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/blkdev.h>
11 #include <linux/elevator.h>
12 #include <linux/jiffies.h>
13 #include <linux/rbtree.h>
14 #include <linux/ioprio.h>
15 #include <linux/blktrace_api.h>
20 /* max queue in one round of service */
21 static const int cfq_quantum
= 4;
22 static const int cfq_fifo_expire
[2] = { HZ
/ 4, HZ
/ 8 };
23 /* maximum backwards seek, in KiB */
24 static const int cfq_back_max
= 16 * 1024;
25 /* penalty of a backwards seek */
26 static const int cfq_back_penalty
= 2;
27 static const int cfq_slice_sync
= HZ
/ 10;
28 static int cfq_slice_async
= HZ
/ 25;
29 static const int cfq_slice_async_rq
= 2;
30 static int cfq_slice_idle
= HZ
/ 125;
31 static const int cfq_target_latency
= HZ
* 3/10; /* 300 ms */
32 static const int cfq_hist_divisor
= 4;
35 * offset from end of service tree
37 #define CFQ_IDLE_DELAY (HZ / 5)
40 * below this threshold, we consider thinktime immediate
42 #define CFQ_MIN_TT (2)
45 * Allow merged cfqqs to perform this amount of seeky I/O before
46 * deciding to break the queues up again.
48 #define CFQQ_COOP_TOUT (HZ)
50 #define CFQ_SLICE_SCALE (5)
51 #define CFQ_HW_QUEUE_MIN (5)
54 ((struct cfq_io_context *) (rq)->elevator_private)
55 #define RQ_CFQQ(rq) (struct cfq_queue *) ((rq)->elevator_private2)
57 static struct kmem_cache
*cfq_pool
;
58 static struct kmem_cache
*cfq_ioc_pool
;
60 static DEFINE_PER_CPU(unsigned long, cfq_ioc_count
);
61 static struct completion
*ioc_gone
;
62 static DEFINE_SPINLOCK(ioc_gone_lock
);
64 #define CFQ_PRIO_LISTS IOPRIO_BE_NR
65 #define cfq_class_idle(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
66 #define cfq_class_rt(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
68 #define sample_valid(samples) ((samples) > 80)
71 * Most of our rbtree usage is for sorting with min extraction, so
72 * if we cache the leftmost node we don't have to walk down the tree
73 * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should
74 * move this into the elevator for the rq sorting as well.
81 #define CFQ_RB_ROOT (struct cfq_rb_root) { RB_ROOT, NULL, 0, }
84 * Per process-grouping structure
89 /* various state flags, see below */
92 struct cfq_data
*cfqd
;
93 /* service_tree member */
94 struct rb_node rb_node
;
95 /* service_tree key */
97 /* prio tree member */
98 struct rb_node p_node
;
99 /* prio tree root we belong to, if any */
100 struct rb_root
*p_root
;
101 /* sorted list of pending requests */
102 struct rb_root sort_list
;
103 /* if fifo isn't expired, next request to serve */
104 struct request
*next_rq
;
105 /* requests queued in sort_list */
107 /* currently allocated requests */
109 /* fifo list of requests in sort_list */
110 struct list_head fifo
;
112 unsigned long slice_end
;
114 unsigned int slice_dispatch
;
116 /* pending metadata requests */
118 /* number of requests that are on the dispatch list or inside driver */
121 /* io prio of this group */
122 unsigned short ioprio
, org_ioprio
;
123 unsigned short ioprio_class
, org_ioprio_class
;
125 unsigned int seek_samples
;
128 sector_t last_request_pos
;
129 unsigned long seeky_start
;
133 struct cfq_rb_root
*service_tree
;
134 struct cfq_queue
*new_cfqq
;
138 * First index in the service_trees.
139 * IDLE is handled separately, so it has negative index
148 * Second index in the service_trees.
152 SYNC_NOIDLE_WORKLOAD
= 1,
158 * Per block device queue structure
161 struct request_queue
*queue
;
164 * rr lists of queues with requests, onle rr for each priority class.
165 * Counts are embedded in the cfq_rb_root
167 struct cfq_rb_root service_trees
[2][3];
168 struct cfq_rb_root service_tree_idle
;
170 * The priority currently being served
172 enum wl_prio_t serving_prio
;
173 enum wl_type_t serving_type
;
174 unsigned long workload_expires
;
175 bool noidle_tree_requires_idle
;
178 * Each priority tree is sorted by next_request position. These
179 * trees are used when determining if two or more queues are
180 * interleaving requests (see cfq_close_cooperator).
182 struct rb_root prio_trees
[CFQ_PRIO_LISTS
];
184 unsigned int busy_queues
;
185 unsigned int busy_queues_avg
[2];
191 * queue-depth detection
197 * -1 => indeterminate, (cfq will behave as if NCQ is present, to allow better detection)
198 * 1 => NCQ is present (hw_tag_est_depth is the estimated max depth)
201 int hw_tag_est_depth
;
202 unsigned int hw_tag_samples
;
205 * idle window management
207 struct timer_list idle_slice_timer
;
208 struct work_struct unplug_work
;
210 struct cfq_queue
*active_queue
;
211 struct cfq_io_context
*active_cic
;
214 * async queue for each priority case
216 struct cfq_queue
*async_cfqq
[2][IOPRIO_BE_NR
];
217 struct cfq_queue
*async_idle_cfqq
;
219 sector_t last_position
;
222 * tunables, see top of file
224 unsigned int cfq_quantum
;
225 unsigned int cfq_fifo_expire
[2];
226 unsigned int cfq_back_penalty
;
227 unsigned int cfq_back_max
;
228 unsigned int cfq_slice
[2];
229 unsigned int cfq_slice_async_rq
;
230 unsigned int cfq_slice_idle
;
231 unsigned int cfq_latency
;
233 struct list_head cic_list
;
236 * Fallback dummy cfqq for extreme OOM conditions
238 struct cfq_queue oom_cfqq
;
240 unsigned long last_end_sync_rq
;
243 static struct cfq_rb_root
*service_tree_for(enum wl_prio_t prio
,
245 struct cfq_data
*cfqd
)
247 if (prio
== IDLE_WORKLOAD
)
248 return &cfqd
->service_tree_idle
;
250 return &cfqd
->service_trees
[prio
][type
];
253 enum cfqq_state_flags
{
254 CFQ_CFQQ_FLAG_on_rr
= 0, /* on round-robin busy list */
255 CFQ_CFQQ_FLAG_wait_request
, /* waiting for a request */
256 CFQ_CFQQ_FLAG_must_dispatch
, /* must be allowed a dispatch */
257 CFQ_CFQQ_FLAG_must_alloc_slice
, /* per-slice must_alloc flag */
258 CFQ_CFQQ_FLAG_fifo_expire
, /* FIFO checked in this slice */
259 CFQ_CFQQ_FLAG_idle_window
, /* slice idling enabled */
260 CFQ_CFQQ_FLAG_prio_changed
, /* task priority has changed */
261 CFQ_CFQQ_FLAG_slice_new
, /* no requests dispatched in slice */
262 CFQ_CFQQ_FLAG_sync
, /* synchronous queue */
263 CFQ_CFQQ_FLAG_coop
, /* cfqq is shared */
264 CFQ_CFQQ_FLAG_deep
, /* sync cfqq experienced large depth */
267 #define CFQ_CFQQ_FNS(name) \
268 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \
270 (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name); \
272 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \
274 (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \
276 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \
278 return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \
282 CFQ_CFQQ_FNS(wait_request
);
283 CFQ_CFQQ_FNS(must_dispatch
);
284 CFQ_CFQQ_FNS(must_alloc_slice
);
285 CFQ_CFQQ_FNS(fifo_expire
);
286 CFQ_CFQQ_FNS(idle_window
);
287 CFQ_CFQQ_FNS(prio_changed
);
288 CFQ_CFQQ_FNS(slice_new
);
294 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
295 blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args)
296 #define cfq_log(cfqd, fmt, args...) \
297 blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
299 static inline enum wl_prio_t
cfqq_prio(struct cfq_queue
*cfqq
)
301 if (cfq_class_idle(cfqq
))
302 return IDLE_WORKLOAD
;
303 if (cfq_class_rt(cfqq
))
309 static enum wl_type_t
cfqq_type(struct cfq_queue
*cfqq
)
311 if (!cfq_cfqq_sync(cfqq
))
312 return ASYNC_WORKLOAD
;
313 if (!cfq_cfqq_idle_window(cfqq
))
314 return SYNC_NOIDLE_WORKLOAD
;
315 return SYNC_WORKLOAD
;
318 static inline int cfq_busy_queues_wl(enum wl_prio_t wl
, struct cfq_data
*cfqd
)
320 if (wl
== IDLE_WORKLOAD
)
321 return cfqd
->service_tree_idle
.count
;
323 return cfqd
->service_trees
[wl
][ASYNC_WORKLOAD
].count
324 + cfqd
->service_trees
[wl
][SYNC_NOIDLE_WORKLOAD
].count
325 + cfqd
->service_trees
[wl
][SYNC_WORKLOAD
].count
;
328 static void cfq_dispatch_insert(struct request_queue
*, struct request
*);
329 static struct cfq_queue
*cfq_get_queue(struct cfq_data
*, bool,
330 struct io_context
*, gfp_t
);
331 static struct cfq_io_context
*cfq_cic_lookup(struct cfq_data
*,
332 struct io_context
*);
334 static inline int rq_in_driver(struct cfq_data
*cfqd
)
336 return cfqd
->rq_in_driver
[0] + cfqd
->rq_in_driver
[1];
339 static inline struct cfq_queue
*cic_to_cfqq(struct cfq_io_context
*cic
,
342 return cic
->cfqq
[is_sync
];
345 static inline void cic_set_cfqq(struct cfq_io_context
*cic
,
346 struct cfq_queue
*cfqq
, bool is_sync
)
348 cic
->cfqq
[is_sync
] = cfqq
;
352 * We regard a request as SYNC, if it's either a read or has the SYNC bit
353 * set (in which case it could also be direct WRITE).
355 static inline bool cfq_bio_sync(struct bio
*bio
)
357 return bio_data_dir(bio
) == READ
|| bio_rw_flagged(bio
, BIO_RW_SYNCIO
);
361 * scheduler run of queue, if there are requests pending and no one in the
362 * driver that will restart queueing
364 static inline void cfq_schedule_dispatch(struct cfq_data
*cfqd
)
366 if (cfqd
->busy_queues
) {
367 cfq_log(cfqd
, "schedule dispatch");
368 kblockd_schedule_work(cfqd
->queue
, &cfqd
->unplug_work
);
372 static int cfq_queue_empty(struct request_queue
*q
)
374 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
376 return !cfqd
->busy_queues
;
380 * Scale schedule slice based on io priority. Use the sync time slice only
381 * if a queue is marked sync and has sync io queued. A sync queue with async
382 * io only, should not get full sync slice length.
384 static inline int cfq_prio_slice(struct cfq_data
*cfqd
, bool sync
,
387 const int base_slice
= cfqd
->cfq_slice
[sync
];
389 WARN_ON(prio
>= IOPRIO_BE_NR
);
391 return base_slice
+ (base_slice
/CFQ_SLICE_SCALE
* (4 - prio
));
395 cfq_prio_to_slice(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
397 return cfq_prio_slice(cfqd
, cfq_cfqq_sync(cfqq
), cfqq
->ioprio
);
401 * get averaged number of queues of RT/BE priority.
402 * average is updated, with a formula that gives more weight to higher numbers,
403 * to quickly follows sudden increases and decrease slowly
406 static inline unsigned cfq_get_avg_queues(struct cfq_data
*cfqd
, bool rt
)
408 unsigned min_q
, max_q
;
409 unsigned mult
= cfq_hist_divisor
- 1;
410 unsigned round
= cfq_hist_divisor
/ 2;
411 unsigned busy
= cfq_busy_queues_wl(rt
, cfqd
);
413 min_q
= min(cfqd
->busy_queues_avg
[rt
], busy
);
414 max_q
= max(cfqd
->busy_queues_avg
[rt
], busy
);
415 cfqd
->busy_queues_avg
[rt
] = (mult
* max_q
+ min_q
+ round
) /
417 return cfqd
->busy_queues_avg
[rt
];
421 cfq_set_prio_slice(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
423 unsigned slice
= cfq_prio_to_slice(cfqd
, cfqq
);
424 if (cfqd
->cfq_latency
) {
425 /* interested queues (we consider only the ones with the same
427 unsigned iq
= cfq_get_avg_queues(cfqd
, cfq_class_rt(cfqq
));
428 unsigned sync_slice
= cfqd
->cfq_slice
[1];
429 unsigned expect_latency
= sync_slice
* iq
;
430 if (expect_latency
> cfq_target_latency
) {
431 unsigned base_low_slice
= 2 * cfqd
->cfq_slice_idle
;
432 /* scale low_slice according to IO priority
433 * and sync vs async */
435 min(slice
, base_low_slice
* slice
/ sync_slice
);
436 /* the adapted slice value is scaled to fit all iqs
437 * into the target latency */
438 slice
= max(slice
* cfq_target_latency
/ expect_latency
,
442 cfqq
->slice_end
= jiffies
+ slice
;
443 cfq_log_cfqq(cfqd
, cfqq
, "set_slice=%lu", cfqq
->slice_end
- jiffies
);
447 * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
448 * isn't valid until the first request from the dispatch is activated
449 * and the slice time set.
451 static inline bool cfq_slice_used(struct cfq_queue
*cfqq
)
453 if (cfq_cfqq_slice_new(cfqq
))
455 if (time_before(jiffies
, cfqq
->slice_end
))
462 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
463 * We choose the request that is closest to the head right now. Distance
464 * behind the head is penalized and only allowed to a certain extent.
466 static struct request
*
467 cfq_choose_req(struct cfq_data
*cfqd
, struct request
*rq1
, struct request
*rq2
, sector_t last
)
469 sector_t s1
, s2
, d1
= 0, d2
= 0;
470 unsigned long back_max
;
471 #define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */
472 #define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */
473 unsigned wrap
= 0; /* bit mask: requests behind the disk head? */
475 if (rq1
== NULL
|| rq1
== rq2
)
480 if (rq_is_sync(rq1
) && !rq_is_sync(rq2
))
482 else if (rq_is_sync(rq2
) && !rq_is_sync(rq1
))
484 if (rq_is_meta(rq1
) && !rq_is_meta(rq2
))
486 else if (rq_is_meta(rq2
) && !rq_is_meta(rq1
))
489 s1
= blk_rq_pos(rq1
);
490 s2
= blk_rq_pos(rq2
);
493 * by definition, 1KiB is 2 sectors
495 back_max
= cfqd
->cfq_back_max
* 2;
498 * Strict one way elevator _except_ in the case where we allow
499 * short backward seeks which are biased as twice the cost of a
500 * similar forward seek.
504 else if (s1
+ back_max
>= last
)
505 d1
= (last
- s1
) * cfqd
->cfq_back_penalty
;
507 wrap
|= CFQ_RQ1_WRAP
;
511 else if (s2
+ back_max
>= last
)
512 d2
= (last
- s2
) * cfqd
->cfq_back_penalty
;
514 wrap
|= CFQ_RQ2_WRAP
;
516 /* Found required data */
519 * By doing switch() on the bit mask "wrap" we avoid having to
520 * check two variables for all permutations: --> faster!
523 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
539 case (CFQ_RQ1_WRAP
|CFQ_RQ2_WRAP
): /* both rqs wrapped */
542 * Since both rqs are wrapped,
543 * start with the one that's further behind head
544 * (--> only *one* back seek required),
545 * since back seek takes more time than forward.
555 * The below is leftmost cache rbtree addon
557 static struct cfq_queue
*cfq_rb_first(struct cfq_rb_root
*root
)
560 root
->left
= rb_first(&root
->rb
);
563 return rb_entry(root
->left
, struct cfq_queue
, rb_node
);
568 static void rb_erase_init(struct rb_node
*n
, struct rb_root
*root
)
574 static void cfq_rb_erase(struct rb_node
*n
, struct cfq_rb_root
*root
)
578 rb_erase_init(n
, &root
->rb
);
583 * would be nice to take fifo expire time into account as well
585 static struct request
*
586 cfq_find_next_rq(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
587 struct request
*last
)
589 struct rb_node
*rbnext
= rb_next(&last
->rb_node
);
590 struct rb_node
*rbprev
= rb_prev(&last
->rb_node
);
591 struct request
*next
= NULL
, *prev
= NULL
;
593 BUG_ON(RB_EMPTY_NODE(&last
->rb_node
));
596 prev
= rb_entry_rq(rbprev
);
599 next
= rb_entry_rq(rbnext
);
601 rbnext
= rb_first(&cfqq
->sort_list
);
602 if (rbnext
&& rbnext
!= &last
->rb_node
)
603 next
= rb_entry_rq(rbnext
);
606 return cfq_choose_req(cfqd
, next
, prev
, blk_rq_pos(last
));
609 static unsigned long cfq_slice_offset(struct cfq_data
*cfqd
,
610 struct cfq_queue
*cfqq
)
613 * just an approximation, should be ok.
615 return (cfqd
->busy_queues
- 1) * (cfq_prio_slice(cfqd
, 1, 0) -
616 cfq_prio_slice(cfqd
, cfq_cfqq_sync(cfqq
), cfqq
->ioprio
));
620 * The cfqd->service_trees holds all pending cfq_queue's that have
621 * requests waiting to be processed. It is sorted in the order that
622 * we will service the queues.
624 static void cfq_service_tree_add(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
627 struct rb_node
**p
, *parent
;
628 struct cfq_queue
*__cfqq
;
629 unsigned long rb_key
;
630 struct cfq_rb_root
*service_tree
;
633 service_tree
= service_tree_for(cfqq_prio(cfqq
), cfqq_type(cfqq
), cfqd
);
634 if (cfq_class_idle(cfqq
)) {
635 rb_key
= CFQ_IDLE_DELAY
;
636 parent
= rb_last(&service_tree
->rb
);
637 if (parent
&& parent
!= &cfqq
->rb_node
) {
638 __cfqq
= rb_entry(parent
, struct cfq_queue
, rb_node
);
639 rb_key
+= __cfqq
->rb_key
;
642 } else if (!add_front
) {
644 * Get our rb key offset. Subtract any residual slice
645 * value carried from last service. A negative resid
646 * count indicates slice overrun, and this should position
647 * the next service time further away in the tree.
649 rb_key
= cfq_slice_offset(cfqd
, cfqq
) + jiffies
;
650 rb_key
-= cfqq
->slice_resid
;
651 cfqq
->slice_resid
= 0;
654 __cfqq
= cfq_rb_first(service_tree
);
655 rb_key
+= __cfqq
? __cfqq
->rb_key
: jiffies
;
658 if (!RB_EMPTY_NODE(&cfqq
->rb_node
)) {
660 * same position, nothing more to do
662 if (rb_key
== cfqq
->rb_key
&&
663 cfqq
->service_tree
== service_tree
)
666 cfq_rb_erase(&cfqq
->rb_node
, cfqq
->service_tree
);
667 cfqq
->service_tree
= NULL
;
672 cfqq
->service_tree
= service_tree
;
673 p
= &service_tree
->rb
.rb_node
;
678 __cfqq
= rb_entry(parent
, struct cfq_queue
, rb_node
);
681 * sort by key, that represents service time.
683 if (time_before(rb_key
, __cfqq
->rb_key
))
694 service_tree
->left
= &cfqq
->rb_node
;
696 cfqq
->rb_key
= rb_key
;
697 rb_link_node(&cfqq
->rb_node
, parent
, p
);
698 rb_insert_color(&cfqq
->rb_node
, &service_tree
->rb
);
699 service_tree
->count
++;
702 static struct cfq_queue
*
703 cfq_prio_tree_lookup(struct cfq_data
*cfqd
, struct rb_root
*root
,
704 sector_t sector
, struct rb_node
**ret_parent
,
705 struct rb_node
***rb_link
)
707 struct rb_node
**p
, *parent
;
708 struct cfq_queue
*cfqq
= NULL
;
716 cfqq
= rb_entry(parent
, struct cfq_queue
, p_node
);
719 * Sort strictly based on sector. Smallest to the left,
720 * largest to the right.
722 if (sector
> blk_rq_pos(cfqq
->next_rq
))
724 else if (sector
< blk_rq_pos(cfqq
->next_rq
))
732 *ret_parent
= parent
;
738 static void cfq_prio_tree_add(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
740 struct rb_node
**p
, *parent
;
741 struct cfq_queue
*__cfqq
;
744 rb_erase(&cfqq
->p_node
, cfqq
->p_root
);
748 if (cfq_class_idle(cfqq
))
753 cfqq
->p_root
= &cfqd
->prio_trees
[cfqq
->org_ioprio
];
754 __cfqq
= cfq_prio_tree_lookup(cfqd
, cfqq
->p_root
,
755 blk_rq_pos(cfqq
->next_rq
), &parent
, &p
);
757 rb_link_node(&cfqq
->p_node
, parent
, p
);
758 rb_insert_color(&cfqq
->p_node
, cfqq
->p_root
);
764 * Update cfqq's position in the service tree.
766 static void cfq_resort_rr_list(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
769 * Resorting requires the cfqq to be on the RR list already.
771 if (cfq_cfqq_on_rr(cfqq
)) {
772 cfq_service_tree_add(cfqd
, cfqq
, 0);
773 cfq_prio_tree_add(cfqd
, cfqq
);
778 * add to busy list of queues for service, trying to be fair in ordering
779 * the pending list according to last request service
781 static void cfq_add_cfqq_rr(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
783 cfq_log_cfqq(cfqd
, cfqq
, "add_to_rr");
784 BUG_ON(cfq_cfqq_on_rr(cfqq
));
785 cfq_mark_cfqq_on_rr(cfqq
);
788 cfq_resort_rr_list(cfqd
, cfqq
);
792 * Called when the cfqq no longer has requests pending, remove it from
795 static void cfq_del_cfqq_rr(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
797 cfq_log_cfqq(cfqd
, cfqq
, "del_from_rr");
798 BUG_ON(!cfq_cfqq_on_rr(cfqq
));
799 cfq_clear_cfqq_on_rr(cfqq
);
801 if (!RB_EMPTY_NODE(&cfqq
->rb_node
)) {
802 cfq_rb_erase(&cfqq
->rb_node
, cfqq
->service_tree
);
803 cfqq
->service_tree
= NULL
;
806 rb_erase(&cfqq
->p_node
, cfqq
->p_root
);
810 BUG_ON(!cfqd
->busy_queues
);
815 * rb tree support functions
817 static void cfq_del_rq_rb(struct request
*rq
)
819 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
820 struct cfq_data
*cfqd
= cfqq
->cfqd
;
821 const int sync
= rq_is_sync(rq
);
823 BUG_ON(!cfqq
->queued
[sync
]);
824 cfqq
->queued
[sync
]--;
826 elv_rb_del(&cfqq
->sort_list
, rq
);
828 if (cfq_cfqq_on_rr(cfqq
) && RB_EMPTY_ROOT(&cfqq
->sort_list
))
829 cfq_del_cfqq_rr(cfqd
, cfqq
);
832 static void cfq_add_rq_rb(struct request
*rq
)
834 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
835 struct cfq_data
*cfqd
= cfqq
->cfqd
;
836 struct request
*__alias
, *prev
;
838 cfqq
->queued
[rq_is_sync(rq
)]++;
841 * looks a little odd, but the first insert might return an alias.
842 * if that happens, put the alias on the dispatch list
844 while ((__alias
= elv_rb_add(&cfqq
->sort_list
, rq
)) != NULL
)
845 cfq_dispatch_insert(cfqd
->queue
, __alias
);
847 if (!cfq_cfqq_on_rr(cfqq
))
848 cfq_add_cfqq_rr(cfqd
, cfqq
);
851 * check if this request is a better next-serve candidate
853 prev
= cfqq
->next_rq
;
854 cfqq
->next_rq
= cfq_choose_req(cfqd
, cfqq
->next_rq
, rq
, cfqd
->last_position
);
857 * adjust priority tree position, if ->next_rq changes
859 if (prev
!= cfqq
->next_rq
)
860 cfq_prio_tree_add(cfqd
, cfqq
);
862 BUG_ON(!cfqq
->next_rq
);
865 static void cfq_reposition_rq_rb(struct cfq_queue
*cfqq
, struct request
*rq
)
867 elv_rb_del(&cfqq
->sort_list
, rq
);
868 cfqq
->queued
[rq_is_sync(rq
)]--;
872 static struct request
*
873 cfq_find_rq_fmerge(struct cfq_data
*cfqd
, struct bio
*bio
)
875 struct task_struct
*tsk
= current
;
876 struct cfq_io_context
*cic
;
877 struct cfq_queue
*cfqq
;
879 cic
= cfq_cic_lookup(cfqd
, tsk
->io_context
);
883 cfqq
= cic_to_cfqq(cic
, cfq_bio_sync(bio
));
885 sector_t sector
= bio
->bi_sector
+ bio_sectors(bio
);
887 return elv_rb_find(&cfqq
->sort_list
, sector
);
893 static void cfq_activate_request(struct request_queue
*q
, struct request
*rq
)
895 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
897 cfqd
->rq_in_driver
[rq_is_sync(rq
)]++;
898 cfq_log_cfqq(cfqd
, RQ_CFQQ(rq
), "activate rq, drv=%d",
901 cfqd
->last_position
= blk_rq_pos(rq
) + blk_rq_sectors(rq
);
904 static void cfq_deactivate_request(struct request_queue
*q
, struct request
*rq
)
906 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
907 const int sync
= rq_is_sync(rq
);
909 WARN_ON(!cfqd
->rq_in_driver
[sync
]);
910 cfqd
->rq_in_driver
[sync
]--;
911 cfq_log_cfqq(cfqd
, RQ_CFQQ(rq
), "deactivate rq, drv=%d",
915 static void cfq_remove_request(struct request
*rq
)
917 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
919 if (cfqq
->next_rq
== rq
)
920 cfqq
->next_rq
= cfq_find_next_rq(cfqq
->cfqd
, cfqq
, rq
);
922 list_del_init(&rq
->queuelist
);
925 cfqq
->cfqd
->rq_queued
--;
926 if (rq_is_meta(rq
)) {
927 WARN_ON(!cfqq
->meta_pending
);
928 cfqq
->meta_pending
--;
932 static int cfq_merge(struct request_queue
*q
, struct request
**req
,
935 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
936 struct request
*__rq
;
938 __rq
= cfq_find_rq_fmerge(cfqd
, bio
);
939 if (__rq
&& elv_rq_merge_ok(__rq
, bio
)) {
941 return ELEVATOR_FRONT_MERGE
;
944 return ELEVATOR_NO_MERGE
;
947 static void cfq_merged_request(struct request_queue
*q
, struct request
*req
,
950 if (type
== ELEVATOR_FRONT_MERGE
) {
951 struct cfq_queue
*cfqq
= RQ_CFQQ(req
);
953 cfq_reposition_rq_rb(cfqq
, req
);
958 cfq_merged_requests(struct request_queue
*q
, struct request
*rq
,
959 struct request
*next
)
961 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
963 * reposition in fifo if next is older than rq
965 if (!list_empty(&rq
->queuelist
) && !list_empty(&next
->queuelist
) &&
966 time_before(rq_fifo_time(next
), rq_fifo_time(rq
))) {
967 list_move(&rq
->queuelist
, &next
->queuelist
);
968 rq_set_fifo_time(rq
, rq_fifo_time(next
));
971 if (cfqq
->next_rq
== next
)
973 cfq_remove_request(next
);
976 static int cfq_allow_merge(struct request_queue
*q
, struct request
*rq
,
979 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
980 struct cfq_io_context
*cic
;
981 struct cfq_queue
*cfqq
;
984 * Disallow merge of a sync bio into an async request.
986 if (cfq_bio_sync(bio
) && !rq_is_sync(rq
))
990 * Lookup the cfqq that this bio will be queued with. Allow
991 * merge only if rq is queued there.
993 cic
= cfq_cic_lookup(cfqd
, current
->io_context
);
997 cfqq
= cic_to_cfqq(cic
, cfq_bio_sync(bio
));
998 return cfqq
== RQ_CFQQ(rq
);
1001 static void __cfq_set_active_queue(struct cfq_data
*cfqd
,
1002 struct cfq_queue
*cfqq
)
1005 cfq_log_cfqq(cfqd
, cfqq
, "set_active");
1006 cfqq
->slice_end
= 0;
1007 cfqq
->slice_dispatch
= 0;
1009 cfq_clear_cfqq_wait_request(cfqq
);
1010 cfq_clear_cfqq_must_dispatch(cfqq
);
1011 cfq_clear_cfqq_must_alloc_slice(cfqq
);
1012 cfq_clear_cfqq_fifo_expire(cfqq
);
1013 cfq_mark_cfqq_slice_new(cfqq
);
1015 del_timer(&cfqd
->idle_slice_timer
);
1018 cfqd
->active_queue
= cfqq
;
1022 * current cfqq expired its slice (or was too idle), select new one
1025 __cfq_slice_expired(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
1028 cfq_log_cfqq(cfqd
, cfqq
, "slice expired t=%d", timed_out
);
1030 if (cfq_cfqq_wait_request(cfqq
))
1031 del_timer(&cfqd
->idle_slice_timer
);
1033 cfq_clear_cfqq_wait_request(cfqq
);
1036 * store what was left of this slice, if the queue idled/timed out
1038 if (timed_out
&& !cfq_cfqq_slice_new(cfqq
)) {
1039 cfqq
->slice_resid
= cfqq
->slice_end
- jiffies
;
1040 cfq_log_cfqq(cfqd
, cfqq
, "resid=%ld", cfqq
->slice_resid
);
1043 cfq_resort_rr_list(cfqd
, cfqq
);
1045 if (cfqq
== cfqd
->active_queue
)
1046 cfqd
->active_queue
= NULL
;
1048 if (cfqd
->active_cic
) {
1049 put_io_context(cfqd
->active_cic
->ioc
);
1050 cfqd
->active_cic
= NULL
;
1054 static inline void cfq_slice_expired(struct cfq_data
*cfqd
, bool timed_out
)
1056 struct cfq_queue
*cfqq
= cfqd
->active_queue
;
1059 __cfq_slice_expired(cfqd
, cfqq
, timed_out
);
1063 * Get next queue for service. Unless we have a queue preemption,
1064 * we'll simply select the first cfqq in the service tree.
1066 static struct cfq_queue
*cfq_get_next_queue(struct cfq_data
*cfqd
)
1068 struct cfq_rb_root
*service_tree
=
1069 service_tree_for(cfqd
->serving_prio
, cfqd
->serving_type
, cfqd
);
1071 if (RB_EMPTY_ROOT(&service_tree
->rb
))
1073 return cfq_rb_first(service_tree
);
1077 * Get and set a new active queue for service.
1079 static struct cfq_queue
*cfq_set_active_queue(struct cfq_data
*cfqd
,
1080 struct cfq_queue
*cfqq
)
1083 cfqq
= cfq_get_next_queue(cfqd
);
1085 __cfq_set_active_queue(cfqd
, cfqq
);
1089 static inline sector_t
cfq_dist_from_last(struct cfq_data
*cfqd
,
1092 if (blk_rq_pos(rq
) >= cfqd
->last_position
)
1093 return blk_rq_pos(rq
) - cfqd
->last_position
;
1095 return cfqd
->last_position
- blk_rq_pos(rq
);
1098 #define CFQQ_SEEK_THR 8 * 1024
1099 #define CFQQ_SEEKY(cfqq) ((cfqq)->seek_mean > CFQQ_SEEK_THR)
1101 static inline int cfq_rq_close(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
1104 sector_t sdist
= cfqq
->seek_mean
;
1106 if (!sample_valid(cfqq
->seek_samples
))
1107 sdist
= CFQQ_SEEK_THR
;
1109 return cfq_dist_from_last(cfqd
, rq
) <= sdist
;
1112 static struct cfq_queue
*cfqq_close(struct cfq_data
*cfqd
,
1113 struct cfq_queue
*cur_cfqq
)
1115 struct rb_root
*root
= &cfqd
->prio_trees
[cur_cfqq
->org_ioprio
];
1116 struct rb_node
*parent
, *node
;
1117 struct cfq_queue
*__cfqq
;
1118 sector_t sector
= cfqd
->last_position
;
1120 if (RB_EMPTY_ROOT(root
))
1124 * First, if we find a request starting at the end of the last
1125 * request, choose it.
1127 __cfqq
= cfq_prio_tree_lookup(cfqd
, root
, sector
, &parent
, NULL
);
1132 * If the exact sector wasn't found, the parent of the NULL leaf
1133 * will contain the closest sector.
1135 __cfqq
= rb_entry(parent
, struct cfq_queue
, p_node
);
1136 if (cfq_rq_close(cfqd
, cur_cfqq
, __cfqq
->next_rq
))
1139 if (blk_rq_pos(__cfqq
->next_rq
) < sector
)
1140 node
= rb_next(&__cfqq
->p_node
);
1142 node
= rb_prev(&__cfqq
->p_node
);
1146 __cfqq
= rb_entry(node
, struct cfq_queue
, p_node
);
1147 if (cfq_rq_close(cfqd
, cur_cfqq
, __cfqq
->next_rq
))
1155 * cur_cfqq - passed in so that we don't decide that the current queue is
1156 * closely cooperating with itself.
1158 * So, basically we're assuming that that cur_cfqq has dispatched at least
1159 * one request, and that cfqd->last_position reflects a position on the disk
1160 * associated with the I/O issued by cur_cfqq. I'm not sure this is a valid
1163 static struct cfq_queue
*cfq_close_cooperator(struct cfq_data
*cfqd
,
1164 struct cfq_queue
*cur_cfqq
)
1166 struct cfq_queue
*cfqq
;
1168 if (!cfq_cfqq_sync(cur_cfqq
))
1170 if (CFQQ_SEEKY(cur_cfqq
))
1174 * We should notice if some of the queues are cooperating, eg
1175 * working closely on the same area of the disk. In that case,
1176 * we can group them together and don't waste time idling.
1178 cfqq
= cfqq_close(cfqd
, cur_cfqq
);
1183 * It only makes sense to merge sync queues.
1185 if (!cfq_cfqq_sync(cfqq
))
1187 if (CFQQ_SEEKY(cfqq
))
1191 * Do not merge queues of different priority classes
1193 if (cfq_class_rt(cfqq
) != cfq_class_rt(cur_cfqq
))
1200 * Determine whether we should enforce idle window for this queue.
1203 static bool cfq_should_idle(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
1205 enum wl_prio_t prio
= cfqq_prio(cfqq
);
1206 struct cfq_rb_root
*service_tree
= cfqq
->service_tree
;
1208 /* We never do for idle class queues. */
1209 if (prio
== IDLE_WORKLOAD
)
1212 /* We do for queues that were marked with idle window flag. */
1213 if (cfq_cfqq_idle_window(cfqq
))
1217 * Otherwise, we do only if they are the last ones
1218 * in their service tree.
1221 service_tree
= service_tree_for(prio
, cfqq_type(cfqq
), cfqd
);
1223 if (service_tree
->count
== 0)
1226 return (service_tree
->count
== 1 && cfq_rb_first(service_tree
) == cfqq
);
1229 static void cfq_arm_slice_timer(struct cfq_data
*cfqd
)
1231 struct cfq_queue
*cfqq
= cfqd
->active_queue
;
1232 struct cfq_io_context
*cic
;
1236 * SSD device without seek penalty, disable idling. But only do so
1237 * for devices that support queuing, otherwise we still have a problem
1238 * with sync vs async workloads.
1240 if (blk_queue_nonrot(cfqd
->queue
) && cfqd
->hw_tag
)
1243 WARN_ON(!RB_EMPTY_ROOT(&cfqq
->sort_list
));
1244 WARN_ON(cfq_cfqq_slice_new(cfqq
));
1247 * idle is disabled, either manually or by past process history
1249 if (!cfqd
->cfq_slice_idle
|| !cfq_should_idle(cfqd
, cfqq
))
1253 * still active requests from this queue, don't idle
1255 if (cfqq
->dispatched
)
1259 * task has exited, don't wait
1261 cic
= cfqd
->active_cic
;
1262 if (!cic
|| !atomic_read(&cic
->ioc
->nr_tasks
))
1266 * If our average think time is larger than the remaining time
1267 * slice, then don't idle. This avoids overrunning the allotted
1270 if (sample_valid(cic
->ttime_samples
) &&
1271 (cfqq
->slice_end
- jiffies
< cic
->ttime_mean
))
1274 cfq_mark_cfqq_wait_request(cfqq
);
1276 sl
= cfqd
->cfq_slice_idle
;
1278 mod_timer(&cfqd
->idle_slice_timer
, jiffies
+ sl
);
1279 cfq_log_cfqq(cfqd
, cfqq
, "arm_idle: %lu", sl
);
1283 * Move request from internal lists to the request queue dispatch list.
1285 static void cfq_dispatch_insert(struct request_queue
*q
, struct request
*rq
)
1287 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
1288 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
1290 cfq_log_cfqq(cfqd
, cfqq
, "dispatch_insert");
1292 cfqq
->next_rq
= cfq_find_next_rq(cfqd
, cfqq
, rq
);
1293 cfq_remove_request(rq
);
1295 elv_dispatch_sort(q
, rq
);
1297 if (cfq_cfqq_sync(cfqq
))
1298 cfqd
->sync_flight
++;
1302 * return expired entry, or NULL to just start from scratch in rbtree
1304 static struct request
*cfq_check_fifo(struct cfq_queue
*cfqq
)
1306 struct request
*rq
= NULL
;
1308 if (cfq_cfqq_fifo_expire(cfqq
))
1311 cfq_mark_cfqq_fifo_expire(cfqq
);
1313 if (list_empty(&cfqq
->fifo
))
1316 rq
= rq_entry_fifo(cfqq
->fifo
.next
);
1317 if (time_before(jiffies
, rq_fifo_time(rq
)))
1320 cfq_log_cfqq(cfqq
->cfqd
, cfqq
, "fifo=%p", rq
);
1325 cfq_prio_to_maxrq(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
1327 const int base_rq
= cfqd
->cfq_slice_async_rq
;
1329 WARN_ON(cfqq
->ioprio
>= IOPRIO_BE_NR
);
1331 return 2 * (base_rq
+ base_rq
* (CFQ_PRIO_LISTS
- 1 - cfqq
->ioprio
));
1335 * Must be called with the queue_lock held.
1337 static int cfqq_process_refs(struct cfq_queue
*cfqq
)
1339 int process_refs
, io_refs
;
1341 io_refs
= cfqq
->allocated
[READ
] + cfqq
->allocated
[WRITE
];
1342 process_refs
= atomic_read(&cfqq
->ref
) - io_refs
;
1343 BUG_ON(process_refs
< 0);
1344 return process_refs
;
1347 static void cfq_setup_merge(struct cfq_queue
*cfqq
, struct cfq_queue
*new_cfqq
)
1349 int process_refs
, new_process_refs
;
1350 struct cfq_queue
*__cfqq
;
1352 /* Avoid a circular list and skip interim queue merges */
1353 while ((__cfqq
= new_cfqq
->new_cfqq
)) {
1359 process_refs
= cfqq_process_refs(cfqq
);
1361 * If the process for the cfqq has gone away, there is no
1362 * sense in merging the queues.
1364 if (process_refs
== 0)
1368 * Merge in the direction of the lesser amount of work.
1370 new_process_refs
= cfqq_process_refs(new_cfqq
);
1371 if (new_process_refs
>= process_refs
) {
1372 cfqq
->new_cfqq
= new_cfqq
;
1373 atomic_add(process_refs
, &new_cfqq
->ref
);
1375 new_cfqq
->new_cfqq
= cfqq
;
1376 atomic_add(new_process_refs
, &cfqq
->ref
);
1380 static enum wl_type_t
cfq_choose_wl(struct cfq_data
*cfqd
, enum wl_prio_t prio
,
1383 struct cfq_queue
*queue
;
1385 bool key_valid
= false;
1386 unsigned long lowest_key
= 0;
1387 enum wl_type_t cur_best
= SYNC_NOIDLE_WORKLOAD
;
1391 * When priorities switched, we prefer starting
1392 * from SYNC_NOIDLE (first choice), or just SYNC
1395 if (service_tree_for(prio
, cur_best
, cfqd
)->count
)
1397 cur_best
= SYNC_WORKLOAD
;
1398 if (service_tree_for(prio
, cur_best
, cfqd
)->count
)
1401 return ASYNC_WORKLOAD
;
1404 for (i
= 0; i
< 3; ++i
) {
1405 /* otherwise, select the one with lowest rb_key */
1406 queue
= cfq_rb_first(service_tree_for(prio
, i
, cfqd
));
1408 (!key_valid
|| time_before(queue
->rb_key
, lowest_key
))) {
1409 lowest_key
= queue
->rb_key
;
1418 static void choose_service_tree(struct cfq_data
*cfqd
)
1420 enum wl_prio_t previous_prio
= cfqd
->serving_prio
;
1425 /* Choose next priority. RT > BE > IDLE */
1426 if (cfq_busy_queues_wl(RT_WORKLOAD
, cfqd
))
1427 cfqd
->serving_prio
= RT_WORKLOAD
;
1428 else if (cfq_busy_queues_wl(BE_WORKLOAD
, cfqd
))
1429 cfqd
->serving_prio
= BE_WORKLOAD
;
1431 cfqd
->serving_prio
= IDLE_WORKLOAD
;
1432 cfqd
->workload_expires
= jiffies
+ 1;
1437 * For RT and BE, we have to choose also the type
1438 * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
1441 prio_changed
= (cfqd
->serving_prio
!= previous_prio
);
1442 count
= service_tree_for(cfqd
->serving_prio
, cfqd
->serving_type
, cfqd
)
1446 * If priority didn't change, check workload expiration,
1447 * and that we still have other queues ready
1449 if (!prio_changed
&& count
&&
1450 !time_after(jiffies
, cfqd
->workload_expires
))
1453 /* otherwise select new workload type */
1454 cfqd
->serving_type
=
1455 cfq_choose_wl(cfqd
, cfqd
->serving_prio
, prio_changed
);
1456 count
= service_tree_for(cfqd
->serving_prio
, cfqd
->serving_type
, cfqd
)
1460 * the workload slice is computed as a fraction of target latency
1461 * proportional to the number of queues in that workload, over
1462 * all the queues in the same priority class
1464 slice
= cfq_target_latency
* count
/
1465 max_t(unsigned, cfqd
->busy_queues_avg
[cfqd
->serving_prio
],
1466 cfq_busy_queues_wl(cfqd
->serving_prio
, cfqd
));
1468 if (cfqd
->serving_type
== ASYNC_WORKLOAD
)
1469 /* async workload slice is scaled down according to
1470 * the sync/async slice ratio. */
1471 slice
= slice
* cfqd
->cfq_slice
[0] / cfqd
->cfq_slice
[1];
1473 /* sync workload slice is at least 2 * cfq_slice_idle */
1474 slice
= max(slice
, 2 * cfqd
->cfq_slice_idle
);
1476 slice
= max_t(unsigned, slice
, CFQ_MIN_TT
);
1477 cfqd
->workload_expires
= jiffies
+ slice
;
1478 cfqd
->noidle_tree_requires_idle
= false;
1482 * Select a queue for service. If we have a current active queue,
1483 * check whether to continue servicing it, or retrieve and set a new one.
1485 static struct cfq_queue
*cfq_select_queue(struct cfq_data
*cfqd
)
1487 struct cfq_queue
*cfqq
, *new_cfqq
= NULL
;
1489 cfqq
= cfqd
->active_queue
;
1494 * The active queue has run out of time, expire it and select new.
1496 if (cfq_slice_used(cfqq
) && !cfq_cfqq_must_dispatch(cfqq
))
1500 * The active queue has requests and isn't expired, allow it to
1503 if (!RB_EMPTY_ROOT(&cfqq
->sort_list
))
1507 * If another queue has a request waiting within our mean seek
1508 * distance, let it run. The expire code will check for close
1509 * cooperators and put the close queue at the front of the service
1510 * tree. If possible, merge the expiring queue with the new cfqq.
1512 new_cfqq
= cfq_close_cooperator(cfqd
, cfqq
);
1514 if (!cfqq
->new_cfqq
)
1515 cfq_setup_merge(cfqq
, new_cfqq
);
1520 * No requests pending. If the active queue still has requests in
1521 * flight or is idling for a new request, allow either of these
1522 * conditions to happen (or time out) before selecting a new queue.
1524 if (timer_pending(&cfqd
->idle_slice_timer
) ||
1525 (cfqq
->dispatched
&& cfq_should_idle(cfqd
, cfqq
))) {
1531 cfq_slice_expired(cfqd
, 0);
1534 * Current queue expired. Check if we have to switch to a new
1538 choose_service_tree(cfqd
);
1540 cfqq
= cfq_set_active_queue(cfqd
, new_cfqq
);
1545 static int __cfq_forced_dispatch_cfqq(struct cfq_queue
*cfqq
)
1549 while (cfqq
->next_rq
) {
1550 cfq_dispatch_insert(cfqq
->cfqd
->queue
, cfqq
->next_rq
);
1554 BUG_ON(!list_empty(&cfqq
->fifo
));
1559 * Drain our current requests. Used for barriers and when switching
1560 * io schedulers on-the-fly.
1562 static int cfq_forced_dispatch(struct cfq_data
*cfqd
)
1564 struct cfq_queue
*cfqq
;
1567 for (i
= 0; i
< 2; ++i
)
1568 for (j
= 0; j
< 3; ++j
)
1569 while ((cfqq
= cfq_rb_first(&cfqd
->service_trees
[i
][j
]))
1571 dispatched
+= __cfq_forced_dispatch_cfqq(cfqq
);
1573 while ((cfqq
= cfq_rb_first(&cfqd
->service_tree_idle
)) != NULL
)
1574 dispatched
+= __cfq_forced_dispatch_cfqq(cfqq
);
1576 cfq_slice_expired(cfqd
, 0);
1578 BUG_ON(cfqd
->busy_queues
);
1580 cfq_log(cfqd
, "forced_dispatch=%d", dispatched
);
1584 static bool cfq_may_dispatch(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
1586 unsigned int max_dispatch
;
1589 * Drain async requests before we start sync IO
1591 if (cfq_should_idle(cfqd
, cfqq
) && cfqd
->rq_in_driver
[BLK_RW_ASYNC
])
1595 * If this is an async queue and we have sync IO in flight, let it wait
1597 if (cfqd
->sync_flight
&& !cfq_cfqq_sync(cfqq
))
1600 max_dispatch
= cfqd
->cfq_quantum
;
1601 if (cfq_class_idle(cfqq
))
1605 * Does this cfqq already have too much IO in flight?
1607 if (cfqq
->dispatched
>= max_dispatch
) {
1609 * idle queue must always only have a single IO in flight
1611 if (cfq_class_idle(cfqq
))
1615 * We have other queues, don't allow more IO from this one
1617 if (cfqd
->busy_queues
> 1)
1621 * Sole queue user, no limit
1627 * Async queues must wait a bit before being allowed dispatch.
1628 * We also ramp up the dispatch depth gradually for async IO,
1629 * based on the last sync IO we serviced
1631 if (!cfq_cfqq_sync(cfqq
) && cfqd
->cfq_latency
) {
1632 unsigned long last_sync
= jiffies
- cfqd
->last_end_sync_rq
;
1635 depth
= last_sync
/ cfqd
->cfq_slice
[1];
1636 if (!depth
&& !cfqq
->dispatched
)
1638 if (depth
< max_dispatch
)
1639 max_dispatch
= depth
;
1643 * If we're below the current max, allow a dispatch
1645 return cfqq
->dispatched
< max_dispatch
;
1649 * Dispatch a request from cfqq, moving them to the request queue
1652 static bool cfq_dispatch_request(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
1656 BUG_ON(RB_EMPTY_ROOT(&cfqq
->sort_list
));
1658 if (!cfq_may_dispatch(cfqd
, cfqq
))
1662 * follow expired path, else get first next available
1664 rq
= cfq_check_fifo(cfqq
);
1669 * insert request into driver dispatch list
1671 cfq_dispatch_insert(cfqd
->queue
, rq
);
1673 if (!cfqd
->active_cic
) {
1674 struct cfq_io_context
*cic
= RQ_CIC(rq
);
1676 atomic_long_inc(&cic
->ioc
->refcount
);
1677 cfqd
->active_cic
= cic
;
1684 * Find the cfqq that we need to service and move a request from that to the
1687 static int cfq_dispatch_requests(struct request_queue
*q
, int force
)
1689 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
1690 struct cfq_queue
*cfqq
;
1692 if (!cfqd
->busy_queues
)
1695 if (unlikely(force
))
1696 return cfq_forced_dispatch(cfqd
);
1698 cfqq
= cfq_select_queue(cfqd
);
1703 * Dispatch a request from this cfqq, if it is allowed
1705 if (!cfq_dispatch_request(cfqd
, cfqq
))
1708 cfqq
->slice_dispatch
++;
1709 cfq_clear_cfqq_must_dispatch(cfqq
);
1712 * expire an async queue immediately if it has used up its slice. idle
1713 * queue always expire after 1 dispatch round.
1715 if (cfqd
->busy_queues
> 1 && ((!cfq_cfqq_sync(cfqq
) &&
1716 cfqq
->slice_dispatch
>= cfq_prio_to_maxrq(cfqd
, cfqq
)) ||
1717 cfq_class_idle(cfqq
))) {
1718 cfqq
->slice_end
= jiffies
+ 1;
1719 cfq_slice_expired(cfqd
, 0);
1722 cfq_log_cfqq(cfqd
, cfqq
, "dispatched a request");
1727 * task holds one reference to the queue, dropped when task exits. each rq
1728 * in-flight on this queue also holds a reference, dropped when rq is freed.
1730 * queue lock must be held here.
1732 static void cfq_put_queue(struct cfq_queue
*cfqq
)
1734 struct cfq_data
*cfqd
= cfqq
->cfqd
;
1736 BUG_ON(atomic_read(&cfqq
->ref
) <= 0);
1738 if (!atomic_dec_and_test(&cfqq
->ref
))
1741 cfq_log_cfqq(cfqd
, cfqq
, "put_queue");
1742 BUG_ON(rb_first(&cfqq
->sort_list
));
1743 BUG_ON(cfqq
->allocated
[READ
] + cfqq
->allocated
[WRITE
]);
1744 BUG_ON(cfq_cfqq_on_rr(cfqq
));
1746 if (unlikely(cfqd
->active_queue
== cfqq
)) {
1747 __cfq_slice_expired(cfqd
, cfqq
, 0);
1748 cfq_schedule_dispatch(cfqd
);
1751 kmem_cache_free(cfq_pool
, cfqq
);
1755 * Must always be called with the rcu_read_lock() held
1758 __call_for_each_cic(struct io_context
*ioc
,
1759 void (*func
)(struct io_context
*, struct cfq_io_context
*))
1761 struct cfq_io_context
*cic
;
1762 struct hlist_node
*n
;
1764 hlist_for_each_entry_rcu(cic
, n
, &ioc
->cic_list
, cic_list
)
1769 * Call func for each cic attached to this ioc.
1772 call_for_each_cic(struct io_context
*ioc
,
1773 void (*func
)(struct io_context
*, struct cfq_io_context
*))
1776 __call_for_each_cic(ioc
, func
);
1780 static void cfq_cic_free_rcu(struct rcu_head
*head
)
1782 struct cfq_io_context
*cic
;
1784 cic
= container_of(head
, struct cfq_io_context
, rcu_head
);
1786 kmem_cache_free(cfq_ioc_pool
, cic
);
1787 elv_ioc_count_dec(cfq_ioc_count
);
1791 * CFQ scheduler is exiting, grab exit lock and check
1792 * the pending io context count. If it hits zero,
1793 * complete ioc_gone and set it back to NULL
1795 spin_lock(&ioc_gone_lock
);
1796 if (ioc_gone
&& !elv_ioc_count_read(cfq_ioc_count
)) {
1800 spin_unlock(&ioc_gone_lock
);
1804 static void cfq_cic_free(struct cfq_io_context
*cic
)
1806 call_rcu(&cic
->rcu_head
, cfq_cic_free_rcu
);
1809 static void cic_free_func(struct io_context
*ioc
, struct cfq_io_context
*cic
)
1811 unsigned long flags
;
1813 BUG_ON(!cic
->dead_key
);
1815 spin_lock_irqsave(&ioc
->lock
, flags
);
1816 radix_tree_delete(&ioc
->radix_root
, cic
->dead_key
);
1817 hlist_del_rcu(&cic
->cic_list
);
1818 spin_unlock_irqrestore(&ioc
->lock
, flags
);
1824 * Must be called with rcu_read_lock() held or preemption otherwise disabled.
1825 * Only two callers of this - ->dtor() which is called with the rcu_read_lock(),
1826 * and ->trim() which is called with the task lock held
1828 static void cfq_free_io_context(struct io_context
*ioc
)
1831 * ioc->refcount is zero here, or we are called from elv_unregister(),
1832 * so no more cic's are allowed to be linked into this ioc. So it
1833 * should be ok to iterate over the known list, we will see all cic's
1834 * since no new ones are added.
1836 __call_for_each_cic(ioc
, cic_free_func
);
1839 static void cfq_exit_cfqq(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
1841 struct cfq_queue
*__cfqq
, *next
;
1843 if (unlikely(cfqq
== cfqd
->active_queue
)) {
1844 __cfq_slice_expired(cfqd
, cfqq
, 0);
1845 cfq_schedule_dispatch(cfqd
);
1849 * If this queue was scheduled to merge with another queue, be
1850 * sure to drop the reference taken on that queue (and others in
1851 * the merge chain). See cfq_setup_merge and cfq_merge_cfqqs.
1853 __cfqq
= cfqq
->new_cfqq
;
1855 if (__cfqq
== cfqq
) {
1856 WARN(1, "cfqq->new_cfqq loop detected\n");
1859 next
= __cfqq
->new_cfqq
;
1860 cfq_put_queue(__cfqq
);
1864 cfq_put_queue(cfqq
);
1867 static void __cfq_exit_single_io_context(struct cfq_data
*cfqd
,
1868 struct cfq_io_context
*cic
)
1870 struct io_context
*ioc
= cic
->ioc
;
1872 list_del_init(&cic
->queue_list
);
1875 * Make sure key == NULL is seen for dead queues
1878 cic
->dead_key
= (unsigned long) cic
->key
;
1881 if (ioc
->ioc_data
== cic
)
1882 rcu_assign_pointer(ioc
->ioc_data
, NULL
);
1884 if (cic
->cfqq
[BLK_RW_ASYNC
]) {
1885 cfq_exit_cfqq(cfqd
, cic
->cfqq
[BLK_RW_ASYNC
]);
1886 cic
->cfqq
[BLK_RW_ASYNC
] = NULL
;
1889 if (cic
->cfqq
[BLK_RW_SYNC
]) {
1890 cfq_exit_cfqq(cfqd
, cic
->cfqq
[BLK_RW_SYNC
]);
1891 cic
->cfqq
[BLK_RW_SYNC
] = NULL
;
1895 static void cfq_exit_single_io_context(struct io_context
*ioc
,
1896 struct cfq_io_context
*cic
)
1898 struct cfq_data
*cfqd
= cic
->key
;
1901 struct request_queue
*q
= cfqd
->queue
;
1902 unsigned long flags
;
1904 spin_lock_irqsave(q
->queue_lock
, flags
);
1907 * Ensure we get a fresh copy of the ->key to prevent
1908 * race between exiting task and queue
1910 smp_read_barrier_depends();
1912 __cfq_exit_single_io_context(cfqd
, cic
);
1914 spin_unlock_irqrestore(q
->queue_lock
, flags
);
1919 * The process that ioc belongs to has exited, we need to clean up
1920 * and put the internal structures we have that belongs to that process.
1922 static void cfq_exit_io_context(struct io_context
*ioc
)
1924 call_for_each_cic(ioc
, cfq_exit_single_io_context
);
1927 static struct cfq_io_context
*
1928 cfq_alloc_io_context(struct cfq_data
*cfqd
, gfp_t gfp_mask
)
1930 struct cfq_io_context
*cic
;
1932 cic
= kmem_cache_alloc_node(cfq_ioc_pool
, gfp_mask
| __GFP_ZERO
,
1935 cic
->last_end_request
= jiffies
;
1936 INIT_LIST_HEAD(&cic
->queue_list
);
1937 INIT_HLIST_NODE(&cic
->cic_list
);
1938 cic
->dtor
= cfq_free_io_context
;
1939 cic
->exit
= cfq_exit_io_context
;
1940 elv_ioc_count_inc(cfq_ioc_count
);
1946 static void cfq_init_prio_data(struct cfq_queue
*cfqq
, struct io_context
*ioc
)
1948 struct task_struct
*tsk
= current
;
1951 if (!cfq_cfqq_prio_changed(cfqq
))
1954 ioprio_class
= IOPRIO_PRIO_CLASS(ioc
->ioprio
);
1955 switch (ioprio_class
) {
1957 printk(KERN_ERR
"cfq: bad prio %x\n", ioprio_class
);
1958 case IOPRIO_CLASS_NONE
:
1960 * no prio set, inherit CPU scheduling settings
1962 cfqq
->ioprio
= task_nice_ioprio(tsk
);
1963 cfqq
->ioprio_class
= task_nice_ioclass(tsk
);
1965 case IOPRIO_CLASS_RT
:
1966 cfqq
->ioprio
= task_ioprio(ioc
);
1967 cfqq
->ioprio_class
= IOPRIO_CLASS_RT
;
1969 case IOPRIO_CLASS_BE
:
1970 cfqq
->ioprio
= task_ioprio(ioc
);
1971 cfqq
->ioprio_class
= IOPRIO_CLASS_BE
;
1973 case IOPRIO_CLASS_IDLE
:
1974 cfqq
->ioprio_class
= IOPRIO_CLASS_IDLE
;
1976 cfq_clear_cfqq_idle_window(cfqq
);
1981 * keep track of original prio settings in case we have to temporarily
1982 * elevate the priority of this queue
1984 cfqq
->org_ioprio
= cfqq
->ioprio
;
1985 cfqq
->org_ioprio_class
= cfqq
->ioprio_class
;
1986 cfq_clear_cfqq_prio_changed(cfqq
);
1989 static void changed_ioprio(struct io_context
*ioc
, struct cfq_io_context
*cic
)
1991 struct cfq_data
*cfqd
= cic
->key
;
1992 struct cfq_queue
*cfqq
;
1993 unsigned long flags
;
1995 if (unlikely(!cfqd
))
1998 spin_lock_irqsave(cfqd
->queue
->queue_lock
, flags
);
2000 cfqq
= cic
->cfqq
[BLK_RW_ASYNC
];
2002 struct cfq_queue
*new_cfqq
;
2003 new_cfqq
= cfq_get_queue(cfqd
, BLK_RW_ASYNC
, cic
->ioc
,
2006 cic
->cfqq
[BLK_RW_ASYNC
] = new_cfqq
;
2007 cfq_put_queue(cfqq
);
2011 cfqq
= cic
->cfqq
[BLK_RW_SYNC
];
2013 cfq_mark_cfqq_prio_changed(cfqq
);
2015 spin_unlock_irqrestore(cfqd
->queue
->queue_lock
, flags
);
2018 static void cfq_ioc_set_ioprio(struct io_context
*ioc
)
2020 call_for_each_cic(ioc
, changed_ioprio
);
2021 ioc
->ioprio_changed
= 0;
2024 static void cfq_init_cfqq(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
2025 pid_t pid
, bool is_sync
)
2027 RB_CLEAR_NODE(&cfqq
->rb_node
);
2028 RB_CLEAR_NODE(&cfqq
->p_node
);
2029 INIT_LIST_HEAD(&cfqq
->fifo
);
2031 atomic_set(&cfqq
->ref
, 0);
2034 cfq_mark_cfqq_prio_changed(cfqq
);
2037 if (!cfq_class_idle(cfqq
))
2038 cfq_mark_cfqq_idle_window(cfqq
);
2039 cfq_mark_cfqq_sync(cfqq
);
2044 static struct cfq_queue
*
2045 cfq_find_alloc_queue(struct cfq_data
*cfqd
, bool is_sync
,
2046 struct io_context
*ioc
, gfp_t gfp_mask
)
2048 struct cfq_queue
*cfqq
, *new_cfqq
= NULL
;
2049 struct cfq_io_context
*cic
;
2052 cic
= cfq_cic_lookup(cfqd
, ioc
);
2053 /* cic always exists here */
2054 cfqq
= cic_to_cfqq(cic
, is_sync
);
2057 * Always try a new alloc if we fell back to the OOM cfqq
2058 * originally, since it should just be a temporary situation.
2060 if (!cfqq
|| cfqq
== &cfqd
->oom_cfqq
) {
2065 } else if (gfp_mask
& __GFP_WAIT
) {
2066 spin_unlock_irq(cfqd
->queue
->queue_lock
);
2067 new_cfqq
= kmem_cache_alloc_node(cfq_pool
,
2068 gfp_mask
| __GFP_ZERO
,
2070 spin_lock_irq(cfqd
->queue
->queue_lock
);
2074 cfqq
= kmem_cache_alloc_node(cfq_pool
,
2075 gfp_mask
| __GFP_ZERO
,
2080 cfq_init_cfqq(cfqd
, cfqq
, current
->pid
, is_sync
);
2081 cfq_init_prio_data(cfqq
, ioc
);
2082 cfq_log_cfqq(cfqd
, cfqq
, "alloced");
2084 cfqq
= &cfqd
->oom_cfqq
;
2088 kmem_cache_free(cfq_pool
, new_cfqq
);
2093 static struct cfq_queue
**
2094 cfq_async_queue_prio(struct cfq_data
*cfqd
, int ioprio_class
, int ioprio
)
2096 switch (ioprio_class
) {
2097 case IOPRIO_CLASS_RT
:
2098 return &cfqd
->async_cfqq
[0][ioprio
];
2099 case IOPRIO_CLASS_BE
:
2100 return &cfqd
->async_cfqq
[1][ioprio
];
2101 case IOPRIO_CLASS_IDLE
:
2102 return &cfqd
->async_idle_cfqq
;
2108 static struct cfq_queue
*
2109 cfq_get_queue(struct cfq_data
*cfqd
, bool is_sync
, struct io_context
*ioc
,
2112 const int ioprio
= task_ioprio(ioc
);
2113 const int ioprio_class
= task_ioprio_class(ioc
);
2114 struct cfq_queue
**async_cfqq
= NULL
;
2115 struct cfq_queue
*cfqq
= NULL
;
2118 async_cfqq
= cfq_async_queue_prio(cfqd
, ioprio_class
, ioprio
);
2123 cfqq
= cfq_find_alloc_queue(cfqd
, is_sync
, ioc
, gfp_mask
);
2126 * pin the queue now that it's allocated, scheduler exit will prune it
2128 if (!is_sync
&& !(*async_cfqq
)) {
2129 atomic_inc(&cfqq
->ref
);
2133 atomic_inc(&cfqq
->ref
);
2138 * We drop cfq io contexts lazily, so we may find a dead one.
2141 cfq_drop_dead_cic(struct cfq_data
*cfqd
, struct io_context
*ioc
,
2142 struct cfq_io_context
*cic
)
2144 unsigned long flags
;
2146 WARN_ON(!list_empty(&cic
->queue_list
));
2148 spin_lock_irqsave(&ioc
->lock
, flags
);
2150 BUG_ON(ioc
->ioc_data
== cic
);
2152 radix_tree_delete(&ioc
->radix_root
, (unsigned long) cfqd
);
2153 hlist_del_rcu(&cic
->cic_list
);
2154 spin_unlock_irqrestore(&ioc
->lock
, flags
);
2159 static struct cfq_io_context
*
2160 cfq_cic_lookup(struct cfq_data
*cfqd
, struct io_context
*ioc
)
2162 struct cfq_io_context
*cic
;
2163 unsigned long flags
;
2172 * we maintain a last-hit cache, to avoid browsing over the tree
2174 cic
= rcu_dereference(ioc
->ioc_data
);
2175 if (cic
&& cic
->key
== cfqd
) {
2181 cic
= radix_tree_lookup(&ioc
->radix_root
, (unsigned long) cfqd
);
2185 /* ->key must be copied to avoid race with cfq_exit_queue() */
2188 cfq_drop_dead_cic(cfqd
, ioc
, cic
);
2193 spin_lock_irqsave(&ioc
->lock
, flags
);
2194 rcu_assign_pointer(ioc
->ioc_data
, cic
);
2195 spin_unlock_irqrestore(&ioc
->lock
, flags
);
2203 * Add cic into ioc, using cfqd as the search key. This enables us to lookup
2204 * the process specific cfq io context when entered from the block layer.
2205 * Also adds the cic to a per-cfqd list, used when this queue is removed.
2207 static int cfq_cic_link(struct cfq_data
*cfqd
, struct io_context
*ioc
,
2208 struct cfq_io_context
*cic
, gfp_t gfp_mask
)
2210 unsigned long flags
;
2213 ret
= radix_tree_preload(gfp_mask
);
2218 spin_lock_irqsave(&ioc
->lock
, flags
);
2219 ret
= radix_tree_insert(&ioc
->radix_root
,
2220 (unsigned long) cfqd
, cic
);
2222 hlist_add_head_rcu(&cic
->cic_list
, &ioc
->cic_list
);
2223 spin_unlock_irqrestore(&ioc
->lock
, flags
);
2225 radix_tree_preload_end();
2228 spin_lock_irqsave(cfqd
->queue
->queue_lock
, flags
);
2229 list_add(&cic
->queue_list
, &cfqd
->cic_list
);
2230 spin_unlock_irqrestore(cfqd
->queue
->queue_lock
, flags
);
2235 printk(KERN_ERR
"cfq: cic link failed!\n");
2241 * Setup general io context and cfq io context. There can be several cfq
2242 * io contexts per general io context, if this process is doing io to more
2243 * than one device managed by cfq.
2245 static struct cfq_io_context
*
2246 cfq_get_io_context(struct cfq_data
*cfqd
, gfp_t gfp_mask
)
2248 struct io_context
*ioc
= NULL
;
2249 struct cfq_io_context
*cic
;
2251 might_sleep_if(gfp_mask
& __GFP_WAIT
);
2253 ioc
= get_io_context(gfp_mask
, cfqd
->queue
->node
);
2257 cic
= cfq_cic_lookup(cfqd
, ioc
);
2261 cic
= cfq_alloc_io_context(cfqd
, gfp_mask
);
2265 if (cfq_cic_link(cfqd
, ioc
, cic
, gfp_mask
))
2269 smp_read_barrier_depends();
2270 if (unlikely(ioc
->ioprio_changed
))
2271 cfq_ioc_set_ioprio(ioc
);
2277 put_io_context(ioc
);
2282 cfq_update_io_thinktime(struct cfq_data
*cfqd
, struct cfq_io_context
*cic
)
2284 unsigned long elapsed
= jiffies
- cic
->last_end_request
;
2285 unsigned long ttime
= min(elapsed
, 2UL * cfqd
->cfq_slice_idle
);
2287 cic
->ttime_samples
= (7*cic
->ttime_samples
+ 256) / 8;
2288 cic
->ttime_total
= (7*cic
->ttime_total
+ 256*ttime
) / 8;
2289 cic
->ttime_mean
= (cic
->ttime_total
+ 128) / cic
->ttime_samples
;
2293 cfq_update_io_seektime(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
2299 if (!cfqq
->last_request_pos
)
2301 else if (cfqq
->last_request_pos
< blk_rq_pos(rq
))
2302 sdist
= blk_rq_pos(rq
) - cfqq
->last_request_pos
;
2304 sdist
= cfqq
->last_request_pos
- blk_rq_pos(rq
);
2307 * Don't allow the seek distance to get too large from the
2308 * odd fragment, pagein, etc
2310 if (cfqq
->seek_samples
<= 60) /* second&third seek */
2311 sdist
= min(sdist
, (cfqq
->seek_mean
* 4) + 2*1024*1024);
2313 sdist
= min(sdist
, (cfqq
->seek_mean
* 4) + 2*1024*64);
2315 cfqq
->seek_samples
= (7*cfqq
->seek_samples
+ 256) / 8;
2316 cfqq
->seek_total
= (7*cfqq
->seek_total
+ (u64
)256*sdist
) / 8;
2317 total
= cfqq
->seek_total
+ (cfqq
->seek_samples
/2);
2318 do_div(total
, cfqq
->seek_samples
);
2319 cfqq
->seek_mean
= (sector_t
)total
;
2322 * If this cfqq is shared between multiple processes, check to
2323 * make sure that those processes are still issuing I/Os within
2324 * the mean seek distance. If not, it may be time to break the
2325 * queues apart again.
2327 if (cfq_cfqq_coop(cfqq
)) {
2328 if (CFQQ_SEEKY(cfqq
) && !cfqq
->seeky_start
)
2329 cfqq
->seeky_start
= jiffies
;
2330 else if (!CFQQ_SEEKY(cfqq
))
2331 cfqq
->seeky_start
= 0;
2336 * Disable idle window if the process thinks too long or seeks so much that
2340 cfq_update_idle_window(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
2341 struct cfq_io_context
*cic
)
2343 int old_idle
, enable_idle
;
2346 * Don't idle for async or idle io prio class
2348 if (!cfq_cfqq_sync(cfqq
) || cfq_class_idle(cfqq
))
2351 enable_idle
= old_idle
= cfq_cfqq_idle_window(cfqq
);
2353 if (cfqq
->queued
[0] + cfqq
->queued
[1] >= 4)
2354 cfq_mark_cfqq_deep(cfqq
);
2356 if (!atomic_read(&cic
->ioc
->nr_tasks
) || !cfqd
->cfq_slice_idle
||
2357 (!cfq_cfqq_deep(cfqq
) && sample_valid(cfqq
->seek_samples
)
2358 && CFQQ_SEEKY(cfqq
)))
2360 else if (sample_valid(cic
->ttime_samples
)) {
2361 if (cic
->ttime_mean
> cfqd
->cfq_slice_idle
)
2367 if (old_idle
!= enable_idle
) {
2368 cfq_log_cfqq(cfqd
, cfqq
, "idle=%d", enable_idle
);
2370 cfq_mark_cfqq_idle_window(cfqq
);
2372 cfq_clear_cfqq_idle_window(cfqq
);
2377 * Check if new_cfqq should preempt the currently active queue. Return 0 for
2378 * no or if we aren't sure, a 1 will cause a preempt.
2381 cfq_should_preempt(struct cfq_data
*cfqd
, struct cfq_queue
*new_cfqq
,
2384 struct cfq_queue
*cfqq
;
2386 cfqq
= cfqd
->active_queue
;
2390 if (cfq_slice_used(cfqq
))
2393 if (cfq_class_idle(new_cfqq
))
2396 if (cfq_class_idle(cfqq
))
2399 if (cfqd
->serving_type
== SYNC_NOIDLE_WORKLOAD
&&
2400 cfqq_type(new_cfqq
) == SYNC_NOIDLE_WORKLOAD
&&
2401 new_cfqq
->service_tree
->count
== 1)
2405 * if the new request is sync, but the currently running queue is
2406 * not, let the sync request have priority.
2408 if (rq_is_sync(rq
) && !cfq_cfqq_sync(cfqq
))
2412 * So both queues are sync. Let the new request get disk time if
2413 * it's a metadata request and the current queue is doing regular IO.
2415 if (rq_is_meta(rq
) && !cfqq
->meta_pending
)
2419 * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
2421 if (cfq_class_rt(new_cfqq
) && !cfq_class_rt(cfqq
))
2424 if (!cfqd
->active_cic
|| !cfq_cfqq_wait_request(cfqq
))
2428 * if this request is as-good as one we would expect from the
2429 * current cfqq, let it preempt
2431 if (cfq_rq_close(cfqd
, cfqq
, rq
))
2438 * cfqq preempts the active queue. if we allowed preempt with no slice left,
2439 * let it have half of its nominal slice.
2441 static void cfq_preempt_queue(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
2443 cfq_log_cfqq(cfqd
, cfqq
, "preempt");
2444 cfq_slice_expired(cfqd
, 1);
2447 * Put the new queue at the front of the of the current list,
2448 * so we know that it will be selected next.
2450 BUG_ON(!cfq_cfqq_on_rr(cfqq
));
2452 cfq_service_tree_add(cfqd
, cfqq
, 1);
2454 cfqq
->slice_end
= 0;
2455 cfq_mark_cfqq_slice_new(cfqq
);
2459 * Called when a new fs request (rq) is added (to cfqq). Check if there's
2460 * something we should do about it
2463 cfq_rq_enqueued(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
2466 struct cfq_io_context
*cic
= RQ_CIC(rq
);
2470 cfqq
->meta_pending
++;
2472 cfq_update_io_thinktime(cfqd
, cic
);
2473 cfq_update_io_seektime(cfqd
, cfqq
, rq
);
2474 cfq_update_idle_window(cfqd
, cfqq
, cic
);
2476 cfqq
->last_request_pos
= blk_rq_pos(rq
) + blk_rq_sectors(rq
);
2478 if (cfqq
== cfqd
->active_queue
) {
2480 * Remember that we saw a request from this process, but
2481 * don't start queuing just yet. Otherwise we risk seeing lots
2482 * of tiny requests, because we disrupt the normal plugging
2483 * and merging. If the request is already larger than a single
2484 * page, let it rip immediately. For that case we assume that
2485 * merging is already done. Ditto for a busy system that
2486 * has other work pending, don't risk delaying until the
2487 * idle timer unplug to continue working.
2489 if (cfq_cfqq_wait_request(cfqq
)) {
2490 if (blk_rq_bytes(rq
) > PAGE_CACHE_SIZE
||
2491 cfqd
->busy_queues
> 1) {
2492 del_timer(&cfqd
->idle_slice_timer
);
2493 __blk_run_queue(cfqd
->queue
);
2495 cfq_mark_cfqq_must_dispatch(cfqq
);
2497 } else if (cfq_should_preempt(cfqd
, cfqq
, rq
)) {
2499 * not the active queue - expire current slice if it is
2500 * idle and has expired it's mean thinktime or this new queue
2501 * has some old slice time left and is of higher priority or
2502 * this new queue is RT and the current one is BE
2504 cfq_preempt_queue(cfqd
, cfqq
);
2505 __blk_run_queue(cfqd
->queue
);
2509 static void cfq_insert_request(struct request_queue
*q
, struct request
*rq
)
2511 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
2512 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
2514 cfq_log_cfqq(cfqd
, cfqq
, "insert_request");
2515 cfq_init_prio_data(cfqq
, RQ_CIC(rq
)->ioc
);
2517 rq_set_fifo_time(rq
, jiffies
+ cfqd
->cfq_fifo_expire
[rq_is_sync(rq
)]);
2518 list_add_tail(&rq
->queuelist
, &cfqq
->fifo
);
2521 cfq_rq_enqueued(cfqd
, cfqq
, rq
);
2525 * Update hw_tag based on peak queue depth over 50 samples under
2528 static void cfq_update_hw_tag(struct cfq_data
*cfqd
)
2530 struct cfq_queue
*cfqq
= cfqd
->active_queue
;
2532 if (rq_in_driver(cfqd
) > cfqd
->hw_tag_est_depth
)
2533 cfqd
->hw_tag_est_depth
= rq_in_driver(cfqd
);
2535 if (cfqd
->hw_tag
== 1)
2538 if (cfqd
->rq_queued
<= CFQ_HW_QUEUE_MIN
&&
2539 rq_in_driver(cfqd
) <= CFQ_HW_QUEUE_MIN
)
2543 * If active queue hasn't enough requests and can idle, cfq might not
2544 * dispatch sufficient requests to hardware. Don't zero hw_tag in this
2547 if (cfqq
&& cfq_cfqq_idle_window(cfqq
) &&
2548 cfqq
->dispatched
+ cfqq
->queued
[0] + cfqq
->queued
[1] <
2549 CFQ_HW_QUEUE_MIN
&& rq_in_driver(cfqd
) < CFQ_HW_QUEUE_MIN
)
2552 if (cfqd
->hw_tag_samples
++ < 50)
2555 if (cfqd
->hw_tag_est_depth
>= CFQ_HW_QUEUE_MIN
)
2561 static void cfq_completed_request(struct request_queue
*q
, struct request
*rq
)
2563 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
2564 struct cfq_data
*cfqd
= cfqq
->cfqd
;
2565 const int sync
= rq_is_sync(rq
);
2569 cfq_log_cfqq(cfqd
, cfqq
, "complete");
2571 cfq_update_hw_tag(cfqd
);
2573 WARN_ON(!cfqd
->rq_in_driver
[sync
]);
2574 WARN_ON(!cfqq
->dispatched
);
2575 cfqd
->rq_in_driver
[sync
]--;
2578 if (cfq_cfqq_sync(cfqq
))
2579 cfqd
->sync_flight
--;
2582 RQ_CIC(rq
)->last_end_request
= now
;
2583 cfqd
->last_end_sync_rq
= now
;
2587 * If this is the active queue, check if it needs to be expired,
2588 * or if we want to idle in case it has no pending requests.
2590 if (cfqd
->active_queue
== cfqq
) {
2591 const bool cfqq_empty
= RB_EMPTY_ROOT(&cfqq
->sort_list
);
2593 if (cfq_cfqq_slice_new(cfqq
)) {
2594 cfq_set_prio_slice(cfqd
, cfqq
);
2595 cfq_clear_cfqq_slice_new(cfqq
);
2598 * Idling is not enabled on:
2600 * - idle-priority queues
2602 * - queues with still some requests queued
2603 * - when there is a close cooperator
2605 if (cfq_slice_used(cfqq
) || cfq_class_idle(cfqq
))
2606 cfq_slice_expired(cfqd
, 1);
2607 else if (sync
&& cfqq_empty
&&
2608 !cfq_close_cooperator(cfqd
, cfqq
)) {
2609 cfqd
->noidle_tree_requires_idle
|= !rq_noidle(rq
);
2611 * Idling is enabled for SYNC_WORKLOAD.
2612 * SYNC_NOIDLE_WORKLOAD idles at the end of the tree
2613 * only if we processed at least one !rq_noidle request
2615 if (cfqd
->serving_type
== SYNC_WORKLOAD
2616 || cfqd
->noidle_tree_requires_idle
)
2617 cfq_arm_slice_timer(cfqd
);
2621 if (!rq_in_driver(cfqd
))
2622 cfq_schedule_dispatch(cfqd
);
2626 * we temporarily boost lower priority queues if they are holding fs exclusive
2627 * resources. they are boosted to normal prio (CLASS_BE/4)
2629 static void cfq_prio_boost(struct cfq_queue
*cfqq
)
2631 if (has_fs_excl()) {
2633 * boost idle prio on transactions that would lock out other
2634 * users of the filesystem
2636 if (cfq_class_idle(cfqq
))
2637 cfqq
->ioprio_class
= IOPRIO_CLASS_BE
;
2638 if (cfqq
->ioprio
> IOPRIO_NORM
)
2639 cfqq
->ioprio
= IOPRIO_NORM
;
2642 * unboost the queue (if needed)
2644 cfqq
->ioprio_class
= cfqq
->org_ioprio_class
;
2645 cfqq
->ioprio
= cfqq
->org_ioprio
;
2649 static inline int __cfq_may_queue(struct cfq_queue
*cfqq
)
2651 if (cfq_cfqq_wait_request(cfqq
) && !cfq_cfqq_must_alloc_slice(cfqq
)) {
2652 cfq_mark_cfqq_must_alloc_slice(cfqq
);
2653 return ELV_MQUEUE_MUST
;
2656 return ELV_MQUEUE_MAY
;
2659 static int cfq_may_queue(struct request_queue
*q
, int rw
)
2661 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
2662 struct task_struct
*tsk
= current
;
2663 struct cfq_io_context
*cic
;
2664 struct cfq_queue
*cfqq
;
2667 * don't force setup of a queue from here, as a call to may_queue
2668 * does not necessarily imply that a request actually will be queued.
2669 * so just lookup a possibly existing queue, or return 'may queue'
2672 cic
= cfq_cic_lookup(cfqd
, tsk
->io_context
);
2674 return ELV_MQUEUE_MAY
;
2676 cfqq
= cic_to_cfqq(cic
, rw_is_sync(rw
));
2678 cfq_init_prio_data(cfqq
, cic
->ioc
);
2679 cfq_prio_boost(cfqq
);
2681 return __cfq_may_queue(cfqq
);
2684 return ELV_MQUEUE_MAY
;
2688 * queue lock held here
2690 static void cfq_put_request(struct request
*rq
)
2692 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
2695 const int rw
= rq_data_dir(rq
);
2697 BUG_ON(!cfqq
->allocated
[rw
]);
2698 cfqq
->allocated
[rw
]--;
2700 put_io_context(RQ_CIC(rq
)->ioc
);
2702 rq
->elevator_private
= NULL
;
2703 rq
->elevator_private2
= NULL
;
2705 cfq_put_queue(cfqq
);
2709 static struct cfq_queue
*
2710 cfq_merge_cfqqs(struct cfq_data
*cfqd
, struct cfq_io_context
*cic
,
2711 struct cfq_queue
*cfqq
)
2713 cfq_log_cfqq(cfqd
, cfqq
, "merging with queue %p", cfqq
->new_cfqq
);
2714 cic_set_cfqq(cic
, cfqq
->new_cfqq
, 1);
2715 cfq_mark_cfqq_coop(cfqq
->new_cfqq
);
2716 cfq_put_queue(cfqq
);
2717 return cic_to_cfqq(cic
, 1);
2720 static int should_split_cfqq(struct cfq_queue
*cfqq
)
2722 if (cfqq
->seeky_start
&&
2723 time_after(jiffies
, cfqq
->seeky_start
+ CFQQ_COOP_TOUT
))
2729 * Returns NULL if a new cfqq should be allocated, or the old cfqq if this
2730 * was the last process referring to said cfqq.
2732 static struct cfq_queue
*
2733 split_cfqq(struct cfq_io_context
*cic
, struct cfq_queue
*cfqq
)
2735 if (cfqq_process_refs(cfqq
) == 1) {
2736 cfqq
->seeky_start
= 0;
2737 cfqq
->pid
= current
->pid
;
2738 cfq_clear_cfqq_coop(cfqq
);
2742 cic_set_cfqq(cic
, NULL
, 1);
2743 cfq_put_queue(cfqq
);
2747 * Allocate cfq data structures associated with this request.
2750 cfq_set_request(struct request_queue
*q
, struct request
*rq
, gfp_t gfp_mask
)
2752 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
2753 struct cfq_io_context
*cic
;
2754 const int rw
= rq_data_dir(rq
);
2755 const bool is_sync
= rq_is_sync(rq
);
2756 struct cfq_queue
*cfqq
;
2757 unsigned long flags
;
2759 might_sleep_if(gfp_mask
& __GFP_WAIT
);
2761 cic
= cfq_get_io_context(cfqd
, gfp_mask
);
2763 spin_lock_irqsave(q
->queue_lock
, flags
);
2769 cfqq
= cic_to_cfqq(cic
, is_sync
);
2770 if (!cfqq
|| cfqq
== &cfqd
->oom_cfqq
) {
2771 cfqq
= cfq_get_queue(cfqd
, is_sync
, cic
->ioc
, gfp_mask
);
2772 cic_set_cfqq(cic
, cfqq
, is_sync
);
2775 * If the queue was seeky for too long, break it apart.
2777 if (cfq_cfqq_coop(cfqq
) && should_split_cfqq(cfqq
)) {
2778 cfq_log_cfqq(cfqd
, cfqq
, "breaking apart cfqq");
2779 cfqq
= split_cfqq(cic
, cfqq
);
2785 * Check to see if this queue is scheduled to merge with
2786 * another, closely cooperating queue. The merging of
2787 * queues happens here as it must be done in process context.
2788 * The reference on new_cfqq was taken in merge_cfqqs.
2791 cfqq
= cfq_merge_cfqqs(cfqd
, cic
, cfqq
);
2794 cfqq
->allocated
[rw
]++;
2795 atomic_inc(&cfqq
->ref
);
2797 spin_unlock_irqrestore(q
->queue_lock
, flags
);
2799 rq
->elevator_private
= cic
;
2800 rq
->elevator_private2
= cfqq
;
2805 put_io_context(cic
->ioc
);
2807 cfq_schedule_dispatch(cfqd
);
2808 spin_unlock_irqrestore(q
->queue_lock
, flags
);
2809 cfq_log(cfqd
, "set_request fail");
2813 static void cfq_kick_queue(struct work_struct
*work
)
2815 struct cfq_data
*cfqd
=
2816 container_of(work
, struct cfq_data
, unplug_work
);
2817 struct request_queue
*q
= cfqd
->queue
;
2819 spin_lock_irq(q
->queue_lock
);
2820 __blk_run_queue(cfqd
->queue
);
2821 spin_unlock_irq(q
->queue_lock
);
2825 * Timer running if the active_queue is currently idling inside its time slice
2827 static void cfq_idle_slice_timer(unsigned long data
)
2829 struct cfq_data
*cfqd
= (struct cfq_data
*) data
;
2830 struct cfq_queue
*cfqq
;
2831 unsigned long flags
;
2834 cfq_log(cfqd
, "idle timer fired");
2836 spin_lock_irqsave(cfqd
->queue
->queue_lock
, flags
);
2838 cfqq
= cfqd
->active_queue
;
2843 * We saw a request before the queue expired, let it through
2845 if (cfq_cfqq_must_dispatch(cfqq
))
2851 if (cfq_slice_used(cfqq
))
2855 * only expire and reinvoke request handler, if there are
2856 * other queues with pending requests
2858 if (!cfqd
->busy_queues
)
2862 * not expired and it has a request pending, let it dispatch
2864 if (!RB_EMPTY_ROOT(&cfqq
->sort_list
))
2868 * Queue depth flag is reset only when the idle didn't succeed
2870 cfq_clear_cfqq_deep(cfqq
);
2873 cfq_slice_expired(cfqd
, timed_out
);
2875 cfq_schedule_dispatch(cfqd
);
2877 spin_unlock_irqrestore(cfqd
->queue
->queue_lock
, flags
);
2880 static void cfq_shutdown_timer_wq(struct cfq_data
*cfqd
)
2882 del_timer_sync(&cfqd
->idle_slice_timer
);
2883 cancel_work_sync(&cfqd
->unplug_work
);
2886 static void cfq_put_async_queues(struct cfq_data
*cfqd
)
2890 for (i
= 0; i
< IOPRIO_BE_NR
; i
++) {
2891 if (cfqd
->async_cfqq
[0][i
])
2892 cfq_put_queue(cfqd
->async_cfqq
[0][i
]);
2893 if (cfqd
->async_cfqq
[1][i
])
2894 cfq_put_queue(cfqd
->async_cfqq
[1][i
]);
2897 if (cfqd
->async_idle_cfqq
)
2898 cfq_put_queue(cfqd
->async_idle_cfqq
);
2901 static void cfq_exit_queue(struct elevator_queue
*e
)
2903 struct cfq_data
*cfqd
= e
->elevator_data
;
2904 struct request_queue
*q
= cfqd
->queue
;
2906 cfq_shutdown_timer_wq(cfqd
);
2908 spin_lock_irq(q
->queue_lock
);
2910 if (cfqd
->active_queue
)
2911 __cfq_slice_expired(cfqd
, cfqd
->active_queue
, 0);
2913 while (!list_empty(&cfqd
->cic_list
)) {
2914 struct cfq_io_context
*cic
= list_entry(cfqd
->cic_list
.next
,
2915 struct cfq_io_context
,
2918 __cfq_exit_single_io_context(cfqd
, cic
);
2921 cfq_put_async_queues(cfqd
);
2923 spin_unlock_irq(q
->queue_lock
);
2925 cfq_shutdown_timer_wq(cfqd
);
2930 static void *cfq_init_queue(struct request_queue
*q
)
2932 struct cfq_data
*cfqd
;
2935 cfqd
= kmalloc_node(sizeof(*cfqd
), GFP_KERNEL
| __GFP_ZERO
, q
->node
);
2939 for (i
= 0; i
< 2; ++i
)
2940 for (j
= 0; j
< 3; ++j
)
2941 cfqd
->service_trees
[i
][j
] = CFQ_RB_ROOT
;
2942 cfqd
->service_tree_idle
= CFQ_RB_ROOT
;
2945 * Not strictly needed (since RB_ROOT just clears the node and we
2946 * zeroed cfqd on alloc), but better be safe in case someone decides
2947 * to add magic to the rb code
2949 for (i
= 0; i
< CFQ_PRIO_LISTS
; i
++)
2950 cfqd
->prio_trees
[i
] = RB_ROOT
;
2953 * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
2954 * Grab a permanent reference to it, so that the normal code flow
2955 * will not attempt to free it.
2957 cfq_init_cfqq(cfqd
, &cfqd
->oom_cfqq
, 1, 0);
2958 atomic_inc(&cfqd
->oom_cfqq
.ref
);
2960 INIT_LIST_HEAD(&cfqd
->cic_list
);
2964 init_timer(&cfqd
->idle_slice_timer
);
2965 cfqd
->idle_slice_timer
.function
= cfq_idle_slice_timer
;
2966 cfqd
->idle_slice_timer
.data
= (unsigned long) cfqd
;
2968 INIT_WORK(&cfqd
->unplug_work
, cfq_kick_queue
);
2970 cfqd
->cfq_quantum
= cfq_quantum
;
2971 cfqd
->cfq_fifo_expire
[0] = cfq_fifo_expire
[0];
2972 cfqd
->cfq_fifo_expire
[1] = cfq_fifo_expire
[1];
2973 cfqd
->cfq_back_max
= cfq_back_max
;
2974 cfqd
->cfq_back_penalty
= cfq_back_penalty
;
2975 cfqd
->cfq_slice
[0] = cfq_slice_async
;
2976 cfqd
->cfq_slice
[1] = cfq_slice_sync
;
2977 cfqd
->cfq_slice_async_rq
= cfq_slice_async_rq
;
2978 cfqd
->cfq_slice_idle
= cfq_slice_idle
;
2979 cfqd
->cfq_latency
= 1;
2981 cfqd
->last_end_sync_rq
= jiffies
;
2985 static void cfq_slab_kill(void)
2988 * Caller already ensured that pending RCU callbacks are completed,
2989 * so we should have no busy allocations at this point.
2992 kmem_cache_destroy(cfq_pool
);
2994 kmem_cache_destroy(cfq_ioc_pool
);
2997 static int __init
cfq_slab_setup(void)
2999 cfq_pool
= KMEM_CACHE(cfq_queue
, 0);
3003 cfq_ioc_pool
= KMEM_CACHE(cfq_io_context
, 0);
3014 * sysfs parts below -->
3017 cfq_var_show(unsigned int var
, char *page
)
3019 return sprintf(page
, "%d\n", var
);
3023 cfq_var_store(unsigned int *var
, const char *page
, size_t count
)
3025 char *p
= (char *) page
;
3027 *var
= simple_strtoul(p
, &p
, 10);
3031 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
3032 static ssize_t __FUNC(struct elevator_queue *e, char *page) \
3034 struct cfq_data *cfqd = e->elevator_data; \
3035 unsigned int __data = __VAR; \
3037 __data = jiffies_to_msecs(__data); \
3038 return cfq_var_show(__data, (page)); \
3040 SHOW_FUNCTION(cfq_quantum_show
, cfqd
->cfq_quantum
, 0);
3041 SHOW_FUNCTION(cfq_fifo_expire_sync_show
, cfqd
->cfq_fifo_expire
[1], 1);
3042 SHOW_FUNCTION(cfq_fifo_expire_async_show
, cfqd
->cfq_fifo_expire
[0], 1);
3043 SHOW_FUNCTION(cfq_back_seek_max_show
, cfqd
->cfq_back_max
, 0);
3044 SHOW_FUNCTION(cfq_back_seek_penalty_show
, cfqd
->cfq_back_penalty
, 0);
3045 SHOW_FUNCTION(cfq_slice_idle_show
, cfqd
->cfq_slice_idle
, 1);
3046 SHOW_FUNCTION(cfq_slice_sync_show
, cfqd
->cfq_slice
[1], 1);
3047 SHOW_FUNCTION(cfq_slice_async_show
, cfqd
->cfq_slice
[0], 1);
3048 SHOW_FUNCTION(cfq_slice_async_rq_show
, cfqd
->cfq_slice_async_rq
, 0);
3049 SHOW_FUNCTION(cfq_low_latency_show
, cfqd
->cfq_latency
, 0);
3050 #undef SHOW_FUNCTION
3052 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
3053 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
3055 struct cfq_data *cfqd = e->elevator_data; \
3056 unsigned int __data; \
3057 int ret = cfq_var_store(&__data, (page), count); \
3058 if (__data < (MIN)) \
3060 else if (__data > (MAX)) \
3063 *(__PTR) = msecs_to_jiffies(__data); \
3065 *(__PTR) = __data; \
3068 STORE_FUNCTION(cfq_quantum_store
, &cfqd
->cfq_quantum
, 1, UINT_MAX
, 0);
3069 STORE_FUNCTION(cfq_fifo_expire_sync_store
, &cfqd
->cfq_fifo_expire
[1], 1,
3071 STORE_FUNCTION(cfq_fifo_expire_async_store
, &cfqd
->cfq_fifo_expire
[0], 1,
3073 STORE_FUNCTION(cfq_back_seek_max_store
, &cfqd
->cfq_back_max
, 0, UINT_MAX
, 0);
3074 STORE_FUNCTION(cfq_back_seek_penalty_store
, &cfqd
->cfq_back_penalty
, 1,
3076 STORE_FUNCTION(cfq_slice_idle_store
, &cfqd
->cfq_slice_idle
, 0, UINT_MAX
, 1);
3077 STORE_FUNCTION(cfq_slice_sync_store
, &cfqd
->cfq_slice
[1], 1, UINT_MAX
, 1);
3078 STORE_FUNCTION(cfq_slice_async_store
, &cfqd
->cfq_slice
[0], 1, UINT_MAX
, 1);
3079 STORE_FUNCTION(cfq_slice_async_rq_store
, &cfqd
->cfq_slice_async_rq
, 1,
3081 STORE_FUNCTION(cfq_low_latency_store
, &cfqd
->cfq_latency
, 0, 1, 0);
3082 #undef STORE_FUNCTION
3084 #define CFQ_ATTR(name) \
3085 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
3087 static struct elv_fs_entry cfq_attrs
[] = {
3089 CFQ_ATTR(fifo_expire_sync
),
3090 CFQ_ATTR(fifo_expire_async
),
3091 CFQ_ATTR(back_seek_max
),
3092 CFQ_ATTR(back_seek_penalty
),
3093 CFQ_ATTR(slice_sync
),
3094 CFQ_ATTR(slice_async
),
3095 CFQ_ATTR(slice_async_rq
),
3096 CFQ_ATTR(slice_idle
),
3097 CFQ_ATTR(low_latency
),
3101 static struct elevator_type iosched_cfq
= {
3103 .elevator_merge_fn
= cfq_merge
,
3104 .elevator_merged_fn
= cfq_merged_request
,
3105 .elevator_merge_req_fn
= cfq_merged_requests
,
3106 .elevator_allow_merge_fn
= cfq_allow_merge
,
3107 .elevator_dispatch_fn
= cfq_dispatch_requests
,
3108 .elevator_add_req_fn
= cfq_insert_request
,
3109 .elevator_activate_req_fn
= cfq_activate_request
,
3110 .elevator_deactivate_req_fn
= cfq_deactivate_request
,
3111 .elevator_queue_empty_fn
= cfq_queue_empty
,
3112 .elevator_completed_req_fn
= cfq_completed_request
,
3113 .elevator_former_req_fn
= elv_rb_former_request
,
3114 .elevator_latter_req_fn
= elv_rb_latter_request
,
3115 .elevator_set_req_fn
= cfq_set_request
,
3116 .elevator_put_req_fn
= cfq_put_request
,
3117 .elevator_may_queue_fn
= cfq_may_queue
,
3118 .elevator_init_fn
= cfq_init_queue
,
3119 .elevator_exit_fn
= cfq_exit_queue
,
3120 .trim
= cfq_free_io_context
,
3122 .elevator_attrs
= cfq_attrs
,
3123 .elevator_name
= "cfq",
3124 .elevator_owner
= THIS_MODULE
,
3127 static int __init
cfq_init(void)
3130 * could be 0 on HZ < 1000 setups
3132 if (!cfq_slice_async
)
3133 cfq_slice_async
= 1;
3134 if (!cfq_slice_idle
)
3137 if (cfq_slab_setup())
3140 elv_register(&iosched_cfq
);
3145 static void __exit
cfq_exit(void)
3147 DECLARE_COMPLETION_ONSTACK(all_gone
);
3148 elv_unregister(&iosched_cfq
);
3149 ioc_gone
= &all_gone
;
3150 /* ioc_gone's update must be visible before reading ioc_count */
3154 * this also protects us from entering cfq_slab_kill() with
3155 * pending RCU callbacks
3157 if (elv_ioc_count_read(cfq_ioc_count
))
3158 wait_for_completion(&all_gone
);
3162 module_init(cfq_init
);
3163 module_exit(cfq_exit
);
3165 MODULE_AUTHOR("Jens Axboe");
3166 MODULE_LICENSE("GPL");
3167 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");