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, cfq_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 process-grouping structure
79 /* various state flags, see below */
82 struct cfq_data *cfqd;
83 /* service_tree member */
84 struct rb_node rb_node;
85 /* service_tree key */
87 /* prio tree member */
88 struct rb_node p_node;
89 /* prio tree root we belong to, if any */
90 struct rb_root *p_root;
91 /* sorted list of pending requests */
92 struct rb_root sort_list;
93 /* if fifo isn't expired, next request to serve */
94 struct request *next_rq;
95 /* requests queued in sort_list */
97 /* currently allocated requests */
99 /* fifo list of requests in sort_list */
100 struct list_head fifo;
102 unsigned long slice_end;
104 unsigned int slice_dispatch;
106 /* pending metadata requests */
108 /* number of requests that are on the dispatch list or inside driver */
111 /* io prio of this group */
112 unsigned short ioprio, org_ioprio;
113 unsigned short ioprio_class, org_ioprio_class;
119 * Per block device queue structure
122 struct request_queue *queue;
125 * rr list of queues with requests and the count of them
127 struct cfq_rb_root service_tree;
130 * Each priority tree is sorted by next_request position. These
131 * trees are used when determining if two or more queues are
132 * interleaving requests (see cfq_close_cooperator).
134 struct rb_root prio_trees[CFQ_PRIO_LISTS];
136 unsigned int busy_queues;
142 * queue-depth detection
147 int rq_in_driver_peak;
150 * idle window management
152 struct timer_list idle_slice_timer;
153 struct work_struct unplug_work;
155 struct cfq_queue *active_queue;
156 struct cfq_io_context *active_cic;
159 * async queue for each priority case
161 struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
162 struct cfq_queue *async_idle_cfqq;
164 sector_t last_position;
167 * tunables, see top of file
169 unsigned int cfq_quantum;
170 unsigned int cfq_fifo_expire[2];
171 unsigned int cfq_back_penalty;
172 unsigned int cfq_back_max;
173 unsigned int cfq_slice[2];
174 unsigned int cfq_slice_async_rq;
175 unsigned int cfq_slice_idle;
176 unsigned int cfq_desktop;
178 struct list_head cic_list;
181 * Fallback dummy cfqq for extreme OOM conditions
183 struct cfq_queue oom_cfqq;
185 unsigned long last_end_sync_rq;
188 enum cfqq_state_flags {
189 CFQ_CFQQ_FLAG_on_rr = 0, /* on round-robin busy list */
190 CFQ_CFQQ_FLAG_wait_request, /* waiting for a request */
191 CFQ_CFQQ_FLAG_must_dispatch, /* must be allowed a dispatch */
192 CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
193 CFQ_CFQQ_FLAG_fifo_expire, /* FIFO checked in this slice */
194 CFQ_CFQQ_FLAG_idle_window, /* slice idling enabled */
195 CFQ_CFQQ_FLAG_prio_changed, /* task priority has changed */
196 CFQ_CFQQ_FLAG_slice_new, /* no requests dispatched in slice */
197 CFQ_CFQQ_FLAG_sync, /* synchronous queue */
198 CFQ_CFQQ_FLAG_coop, /* has done a coop jump of the queue */
201 #define CFQ_CFQQ_FNS(name) \
202 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \
204 (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name); \
206 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \
208 (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \
210 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \
212 return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \
216 CFQ_CFQQ_FNS(wait_request);
217 CFQ_CFQQ_FNS(must_dispatch);
218 CFQ_CFQQ_FNS(must_alloc_slice);
219 CFQ_CFQQ_FNS(fifo_expire);
220 CFQ_CFQQ_FNS(idle_window);
221 CFQ_CFQQ_FNS(prio_changed);
222 CFQ_CFQQ_FNS(slice_new);
227 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
228 blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args)
229 #define cfq_log(cfqd, fmt, args...) \
230 blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
232 static void cfq_dispatch_insert(struct request_queue *, struct request *);
233 static struct cfq_queue *cfq_get_queue(struct cfq_data *, int,
234 struct io_context *, gfp_t);
235 static struct cfq_io_context *cfq_cic_lookup(struct cfq_data *,
236 struct io_context *);
238 static inline int rq_in_driver(struct cfq_data *cfqd)
240 return cfqd->rq_in_driver[0] + cfqd->rq_in_driver[1];
243 static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_context *cic,
246 return cic->cfqq[!!is_sync];
249 static inline void cic_set_cfqq(struct cfq_io_context *cic,
250 struct cfq_queue *cfqq, int is_sync)
252 cic->cfqq[!!is_sync] = cfqq;
256 * We regard a request as SYNC, if it's either a read or has the SYNC bit
257 * set (in which case it could also be direct WRITE).
259 static inline int cfq_bio_sync(struct bio *bio)
261 if (bio_data_dir(bio) == READ || bio_rw_flagged(bio, BIO_RW_SYNCIO))
268 * scheduler run of queue, if there are requests pending and no one in the
269 * driver that will restart queueing
271 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
273 if (cfqd->busy_queues) {
274 cfq_log(cfqd, "schedule dispatch");
275 kblockd_schedule_work(cfqd->queue, &cfqd->unplug_work);
279 static int cfq_queue_empty(struct request_queue *q)
281 struct cfq_data *cfqd = q->elevator->elevator_data;
283 return !cfqd->busy_queues;
287 * Scale schedule slice based on io priority. Use the sync time slice only
288 * if a queue is marked sync and has sync io queued. A sync queue with async
289 * io only, should not get full sync slice length.
291 static inline int cfq_prio_slice(struct cfq_data *cfqd, int sync,
294 const int base_slice = cfqd->cfq_slice[sync];
296 WARN_ON(prio >= IOPRIO_BE_NR);
298 return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
302 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
304 return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
308 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
310 cfqq->slice_end = cfq_prio_to_slice(cfqd, cfqq) + jiffies;
311 cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
315 * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
316 * isn't valid until the first request from the dispatch is activated
317 * and the slice time set.
319 static inline int cfq_slice_used(struct cfq_queue *cfqq)
321 if (cfq_cfqq_slice_new(cfqq))
323 if (time_before(jiffies, cfqq->slice_end))
330 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
331 * We choose the request that is closest to the head right now. Distance
332 * behind the head is penalized and only allowed to a certain extent.
334 static struct request *
335 cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2)
337 sector_t last, s1, s2, d1 = 0, d2 = 0;
338 unsigned long back_max;
339 #define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */
340 #define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */
341 unsigned wrap = 0; /* bit mask: requests behind the disk head? */
343 if (rq1 == NULL || rq1 == rq2)
348 if (rq_is_sync(rq1) && !rq_is_sync(rq2))
350 else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
352 if (rq_is_meta(rq1) && !rq_is_meta(rq2))
354 else if (rq_is_meta(rq2) && !rq_is_meta(rq1))
357 s1 = blk_rq_pos(rq1);
358 s2 = blk_rq_pos(rq2);
360 last = cfqd->last_position;
363 * by definition, 1KiB is 2 sectors
365 back_max = cfqd->cfq_back_max * 2;
368 * Strict one way elevator _except_ in the case where we allow
369 * short backward seeks which are biased as twice the cost of a
370 * similar forward seek.
374 else if (s1 + back_max >= last)
375 d1 = (last - s1) * cfqd->cfq_back_penalty;
377 wrap |= CFQ_RQ1_WRAP;
381 else if (s2 + back_max >= last)
382 d2 = (last - s2) * cfqd->cfq_back_penalty;
384 wrap |= CFQ_RQ2_WRAP;
386 /* Found required data */
389 * By doing switch() on the bit mask "wrap" we avoid having to
390 * check two variables for all permutations: --> faster!
393 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
409 case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
412 * Since both rqs are wrapped,
413 * start with the one that's further behind head
414 * (--> only *one* back seek required),
415 * since back seek takes more time than forward.
425 * The below is leftmost cache rbtree addon
427 static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
430 root->left = rb_first(&root->rb);
433 return rb_entry(root->left, struct cfq_queue, rb_node);
438 static void rb_erase_init(struct rb_node *n, struct rb_root *root)
444 static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
448 rb_erase_init(n, &root->rb);
452 * would be nice to take fifo expire time into account as well
454 static struct request *
455 cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
456 struct request *last)
458 struct rb_node *rbnext = rb_next(&last->rb_node);
459 struct rb_node *rbprev = rb_prev(&last->rb_node);
460 struct request *next = NULL, *prev = NULL;
462 BUG_ON(RB_EMPTY_NODE(&last->rb_node));
465 prev = rb_entry_rq(rbprev);
468 next = rb_entry_rq(rbnext);
470 rbnext = rb_first(&cfqq->sort_list);
471 if (rbnext && rbnext != &last->rb_node)
472 next = rb_entry_rq(rbnext);
475 return cfq_choose_req(cfqd, next, prev);
478 static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
479 struct cfq_queue *cfqq)
482 * just an approximation, should be ok.
484 return (cfqd->busy_queues - 1) * (cfq_prio_slice(cfqd, 1, 0) -
485 cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
489 * The cfqd->service_tree holds all pending cfq_queue's that have
490 * requests waiting to be processed. It is sorted in the order that
491 * we will service the queues.
493 static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
496 struct rb_node **p, *parent;
497 struct cfq_queue *__cfqq;
498 unsigned long rb_key;
501 if (cfq_class_idle(cfqq)) {
502 rb_key = CFQ_IDLE_DELAY;
503 parent = rb_last(&cfqd->service_tree.rb);
504 if (parent && parent != &cfqq->rb_node) {
505 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
506 rb_key += __cfqq->rb_key;
509 } else if (!add_front) {
510 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
511 rb_key += cfqq->slice_resid;
512 cfqq->slice_resid = 0;
516 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
518 * same position, nothing more to do
520 if (rb_key == cfqq->rb_key)
523 cfq_rb_erase(&cfqq->rb_node, &cfqd->service_tree);
528 p = &cfqd->service_tree.rb.rb_node;
533 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
536 * sort RT queues first, we always want to give
537 * preference to them. IDLE queues goes to the back.
538 * after that, sort on the next service time.
540 if (cfq_class_rt(cfqq) > cfq_class_rt(__cfqq))
542 else if (cfq_class_rt(cfqq) < cfq_class_rt(__cfqq))
544 else if (cfq_class_idle(cfqq) < cfq_class_idle(__cfqq))
546 else if (cfq_class_idle(cfqq) > cfq_class_idle(__cfqq))
548 else if (rb_key < __cfqq->rb_key)
553 if (n == &(*p)->rb_right)
560 cfqd->service_tree.left = &cfqq->rb_node;
562 cfqq->rb_key = rb_key;
563 rb_link_node(&cfqq->rb_node, parent, p);
564 rb_insert_color(&cfqq->rb_node, &cfqd->service_tree.rb);
567 static struct cfq_queue *
568 cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
569 sector_t sector, struct rb_node **ret_parent,
570 struct rb_node ***rb_link)
572 struct rb_node **p, *parent;
573 struct cfq_queue *cfqq = NULL;
581 cfqq = rb_entry(parent, struct cfq_queue, p_node);
584 * Sort strictly based on sector. Smallest to the left,
585 * largest to the right.
587 if (sector > blk_rq_pos(cfqq->next_rq))
589 else if (sector < blk_rq_pos(cfqq->next_rq))
597 *ret_parent = parent;
603 static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
605 struct rb_node **p, *parent;
606 struct cfq_queue *__cfqq;
609 rb_erase(&cfqq->p_node, cfqq->p_root);
613 if (cfq_class_idle(cfqq))
618 cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
619 __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
620 blk_rq_pos(cfqq->next_rq), &parent, &p);
622 rb_link_node(&cfqq->p_node, parent, p);
623 rb_insert_color(&cfqq->p_node, cfqq->p_root);
629 * Update cfqq's position in the service tree.
631 static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
634 * Resorting requires the cfqq to be on the RR list already.
636 if (cfq_cfqq_on_rr(cfqq)) {
637 cfq_service_tree_add(cfqd, cfqq, 0);
638 cfq_prio_tree_add(cfqd, cfqq);
643 * add to busy list of queues for service, trying to be fair in ordering
644 * the pending list according to last request service
646 static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
648 cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
649 BUG_ON(cfq_cfqq_on_rr(cfqq));
650 cfq_mark_cfqq_on_rr(cfqq);
653 cfq_resort_rr_list(cfqd, cfqq);
657 * Called when the cfqq no longer has requests pending, remove it from
660 static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
662 cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
663 BUG_ON(!cfq_cfqq_on_rr(cfqq));
664 cfq_clear_cfqq_on_rr(cfqq);
666 if (!RB_EMPTY_NODE(&cfqq->rb_node))
667 cfq_rb_erase(&cfqq->rb_node, &cfqd->service_tree);
669 rb_erase(&cfqq->p_node, cfqq->p_root);
673 BUG_ON(!cfqd->busy_queues);
678 * rb tree support functions
680 static void cfq_del_rq_rb(struct request *rq)
682 struct cfq_queue *cfqq = RQ_CFQQ(rq);
683 struct cfq_data *cfqd = cfqq->cfqd;
684 const int sync = rq_is_sync(rq);
686 BUG_ON(!cfqq->queued[sync]);
687 cfqq->queued[sync]--;
689 elv_rb_del(&cfqq->sort_list, rq);
691 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
692 cfq_del_cfqq_rr(cfqd, cfqq);
695 static void cfq_add_rq_rb(struct request *rq)
697 struct cfq_queue *cfqq = RQ_CFQQ(rq);
698 struct cfq_data *cfqd = cfqq->cfqd;
699 struct request *__alias, *prev;
701 cfqq->queued[rq_is_sync(rq)]++;
704 * looks a little odd, but the first insert might return an alias.
705 * if that happens, put the alias on the dispatch list
707 while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL)
708 cfq_dispatch_insert(cfqd->queue, __alias);
710 if (!cfq_cfqq_on_rr(cfqq))
711 cfq_add_cfqq_rr(cfqd, cfqq);
714 * check if this request is a better next-serve candidate
716 prev = cfqq->next_rq;
717 cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq);
720 * adjust priority tree position, if ->next_rq changes
722 if (prev != cfqq->next_rq)
723 cfq_prio_tree_add(cfqd, cfqq);
725 BUG_ON(!cfqq->next_rq);
728 static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
730 elv_rb_del(&cfqq->sort_list, rq);
731 cfqq->queued[rq_is_sync(rq)]--;
735 static struct request *
736 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
738 struct task_struct *tsk = current;
739 struct cfq_io_context *cic;
740 struct cfq_queue *cfqq;
742 cic = cfq_cic_lookup(cfqd, tsk->io_context);
746 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
748 sector_t sector = bio->bi_sector + bio_sectors(bio);
750 return elv_rb_find(&cfqq->sort_list, sector);
756 static void cfq_activate_request(struct request_queue *q, struct request *rq)
758 struct cfq_data *cfqd = q->elevator->elevator_data;
760 cfqd->rq_in_driver[rq_is_sync(rq)]++;
761 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
764 cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
767 static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
769 struct cfq_data *cfqd = q->elevator->elevator_data;
770 const int sync = rq_is_sync(rq);
772 WARN_ON(!cfqd->rq_in_driver[sync]);
773 cfqd->rq_in_driver[sync]--;
774 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
778 static void cfq_remove_request(struct request *rq)
780 struct cfq_queue *cfqq = RQ_CFQQ(rq);
782 if (cfqq->next_rq == rq)
783 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
785 list_del_init(&rq->queuelist);
788 cfqq->cfqd->rq_queued--;
789 if (rq_is_meta(rq)) {
790 WARN_ON(!cfqq->meta_pending);
791 cfqq->meta_pending--;
795 static int cfq_merge(struct request_queue *q, struct request **req,
798 struct cfq_data *cfqd = q->elevator->elevator_data;
799 struct request *__rq;
801 __rq = cfq_find_rq_fmerge(cfqd, bio);
802 if (__rq && elv_rq_merge_ok(__rq, bio)) {
804 return ELEVATOR_FRONT_MERGE;
807 return ELEVATOR_NO_MERGE;
810 static void cfq_merged_request(struct request_queue *q, struct request *req,
813 if (type == ELEVATOR_FRONT_MERGE) {
814 struct cfq_queue *cfqq = RQ_CFQQ(req);
816 cfq_reposition_rq_rb(cfqq, req);
821 cfq_merged_requests(struct request_queue *q, struct request *rq,
822 struct request *next)
825 * reposition in fifo if next is older than rq
827 if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
828 time_before(next->start_time, rq->start_time))
829 list_move(&rq->queuelist, &next->queuelist);
831 cfq_remove_request(next);
834 static int cfq_allow_merge(struct request_queue *q, struct request *rq,
837 struct cfq_data *cfqd = q->elevator->elevator_data;
838 struct cfq_io_context *cic;
839 struct cfq_queue *cfqq;
842 * Disallow merge of a sync bio into an async request.
844 if (cfq_bio_sync(bio) && !rq_is_sync(rq))
848 * Lookup the cfqq that this bio will be queued with. Allow
849 * merge only if rq is queued there.
851 cic = cfq_cic_lookup(cfqd, current->io_context);
855 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
856 if (cfqq == RQ_CFQQ(rq))
862 static void __cfq_set_active_queue(struct cfq_data *cfqd,
863 struct cfq_queue *cfqq)
866 cfq_log_cfqq(cfqd, cfqq, "set_active");
868 cfqq->slice_dispatch = 0;
870 cfq_clear_cfqq_wait_request(cfqq);
871 cfq_clear_cfqq_must_dispatch(cfqq);
872 cfq_clear_cfqq_must_alloc_slice(cfqq);
873 cfq_clear_cfqq_fifo_expire(cfqq);
874 cfq_mark_cfqq_slice_new(cfqq);
876 del_timer(&cfqd->idle_slice_timer);
879 cfqd->active_queue = cfqq;
883 * current cfqq expired its slice (or was too idle), select new one
886 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
889 cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
891 if (cfq_cfqq_wait_request(cfqq))
892 del_timer(&cfqd->idle_slice_timer);
894 cfq_clear_cfqq_wait_request(cfqq);
897 * store what was left of this slice, if the queue idled/timed out
899 if (timed_out && !cfq_cfqq_slice_new(cfqq)) {
900 cfqq->slice_resid = cfqq->slice_end - jiffies;
901 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
904 cfq_resort_rr_list(cfqd, cfqq);
906 if (cfqq == cfqd->active_queue)
907 cfqd->active_queue = NULL;
909 if (cfqd->active_cic) {
910 put_io_context(cfqd->active_cic->ioc);
911 cfqd->active_cic = NULL;
915 static inline void cfq_slice_expired(struct cfq_data *cfqd, int timed_out)
917 struct cfq_queue *cfqq = cfqd->active_queue;
920 __cfq_slice_expired(cfqd, cfqq, timed_out);
924 * Get next queue for service. Unless we have a queue preemption,
925 * we'll simply select the first cfqq in the service tree.
927 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
929 if (RB_EMPTY_ROOT(&cfqd->service_tree.rb))
932 return cfq_rb_first(&cfqd->service_tree);
936 * Get and set a new active queue for service.
938 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
939 struct cfq_queue *cfqq)
942 cfqq = cfq_get_next_queue(cfqd);
944 cfq_clear_cfqq_coop(cfqq);
947 __cfq_set_active_queue(cfqd, cfqq);
951 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
954 if (blk_rq_pos(rq) >= cfqd->last_position)
955 return blk_rq_pos(rq) - cfqd->last_position;
957 return cfqd->last_position - blk_rq_pos(rq);
960 #define CIC_SEEK_THR 8 * 1024
961 #define CIC_SEEKY(cic) ((cic)->seek_mean > CIC_SEEK_THR)
963 static inline int cfq_rq_close(struct cfq_data *cfqd, struct request *rq)
965 struct cfq_io_context *cic = cfqd->active_cic;
966 sector_t sdist = cic->seek_mean;
968 if (!sample_valid(cic->seek_samples))
969 sdist = CIC_SEEK_THR;
971 return cfq_dist_from_last(cfqd, rq) <= sdist;
974 static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
975 struct cfq_queue *cur_cfqq)
977 struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
978 struct rb_node *parent, *node;
979 struct cfq_queue *__cfqq;
980 sector_t sector = cfqd->last_position;
982 if (RB_EMPTY_ROOT(root))
986 * First, if we find a request starting at the end of the last
987 * request, choose it.
989 __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
994 * If the exact sector wasn't found, the parent of the NULL leaf
995 * will contain the closest sector.
997 __cfqq = rb_entry(parent, struct cfq_queue, p_node);
998 if (cfq_rq_close(cfqd, __cfqq->next_rq))
1001 if (blk_rq_pos(__cfqq->next_rq) < sector)
1002 node = rb_next(&__cfqq->p_node);
1004 node = rb_prev(&__cfqq->p_node);
1008 __cfqq = rb_entry(node, struct cfq_queue, p_node);
1009 if (cfq_rq_close(cfqd, __cfqq->next_rq))
1017 * cur_cfqq - passed in so that we don't decide that the current queue is
1018 * closely cooperating with itself.
1020 * So, basically we're assuming that that cur_cfqq has dispatched at least
1021 * one request, and that cfqd->last_position reflects a position on the disk
1022 * associated with the I/O issued by cur_cfqq. I'm not sure this is a valid
1025 static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
1026 struct cfq_queue *cur_cfqq,
1029 struct cfq_queue *cfqq;
1032 * A valid cfq_io_context is necessary to compare requests against
1033 * the seek_mean of the current cfqq.
1035 if (!cfqd->active_cic)
1039 * We should notice if some of the queues are cooperating, eg
1040 * working closely on the same area of the disk. In that case,
1041 * we can group them together and don't waste time idling.
1043 cfqq = cfqq_close(cfqd, cur_cfqq);
1047 if (cfq_cfqq_coop(cfqq))
1051 cfq_mark_cfqq_coop(cfqq);
1055 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
1057 struct cfq_queue *cfqq = cfqd->active_queue;
1058 struct cfq_io_context *cic;
1062 * SSD device without seek penalty, disable idling. But only do so
1063 * for devices that support queuing, otherwise we still have a problem
1064 * with sync vs async workloads.
1066 if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
1069 WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
1070 WARN_ON(cfq_cfqq_slice_new(cfqq));
1073 * idle is disabled, either manually or by past process history
1075 if (!cfqd->cfq_slice_idle || !cfq_cfqq_idle_window(cfqq))
1079 * still requests with the driver, don't idle
1081 if (rq_in_driver(cfqd))
1085 * task has exited, don't wait
1087 cic = cfqd->active_cic;
1088 if (!cic || !atomic_read(&cic->ioc->nr_tasks))
1091 cfq_mark_cfqq_wait_request(cfqq);
1094 * we don't want to idle for seeks, but we do want to allow
1095 * fair distribution of slice time for a process doing back-to-back
1096 * seeks. so allow a little bit of time for him to submit a new rq
1098 sl = cfqd->cfq_slice_idle;
1099 if (sample_valid(cic->seek_samples) && CIC_SEEKY(cic))
1100 sl = min(sl, msecs_to_jiffies(CFQ_MIN_TT));
1102 mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
1103 cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu", sl);
1107 * Move request from internal lists to the request queue dispatch list.
1109 static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
1111 struct cfq_data *cfqd = q->elevator->elevator_data;
1112 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1114 cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
1116 cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
1117 cfq_remove_request(rq);
1119 elv_dispatch_sort(q, rq);
1121 if (cfq_cfqq_sync(cfqq))
1122 cfqd->sync_flight++;
1126 * return expired entry, or NULL to just start from scratch in rbtree
1128 static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
1130 struct cfq_data *cfqd = cfqq->cfqd;
1134 if (cfq_cfqq_fifo_expire(cfqq))
1137 cfq_mark_cfqq_fifo_expire(cfqq);
1139 if (list_empty(&cfqq->fifo))
1142 fifo = cfq_cfqq_sync(cfqq);
1143 rq = rq_entry_fifo(cfqq->fifo.next);
1145 if (time_before(jiffies, rq->start_time + cfqd->cfq_fifo_expire[fifo]))
1148 cfq_log_cfqq(cfqd, cfqq, "fifo=%p", rq);
1153 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1155 const int base_rq = cfqd->cfq_slice_async_rq;
1157 WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
1159 return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
1163 * Select a queue for service. If we have a current active queue,
1164 * check whether to continue servicing it, or retrieve and set a new one.
1166 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
1168 struct cfq_queue *cfqq, *new_cfqq = NULL;
1170 cfqq = cfqd->active_queue;
1175 * The active queue has run out of time, expire it and select new.
1177 if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq))
1181 * The active queue has requests and isn't expired, allow it to
1184 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
1188 * If another queue has a request waiting within our mean seek
1189 * distance, let it run. The expire code will check for close
1190 * cooperators and put the close queue at the front of the service
1193 new_cfqq = cfq_close_cooperator(cfqd, cfqq, 0);
1198 * No requests pending. If the active queue still has requests in
1199 * flight or is idling for a new request, allow either of these
1200 * conditions to happen (or time out) before selecting a new queue.
1202 if (timer_pending(&cfqd->idle_slice_timer) ||
1203 (cfqq->dispatched && cfq_cfqq_idle_window(cfqq))) {
1209 cfq_slice_expired(cfqd, 0);
1211 cfqq = cfq_set_active_queue(cfqd, new_cfqq);
1216 static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
1220 while (cfqq->next_rq) {
1221 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
1225 BUG_ON(!list_empty(&cfqq->fifo));
1230 * Drain our current requests. Used for barriers and when switching
1231 * io schedulers on-the-fly.
1233 static int cfq_forced_dispatch(struct cfq_data *cfqd)
1235 struct cfq_queue *cfqq;
1238 while ((cfqq = cfq_rb_first(&cfqd->service_tree)) != NULL)
1239 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
1241 cfq_slice_expired(cfqd, 0);
1243 BUG_ON(cfqd->busy_queues);
1245 cfq_log(cfqd, "forced_dispatch=%d", dispatched);
1250 * Dispatch a request from cfqq, moving them to the request queue
1253 static void cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1257 BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
1260 * follow expired path, else get first next available
1262 rq = cfq_check_fifo(cfqq);
1267 * insert request into driver dispatch list
1269 cfq_dispatch_insert(cfqd->queue, rq);
1271 if (!cfqd->active_cic) {
1272 struct cfq_io_context *cic = RQ_CIC(rq);
1274 atomic_long_inc(&cic->ioc->refcount);
1275 cfqd->active_cic = cic;
1280 * Find the cfqq that we need to service and move a request from that to the
1283 static int cfq_dispatch_requests(struct request_queue *q, int force)
1285 struct cfq_data *cfqd = q->elevator->elevator_data;
1286 struct cfq_queue *cfqq;
1287 unsigned int max_dispatch;
1289 if (!cfqd->busy_queues)
1292 if (unlikely(force))
1293 return cfq_forced_dispatch(cfqd);
1295 cfqq = cfq_select_queue(cfqd);
1300 * Drain async requests before we start sync IO
1302 if (cfq_cfqq_idle_window(cfqq) && cfqd->rq_in_driver[BLK_RW_ASYNC])
1306 * If this is an async queue and we have sync IO in flight, let it wait
1308 if (cfqd->sync_flight && !cfq_cfqq_sync(cfqq))
1311 max_dispatch = cfqd->cfq_quantum;
1312 if (cfq_class_idle(cfqq))
1316 * Does this cfqq already have too much IO in flight?
1318 if (cfqq->dispatched >= max_dispatch) {
1319 unsigned long load_at = cfqd->last_end_sync_rq + cfq_slice_sync;
1322 * idle queue must always only have a single IO in flight
1324 if (cfq_class_idle(cfqq))
1328 * We have other queues, don't allow more IO from this one
1330 if (cfqd->busy_queues > 1)
1334 * If a sync request has completed recently, don't overload
1335 * the dispatch queue yet with async requests.
1337 if (cfqd->cfq_desktop && !cfq_cfqq_sync(cfqq)
1338 && time_before(jiffies, load_at))
1342 * we are the only queue, allow up to 4 times of 'quantum'
1344 if (cfqq->dispatched >= 4 * max_dispatch)
1349 * Dispatch a request from this cfqq
1351 cfq_dispatch_request(cfqd, cfqq);
1352 cfqq->slice_dispatch++;
1353 cfq_clear_cfqq_must_dispatch(cfqq);
1356 * expire an async queue immediately if it has used up its slice. idle
1357 * queue always expire after 1 dispatch round.
1359 if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
1360 cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
1361 cfq_class_idle(cfqq))) {
1362 cfqq->slice_end = jiffies + 1;
1363 cfq_slice_expired(cfqd, 0);
1366 cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
1371 * task holds one reference to the queue, dropped when task exits. each rq
1372 * in-flight on this queue also holds a reference, dropped when rq is freed.
1374 * queue lock must be held here.
1376 static void cfq_put_queue(struct cfq_queue *cfqq)
1378 struct cfq_data *cfqd = cfqq->cfqd;
1380 BUG_ON(atomic_read(&cfqq->ref) <= 0);
1382 if (!atomic_dec_and_test(&cfqq->ref))
1385 cfq_log_cfqq(cfqd, cfqq, "put_queue");
1386 BUG_ON(rb_first(&cfqq->sort_list));
1387 BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
1388 BUG_ON(cfq_cfqq_on_rr(cfqq));
1390 if (unlikely(cfqd->active_queue == cfqq)) {
1391 __cfq_slice_expired(cfqd, cfqq, 0);
1392 cfq_schedule_dispatch(cfqd);
1395 kmem_cache_free(cfq_pool, cfqq);
1399 * Must always be called with the rcu_read_lock() held
1402 __call_for_each_cic(struct io_context *ioc,
1403 void (*func)(struct io_context *, struct cfq_io_context *))
1405 struct cfq_io_context *cic;
1406 struct hlist_node *n;
1408 hlist_for_each_entry_rcu(cic, n, &ioc->cic_list, cic_list)
1413 * Call func for each cic attached to this ioc.
1416 call_for_each_cic(struct io_context *ioc,
1417 void (*func)(struct io_context *, struct cfq_io_context *))
1420 __call_for_each_cic(ioc, func);
1424 static void cfq_cic_free_rcu(struct rcu_head *head)
1426 struct cfq_io_context *cic;
1428 cic = container_of(head, struct cfq_io_context, rcu_head);
1430 kmem_cache_free(cfq_ioc_pool, cic);
1431 elv_ioc_count_dec(cfq_ioc_count);
1435 * CFQ scheduler is exiting, grab exit lock and check
1436 * the pending io context count. If it hits zero,
1437 * complete ioc_gone and set it back to NULL
1439 spin_lock(&ioc_gone_lock);
1440 if (ioc_gone && !elv_ioc_count_read(cfq_ioc_count)) {
1444 spin_unlock(&ioc_gone_lock);
1448 static void cfq_cic_free(struct cfq_io_context *cic)
1450 call_rcu(&cic->rcu_head, cfq_cic_free_rcu);
1453 static void cic_free_func(struct io_context *ioc, struct cfq_io_context *cic)
1455 unsigned long flags;
1457 BUG_ON(!cic->dead_key);
1459 spin_lock_irqsave(&ioc->lock, flags);
1460 radix_tree_delete(&ioc->radix_root, cic->dead_key);
1461 hlist_del_rcu(&cic->cic_list);
1462 spin_unlock_irqrestore(&ioc->lock, flags);
1468 * Must be called with rcu_read_lock() held or preemption otherwise disabled.
1469 * Only two callers of this - ->dtor() which is called with the rcu_read_lock(),
1470 * and ->trim() which is called with the task lock held
1472 static void cfq_free_io_context(struct io_context *ioc)
1475 * ioc->refcount is zero here, or we are called from elv_unregister(),
1476 * so no more cic's are allowed to be linked into this ioc. So it
1477 * should be ok to iterate over the known list, we will see all cic's
1478 * since no new ones are added.
1480 __call_for_each_cic(ioc, cic_free_func);
1483 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1485 if (unlikely(cfqq == cfqd->active_queue)) {
1486 __cfq_slice_expired(cfqd, cfqq, 0);
1487 cfq_schedule_dispatch(cfqd);
1490 cfq_put_queue(cfqq);
1493 static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
1494 struct cfq_io_context *cic)
1496 struct io_context *ioc = cic->ioc;
1498 list_del_init(&cic->queue_list);
1501 * Make sure key == NULL is seen for dead queues
1504 cic->dead_key = (unsigned long) cic->key;
1507 if (ioc->ioc_data == cic)
1508 rcu_assign_pointer(ioc->ioc_data, NULL);
1510 if (cic->cfqq[BLK_RW_ASYNC]) {
1511 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]);
1512 cic->cfqq[BLK_RW_ASYNC] = NULL;
1515 if (cic->cfqq[BLK_RW_SYNC]) {
1516 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]);
1517 cic->cfqq[BLK_RW_SYNC] = NULL;
1521 static void cfq_exit_single_io_context(struct io_context *ioc,
1522 struct cfq_io_context *cic)
1524 struct cfq_data *cfqd = cic->key;
1527 struct request_queue *q = cfqd->queue;
1528 unsigned long flags;
1530 spin_lock_irqsave(q->queue_lock, flags);
1533 * Ensure we get a fresh copy of the ->key to prevent
1534 * race between exiting task and queue
1536 smp_read_barrier_depends();
1538 __cfq_exit_single_io_context(cfqd, cic);
1540 spin_unlock_irqrestore(q->queue_lock, flags);
1545 * The process that ioc belongs to has exited, we need to clean up
1546 * and put the internal structures we have that belongs to that process.
1548 static void cfq_exit_io_context(struct io_context *ioc)
1550 call_for_each_cic(ioc, cfq_exit_single_io_context);
1553 static struct cfq_io_context *
1554 cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
1556 struct cfq_io_context *cic;
1558 cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO,
1561 cic->last_end_request = jiffies;
1562 INIT_LIST_HEAD(&cic->queue_list);
1563 INIT_HLIST_NODE(&cic->cic_list);
1564 cic->dtor = cfq_free_io_context;
1565 cic->exit = cfq_exit_io_context;
1566 elv_ioc_count_inc(cfq_ioc_count);
1572 static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc)
1574 struct task_struct *tsk = current;
1577 if (!cfq_cfqq_prio_changed(cfqq))
1580 ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio);
1581 switch (ioprio_class) {
1583 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
1584 case IOPRIO_CLASS_NONE:
1586 * no prio set, inherit CPU scheduling settings
1588 cfqq->ioprio = task_nice_ioprio(tsk);
1589 cfqq->ioprio_class = task_nice_ioclass(tsk);
1591 case IOPRIO_CLASS_RT:
1592 cfqq->ioprio = task_ioprio(ioc);
1593 cfqq->ioprio_class = IOPRIO_CLASS_RT;
1595 case IOPRIO_CLASS_BE:
1596 cfqq->ioprio = task_ioprio(ioc);
1597 cfqq->ioprio_class = IOPRIO_CLASS_BE;
1599 case IOPRIO_CLASS_IDLE:
1600 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
1602 cfq_clear_cfqq_idle_window(cfqq);
1607 * keep track of original prio settings in case we have to temporarily
1608 * elevate the priority of this queue
1610 cfqq->org_ioprio = cfqq->ioprio;
1611 cfqq->org_ioprio_class = cfqq->ioprio_class;
1612 cfq_clear_cfqq_prio_changed(cfqq);
1615 static void changed_ioprio(struct io_context *ioc, struct cfq_io_context *cic)
1617 struct cfq_data *cfqd = cic->key;
1618 struct cfq_queue *cfqq;
1619 unsigned long flags;
1621 if (unlikely(!cfqd))
1624 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1626 cfqq = cic->cfqq[BLK_RW_ASYNC];
1628 struct cfq_queue *new_cfqq;
1629 new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic->ioc,
1632 cic->cfqq[BLK_RW_ASYNC] = new_cfqq;
1633 cfq_put_queue(cfqq);
1637 cfqq = cic->cfqq[BLK_RW_SYNC];
1639 cfq_mark_cfqq_prio_changed(cfqq);
1641 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1644 static void cfq_ioc_set_ioprio(struct io_context *ioc)
1646 call_for_each_cic(ioc, changed_ioprio);
1647 ioc->ioprio_changed = 0;
1650 static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1651 pid_t pid, int is_sync)
1653 RB_CLEAR_NODE(&cfqq->rb_node);
1654 RB_CLEAR_NODE(&cfqq->p_node);
1655 INIT_LIST_HEAD(&cfqq->fifo);
1657 atomic_set(&cfqq->ref, 0);
1660 cfq_mark_cfqq_prio_changed(cfqq);
1663 if (!cfq_class_idle(cfqq))
1664 cfq_mark_cfqq_idle_window(cfqq);
1665 cfq_mark_cfqq_sync(cfqq);
1670 static struct cfq_queue *
1671 cfq_find_alloc_queue(struct cfq_data *cfqd, int is_sync,
1672 struct io_context *ioc, gfp_t gfp_mask)
1674 struct cfq_queue *cfqq, *new_cfqq = NULL;
1675 struct cfq_io_context *cic;
1678 cic = cfq_cic_lookup(cfqd, ioc);
1679 /* cic always exists here */
1680 cfqq = cic_to_cfqq(cic, is_sync);
1683 * Always try a new alloc if we fell back to the OOM cfqq
1684 * originally, since it should just be a temporary situation.
1686 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
1691 } else if (gfp_mask & __GFP_WAIT) {
1692 spin_unlock_irq(cfqd->queue->queue_lock);
1693 new_cfqq = kmem_cache_alloc_node(cfq_pool,
1694 gfp_mask | __GFP_ZERO,
1696 spin_lock_irq(cfqd->queue->queue_lock);
1700 cfqq = kmem_cache_alloc_node(cfq_pool,
1701 gfp_mask | __GFP_ZERO,
1706 cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
1707 cfq_init_prio_data(cfqq, ioc);
1708 cfq_log_cfqq(cfqd, cfqq, "alloced");
1710 cfqq = &cfqd->oom_cfqq;
1714 kmem_cache_free(cfq_pool, new_cfqq);
1719 static struct cfq_queue **
1720 cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
1722 switch (ioprio_class) {
1723 case IOPRIO_CLASS_RT:
1724 return &cfqd->async_cfqq[0][ioprio];
1725 case IOPRIO_CLASS_BE:
1726 return &cfqd->async_cfqq[1][ioprio];
1727 case IOPRIO_CLASS_IDLE:
1728 return &cfqd->async_idle_cfqq;
1734 static struct cfq_queue *
1735 cfq_get_queue(struct cfq_data *cfqd, int is_sync, struct io_context *ioc,
1738 const int ioprio = task_ioprio(ioc);
1739 const int ioprio_class = task_ioprio_class(ioc);
1740 struct cfq_queue **async_cfqq = NULL;
1741 struct cfq_queue *cfqq = NULL;
1744 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
1749 cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask);
1752 * pin the queue now that it's allocated, scheduler exit will prune it
1754 if (!is_sync && !(*async_cfqq)) {
1755 atomic_inc(&cfqq->ref);
1759 atomic_inc(&cfqq->ref);
1764 * We drop cfq io contexts lazily, so we may find a dead one.
1767 cfq_drop_dead_cic(struct cfq_data *cfqd, struct io_context *ioc,
1768 struct cfq_io_context *cic)
1770 unsigned long flags;
1772 WARN_ON(!list_empty(&cic->queue_list));
1774 spin_lock_irqsave(&ioc->lock, flags);
1776 BUG_ON(ioc->ioc_data == cic);
1778 radix_tree_delete(&ioc->radix_root, (unsigned long) cfqd);
1779 hlist_del_rcu(&cic->cic_list);
1780 spin_unlock_irqrestore(&ioc->lock, flags);
1785 static struct cfq_io_context *
1786 cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc)
1788 struct cfq_io_context *cic;
1789 unsigned long flags;
1798 * we maintain a last-hit cache, to avoid browsing over the tree
1800 cic = rcu_dereference(ioc->ioc_data);
1801 if (cic && cic->key == cfqd) {
1807 cic = radix_tree_lookup(&ioc->radix_root, (unsigned long) cfqd);
1811 /* ->key must be copied to avoid race with cfq_exit_queue() */
1814 cfq_drop_dead_cic(cfqd, ioc, cic);
1819 spin_lock_irqsave(&ioc->lock, flags);
1820 rcu_assign_pointer(ioc->ioc_data, cic);
1821 spin_unlock_irqrestore(&ioc->lock, flags);
1829 * Add cic into ioc, using cfqd as the search key. This enables us to lookup
1830 * the process specific cfq io context when entered from the block layer.
1831 * Also adds the cic to a per-cfqd list, used when this queue is removed.
1833 static int cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
1834 struct cfq_io_context *cic, gfp_t gfp_mask)
1836 unsigned long flags;
1839 ret = radix_tree_preload(gfp_mask);
1844 spin_lock_irqsave(&ioc->lock, flags);
1845 ret = radix_tree_insert(&ioc->radix_root,
1846 (unsigned long) cfqd, cic);
1848 hlist_add_head_rcu(&cic->cic_list, &ioc->cic_list);
1849 spin_unlock_irqrestore(&ioc->lock, flags);
1851 radix_tree_preload_end();
1854 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1855 list_add(&cic->queue_list, &cfqd->cic_list);
1856 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1861 printk(KERN_ERR "cfq: cic link failed!\n");
1867 * Setup general io context and cfq io context. There can be several cfq
1868 * io contexts per general io context, if this process is doing io to more
1869 * than one device managed by cfq.
1871 static struct cfq_io_context *
1872 cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
1874 struct io_context *ioc = NULL;
1875 struct cfq_io_context *cic;
1877 might_sleep_if(gfp_mask & __GFP_WAIT);
1879 ioc = get_io_context(gfp_mask, cfqd->queue->node);
1883 cic = cfq_cic_lookup(cfqd, ioc);
1887 cic = cfq_alloc_io_context(cfqd, gfp_mask);
1891 if (cfq_cic_link(cfqd, ioc, cic, gfp_mask))
1895 smp_read_barrier_depends();
1896 if (unlikely(ioc->ioprio_changed))
1897 cfq_ioc_set_ioprio(ioc);
1903 put_io_context(ioc);
1908 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
1910 unsigned long elapsed = jiffies - cic->last_end_request;
1911 unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
1913 cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
1914 cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
1915 cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
1919 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_io_context *cic,
1925 if (!cic->last_request_pos)
1927 else if (cic->last_request_pos < blk_rq_pos(rq))
1928 sdist = blk_rq_pos(rq) - cic->last_request_pos;
1930 sdist = cic->last_request_pos - blk_rq_pos(rq);
1933 * Don't allow the seek distance to get too large from the
1934 * odd fragment, pagein, etc
1936 if (cic->seek_samples <= 60) /* second&third seek */
1937 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*1024);
1939 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*64);
1941 cic->seek_samples = (7*cic->seek_samples + 256) / 8;
1942 cic->seek_total = (7*cic->seek_total + (u64)256*sdist) / 8;
1943 total = cic->seek_total + (cic->seek_samples/2);
1944 do_div(total, cic->seek_samples);
1945 cic->seek_mean = (sector_t)total;
1949 * Disable idle window if the process thinks too long or seeks so much that
1953 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1954 struct cfq_io_context *cic)
1956 int old_idle, enable_idle;
1959 * Don't idle for async or idle io prio class
1961 if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
1964 enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
1966 if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle ||
1967 (!cfqd->cfq_desktop && cfqd->hw_tag && CIC_SEEKY(cic)))
1969 else if (sample_valid(cic->ttime_samples)) {
1970 if (cic->ttime_mean > cfqd->cfq_slice_idle)
1976 if (old_idle != enable_idle) {
1977 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
1979 cfq_mark_cfqq_idle_window(cfqq);
1981 cfq_clear_cfqq_idle_window(cfqq);
1986 * Check if new_cfqq should preempt the currently active queue. Return 0 for
1987 * no or if we aren't sure, a 1 will cause a preempt.
1990 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
1993 struct cfq_queue *cfqq;
1995 cfqq = cfqd->active_queue;
1999 if (cfq_slice_used(cfqq))
2002 if (cfq_class_idle(new_cfqq))
2005 if (cfq_class_idle(cfqq))
2009 * if the new request is sync, but the currently running queue is
2010 * not, let the sync request have priority.
2012 if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
2016 * So both queues are sync. Let the new request get disk time if
2017 * it's a metadata request and the current queue is doing regular IO.
2019 if (rq_is_meta(rq) && !cfqq->meta_pending)
2023 * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
2025 if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
2028 if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
2032 * if this request is as-good as one we would expect from the
2033 * current cfqq, let it preempt
2035 if (cfq_rq_close(cfqd, rq))
2042 * cfqq preempts the active queue. if we allowed preempt with no slice left,
2043 * let it have half of its nominal slice.
2045 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2047 cfq_log_cfqq(cfqd, cfqq, "preempt");
2048 cfq_slice_expired(cfqd, 1);
2051 * Put the new queue at the front of the of the current list,
2052 * so we know that it will be selected next.
2054 BUG_ON(!cfq_cfqq_on_rr(cfqq));
2056 cfq_service_tree_add(cfqd, cfqq, 1);
2058 cfqq->slice_end = 0;
2059 cfq_mark_cfqq_slice_new(cfqq);
2063 * Called when a new fs request (rq) is added (to cfqq). Check if there's
2064 * something we should do about it
2067 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2070 struct cfq_io_context *cic = RQ_CIC(rq);
2074 cfqq->meta_pending++;
2076 cfq_update_io_thinktime(cfqd, cic);
2077 cfq_update_io_seektime(cfqd, cic, rq);
2078 cfq_update_idle_window(cfqd, cfqq, cic);
2080 cic->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
2082 if (cfqq == cfqd->active_queue) {
2084 * Remember that we saw a request from this process, but
2085 * don't start queuing just yet. Otherwise we risk seeing lots
2086 * of tiny requests, because we disrupt the normal plugging
2087 * and merging. If the request is already larger than a single
2088 * page, let it rip immediately. For that case we assume that
2089 * merging is already done. Ditto for a busy system that
2090 * has other work pending, don't risk delaying until the
2091 * idle timer unplug to continue working.
2093 if (cfq_cfqq_wait_request(cfqq)) {
2094 if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
2095 cfqd->busy_queues > 1) {
2096 del_timer(&cfqd->idle_slice_timer);
2097 __blk_run_queue(cfqd->queue);
2099 cfq_mark_cfqq_must_dispatch(cfqq);
2101 } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
2103 * not the active queue - expire current slice if it is
2104 * idle and has expired it's mean thinktime or this new queue
2105 * has some old slice time left and is of higher priority or
2106 * this new queue is RT and the current one is BE
2108 cfq_preempt_queue(cfqd, cfqq);
2109 __blk_run_queue(cfqd->queue);
2113 static void cfq_insert_request(struct request_queue *q, struct request *rq)
2115 struct cfq_data *cfqd = q->elevator->elevator_data;
2116 struct cfq_queue *cfqq = RQ_CFQQ(rq);
2118 cfq_log_cfqq(cfqd, cfqq, "insert_request");
2119 cfq_init_prio_data(cfqq, RQ_CIC(rq)->ioc);
2123 list_add_tail(&rq->queuelist, &cfqq->fifo);
2125 cfq_rq_enqueued(cfqd, cfqq, rq);
2129 * Update hw_tag based on peak queue depth over 50 samples under
2132 static void cfq_update_hw_tag(struct cfq_data *cfqd)
2134 if (rq_in_driver(cfqd) > cfqd->rq_in_driver_peak)
2135 cfqd->rq_in_driver_peak = rq_in_driver(cfqd);
2137 if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
2138 rq_in_driver(cfqd) <= CFQ_HW_QUEUE_MIN)
2141 if (cfqd->hw_tag_samples++ < 50)
2144 if (cfqd->rq_in_driver_peak >= CFQ_HW_QUEUE_MIN)
2149 cfqd->hw_tag_samples = 0;
2150 cfqd->rq_in_driver_peak = 0;
2153 static void cfq_completed_request(struct request_queue *q, struct request *rq)
2155 struct cfq_queue *cfqq = RQ_CFQQ(rq);
2156 struct cfq_data *cfqd = cfqq->cfqd;
2157 const int sync = rq_is_sync(rq);
2161 cfq_log_cfqq(cfqd, cfqq, "complete");
2163 cfq_update_hw_tag(cfqd);
2165 WARN_ON(!cfqd->rq_in_driver[sync]);
2166 WARN_ON(!cfqq->dispatched);
2167 cfqd->rq_in_driver[sync]--;
2170 if (cfq_cfqq_sync(cfqq))
2171 cfqd->sync_flight--;
2174 RQ_CIC(rq)->last_end_request = now;
2175 cfqd->last_end_sync_rq = now;
2179 * If this is the active queue, check if it needs to be expired,
2180 * or if we want to idle in case it has no pending requests.
2182 if (cfqd->active_queue == cfqq) {
2183 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
2185 if (cfq_cfqq_slice_new(cfqq)) {
2186 cfq_set_prio_slice(cfqd, cfqq);
2187 cfq_clear_cfqq_slice_new(cfqq);
2190 * If there are no requests waiting in this queue, and
2191 * there are other queues ready to issue requests, AND
2192 * those other queues are issuing requests within our
2193 * mean seek distance, give them a chance to run instead
2196 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
2197 cfq_slice_expired(cfqd, 1);
2198 else if (cfqq_empty && !cfq_close_cooperator(cfqd, cfqq, 1) &&
2199 sync && !rq_noidle(rq))
2200 cfq_arm_slice_timer(cfqd);
2203 if (!rq_in_driver(cfqd))
2204 cfq_schedule_dispatch(cfqd);
2208 * we temporarily boost lower priority queues if they are holding fs exclusive
2209 * resources. they are boosted to normal prio (CLASS_BE/4)
2211 static void cfq_prio_boost(struct cfq_queue *cfqq)
2213 if (has_fs_excl()) {
2215 * boost idle prio on transactions that would lock out other
2216 * users of the filesystem
2218 if (cfq_class_idle(cfqq))
2219 cfqq->ioprio_class = IOPRIO_CLASS_BE;
2220 if (cfqq->ioprio > IOPRIO_NORM)
2221 cfqq->ioprio = IOPRIO_NORM;
2224 * check if we need to unboost the queue
2226 if (cfqq->ioprio_class != cfqq->org_ioprio_class)
2227 cfqq->ioprio_class = cfqq->org_ioprio_class;
2228 if (cfqq->ioprio != cfqq->org_ioprio)
2229 cfqq->ioprio = cfqq->org_ioprio;
2233 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
2235 if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
2236 cfq_mark_cfqq_must_alloc_slice(cfqq);
2237 return ELV_MQUEUE_MUST;
2240 return ELV_MQUEUE_MAY;
2243 static int cfq_may_queue(struct request_queue *q, int rw)
2245 struct cfq_data *cfqd = q->elevator->elevator_data;
2246 struct task_struct *tsk = current;
2247 struct cfq_io_context *cic;
2248 struct cfq_queue *cfqq;
2251 * don't force setup of a queue from here, as a call to may_queue
2252 * does not necessarily imply that a request actually will be queued.
2253 * so just lookup a possibly existing queue, or return 'may queue'
2256 cic = cfq_cic_lookup(cfqd, tsk->io_context);
2258 return ELV_MQUEUE_MAY;
2260 cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
2262 cfq_init_prio_data(cfqq, cic->ioc);
2263 cfq_prio_boost(cfqq);
2265 return __cfq_may_queue(cfqq);
2268 return ELV_MQUEUE_MAY;
2272 * queue lock held here
2274 static void cfq_put_request(struct request *rq)
2276 struct cfq_queue *cfqq = RQ_CFQQ(rq);
2279 const int rw = rq_data_dir(rq);
2281 BUG_ON(!cfqq->allocated[rw]);
2282 cfqq->allocated[rw]--;
2284 put_io_context(RQ_CIC(rq)->ioc);
2286 rq->elevator_private = NULL;
2287 rq->elevator_private2 = NULL;
2289 cfq_put_queue(cfqq);
2294 * Allocate cfq data structures associated with this request.
2297 cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
2299 struct cfq_data *cfqd = q->elevator->elevator_data;
2300 struct cfq_io_context *cic;
2301 const int rw = rq_data_dir(rq);
2302 const int is_sync = rq_is_sync(rq);
2303 struct cfq_queue *cfqq;
2304 unsigned long flags;
2306 might_sleep_if(gfp_mask & __GFP_WAIT);
2308 cic = cfq_get_io_context(cfqd, gfp_mask);
2310 spin_lock_irqsave(q->queue_lock, flags);
2315 cfqq = cic_to_cfqq(cic, is_sync);
2316 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
2317 cfqq = cfq_get_queue(cfqd, is_sync, cic->ioc, gfp_mask);
2318 cic_set_cfqq(cic, cfqq, is_sync);
2321 cfqq->allocated[rw]++;
2322 atomic_inc(&cfqq->ref);
2324 spin_unlock_irqrestore(q->queue_lock, flags);
2326 rq->elevator_private = cic;
2327 rq->elevator_private2 = cfqq;
2332 put_io_context(cic->ioc);
2334 cfq_schedule_dispatch(cfqd);
2335 spin_unlock_irqrestore(q->queue_lock, flags);
2336 cfq_log(cfqd, "set_request fail");
2340 static void cfq_kick_queue(struct work_struct *work)
2342 struct cfq_data *cfqd =
2343 container_of(work, struct cfq_data, unplug_work);
2344 struct request_queue *q = cfqd->queue;
2346 spin_lock_irq(q->queue_lock);
2347 __blk_run_queue(cfqd->queue);
2348 spin_unlock_irq(q->queue_lock);
2352 * Timer running if the active_queue is currently idling inside its time slice
2354 static void cfq_idle_slice_timer(unsigned long data)
2356 struct cfq_data *cfqd = (struct cfq_data *) data;
2357 struct cfq_queue *cfqq;
2358 unsigned long flags;
2361 cfq_log(cfqd, "idle timer fired");
2363 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2365 cfqq = cfqd->active_queue;
2370 * We saw a request before the queue expired, let it through
2372 if (cfq_cfqq_must_dispatch(cfqq))
2378 if (cfq_slice_used(cfqq))
2382 * only expire and reinvoke request handler, if there are
2383 * other queues with pending requests
2385 if (!cfqd->busy_queues)
2389 * not expired and it has a request pending, let it dispatch
2391 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
2395 cfq_slice_expired(cfqd, timed_out);
2397 cfq_schedule_dispatch(cfqd);
2399 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2402 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
2404 del_timer_sync(&cfqd->idle_slice_timer);
2405 cancel_work_sync(&cfqd->unplug_work);
2408 static void cfq_put_async_queues(struct cfq_data *cfqd)
2412 for (i = 0; i < IOPRIO_BE_NR; i++) {
2413 if (cfqd->async_cfqq[0][i])
2414 cfq_put_queue(cfqd->async_cfqq[0][i]);
2415 if (cfqd->async_cfqq[1][i])
2416 cfq_put_queue(cfqd->async_cfqq[1][i]);
2419 if (cfqd->async_idle_cfqq)
2420 cfq_put_queue(cfqd->async_idle_cfqq);
2423 static void cfq_exit_queue(struct elevator_queue *e)
2425 struct cfq_data *cfqd = e->elevator_data;
2426 struct request_queue *q = cfqd->queue;
2428 cfq_shutdown_timer_wq(cfqd);
2430 spin_lock_irq(q->queue_lock);
2432 if (cfqd->active_queue)
2433 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
2435 while (!list_empty(&cfqd->cic_list)) {
2436 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
2437 struct cfq_io_context,
2440 __cfq_exit_single_io_context(cfqd, cic);
2443 cfq_put_async_queues(cfqd);
2445 spin_unlock_irq(q->queue_lock);
2447 cfq_shutdown_timer_wq(cfqd);
2452 static void *cfq_init_queue(struct request_queue *q)
2454 struct cfq_data *cfqd;
2457 cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
2461 cfqd->service_tree = CFQ_RB_ROOT;
2464 * Not strictly needed (since RB_ROOT just clears the node and we
2465 * zeroed cfqd on alloc), but better be safe in case someone decides
2466 * to add magic to the rb code
2468 for (i = 0; i < CFQ_PRIO_LISTS; i++)
2469 cfqd->prio_trees[i] = RB_ROOT;
2472 * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
2473 * Grab a permanent reference to it, so that the normal code flow
2474 * will not attempt to free it.
2476 cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
2477 atomic_inc(&cfqd->oom_cfqq.ref);
2479 INIT_LIST_HEAD(&cfqd->cic_list);
2483 init_timer(&cfqd->idle_slice_timer);
2484 cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
2485 cfqd->idle_slice_timer.data = (unsigned long) cfqd;
2487 INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
2489 cfqd->cfq_quantum = cfq_quantum;
2490 cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
2491 cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
2492 cfqd->cfq_back_max = cfq_back_max;
2493 cfqd->cfq_back_penalty = cfq_back_penalty;
2494 cfqd->cfq_slice[0] = cfq_slice_async;
2495 cfqd->cfq_slice[1] = cfq_slice_sync;
2496 cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
2497 cfqd->cfq_slice_idle = cfq_slice_idle;
2498 cfqd->cfq_desktop = 1;
2500 cfqd->last_end_sync_rq = jiffies;
2504 static void cfq_slab_kill(void)
2507 * Caller already ensured that pending RCU callbacks are completed,
2508 * so we should have no busy allocations at this point.
2511 kmem_cache_destroy(cfq_pool);
2513 kmem_cache_destroy(cfq_ioc_pool);
2516 static int __init cfq_slab_setup(void)
2518 cfq_pool = KMEM_CACHE(cfq_queue, 0);
2522 cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0);
2533 * sysfs parts below -->
2536 cfq_var_show(unsigned int var, char *page)
2538 return sprintf(page, "%d\n", var);
2542 cfq_var_store(unsigned int *var, const char *page, size_t count)
2544 char *p = (char *) page;
2546 *var = simple_strtoul(p, &p, 10);
2550 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
2551 static ssize_t __FUNC(struct elevator_queue *e, char *page) \
2553 struct cfq_data *cfqd = e->elevator_data; \
2554 unsigned int __data = __VAR; \
2556 __data = jiffies_to_msecs(__data); \
2557 return cfq_var_show(__data, (page)); \
2559 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
2560 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
2561 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
2562 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
2563 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
2564 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
2565 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
2566 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
2567 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
2568 SHOW_FUNCTION(cfq_desktop_show, cfqd->cfq_desktop, 0);
2569 #undef SHOW_FUNCTION
2571 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
2572 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
2574 struct cfq_data *cfqd = e->elevator_data; \
2575 unsigned int __data; \
2576 int ret = cfq_var_store(&__data, (page), count); \
2577 if (__data < (MIN)) \
2579 else if (__data > (MAX)) \
2582 *(__PTR) = msecs_to_jiffies(__data); \
2584 *(__PTR) = __data; \
2587 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
2588 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
2590 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
2592 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
2593 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
2595 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
2596 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
2597 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
2598 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
2600 STORE_FUNCTION(cfq_desktop_store, &cfqd->cfq_desktop, 0, 1, 0);
2601 #undef STORE_FUNCTION
2603 #define CFQ_ATTR(name) \
2604 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
2606 static struct elv_fs_entry cfq_attrs[] = {
2608 CFQ_ATTR(fifo_expire_sync),
2609 CFQ_ATTR(fifo_expire_async),
2610 CFQ_ATTR(back_seek_max),
2611 CFQ_ATTR(back_seek_penalty),
2612 CFQ_ATTR(slice_sync),
2613 CFQ_ATTR(slice_async),
2614 CFQ_ATTR(slice_async_rq),
2615 CFQ_ATTR(slice_idle),
2620 static struct elevator_type iosched_cfq = {
2622 .elevator_merge_fn = cfq_merge,
2623 .elevator_merged_fn = cfq_merged_request,
2624 .elevator_merge_req_fn = cfq_merged_requests,
2625 .elevator_allow_merge_fn = cfq_allow_merge,
2626 .elevator_dispatch_fn = cfq_dispatch_requests,
2627 .elevator_add_req_fn = cfq_insert_request,
2628 .elevator_activate_req_fn = cfq_activate_request,
2629 .elevator_deactivate_req_fn = cfq_deactivate_request,
2630 .elevator_queue_empty_fn = cfq_queue_empty,
2631 .elevator_completed_req_fn = cfq_completed_request,
2632 .elevator_former_req_fn = elv_rb_former_request,
2633 .elevator_latter_req_fn = elv_rb_latter_request,
2634 .elevator_set_req_fn = cfq_set_request,
2635 .elevator_put_req_fn = cfq_put_request,
2636 .elevator_may_queue_fn = cfq_may_queue,
2637 .elevator_init_fn = cfq_init_queue,
2638 .elevator_exit_fn = cfq_exit_queue,
2639 .trim = cfq_free_io_context,
2641 .elevator_attrs = cfq_attrs,
2642 .elevator_name = "cfq",
2643 .elevator_owner = THIS_MODULE,
2646 static int __init cfq_init(void)
2649 * could be 0 on HZ < 1000 setups
2651 if (!cfq_slice_async)
2652 cfq_slice_async = 1;
2653 if (!cfq_slice_idle)
2656 if (cfq_slab_setup())
2659 elv_register(&iosched_cfq);
2664 static void __exit cfq_exit(void)
2666 DECLARE_COMPLETION_ONSTACK(all_gone);
2667 elv_unregister(&iosched_cfq);
2668 ioc_gone = &all_gone;
2669 /* ioc_gone's update must be visible before reading ioc_count */
2673 * this also protects us from entering cfq_slab_kill() with
2674 * pending RCU callbacks
2676 if (elv_ioc_count_read(cfq_ioc_count))
2677 wait_for_completion(&all_gone);
2681 module_init(cfq_init);
2682 module_exit(cfq_exit);
2684 MODULE_AUTHOR("Jens Axboe");
2685 MODULE_LICENSE("GPL");
2686 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");