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
;
224 * Per block device queue structure
227 struct request_queue
*queue
;
228 /* Root service tree for cfq_groups */
229 struct cfq_rb_root grp_service_tree
;
230 struct cfq_group
*root_group
;
233 * The priority currently being served
235 enum wl_prio_t serving_prio
;
236 enum wl_type_t serving_type
;
237 unsigned long workload_expires
;
238 struct cfq_group
*serving_group
;
241 * Each priority tree is sorted by next_request position. These
242 * trees are used when determining if two or more queues are
243 * interleaving requests (see cfq_close_cooperator).
245 struct rb_root prio_trees
[CFQ_PRIO_LISTS
];
247 unsigned int busy_queues
;
248 unsigned int busy_sync_queues
;
254 * queue-depth detection
260 * -1 => indeterminate, (cfq will behave as if NCQ is present, to allow better detection)
261 * 1 => NCQ is present (hw_tag_est_depth is the estimated max depth)
264 int hw_tag_est_depth
;
265 unsigned int hw_tag_samples
;
268 * idle window management
270 struct timer_list idle_slice_timer
;
271 struct work_struct unplug_work
;
273 struct cfq_queue
*active_queue
;
274 struct cfq_io_cq
*active_cic
;
277 * async queue for each priority case
279 struct cfq_queue
*async_cfqq
[2][IOPRIO_BE_NR
];
280 struct cfq_queue
*async_idle_cfqq
;
282 sector_t last_position
;
285 * tunables, see top of file
287 unsigned int cfq_quantum
;
288 unsigned int cfq_fifo_expire
[2];
289 unsigned int cfq_back_penalty
;
290 unsigned int cfq_back_max
;
291 unsigned int cfq_slice
[2];
292 unsigned int cfq_slice_async_rq
;
293 unsigned int cfq_slice_idle
;
294 unsigned int cfq_group_idle
;
295 unsigned int cfq_latency
;
298 * Fallback dummy cfqq for extreme OOM conditions
300 struct cfq_queue oom_cfqq
;
302 unsigned long last_delayed_sync
;
305 static inline struct cfq_group
*blkg_to_cfqg(struct blkio_group
*blkg
)
307 return blkg_to_pdata(blkg
, &blkio_policy_cfq
);
310 static inline struct blkio_group
*cfqg_to_blkg(struct cfq_group
*cfqg
)
312 return pdata_to_blkg(cfqg
, &blkio_policy_cfq
);
315 static struct cfq_group
*cfq_get_next_cfqg(struct cfq_data
*cfqd
);
317 static struct cfq_rb_root
*service_tree_for(struct cfq_group
*cfqg
,
324 if (prio
== IDLE_WORKLOAD
)
325 return &cfqg
->service_tree_idle
;
327 return &cfqg
->service_trees
[prio
][type
];
330 enum cfqq_state_flags
{
331 CFQ_CFQQ_FLAG_on_rr
= 0, /* on round-robin busy list */
332 CFQ_CFQQ_FLAG_wait_request
, /* waiting for a request */
333 CFQ_CFQQ_FLAG_must_dispatch
, /* must be allowed a dispatch */
334 CFQ_CFQQ_FLAG_must_alloc_slice
, /* per-slice must_alloc flag */
335 CFQ_CFQQ_FLAG_fifo_expire
, /* FIFO checked in this slice */
336 CFQ_CFQQ_FLAG_idle_window
, /* slice idling enabled */
337 CFQ_CFQQ_FLAG_prio_changed
, /* task priority has changed */
338 CFQ_CFQQ_FLAG_slice_new
, /* no requests dispatched in slice */
339 CFQ_CFQQ_FLAG_sync
, /* synchronous queue */
340 CFQ_CFQQ_FLAG_coop
, /* cfqq is shared */
341 CFQ_CFQQ_FLAG_split_coop
, /* shared cfqq will be splitted */
342 CFQ_CFQQ_FLAG_deep
, /* sync cfqq experienced large depth */
343 CFQ_CFQQ_FLAG_wait_busy
, /* Waiting for next request */
346 #define CFQ_CFQQ_FNS(name) \
347 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \
349 (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name); \
351 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \
353 (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \
355 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \
357 return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \
361 CFQ_CFQQ_FNS(wait_request
);
362 CFQ_CFQQ_FNS(must_dispatch
);
363 CFQ_CFQQ_FNS(must_alloc_slice
);
364 CFQ_CFQQ_FNS(fifo_expire
);
365 CFQ_CFQQ_FNS(idle_window
);
366 CFQ_CFQQ_FNS(prio_changed
);
367 CFQ_CFQQ_FNS(slice_new
);
370 CFQ_CFQQ_FNS(split_coop
);
372 CFQ_CFQQ_FNS(wait_busy
);
375 #ifdef CONFIG_CFQ_GROUP_IOSCHED
376 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
377 blk_add_trace_msg((cfqd)->queue, "cfq%d%c %s " fmt, (cfqq)->pid, \
378 cfq_cfqq_sync((cfqq)) ? 'S' : 'A', \
379 blkg_path(cfqg_to_blkg((cfqq)->cfqg)), ##args)
381 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...) \
382 blk_add_trace_msg((cfqd)->queue, "%s " fmt, \
383 blkg_path(cfqg_to_blkg((cfqg))), ##args) \
386 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
387 blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args)
388 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...) do {} while (0)
390 #define cfq_log(cfqd, fmt, args...) \
391 blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
393 /* Traverses through cfq group service trees */
394 #define for_each_cfqg_st(cfqg, i, j, st) \
395 for (i = 0; i <= IDLE_WORKLOAD; i++) \
396 for (j = 0, st = i < IDLE_WORKLOAD ? &cfqg->service_trees[i][j]\
397 : &cfqg->service_tree_idle; \
398 (i < IDLE_WORKLOAD && j <= SYNC_WORKLOAD) || \
399 (i == IDLE_WORKLOAD && j == 0); \
400 j++, st = i < IDLE_WORKLOAD ? \
401 &cfqg->service_trees[i][j]: NULL) \
403 static inline bool cfq_io_thinktime_big(struct cfq_data *cfqd,
404 struct cfq_ttime
*ttime
, bool group_idle
)
407 if (!sample_valid(ttime
->ttime_samples
))
410 slice
= cfqd
->cfq_group_idle
;
412 slice
= cfqd
->cfq_slice_idle
;
413 return ttime
->ttime_mean
> slice
;
416 static inline bool iops_mode(struct cfq_data
*cfqd
)
419 * If we are not idling on queues and it is a NCQ drive, parallel
420 * execution of requests is on and measuring time is not possible
421 * in most of the cases until and unless we drive shallower queue
422 * depths and that becomes a performance bottleneck. In such cases
423 * switch to start providing fairness in terms of number of IOs.
425 if (!cfqd
->cfq_slice_idle
&& cfqd
->hw_tag
)
431 static inline enum wl_prio_t
cfqq_prio(struct cfq_queue
*cfqq
)
433 if (cfq_class_idle(cfqq
))
434 return IDLE_WORKLOAD
;
435 if (cfq_class_rt(cfqq
))
441 static enum wl_type_t
cfqq_type(struct cfq_queue
*cfqq
)
443 if (!cfq_cfqq_sync(cfqq
))
444 return ASYNC_WORKLOAD
;
445 if (!cfq_cfqq_idle_window(cfqq
))
446 return SYNC_NOIDLE_WORKLOAD
;
447 return SYNC_WORKLOAD
;
450 static inline int cfq_group_busy_queues_wl(enum wl_prio_t wl
,
451 struct cfq_data
*cfqd
,
452 struct cfq_group
*cfqg
)
454 if (wl
== IDLE_WORKLOAD
)
455 return cfqg
->service_tree_idle
.count
;
457 return cfqg
->service_trees
[wl
][ASYNC_WORKLOAD
].count
458 + cfqg
->service_trees
[wl
][SYNC_NOIDLE_WORKLOAD
].count
459 + cfqg
->service_trees
[wl
][SYNC_WORKLOAD
].count
;
462 static inline int cfqg_busy_async_queues(struct cfq_data
*cfqd
,
463 struct cfq_group
*cfqg
)
465 return cfqg
->service_trees
[RT_WORKLOAD
][ASYNC_WORKLOAD
].count
466 + cfqg
->service_trees
[BE_WORKLOAD
][ASYNC_WORKLOAD
].count
;
469 static void cfq_dispatch_insert(struct request_queue
*, struct request
*);
470 static struct cfq_queue
*cfq_get_queue(struct cfq_data
*, bool,
471 struct io_context
*, gfp_t
);
473 static inline struct cfq_io_cq
*icq_to_cic(struct io_cq
*icq
)
475 /* cic->icq is the first member, %NULL will convert to %NULL */
476 return container_of(icq
, struct cfq_io_cq
, icq
);
479 static inline struct cfq_io_cq
*cfq_cic_lookup(struct cfq_data
*cfqd
,
480 struct io_context
*ioc
)
483 return icq_to_cic(ioc_lookup_icq(ioc
, cfqd
->queue
));
487 static inline struct cfq_queue
*cic_to_cfqq(struct cfq_io_cq
*cic
, bool is_sync
)
489 return cic
->cfqq
[is_sync
];
492 static inline void cic_set_cfqq(struct cfq_io_cq
*cic
, struct cfq_queue
*cfqq
,
495 cic
->cfqq
[is_sync
] = cfqq
;
498 static inline struct cfq_data
*cic_to_cfqd(struct cfq_io_cq
*cic
)
500 return cic
->icq
.q
->elevator
->elevator_data
;
504 * We regard a request as SYNC, if it's either a read or has the SYNC bit
505 * set (in which case it could also be direct WRITE).
507 static inline bool cfq_bio_sync(struct bio
*bio
)
509 return bio_data_dir(bio
) == READ
|| (bio
->bi_rw
& REQ_SYNC
);
513 * scheduler run of queue, if there are requests pending and no one in the
514 * driver that will restart queueing
516 static inline void cfq_schedule_dispatch(struct cfq_data
*cfqd
)
518 if (cfqd
->busy_queues
) {
519 cfq_log(cfqd
, "schedule dispatch");
520 kblockd_schedule_work(cfqd
->queue
, &cfqd
->unplug_work
);
525 * Scale schedule slice based on io priority. Use the sync time slice only
526 * if a queue is marked sync and has sync io queued. A sync queue with async
527 * io only, should not get full sync slice length.
529 static inline int cfq_prio_slice(struct cfq_data
*cfqd
, bool sync
,
532 const int base_slice
= cfqd
->cfq_slice
[sync
];
534 WARN_ON(prio
>= IOPRIO_BE_NR
);
536 return base_slice
+ (base_slice
/CFQ_SLICE_SCALE
* (4 - prio
));
540 cfq_prio_to_slice(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
542 return cfq_prio_slice(cfqd
, cfq_cfqq_sync(cfqq
), cfqq
->ioprio
);
545 static inline u64
cfq_scale_slice(unsigned long delta
, struct cfq_group
*cfqg
)
547 u64 d
= delta
<< CFQ_SERVICE_SHIFT
;
549 d
= d
* BLKIO_WEIGHT_DEFAULT
;
550 do_div(d
, cfqg
->weight
);
554 static inline u64
max_vdisktime(u64 min_vdisktime
, u64 vdisktime
)
556 s64 delta
= (s64
)(vdisktime
- min_vdisktime
);
558 min_vdisktime
= vdisktime
;
560 return min_vdisktime
;
563 static inline u64
min_vdisktime(u64 min_vdisktime
, u64 vdisktime
)
565 s64 delta
= (s64
)(vdisktime
- min_vdisktime
);
567 min_vdisktime
= vdisktime
;
569 return min_vdisktime
;
572 static void update_min_vdisktime(struct cfq_rb_root
*st
)
574 struct cfq_group
*cfqg
;
577 cfqg
= rb_entry_cfqg(st
->left
);
578 st
->min_vdisktime
= max_vdisktime(st
->min_vdisktime
,
584 * get averaged number of queues of RT/BE priority.
585 * average is updated, with a formula that gives more weight to higher numbers,
586 * to quickly follows sudden increases and decrease slowly
589 static inline unsigned cfq_group_get_avg_queues(struct cfq_data
*cfqd
,
590 struct cfq_group
*cfqg
, bool rt
)
592 unsigned min_q
, max_q
;
593 unsigned mult
= cfq_hist_divisor
- 1;
594 unsigned round
= cfq_hist_divisor
/ 2;
595 unsigned busy
= cfq_group_busy_queues_wl(rt
, cfqd
, cfqg
);
597 min_q
= min(cfqg
->busy_queues_avg
[rt
], busy
);
598 max_q
= max(cfqg
->busy_queues_avg
[rt
], busy
);
599 cfqg
->busy_queues_avg
[rt
] = (mult
* max_q
+ min_q
+ round
) /
601 return cfqg
->busy_queues_avg
[rt
];
604 static inline unsigned
605 cfq_group_slice(struct cfq_data
*cfqd
, struct cfq_group
*cfqg
)
607 struct cfq_rb_root
*st
= &cfqd
->grp_service_tree
;
609 return cfq_target_latency
* cfqg
->weight
/ st
->total_weight
;
612 static inline unsigned
613 cfq_scaled_cfqq_slice(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
615 unsigned slice
= cfq_prio_to_slice(cfqd
, cfqq
);
616 if (cfqd
->cfq_latency
) {
618 * interested queues (we consider only the ones with the same
619 * priority class in the cfq group)
621 unsigned iq
= cfq_group_get_avg_queues(cfqd
, cfqq
->cfqg
,
623 unsigned sync_slice
= cfqd
->cfq_slice
[1];
624 unsigned expect_latency
= sync_slice
* iq
;
625 unsigned group_slice
= cfq_group_slice(cfqd
, cfqq
->cfqg
);
627 if (expect_latency
> group_slice
) {
628 unsigned base_low_slice
= 2 * cfqd
->cfq_slice_idle
;
629 /* scale low_slice according to IO priority
630 * and sync vs async */
632 min(slice
, base_low_slice
* slice
/ sync_slice
);
633 /* the adapted slice value is scaled to fit all iqs
634 * into the target latency */
635 slice
= max(slice
* group_slice
/ expect_latency
,
643 cfq_set_prio_slice(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
645 unsigned slice
= cfq_scaled_cfqq_slice(cfqd
, cfqq
);
647 cfqq
->slice_start
= jiffies
;
648 cfqq
->slice_end
= jiffies
+ slice
;
649 cfqq
->allocated_slice
= slice
;
650 cfq_log_cfqq(cfqd
, cfqq
, "set_slice=%lu", cfqq
->slice_end
- jiffies
);
654 * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
655 * isn't valid until the first request from the dispatch is activated
656 * and the slice time set.
658 static inline bool cfq_slice_used(struct cfq_queue
*cfqq
)
660 if (cfq_cfqq_slice_new(cfqq
))
662 if (time_before(jiffies
, cfqq
->slice_end
))
669 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
670 * We choose the request that is closest to the head right now. Distance
671 * behind the head is penalized and only allowed to a certain extent.
673 static struct request
*
674 cfq_choose_req(struct cfq_data
*cfqd
, struct request
*rq1
, struct request
*rq2
, sector_t last
)
676 sector_t s1
, s2
, d1
= 0, d2
= 0;
677 unsigned long back_max
;
678 #define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */
679 #define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */
680 unsigned wrap
= 0; /* bit mask: requests behind the disk head? */
682 if (rq1
== NULL
|| rq1
== rq2
)
687 if (rq_is_sync(rq1
) != rq_is_sync(rq2
))
688 return rq_is_sync(rq1
) ? rq1
: rq2
;
690 if ((rq1
->cmd_flags
^ rq2
->cmd_flags
) & REQ_PRIO
)
691 return rq1
->cmd_flags
& REQ_PRIO
? rq1
: rq2
;
693 s1
= blk_rq_pos(rq1
);
694 s2
= blk_rq_pos(rq2
);
697 * by definition, 1KiB is 2 sectors
699 back_max
= cfqd
->cfq_back_max
* 2;
702 * Strict one way elevator _except_ in the case where we allow
703 * short backward seeks which are biased as twice the cost of a
704 * similar forward seek.
708 else if (s1
+ back_max
>= last
)
709 d1
= (last
- s1
) * cfqd
->cfq_back_penalty
;
711 wrap
|= CFQ_RQ1_WRAP
;
715 else if (s2
+ back_max
>= last
)
716 d2
= (last
- s2
) * cfqd
->cfq_back_penalty
;
718 wrap
|= CFQ_RQ2_WRAP
;
720 /* Found required data */
723 * By doing switch() on the bit mask "wrap" we avoid having to
724 * check two variables for all permutations: --> faster!
727 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
743 case (CFQ_RQ1_WRAP
|CFQ_RQ2_WRAP
): /* both rqs wrapped */
746 * Since both rqs are wrapped,
747 * start with the one that's further behind head
748 * (--> only *one* back seek required),
749 * since back seek takes more time than forward.
759 * The below is leftmost cache rbtree addon
761 static struct cfq_queue
*cfq_rb_first(struct cfq_rb_root
*root
)
763 /* Service tree is empty */
768 root
->left
= rb_first(&root
->rb
);
771 return rb_entry(root
->left
, struct cfq_queue
, rb_node
);
776 static struct cfq_group
*cfq_rb_first_group(struct cfq_rb_root
*root
)
779 root
->left
= rb_first(&root
->rb
);
782 return rb_entry_cfqg(root
->left
);
787 static void rb_erase_init(struct rb_node
*n
, struct rb_root
*root
)
793 static void cfq_rb_erase(struct rb_node
*n
, struct cfq_rb_root
*root
)
797 rb_erase_init(n
, &root
->rb
);
802 * would be nice to take fifo expire time into account as well
804 static struct request
*
805 cfq_find_next_rq(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
806 struct request
*last
)
808 struct rb_node
*rbnext
= rb_next(&last
->rb_node
);
809 struct rb_node
*rbprev
= rb_prev(&last
->rb_node
);
810 struct request
*next
= NULL
, *prev
= NULL
;
812 BUG_ON(RB_EMPTY_NODE(&last
->rb_node
));
815 prev
= rb_entry_rq(rbprev
);
818 next
= rb_entry_rq(rbnext
);
820 rbnext
= rb_first(&cfqq
->sort_list
);
821 if (rbnext
&& rbnext
!= &last
->rb_node
)
822 next
= rb_entry_rq(rbnext
);
825 return cfq_choose_req(cfqd
, next
, prev
, blk_rq_pos(last
));
828 static unsigned long cfq_slice_offset(struct cfq_data
*cfqd
,
829 struct cfq_queue
*cfqq
)
832 * just an approximation, should be ok.
834 return (cfqq
->cfqg
->nr_cfqq
- 1) * (cfq_prio_slice(cfqd
, 1, 0) -
835 cfq_prio_slice(cfqd
, cfq_cfqq_sync(cfqq
), cfqq
->ioprio
));
839 cfqg_key(struct cfq_rb_root
*st
, struct cfq_group
*cfqg
)
841 return cfqg
->vdisktime
- st
->min_vdisktime
;
845 __cfq_group_service_tree_add(struct cfq_rb_root
*st
, struct cfq_group
*cfqg
)
847 struct rb_node
**node
= &st
->rb
.rb_node
;
848 struct rb_node
*parent
= NULL
;
849 struct cfq_group
*__cfqg
;
850 s64 key
= cfqg_key(st
, cfqg
);
853 while (*node
!= NULL
) {
855 __cfqg
= rb_entry_cfqg(parent
);
857 if (key
< cfqg_key(st
, __cfqg
))
858 node
= &parent
->rb_left
;
860 node
= &parent
->rb_right
;
866 st
->left
= &cfqg
->rb_node
;
868 rb_link_node(&cfqg
->rb_node
, parent
, node
);
869 rb_insert_color(&cfqg
->rb_node
, &st
->rb
);
873 cfq_update_group_weight(struct cfq_group
*cfqg
)
875 BUG_ON(!RB_EMPTY_NODE(&cfqg
->rb_node
));
876 if (cfqg
->needs_update
) {
877 cfqg
->weight
= cfqg
->new_weight
;
878 cfqg
->needs_update
= false;
883 cfq_group_service_tree_add(struct cfq_rb_root
*st
, struct cfq_group
*cfqg
)
885 BUG_ON(!RB_EMPTY_NODE(&cfqg
->rb_node
));
887 cfq_update_group_weight(cfqg
);
888 __cfq_group_service_tree_add(st
, cfqg
);
889 st
->total_weight
+= cfqg
->weight
;
893 cfq_group_notify_queue_add(struct cfq_data
*cfqd
, struct cfq_group
*cfqg
)
895 struct cfq_rb_root
*st
= &cfqd
->grp_service_tree
;
896 struct cfq_group
*__cfqg
;
900 if (!RB_EMPTY_NODE(&cfqg
->rb_node
))
904 * Currently put the group at the end. Later implement something
905 * so that groups get lesser vtime based on their weights, so that
906 * if group does not loose all if it was not continuously backlogged.
908 n
= rb_last(&st
->rb
);
910 __cfqg
= rb_entry_cfqg(n
);
911 cfqg
->vdisktime
= __cfqg
->vdisktime
+ CFQ_IDLE_DELAY
;
913 cfqg
->vdisktime
= st
->min_vdisktime
;
914 cfq_group_service_tree_add(st
, cfqg
);
918 cfq_group_service_tree_del(struct cfq_rb_root
*st
, struct cfq_group
*cfqg
)
920 st
->total_weight
-= cfqg
->weight
;
921 if (!RB_EMPTY_NODE(&cfqg
->rb_node
))
922 cfq_rb_erase(&cfqg
->rb_node
, st
);
926 cfq_group_notify_queue_del(struct cfq_data
*cfqd
, struct cfq_group
*cfqg
)
928 struct cfq_rb_root
*st
= &cfqd
->grp_service_tree
;
930 BUG_ON(cfqg
->nr_cfqq
< 1);
933 /* If there are other cfq queues under this group, don't delete it */
937 cfq_log_cfqg(cfqd
, cfqg
, "del_from_rr group");
938 cfq_group_service_tree_del(st
, cfqg
);
939 cfqg
->saved_workload_slice
= 0;
940 cfq_blkiocg_update_dequeue_stats(cfqg_to_blkg(cfqg
),
941 &blkio_policy_cfq
, 1);
944 static inline unsigned int cfq_cfqq_slice_usage(struct cfq_queue
*cfqq
,
945 unsigned int *unaccounted_time
)
947 unsigned int slice_used
;
950 * Queue got expired before even a single request completed or
951 * got expired immediately after first request completion.
953 if (!cfqq
->slice_start
|| cfqq
->slice_start
== jiffies
) {
955 * Also charge the seek time incurred to the group, otherwise
956 * if there are mutiple queues in the group, each can dispatch
957 * a single request on seeky media and cause lots of seek time
958 * and group will never know it.
960 slice_used
= max_t(unsigned, (jiffies
- cfqq
->dispatch_start
),
963 slice_used
= jiffies
- cfqq
->slice_start
;
964 if (slice_used
> cfqq
->allocated_slice
) {
965 *unaccounted_time
= slice_used
- cfqq
->allocated_slice
;
966 slice_used
= cfqq
->allocated_slice
;
968 if (time_after(cfqq
->slice_start
, cfqq
->dispatch_start
))
969 *unaccounted_time
+= cfqq
->slice_start
-
970 cfqq
->dispatch_start
;
976 static void cfq_group_served(struct cfq_data
*cfqd
, struct cfq_group
*cfqg
,
977 struct cfq_queue
*cfqq
)
979 struct cfq_rb_root
*st
= &cfqd
->grp_service_tree
;
980 unsigned int used_sl
, charge
, unaccounted_sl
= 0;
981 int nr_sync
= cfqg
->nr_cfqq
- cfqg_busy_async_queues(cfqd
, cfqg
)
982 - cfqg
->service_tree_idle
.count
;
985 used_sl
= charge
= cfq_cfqq_slice_usage(cfqq
, &unaccounted_sl
);
988 charge
= cfqq
->slice_dispatch
;
989 else if (!cfq_cfqq_sync(cfqq
) && !nr_sync
)
990 charge
= cfqq
->allocated_slice
;
992 /* Can't update vdisktime while group is on service tree */
993 cfq_group_service_tree_del(st
, cfqg
);
994 cfqg
->vdisktime
+= cfq_scale_slice(charge
, cfqg
);
995 /* If a new weight was requested, update now, off tree */
996 cfq_group_service_tree_add(st
, cfqg
);
998 /* This group is being expired. Save the context */
999 if (time_after(cfqd
->workload_expires
, jiffies
)) {
1000 cfqg
->saved_workload_slice
= cfqd
->workload_expires
1002 cfqg
->saved_workload
= cfqd
->serving_type
;
1003 cfqg
->saved_serving_prio
= cfqd
->serving_prio
;
1005 cfqg
->saved_workload_slice
= 0;
1007 cfq_log_cfqg(cfqd
, cfqg
, "served: vt=%llu min_vt=%llu", cfqg
->vdisktime
,
1009 cfq_log_cfqq(cfqq
->cfqd
, cfqq
,
1010 "sl_used=%u disp=%u charge=%u iops=%u sect=%lu",
1011 used_sl
, cfqq
->slice_dispatch
, charge
,
1012 iops_mode(cfqd
), cfqq
->nr_sectors
);
1013 cfq_blkiocg_update_timeslice_used(cfqg_to_blkg(cfqg
), &blkio_policy_cfq
,
1014 used_sl
, unaccounted_sl
);
1015 cfq_blkiocg_set_start_empty_time(cfqg_to_blkg(cfqg
), &blkio_policy_cfq
);
1019 * cfq_init_cfqg_base - initialize base part of a cfq_group
1020 * @cfqg: cfq_group to initialize
1022 * Initialize the base part which is used whether %CONFIG_CFQ_GROUP_IOSCHED
1023 * is enabled or not.
1025 static void cfq_init_cfqg_base(struct cfq_group
*cfqg
)
1027 struct cfq_rb_root
*st
;
1030 for_each_cfqg_st(cfqg
, i
, j
, st
)
1032 RB_CLEAR_NODE(&cfqg
->rb_node
);
1034 cfqg
->ttime
.last_end_request
= jiffies
;
1037 #ifdef CONFIG_CFQ_GROUP_IOSCHED
1038 static void cfq_update_blkio_group_weight(struct request_queue
*q
,
1039 struct blkio_group
*blkg
,
1040 unsigned int weight
)
1042 struct cfq_group
*cfqg
= blkg_to_cfqg(blkg
);
1044 cfqg
->new_weight
= weight
;
1045 cfqg
->needs_update
= true;
1048 static void cfq_init_blkio_group(struct blkio_group
*blkg
)
1050 struct cfq_group
*cfqg
= blkg_to_cfqg(blkg
);
1052 cfq_init_cfqg_base(cfqg
);
1053 cfqg
->weight
= blkg
->blkcg
->weight
;
1057 * Search for the cfq group current task belongs to. request_queue lock must
1060 static struct cfq_group
*cfq_lookup_create_cfqg(struct cfq_data
*cfqd
,
1061 struct blkio_cgroup
*blkcg
)
1063 struct request_queue
*q
= cfqd
->queue
;
1064 struct cfq_group
*cfqg
= NULL
;
1066 /* avoid lookup for the common case where there's no blkio cgroup */
1067 if (blkcg
== &blkio_root_cgroup
) {
1068 cfqg
= cfqd
->root_group
;
1070 struct blkio_group
*blkg
;
1072 blkg
= blkg_lookup_create(blkcg
, q
, BLKIO_POLICY_PROP
, false);
1074 cfqg
= blkg_to_cfqg(blkg
);
1080 static void cfq_link_cfqq_cfqg(struct cfq_queue
*cfqq
, struct cfq_group
*cfqg
)
1082 /* Currently, all async queues are mapped to root group */
1083 if (!cfq_cfqq_sync(cfqq
))
1084 cfqg
= cfqq
->cfqd
->root_group
;
1087 /* cfqq reference on cfqg */
1088 blkg_get(cfqg_to_blkg(cfqg
));
1091 #else /* GROUP_IOSCHED */
1092 static struct cfq_group
*cfq_lookup_create_cfqg(struct cfq_data
*cfqd
,
1093 struct blkio_cgroup
*blkcg
)
1095 return cfqd
->root_group
;
1099 cfq_link_cfqq_cfqg(struct cfq_queue
*cfqq
, struct cfq_group
*cfqg
) {
1103 #endif /* GROUP_IOSCHED */
1106 * The cfqd->service_trees holds all pending cfq_queue's that have
1107 * requests waiting to be processed. It is sorted in the order that
1108 * we will service the queues.
1110 static void cfq_service_tree_add(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
1113 struct rb_node
**p
, *parent
;
1114 struct cfq_queue
*__cfqq
;
1115 unsigned long rb_key
;
1116 struct cfq_rb_root
*service_tree
;
1120 service_tree
= service_tree_for(cfqq
->cfqg
, cfqq_prio(cfqq
),
1122 if (cfq_class_idle(cfqq
)) {
1123 rb_key
= CFQ_IDLE_DELAY
;
1124 parent
= rb_last(&service_tree
->rb
);
1125 if (parent
&& parent
!= &cfqq
->rb_node
) {
1126 __cfqq
= rb_entry(parent
, struct cfq_queue
, rb_node
);
1127 rb_key
+= __cfqq
->rb_key
;
1130 } else if (!add_front
) {
1132 * Get our rb key offset. Subtract any residual slice
1133 * value carried from last service. A negative resid
1134 * count indicates slice overrun, and this should position
1135 * the next service time further away in the tree.
1137 rb_key
= cfq_slice_offset(cfqd
, cfqq
) + jiffies
;
1138 rb_key
-= cfqq
->slice_resid
;
1139 cfqq
->slice_resid
= 0;
1142 __cfqq
= cfq_rb_first(service_tree
);
1143 rb_key
+= __cfqq
? __cfqq
->rb_key
: jiffies
;
1146 if (!RB_EMPTY_NODE(&cfqq
->rb_node
)) {
1149 * same position, nothing more to do
1151 if (rb_key
== cfqq
->rb_key
&&
1152 cfqq
->service_tree
== service_tree
)
1155 cfq_rb_erase(&cfqq
->rb_node
, cfqq
->service_tree
);
1156 cfqq
->service_tree
= NULL
;
1161 cfqq
->service_tree
= service_tree
;
1162 p
= &service_tree
->rb
.rb_node
;
1167 __cfqq
= rb_entry(parent
, struct cfq_queue
, rb_node
);
1170 * sort by key, that represents service time.
1172 if (time_before(rb_key
, __cfqq
->rb_key
))
1175 n
= &(*p
)->rb_right
;
1183 service_tree
->left
= &cfqq
->rb_node
;
1185 cfqq
->rb_key
= rb_key
;
1186 rb_link_node(&cfqq
->rb_node
, parent
, p
);
1187 rb_insert_color(&cfqq
->rb_node
, &service_tree
->rb
);
1188 service_tree
->count
++;
1189 if (add_front
|| !new_cfqq
)
1191 cfq_group_notify_queue_add(cfqd
, cfqq
->cfqg
);
1194 static struct cfq_queue
*
1195 cfq_prio_tree_lookup(struct cfq_data
*cfqd
, struct rb_root
*root
,
1196 sector_t sector
, struct rb_node
**ret_parent
,
1197 struct rb_node
***rb_link
)
1199 struct rb_node
**p
, *parent
;
1200 struct cfq_queue
*cfqq
= NULL
;
1208 cfqq
= rb_entry(parent
, struct cfq_queue
, p_node
);
1211 * Sort strictly based on sector. Smallest to the left,
1212 * largest to the right.
1214 if (sector
> blk_rq_pos(cfqq
->next_rq
))
1215 n
= &(*p
)->rb_right
;
1216 else if (sector
< blk_rq_pos(cfqq
->next_rq
))
1224 *ret_parent
= parent
;
1230 static void cfq_prio_tree_add(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
1232 struct rb_node
**p
, *parent
;
1233 struct cfq_queue
*__cfqq
;
1236 rb_erase(&cfqq
->p_node
, cfqq
->p_root
);
1237 cfqq
->p_root
= NULL
;
1240 if (cfq_class_idle(cfqq
))
1245 cfqq
->p_root
= &cfqd
->prio_trees
[cfqq
->org_ioprio
];
1246 __cfqq
= cfq_prio_tree_lookup(cfqd
, cfqq
->p_root
,
1247 blk_rq_pos(cfqq
->next_rq
), &parent
, &p
);
1249 rb_link_node(&cfqq
->p_node
, parent
, p
);
1250 rb_insert_color(&cfqq
->p_node
, cfqq
->p_root
);
1252 cfqq
->p_root
= NULL
;
1256 * Update cfqq's position in the service tree.
1258 static void cfq_resort_rr_list(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
1261 * Resorting requires the cfqq to be on the RR list already.
1263 if (cfq_cfqq_on_rr(cfqq
)) {
1264 cfq_service_tree_add(cfqd
, cfqq
, 0);
1265 cfq_prio_tree_add(cfqd
, cfqq
);
1270 * add to busy list of queues for service, trying to be fair in ordering
1271 * the pending list according to last request service
1273 static void cfq_add_cfqq_rr(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
1275 cfq_log_cfqq(cfqd
, cfqq
, "add_to_rr");
1276 BUG_ON(cfq_cfqq_on_rr(cfqq
));
1277 cfq_mark_cfqq_on_rr(cfqq
);
1278 cfqd
->busy_queues
++;
1279 if (cfq_cfqq_sync(cfqq
))
1280 cfqd
->busy_sync_queues
++;
1282 cfq_resort_rr_list(cfqd
, cfqq
);
1286 * Called when the cfqq no longer has requests pending, remove it from
1289 static void cfq_del_cfqq_rr(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
1291 cfq_log_cfqq(cfqd
, cfqq
, "del_from_rr");
1292 BUG_ON(!cfq_cfqq_on_rr(cfqq
));
1293 cfq_clear_cfqq_on_rr(cfqq
);
1295 if (!RB_EMPTY_NODE(&cfqq
->rb_node
)) {
1296 cfq_rb_erase(&cfqq
->rb_node
, cfqq
->service_tree
);
1297 cfqq
->service_tree
= NULL
;
1300 rb_erase(&cfqq
->p_node
, cfqq
->p_root
);
1301 cfqq
->p_root
= NULL
;
1304 cfq_group_notify_queue_del(cfqd
, cfqq
->cfqg
);
1305 BUG_ON(!cfqd
->busy_queues
);
1306 cfqd
->busy_queues
--;
1307 if (cfq_cfqq_sync(cfqq
))
1308 cfqd
->busy_sync_queues
--;
1312 * rb tree support functions
1314 static void cfq_del_rq_rb(struct request
*rq
)
1316 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
1317 const int sync
= rq_is_sync(rq
);
1319 BUG_ON(!cfqq
->queued
[sync
]);
1320 cfqq
->queued
[sync
]--;
1322 elv_rb_del(&cfqq
->sort_list
, rq
);
1324 if (cfq_cfqq_on_rr(cfqq
) && RB_EMPTY_ROOT(&cfqq
->sort_list
)) {
1326 * Queue will be deleted from service tree when we actually
1327 * expire it later. Right now just remove it from prio tree
1331 rb_erase(&cfqq
->p_node
, cfqq
->p_root
);
1332 cfqq
->p_root
= NULL
;
1337 static void cfq_add_rq_rb(struct request
*rq
)
1339 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
1340 struct cfq_data
*cfqd
= cfqq
->cfqd
;
1341 struct request
*prev
;
1343 cfqq
->queued
[rq_is_sync(rq
)]++;
1345 elv_rb_add(&cfqq
->sort_list
, rq
);
1347 if (!cfq_cfqq_on_rr(cfqq
))
1348 cfq_add_cfqq_rr(cfqd
, cfqq
);
1351 * check if this request is a better next-serve candidate
1353 prev
= cfqq
->next_rq
;
1354 cfqq
->next_rq
= cfq_choose_req(cfqd
, cfqq
->next_rq
, rq
, cfqd
->last_position
);
1357 * adjust priority tree position, if ->next_rq changes
1359 if (prev
!= cfqq
->next_rq
)
1360 cfq_prio_tree_add(cfqd
, cfqq
);
1362 BUG_ON(!cfqq
->next_rq
);
1365 static void cfq_reposition_rq_rb(struct cfq_queue
*cfqq
, struct request
*rq
)
1367 elv_rb_del(&cfqq
->sort_list
, rq
);
1368 cfqq
->queued
[rq_is_sync(rq
)]--;
1369 cfq_blkiocg_update_io_remove_stats(cfqg_to_blkg(RQ_CFQG(rq
)),
1370 &blkio_policy_cfq
, rq_data_dir(rq
),
1373 cfq_blkiocg_update_io_add_stats(cfqg_to_blkg(RQ_CFQG(rq
)),
1375 cfqg_to_blkg(cfqq
->cfqd
->serving_group
),
1376 rq_data_dir(rq
), rq_is_sync(rq
));
1379 static struct request
*
1380 cfq_find_rq_fmerge(struct cfq_data
*cfqd
, struct bio
*bio
)
1382 struct task_struct
*tsk
= current
;
1383 struct cfq_io_cq
*cic
;
1384 struct cfq_queue
*cfqq
;
1386 cic
= cfq_cic_lookup(cfqd
, tsk
->io_context
);
1390 cfqq
= cic_to_cfqq(cic
, cfq_bio_sync(bio
));
1392 sector_t sector
= bio
->bi_sector
+ bio_sectors(bio
);
1394 return elv_rb_find(&cfqq
->sort_list
, sector
);
1400 static void cfq_activate_request(struct request_queue
*q
, struct request
*rq
)
1402 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
1404 cfqd
->rq_in_driver
++;
1405 cfq_log_cfqq(cfqd
, RQ_CFQQ(rq
), "activate rq, drv=%d",
1406 cfqd
->rq_in_driver
);
1408 cfqd
->last_position
= blk_rq_pos(rq
) + blk_rq_sectors(rq
);
1411 static void cfq_deactivate_request(struct request_queue
*q
, struct request
*rq
)
1413 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
1415 WARN_ON(!cfqd
->rq_in_driver
);
1416 cfqd
->rq_in_driver
--;
1417 cfq_log_cfqq(cfqd
, RQ_CFQQ(rq
), "deactivate rq, drv=%d",
1418 cfqd
->rq_in_driver
);
1421 static void cfq_remove_request(struct request
*rq
)
1423 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
1425 if (cfqq
->next_rq
== rq
)
1426 cfqq
->next_rq
= cfq_find_next_rq(cfqq
->cfqd
, cfqq
, rq
);
1428 list_del_init(&rq
->queuelist
);
1431 cfqq
->cfqd
->rq_queued
--;
1432 cfq_blkiocg_update_io_remove_stats(cfqg_to_blkg(RQ_CFQG(rq
)),
1433 &blkio_policy_cfq
, rq_data_dir(rq
),
1435 if (rq
->cmd_flags
& REQ_PRIO
) {
1436 WARN_ON(!cfqq
->prio_pending
);
1437 cfqq
->prio_pending
--;
1441 static int cfq_merge(struct request_queue
*q
, struct request
**req
,
1444 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
1445 struct request
*__rq
;
1447 __rq
= cfq_find_rq_fmerge(cfqd
, bio
);
1448 if (__rq
&& elv_rq_merge_ok(__rq
, bio
)) {
1450 return ELEVATOR_FRONT_MERGE
;
1453 return ELEVATOR_NO_MERGE
;
1456 static void cfq_merged_request(struct request_queue
*q
, struct request
*req
,
1459 if (type
== ELEVATOR_FRONT_MERGE
) {
1460 struct cfq_queue
*cfqq
= RQ_CFQQ(req
);
1462 cfq_reposition_rq_rb(cfqq
, req
);
1466 static void cfq_bio_merged(struct request_queue
*q
, struct request
*req
,
1469 cfq_blkiocg_update_io_merged_stats(cfqg_to_blkg(RQ_CFQG(req
)),
1470 &blkio_policy_cfq
, bio_data_dir(bio
),
1475 cfq_merged_requests(struct request_queue
*q
, struct request
*rq
,
1476 struct request
*next
)
1478 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
1479 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
1482 * reposition in fifo if next is older than rq
1484 if (!list_empty(&rq
->queuelist
) && !list_empty(&next
->queuelist
) &&
1485 time_before(rq_fifo_time(next
), rq_fifo_time(rq
))) {
1486 list_move(&rq
->queuelist
, &next
->queuelist
);
1487 rq_set_fifo_time(rq
, rq_fifo_time(next
));
1490 if (cfqq
->next_rq
== next
)
1492 cfq_remove_request(next
);
1493 cfq_blkiocg_update_io_merged_stats(cfqg_to_blkg(RQ_CFQG(rq
)),
1494 &blkio_policy_cfq
, rq_data_dir(next
),
1497 cfqq
= RQ_CFQQ(next
);
1499 * all requests of this queue are merged to other queues, delete it
1500 * from the service tree. If it's the active_queue,
1501 * cfq_dispatch_requests() will choose to expire it or do idle
1503 if (cfq_cfqq_on_rr(cfqq
) && RB_EMPTY_ROOT(&cfqq
->sort_list
) &&
1504 cfqq
!= cfqd
->active_queue
)
1505 cfq_del_cfqq_rr(cfqd
, cfqq
);
1508 static int cfq_allow_merge(struct request_queue
*q
, struct request
*rq
,
1511 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
1512 struct cfq_io_cq
*cic
;
1513 struct cfq_queue
*cfqq
;
1516 * Disallow merge of a sync bio into an async request.
1518 if (cfq_bio_sync(bio
) && !rq_is_sync(rq
))
1522 * Lookup the cfqq that this bio will be queued with and allow
1523 * merge only if rq is queued there.
1525 cic
= cfq_cic_lookup(cfqd
, current
->io_context
);
1529 cfqq
= cic_to_cfqq(cic
, cfq_bio_sync(bio
));
1530 return cfqq
== RQ_CFQQ(rq
);
1533 static inline void cfq_del_timer(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
1535 del_timer(&cfqd
->idle_slice_timer
);
1536 cfq_blkiocg_update_idle_time_stats(cfqg_to_blkg(cfqq
->cfqg
),
1540 static void __cfq_set_active_queue(struct cfq_data
*cfqd
,
1541 struct cfq_queue
*cfqq
)
1544 cfq_log_cfqq(cfqd
, cfqq
, "set_active wl_prio:%d wl_type:%d",
1545 cfqd
->serving_prio
, cfqd
->serving_type
);
1546 cfq_blkiocg_update_avg_queue_size_stats(cfqg_to_blkg(cfqq
->cfqg
),
1548 cfqq
->slice_start
= 0;
1549 cfqq
->dispatch_start
= jiffies
;
1550 cfqq
->allocated_slice
= 0;
1551 cfqq
->slice_end
= 0;
1552 cfqq
->slice_dispatch
= 0;
1553 cfqq
->nr_sectors
= 0;
1555 cfq_clear_cfqq_wait_request(cfqq
);
1556 cfq_clear_cfqq_must_dispatch(cfqq
);
1557 cfq_clear_cfqq_must_alloc_slice(cfqq
);
1558 cfq_clear_cfqq_fifo_expire(cfqq
);
1559 cfq_mark_cfqq_slice_new(cfqq
);
1561 cfq_del_timer(cfqd
, cfqq
);
1564 cfqd
->active_queue
= cfqq
;
1568 * current cfqq expired its slice (or was too idle), select new one
1571 __cfq_slice_expired(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
1574 cfq_log_cfqq(cfqd
, cfqq
, "slice expired t=%d", timed_out
);
1576 if (cfq_cfqq_wait_request(cfqq
))
1577 cfq_del_timer(cfqd
, cfqq
);
1579 cfq_clear_cfqq_wait_request(cfqq
);
1580 cfq_clear_cfqq_wait_busy(cfqq
);
1583 * If this cfqq is shared between multiple processes, check to
1584 * make sure that those processes are still issuing I/Os within
1585 * the mean seek distance. If not, it may be time to break the
1586 * queues apart again.
1588 if (cfq_cfqq_coop(cfqq
) && CFQQ_SEEKY(cfqq
))
1589 cfq_mark_cfqq_split_coop(cfqq
);
1592 * store what was left of this slice, if the queue idled/timed out
1595 if (cfq_cfqq_slice_new(cfqq
))
1596 cfqq
->slice_resid
= cfq_scaled_cfqq_slice(cfqd
, cfqq
);
1598 cfqq
->slice_resid
= cfqq
->slice_end
- jiffies
;
1599 cfq_log_cfqq(cfqd
, cfqq
, "resid=%ld", cfqq
->slice_resid
);
1602 cfq_group_served(cfqd
, cfqq
->cfqg
, cfqq
);
1604 if (cfq_cfqq_on_rr(cfqq
) && RB_EMPTY_ROOT(&cfqq
->sort_list
))
1605 cfq_del_cfqq_rr(cfqd
, cfqq
);
1607 cfq_resort_rr_list(cfqd
, cfqq
);
1609 if (cfqq
== cfqd
->active_queue
)
1610 cfqd
->active_queue
= NULL
;
1612 if (cfqd
->active_cic
) {
1613 put_io_context(cfqd
->active_cic
->icq
.ioc
);
1614 cfqd
->active_cic
= NULL
;
1618 static inline void cfq_slice_expired(struct cfq_data
*cfqd
, bool timed_out
)
1620 struct cfq_queue
*cfqq
= cfqd
->active_queue
;
1623 __cfq_slice_expired(cfqd
, cfqq
, timed_out
);
1627 * Get next queue for service. Unless we have a queue preemption,
1628 * we'll simply select the first cfqq in the service tree.
1630 static struct cfq_queue
*cfq_get_next_queue(struct cfq_data
*cfqd
)
1632 struct cfq_rb_root
*service_tree
=
1633 service_tree_for(cfqd
->serving_group
, cfqd
->serving_prio
,
1634 cfqd
->serving_type
);
1636 if (!cfqd
->rq_queued
)
1639 /* There is nothing to dispatch */
1642 if (RB_EMPTY_ROOT(&service_tree
->rb
))
1644 return cfq_rb_first(service_tree
);
1647 static struct cfq_queue
*cfq_get_next_queue_forced(struct cfq_data
*cfqd
)
1649 struct cfq_group
*cfqg
;
1650 struct cfq_queue
*cfqq
;
1652 struct cfq_rb_root
*st
;
1654 if (!cfqd
->rq_queued
)
1657 cfqg
= cfq_get_next_cfqg(cfqd
);
1661 for_each_cfqg_st(cfqg
, i
, j
, st
)
1662 if ((cfqq
= cfq_rb_first(st
)) != NULL
)
1668 * Get and set a new active queue for service.
1670 static struct cfq_queue
*cfq_set_active_queue(struct cfq_data
*cfqd
,
1671 struct cfq_queue
*cfqq
)
1674 cfqq
= cfq_get_next_queue(cfqd
);
1676 __cfq_set_active_queue(cfqd
, cfqq
);
1680 static inline sector_t
cfq_dist_from_last(struct cfq_data
*cfqd
,
1683 if (blk_rq_pos(rq
) >= cfqd
->last_position
)
1684 return blk_rq_pos(rq
) - cfqd
->last_position
;
1686 return cfqd
->last_position
- blk_rq_pos(rq
);
1689 static inline int cfq_rq_close(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
1692 return cfq_dist_from_last(cfqd
, rq
) <= CFQQ_CLOSE_THR
;
1695 static struct cfq_queue
*cfqq_close(struct cfq_data
*cfqd
,
1696 struct cfq_queue
*cur_cfqq
)
1698 struct rb_root
*root
= &cfqd
->prio_trees
[cur_cfqq
->org_ioprio
];
1699 struct rb_node
*parent
, *node
;
1700 struct cfq_queue
*__cfqq
;
1701 sector_t sector
= cfqd
->last_position
;
1703 if (RB_EMPTY_ROOT(root
))
1707 * First, if we find a request starting at the end of the last
1708 * request, choose it.
1710 __cfqq
= cfq_prio_tree_lookup(cfqd
, root
, sector
, &parent
, NULL
);
1715 * If the exact sector wasn't found, the parent of the NULL leaf
1716 * will contain the closest sector.
1718 __cfqq
= rb_entry(parent
, struct cfq_queue
, p_node
);
1719 if (cfq_rq_close(cfqd
, cur_cfqq
, __cfqq
->next_rq
))
1722 if (blk_rq_pos(__cfqq
->next_rq
) < sector
)
1723 node
= rb_next(&__cfqq
->p_node
);
1725 node
= rb_prev(&__cfqq
->p_node
);
1729 __cfqq
= rb_entry(node
, struct cfq_queue
, p_node
);
1730 if (cfq_rq_close(cfqd
, cur_cfqq
, __cfqq
->next_rq
))
1738 * cur_cfqq - passed in so that we don't decide that the current queue is
1739 * closely cooperating with itself.
1741 * So, basically we're assuming that that cur_cfqq has dispatched at least
1742 * one request, and that cfqd->last_position reflects a position on the disk
1743 * associated with the I/O issued by cur_cfqq. I'm not sure this is a valid
1746 static struct cfq_queue
*cfq_close_cooperator(struct cfq_data
*cfqd
,
1747 struct cfq_queue
*cur_cfqq
)
1749 struct cfq_queue
*cfqq
;
1751 if (cfq_class_idle(cur_cfqq
))
1753 if (!cfq_cfqq_sync(cur_cfqq
))
1755 if (CFQQ_SEEKY(cur_cfqq
))
1759 * Don't search priority tree if it's the only queue in the group.
1761 if (cur_cfqq
->cfqg
->nr_cfqq
== 1)
1765 * We should notice if some of the queues are cooperating, eg
1766 * working closely on the same area of the disk. In that case,
1767 * we can group them together and don't waste time idling.
1769 cfqq
= cfqq_close(cfqd
, cur_cfqq
);
1773 /* If new queue belongs to different cfq_group, don't choose it */
1774 if (cur_cfqq
->cfqg
!= cfqq
->cfqg
)
1778 * It only makes sense to merge sync queues.
1780 if (!cfq_cfqq_sync(cfqq
))
1782 if (CFQQ_SEEKY(cfqq
))
1786 * Do not merge queues of different priority classes
1788 if (cfq_class_rt(cfqq
) != cfq_class_rt(cur_cfqq
))
1795 * Determine whether we should enforce idle window for this queue.
1798 static bool cfq_should_idle(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
1800 enum wl_prio_t prio
= cfqq_prio(cfqq
);
1801 struct cfq_rb_root
*service_tree
= cfqq
->service_tree
;
1803 BUG_ON(!service_tree
);
1804 BUG_ON(!service_tree
->count
);
1806 if (!cfqd
->cfq_slice_idle
)
1809 /* We never do for idle class queues. */
1810 if (prio
== IDLE_WORKLOAD
)
1813 /* We do for queues that were marked with idle window flag. */
1814 if (cfq_cfqq_idle_window(cfqq
) &&
1815 !(blk_queue_nonrot(cfqd
->queue
) && cfqd
->hw_tag
))
1819 * Otherwise, we do only if they are the last ones
1820 * in their service tree.
1822 if (service_tree
->count
== 1 && cfq_cfqq_sync(cfqq
) &&
1823 !cfq_io_thinktime_big(cfqd
, &service_tree
->ttime
, false))
1825 cfq_log_cfqq(cfqd
, cfqq
, "Not idling. st->count:%d",
1826 service_tree
->count
);
1830 static void cfq_arm_slice_timer(struct cfq_data
*cfqd
)
1832 struct cfq_queue
*cfqq
= cfqd
->active_queue
;
1833 struct cfq_io_cq
*cic
;
1834 unsigned long sl
, group_idle
= 0;
1837 * SSD device without seek penalty, disable idling. But only do so
1838 * for devices that support queuing, otherwise we still have a problem
1839 * with sync vs async workloads.
1841 if (blk_queue_nonrot(cfqd
->queue
) && cfqd
->hw_tag
)
1844 WARN_ON(!RB_EMPTY_ROOT(&cfqq
->sort_list
));
1845 WARN_ON(cfq_cfqq_slice_new(cfqq
));
1848 * idle is disabled, either manually or by past process history
1850 if (!cfq_should_idle(cfqd
, cfqq
)) {
1851 /* no queue idling. Check for group idling */
1852 if (cfqd
->cfq_group_idle
)
1853 group_idle
= cfqd
->cfq_group_idle
;
1859 * still active requests from this queue, don't idle
1861 if (cfqq
->dispatched
)
1865 * task has exited, don't wait
1867 cic
= cfqd
->active_cic
;
1868 if (!cic
|| !atomic_read(&cic
->icq
.ioc
->nr_tasks
))
1872 * If our average think time is larger than the remaining time
1873 * slice, then don't idle. This avoids overrunning the allotted
1876 if (sample_valid(cic
->ttime
.ttime_samples
) &&
1877 (cfqq
->slice_end
- jiffies
< cic
->ttime
.ttime_mean
)) {
1878 cfq_log_cfqq(cfqd
, cfqq
, "Not idling. think_time:%lu",
1879 cic
->ttime
.ttime_mean
);
1883 /* There are other queues in the group, don't do group idle */
1884 if (group_idle
&& cfqq
->cfqg
->nr_cfqq
> 1)
1887 cfq_mark_cfqq_wait_request(cfqq
);
1890 sl
= cfqd
->cfq_group_idle
;
1892 sl
= cfqd
->cfq_slice_idle
;
1894 mod_timer(&cfqd
->idle_slice_timer
, jiffies
+ sl
);
1895 cfq_blkiocg_update_set_idle_time_stats(cfqg_to_blkg(cfqq
->cfqg
),
1897 cfq_log_cfqq(cfqd
, cfqq
, "arm_idle: %lu group_idle: %d", sl
,
1898 group_idle
? 1 : 0);
1902 * Move request from internal lists to the request queue dispatch list.
1904 static void cfq_dispatch_insert(struct request_queue
*q
, struct request
*rq
)
1906 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
1907 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
1909 cfq_log_cfqq(cfqd
, cfqq
, "dispatch_insert");
1911 cfqq
->next_rq
= cfq_find_next_rq(cfqd
, cfqq
, rq
);
1912 cfq_remove_request(rq
);
1914 (RQ_CFQG(rq
))->dispatched
++;
1915 elv_dispatch_sort(q
, rq
);
1917 cfqd
->rq_in_flight
[cfq_cfqq_sync(cfqq
)]++;
1918 cfqq
->nr_sectors
+= blk_rq_sectors(rq
);
1919 cfq_blkiocg_update_dispatch_stats(cfqg_to_blkg(cfqq
->cfqg
),
1920 &blkio_policy_cfq
, blk_rq_bytes(rq
),
1921 rq_data_dir(rq
), rq_is_sync(rq
));
1925 * return expired entry, or NULL to just start from scratch in rbtree
1927 static struct request
*cfq_check_fifo(struct cfq_queue
*cfqq
)
1929 struct request
*rq
= NULL
;
1931 if (cfq_cfqq_fifo_expire(cfqq
))
1934 cfq_mark_cfqq_fifo_expire(cfqq
);
1936 if (list_empty(&cfqq
->fifo
))
1939 rq
= rq_entry_fifo(cfqq
->fifo
.next
);
1940 if (time_before(jiffies
, rq_fifo_time(rq
)))
1943 cfq_log_cfqq(cfqq
->cfqd
, cfqq
, "fifo=%p", rq
);
1948 cfq_prio_to_maxrq(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
1950 const int base_rq
= cfqd
->cfq_slice_async_rq
;
1952 WARN_ON(cfqq
->ioprio
>= IOPRIO_BE_NR
);
1954 return 2 * base_rq
* (IOPRIO_BE_NR
- cfqq
->ioprio
);
1958 * Must be called with the queue_lock held.
1960 static int cfqq_process_refs(struct cfq_queue
*cfqq
)
1962 int process_refs
, io_refs
;
1964 io_refs
= cfqq
->allocated
[READ
] + cfqq
->allocated
[WRITE
];
1965 process_refs
= cfqq
->ref
- io_refs
;
1966 BUG_ON(process_refs
< 0);
1967 return process_refs
;
1970 static void cfq_setup_merge(struct cfq_queue
*cfqq
, struct cfq_queue
*new_cfqq
)
1972 int process_refs
, new_process_refs
;
1973 struct cfq_queue
*__cfqq
;
1976 * If there are no process references on the new_cfqq, then it is
1977 * unsafe to follow the ->new_cfqq chain as other cfqq's in the
1978 * chain may have dropped their last reference (not just their
1979 * last process reference).
1981 if (!cfqq_process_refs(new_cfqq
))
1984 /* Avoid a circular list and skip interim queue merges */
1985 while ((__cfqq
= new_cfqq
->new_cfqq
)) {
1991 process_refs
= cfqq_process_refs(cfqq
);
1992 new_process_refs
= cfqq_process_refs(new_cfqq
);
1994 * If the process for the cfqq has gone away, there is no
1995 * sense in merging the queues.
1997 if (process_refs
== 0 || new_process_refs
== 0)
2001 * Merge in the direction of the lesser amount of work.
2003 if (new_process_refs
>= process_refs
) {
2004 cfqq
->new_cfqq
= new_cfqq
;
2005 new_cfqq
->ref
+= process_refs
;
2007 new_cfqq
->new_cfqq
= cfqq
;
2008 cfqq
->ref
+= new_process_refs
;
2012 static enum wl_type_t
cfq_choose_wl(struct cfq_data
*cfqd
,
2013 struct cfq_group
*cfqg
, enum wl_prio_t prio
)
2015 struct cfq_queue
*queue
;
2017 bool key_valid
= false;
2018 unsigned long lowest_key
= 0;
2019 enum wl_type_t cur_best
= SYNC_NOIDLE_WORKLOAD
;
2021 for (i
= 0; i
<= SYNC_WORKLOAD
; ++i
) {
2022 /* select the one with lowest rb_key */
2023 queue
= cfq_rb_first(service_tree_for(cfqg
, prio
, i
));
2025 (!key_valid
|| time_before(queue
->rb_key
, lowest_key
))) {
2026 lowest_key
= queue
->rb_key
;
2035 static void choose_service_tree(struct cfq_data
*cfqd
, struct cfq_group
*cfqg
)
2039 struct cfq_rb_root
*st
;
2040 unsigned group_slice
;
2041 enum wl_prio_t original_prio
= cfqd
->serving_prio
;
2043 /* Choose next priority. RT > BE > IDLE */
2044 if (cfq_group_busy_queues_wl(RT_WORKLOAD
, cfqd
, cfqg
))
2045 cfqd
->serving_prio
= RT_WORKLOAD
;
2046 else if (cfq_group_busy_queues_wl(BE_WORKLOAD
, cfqd
, cfqg
))
2047 cfqd
->serving_prio
= BE_WORKLOAD
;
2049 cfqd
->serving_prio
= IDLE_WORKLOAD
;
2050 cfqd
->workload_expires
= jiffies
+ 1;
2054 if (original_prio
!= cfqd
->serving_prio
)
2058 * For RT and BE, we have to choose also the type
2059 * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
2062 st
= service_tree_for(cfqg
, cfqd
->serving_prio
, cfqd
->serving_type
);
2066 * check workload expiration, and that we still have other queues ready
2068 if (count
&& !time_after(jiffies
, cfqd
->workload_expires
))
2072 /* otherwise select new workload type */
2073 cfqd
->serving_type
=
2074 cfq_choose_wl(cfqd
, cfqg
, cfqd
->serving_prio
);
2075 st
= service_tree_for(cfqg
, cfqd
->serving_prio
, cfqd
->serving_type
);
2079 * the workload slice is computed as a fraction of target latency
2080 * proportional to the number of queues in that workload, over
2081 * all the queues in the same priority class
2083 group_slice
= cfq_group_slice(cfqd
, cfqg
);
2085 slice
= group_slice
* count
/
2086 max_t(unsigned, cfqg
->busy_queues_avg
[cfqd
->serving_prio
],
2087 cfq_group_busy_queues_wl(cfqd
->serving_prio
, cfqd
, cfqg
));
2089 if (cfqd
->serving_type
== ASYNC_WORKLOAD
) {
2093 * Async queues are currently system wide. Just taking
2094 * proportion of queues with-in same group will lead to higher
2095 * async ratio system wide as generally root group is going
2096 * to have higher weight. A more accurate thing would be to
2097 * calculate system wide asnc/sync ratio.
2099 tmp
= cfq_target_latency
* cfqg_busy_async_queues(cfqd
, cfqg
);
2100 tmp
= tmp
/cfqd
->busy_queues
;
2101 slice
= min_t(unsigned, slice
, tmp
);
2103 /* async workload slice is scaled down according to
2104 * the sync/async slice ratio. */
2105 slice
= slice
* cfqd
->cfq_slice
[0] / cfqd
->cfq_slice
[1];
2107 /* sync workload slice is at least 2 * cfq_slice_idle */
2108 slice
= max(slice
, 2 * cfqd
->cfq_slice_idle
);
2110 slice
= max_t(unsigned, slice
, CFQ_MIN_TT
);
2111 cfq_log(cfqd
, "workload slice:%d", slice
);
2112 cfqd
->workload_expires
= jiffies
+ slice
;
2115 static struct cfq_group
*cfq_get_next_cfqg(struct cfq_data
*cfqd
)
2117 struct cfq_rb_root
*st
= &cfqd
->grp_service_tree
;
2118 struct cfq_group
*cfqg
;
2120 if (RB_EMPTY_ROOT(&st
->rb
))
2122 cfqg
= cfq_rb_first_group(st
);
2123 update_min_vdisktime(st
);
2127 static void cfq_choose_cfqg(struct cfq_data
*cfqd
)
2129 struct cfq_group
*cfqg
= cfq_get_next_cfqg(cfqd
);
2131 cfqd
->serving_group
= cfqg
;
2133 /* Restore the workload type data */
2134 if (cfqg
->saved_workload_slice
) {
2135 cfqd
->workload_expires
= jiffies
+ cfqg
->saved_workload_slice
;
2136 cfqd
->serving_type
= cfqg
->saved_workload
;
2137 cfqd
->serving_prio
= cfqg
->saved_serving_prio
;
2139 cfqd
->workload_expires
= jiffies
- 1;
2141 choose_service_tree(cfqd
, cfqg
);
2145 * Select a queue for service. If we have a current active queue,
2146 * check whether to continue servicing it, or retrieve and set a new one.
2148 static struct cfq_queue
*cfq_select_queue(struct cfq_data
*cfqd
)
2150 struct cfq_queue
*cfqq
, *new_cfqq
= NULL
;
2152 cfqq
= cfqd
->active_queue
;
2156 if (!cfqd
->rq_queued
)
2160 * We were waiting for group to get backlogged. Expire the queue
2162 if (cfq_cfqq_wait_busy(cfqq
) && !RB_EMPTY_ROOT(&cfqq
->sort_list
))
2166 * The active queue has run out of time, expire it and select new.
2168 if (cfq_slice_used(cfqq
) && !cfq_cfqq_must_dispatch(cfqq
)) {
2170 * If slice had not expired at the completion of last request
2171 * we might not have turned on wait_busy flag. Don't expire
2172 * the queue yet. Allow the group to get backlogged.
2174 * The very fact that we have used the slice, that means we
2175 * have been idling all along on this queue and it should be
2176 * ok to wait for this request to complete.
2178 if (cfqq
->cfqg
->nr_cfqq
== 1 && RB_EMPTY_ROOT(&cfqq
->sort_list
)
2179 && cfqq
->dispatched
&& cfq_should_idle(cfqd
, cfqq
)) {
2183 goto check_group_idle
;
2187 * The active queue has requests and isn't expired, allow it to
2190 if (!RB_EMPTY_ROOT(&cfqq
->sort_list
))
2194 * If another queue has a request waiting within our mean seek
2195 * distance, let it run. The expire code will check for close
2196 * cooperators and put the close queue at the front of the service
2197 * tree. If possible, merge the expiring queue with the new cfqq.
2199 new_cfqq
= cfq_close_cooperator(cfqd
, cfqq
);
2201 if (!cfqq
->new_cfqq
)
2202 cfq_setup_merge(cfqq
, new_cfqq
);
2207 * No requests pending. If the active queue still has requests in
2208 * flight or is idling for a new request, allow either of these
2209 * conditions to happen (or time out) before selecting a new queue.
2211 if (timer_pending(&cfqd
->idle_slice_timer
)) {
2217 * This is a deep seek queue, but the device is much faster than
2218 * the queue can deliver, don't idle
2220 if (CFQQ_SEEKY(cfqq
) && cfq_cfqq_idle_window(cfqq
) &&
2221 (cfq_cfqq_slice_new(cfqq
) ||
2222 (cfqq
->slice_end
- jiffies
> jiffies
- cfqq
->slice_start
))) {
2223 cfq_clear_cfqq_deep(cfqq
);
2224 cfq_clear_cfqq_idle_window(cfqq
);
2227 if (cfqq
->dispatched
&& cfq_should_idle(cfqd
, cfqq
)) {
2233 * If group idle is enabled and there are requests dispatched from
2234 * this group, wait for requests to complete.
2237 if (cfqd
->cfq_group_idle
&& cfqq
->cfqg
->nr_cfqq
== 1 &&
2238 cfqq
->cfqg
->dispatched
&&
2239 !cfq_io_thinktime_big(cfqd
, &cfqq
->cfqg
->ttime
, true)) {
2245 cfq_slice_expired(cfqd
, 0);
2248 * Current queue expired. Check if we have to switch to a new
2252 cfq_choose_cfqg(cfqd
);
2254 cfqq
= cfq_set_active_queue(cfqd
, new_cfqq
);
2259 static int __cfq_forced_dispatch_cfqq(struct cfq_queue
*cfqq
)
2263 while (cfqq
->next_rq
) {
2264 cfq_dispatch_insert(cfqq
->cfqd
->queue
, cfqq
->next_rq
);
2268 BUG_ON(!list_empty(&cfqq
->fifo
));
2270 /* By default cfqq is not expired if it is empty. Do it explicitly */
2271 __cfq_slice_expired(cfqq
->cfqd
, cfqq
, 0);
2276 * Drain our current requests. Used for barriers and when switching
2277 * io schedulers on-the-fly.
2279 static int cfq_forced_dispatch(struct cfq_data
*cfqd
)
2281 struct cfq_queue
*cfqq
;
2284 /* Expire the timeslice of the current active queue first */
2285 cfq_slice_expired(cfqd
, 0);
2286 while ((cfqq
= cfq_get_next_queue_forced(cfqd
)) != NULL
) {
2287 __cfq_set_active_queue(cfqd
, cfqq
);
2288 dispatched
+= __cfq_forced_dispatch_cfqq(cfqq
);
2291 BUG_ON(cfqd
->busy_queues
);
2293 cfq_log(cfqd
, "forced_dispatch=%d", dispatched
);
2297 static inline bool cfq_slice_used_soon(struct cfq_data
*cfqd
,
2298 struct cfq_queue
*cfqq
)
2300 /* the queue hasn't finished any request, can't estimate */
2301 if (cfq_cfqq_slice_new(cfqq
))
2303 if (time_after(jiffies
+ cfqd
->cfq_slice_idle
* cfqq
->dispatched
,
2310 static bool cfq_may_dispatch(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
2312 unsigned int max_dispatch
;
2315 * Drain async requests before we start sync IO
2317 if (cfq_should_idle(cfqd
, cfqq
) && cfqd
->rq_in_flight
[BLK_RW_ASYNC
])
2321 * If this is an async queue and we have sync IO in flight, let it wait
2323 if (cfqd
->rq_in_flight
[BLK_RW_SYNC
] && !cfq_cfqq_sync(cfqq
))
2326 max_dispatch
= max_t(unsigned int, cfqd
->cfq_quantum
/ 2, 1);
2327 if (cfq_class_idle(cfqq
))
2331 * Does this cfqq already have too much IO in flight?
2333 if (cfqq
->dispatched
>= max_dispatch
) {
2334 bool promote_sync
= false;
2336 * idle queue must always only have a single IO in flight
2338 if (cfq_class_idle(cfqq
))
2342 * If there is only one sync queue
2343 * we can ignore async queue here and give the sync
2344 * queue no dispatch limit. The reason is a sync queue can
2345 * preempt async queue, limiting the sync queue doesn't make
2346 * sense. This is useful for aiostress test.
2348 if (cfq_cfqq_sync(cfqq
) && cfqd
->busy_sync_queues
== 1)
2349 promote_sync
= true;
2352 * We have other queues, don't allow more IO from this one
2354 if (cfqd
->busy_queues
> 1 && cfq_slice_used_soon(cfqd
, cfqq
) &&
2359 * Sole queue user, no limit
2361 if (cfqd
->busy_queues
== 1 || promote_sync
)
2365 * Normally we start throttling cfqq when cfq_quantum/2
2366 * requests have been dispatched. But we can drive
2367 * deeper queue depths at the beginning of slice
2368 * subjected to upper limit of cfq_quantum.
2370 max_dispatch
= cfqd
->cfq_quantum
;
2374 * Async queues must wait a bit before being allowed dispatch.
2375 * We also ramp up the dispatch depth gradually for async IO,
2376 * based on the last sync IO we serviced
2378 if (!cfq_cfqq_sync(cfqq
) && cfqd
->cfq_latency
) {
2379 unsigned long last_sync
= jiffies
- cfqd
->last_delayed_sync
;
2382 depth
= last_sync
/ cfqd
->cfq_slice
[1];
2383 if (!depth
&& !cfqq
->dispatched
)
2385 if (depth
< max_dispatch
)
2386 max_dispatch
= depth
;
2390 * If we're below the current max, allow a dispatch
2392 return cfqq
->dispatched
< max_dispatch
;
2396 * Dispatch a request from cfqq, moving them to the request queue
2399 static bool cfq_dispatch_request(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
2403 BUG_ON(RB_EMPTY_ROOT(&cfqq
->sort_list
));
2405 if (!cfq_may_dispatch(cfqd
, cfqq
))
2409 * follow expired path, else get first next available
2411 rq
= cfq_check_fifo(cfqq
);
2416 * insert request into driver dispatch list
2418 cfq_dispatch_insert(cfqd
->queue
, rq
);
2420 if (!cfqd
->active_cic
) {
2421 struct cfq_io_cq
*cic
= RQ_CIC(rq
);
2423 atomic_long_inc(&cic
->icq
.ioc
->refcount
);
2424 cfqd
->active_cic
= cic
;
2431 * Find the cfqq that we need to service and move a request from that to the
2434 static int cfq_dispatch_requests(struct request_queue
*q
, int force
)
2436 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
2437 struct cfq_queue
*cfqq
;
2439 if (!cfqd
->busy_queues
)
2442 if (unlikely(force
))
2443 return cfq_forced_dispatch(cfqd
);
2445 cfqq
= cfq_select_queue(cfqd
);
2450 * Dispatch a request from this cfqq, if it is allowed
2452 if (!cfq_dispatch_request(cfqd
, cfqq
))
2455 cfqq
->slice_dispatch
++;
2456 cfq_clear_cfqq_must_dispatch(cfqq
);
2459 * expire an async queue immediately if it has used up its slice. idle
2460 * queue always expire after 1 dispatch round.
2462 if (cfqd
->busy_queues
> 1 && ((!cfq_cfqq_sync(cfqq
) &&
2463 cfqq
->slice_dispatch
>= cfq_prio_to_maxrq(cfqd
, cfqq
)) ||
2464 cfq_class_idle(cfqq
))) {
2465 cfqq
->slice_end
= jiffies
+ 1;
2466 cfq_slice_expired(cfqd
, 0);
2469 cfq_log_cfqq(cfqd
, cfqq
, "dispatched a request");
2474 * task holds one reference to the queue, dropped when task exits. each rq
2475 * in-flight on this queue also holds a reference, dropped when rq is freed.
2477 * Each cfq queue took a reference on the parent group. Drop it now.
2478 * queue lock must be held here.
2480 static void cfq_put_queue(struct cfq_queue
*cfqq
)
2482 struct cfq_data
*cfqd
= cfqq
->cfqd
;
2483 struct cfq_group
*cfqg
;
2485 BUG_ON(cfqq
->ref
<= 0);
2491 cfq_log_cfqq(cfqd
, cfqq
, "put_queue");
2492 BUG_ON(rb_first(&cfqq
->sort_list
));
2493 BUG_ON(cfqq
->allocated
[READ
] + cfqq
->allocated
[WRITE
]);
2496 if (unlikely(cfqd
->active_queue
== cfqq
)) {
2497 __cfq_slice_expired(cfqd
, cfqq
, 0);
2498 cfq_schedule_dispatch(cfqd
);
2501 BUG_ON(cfq_cfqq_on_rr(cfqq
));
2502 kmem_cache_free(cfq_pool
, cfqq
);
2503 blkg_put(cfqg_to_blkg(cfqg
));
2506 static void cfq_put_cooperator(struct cfq_queue
*cfqq
)
2508 struct cfq_queue
*__cfqq
, *next
;
2511 * If this queue was scheduled to merge with another queue, be
2512 * sure to drop the reference taken on that queue (and others in
2513 * the merge chain). See cfq_setup_merge and cfq_merge_cfqqs.
2515 __cfqq
= cfqq
->new_cfqq
;
2517 if (__cfqq
== cfqq
) {
2518 WARN(1, "cfqq->new_cfqq loop detected\n");
2521 next
= __cfqq
->new_cfqq
;
2522 cfq_put_queue(__cfqq
);
2527 static void cfq_exit_cfqq(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
2529 if (unlikely(cfqq
== cfqd
->active_queue
)) {
2530 __cfq_slice_expired(cfqd
, cfqq
, 0);
2531 cfq_schedule_dispatch(cfqd
);
2534 cfq_put_cooperator(cfqq
);
2536 cfq_put_queue(cfqq
);
2539 static void cfq_init_icq(struct io_cq
*icq
)
2541 struct cfq_io_cq
*cic
= icq_to_cic(icq
);
2543 cic
->ttime
.last_end_request
= jiffies
;
2546 static void cfq_exit_icq(struct io_cq
*icq
)
2548 struct cfq_io_cq
*cic
= icq_to_cic(icq
);
2549 struct cfq_data
*cfqd
= cic_to_cfqd(cic
);
2551 if (cic
->cfqq
[BLK_RW_ASYNC
]) {
2552 cfq_exit_cfqq(cfqd
, cic
->cfqq
[BLK_RW_ASYNC
]);
2553 cic
->cfqq
[BLK_RW_ASYNC
] = NULL
;
2556 if (cic
->cfqq
[BLK_RW_SYNC
]) {
2557 cfq_exit_cfqq(cfqd
, cic
->cfqq
[BLK_RW_SYNC
]);
2558 cic
->cfqq
[BLK_RW_SYNC
] = NULL
;
2562 static void cfq_init_prio_data(struct cfq_queue
*cfqq
, struct io_context
*ioc
)
2564 struct task_struct
*tsk
= current
;
2567 if (!cfq_cfqq_prio_changed(cfqq
))
2570 ioprio_class
= IOPRIO_PRIO_CLASS(ioc
->ioprio
);
2571 switch (ioprio_class
) {
2573 printk(KERN_ERR
"cfq: bad prio %x\n", ioprio_class
);
2574 case IOPRIO_CLASS_NONE
:
2576 * no prio set, inherit CPU scheduling settings
2578 cfqq
->ioprio
= task_nice_ioprio(tsk
);
2579 cfqq
->ioprio_class
= task_nice_ioclass(tsk
);
2581 case IOPRIO_CLASS_RT
:
2582 cfqq
->ioprio
= task_ioprio(ioc
);
2583 cfqq
->ioprio_class
= IOPRIO_CLASS_RT
;
2585 case IOPRIO_CLASS_BE
:
2586 cfqq
->ioprio
= task_ioprio(ioc
);
2587 cfqq
->ioprio_class
= IOPRIO_CLASS_BE
;
2589 case IOPRIO_CLASS_IDLE
:
2590 cfqq
->ioprio_class
= IOPRIO_CLASS_IDLE
;
2592 cfq_clear_cfqq_idle_window(cfqq
);
2597 * keep track of original prio settings in case we have to temporarily
2598 * elevate the priority of this queue
2600 cfqq
->org_ioprio
= cfqq
->ioprio
;
2601 cfq_clear_cfqq_prio_changed(cfqq
);
2604 static void changed_ioprio(struct cfq_io_cq
*cic
)
2606 struct cfq_data
*cfqd
= cic_to_cfqd(cic
);
2607 struct cfq_queue
*cfqq
;
2609 if (unlikely(!cfqd
))
2612 cfqq
= cic
->cfqq
[BLK_RW_ASYNC
];
2614 struct cfq_queue
*new_cfqq
;
2615 new_cfqq
= cfq_get_queue(cfqd
, BLK_RW_ASYNC
, cic
->icq
.ioc
,
2618 cic
->cfqq
[BLK_RW_ASYNC
] = new_cfqq
;
2619 cfq_put_queue(cfqq
);
2623 cfqq
= cic
->cfqq
[BLK_RW_SYNC
];
2625 cfq_mark_cfqq_prio_changed(cfqq
);
2628 static void cfq_init_cfqq(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
2629 pid_t pid
, bool is_sync
)
2631 RB_CLEAR_NODE(&cfqq
->rb_node
);
2632 RB_CLEAR_NODE(&cfqq
->p_node
);
2633 INIT_LIST_HEAD(&cfqq
->fifo
);
2638 cfq_mark_cfqq_prio_changed(cfqq
);
2641 if (!cfq_class_idle(cfqq
))
2642 cfq_mark_cfqq_idle_window(cfqq
);
2643 cfq_mark_cfqq_sync(cfqq
);
2648 #ifdef CONFIG_CFQ_GROUP_IOSCHED
2649 static void changed_cgroup(struct cfq_io_cq
*cic
)
2651 struct cfq_queue
*sync_cfqq
= cic_to_cfqq(cic
, 1);
2652 struct cfq_data
*cfqd
= cic_to_cfqd(cic
);
2653 struct request_queue
*q
;
2655 if (unlikely(!cfqd
))
2662 * Drop reference to sync queue. A new sync queue will be
2663 * assigned in new group upon arrival of a fresh request.
2665 cfq_log_cfqq(cfqd
, sync_cfqq
, "changed cgroup");
2666 cic_set_cfqq(cic
, NULL
, 1);
2667 cfq_put_queue(sync_cfqq
);
2670 #endif /* CONFIG_CFQ_GROUP_IOSCHED */
2672 static struct cfq_queue
*
2673 cfq_find_alloc_queue(struct cfq_data
*cfqd
, bool is_sync
,
2674 struct io_context
*ioc
, gfp_t gfp_mask
)
2676 struct blkio_cgroup
*blkcg
;
2677 struct cfq_queue
*cfqq
, *new_cfqq
= NULL
;
2678 struct cfq_io_cq
*cic
;
2679 struct cfq_group
*cfqg
;
2684 blkcg
= task_blkio_cgroup(current
);
2686 cfqg
= cfq_lookup_create_cfqg(cfqd
, blkcg
);
2688 cic
= cfq_cic_lookup(cfqd
, ioc
);
2689 /* cic always exists here */
2690 cfqq
= cic_to_cfqq(cic
, is_sync
);
2693 * Always try a new alloc if we fell back to the OOM cfqq
2694 * originally, since it should just be a temporary situation.
2696 if (!cfqq
|| cfqq
== &cfqd
->oom_cfqq
) {
2701 } else if (gfp_mask
& __GFP_WAIT
) {
2703 spin_unlock_irq(cfqd
->queue
->queue_lock
);
2704 new_cfqq
= kmem_cache_alloc_node(cfq_pool
,
2705 gfp_mask
| __GFP_ZERO
,
2707 spin_lock_irq(cfqd
->queue
->queue_lock
);
2711 cfqq
= kmem_cache_alloc_node(cfq_pool
,
2712 gfp_mask
| __GFP_ZERO
,
2717 cfq_init_cfqq(cfqd
, cfqq
, current
->pid
, is_sync
);
2718 cfq_init_prio_data(cfqq
, ioc
);
2719 cfq_link_cfqq_cfqg(cfqq
, cfqg
);
2720 cfq_log_cfqq(cfqd
, cfqq
, "alloced");
2722 cfqq
= &cfqd
->oom_cfqq
;
2726 kmem_cache_free(cfq_pool
, new_cfqq
);
2732 static struct cfq_queue
**
2733 cfq_async_queue_prio(struct cfq_data
*cfqd
, int ioprio_class
, int ioprio
)
2735 switch (ioprio_class
) {
2736 case IOPRIO_CLASS_RT
:
2737 return &cfqd
->async_cfqq
[0][ioprio
];
2738 case IOPRIO_CLASS_BE
:
2739 return &cfqd
->async_cfqq
[1][ioprio
];
2740 case IOPRIO_CLASS_IDLE
:
2741 return &cfqd
->async_idle_cfqq
;
2747 static struct cfq_queue
*
2748 cfq_get_queue(struct cfq_data
*cfqd
, bool is_sync
, struct io_context
*ioc
,
2751 const int ioprio
= task_ioprio(ioc
);
2752 const int ioprio_class
= task_ioprio_class(ioc
);
2753 struct cfq_queue
**async_cfqq
= NULL
;
2754 struct cfq_queue
*cfqq
= NULL
;
2757 async_cfqq
= cfq_async_queue_prio(cfqd
, ioprio_class
, ioprio
);
2762 cfqq
= cfq_find_alloc_queue(cfqd
, is_sync
, ioc
, gfp_mask
);
2765 * pin the queue now that it's allocated, scheduler exit will prune it
2767 if (!is_sync
&& !(*async_cfqq
)) {
2777 __cfq_update_io_thinktime(struct cfq_ttime
*ttime
, unsigned long slice_idle
)
2779 unsigned long elapsed
= jiffies
- ttime
->last_end_request
;
2780 elapsed
= min(elapsed
, 2UL * slice_idle
);
2782 ttime
->ttime_samples
= (7*ttime
->ttime_samples
+ 256) / 8;
2783 ttime
->ttime_total
= (7*ttime
->ttime_total
+ 256*elapsed
) / 8;
2784 ttime
->ttime_mean
= (ttime
->ttime_total
+ 128) / ttime
->ttime_samples
;
2788 cfq_update_io_thinktime(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
2789 struct cfq_io_cq
*cic
)
2791 if (cfq_cfqq_sync(cfqq
)) {
2792 __cfq_update_io_thinktime(&cic
->ttime
, cfqd
->cfq_slice_idle
);
2793 __cfq_update_io_thinktime(&cfqq
->service_tree
->ttime
,
2794 cfqd
->cfq_slice_idle
);
2796 #ifdef CONFIG_CFQ_GROUP_IOSCHED
2797 __cfq_update_io_thinktime(&cfqq
->cfqg
->ttime
, cfqd
->cfq_group_idle
);
2802 cfq_update_io_seektime(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
2806 sector_t n_sec
= blk_rq_sectors(rq
);
2807 if (cfqq
->last_request_pos
) {
2808 if (cfqq
->last_request_pos
< blk_rq_pos(rq
))
2809 sdist
= blk_rq_pos(rq
) - cfqq
->last_request_pos
;
2811 sdist
= cfqq
->last_request_pos
- blk_rq_pos(rq
);
2814 cfqq
->seek_history
<<= 1;
2815 if (blk_queue_nonrot(cfqd
->queue
))
2816 cfqq
->seek_history
|= (n_sec
< CFQQ_SECT_THR_NONROT
);
2818 cfqq
->seek_history
|= (sdist
> CFQQ_SEEK_THR
);
2822 * Disable idle window if the process thinks too long or seeks so much that
2826 cfq_update_idle_window(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
2827 struct cfq_io_cq
*cic
)
2829 int old_idle
, enable_idle
;
2832 * Don't idle for async or idle io prio class
2834 if (!cfq_cfqq_sync(cfqq
) || cfq_class_idle(cfqq
))
2837 enable_idle
= old_idle
= cfq_cfqq_idle_window(cfqq
);
2839 if (cfqq
->queued
[0] + cfqq
->queued
[1] >= 4)
2840 cfq_mark_cfqq_deep(cfqq
);
2842 if (cfqq
->next_rq
&& (cfqq
->next_rq
->cmd_flags
& REQ_NOIDLE
))
2844 else if (!atomic_read(&cic
->icq
.ioc
->nr_tasks
) ||
2845 !cfqd
->cfq_slice_idle
||
2846 (!cfq_cfqq_deep(cfqq
) && CFQQ_SEEKY(cfqq
)))
2848 else if (sample_valid(cic
->ttime
.ttime_samples
)) {
2849 if (cic
->ttime
.ttime_mean
> cfqd
->cfq_slice_idle
)
2855 if (old_idle
!= enable_idle
) {
2856 cfq_log_cfqq(cfqd
, cfqq
, "idle=%d", enable_idle
);
2858 cfq_mark_cfqq_idle_window(cfqq
);
2860 cfq_clear_cfqq_idle_window(cfqq
);
2865 * Check if new_cfqq should preempt the currently active queue. Return 0 for
2866 * no or if we aren't sure, a 1 will cause a preempt.
2869 cfq_should_preempt(struct cfq_data
*cfqd
, struct cfq_queue
*new_cfqq
,
2872 struct cfq_queue
*cfqq
;
2874 cfqq
= cfqd
->active_queue
;
2878 if (cfq_class_idle(new_cfqq
))
2881 if (cfq_class_idle(cfqq
))
2885 * Don't allow a non-RT request to preempt an ongoing RT cfqq timeslice.
2887 if (cfq_class_rt(cfqq
) && !cfq_class_rt(new_cfqq
))
2891 * if the new request is sync, but the currently running queue is
2892 * not, let the sync request have priority.
2894 if (rq_is_sync(rq
) && !cfq_cfqq_sync(cfqq
))
2897 if (new_cfqq
->cfqg
!= cfqq
->cfqg
)
2900 if (cfq_slice_used(cfqq
))
2903 /* Allow preemption only if we are idling on sync-noidle tree */
2904 if (cfqd
->serving_type
== SYNC_NOIDLE_WORKLOAD
&&
2905 cfqq_type(new_cfqq
) == SYNC_NOIDLE_WORKLOAD
&&
2906 new_cfqq
->service_tree
->count
== 2 &&
2907 RB_EMPTY_ROOT(&cfqq
->sort_list
))
2911 * So both queues are sync. Let the new request get disk time if
2912 * it's a metadata request and the current queue is doing regular IO.
2914 if ((rq
->cmd_flags
& REQ_PRIO
) && !cfqq
->prio_pending
)
2918 * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
2920 if (cfq_class_rt(new_cfqq
) && !cfq_class_rt(cfqq
))
2923 /* An idle queue should not be idle now for some reason */
2924 if (RB_EMPTY_ROOT(&cfqq
->sort_list
) && !cfq_should_idle(cfqd
, cfqq
))
2927 if (!cfqd
->active_cic
|| !cfq_cfqq_wait_request(cfqq
))
2931 * if this request is as-good as one we would expect from the
2932 * current cfqq, let it preempt
2934 if (cfq_rq_close(cfqd
, cfqq
, rq
))
2941 * cfqq preempts the active queue. if we allowed preempt with no slice left,
2942 * let it have half of its nominal slice.
2944 static void cfq_preempt_queue(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
2946 enum wl_type_t old_type
= cfqq_type(cfqd
->active_queue
);
2948 cfq_log_cfqq(cfqd
, cfqq
, "preempt");
2949 cfq_slice_expired(cfqd
, 1);
2952 * workload type is changed, don't save slice, otherwise preempt
2955 if (old_type
!= cfqq_type(cfqq
))
2956 cfqq
->cfqg
->saved_workload_slice
= 0;
2959 * Put the new queue at the front of the of the current list,
2960 * so we know that it will be selected next.
2962 BUG_ON(!cfq_cfqq_on_rr(cfqq
));
2964 cfq_service_tree_add(cfqd
, cfqq
, 1);
2966 cfqq
->slice_end
= 0;
2967 cfq_mark_cfqq_slice_new(cfqq
);
2971 * Called when a new fs request (rq) is added (to cfqq). Check if there's
2972 * something we should do about it
2975 cfq_rq_enqueued(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
2978 struct cfq_io_cq
*cic
= RQ_CIC(rq
);
2981 if (rq
->cmd_flags
& REQ_PRIO
)
2982 cfqq
->prio_pending
++;
2984 cfq_update_io_thinktime(cfqd
, cfqq
, cic
);
2985 cfq_update_io_seektime(cfqd
, cfqq
, rq
);
2986 cfq_update_idle_window(cfqd
, cfqq
, cic
);
2988 cfqq
->last_request_pos
= blk_rq_pos(rq
) + blk_rq_sectors(rq
);
2990 if (cfqq
== cfqd
->active_queue
) {
2992 * Remember that we saw a request from this process, but
2993 * don't start queuing just yet. Otherwise we risk seeing lots
2994 * of tiny requests, because we disrupt the normal plugging
2995 * and merging. If the request is already larger than a single
2996 * page, let it rip immediately. For that case we assume that
2997 * merging is already done. Ditto for a busy system that
2998 * has other work pending, don't risk delaying until the
2999 * idle timer unplug to continue working.
3001 if (cfq_cfqq_wait_request(cfqq
)) {
3002 if (blk_rq_bytes(rq
) > PAGE_CACHE_SIZE
||
3003 cfqd
->busy_queues
> 1) {
3004 cfq_del_timer(cfqd
, cfqq
);
3005 cfq_clear_cfqq_wait_request(cfqq
);
3006 __blk_run_queue(cfqd
->queue
);
3008 cfq_blkiocg_update_idle_time_stats(
3009 cfqg_to_blkg(cfqq
->cfqg
),
3011 cfq_mark_cfqq_must_dispatch(cfqq
);
3014 } else if (cfq_should_preempt(cfqd
, cfqq
, rq
)) {
3016 * not the active queue - expire current slice if it is
3017 * idle and has expired it's mean thinktime or this new queue
3018 * has some old slice time left and is of higher priority or
3019 * this new queue is RT and the current one is BE
3021 cfq_preempt_queue(cfqd
, cfqq
);
3022 __blk_run_queue(cfqd
->queue
);
3026 static void cfq_insert_request(struct request_queue
*q
, struct request
*rq
)
3028 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
3029 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
3031 cfq_log_cfqq(cfqd
, cfqq
, "insert_request");
3032 cfq_init_prio_data(cfqq
, RQ_CIC(rq
)->icq
.ioc
);
3034 rq_set_fifo_time(rq
, jiffies
+ cfqd
->cfq_fifo_expire
[rq_is_sync(rq
)]);
3035 list_add_tail(&rq
->queuelist
, &cfqq
->fifo
);
3037 cfq_blkiocg_update_io_add_stats(cfqg_to_blkg(RQ_CFQG(rq
)),
3039 cfqg_to_blkg(cfqd
->serving_group
),
3040 rq_data_dir(rq
), rq_is_sync(rq
));
3041 cfq_rq_enqueued(cfqd
, cfqq
, rq
);
3045 * Update hw_tag based on peak queue depth over 50 samples under
3048 static void cfq_update_hw_tag(struct cfq_data
*cfqd
)
3050 struct cfq_queue
*cfqq
= cfqd
->active_queue
;
3052 if (cfqd
->rq_in_driver
> cfqd
->hw_tag_est_depth
)
3053 cfqd
->hw_tag_est_depth
= cfqd
->rq_in_driver
;
3055 if (cfqd
->hw_tag
== 1)
3058 if (cfqd
->rq_queued
<= CFQ_HW_QUEUE_MIN
&&
3059 cfqd
->rq_in_driver
<= CFQ_HW_QUEUE_MIN
)
3063 * If active queue hasn't enough requests and can idle, cfq might not
3064 * dispatch sufficient requests to hardware. Don't zero hw_tag in this
3067 if (cfqq
&& cfq_cfqq_idle_window(cfqq
) &&
3068 cfqq
->dispatched
+ cfqq
->queued
[0] + cfqq
->queued
[1] <
3069 CFQ_HW_QUEUE_MIN
&& cfqd
->rq_in_driver
< CFQ_HW_QUEUE_MIN
)
3072 if (cfqd
->hw_tag_samples
++ < 50)
3075 if (cfqd
->hw_tag_est_depth
>= CFQ_HW_QUEUE_MIN
)
3081 static bool cfq_should_wait_busy(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
3083 struct cfq_io_cq
*cic
= cfqd
->active_cic
;
3085 /* If the queue already has requests, don't wait */
3086 if (!RB_EMPTY_ROOT(&cfqq
->sort_list
))
3089 /* If there are other queues in the group, don't wait */
3090 if (cfqq
->cfqg
->nr_cfqq
> 1)
3093 /* the only queue in the group, but think time is big */
3094 if (cfq_io_thinktime_big(cfqd
, &cfqq
->cfqg
->ttime
, true))
3097 if (cfq_slice_used(cfqq
))
3100 /* if slice left is less than think time, wait busy */
3101 if (cic
&& sample_valid(cic
->ttime
.ttime_samples
)
3102 && (cfqq
->slice_end
- jiffies
< cic
->ttime
.ttime_mean
))
3106 * If think times is less than a jiffy than ttime_mean=0 and above
3107 * will not be true. It might happen that slice has not expired yet
3108 * but will expire soon (4-5 ns) during select_queue(). To cover the
3109 * case where think time is less than a jiffy, mark the queue wait
3110 * busy if only 1 jiffy is left in the slice.
3112 if (cfqq
->slice_end
- jiffies
== 1)
3118 static void cfq_completed_request(struct request_queue
*q
, struct request
*rq
)
3120 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
3121 struct cfq_data
*cfqd
= cfqq
->cfqd
;
3122 const int sync
= rq_is_sync(rq
);
3126 cfq_log_cfqq(cfqd
, cfqq
, "complete rqnoidle %d",
3127 !!(rq
->cmd_flags
& REQ_NOIDLE
));
3129 cfq_update_hw_tag(cfqd
);
3131 WARN_ON(!cfqd
->rq_in_driver
);
3132 WARN_ON(!cfqq
->dispatched
);
3133 cfqd
->rq_in_driver
--;
3135 (RQ_CFQG(rq
))->dispatched
--;
3136 cfq_blkiocg_update_completion_stats(cfqg_to_blkg(cfqq
->cfqg
),
3137 &blkio_policy_cfq
, rq_start_time_ns(rq
),
3138 rq_io_start_time_ns(rq
), rq_data_dir(rq
),
3141 cfqd
->rq_in_flight
[cfq_cfqq_sync(cfqq
)]--;
3144 struct cfq_rb_root
*service_tree
;
3146 RQ_CIC(rq
)->ttime
.last_end_request
= now
;
3148 if (cfq_cfqq_on_rr(cfqq
))
3149 service_tree
= cfqq
->service_tree
;
3151 service_tree
= service_tree_for(cfqq
->cfqg
,
3152 cfqq_prio(cfqq
), cfqq_type(cfqq
));
3153 service_tree
->ttime
.last_end_request
= now
;
3154 if (!time_after(rq
->start_time
+ cfqd
->cfq_fifo_expire
[1], now
))
3155 cfqd
->last_delayed_sync
= now
;
3158 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3159 cfqq
->cfqg
->ttime
.last_end_request
= now
;
3163 * If this is the active queue, check if it needs to be expired,
3164 * or if we want to idle in case it has no pending requests.
3166 if (cfqd
->active_queue
== cfqq
) {
3167 const bool cfqq_empty
= RB_EMPTY_ROOT(&cfqq
->sort_list
);
3169 if (cfq_cfqq_slice_new(cfqq
)) {
3170 cfq_set_prio_slice(cfqd
, cfqq
);
3171 cfq_clear_cfqq_slice_new(cfqq
);
3175 * Should we wait for next request to come in before we expire
3178 if (cfq_should_wait_busy(cfqd
, cfqq
)) {
3179 unsigned long extend_sl
= cfqd
->cfq_slice_idle
;
3180 if (!cfqd
->cfq_slice_idle
)
3181 extend_sl
= cfqd
->cfq_group_idle
;
3182 cfqq
->slice_end
= jiffies
+ extend_sl
;
3183 cfq_mark_cfqq_wait_busy(cfqq
);
3184 cfq_log_cfqq(cfqd
, cfqq
, "will busy wait");
3188 * Idling is not enabled on:
3190 * - idle-priority queues
3192 * - queues with still some requests queued
3193 * - when there is a close cooperator
3195 if (cfq_slice_used(cfqq
) || cfq_class_idle(cfqq
))
3196 cfq_slice_expired(cfqd
, 1);
3197 else if (sync
&& cfqq_empty
&&
3198 !cfq_close_cooperator(cfqd
, cfqq
)) {
3199 cfq_arm_slice_timer(cfqd
);
3203 if (!cfqd
->rq_in_driver
)
3204 cfq_schedule_dispatch(cfqd
);
3207 static inline int __cfq_may_queue(struct cfq_queue
*cfqq
)
3209 if (cfq_cfqq_wait_request(cfqq
) && !cfq_cfqq_must_alloc_slice(cfqq
)) {
3210 cfq_mark_cfqq_must_alloc_slice(cfqq
);
3211 return ELV_MQUEUE_MUST
;
3214 return ELV_MQUEUE_MAY
;
3217 static int cfq_may_queue(struct request_queue
*q
, int rw
)
3219 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
3220 struct task_struct
*tsk
= current
;
3221 struct cfq_io_cq
*cic
;
3222 struct cfq_queue
*cfqq
;
3225 * don't force setup of a queue from here, as a call to may_queue
3226 * does not necessarily imply that a request actually will be queued.
3227 * so just lookup a possibly existing queue, or return 'may queue'
3230 cic
= cfq_cic_lookup(cfqd
, tsk
->io_context
);
3232 return ELV_MQUEUE_MAY
;
3234 cfqq
= cic_to_cfqq(cic
, rw_is_sync(rw
));
3236 cfq_init_prio_data(cfqq
, cic
->icq
.ioc
);
3238 return __cfq_may_queue(cfqq
);
3241 return ELV_MQUEUE_MAY
;
3245 * queue lock held here
3247 static void cfq_put_request(struct request
*rq
)
3249 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
3252 const int rw
= rq_data_dir(rq
);
3254 BUG_ON(!cfqq
->allocated
[rw
]);
3255 cfqq
->allocated
[rw
]--;
3257 /* Put down rq reference on cfqg */
3258 blkg_put(cfqg_to_blkg(RQ_CFQG(rq
)));
3259 rq
->elv
.priv
[0] = NULL
;
3260 rq
->elv
.priv
[1] = NULL
;
3262 cfq_put_queue(cfqq
);
3266 static struct cfq_queue
*
3267 cfq_merge_cfqqs(struct cfq_data
*cfqd
, struct cfq_io_cq
*cic
,
3268 struct cfq_queue
*cfqq
)
3270 cfq_log_cfqq(cfqd
, cfqq
, "merging with queue %p", cfqq
->new_cfqq
);
3271 cic_set_cfqq(cic
, cfqq
->new_cfqq
, 1);
3272 cfq_mark_cfqq_coop(cfqq
->new_cfqq
);
3273 cfq_put_queue(cfqq
);
3274 return cic_to_cfqq(cic
, 1);
3278 * Returns NULL if a new cfqq should be allocated, or the old cfqq if this
3279 * was the last process referring to said cfqq.
3281 static struct cfq_queue
*
3282 split_cfqq(struct cfq_io_cq
*cic
, struct cfq_queue
*cfqq
)
3284 if (cfqq_process_refs(cfqq
) == 1) {
3285 cfqq
->pid
= current
->pid
;
3286 cfq_clear_cfqq_coop(cfqq
);
3287 cfq_clear_cfqq_split_coop(cfqq
);
3291 cic_set_cfqq(cic
, NULL
, 1);
3293 cfq_put_cooperator(cfqq
);
3295 cfq_put_queue(cfqq
);
3299 * Allocate cfq data structures associated with this request.
3302 cfq_set_request(struct request_queue
*q
, struct request
*rq
, gfp_t gfp_mask
)
3304 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
3305 struct cfq_io_cq
*cic
= icq_to_cic(rq
->elv
.icq
);
3306 const int rw
= rq_data_dir(rq
);
3307 const bool is_sync
= rq_is_sync(rq
);
3308 struct cfq_queue
*cfqq
;
3309 unsigned int changed
;
3311 might_sleep_if(gfp_mask
& __GFP_WAIT
);
3313 spin_lock_irq(q
->queue_lock
);
3315 /* handle changed notifications */
3316 changed
= icq_get_changed(&cic
->icq
);
3317 if (unlikely(changed
& ICQ_IOPRIO_CHANGED
))
3318 changed_ioprio(cic
);
3319 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3320 if (unlikely(changed
& ICQ_CGROUP_CHANGED
))
3321 changed_cgroup(cic
);
3325 cfqq
= cic_to_cfqq(cic
, is_sync
);
3326 if (!cfqq
|| cfqq
== &cfqd
->oom_cfqq
) {
3327 cfqq
= cfq_get_queue(cfqd
, is_sync
, cic
->icq
.ioc
, gfp_mask
);
3328 cic_set_cfqq(cic
, cfqq
, is_sync
);
3331 * If the queue was seeky for too long, break it apart.
3333 if (cfq_cfqq_coop(cfqq
) && cfq_cfqq_split_coop(cfqq
)) {
3334 cfq_log_cfqq(cfqd
, cfqq
, "breaking apart cfqq");
3335 cfqq
= split_cfqq(cic
, cfqq
);
3341 * Check to see if this queue is scheduled to merge with
3342 * another, closely cooperating queue. The merging of
3343 * queues happens here as it must be done in process context.
3344 * The reference on new_cfqq was taken in merge_cfqqs.
3347 cfqq
= cfq_merge_cfqqs(cfqd
, cic
, cfqq
);
3350 cfqq
->allocated
[rw
]++;
3353 blkg_get(cfqg_to_blkg(cfqq
->cfqg
));
3354 rq
->elv
.priv
[0] = cfqq
;
3355 rq
->elv
.priv
[1] = cfqq
->cfqg
;
3356 spin_unlock_irq(q
->queue_lock
);
3360 static void cfq_kick_queue(struct work_struct
*work
)
3362 struct cfq_data
*cfqd
=
3363 container_of(work
, struct cfq_data
, unplug_work
);
3364 struct request_queue
*q
= cfqd
->queue
;
3366 spin_lock_irq(q
->queue_lock
);
3367 __blk_run_queue(cfqd
->queue
);
3368 spin_unlock_irq(q
->queue_lock
);
3372 * Timer running if the active_queue is currently idling inside its time slice
3374 static void cfq_idle_slice_timer(unsigned long data
)
3376 struct cfq_data
*cfqd
= (struct cfq_data
*) data
;
3377 struct cfq_queue
*cfqq
;
3378 unsigned long flags
;
3381 cfq_log(cfqd
, "idle timer fired");
3383 spin_lock_irqsave(cfqd
->queue
->queue_lock
, flags
);
3385 cfqq
= cfqd
->active_queue
;
3390 * We saw a request before the queue expired, let it through
3392 if (cfq_cfqq_must_dispatch(cfqq
))
3398 if (cfq_slice_used(cfqq
))
3402 * only expire and reinvoke request handler, if there are
3403 * other queues with pending requests
3405 if (!cfqd
->busy_queues
)
3409 * not expired and it has a request pending, let it dispatch
3411 if (!RB_EMPTY_ROOT(&cfqq
->sort_list
))
3415 * Queue depth flag is reset only when the idle didn't succeed
3417 cfq_clear_cfqq_deep(cfqq
);
3420 cfq_slice_expired(cfqd
, timed_out
);
3422 cfq_schedule_dispatch(cfqd
);
3424 spin_unlock_irqrestore(cfqd
->queue
->queue_lock
, flags
);
3427 static void cfq_shutdown_timer_wq(struct cfq_data
*cfqd
)
3429 del_timer_sync(&cfqd
->idle_slice_timer
);
3430 cancel_work_sync(&cfqd
->unplug_work
);
3433 static void cfq_put_async_queues(struct cfq_data
*cfqd
)
3437 for (i
= 0; i
< IOPRIO_BE_NR
; i
++) {
3438 if (cfqd
->async_cfqq
[0][i
])
3439 cfq_put_queue(cfqd
->async_cfqq
[0][i
]);
3440 if (cfqd
->async_cfqq
[1][i
])
3441 cfq_put_queue(cfqd
->async_cfqq
[1][i
]);
3444 if (cfqd
->async_idle_cfqq
)
3445 cfq_put_queue(cfqd
->async_idle_cfqq
);
3448 static void cfq_exit_queue(struct elevator_queue
*e
)
3450 struct cfq_data
*cfqd
= e
->elevator_data
;
3451 struct request_queue
*q
= cfqd
->queue
;
3454 cfq_shutdown_timer_wq(cfqd
);
3456 spin_lock_irq(q
->queue_lock
);
3458 if (cfqd
->active_queue
)
3459 __cfq_slice_expired(cfqd
, cfqd
->active_queue
, 0);
3461 cfq_put_async_queues(cfqd
);
3463 spin_unlock_irq(q
->queue_lock
);
3465 #ifdef CONFIG_BLK_CGROUP
3467 * If there are groups which we could not unlink from blkcg list,
3468 * wait for a rcu period for them to be freed.
3470 spin_lock_irq(q
->queue_lock
);
3472 spin_unlock_irq(q
->queue_lock
);
3474 cfq_shutdown_timer_wq(cfqd
);
3477 * Wait for cfqg->blkg->key accessors to exit their grace periods.
3478 * Do this wait only if there are other unlinked groups out
3479 * there. This can happen if cgroup deletion path claimed the
3480 * responsibility of cleaning up a group before queue cleanup code
3483 * Do not call synchronize_rcu() unconditionally as there are drivers
3484 * which create/delete request queue hundreds of times during scan/boot
3485 * and synchronize_rcu() can take significant time and slow down boot.
3490 #ifndef CONFIG_CFQ_GROUP_IOSCHED
3491 kfree(cfqd
->root_group
);
3493 update_root_blkg_pd(q
, BLKIO_POLICY_PROP
);
3497 static int cfq_init_queue(struct request_queue
*q
)
3499 struct cfq_data
*cfqd
;
3500 struct blkio_group
*blkg __maybe_unused
;
3503 cfqd
= kmalloc_node(sizeof(*cfqd
), GFP_KERNEL
| __GFP_ZERO
, q
->node
);
3508 q
->elevator
->elevator_data
= cfqd
;
3510 /* Init root service tree */
3511 cfqd
->grp_service_tree
= CFQ_RB_ROOT
;
3513 /* Init root group and prefer root group over other groups by default */
3514 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3516 spin_lock_irq(q
->queue_lock
);
3518 blkg
= blkg_lookup_create(&blkio_root_cgroup
, q
, BLKIO_POLICY_PROP
,
3521 cfqd
->root_group
= blkg_to_cfqg(blkg
);
3523 spin_unlock_irq(q
->queue_lock
);
3526 cfqd
->root_group
= kzalloc_node(sizeof(*cfqd
->root_group
),
3527 GFP_KERNEL
, cfqd
->queue
->node
);
3528 if (cfqd
->root_group
)
3529 cfq_init_cfqg_base(cfqd
->root_group
);
3531 if (!cfqd
->root_group
) {
3536 cfqd
->root_group
->weight
= 2*BLKIO_WEIGHT_DEFAULT
;
3539 * Not strictly needed (since RB_ROOT just clears the node and we
3540 * zeroed cfqd on alloc), but better be safe in case someone decides
3541 * to add magic to the rb code
3543 for (i
= 0; i
< CFQ_PRIO_LISTS
; i
++)
3544 cfqd
->prio_trees
[i
] = RB_ROOT
;
3547 * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
3548 * Grab a permanent reference to it, so that the normal code flow
3549 * will not attempt to free it. oom_cfqq is linked to root_group
3550 * but shouldn't hold a reference as it'll never be unlinked. Lose
3551 * the reference from linking right away.
3553 cfq_init_cfqq(cfqd
, &cfqd
->oom_cfqq
, 1, 0);
3554 cfqd
->oom_cfqq
.ref
++;
3556 spin_lock_irq(q
->queue_lock
);
3557 cfq_link_cfqq_cfqg(&cfqd
->oom_cfqq
, cfqd
->root_group
);
3558 blkg_put(cfqg_to_blkg(cfqd
->root_group
));
3559 spin_unlock_irq(q
->queue_lock
);
3561 init_timer(&cfqd
->idle_slice_timer
);
3562 cfqd
->idle_slice_timer
.function
= cfq_idle_slice_timer
;
3563 cfqd
->idle_slice_timer
.data
= (unsigned long) cfqd
;
3565 INIT_WORK(&cfqd
->unplug_work
, cfq_kick_queue
);
3567 cfqd
->cfq_quantum
= cfq_quantum
;
3568 cfqd
->cfq_fifo_expire
[0] = cfq_fifo_expire
[0];
3569 cfqd
->cfq_fifo_expire
[1] = cfq_fifo_expire
[1];
3570 cfqd
->cfq_back_max
= cfq_back_max
;
3571 cfqd
->cfq_back_penalty
= cfq_back_penalty
;
3572 cfqd
->cfq_slice
[0] = cfq_slice_async
;
3573 cfqd
->cfq_slice
[1] = cfq_slice_sync
;
3574 cfqd
->cfq_slice_async_rq
= cfq_slice_async_rq
;
3575 cfqd
->cfq_slice_idle
= cfq_slice_idle
;
3576 cfqd
->cfq_group_idle
= cfq_group_idle
;
3577 cfqd
->cfq_latency
= 1;
3580 * we optimistically start assuming sync ops weren't delayed in last
3581 * second, in order to have larger depth for async operations.
3583 cfqd
->last_delayed_sync
= jiffies
- HZ
;
3588 * sysfs parts below -->
3591 cfq_var_show(unsigned int var
, char *page
)
3593 return sprintf(page
, "%d\n", var
);
3597 cfq_var_store(unsigned int *var
, const char *page
, size_t count
)
3599 char *p
= (char *) page
;
3601 *var
= simple_strtoul(p
, &p
, 10);
3605 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
3606 static ssize_t __FUNC(struct elevator_queue *e, char *page) \
3608 struct cfq_data *cfqd = e->elevator_data; \
3609 unsigned int __data = __VAR; \
3611 __data = jiffies_to_msecs(__data); \
3612 return cfq_var_show(__data, (page)); \
3614 SHOW_FUNCTION(cfq_quantum_show
, cfqd
->cfq_quantum
, 0);
3615 SHOW_FUNCTION(cfq_fifo_expire_sync_show
, cfqd
->cfq_fifo_expire
[1], 1);
3616 SHOW_FUNCTION(cfq_fifo_expire_async_show
, cfqd
->cfq_fifo_expire
[0], 1);
3617 SHOW_FUNCTION(cfq_back_seek_max_show
, cfqd
->cfq_back_max
, 0);
3618 SHOW_FUNCTION(cfq_back_seek_penalty_show
, cfqd
->cfq_back_penalty
, 0);
3619 SHOW_FUNCTION(cfq_slice_idle_show
, cfqd
->cfq_slice_idle
, 1);
3620 SHOW_FUNCTION(cfq_group_idle_show
, cfqd
->cfq_group_idle
, 1);
3621 SHOW_FUNCTION(cfq_slice_sync_show
, cfqd
->cfq_slice
[1], 1);
3622 SHOW_FUNCTION(cfq_slice_async_show
, cfqd
->cfq_slice
[0], 1);
3623 SHOW_FUNCTION(cfq_slice_async_rq_show
, cfqd
->cfq_slice_async_rq
, 0);
3624 SHOW_FUNCTION(cfq_low_latency_show
, cfqd
->cfq_latency
, 0);
3625 #undef SHOW_FUNCTION
3627 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
3628 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
3630 struct cfq_data *cfqd = e->elevator_data; \
3631 unsigned int __data; \
3632 int ret = cfq_var_store(&__data, (page), count); \
3633 if (__data < (MIN)) \
3635 else if (__data > (MAX)) \
3638 *(__PTR) = msecs_to_jiffies(__data); \
3640 *(__PTR) = __data; \
3643 STORE_FUNCTION(cfq_quantum_store
, &cfqd
->cfq_quantum
, 1, UINT_MAX
, 0);
3644 STORE_FUNCTION(cfq_fifo_expire_sync_store
, &cfqd
->cfq_fifo_expire
[1], 1,
3646 STORE_FUNCTION(cfq_fifo_expire_async_store
, &cfqd
->cfq_fifo_expire
[0], 1,
3648 STORE_FUNCTION(cfq_back_seek_max_store
, &cfqd
->cfq_back_max
, 0, UINT_MAX
, 0);
3649 STORE_FUNCTION(cfq_back_seek_penalty_store
, &cfqd
->cfq_back_penalty
, 1,
3651 STORE_FUNCTION(cfq_slice_idle_store
, &cfqd
->cfq_slice_idle
, 0, UINT_MAX
, 1);
3652 STORE_FUNCTION(cfq_group_idle_store
, &cfqd
->cfq_group_idle
, 0, UINT_MAX
, 1);
3653 STORE_FUNCTION(cfq_slice_sync_store
, &cfqd
->cfq_slice
[1], 1, UINT_MAX
, 1);
3654 STORE_FUNCTION(cfq_slice_async_store
, &cfqd
->cfq_slice
[0], 1, UINT_MAX
, 1);
3655 STORE_FUNCTION(cfq_slice_async_rq_store
, &cfqd
->cfq_slice_async_rq
, 1,
3657 STORE_FUNCTION(cfq_low_latency_store
, &cfqd
->cfq_latency
, 0, 1, 0);
3658 #undef STORE_FUNCTION
3660 #define CFQ_ATTR(name) \
3661 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
3663 static struct elv_fs_entry cfq_attrs
[] = {
3665 CFQ_ATTR(fifo_expire_sync
),
3666 CFQ_ATTR(fifo_expire_async
),
3667 CFQ_ATTR(back_seek_max
),
3668 CFQ_ATTR(back_seek_penalty
),
3669 CFQ_ATTR(slice_sync
),
3670 CFQ_ATTR(slice_async
),
3671 CFQ_ATTR(slice_async_rq
),
3672 CFQ_ATTR(slice_idle
),
3673 CFQ_ATTR(group_idle
),
3674 CFQ_ATTR(low_latency
),
3678 static struct elevator_type iosched_cfq
= {
3680 .elevator_merge_fn
= cfq_merge
,
3681 .elevator_merged_fn
= cfq_merged_request
,
3682 .elevator_merge_req_fn
= cfq_merged_requests
,
3683 .elevator_allow_merge_fn
= cfq_allow_merge
,
3684 .elevator_bio_merged_fn
= cfq_bio_merged
,
3685 .elevator_dispatch_fn
= cfq_dispatch_requests
,
3686 .elevator_add_req_fn
= cfq_insert_request
,
3687 .elevator_activate_req_fn
= cfq_activate_request
,
3688 .elevator_deactivate_req_fn
= cfq_deactivate_request
,
3689 .elevator_completed_req_fn
= cfq_completed_request
,
3690 .elevator_former_req_fn
= elv_rb_former_request
,
3691 .elevator_latter_req_fn
= elv_rb_latter_request
,
3692 .elevator_init_icq_fn
= cfq_init_icq
,
3693 .elevator_exit_icq_fn
= cfq_exit_icq
,
3694 .elevator_set_req_fn
= cfq_set_request
,
3695 .elevator_put_req_fn
= cfq_put_request
,
3696 .elevator_may_queue_fn
= cfq_may_queue
,
3697 .elevator_init_fn
= cfq_init_queue
,
3698 .elevator_exit_fn
= cfq_exit_queue
,
3700 .icq_size
= sizeof(struct cfq_io_cq
),
3701 .icq_align
= __alignof__(struct cfq_io_cq
),
3702 .elevator_attrs
= cfq_attrs
,
3703 .elevator_name
= "cfq",
3704 .elevator_owner
= THIS_MODULE
,
3707 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3708 static struct blkio_policy_type blkio_policy_cfq
= {
3710 .blkio_init_group_fn
= cfq_init_blkio_group
,
3711 .blkio_update_group_weight_fn
= cfq_update_blkio_group_weight
,
3713 .plid
= BLKIO_POLICY_PROP
,
3714 .pdata_size
= sizeof(struct cfq_group
),
3718 static int __init
cfq_init(void)
3723 * could be 0 on HZ < 1000 setups
3725 if (!cfq_slice_async
)
3726 cfq_slice_async
= 1;
3727 if (!cfq_slice_idle
)
3730 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3731 if (!cfq_group_idle
)
3736 cfq_pool
= KMEM_CACHE(cfq_queue
, 0);
3740 ret
= elv_register(&iosched_cfq
);
3742 kmem_cache_destroy(cfq_pool
);
3746 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3747 blkio_policy_register(&blkio_policy_cfq
);
3752 static void __exit
cfq_exit(void)
3754 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3755 blkio_policy_unregister(&blkio_policy_cfq
);
3757 elv_unregister(&iosched_cfq
);
3758 kmem_cache_destroy(cfq_pool
);
3761 module_init(cfq_init
);
3762 module_exit(cfq_exit
);
3764 MODULE_AUTHOR("Jens Axboe");
3765 MODULE_LICENSE("GPL");
3766 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");