
;
2540 if (__cfqq
== cfqq
) {
2541 WARN(1, "cfqq->new_cfqq loop detected\n");
2544 next
= __cfqq
->new_cfqq
;
2545 cfq_put_queue(__cfqq
);
2550 static void cfq_exit_cfqq(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
2552 if (unlikely(cfqq
== cfqd
->active_queue
)) {
2553 __cfq_slice_expired(cfqd
, cfqq
, 0);
2554 cfq_schedule_dispatch(cfqd
);
2557 cfq_put_cooperator(cfqq
);
2559 cfq_put_queue(cfqq
);
2562 static void cfq_init_icq(struct io_cq
*icq
)
2564 struct cfq_io_cq
*cic
= icq_to_cic(icq
);
2566 cic
->ttime
.last_end_request
= jiffies
;
2569 static void cfq_exit_icq(struct io_cq
*icq
)
2571 struct cfq_io_cq
*cic
= icq_to_cic(icq
);
2572 struct cfq_data
*cfqd
= cic_to_cfqd(cic
);
2574 if (cic
->cfqq
[BLK_RW_ASYNC
]) {
2575 cfq_exit_cfqq(cfqd
, cic
->cfqq
[BLK_RW_ASYNC
]);
2576 cic
->cfqq
[BLK_RW_ASYNC
] = NULL
;
2579 if (cic
->cfqq
[BLK_RW_SYNC
]) {
2580 cfq_exit_cfqq(cfqd
, cic
->cfqq
[BLK_RW_SYNC
]);
2581 cic
->cfqq
[BLK_RW_SYNC
] = NULL
;
2585 static void cfq_init_prio_data(struct cfq_queue
*cfqq
, struct cfq_io_cq
*cic
)
2587 struct task_struct
*tsk
= current
;
2590 if (!cfq_cfqq_prio_changed(cfqq
))
2593 ioprio_class
= IOPRIO_PRIO_CLASS(cic
->ioprio
);
2594 switch (ioprio_class
) {
2596 printk(KERN_ERR
"cfq: bad prio %x\n", ioprio_class
);
2597 case IOPRIO_CLASS_NONE
:
2599 * no prio set, inherit CPU scheduling settings
2601 cfqq
->ioprio
= task_nice_ioprio(tsk
);
2602 cfqq
->ioprio_class
= task_nice_ioclass(tsk
);
2604 case IOPRIO_CLASS_RT
:
2605 cfqq
->ioprio
= IOPRIO_PRIO_DATA(cic
->ioprio
);
2606 cfqq
->ioprio_class
= IOPRIO_CLASS_RT
;
2608 case IOPRIO_CLASS_BE
:
2609 cfqq
->ioprio
= IOPRIO_PRIO_DATA(cic
->ioprio
);
2610 cfqq
->ioprio_class
= IOPRIO_CLASS_BE
;
2612 case IOPRIO_CLASS_IDLE
:
2613 cfqq
->ioprio_class
= IOPRIO_CLASS_IDLE
;
2615 cfq_clear_cfqq_idle_window(cfqq
);
2620 * keep track of original prio settings in case we have to temporarily
2621 * elevate the priority of this queue
2623 cfqq
->org_ioprio
= cfqq
->ioprio
;
2624 cfq_clear_cfqq_prio_changed(cfqq
);
2627 static void check_ioprio_changed(struct cfq_io_cq
*cic
, struct bio
*bio
)
2629 int ioprio
= cic
->icq
.ioc
->ioprio
;
2630 struct cfq_data
*cfqd
= cic_to_cfqd(cic
);
2631 struct cfq_queue
*cfqq
;
2634 * Check whether ioprio has changed. The condition may trigger
2635 * spuriously on a newly created cic but there's no harm.
2637 if (unlikely(!cfqd
) || likely(cic
->ioprio
== ioprio
))
2640 cfqq
= cic
->cfqq
[BLK_RW_ASYNC
];
2642 struct cfq_queue
*new_cfqq
;
2643 new_cfqq
= cfq_get_queue(cfqd
, BLK_RW_ASYNC
, cic
, bio
,
2646 cic
->cfqq
[BLK_RW_ASYNC
] = new_cfqq
;
2647 cfq_put_queue(cfqq
);
2651 cfqq
= cic
->cfqq
[BLK_RW_SYNC
];
2653 cfq_mark_cfqq_prio_changed(cfqq
);
2655 cic
->ioprio
= ioprio
;
2658 static void cfq_init_cfqq(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
2659 pid_t pid
, bool is_sync
)
2661 RB_CLEAR_NODE(&cfqq
->rb_node
);
2662 RB_CLEAR_NODE(&cfqq
->p_node
);
2663 INIT_LIST_HEAD(&cfqq
->fifo
);
2668 cfq_mark_cfqq_prio_changed(cfqq
);
2671 if (!cfq_class_idle(cfqq
))
2672 cfq_mark_cfqq_idle_window(cfqq
);
2673 cfq_mark_cfqq_sync(cfqq
);
2678 #ifdef CONFIG_CFQ_GROUP_IOSCHED
2679 static void check_blkcg_changed(struct cfq_io_cq
*cic
, struct bio
*bio
)
2681 struct cfq_data
*cfqd
= cic_to_cfqd(cic
);
2682 struct cfq_queue
*sync_cfqq
;
2686 id
= bio_blkio_cgroup(bio
)->id
;
2690 * Check whether blkcg has changed. The condition may trigger
2691 * spuriously on a newly created cic but there's no harm.
2693 if (unlikely(!cfqd
) || likely(cic
->blkcg_id
== id
))
2696 sync_cfqq
= cic_to_cfqq(cic
, 1);
2699 * Drop reference to sync queue. A new sync queue will be
2700 * assigned in new group upon arrival of a fresh request.
2702 cfq_log_cfqq(cfqd
, sync_cfqq
, "changed cgroup");
2703 cic_set_cfqq(cic
, NULL
, 1);
2704 cfq_put_queue(sync_cfqq
);
2710 static inline void check_blkcg_changed(struct cfq_io_cq
*cic
, struct bio
*bio
) { }
2711 #endif /* CONFIG_CFQ_GROUP_IOSCHED */
2713 static struct cfq_queue
*
2714 cfq_find_alloc_queue(struct cfq_data
*cfqd
, bool is_sync
, struct cfq_io_cq
*cic
,
2715 struct bio
*bio
, gfp_t gfp_mask
)
2717 struct blkio_cgroup
*blkcg
;
2718 struct cfq_queue
*cfqq
, *new_cfqq
= NULL
;
2719 struct cfq_group
*cfqg
;
2724 blkcg
= bio_blkio_cgroup(bio
);
2725 cfqg
= cfq_lookup_create_cfqg(cfqd
, blkcg
);
2726 cfqq
= cic_to_cfqq(cic
, is_sync
);
2729 * Always try a new alloc if we fell back to the OOM cfqq
2730 * originally, since it should just be a temporary situation.
2732 if (!cfqq
|| cfqq
== &cfqd
->oom_cfqq
) {
2737 } else if (gfp_mask
& __GFP_WAIT
) {
2739 spin_unlock_irq(cfqd
->queue
->queue_lock
);
2740 new_cfqq
= kmem_cache_alloc_node(cfq_pool
,
2741 gfp_mask
| __GFP_ZERO
,
2743 spin_lock_irq(cfqd
->queue
->queue_lock
);
2747 cfqq
= kmem_cache_alloc_node(cfq_pool
,
2748 gfp_mask
| __GFP_ZERO
,
2753 cfq_init_cfqq(cfqd
, cfqq
, current
->pid
, is_sync
);
2754 cfq_init_prio_data(cfqq
, cic
);
2755 cfq_link_cfqq_cfqg(cfqq
, cfqg
);
2756 cfq_log_cfqq(cfqd
, cfqq
, "alloced");
2758 cfqq
= &cfqd
->oom_cfqq
;
2762 kmem_cache_free(cfq_pool
, new_cfqq
);
2768 static struct cfq_queue
**
2769 cfq_async_queue_prio(struct cfq_data
*cfqd
, int ioprio_class
, int ioprio
)
2771 switch (ioprio_class
) {
2772 case IOPRIO_CLASS_RT
:
2773 return &cfqd
->async_cfqq
[0][ioprio
];
2774 case IOPRIO_CLASS_NONE
:
2775 ioprio
= IOPRIO_NORM
;
2777 case IOPRIO_CLASS_BE
:
2778 return &cfqd
->async_cfqq
[1][ioprio
];
2779 case IOPRIO_CLASS_IDLE
:
2780 return &cfqd
->async_idle_cfqq
;
2786 static struct cfq_queue
*
2787 cfq_get_queue(struct cfq_data
*cfqd
, bool is_sync
, struct cfq_io_cq
*cic
,
2788 struct bio
*bio
, gfp_t gfp_mask
)
2790 const int ioprio_class
= IOPRIO_PRIO_CLASS(cic
->ioprio
);
2791 const int ioprio
= IOPRIO_PRIO_DATA(cic
->ioprio
);
2792 struct cfq_queue
**async_cfqq
= NULL
;
2793 struct cfq_queue
*cfqq
= NULL
;
2796 async_cfqq
= cfq_async_queue_prio(cfqd
, ioprio_class
, ioprio
);
2801 cfqq
= cfq_find_alloc_queue(cfqd
, is_sync
, cic
, bio
, gfp_mask
);
2804 * pin the queue now that it's allocated, scheduler exit will prune it
2806 if (!is_sync
&& !(*async_cfqq
)) {
2816 __cfq_update_io_thinktime(struct cfq_ttime
*ttime
, unsigned long slice_idle
)
2818 unsigned long elapsed
= jiffies
- ttime
->last_end_request
;
2819 elapsed
= min(elapsed
, 2UL * slice_idle
);
2821 ttime
->ttime_samples
= (7*ttime
->ttime_samples
+ 256) / 8;
2822 ttime
->ttime_total
= (7*ttime
->ttime_total
+ 256*elapsed
) / 8;
2823 ttime
->ttime_mean
= (ttime
->ttime_total
+ 128) / ttime
->ttime_samples
;
2827 cfq_update_io_thinktime(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
2828 struct cfq_io_cq
*cic
)
2830 if (cfq_cfqq_sync(cfqq
)) {
2831 __cfq_update_io_thinktime(&cic
->ttime
, cfqd
->cfq_slice_idle
);
2832 __cfq_update_io_thinktime(&cfqq
->service_tree
->ttime
,
2833 cfqd
->cfq_slice_idle
);
2835 #ifdef CONFIG_CFQ_GROUP_IOSCHED
2836 __cfq_update_io_thinktime(&cfqq
->cfqg
->ttime
, cfqd
->cfq_group_idle
);
2841 cfq_update_io_seektime(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
2845 sector_t n_sec
= blk_rq_sectors(rq
);
2846 if (cfqq
->last_request_pos
) {
2847 if (cfqq
->last_request_pos
< blk_rq_pos(rq
))
2848 sdist
= blk_rq_pos(rq
) - cfqq
->last_request_pos
;
2850 sdist
= cfqq
->last_request_pos
- blk_rq_pos(rq
);
2853 cfqq
->seek_history
<<= 1;
2854 if (blk_queue_nonrot(cfqd
->queue
))
2855 cfqq
->seek_history
|= (n_sec
< CFQQ_SECT_THR_NONROT
);
2857 cfqq
->seek_history
|= (sdist
> CFQQ_SEEK_THR
);
2861 * Disable idle window if the process thinks too long or seeks so much that
2865 cfq_update_idle_window(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
2866 struct cfq_io_cq
*cic
)
2868 int old_idle
, enable_idle
;
2871 * Don't idle for async or idle io prio class
2873 if (!cfq_cfqq_sync(cfqq
) || cfq_class_idle(cfqq
))
2876 enable_idle
= old_idle
= cfq_cfqq_idle_window(cfqq
);
2878 if (cfqq
->queued
[0] + cfqq
->queued
[1] >= 4)
2879 cfq_mark_cfqq_deep(cfqq
);
2881 if (cfqq
->next_rq
&& (cfqq
->next_rq
->cmd_flags
& REQ_NOIDLE
))
2883 else if (!atomic_read(&cic
->icq
.ioc
->active_ref
) ||
2884 !cfqd
->cfq_slice_idle
||
2885 (!cfq_cfqq_deep(cfqq
) && CFQQ_SEEKY(cfqq
)))
2887 else if (sample_valid(cic
->ttime
.ttime_samples
)) {
2888 if (cic
->ttime
.ttime_mean
> cfqd
->cfq_slice_idle
)
2894 if (old_idle
!= enable_idle
) {
2895 cfq_log_cfqq(cfqd
, cfqq
, "idle=%d", enable_idle
);
2897 cfq_mark_cfqq_idle_window(cfqq
);
2899 cfq_clear_cfqq_idle_window(cfqq
);
2904 * Check if new_cfqq should preempt the currently active queue. Return 0 for
2905 * no or if we aren't sure, a 1 will cause a preempt.
2908 cfq_should_preempt(struct cfq_data
*cfqd
, struct cfq_queue
*new_cfqq
,
2911 struct cfq_queue
*cfqq
;
2913 cfqq
= cfqd
->active_queue
;
2917 if (cfq_class_idle(new_cfqq
))
2920 if (cfq_class_idle(cfqq
))
2924 * Don't allow a non-RT request to preempt an ongoing RT cfqq timeslice.
2926 if (cfq_class_rt(cfqq
) && !cfq_class_rt(new_cfqq
))
2930 * if the new request is sync, but the currently running queue is
2931 * not, let the sync request have priority.
2933 if (rq_is_sync(rq
) && !cfq_cfqq_sync(cfqq
))
2936 if (new_cfqq
->cfqg
!= cfqq
->cfqg
)
2939 if (cfq_slice_used(cfqq
))
2942 /* Allow preemption only if we are idling on sync-noidle tree */
2943 if (cfqd
->serving_type
== SYNC_NOIDLE_WORKLOAD
&&
2944 cfqq_type(new_cfqq
) == SYNC_NOIDLE_WORKLOAD
&&
2945 new_cfqq
->service_tree
->count
== 2 &&
2946 RB_EMPTY_ROOT(&cfqq
->sort_list
))
2950 * So both queues are sync. Let the new request get disk time if
2951 * it's a metadata request and the current queue is doing regular IO.
2953 if ((rq
->cmd_flags
& REQ_PRIO
) && !cfqq
->prio_pending
)
2957 * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
2959 if (cfq_class_rt(new_cfqq
) && !cfq_class_rt(cfqq
))
2962 /* An idle queue should not be idle now for some reason */
2963 if (RB_EMPTY_ROOT(&cfqq
->sort_list
) && !cfq_should_idle(cfqd
, cfqq
))
2966 if (!cfqd
->active_cic
|| !cfq_cfqq_wait_request(cfqq
))
2970 * if this request is as-good as one we would expect from the
2971 * current cfqq, let it preempt
2973 if (cfq_rq_close(cfqd
, cfqq
, rq
))
2980 * cfqq preempts the active queue. if we allowed preempt with no slice left,
2981 * let it have half of its nominal slice.
2983 static void cfq_preempt_queue(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
2985 enum wl_type_t old_type
= cfqq_type(cfqd
->active_queue
);
2987 cfq_log_cfqq(cfqd
, cfqq
, "preempt");
2988 cfq_slice_expired(cfqd
, 1);
2991 * workload type is changed, don't save slice, otherwise preempt
2994 if (old_type
!= cfqq_type(cfqq
))
2995 cfqq
->cfqg
->saved_workload_slice
= 0;
2998 * Put the new queue at the front of the of the current list,
2999 * so we know that it will be selected next.
3001 BUG_ON(!cfq_cfqq_on_rr(cfqq
));
3003 cfq_service_tree_add(cfqd
, cfqq
, 1);
3005 cfqq
->slice_end
= 0;
3006 cfq_mark_cfqq_slice_new(cfqq
);
3010 * Called when a new fs request (rq) is added (to cfqq). Check if there's
3011 * something we should do about it
3014 cfq_rq_enqueued(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
3017 struct cfq_io_cq
*cic
= RQ_CIC(rq
);
3020 if (rq
->cmd_flags
& REQ_PRIO
)
3021 cfqq
->prio_pending
++;
3023 cfq_update_io_thinktime(cfqd
, cfqq
, cic
);
3024 cfq_update_io_seektime(cfqd
, cfqq
, rq
);
3025 cfq_update_idle_window(cfqd
, cfqq
, cic
);
3027 cfqq
->last_request_pos
= blk_rq_pos(rq
) + blk_rq_sectors(rq
);
3029 if (cfqq
== cfqd
->active_queue
) {
3031 * Remember that we saw a request from this process, but
3032 * don't start queuing just yet. Otherwise we risk seeing lots
3033 * of tiny requests, because we disrupt the normal plugging
3034 * and merging. If the request is already larger than a single
3035 * page, let it rip immediately. For that case we assume that
3036 * merging is already done. Ditto for a busy system that
3037 * has other work pending, don't risk delaying until the
3038 * idle timer unplug to continue working.
3040 if (cfq_cfqq_wait_request(cfqq
)) {
3041 if (blk_rq_bytes(rq
) > PAGE_CACHE_SIZE
||
3042 cfqd
->busy_queues
> 1) {
3043 cfq_del_timer(cfqd
, cfqq
);
3044 cfq_clear_cfqq_wait_request(cfqq
);
3045 __blk_run_queue(cfqd
->queue
);
3047 cfq_blkiocg_update_idle_time_stats(
3048 cfqg_to_blkg(cfqq
->cfqg
),
3050 cfq_mark_cfqq_must_dispatch(cfqq
);
3053 } else if (cfq_should_preempt(cfqd
, cfqq
, rq
)) {
3055 * not the active queue - expire current slice if it is
3056 * idle and has expired it's mean thinktime or this new queue
3057 * has some old slice time left and is of higher priority or
3058 * this new queue is RT and the current one is BE
3060 cfq_preempt_queue(cfqd
, cfqq
);
3061 __blk_run_queue(cfqd
->queue
);
3065 static void cfq_insert_request(struct request_queue
*q
, struct request
*rq
)
3067 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
3068 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
3070 cfq_log_cfqq(cfqd
, cfqq
, "insert_request");
3071 cfq_init_prio_data(cfqq
, RQ_CIC(rq
));
3073 rq_set_fifo_time(rq
, jiffies
+ cfqd
->cfq_fifo_expire
[rq_is_sync(rq
)]);
3074 list_add_tail(&rq
->queuelist
, &cfqq
->fifo
);
3076 cfq_blkiocg_update_io_add_stats(cfqg_to_blkg(RQ_CFQG(rq
)),
3078 cfqg_to_blkg(cfqd
->serving_group
),
3079 rq_data_dir(rq
), rq_is_sync(rq
));
3080 cfq_rq_enqueued(cfqd
, cfqq
, rq
);
3084 * Update hw_tag based on peak queue depth over 50 samples under
3087 static void cfq_update_hw_tag(struct cfq_data
*cfqd
)
3089 struct cfq_queue
*cfqq
= cfqd
->active_queue
;
3091 if (cfqd
->rq_in_driver
> cfqd
->hw_tag_est_depth
)
3092 cfqd
->hw_tag_est_depth
= cfqd
->rq_in_driver
;
3094 if (cfqd
->hw_tag
== 1)
3097 if (cfqd
->rq_queued
<= CFQ_HW_QUEUE_MIN
&&
3098 cfqd
->rq_in_driver
<= CFQ_HW_QUEUE_MIN
)
3102 * If active queue hasn't enough requests and can idle, cfq might not
3103 * dispatch sufficient requests to hardware. Don't zero hw_tag in this
3106 if (cfqq
&& cfq_cfqq_idle_window(cfqq
) &&
3107 cfqq
->dispatched
+ cfqq
->queued
[0] + cfqq
->queued
[1] <
3108 CFQ_HW_QUEUE_MIN
&& cfqd
->rq_in_driver
< CFQ_HW_QUEUE_MIN
)
3111 if (cfqd
->hw_tag_samples
++ < 50)
3114 if (cfqd
->hw_tag_est_depth
>= CFQ_HW_QUEUE_MIN
)
3120 static bool cfq_should_wait_busy(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
3122 struct cfq_io_cq
*cic
= cfqd
->active_cic
;
3124 /* If the queue already has requests, don't wait */
3125 if (!RB_EMPTY_ROOT(&cfqq
->sort_list
))
3128 /* If there are other queues in the group, don't wait */
3129 if (cfqq
->cfqg
->nr_cfqq
> 1)
3132 /* the only queue in the group, but think time is big */
3133 if (cfq_io_thinktime_big(cfqd
, &cfqq
->cfqg
->ttime
, true))
3136 if (cfq_slice_used(cfqq
))
3139 /* if slice left is less than think time, wait busy */
3140 if (cic
&& sample_valid(cic
->ttime
.ttime_samples
)
3141 && (cfqq
->slice_end
- jiffies
< cic
->ttime
.ttime_mean
))
3145 * If think times is less than a jiffy than ttime_mean=0 and above
3146 * will not be true. It might happen that slice has not expired yet
3147 * but will expire soon (4-5 ns) during select_queue(). To cover the
3148 * case where think time is less than a jiffy, mark the queue wait
3149 * busy if only 1 jiffy is left in the slice.
3151 if (cfqq
->slice_end
- jiffies
== 1)
3157 static void cfq_completed_request(struct request_queue
*q
, struct request
*rq
)
3159 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
3160 struct cfq_data
*cfqd
= cfqq
->cfqd
;
3161 const int sync
= rq_is_sync(rq
);
3165 cfq_log_cfqq(cfqd
, cfqq
, "complete rqnoidle %d",
3166 !!(rq
->cmd_flags
& REQ_NOIDLE
));
3168 cfq_update_hw_tag(cfqd
);
3170 WARN_ON(!cfqd
->rq_in_driver
);
3171 WARN_ON(!cfqq
->dispatched
);
3172 cfqd
->rq_in_driver
--;
3174 (RQ_CFQG(rq
))->dispatched
--;
3175 cfq_blkiocg_update_completion_stats(cfqg_to_blkg(cfqq
->cfqg
),
3176 &blkio_policy_cfq
, rq_start_time_ns(rq
),
3177 rq_io_start_time_ns(rq
), rq_data_dir(rq
),
3180 cfqd
->rq_in_flight
[cfq_cfqq_sync(cfqq
)]--;
3183 struct cfq_rb_root
*service_tree
;
3185 RQ_CIC(rq
)->ttime
.last_end_request
= now
;
3187 if (cfq_cfqq_on_rr(cfqq
))
3188 service_tree
= cfqq
->service_tree
;
3190 service_tree
= service_tree_for(cfqq
->cfqg
,
3191 cfqq_prio(cfqq
), cfqq_type(cfqq
));
3192 service_tree
->ttime
.last_end_request
= now
;
3193 if (!time_after(rq
->start_time
+ cfqd
->cfq_fifo_expire
[1], now
))
3194 cfqd
->last_delayed_sync
= now
;
3197 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3198 cfqq
->cfqg
->ttime
.last_end_request
= now
;
3202 * If this is the active queue, check if it needs to be expired,
3203 * or if we want to idle in case it has no pending requests.
3205 if (cfqd
->active_queue
== cfqq
) {
3206 const bool cfqq_empty
= RB_EMPTY_ROOT(&cfqq
->sort_list
);
3208 if (cfq_cfqq_slice_new(cfqq
)) {
3209 cfq_set_prio_slice(cfqd
, cfqq
);
3210 cfq_clear_cfqq_slice_new(cfqq
);
3214 * Should we wait for next request to come in before we expire
3217 if (cfq_should_wait_busy(cfqd
, cfqq
)) {
3218 unsigned long extend_sl
= cfqd
->cfq_slice_idle
;
3219 if (!cfqd
->cfq_slice_idle
)
3220 extend_sl
= cfqd
->cfq_group_idle
;
3221 cfqq
->slice_end
= jiffies
+ extend_sl
;
3222 cfq_mark_cfqq_wait_busy(cfqq
);
3223 cfq_log_cfqq(cfqd
, cfqq
, "will busy wait");
3227 * Idling is not enabled on:
3229 * - idle-priority queues
3231 * - queues with still some requests queued
3232 * - when there is a close cooperator
3234 if (cfq_slice_used(cfqq
) || cfq_class_idle(cfqq
))
3235 cfq_slice_expired(cfqd
, 1);
3236 else if (sync
&& cfqq_empty
&&
3237 !cfq_close_cooperator(cfqd
, cfqq
)) {
3238 cfq_arm_slice_timer(cfqd
);
3242 if (!cfqd
->rq_in_driver
)
3243 cfq_schedule_dispatch(cfqd
);
3246 static inline int __cfq_may_queue(struct cfq_queue
*cfqq
)
3248 if (cfq_cfqq_wait_request(cfqq
) && !cfq_cfqq_must_alloc_slice(cfqq
)) {
3249 cfq_mark_cfqq_must_alloc_slice(cfqq
);
3250 return ELV_MQUEUE_MUST
;
3253 return ELV_MQUEUE_MAY
;
3256 static int cfq_may_queue(struct request_queue
*q
, int rw
)
3258 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
3259 struct task_struct
*tsk
= current
;
3260 struct cfq_io_cq
*cic
;
3261 struct cfq_queue
*cfqq
;
3264 * don't force setup of a queue from here, as a call to may_queue
3265 * does not necessarily imply that a request actually will be queued.
3266 * so just lookup a possibly existing queue, or return 'may queue'
3269 cic
= cfq_cic_lookup(cfqd
, tsk
->io_context
);
3271 return ELV_MQUEUE_MAY
;
3273 cfqq
= cic_to_cfqq(cic
, rw_is_sync(rw
));
3275 cfq_init_prio_data(cfqq
, cic
);
3277 return __cfq_may_queue(cfqq
);
3280 return ELV_MQUEUE_MAY
;
3284 * queue lock held here
3286 static void cfq_put_request(struct request
*rq
)
3288 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
3291 const int rw
= rq_data_dir(rq
);
3293 BUG_ON(!cfqq
->allocated
[rw
]);
3294 cfqq
->allocated
[rw
]--;
3296 /* Put down rq reference on cfqg */
3297 cfqg_put(RQ_CFQG(rq
));
3298 rq
->elv
.priv
[0] = NULL
;
3299 rq
->elv
.priv
[1] = NULL
;
3301 cfq_put_queue(cfqq
);
3305 static struct cfq_queue
*
3306 cfq_merge_cfqqs(struct cfq_data
*cfqd
, struct cfq_io_cq
*cic
,
3307 struct cfq_queue
*cfqq
)
3309 cfq_log_cfqq(cfqd
, cfqq
, "merging with queue %p", cfqq
->new_cfqq
);
3310 cic_set_cfqq(cic
, cfqq
->new_cfqq
, 1);
3311 cfq_mark_cfqq_coop(cfqq
->new_cfqq
);
3312 cfq_put_queue(cfqq
);
3313 return cic_to_cfqq(cic
, 1);
3317 * Returns NULL if a new cfqq should be allocated, or the old cfqq if this
3318 * was the last process referring to said cfqq.
3320 static struct cfq_queue
*
3321 split_cfqq(struct cfq_io_cq
*cic
, struct cfq_queue
*cfqq
)
3323 if (cfqq_process_refs(cfqq
) == 1) {
3324 cfqq
->pid
= current
->pid
;
3325 cfq_clear_cfqq_coop(cfqq
);
3326 cfq_clear_cfqq_split_coop(cfqq
);
3330 cic_set_cfqq(cic
, NULL
, 1);
3332 cfq_put_cooperator(cfqq
);
3334 cfq_put_queue(cfqq
);
3338 * Allocate cfq data structures associated with this request.
3341 cfq_set_request(struct request_queue
*q
, struct request
*rq
, struct bio
*bio
,
3344 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
3345 struct cfq_io_cq
*cic
= icq_to_cic(rq
->elv
.icq
);
3346 const int rw
= rq_data_dir(rq
);
3347 const bool is_sync
= rq_is_sync(rq
);
3348 struct cfq_queue
*cfqq
;
3350 might_sleep_if(gfp_mask
& __GFP_WAIT
);
3352 spin_lock_irq(q
->queue_lock
);
3354 check_ioprio_changed(cic
, bio
);
3355 check_blkcg_changed(cic
, bio
);
3357 cfqq
= cic_to_cfqq(cic
, is_sync
);
3358 if (!cfqq
|| cfqq
== &cfqd
->oom_cfqq
) {
3359 cfqq
= cfq_get_queue(cfqd
, is_sync
, cic
, bio
, gfp_mask
);
3360 cic_set_cfqq(cic
, cfqq
, is_sync
);
3363 * If the queue was seeky for too long, break it apart.
3365 if (cfq_cfqq_coop(cfqq
) && cfq_cfqq_split_coop(cfqq
)) {
3366 cfq_log_cfqq(cfqd
, cfqq
, "breaking apart cfqq");
3367 cfqq
= split_cfqq(cic
, cfqq
);
3373 * Check to see if this queue is scheduled to merge with
3374 * another, closely cooperating queue. The merging of
3375 * queues happens here as it must be done in process context.
3376 * The reference on new_cfqq was taken in merge_cfqqs.
3379 cfqq
= cfq_merge_cfqqs(cfqd
, cic
, cfqq
);
3382 cfqq
->allocated
[rw
]++;
3385 cfqg_get(cfqq
->cfqg
);
3386 rq
->elv
.priv
[0] = cfqq
;
3387 rq
->elv
.priv
[1] = cfqq
->cfqg
;
3388 spin_unlock_irq(q
->queue_lock
);
3392 static void cfq_kick_queue(struct work_struct
*work
)
3394 struct cfq_data
*cfqd
=
3395 container_of(work
, struct cfq_data
, unplug_work
);
3396 struct request_queue
*q
= cfqd
->queue
;
3398 spin_lock_irq(q
->queue_lock
);
3399 __blk_run_queue(cfqd
->queue
);
3400 spin_unlock_irq(q
->queue_lock
);
3404 * Timer running if the active_queue is currently idling inside its time slice
3406 static void cfq_idle_slice_timer(unsigned long data
)
3408 struct cfq_data
*cfqd
= (struct cfq_data
*) data
;
3409 struct cfq_queue
*cfqq
;
3410 unsigned long flags
;
3413 cfq_log(cfqd
, "idle timer fired");
3415 spin_lock_irqsave(cfqd
->queue
->queue_lock
, flags
);
3417 cfqq
= cfqd
->active_queue
;
3422 * We saw a request before the queue expired, let it through
3424 if (cfq_cfqq_must_dispatch(cfqq
))
3430 if (cfq_slice_used(cfqq
))
3434 * only expire and reinvoke request handler, if there are
3435 * other queues with pending requests
3437 if (!cfqd
->busy_queues
)
3441 * not expired and it has a request pending, let it dispatch
3443 if (!RB_EMPTY_ROOT(&cfqq
->sort_list
))
3447 * Queue depth flag is reset only when the idle didn't succeed
3449 cfq_clear_cfqq_deep(cfqq
);
3452 cfq_slice_expired(cfqd
, timed_out
);
3454 cfq_schedule_dispatch(cfqd
);
3456 spin_unlock_irqrestore(cfqd
->queue
->queue_lock
, flags
);
3459 static void cfq_shutdown_timer_wq(struct cfq_data
*cfqd
)
3461 del_timer_sync(&cfqd
->idle_slice_timer
);
3462 cancel_work_sync(&cfqd
->unplug_work
);
3465 static void cfq_put_async_queues(struct cfq_data
*cfqd
)
3469 for (i
= 0; i
< IOPRIO_BE_NR
; i
++) {
3470 if (cfqd
->async_cfqq
[0][i
])
3471 cfq_put_queue(cfqd
->async_cfqq
[0][i
]);
3472 if (cfqd
->async_cfqq
[1][i
])
3473 cfq_put_queue(cfqd
->async_cfqq
[1][i
]);
3476 if (cfqd
->async_idle_cfqq
)
3477 cfq_put_queue(cfqd
->async_idle_cfqq
);
3480 static void cfq_exit_queue(struct elevator_queue
*e
)
3482 struct cfq_data
*cfqd
= e
->elevator_data
;
3483 struct request_queue
*q
= cfqd
->queue
;
3485 cfq_shutdown_timer_wq(cfqd
);
3487 spin_lock_irq(q
->queue_lock
);
3489 if (cfqd
->active_queue
)
3490 __cfq_slice_expired(cfqd
, cfqd
->active_queue
, 0);
3492 cfq_put_async_queues(cfqd
);
3494 spin_unlock_irq(q
->queue_lock
);
3496 cfq_shutdown_timer_wq(cfqd
);
3498 #ifndef CONFIG_CFQ_GROUP_IOSCHED
3499 kfree(cfqd
->root_group
);
3501 update_root_blkg_pd(q
, BLKIO_POLICY_PROP
);
3505 static int cfq_init_queue(struct request_queue
*q
)
3507 struct cfq_data
*cfqd
;
3508 struct blkio_group
*blkg __maybe_unused
;
3511 cfqd
= kmalloc_node(sizeof(*cfqd
), GFP_KERNEL
| __GFP_ZERO
, q
->node
);
3516 q
->elevator
->elevator_data
= cfqd
;
3518 /* Init root service tree */
3519 cfqd
->grp_service_tree
= CFQ_RB_ROOT
;
3521 /* Init root group and prefer root group over other groups by default */
3522 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3524 spin_lock_irq(q
->queue_lock
);
3526 blkg
= blkg_lookup_create(&blkio_root_cgroup
, q
, true);
3528 cfqd
->root_group
= blkg_to_cfqg(blkg
);
3530 spin_unlock_irq(q
->queue_lock
);
3533 cfqd
->root_group
= kzalloc_node(sizeof(*cfqd
->root_group
),
3534 GFP_KERNEL
, cfqd
->queue
->node
);
3535 if (cfqd
->root_group
)
3536 cfq_init_cfqg_base(cfqd
->root_group
);
3538 if (!cfqd
->root_group
) {
3543 cfqd
->root_group
->weight
= 2*BLKIO_WEIGHT_DEFAULT
;
3546 * Not strictly needed (since RB_ROOT just clears the node and we
3547 * zeroed cfqd on alloc), but better be safe in case someone decides
3548 * to add magic to the rb code
3550 for (i
= 0; i
< CFQ_PRIO_LISTS
; i
++)
3551 cfqd
->prio_trees
[i
] = RB_ROOT
;
3554 * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
3555 * Grab a permanent reference to it, so that the normal code flow
3556 * will not attempt to free it. oom_cfqq is linked to root_group
3557 * but shouldn't hold a reference as it'll never be unlinked. Lose
3558 * the reference from linking right away.
3560 cfq_init_cfqq(cfqd
, &cfqd
->oom_cfqq
, 1, 0);
3561 cfqd
->oom_cfqq
.ref
++;
3563 spin_lock_irq(q
->queue_lock
);
3564 cfq_link_cfqq_cfqg(&cfqd
->oom_cfqq
, cfqd
->root_group
);
3565 cfqg_put(cfqd
->root_group
);
3566 spin_unlock_irq(q
->queue_lock
);
3568 init_timer(&cfqd
->idle_slice_timer
);
3569 cfqd
->idle_slice_timer
.function
= cfq_idle_slice_timer
;
3570 cfqd
->idle_slice_timer
.data
= (unsigned long) cfqd
;
3572 INIT_WORK(&cfqd
->unplug_work
, cfq_kick_queue
);
3574 cfqd
->cfq_quantum
= cfq_quantum
;
3575 cfqd
->cfq_fifo_expire
[0] = cfq_fifo_expire
[0];
3576 cfqd
->cfq_fifo_expire
[1] = cfq_fifo_expire
[1];
3577 cfqd
->cfq_back_max
= cfq_back_max
;
3578 cfqd
->cfq_back_penalty
= cfq_back_penalty
;
3579 cfqd
->cfq_slice
[0] = cfq_slice_async
;
3580 cfqd
->cfq_slice
[1] = cfq_slice_sync
;
3581 cfqd
->cfq_slice_async_rq
= cfq_slice_async_rq
;
3582 cfqd
->cfq_slice_idle
= cfq_slice_idle
;
3583 cfqd
->cfq_group_idle
= cfq_group_idle
;
3584 cfqd
->cfq_latency
= 1;
3587 * we optimistically start assuming sync ops weren't delayed in last
3588 * second, in order to have larger depth for async operations.
3590 cfqd
->last_delayed_sync
= jiffies
- HZ
;
3595 * sysfs parts below -->
3598 cfq_var_show(unsigned int var
, char *page
)
3600 return sprintf(page
, "%d\n", var
);
3604 cfq_var_store(unsigned int *var
, const char *page
, size_t count
)
3606 char *p
= (char *) page
;
3608 *var
= simple_strtoul(p
, &p
, 10);
3612 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
3613 static ssize_t __FUNC(struct elevator_queue *e, char *page) \
3615 struct cfq_data *cfqd = e->elevator_data; \
3616 unsigned int __data = __VAR; \
3618 __data = jiffies_to_msecs(__data); \
3619 return cfq_var_show(__data, (page)); \
3621 SHOW_FUNCTION(cfq_quantum_show
, cfqd
->cfq_quantum
, 0);
3622 SHOW_FUNCTION(cfq_fifo_expire_sync_show
, cfqd
->cfq_fifo_expire
[1], 1);
3623 SHOW_FUNCTION(cfq_fifo_expire_async_show
, cfqd
->cfq_fifo_expire
[0], 1);
3624 SHOW_FUNCTION(cfq_back_seek_max_show
, cfqd
->cfq_back_max
, 0);
3625 SHOW_FUNCTION(cfq_back_seek_penalty_show
, cfqd
->cfq_back_penalty
, 0);
3626 SHOW_FUNCTION(cfq_slice_idle_show
, cfqd
->cfq_slice_idle
, 1);
3627 SHOW_FUNCTION(cfq_group_idle_show
, cfqd
->cfq_group_idle
, 1);
3628 SHOW_FUNCTION(cfq_slice_sync_show
, cfqd
->cfq_slice
[1], 1);
3629 SHOW_FUNCTION(cfq_slice_async_show
, cfqd
->cfq_slice
[0], 1);
3630 SHOW_FUNCTION(cfq_slice_async_rq_show
, cfqd
->cfq_slice_async_rq
, 0);
3631 SHOW_FUNCTION(cfq_low_latency_show
, cfqd
->cfq_latency
, 0);
3632 #undef SHOW_FUNCTION
3634 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
3635 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
3637 struct cfq_data *cfqd = e->elevator_data; \
3638 unsigned int __data; \
3639 int ret = cfq_var_store(&__data, (page), count); \
3640 if (__data < (MIN)) \
3642 else if (__data > (MAX)) \
3645 *(__PTR) = msecs_to_jiffies(__data); \
3647 *(__PTR) = __data; \
3650 STORE_FUNCTION(cfq_quantum_store
, &cfqd
->cfq_quantum
, 1, UINT_MAX
, 0);
3651 STORE_FUNCTION(cfq_fifo_expire_sync_store
, &cfqd
->cfq_fifo_expire
[1], 1,
3653 STORE_FUNCTION(cfq_fifo_expire_async_store
, &cfqd
->cfq_fifo_expire
[0], 1,
3655 STORE_FUNCTION(cfq_back_seek_max_store
, &cfqd
->cfq_back_max
, 0, UINT_MAX
, 0);
3656 STORE_FUNCTION(cfq_back_seek_penalty_store
, &cfqd
->cfq_back_penalty
, 1,
3658 STORE_FUNCTION(cfq_slice_idle_store
, &cfqd
->cfq_slice_idle
, 0, UINT_MAX
, 1);
3659 STORE_FUNCTION(cfq_group_idle_store
, &cfqd
->cfq_group_idle
, 0, UINT_MAX
, 1);
3660 STORE_FUNCTION(cfq_slice_sync_store
, &cfqd
->cfq_slice
[1], 1, UINT_MAX
, 1);
3661 STORE_FUNCTION(cfq_slice_async_store
, &cfqd
->cfq_slice
[0], 1, UINT_MAX
, 1);
3662 STORE_FUNCTION(cfq_slice_async_rq_store
, &cfqd
->cfq_slice_async_rq
, 1,
3664 STORE_FUNCTION(cfq_low_latency_store
, &cfqd
->cfq_latency
, 0, 1, 0);
3665 #undef STORE_FUNCTION
3667 #define CFQ_ATTR(name) \
3668 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
3670 static struct elv_fs_entry cfq_attrs
[] = {
3672 CFQ_ATTR(fifo_expire_sync
),
3673 CFQ_ATTR(fifo_expire_async
),
3674 CFQ_ATTR(back_seek_max
),
3675 CFQ_ATTR(back_seek_penalty
),
3676 CFQ_ATTR(slice_sync
),
3677 CFQ_ATTR(slice_async
),
3678 CFQ_ATTR(slice_async_rq
),
3679 CFQ_ATTR(slice_idle
),
3680 CFQ_ATTR(group_idle
),
3681 CFQ_ATTR(low_latency
),
3685 static struct elevator_type iosched_cfq
= {
3687 .elevator_merge_fn
= cfq_merge
,
3688 .elevator_merged_fn
= cfq_merged_request
,
3689 .elevator_merge_req_fn
= cfq_merged_requests
,
3690 .elevator_allow_merge_fn
= cfq_allow_merge
,
3691 .elevator_bio_merged_fn
= cfq_bio_merged
,
3692 .elevator_dispatch_fn
= cfq_dispatch_requests
,
3693 .elevator_add_req_fn
= cfq_insert_request
,
3694 .elevator_activate_req_fn
= cfq_activate_request
,
3695 .elevator_deactivate_req_fn
= cfq_deactivate_request
,
3696 .elevator_completed_req_fn
= cfq_completed_request
,
3697 .elevator_former_req_fn
= elv_rb_former_request
,
3698 .elevator_latter_req_fn
= elv_rb_latter_request
,
3699 .elevator_init_icq_fn
= cfq_init_icq
,
3700 .elevator_exit_icq_fn
= cfq_exit_icq
,
3701 .elevator_set_req_fn
= cfq_set_request
,
3702 .elevator_put_req_fn
= cfq_put_request
,
3703 .elevator_may_queue_fn
= cfq_may_queue
,
3704 .elevator_init_fn
= cfq_init_queue
,
3705 .elevator_exit_fn
= cfq_exit_queue
,
3707 .icq_size
= sizeof(struct cfq_io_cq
),
3708 .icq_align
= __alignof__(struct cfq_io_cq
),
3709 .elevator_attrs
= cfq_attrs
,
3710 .elevator_name
= "cfq",
3711 .elevator_owner
= THIS_MODULE
,
3714 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3715 static struct blkio_policy_type blkio_policy_cfq
= {
3717 .blkio_init_group_fn
= cfq_init_blkio_group
,
3718 .blkio_update_group_weight_fn
= cfq_update_blkio_group_weight
,
3720 .plid
= BLKIO_POLICY_PROP
,
3721 .pdata_size
= sizeof(struct cfq_group
),
3725 static int __init
cfq_init(void)
3730 * could be 0 on HZ < 1000 setups
3732 if (!cfq_slice_async
)
3733 cfq_slice_async
= 1;
3734 if (!cfq_slice_idle
)
3737 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3738 if (!cfq_group_idle
)
3743 cfq_pool
= KMEM_CACHE(cfq_queue
, 0);
3747 ret
= elv_register(&iosched_cfq
);
3749 kmem_cache_destroy(cfq_pool
);
3753 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3754 blkio_policy_register(&blkio_policy_cfq
);
3759 static void __exit
cfq_exit(void)
3761 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3762 blkio_policy_unregister(&blkio_policy_cfq
);
3764 elv_unregister(&iosched_cfq
);
3765 kmem_cache_destroy(cfq_pool
);
3768 module_init(cfq_init
);
3769 module_exit(cfq_exit
);
3771 MODULE_AUTHOR("Jens Axboe");
3772 MODULE_LICENSE("GPL");
3773 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");