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/rbtree.h>
13 #include <linux/ioprio.h>
14 #include <linux/blktrace_api.h>
19 /* max queue in one round of service */
20 static const int cfq_quantum = 4;
21 static const int cfq_fifo_expire[2] = { HZ / 4, HZ / 8 };
22 /* maximum backwards seek, in KiB */
23 static const int cfq_back_max = 16 * 1024;
24 /* penalty of a backwards seek */
25 static const int cfq_back_penalty = 2;
26 static const int cfq_slice_sync = HZ / 10;
27 static int cfq_slice_async = HZ / 25;
28 static const int cfq_slice_async_rq = 2;
29 static int cfq_slice_idle = HZ / 125;
32 * offset from end of service tree
34 #define CFQ_IDLE_DELAY (HZ / 5)
37 * below this threshold, we consider thinktime immediate
39 #define CFQ_MIN_TT (2)
41 #define CFQ_SLICE_SCALE (5)
42 #define CFQ_HW_QUEUE_MIN (5)
45 ((struct cfq_io_context *) (rq)->elevator_private)
46 #define RQ_CFQQ(rq) (struct cfq_queue *) ((rq)->elevator_private2)
48 static struct kmem_cache *cfq_pool;
49 static struct kmem_cache *cfq_ioc_pool;
51 static DEFINE_PER_CPU(unsigned long, ioc_count);
52 static struct completion *ioc_gone;
53 static DEFINE_SPINLOCK(ioc_gone_lock);
55 #define CFQ_PRIO_LISTS IOPRIO_BE_NR
56 #define cfq_class_idle(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
57 #define cfq_class_rt(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
59 #define sample_valid(samples) ((samples) > 80)
62 * Most of our rbtree usage is for sorting with min extraction, so
63 * if we cache the leftmost node we don't have to walk down the tree
64 * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should
65 * move this into the elevator for the rq sorting as well.
71 #define CFQ_RB_ROOT (struct cfq_rb_root) { RB_ROOT, NULL, }
74 * Per block device queue structure
77 struct request_queue *queue;
80 * rr list of queues with requests and the count of them
82 struct cfq_rb_root service_tree;
85 * Each priority tree is sorted by next_request position. These
86 * trees are used when determining if two or more queues are
87 * interleaving requests (see cfq_close_cooperator).
89 struct rb_root prio_trees[CFQ_PRIO_LISTS];
91 unsigned int busy_queues;
93 * Used to track any pending rt requests so we can pre-empt current
94 * non-RT cfqq in service when this value is non-zero.
96 unsigned int busy_rt_queues;
102 * queue-depth detection
107 int rq_in_driver_peak;
110 * idle window management
112 struct timer_list idle_slice_timer;
113 struct work_struct unplug_work;
115 struct cfq_queue *active_queue;
116 struct cfq_io_context *active_cic;
119 * async queue for each priority case
121 struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
122 struct cfq_queue *async_idle_cfqq;
124 sector_t last_position;
125 unsigned long last_end_request;
128 * tunables, see top of file
130 unsigned int cfq_quantum;
131 unsigned int cfq_fifo_expire[2];
132 unsigned int cfq_back_penalty;
133 unsigned int cfq_back_max;
134 unsigned int cfq_slice[2];
135 unsigned int cfq_slice_async_rq;
136 unsigned int cfq_slice_idle;
138 struct list_head cic_list;
142 * Per process-grouping structure
145 /* reference count */
147 /* various state flags, see below */
149 /* parent cfq_data */
150 struct cfq_data *cfqd;
151 /* service_tree member */
152 struct rb_node rb_node;
153 /* service_tree key */
154 unsigned long rb_key;
155 /* prio tree member */
156 struct rb_node p_node;
157 /* sorted list of pending requests */
158 struct rb_root sort_list;
159 /* if fifo isn't expired, next request to serve */
160 struct request *next_rq;
161 /* requests queued in sort_list */
163 /* currently allocated requests */
165 /* fifo list of requests in sort_list */
166 struct list_head fifo;
168 unsigned long slice_end;
170 unsigned int slice_dispatch;
172 /* pending metadata requests */
174 /* number of requests that are on the dispatch list or inside driver */
177 /* io prio of this group */
178 unsigned short ioprio, org_ioprio;
179 unsigned short ioprio_class, org_ioprio_class;
184 enum cfqq_state_flags {
185 CFQ_CFQQ_FLAG_on_rr = 0, /* on round-robin busy list */
186 CFQ_CFQQ_FLAG_wait_request, /* waiting for a request */
187 CFQ_CFQQ_FLAG_must_dispatch, /* must be allowed a dispatch */
188 CFQ_CFQQ_FLAG_must_alloc, /* must be allowed rq alloc */
189 CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
190 CFQ_CFQQ_FLAG_fifo_expire, /* FIFO checked in this slice */
191 CFQ_CFQQ_FLAG_idle_window, /* slice idling enabled */
192 CFQ_CFQQ_FLAG_prio_changed, /* task priority has changed */
193 CFQ_CFQQ_FLAG_slice_new, /* no requests dispatched in slice */
194 CFQ_CFQQ_FLAG_sync, /* synchronous queue */
195 CFQ_CFQQ_FLAG_coop, /* has done a coop jump of the queue */
198 #define CFQ_CFQQ_FNS(name) \
199 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \
201 (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name); \
203 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \
205 (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \
207 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \
209 return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \
213 CFQ_CFQQ_FNS(wait_request);
214 CFQ_CFQQ_FNS(must_dispatch);
215 CFQ_CFQQ_FNS(must_alloc);
216 CFQ_CFQQ_FNS(must_alloc_slice);
217 CFQ_CFQQ_FNS(fifo_expire);
218 CFQ_CFQQ_FNS(idle_window);
219 CFQ_CFQQ_FNS(prio_changed);
220 CFQ_CFQQ_FNS(slice_new);
225 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
226 blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args)
227 #define cfq_log(cfqd, fmt, args...) \
228 blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
230 static void cfq_dispatch_insert(struct request_queue *, struct request *);
231 static struct cfq_queue *cfq_get_queue(struct cfq_data *, int,
232 struct io_context *, gfp_t);
233 static struct cfq_io_context *cfq_cic_lookup(struct cfq_data *,
234 struct io_context *);
236 static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_context *cic,
239 return cic->cfqq[!!is_sync];
242 static inline void cic_set_cfqq(struct cfq_io_context *cic,
243 struct cfq_queue *cfqq, int is_sync)
245 cic->cfqq[!!is_sync] = cfqq;
249 * We regard a request as SYNC, if it's either a read or has the SYNC bit
250 * set (in which case it could also be direct WRITE).
252 static inline int cfq_bio_sync(struct bio *bio)
254 if (bio_data_dir(bio) == READ || bio_sync(bio))
261 * scheduler run of queue, if there are requests pending and no one in the
262 * driver that will restart queueing
264 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
266 if (cfqd->busy_queues) {
267 cfq_log(cfqd, "schedule dispatch");
268 kblockd_schedule_work(cfqd->queue, &cfqd->unplug_work);
272 static int cfq_queue_empty(struct request_queue *q)
274 struct cfq_data *cfqd = q->elevator->elevator_data;
276 return !cfqd->busy_queues;
280 * Scale schedule slice based on io priority. Use the sync time slice only
281 * if a queue is marked sync and has sync io queued. A sync queue with async
282 * io only, should not get full sync slice length.
284 static inline int cfq_prio_slice(struct cfq_data *cfqd, int sync,
287 const int base_slice = cfqd->cfq_slice[sync];
289 WARN_ON(prio >= IOPRIO_BE_NR);
291 return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
295 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
297 return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
301 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
303 cfqq->slice_end = cfq_prio_to_slice(cfqd, cfqq) + jiffies;
304 cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
308 * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
309 * isn't valid until the first request from the dispatch is activated
310 * and the slice time set.
312 static inline int cfq_slice_used(struct cfq_queue *cfqq)
314 if (cfq_cfqq_slice_new(cfqq))
316 if (time_before(jiffies, cfqq->slice_end))
323 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
324 * We choose the request that is closest to the head right now. Distance
325 * behind the head is penalized and only allowed to a certain extent.
327 static struct request *
328 cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2)
330 sector_t last, s1, s2, d1 = 0, d2 = 0;
331 unsigned long back_max;
332 #define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */
333 #define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */
334 unsigned wrap = 0; /* bit mask: requests behind the disk head? */
336 if (rq1 == NULL || rq1 == rq2)
341 if (rq_is_sync(rq1) && !rq_is_sync(rq2))
343 else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
345 if (rq_is_meta(rq1) && !rq_is_meta(rq2))
347 else if (rq_is_meta(rq2) && !rq_is_meta(rq1))
353 last = cfqd->last_position;
356 * by definition, 1KiB is 2 sectors
358 back_max = cfqd->cfq_back_max * 2;
361 * Strict one way elevator _except_ in the case where we allow
362 * short backward seeks which are biased as twice the cost of a
363 * similar forward seek.
367 else if (s1 + back_max >= last)
368 d1 = (last - s1) * cfqd->cfq_back_penalty;
370 wrap |= CFQ_RQ1_WRAP;
374 else if (s2 + back_max >= last)
375 d2 = (last - s2) * cfqd->cfq_back_penalty;
377 wrap |= CFQ_RQ2_WRAP;
379 /* Found required data */
382 * By doing switch() on the bit mask "wrap" we avoid having to
383 * check two variables for all permutations: --> faster!
386 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
402 case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
405 * Since both rqs are wrapped,
406 * start with the one that's further behind head
407 * (--> only *one* back seek required),
408 * since back seek takes more time than forward.
418 * The below is leftmost cache rbtree addon
420 static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
423 root->left = rb_first(&root->rb);
426 return rb_entry(root->left, struct cfq_queue, rb_node);
431 static void rb_erase_init(struct rb_node *n, struct rb_root *root)
437 static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
441 rb_erase_init(n, &root->rb);
445 * would be nice to take fifo expire time into account as well
447 static struct request *
448 cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
449 struct request *last)
451 struct rb_node *rbnext = rb_next(&last->rb_node);
452 struct rb_node *rbprev = rb_prev(&last->rb_node);
453 struct request *next = NULL, *prev = NULL;
455 BUG_ON(RB_EMPTY_NODE(&last->rb_node));
458 prev = rb_entry_rq(rbprev);
461 next = rb_entry_rq(rbnext);
463 rbnext = rb_first(&cfqq->sort_list);
464 if (rbnext && rbnext != &last->rb_node)
465 next = rb_entry_rq(rbnext);
468 return cfq_choose_req(cfqd, next, prev);
471 static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
472 struct cfq_queue *cfqq)
475 * just an approximation, should be ok.
477 return (cfqd->busy_queues - 1) * (cfq_prio_slice(cfqd, 1, 0) -
478 cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
482 * The cfqd->service_tree holds all pending cfq_queue's that have
483 * requests waiting to be processed. It is sorted in the order that
484 * we will service the queues.
486 static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
489 struct rb_node **p, *parent;
490 struct cfq_queue *__cfqq;
491 unsigned long rb_key;
494 if (cfq_class_idle(cfqq)) {
495 rb_key = CFQ_IDLE_DELAY;
496 parent = rb_last(&cfqd->service_tree.rb);
497 if (parent && parent != &cfqq->rb_node) {
498 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
499 rb_key += __cfqq->rb_key;
502 } else if (!add_front) {
503 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
504 rb_key += cfqq->slice_resid;
505 cfqq->slice_resid = 0;
509 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
511 * same position, nothing more to do
513 if (rb_key == cfqq->rb_key)
516 cfq_rb_erase(&cfqq->rb_node, &cfqd->service_tree);
521 p = &cfqd->service_tree.rb.rb_node;
526 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
529 * sort RT queues first, we always want to give
530 * preference to them. IDLE queues goes to the back.
531 * after that, sort on the next service time.
533 if (cfq_class_rt(cfqq) > cfq_class_rt(__cfqq))
535 else if (cfq_class_rt(cfqq) < cfq_class_rt(__cfqq))
537 else if (cfq_class_idle(cfqq) < cfq_class_idle(__cfqq))
539 else if (cfq_class_idle(cfqq) > cfq_class_idle(__cfqq))
541 else if (rb_key < __cfqq->rb_key)
546 if (n == &(*p)->rb_right)
553 cfqd->service_tree.left = &cfqq->rb_node;
555 cfqq->rb_key = rb_key;
556 rb_link_node(&cfqq->rb_node, parent, p);
557 rb_insert_color(&cfqq->rb_node, &cfqd->service_tree.rb);
560 static struct cfq_queue *
561 cfq_prio_tree_lookup(struct cfq_data *cfqd, int ioprio, sector_t sector,
562 struct rb_node **ret_parent, struct rb_node ***rb_link)
564 struct rb_root *root = &cfqd->prio_trees[ioprio];
565 struct rb_node **p, *parent;
566 struct cfq_queue *cfqq = NULL;
574 cfqq = rb_entry(parent, struct cfq_queue, p_node);
577 * Sort strictly based on sector. Smallest to the left,
578 * largest to the right.
580 if (sector > cfqq->next_rq->sector)
582 else if (sector < cfqq->next_rq->sector)
590 *ret_parent = parent;
596 static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
598 struct rb_root *root = &cfqd->prio_trees[cfqq->ioprio];
599 struct rb_node **p, *parent;
600 struct cfq_queue *__cfqq;
602 if (!RB_EMPTY_NODE(&cfqq->p_node))
603 rb_erase_init(&cfqq->p_node, root);
605 if (cfq_class_idle(cfqq))
610 __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->ioprio, cfqq->next_rq->sector,
613 rb_link_node(&cfqq->p_node, parent, p);
614 rb_insert_color(&cfqq->p_node, root);
619 * Update cfqq's position in the service tree.
621 static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
624 * Resorting requires the cfqq to be on the RR list already.
626 if (cfq_cfqq_on_rr(cfqq)) {
627 cfq_service_tree_add(cfqd, cfqq, 0);
628 cfq_prio_tree_add(cfqd, cfqq);
633 * add to busy list of queues for service, trying to be fair in ordering
634 * the pending list according to last request service
636 static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
638 cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
639 BUG_ON(cfq_cfqq_on_rr(cfqq));
640 cfq_mark_cfqq_on_rr(cfqq);
642 if (cfq_class_rt(cfqq))
643 cfqd->busy_rt_queues++;
645 cfq_resort_rr_list(cfqd, cfqq);
649 * Called when the cfqq no longer has requests pending, remove it from
652 static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
654 cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
655 BUG_ON(!cfq_cfqq_on_rr(cfqq));
656 cfq_clear_cfqq_on_rr(cfqq);
658 if (!RB_EMPTY_NODE(&cfqq->rb_node))
659 cfq_rb_erase(&cfqq->rb_node, &cfqd->service_tree);
660 if (!RB_EMPTY_NODE(&cfqq->p_node))
661 rb_erase_init(&cfqq->p_node, &cfqd->prio_trees[cfqq->ioprio]);
663 BUG_ON(!cfqd->busy_queues);
665 if (cfq_class_rt(cfqq))
666 cfqd->busy_rt_queues--;
670 * rb tree support functions
672 static void cfq_del_rq_rb(struct request *rq)
674 struct cfq_queue *cfqq = RQ_CFQQ(rq);
675 struct cfq_data *cfqd = cfqq->cfqd;
676 const int sync = rq_is_sync(rq);
678 BUG_ON(!cfqq->queued[sync]);
679 cfqq->queued[sync]--;
681 elv_rb_del(&cfqq->sort_list, rq);
683 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
684 cfq_del_cfqq_rr(cfqd, cfqq);
687 static void cfq_add_rq_rb(struct request *rq)
689 struct cfq_queue *cfqq = RQ_CFQQ(rq);
690 struct cfq_data *cfqd = cfqq->cfqd;
691 struct request *__alias, *prev;
693 cfqq->queued[rq_is_sync(rq)]++;
696 * looks a little odd, but the first insert might return an alias.
697 * if that happens, put the alias on the dispatch list
699 while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL)
700 cfq_dispatch_insert(cfqd->queue, __alias);
702 if (!cfq_cfqq_on_rr(cfqq))
703 cfq_add_cfqq_rr(cfqd, cfqq);
706 * check if this request is a better next-serve candidate
708 prev = cfqq->next_rq;
709 cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq);
712 * adjust priority tree position, if ->next_rq changes
714 if (prev != cfqq->next_rq)
715 cfq_prio_tree_add(cfqd, cfqq);
717 BUG_ON(!cfqq->next_rq);
720 static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
722 elv_rb_del(&cfqq->sort_list, rq);
723 cfqq->queued[rq_is_sync(rq)]--;
727 static struct request *
728 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
730 struct task_struct *tsk = current;
731 struct cfq_io_context *cic;
732 struct cfq_queue *cfqq;
734 cic = cfq_cic_lookup(cfqd, tsk->io_context);
738 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
740 sector_t sector = bio->bi_sector + bio_sectors(bio);
742 return elv_rb_find(&cfqq->sort_list, sector);
748 static void cfq_activate_request(struct request_queue *q, struct request *rq)
750 struct cfq_data *cfqd = q->elevator->elevator_data;
752 cfqd->rq_in_driver++;
753 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
756 cfqd->last_position = rq->hard_sector + rq->hard_nr_sectors;
759 static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
761 struct cfq_data *cfqd = q->elevator->elevator_data;
763 WARN_ON(!cfqd->rq_in_driver);
764 cfqd->rq_in_driver--;
765 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
769 static void cfq_remove_request(struct request *rq)
771 struct cfq_queue *cfqq = RQ_CFQQ(rq);
773 if (cfqq->next_rq == rq)
774 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
776 list_del_init(&rq->queuelist);
779 cfqq->cfqd->rq_queued--;
780 if (rq_is_meta(rq)) {
781 WARN_ON(!cfqq->meta_pending);
782 cfqq->meta_pending--;
786 static int cfq_merge(struct request_queue *q, struct request **req,
789 struct cfq_data *cfqd = q->elevator->elevator_data;
790 struct request *__rq;
792 __rq = cfq_find_rq_fmerge(cfqd, bio);
793 if (__rq && elv_rq_merge_ok(__rq, bio)) {
795 return ELEVATOR_FRONT_MERGE;
798 return ELEVATOR_NO_MERGE;
801 static void cfq_merged_request(struct request_queue *q, struct request *req,
804 if (type == ELEVATOR_FRONT_MERGE) {
805 struct cfq_queue *cfqq = RQ_CFQQ(req);
807 cfq_reposition_rq_rb(cfqq, req);
812 cfq_merged_requests(struct request_queue *q, struct request *rq,
813 struct request *next)
816 * reposition in fifo if next is older than rq
818 if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
819 time_before(next->start_time, rq->start_time))
820 list_move(&rq->queuelist, &next->queuelist);
822 cfq_remove_request(next);
825 static int cfq_allow_merge(struct request_queue *q, struct request *rq,
828 struct cfq_data *cfqd = q->elevator->elevator_data;
829 struct cfq_io_context *cic;
830 struct cfq_queue *cfqq;
833 * Disallow merge of a sync bio into an async request.
835 if (cfq_bio_sync(bio) && !rq_is_sync(rq))
839 * Lookup the cfqq that this bio will be queued with. Allow
840 * merge only if rq is queued there.
842 cic = cfq_cic_lookup(cfqd, current->io_context);
846 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
847 if (cfqq == RQ_CFQQ(rq))
853 static void __cfq_set_active_queue(struct cfq_data *cfqd,
854 struct cfq_queue *cfqq)
857 cfq_log_cfqq(cfqd, cfqq, "set_active");
859 cfqq->slice_dispatch = 0;
861 cfq_clear_cfqq_wait_request(cfqq);
862 cfq_clear_cfqq_must_dispatch(cfqq);
863 cfq_clear_cfqq_must_alloc_slice(cfqq);
864 cfq_clear_cfqq_fifo_expire(cfqq);
865 cfq_mark_cfqq_slice_new(cfqq);
867 del_timer(&cfqd->idle_slice_timer);
870 cfqd->active_queue = cfqq;
874 * current cfqq expired its slice (or was too idle), select new one
877 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
880 cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
882 if (cfq_cfqq_wait_request(cfqq))
883 del_timer(&cfqd->idle_slice_timer);
885 cfq_clear_cfqq_wait_request(cfqq);
888 * store what was left of this slice, if the queue idled/timed out
890 if (timed_out && !cfq_cfqq_slice_new(cfqq)) {
891 cfqq->slice_resid = cfqq->slice_end - jiffies;
892 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
895 cfq_resort_rr_list(cfqd, cfqq);
897 if (cfqq == cfqd->active_queue)
898 cfqd->active_queue = NULL;
900 if (cfqd->active_cic) {
901 put_io_context(cfqd->active_cic->ioc);
902 cfqd->active_cic = NULL;
906 static inline void cfq_slice_expired(struct cfq_data *cfqd, int timed_out)
908 struct cfq_queue *cfqq = cfqd->active_queue;
911 __cfq_slice_expired(cfqd, cfqq, timed_out);
915 * Get next queue for service. Unless we have a queue preemption,
916 * we'll simply select the first cfqq in the service tree.
918 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
920 if (RB_EMPTY_ROOT(&cfqd->service_tree.rb))
923 return cfq_rb_first(&cfqd->service_tree);
927 * Get and set a new active queue for service.
929 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
930 struct cfq_queue *cfqq)
933 cfqq = cfq_get_next_queue(cfqd);
935 cfq_clear_cfqq_coop(cfqq);
938 __cfq_set_active_queue(cfqd, cfqq);
942 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
945 if (rq->sector >= cfqd->last_position)
946 return rq->sector - cfqd->last_position;
948 return cfqd->last_position - rq->sector;
951 #define CIC_SEEK_THR 8 * 1024
952 #define CIC_SEEKY(cic) ((cic)->seek_mean > CIC_SEEK_THR)
954 static inline int cfq_rq_close(struct cfq_data *cfqd, struct request *rq)
956 struct cfq_io_context *cic = cfqd->active_cic;
957 sector_t sdist = cic->seek_mean;
959 if (!sample_valid(cic->seek_samples))
960 sdist = CIC_SEEK_THR;
962 return cfq_dist_from_last(cfqd, rq) <= sdist;
965 static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
966 struct cfq_queue *cur_cfqq)
968 struct rb_root *root = &cfqd->prio_trees[cur_cfqq->ioprio];
969 struct rb_node *parent, *node;
970 struct cfq_queue *__cfqq;
971 sector_t sector = cfqd->last_position;
973 if (RB_EMPTY_ROOT(root))
977 * First, if we find a request starting at the end of the last
978 * request, choose it.
980 __cfqq = cfq_prio_tree_lookup(cfqd, cur_cfqq->ioprio,
981 sector, &parent, NULL);
986 * If the exact sector wasn't found, the parent of the NULL leaf
987 * will contain the closest sector.
989 __cfqq = rb_entry(parent, struct cfq_queue, p_node);
990 if (cfq_rq_close(cfqd, __cfqq->next_rq))
993 if (__cfqq->next_rq->sector < sector)
994 node = rb_next(&__cfqq->p_node);
996 node = rb_prev(&__cfqq->p_node);
1000 __cfqq = rb_entry(node, struct cfq_queue, p_node);
1001 if (cfq_rq_close(cfqd, __cfqq->next_rq))
1009 * cur_cfqq - passed in so that we don't decide that the current queue is
1010 * closely cooperating with itself.
1012 * So, basically we're assuming that that cur_cfqq has dispatched at least
1013 * one request, and that cfqd->last_position reflects a position on the disk
1014 * associated with the I/O issued by cur_cfqq. I'm not sure this is a valid
1017 static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
1018 struct cfq_queue *cur_cfqq,
1021 struct cfq_queue *cfqq;
1024 * A valid cfq_io_context is necessary to compare requests against
1025 * the seek_mean of the current cfqq.
1027 if (!cfqd->active_cic)
1031 * We should notice if some of the queues are cooperating, eg
1032 * working closely on the same area of the disk. In that case,
1033 * we can group them together and don't waste time idling.
1035 cfqq = cfqq_close(cfqd, cur_cfqq);
1039 if (cfq_cfqq_coop(cfqq))
1043 cfq_mark_cfqq_coop(cfqq);
1047 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
1049 struct cfq_queue *cfqq = cfqd->active_queue;
1050 struct cfq_io_context *cic;
1054 * SSD device without seek penalty, disable idling. But only do so
1055 * for devices that support queuing, otherwise we still have a problem
1056 * with sync vs async workloads.
1058 if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
1061 WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
1062 WARN_ON(cfq_cfqq_slice_new(cfqq));
1065 * idle is disabled, either manually or by past process history
1067 if (!cfqd->cfq_slice_idle || !cfq_cfqq_idle_window(cfqq))
1071 * still requests with the driver, don't idle
1073 if (cfqd->rq_in_driver)
1077 * task has exited, don't wait
1079 cic = cfqd->active_cic;
1080 if (!cic || !atomic_read(&cic->ioc->nr_tasks))
1083 cfq_mark_cfqq_wait_request(cfqq);
1086 * we don't want to idle for seeks, but we do want to allow
1087 * fair distribution of slice time for a process doing back-to-back
1088 * seeks. so allow a little bit of time for him to submit a new rq
1090 sl = cfqd->cfq_slice_idle;
1091 if (sample_valid(cic->seek_samples) && CIC_SEEKY(cic))
1092 sl = min(sl, msecs_to_jiffies(CFQ_MIN_TT));
1094 mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
1095 cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu", sl);
1099 * Move request from internal lists to the request queue dispatch list.
1101 static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
1103 struct cfq_data *cfqd = q->elevator->elevator_data;
1104 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1106 cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
1108 cfq_remove_request(rq);
1110 elv_dispatch_sort(q, rq);
1112 if (cfq_cfqq_sync(cfqq))
1113 cfqd->sync_flight++;
1117 * return expired entry, or NULL to just start from scratch in rbtree
1119 static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
1121 struct cfq_data *cfqd = cfqq->cfqd;
1125 if (cfq_cfqq_fifo_expire(cfqq))
1128 cfq_mark_cfqq_fifo_expire(cfqq);
1130 if (list_empty(&cfqq->fifo))
1133 fifo = cfq_cfqq_sync(cfqq);
1134 rq = rq_entry_fifo(cfqq->fifo.next);
1136 if (time_before(jiffies, rq->start_time + cfqd->cfq_fifo_expire[fifo]))
1139 cfq_log_cfqq(cfqd, cfqq, "fifo=%p", rq);
1144 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1146 const int base_rq = cfqd->cfq_slice_async_rq;
1148 WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
1150 return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
1154 * Select a queue for service. If we have a current active queue,
1155 * check whether to continue servicing it, or retrieve and set a new one.
1157 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
1159 struct cfq_queue *cfqq, *new_cfqq = NULL;
1161 cfqq = cfqd->active_queue;
1166 * The active queue has run out of time, expire it and select new.
1168 if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq))
1172 * If we have a RT cfqq waiting, then we pre-empt the current non-rt
1175 if (!cfq_class_rt(cfqq) && cfqd->busy_rt_queues) {
1177 * We simulate this as cfqq timed out so that it gets to bank
1178 * the remaining of its time slice.
1180 cfq_log_cfqq(cfqd, cfqq, "preempt");
1181 cfq_slice_expired(cfqd, 1);
1186 * The active queue has requests and isn't expired, allow it to
1189 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
1193 * If another queue has a request waiting within our mean seek
1194 * distance, let it run. The expire code will check for close
1195 * cooperators and put the close queue at the front of the service
1198 new_cfqq = cfq_close_cooperator(cfqd, cfqq, 0);
1203 * No requests pending. If the active queue still has requests in
1204 * flight or is idling for a new request, allow either of these
1205 * conditions to happen (or time out) before selecting a new queue.
1207 if (timer_pending(&cfqd->idle_slice_timer) ||
1208 (cfqq->dispatched && cfq_cfqq_idle_window(cfqq))) {
1214 cfq_slice_expired(cfqd, 0);
1216 cfqq = cfq_set_active_queue(cfqd, new_cfqq);
1221 static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
1225 while (cfqq->next_rq) {
1226 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
1230 BUG_ON(!list_empty(&cfqq->fifo));
1235 * Drain our current requests. Used for barriers and when switching
1236 * io schedulers on-the-fly.
1238 static int cfq_forced_dispatch(struct cfq_data *cfqd)
1240 struct cfq_queue *cfqq;
1243 while ((cfqq = cfq_rb_first(&cfqd->service_tree)) != NULL)
1244 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
1246 cfq_slice_expired(cfqd, 0);
1248 BUG_ON(cfqd->busy_queues);
1250 cfq_log(cfqd, "forced_dispatch=%d\n", dispatched);
1255 * Dispatch a request from cfqq, moving them to the request queue
1258 static void cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1262 BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
1265 * follow expired path, else get first next available
1267 rq = cfq_check_fifo(cfqq);
1272 * insert request into driver dispatch list
1274 cfq_dispatch_insert(cfqd->queue, rq);
1276 if (!cfqd->active_cic) {
1277 struct cfq_io_context *cic = RQ_CIC(rq);
1279 atomic_inc(&cic->ioc->refcount);
1280 cfqd->active_cic = cic;
1285 * Find the cfqq that we need to service and move a request from that to the
1288 static int cfq_dispatch_requests(struct request_queue *q, int force)
1290 struct cfq_data *cfqd = q->elevator->elevator_data;
1291 struct cfq_queue *cfqq;
1292 unsigned int max_dispatch;
1294 if (!cfqd->busy_queues)
1297 if (unlikely(force))
1298 return cfq_forced_dispatch(cfqd);
1300 cfqq = cfq_select_queue(cfqd);
1305 * If this is an async queue and we have sync IO in flight, let it wait
1307 if (cfqd->sync_flight && !cfq_cfqq_sync(cfqq))
1310 max_dispatch = cfqd->cfq_quantum;
1311 if (cfq_class_idle(cfqq))
1315 * Does this cfqq already have too much IO in flight?
1317 if (cfqq->dispatched >= max_dispatch) {
1319 * idle queue must always only have a single IO in flight
1321 if (cfq_class_idle(cfqq))
1325 * We have other queues, don't allow more IO from this one
1327 if (cfqd->busy_queues > 1)
1331 * we are the only queue, allow up to 4 times of 'quantum'
1333 if (cfqq->dispatched >= 4 * max_dispatch)
1338 * Dispatch a request from this cfqq
1340 cfq_dispatch_request(cfqd, cfqq);
1341 cfqq->slice_dispatch++;
1342 cfq_clear_cfqq_must_dispatch(cfqq);
1345 * expire an async queue immediately if it has used up its slice. idle
1346 * queue always expire after 1 dispatch round.
1348 if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
1349 cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
1350 cfq_class_idle(cfqq))) {
1351 cfqq->slice_end = jiffies + 1;
1352 cfq_slice_expired(cfqd, 0);
1355 cfq_log(cfqd, "dispatched a request");
1360 * task holds one reference to the queue, dropped when task exits. each rq
1361 * in-flight on this queue also holds a reference, dropped when rq is freed.
1363 * queue lock must be held here.
1365 static void cfq_put_queue(struct cfq_queue *cfqq)
1367 struct cfq_data *cfqd = cfqq->cfqd;
1369 BUG_ON(atomic_read(&cfqq->ref) <= 0);
1371 if (!atomic_dec_and_test(&cfqq->ref))
1374 cfq_log_cfqq(cfqd, cfqq, "put_queue");
1375 BUG_ON(rb_first(&cfqq->sort_list));
1376 BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
1377 BUG_ON(cfq_cfqq_on_rr(cfqq));
1379 if (unlikely(cfqd->active_queue == cfqq)) {
1380 __cfq_slice_expired(cfqd, cfqq, 0);
1381 cfq_schedule_dispatch(cfqd);
1384 kmem_cache_free(cfq_pool, cfqq);
1388 * Must always be called with the rcu_read_lock() held
1391 __call_for_each_cic(struct io_context *ioc,
1392 void (*func)(struct io_context *, struct cfq_io_context *))
1394 struct cfq_io_context *cic;
1395 struct hlist_node *n;
1397 hlist_for_each_entry_rcu(cic, n, &ioc->cic_list, cic_list)
1402 * Call func for each cic attached to this ioc.
1405 call_for_each_cic(struct io_context *ioc,
1406 void (*func)(struct io_context *, struct cfq_io_context *))
1409 __call_for_each_cic(ioc, func);
1413 static void cfq_cic_free_rcu(struct rcu_head *head)
1415 struct cfq_io_context *cic;
1417 cic = container_of(head, struct cfq_io_context, rcu_head);
1419 kmem_cache_free(cfq_ioc_pool, cic);
1420 elv_ioc_count_dec(ioc_count);
1424 * CFQ scheduler is exiting, grab exit lock and check
1425 * the pending io context count. If it hits zero,
1426 * complete ioc_gone and set it back to NULL
1428 spin_lock(&ioc_gone_lock);
1429 if (ioc_gone && !elv_ioc_count_read(ioc_count)) {
1433 spin_unlock(&ioc_gone_lock);
1437 static void cfq_cic_free(struct cfq_io_context *cic)
1439 call_rcu(&cic->rcu_head, cfq_cic_free_rcu);
1442 static void cic_free_func(struct io_context *ioc, struct cfq_io_context *cic)
1444 unsigned long flags;
1446 BUG_ON(!cic->dead_key);
1448 spin_lock_irqsave(&ioc->lock, flags);
1449 radix_tree_delete(&ioc->radix_root, cic->dead_key);
1450 hlist_del_rcu(&cic->cic_list);
1451 spin_unlock_irqrestore(&ioc->lock, flags);
1457 * Must be called with rcu_read_lock() held or preemption otherwise disabled.
1458 * Only two callers of this - ->dtor() which is called with the rcu_read_lock(),
1459 * and ->trim() which is called with the task lock held
1461 static void cfq_free_io_context(struct io_context *ioc)
1464 * ioc->refcount is zero here, or we are called from elv_unregister(),
1465 * so no more cic's are allowed to be linked into this ioc. So it
1466 * should be ok to iterate over the known list, we will see all cic's
1467 * since no new ones are added.
1469 __call_for_each_cic(ioc, cic_free_func);
1472 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1474 if (unlikely(cfqq == cfqd->active_queue)) {
1475 __cfq_slice_expired(cfqd, cfqq, 0);
1476 cfq_schedule_dispatch(cfqd);
1479 cfq_put_queue(cfqq);
1482 static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
1483 struct cfq_io_context *cic)
1485 struct io_context *ioc = cic->ioc;
1487 list_del_init(&cic->queue_list);
1490 * Make sure key == NULL is seen for dead queues
1493 cic->dead_key = (unsigned long) cic->key;
1496 if (ioc->ioc_data == cic)
1497 rcu_assign_pointer(ioc->ioc_data, NULL);
1499 if (cic->cfqq[BLK_RW_ASYNC]) {
1500 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]);
1501 cic->cfqq[BLK_RW_ASYNC] = NULL;
1504 if (cic->cfqq[BLK_RW_SYNC]) {
1505 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]);
1506 cic->cfqq[BLK_RW_SYNC] = NULL;
1510 static void cfq_exit_single_io_context(struct io_context *ioc,
1511 struct cfq_io_context *cic)
1513 struct cfq_data *cfqd = cic->key;
1516 struct request_queue *q = cfqd->queue;
1517 unsigned long flags;
1519 spin_lock_irqsave(q->queue_lock, flags);
1522 * Ensure we get a fresh copy of the ->key to prevent
1523 * race between exiting task and queue
1525 smp_read_barrier_depends();
1527 __cfq_exit_single_io_context(cfqd, cic);
1529 spin_unlock_irqrestore(q->queue_lock, flags);
1534 * The process that ioc belongs to has exited, we need to clean up
1535 * and put the internal structures we have that belongs to that process.
1537 static void cfq_exit_io_context(struct io_context *ioc)
1539 call_for_each_cic(ioc, cfq_exit_single_io_context);
1542 static struct cfq_io_context *
1543 cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
1545 struct cfq_io_context *cic;
1547 cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO,
1550 cic->last_end_request = jiffies;
1551 INIT_LIST_HEAD(&cic->queue_list);
1552 INIT_HLIST_NODE(&cic->cic_list);
1553 cic->dtor = cfq_free_io_context;
1554 cic->exit = cfq_exit_io_context;
1555 elv_ioc_count_inc(ioc_count);
1561 static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc)
1563 struct task_struct *tsk = current;
1566 if (!cfq_cfqq_prio_changed(cfqq))
1569 ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio);
1570 switch (ioprio_class) {
1572 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
1573 case IOPRIO_CLASS_NONE:
1575 * no prio set, inherit CPU scheduling settings
1577 cfqq->ioprio = task_nice_ioprio(tsk);
1578 cfqq->ioprio_class = task_nice_ioclass(tsk);
1580 case IOPRIO_CLASS_RT:
1581 cfqq->ioprio = task_ioprio(ioc);
1582 cfqq->ioprio_class = IOPRIO_CLASS_RT;
1584 case IOPRIO_CLASS_BE:
1585 cfqq->ioprio = task_ioprio(ioc);
1586 cfqq->ioprio_class = IOPRIO_CLASS_BE;
1588 case IOPRIO_CLASS_IDLE:
1589 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
1591 cfq_clear_cfqq_idle_window(cfqq);
1596 * keep track of original prio settings in case we have to temporarily
1597 * elevate the priority of this queue
1599 cfqq->org_ioprio = cfqq->ioprio;
1600 cfqq->org_ioprio_class = cfqq->ioprio_class;
1601 cfq_clear_cfqq_prio_changed(cfqq);
1604 static void changed_ioprio(struct io_context *ioc, struct cfq_io_context *cic)
1606 struct cfq_data *cfqd = cic->key;
1607 struct cfq_queue *cfqq;
1608 unsigned long flags;
1610 if (unlikely(!cfqd))
1613 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1615 cfqq = cic->cfqq[BLK_RW_ASYNC];
1617 struct cfq_queue *new_cfqq;
1618 new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic->ioc,
1621 cic->cfqq[BLK_RW_ASYNC] = new_cfqq;
1622 cfq_put_queue(cfqq);
1626 cfqq = cic->cfqq[BLK_RW_SYNC];
1628 cfq_mark_cfqq_prio_changed(cfqq);
1630 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1633 static void cfq_ioc_set_ioprio(struct io_context *ioc)
1635 call_for_each_cic(ioc, changed_ioprio);
1636 ioc->ioprio_changed = 0;
1639 static struct cfq_queue *
1640 cfq_find_alloc_queue(struct cfq_data *cfqd, int is_sync,
1641 struct io_context *ioc, gfp_t gfp_mask)
1643 struct cfq_queue *cfqq, *new_cfqq = NULL;
1644 struct cfq_io_context *cic;
1647 cic = cfq_cic_lookup(cfqd, ioc);
1648 /* cic always exists here */
1649 cfqq = cic_to_cfqq(cic, is_sync);
1655 } else if (gfp_mask & __GFP_WAIT) {
1657 * Inform the allocator of the fact that we will
1658 * just repeat this allocation if it fails, to allow
1659 * the allocator to do whatever it needs to attempt to
1662 spin_unlock_irq(cfqd->queue->queue_lock);
1663 new_cfqq = kmem_cache_alloc_node(cfq_pool,
1664 gfp_mask | __GFP_NOFAIL | __GFP_ZERO,
1666 spin_lock_irq(cfqd->queue->queue_lock);
1669 cfqq = kmem_cache_alloc_node(cfq_pool,
1670 gfp_mask | __GFP_ZERO,
1676 RB_CLEAR_NODE(&cfqq->rb_node);
1677 RB_CLEAR_NODE(&cfqq->p_node);
1678 INIT_LIST_HEAD(&cfqq->fifo);
1680 atomic_set(&cfqq->ref, 0);
1683 cfq_mark_cfqq_prio_changed(cfqq);
1685 cfq_init_prio_data(cfqq, ioc);
1688 if (!cfq_class_idle(cfqq))
1689 cfq_mark_cfqq_idle_window(cfqq);
1690 cfq_mark_cfqq_sync(cfqq);
1692 cfqq->pid = current->pid;
1693 cfq_log_cfqq(cfqd, cfqq, "alloced");
1697 kmem_cache_free(cfq_pool, new_cfqq);
1700 WARN_ON((gfp_mask & __GFP_WAIT) && !cfqq);
1704 static struct cfq_queue **
1705 cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
1707 switch (ioprio_class) {
1708 case IOPRIO_CLASS_RT:
1709 return &cfqd->async_cfqq[0][ioprio];
1710 case IOPRIO_CLASS_BE:
1711 return &cfqd->async_cfqq[1][ioprio];
1712 case IOPRIO_CLASS_IDLE:
1713 return &cfqd->async_idle_cfqq;
1719 static struct cfq_queue *
1720 cfq_get_queue(struct cfq_data *cfqd, int is_sync, struct io_context *ioc,
1723 const int ioprio = task_ioprio(ioc);
1724 const int ioprio_class = task_ioprio_class(ioc);
1725 struct cfq_queue **async_cfqq = NULL;
1726 struct cfq_queue *cfqq = NULL;
1729 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
1734 cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask);
1740 * pin the queue now that it's allocated, scheduler exit will prune it
1742 if (!is_sync && !(*async_cfqq)) {
1743 atomic_inc(&cfqq->ref);
1747 atomic_inc(&cfqq->ref);
1752 * We drop cfq io contexts lazily, so we may find a dead one.
1755 cfq_drop_dead_cic(struct cfq_data *cfqd, struct io_context *ioc,
1756 struct cfq_io_context *cic)
1758 unsigned long flags;
1760 WARN_ON(!list_empty(&cic->queue_list));
1762 spin_lock_irqsave(&ioc->lock, flags);
1764 BUG_ON(ioc->ioc_data == cic);
1766 radix_tree_delete(&ioc->radix_root, (unsigned long) cfqd);
1767 hlist_del_rcu(&cic->cic_list);
1768 spin_unlock_irqrestore(&ioc->lock, flags);
1773 static struct cfq_io_context *
1774 cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc)
1776 struct cfq_io_context *cic;
1777 unsigned long flags;
1786 * we maintain a last-hit cache, to avoid browsing over the tree
1788 cic = rcu_dereference(ioc->ioc_data);
1789 if (cic && cic->key == cfqd) {
1795 cic = radix_tree_lookup(&ioc->radix_root, (unsigned long) cfqd);
1799 /* ->key must be copied to avoid race with cfq_exit_queue() */
1802 cfq_drop_dead_cic(cfqd, ioc, cic);
1807 spin_lock_irqsave(&ioc->lock, flags);
1808 rcu_assign_pointer(ioc->ioc_data, cic);
1809 spin_unlock_irqrestore(&ioc->lock, flags);
1817 * Add cic into ioc, using cfqd as the search key. This enables us to lookup
1818 * the process specific cfq io context when entered from the block layer.
1819 * Also adds the cic to a per-cfqd list, used when this queue is removed.
1821 static int cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
1822 struct cfq_io_context *cic, gfp_t gfp_mask)
1824 unsigned long flags;
1827 ret = radix_tree_preload(gfp_mask);
1832 spin_lock_irqsave(&ioc->lock, flags);
1833 ret = radix_tree_insert(&ioc->radix_root,
1834 (unsigned long) cfqd, cic);
1836 hlist_add_head_rcu(&cic->cic_list, &ioc->cic_list);
1837 spin_unlock_irqrestore(&ioc->lock, flags);
1839 radix_tree_preload_end();
1842 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1843 list_add(&cic->queue_list, &cfqd->cic_list);
1844 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1849 printk(KERN_ERR "cfq: cic link failed!\n");
1855 * Setup general io context and cfq io context. There can be several cfq
1856 * io contexts per general io context, if this process is doing io to more
1857 * than one device managed by cfq.
1859 static struct cfq_io_context *
1860 cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
1862 struct io_context *ioc = NULL;
1863 struct cfq_io_context *cic;
1865 might_sleep_if(gfp_mask & __GFP_WAIT);
1867 ioc = get_io_context(gfp_mask, cfqd->queue->node);
1871 cic = cfq_cic_lookup(cfqd, ioc);
1875 cic = cfq_alloc_io_context(cfqd, gfp_mask);
1879 if (cfq_cic_link(cfqd, ioc, cic, gfp_mask))
1883 smp_read_barrier_depends();
1884 if (unlikely(ioc->ioprio_changed))
1885 cfq_ioc_set_ioprio(ioc);
1891 put_io_context(ioc);
1896 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
1898 unsigned long elapsed = jiffies - cic->last_end_request;
1899 unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
1901 cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
1902 cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
1903 cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
1907 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_io_context *cic,
1913 if (!cic->last_request_pos)
1915 else if (cic->last_request_pos < rq->sector)
1916 sdist = rq->sector - cic->last_request_pos;
1918 sdist = cic->last_request_pos - rq->sector;
1921 * Don't allow the seek distance to get too large from the
1922 * odd fragment, pagein, etc
1924 if (cic->seek_samples <= 60) /* second&third seek */
1925 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*1024);
1927 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*64);
1929 cic->seek_samples = (7*cic->seek_samples + 256) / 8;
1930 cic->seek_total = (7*cic->seek_total + (u64)256*sdist) / 8;
1931 total = cic->seek_total + (cic->seek_samples/2);
1932 do_div(total, cic->seek_samples);
1933 cic->seek_mean = (sector_t)total;
1937 * Disable idle window if the process thinks too long or seeks so much that
1941 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1942 struct cfq_io_context *cic)
1944 int old_idle, enable_idle;
1947 * Don't idle for async or idle io prio class
1949 if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
1952 enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
1954 if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle ||
1955 (cfqd->hw_tag && CIC_SEEKY(cic)))
1957 else if (sample_valid(cic->ttime_samples)) {
1958 if (cic->ttime_mean > cfqd->cfq_slice_idle)
1964 if (old_idle != enable_idle) {
1965 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
1967 cfq_mark_cfqq_idle_window(cfqq);
1969 cfq_clear_cfqq_idle_window(cfqq);
1974 * Check if new_cfqq should preempt the currently active queue. Return 0 for
1975 * no or if we aren't sure, a 1 will cause a preempt.
1978 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
1981 struct cfq_queue *cfqq;
1983 cfqq = cfqd->active_queue;
1987 if (cfq_slice_used(cfqq))
1990 if (cfq_class_idle(new_cfqq))
1993 if (cfq_class_idle(cfqq))
1997 * if the new request is sync, but the currently running queue is
1998 * not, let the sync request have priority.
2000 if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
2004 * So both queues are sync. Let the new request get disk time if
2005 * it's a metadata request and the current queue is doing regular IO.
2007 if (rq_is_meta(rq) && !cfqq->meta_pending)
2011 * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
2013 if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
2016 if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
2020 * if this request is as-good as one we would expect from the
2021 * current cfqq, let it preempt
2023 if (cfq_rq_close(cfqd, rq))
2030 * cfqq preempts the active queue. if we allowed preempt with no slice left,
2031 * let it have half of its nominal slice.
2033 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2035 cfq_log_cfqq(cfqd, cfqq, "preempt");
2036 cfq_slice_expired(cfqd, 1);
2039 * Put the new queue at the front of the of the current list,
2040 * so we know that it will be selected next.
2042 BUG_ON(!cfq_cfqq_on_rr(cfqq));
2044 cfq_service_tree_add(cfqd, cfqq, 1);
2046 cfqq->slice_end = 0;
2047 cfq_mark_cfqq_slice_new(cfqq);
2051 * Called when a new fs request (rq) is added (to cfqq). Check if there's
2052 * something we should do about it
2055 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2058 struct cfq_io_context *cic = RQ_CIC(rq);
2062 cfqq->meta_pending++;
2064 cfq_update_io_thinktime(cfqd, cic);
2065 cfq_update_io_seektime(cfqd, cic, rq);
2066 cfq_update_idle_window(cfqd, cfqq, cic);
2068 cic->last_request_pos = rq->sector + rq->nr_sectors;
2070 if (cfqq == cfqd->active_queue) {
2072 * Remember that we saw a request from this process, but
2073 * don't start queuing just yet. Otherwise we risk seeing lots
2074 * of tiny requests, because we disrupt the normal plugging
2075 * and merging. If the request is already larger than a single
2076 * page, let it rip immediately. For that case we assume that
2077 * merging is already done. Ditto for a busy system that
2078 * has other work pending, don't risk delaying until the
2079 * idle timer unplug to continue working.
2081 if (cfq_cfqq_wait_request(cfqq)) {
2082 if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
2083 cfqd->busy_queues > 1) {
2084 del_timer(&cfqd->idle_slice_timer);
2085 blk_start_queueing(cfqd->queue);
2087 cfq_mark_cfqq_must_dispatch(cfqq);
2089 } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
2091 * not the active queue - expire current slice if it is
2092 * idle and has expired it's mean thinktime or this new queue
2093 * has some old slice time left and is of higher priority or
2094 * this new queue is RT and the current one is BE
2096 cfq_preempt_queue(cfqd, cfqq);
2097 blk_start_queueing(cfqd->queue);
2101 static void cfq_insert_request(struct request_queue *q, struct request *rq)
2103 struct cfq_data *cfqd = q->elevator->elevator_data;
2104 struct cfq_queue *cfqq = RQ_CFQQ(rq);
2106 cfq_log_cfqq(cfqd, cfqq, "insert_request");
2107 cfq_init_prio_data(cfqq, RQ_CIC(rq)->ioc);
2111 list_add_tail(&rq->queuelist, &cfqq->fifo);
2113 cfq_rq_enqueued(cfqd, cfqq, rq);
2117 * Update hw_tag based on peak queue depth over 50 samples under
2120 static void cfq_update_hw_tag(struct cfq_data *cfqd)
2122 if (cfqd->rq_in_driver > cfqd->rq_in_driver_peak)
2123 cfqd->rq_in_driver_peak = cfqd->rq_in_driver;
2125 if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
2126 cfqd->rq_in_driver <= CFQ_HW_QUEUE_MIN)
2129 if (cfqd->hw_tag_samples++ < 50)
2132 if (cfqd->rq_in_driver_peak >= CFQ_HW_QUEUE_MIN)
2137 cfqd->hw_tag_samples = 0;
2138 cfqd->rq_in_driver_peak = 0;
2141 static void cfq_completed_request(struct request_queue *q, struct request *rq)
2143 struct cfq_queue *cfqq = RQ_CFQQ(rq);
2144 struct cfq_data *cfqd = cfqq->cfqd;
2145 const int sync = rq_is_sync(rq);
2149 cfq_log_cfqq(cfqd, cfqq, "complete");
2151 cfq_update_hw_tag(cfqd);
2153 WARN_ON(!cfqd->rq_in_driver);
2154 WARN_ON(!cfqq->dispatched);
2155 cfqd->rq_in_driver--;
2158 if (cfq_cfqq_sync(cfqq))
2159 cfqd->sync_flight--;
2161 if (!cfq_class_idle(cfqq))
2162 cfqd->last_end_request = now;
2165 RQ_CIC(rq)->last_end_request = now;
2168 * If this is the active queue, check if it needs to be expired,
2169 * or if we want to idle in case it has no pending requests.
2171 if (cfqd->active_queue == cfqq) {
2172 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
2174 if (cfq_cfqq_slice_new(cfqq)) {
2175 cfq_set_prio_slice(cfqd, cfqq);
2176 cfq_clear_cfqq_slice_new(cfqq);
2179 * If there are no requests waiting in this queue, and
2180 * there are other queues ready to issue requests, AND
2181 * those other queues are issuing requests within our
2182 * mean seek distance, give them a chance to run instead
2185 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
2186 cfq_slice_expired(cfqd, 1);
2187 else if (cfqq_empty && !cfq_close_cooperator(cfqd, cfqq, 1) &&
2188 sync && !rq_noidle(rq))
2189 cfq_arm_slice_timer(cfqd);
2192 if (!cfqd->rq_in_driver)
2193 cfq_schedule_dispatch(cfqd);
2197 * we temporarily boost lower priority queues if they are holding fs exclusive
2198 * resources. they are boosted to normal prio (CLASS_BE/4)
2200 static void cfq_prio_boost(struct cfq_queue *cfqq)
2202 if (has_fs_excl()) {
2204 * boost idle prio on transactions that would lock out other
2205 * users of the filesystem
2207 if (cfq_class_idle(cfqq))
2208 cfqq->ioprio_class = IOPRIO_CLASS_BE;
2209 if (cfqq->ioprio > IOPRIO_NORM)
2210 cfqq->ioprio = IOPRIO_NORM;
2213 * check if we need to unboost the queue
2215 if (cfqq->ioprio_class != cfqq->org_ioprio_class)
2216 cfqq->ioprio_class = cfqq->org_ioprio_class;
2217 if (cfqq->ioprio != cfqq->org_ioprio)
2218 cfqq->ioprio = cfqq->org_ioprio;
2222 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
2224 if ((cfq_cfqq_wait_request(cfqq) || cfq_cfqq_must_alloc(cfqq)) &&
2225 !cfq_cfqq_must_alloc_slice(cfqq)) {
2226 cfq_mark_cfqq_must_alloc_slice(cfqq);
2227 return ELV_MQUEUE_MUST;
2230 return ELV_MQUEUE_MAY;
2233 static int cfq_may_queue(struct request_queue *q, int rw)
2235 struct cfq_data *cfqd = q->elevator->elevator_data;
2236 struct task_struct *tsk = current;
2237 struct cfq_io_context *cic;
2238 struct cfq_queue *cfqq;
2241 * don't force setup of a queue from here, as a call to may_queue
2242 * does not necessarily imply that a request actually will be queued.
2243 * so just lookup a possibly existing queue, or return 'may queue'
2246 cic = cfq_cic_lookup(cfqd, tsk->io_context);
2248 return ELV_MQUEUE_MAY;
2250 cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
2252 cfq_init_prio_data(cfqq, cic->ioc);
2253 cfq_prio_boost(cfqq);
2255 return __cfq_may_queue(cfqq);
2258 return ELV_MQUEUE_MAY;
2262 * queue lock held here
2264 static void cfq_put_request(struct request *rq)
2266 struct cfq_queue *cfqq = RQ_CFQQ(rq);
2269 const int rw = rq_data_dir(rq);
2271 BUG_ON(!cfqq->allocated[rw]);
2272 cfqq->allocated[rw]--;
2274 put_io_context(RQ_CIC(rq)->ioc);
2276 rq->elevator_private = NULL;
2277 rq->elevator_private2 = NULL;
2279 cfq_put_queue(cfqq);
2284 * Allocate cfq data structures associated with this request.
2287 cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
2289 struct cfq_data *cfqd = q->elevator->elevator_data;
2290 struct cfq_io_context *cic;
2291 const int rw = rq_data_dir(rq);
2292 const int is_sync = rq_is_sync(rq);
2293 struct cfq_queue *cfqq;
2294 unsigned long flags;
2296 might_sleep_if(gfp_mask & __GFP_WAIT);
2298 cic = cfq_get_io_context(cfqd, gfp_mask);
2300 spin_lock_irqsave(q->queue_lock, flags);
2305 cfqq = cic_to_cfqq(cic, is_sync);
2307 cfqq = cfq_get_queue(cfqd, is_sync, cic->ioc, gfp_mask);
2312 cic_set_cfqq(cic, cfqq, is_sync);
2315 cfqq->allocated[rw]++;
2316 cfq_clear_cfqq_must_alloc(cfqq);
2317 atomic_inc(&cfqq->ref);
2319 spin_unlock_irqrestore(q->queue_lock, flags);
2321 rq->elevator_private = cic;
2322 rq->elevator_private2 = cfqq;
2327 put_io_context(cic->ioc);
2329 cfq_schedule_dispatch(cfqd);
2330 spin_unlock_irqrestore(q->queue_lock, flags);
2331 cfq_log(cfqd, "set_request fail");
2335 static void cfq_kick_queue(struct work_struct *work)
2337 struct cfq_data *cfqd =
2338 container_of(work, struct cfq_data, unplug_work);
2339 struct request_queue *q = cfqd->queue;
2341 spin_lock_irq(q->queue_lock);
2342 blk_start_queueing(q);
2343 spin_unlock_irq(q->queue_lock);
2347 * Timer running if the active_queue is currently idling inside its time slice
2349 static void cfq_idle_slice_timer(unsigned long data)
2351 struct cfq_data *cfqd = (struct cfq_data *) data;
2352 struct cfq_queue *cfqq;
2353 unsigned long flags;
2356 cfq_log(cfqd, "idle timer fired");
2358 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2360 cfqq = cfqd->active_queue;
2365 * We saw a request before the queue expired, let it through
2367 if (cfq_cfqq_must_dispatch(cfqq))
2373 if (cfq_slice_used(cfqq))
2377 * only expire and reinvoke request handler, if there are
2378 * other queues with pending requests
2380 if (!cfqd->busy_queues)
2384 * not expired and it has a request pending, let it dispatch
2386 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
2390 cfq_slice_expired(cfqd, timed_out);
2392 cfq_schedule_dispatch(cfqd);
2394 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2397 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
2399 del_timer_sync(&cfqd->idle_slice_timer);
2400 cancel_work_sync(&cfqd->unplug_work);
2403 static void cfq_put_async_queues(struct cfq_data *cfqd)
2407 for (i = 0; i < IOPRIO_BE_NR; i++) {
2408 if (cfqd->async_cfqq[0][i])
2409 cfq_put_queue(cfqd->async_cfqq[0][i]);
2410 if (cfqd->async_cfqq[1][i])
2411 cfq_put_queue(cfqd->async_cfqq[1][i]);
2414 if (cfqd->async_idle_cfqq)
2415 cfq_put_queue(cfqd->async_idle_cfqq);
2418 static void cfq_exit_queue(struct elevator_queue *e)
2420 struct cfq_data *cfqd = e->elevator_data;
2421 struct request_queue *q = cfqd->queue;
2423 cfq_shutdown_timer_wq(cfqd);
2425 spin_lock_irq(q->queue_lock);
2427 if (cfqd->active_queue)
2428 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
2430 while (!list_empty(&cfqd->cic_list)) {
2431 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
2432 struct cfq_io_context,
2435 __cfq_exit_single_io_context(cfqd, cic);
2438 cfq_put_async_queues(cfqd);
2440 spin_unlock_irq(q->queue_lock);
2442 cfq_shutdown_timer_wq(cfqd);
2447 static void *cfq_init_queue(struct request_queue *q)
2449 struct cfq_data *cfqd;
2452 cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
2456 cfqd->service_tree = CFQ_RB_ROOT;
2459 * Not strictly needed (since RB_ROOT just clears the node and we
2460 * zeroed cfqd on alloc), but better be safe in case someone decides
2461 * to add magic to the rb code
2463 for (i = 0; i < CFQ_PRIO_LISTS; i++)
2464 cfqd->prio_trees[i] = RB_ROOT;
2466 INIT_LIST_HEAD(&cfqd->cic_list);
2470 init_timer(&cfqd->idle_slice_timer);
2471 cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
2472 cfqd->idle_slice_timer.data = (unsigned long) cfqd;
2474 INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
2476 cfqd->last_end_request = jiffies;
2477 cfqd->cfq_quantum = cfq_quantum;
2478 cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
2479 cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
2480 cfqd->cfq_back_max = cfq_back_max;
2481 cfqd->cfq_back_penalty = cfq_back_penalty;
2482 cfqd->cfq_slice[0] = cfq_slice_async;
2483 cfqd->cfq_slice[1] = cfq_slice_sync;
2484 cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
2485 cfqd->cfq_slice_idle = cfq_slice_idle;
2491 static void cfq_slab_kill(void)
2494 * Caller already ensured that pending RCU callbacks are completed,
2495 * so we should have no busy allocations at this point.
2498 kmem_cache_destroy(cfq_pool);
2500 kmem_cache_destroy(cfq_ioc_pool);
2503 static int __init cfq_slab_setup(void)
2505 cfq_pool = KMEM_CACHE(cfq_queue, 0);
2509 cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0);
2520 * sysfs parts below -->
2523 cfq_var_show(unsigned int var, char *page)
2525 return sprintf(page, "%d\n", var);
2529 cfq_var_store(unsigned int *var, const char *page, size_t count)
2531 char *p = (char *) page;
2533 *var = simple_strtoul(p, &p, 10);
2537 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
2538 static ssize_t __FUNC(struct elevator_queue *e, char *page) \
2540 struct cfq_data *cfqd = e->elevator_data; \
2541 unsigned int __data = __VAR; \
2543 __data = jiffies_to_msecs(__data); \
2544 return cfq_var_show(__data, (page)); \
2546 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
2547 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
2548 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
2549 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
2550 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
2551 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
2552 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
2553 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
2554 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
2555 #undef SHOW_FUNCTION
2557 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
2558 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
2560 struct cfq_data *cfqd = e->elevator_data; \
2561 unsigned int __data; \
2562 int ret = cfq_var_store(&__data, (page), count); \
2563 if (__data < (MIN)) \
2565 else if (__data > (MAX)) \
2568 *(__PTR) = msecs_to_jiffies(__data); \
2570 *(__PTR) = __data; \
2573 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
2574 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
2576 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
2578 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
2579 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
2581 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
2582 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
2583 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
2584 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
2586 #undef STORE_FUNCTION
2588 #define CFQ_ATTR(name) \
2589 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
2591 static struct elv_fs_entry cfq_attrs[] = {
2593 CFQ_ATTR(fifo_expire_sync),
2594 CFQ_ATTR(fifo_expire_async),
2595 CFQ_ATTR(back_seek_max),
2596 CFQ_ATTR(back_seek_penalty),
2597 CFQ_ATTR(slice_sync),
2598 CFQ_ATTR(slice_async),
2599 CFQ_ATTR(slice_async_rq),
2600 CFQ_ATTR(slice_idle),
2604 static struct elevator_type iosched_cfq = {
2606 .elevator_merge_fn = cfq_merge,
2607 .elevator_merged_fn = cfq_merged_request,
2608 .elevator_merge_req_fn = cfq_merged_requests,
2609 .elevator_allow_merge_fn = cfq_allow_merge,
2610 .elevator_dispatch_fn = cfq_dispatch_requests,
2611 .elevator_add_req_fn = cfq_insert_request,
2612 .elevator_activate_req_fn = cfq_activate_request,
2613 .elevator_deactivate_req_fn = cfq_deactivate_request,
2614 .elevator_queue_empty_fn = cfq_queue_empty,
2615 .elevator_completed_req_fn = cfq_completed_request,
2616 .elevator_former_req_fn = elv_rb_former_request,
2617 .elevator_latter_req_fn = elv_rb_latter_request,
2618 .elevator_set_req_fn = cfq_set_request,
2619 .elevator_put_req_fn = cfq_put_request,
2620 .elevator_may_queue_fn = cfq_may_queue,
2621 .elevator_init_fn = cfq_init_queue,
2622 .elevator_exit_fn = cfq_exit_queue,
2623 .trim = cfq_free_io_context,
2625 .elevator_attrs = cfq_attrs,
2626 .elevator_name = "cfq",
2627 .elevator_owner = THIS_MODULE,
2630 static int __init cfq_init(void)
2633 * could be 0 on HZ < 1000 setups
2635 if (!cfq_slice_async)
2636 cfq_slice_async = 1;
2637 if (!cfq_slice_idle)
2640 if (cfq_slab_setup())
2643 elv_register(&iosched_cfq);
2648 static void __exit cfq_exit(void)
2650 DECLARE_COMPLETION_ONSTACK(all_gone);
2651 elv_unregister(&iosched_cfq);
2652 ioc_gone = &all_gone;
2653 /* ioc_gone's update must be visible before reading ioc_count */
2657 * this also protects us from entering cfq_slab_kill() with
2658 * pending RCU callbacks
2660 if (elv_ioc_count_read(ioc_count))
2661 wait_for_completion(&all_gone);
2665 module_init(cfq_init);
2666 module_exit(cfq_exit);
2668 MODULE_AUTHOR("Jens Axboe");
2669 MODULE_LICENSE("GPL");
2670 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");