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 delayed_work 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_latency;
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,
274 if (cfqd->busy_queues) {
275 cfq_log(cfqd, "schedule dispatch");
276 kblockd_schedule_delayed_work(cfqd->queue, &cfqd->unplug_work,
281 static int cfq_queue_empty(struct request_queue *q)
283 struct cfq_data *cfqd = q->elevator->elevator_data;
285 return !cfqd->busy_queues;
289 * Scale schedule slice based on io priority. Use the sync time slice only
290 * if a queue is marked sync and has sync io queued. A sync queue with async
291 * io only, should not get full sync slice length.
293 static inline int cfq_prio_slice(struct cfq_data *cfqd, int sync,
296 const int base_slice = cfqd->cfq_slice[sync];
298 WARN_ON(prio >= IOPRIO_BE_NR);
300 return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
304 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
306 return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
310 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
312 cfqq->slice_end = cfq_prio_to_slice(cfqd, cfqq) + jiffies;
313 cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
317 * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
318 * isn't valid until the first request from the dispatch is activated
319 * and the slice time set.
321 static inline int cfq_slice_used(struct cfq_queue *cfqq)
323 if (cfq_cfqq_slice_new(cfqq))
325 if (time_before(jiffies, cfqq->slice_end))
332 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
333 * We choose the request that is closest to the head right now. Distance
334 * behind the head is penalized and only allowed to a certain extent.
336 static struct request *
337 cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2)
339 sector_t last, s1, s2, d1 = 0, d2 = 0;
340 unsigned long back_max;
341 #define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */
342 #define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */
343 unsigned wrap = 0; /* bit mask: requests behind the disk head? */
345 if (rq1 == NULL || rq1 == rq2)
350 if (rq_is_sync(rq1) && !rq_is_sync(rq2))
352 else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
354 if (rq_is_meta(rq1) && !rq_is_meta(rq2))
356 else if (rq_is_meta(rq2) && !rq_is_meta(rq1))
359 s1 = blk_rq_pos(rq1);
360 s2 = blk_rq_pos(rq2);
362 last = cfqd->last_position;
365 * by definition, 1KiB is 2 sectors
367 back_max = cfqd->cfq_back_max * 2;
370 * Strict one way elevator _except_ in the case where we allow
371 * short backward seeks which are biased as twice the cost of a
372 * similar forward seek.
376 else if (s1 + back_max >= last)
377 d1 = (last - s1) * cfqd->cfq_back_penalty;
379 wrap |= CFQ_RQ1_WRAP;
383 else if (s2 + back_max >= last)
384 d2 = (last - s2) * cfqd->cfq_back_penalty;
386 wrap |= CFQ_RQ2_WRAP;
388 /* Found required data */
391 * By doing switch() on the bit mask "wrap" we avoid having to
392 * check two variables for all permutations: --> faster!
395 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
411 case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
414 * Since both rqs are wrapped,
415 * start with the one that's further behind head
416 * (--> only *one* back seek required),
417 * since back seek takes more time than forward.
427 * The below is leftmost cache rbtree addon
429 static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
432 root->left = rb_first(&root->rb);
435 return rb_entry(root->left, struct cfq_queue, rb_node);
440 static void rb_erase_init(struct rb_node *n, struct rb_root *root)
446 static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
450 rb_erase_init(n, &root->rb);
454 * would be nice to take fifo expire time into account as well
456 static struct request *
457 cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
458 struct request *last)
460 struct rb_node *rbnext = rb_next(&last->rb_node);
461 struct rb_node *rbprev = rb_prev(&last->rb_node);
462 struct request *next = NULL, *prev = NULL;
464 BUG_ON(RB_EMPTY_NODE(&last->rb_node));
467 prev = rb_entry_rq(rbprev);
470 next = rb_entry_rq(rbnext);
472 rbnext = rb_first(&cfqq->sort_list);
473 if (rbnext && rbnext != &last->rb_node)
474 next = rb_entry_rq(rbnext);
477 return cfq_choose_req(cfqd, next, prev);
480 static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
481 struct cfq_queue *cfqq)
484 * just an approximation, should be ok.
486 return (cfqd->busy_queues - 1) * (cfq_prio_slice(cfqd, 1, 0) -
487 cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
491 * The cfqd->service_tree holds all pending cfq_queue's that have
492 * requests waiting to be processed. It is sorted in the order that
493 * we will service the queues.
495 static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
498 struct rb_node **p, *parent;
499 struct cfq_queue *__cfqq;
500 unsigned long rb_key;
503 if (cfq_class_idle(cfqq)) {
504 rb_key = CFQ_IDLE_DELAY;
505 parent = rb_last(&cfqd->service_tree.rb);
506 if (parent && parent != &cfqq->rb_node) {
507 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
508 rb_key += __cfqq->rb_key;
511 } else if (!add_front) {
512 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
513 rb_key += cfqq->slice_resid;
514 cfqq->slice_resid = 0;
518 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
520 * same position, nothing more to do
522 if (rb_key == cfqq->rb_key)
525 cfq_rb_erase(&cfqq->rb_node, &cfqd->service_tree);
530 p = &cfqd->service_tree.rb.rb_node;
535 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
538 * sort RT queues first, we always want to give
539 * preference to them. IDLE queues goes to the back.
540 * after that, sort on the next service time.
542 if (cfq_class_rt(cfqq) > cfq_class_rt(__cfqq))
544 else if (cfq_class_rt(cfqq) < cfq_class_rt(__cfqq))
546 else if (cfq_class_idle(cfqq) < cfq_class_idle(__cfqq))
548 else if (cfq_class_idle(cfqq) > cfq_class_idle(__cfqq))
550 else if (rb_key < __cfqq->rb_key)
555 if (n == &(*p)->rb_right)
562 cfqd->service_tree.left = &cfqq->rb_node;
564 cfqq->rb_key = rb_key;
565 rb_link_node(&cfqq->rb_node, parent, p);
566 rb_insert_color(&cfqq->rb_node, &cfqd->service_tree.rb);
569 static struct cfq_queue *
570 cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
571 sector_t sector, struct rb_node **ret_parent,
572 struct rb_node ***rb_link)
574 struct rb_node **p, *parent;
575 struct cfq_queue *cfqq = NULL;
583 cfqq = rb_entry(parent, struct cfq_queue, p_node);
586 * Sort strictly based on sector. Smallest to the left,
587 * largest to the right.
589 if (sector > blk_rq_pos(cfqq->next_rq))
591 else if (sector < blk_rq_pos(cfqq->next_rq))
599 *ret_parent = parent;
605 static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
607 struct rb_node **p, *parent;
608 struct cfq_queue *__cfqq;
611 rb_erase(&cfqq->p_node, cfqq->p_root);
615 if (cfq_class_idle(cfqq))
620 cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
621 __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
622 blk_rq_pos(cfqq->next_rq), &parent, &p);
624 rb_link_node(&cfqq->p_node, parent, p);
625 rb_insert_color(&cfqq->p_node, cfqq->p_root);
631 * Update cfqq's position in the service tree.
633 static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
636 * Resorting requires the cfqq to be on the RR list already.
638 if (cfq_cfqq_on_rr(cfqq)) {
639 cfq_service_tree_add(cfqd, cfqq, 0);
640 cfq_prio_tree_add(cfqd, cfqq);
645 * add to busy list of queues for service, trying to be fair in ordering
646 * the pending list according to last request service
648 static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
650 cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
651 BUG_ON(cfq_cfqq_on_rr(cfqq));
652 cfq_mark_cfqq_on_rr(cfqq);
655 cfq_resort_rr_list(cfqd, cfqq);
659 * Called when the cfqq no longer has requests pending, remove it from
662 static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
664 cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
665 BUG_ON(!cfq_cfqq_on_rr(cfqq));
666 cfq_clear_cfqq_on_rr(cfqq);
668 if (!RB_EMPTY_NODE(&cfqq->rb_node))
669 cfq_rb_erase(&cfqq->rb_node, &cfqd->service_tree);
671 rb_erase(&cfqq->p_node, cfqq->p_root);
675 BUG_ON(!cfqd->busy_queues);
680 * rb tree support functions
682 static void cfq_del_rq_rb(struct request *rq)
684 struct cfq_queue *cfqq = RQ_CFQQ(rq);
685 struct cfq_data *cfqd = cfqq->cfqd;
686 const int sync = rq_is_sync(rq);
688 BUG_ON(!cfqq->queued[sync]);
689 cfqq->queued[sync]--;
691 elv_rb_del(&cfqq->sort_list, rq);
693 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
694 cfq_del_cfqq_rr(cfqd, cfqq);
697 static void cfq_add_rq_rb(struct request *rq)
699 struct cfq_queue *cfqq = RQ_CFQQ(rq);
700 struct cfq_data *cfqd = cfqq->cfqd;
701 struct request *__alias, *prev;
703 cfqq->queued[rq_is_sync(rq)]++;
706 * looks a little odd, but the first insert might return an alias.
707 * if that happens, put the alias on the dispatch list
709 while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL)
710 cfq_dispatch_insert(cfqd->queue, __alias);
712 if (!cfq_cfqq_on_rr(cfqq))
713 cfq_add_cfqq_rr(cfqd, cfqq);
716 * check if this request is a better next-serve candidate
718 prev = cfqq->next_rq;
719 cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq);
722 * adjust priority tree position, if ->next_rq changes
724 if (prev != cfqq->next_rq)
725 cfq_prio_tree_add(cfqd, cfqq);
727 BUG_ON(!cfqq->next_rq);
730 static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
732 elv_rb_del(&cfqq->sort_list, rq);
733 cfqq->queued[rq_is_sync(rq)]--;
737 static struct request *
738 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
740 struct task_struct *tsk = current;
741 struct cfq_io_context *cic;
742 struct cfq_queue *cfqq;
744 cic = cfq_cic_lookup(cfqd, tsk->io_context);
748 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
750 sector_t sector = bio->bi_sector + bio_sectors(bio);
752 return elv_rb_find(&cfqq->sort_list, sector);
758 static void cfq_activate_request(struct request_queue *q, struct request *rq)
760 struct cfq_data *cfqd = q->elevator->elevator_data;
762 cfqd->rq_in_driver[rq_is_sync(rq)]++;
763 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
766 cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
769 static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
771 struct cfq_data *cfqd = q->elevator->elevator_data;
772 const int sync = rq_is_sync(rq);
774 WARN_ON(!cfqd->rq_in_driver[sync]);
775 cfqd->rq_in_driver[sync]--;
776 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
780 static void cfq_remove_request(struct request *rq)
782 struct cfq_queue *cfqq = RQ_CFQQ(rq);
784 if (cfqq->next_rq == rq)
785 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
787 list_del_init(&rq->queuelist);
790 cfqq->cfqd->rq_queued--;
791 if (rq_is_meta(rq)) {
792 WARN_ON(!cfqq->meta_pending);
793 cfqq->meta_pending--;
797 static int cfq_merge(struct request_queue *q, struct request **req,
800 struct cfq_data *cfqd = q->elevator->elevator_data;
801 struct request *__rq;
803 __rq = cfq_find_rq_fmerge(cfqd, bio);
804 if (__rq && elv_rq_merge_ok(__rq, bio)) {
806 return ELEVATOR_FRONT_MERGE;
809 return ELEVATOR_NO_MERGE;
812 static void cfq_merged_request(struct request_queue *q, struct request *req,
815 if (type == ELEVATOR_FRONT_MERGE) {
816 struct cfq_queue *cfqq = RQ_CFQQ(req);
818 cfq_reposition_rq_rb(cfqq, req);
823 cfq_merged_requests(struct request_queue *q, struct request *rq,
824 struct request *next)
827 * reposition in fifo if next is older than rq
829 if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
830 time_before(rq_fifo_time(next), rq_fifo_time(rq))) {
831 list_move(&rq->queuelist, &next->queuelist);
832 rq_set_fifo_time(rq, rq_fifo_time(next));
835 cfq_remove_request(next);
838 static int cfq_allow_merge(struct request_queue *q, struct request *rq,
841 struct cfq_data *cfqd = q->elevator->elevator_data;
842 struct cfq_io_context *cic;
843 struct cfq_queue *cfqq;
846 * Disallow merge of a sync bio into an async request.
848 if (cfq_bio_sync(bio) && !rq_is_sync(rq))
852 * Lookup the cfqq that this bio will be queued with. Allow
853 * merge only if rq is queued there.
855 cic = cfq_cic_lookup(cfqd, current->io_context);
859 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
860 if (cfqq == RQ_CFQQ(rq))
866 static void __cfq_set_active_queue(struct cfq_data *cfqd,
867 struct cfq_queue *cfqq)
870 cfq_log_cfqq(cfqd, cfqq, "set_active");
872 cfqq->slice_dispatch = 0;
874 cfq_clear_cfqq_wait_request(cfqq);
875 cfq_clear_cfqq_must_dispatch(cfqq);
876 cfq_clear_cfqq_must_alloc_slice(cfqq);
877 cfq_clear_cfqq_fifo_expire(cfqq);
878 cfq_mark_cfqq_slice_new(cfqq);
880 del_timer(&cfqd->idle_slice_timer);
883 cfqd->active_queue = cfqq;
887 * current cfqq expired its slice (or was too idle), select new one
890 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
893 cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
895 if (cfq_cfqq_wait_request(cfqq))
896 del_timer(&cfqd->idle_slice_timer);
898 cfq_clear_cfqq_wait_request(cfqq);
901 * store what was left of this slice, if the queue idled/timed out
903 if (timed_out && !cfq_cfqq_slice_new(cfqq)) {
904 cfqq->slice_resid = cfqq->slice_end - jiffies;
905 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
908 cfq_resort_rr_list(cfqd, cfqq);
910 if (cfqq == cfqd->active_queue)
911 cfqd->active_queue = NULL;
913 if (cfqd->active_cic) {
914 put_io_context(cfqd->active_cic->ioc);
915 cfqd->active_cic = NULL;
919 static inline void cfq_slice_expired(struct cfq_data *cfqd, int timed_out)
921 struct cfq_queue *cfqq = cfqd->active_queue;
924 __cfq_slice_expired(cfqd, cfqq, timed_out);
928 * Get next queue for service. Unless we have a queue preemption,
929 * we'll simply select the first cfqq in the service tree.
931 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
933 if (RB_EMPTY_ROOT(&cfqd->service_tree.rb))
936 return cfq_rb_first(&cfqd->service_tree);
940 * Get and set a new active queue for service.
942 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
943 struct cfq_queue *cfqq)
946 cfqq = cfq_get_next_queue(cfqd);
948 cfq_clear_cfqq_coop(cfqq);
951 __cfq_set_active_queue(cfqd, cfqq);
955 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
958 if (blk_rq_pos(rq) >= cfqd->last_position)
959 return blk_rq_pos(rq) - cfqd->last_position;
961 return cfqd->last_position - blk_rq_pos(rq);
964 #define CIC_SEEK_THR 8 * 1024
965 #define CIC_SEEKY(cic) ((cic)->seek_mean > CIC_SEEK_THR)
967 static inline int cfq_rq_close(struct cfq_data *cfqd, struct request *rq)
969 struct cfq_io_context *cic = cfqd->active_cic;
970 sector_t sdist = cic->seek_mean;
972 if (!sample_valid(cic->seek_samples))
973 sdist = CIC_SEEK_THR;
975 return cfq_dist_from_last(cfqd, rq) <= sdist;
978 static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
979 struct cfq_queue *cur_cfqq)
981 struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
982 struct rb_node *parent, *node;
983 struct cfq_queue *__cfqq;
984 sector_t sector = cfqd->last_position;
986 if (RB_EMPTY_ROOT(root))
990 * First, if we find a request starting at the end of the last
991 * request, choose it.
993 __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
998 * If the exact sector wasn't found, the parent of the NULL leaf
999 * will contain the closest sector.
1001 __cfqq = rb_entry(parent, struct cfq_queue, p_node);
1002 if (cfq_rq_close(cfqd, __cfqq->next_rq))
1005 if (blk_rq_pos(__cfqq->next_rq) < sector)
1006 node = rb_next(&__cfqq->p_node);
1008 node = rb_prev(&__cfqq->p_node);
1012 __cfqq = rb_entry(node, struct cfq_queue, p_node);
1013 if (cfq_rq_close(cfqd, __cfqq->next_rq))
1021 * cur_cfqq - passed in so that we don't decide that the current queue is
1022 * closely cooperating with itself.
1024 * So, basically we're assuming that that cur_cfqq has dispatched at least
1025 * one request, and that cfqd->last_position reflects a position on the disk
1026 * associated with the I/O issued by cur_cfqq. I'm not sure this is a valid
1029 static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
1030 struct cfq_queue *cur_cfqq,
1033 struct cfq_queue *cfqq;
1036 * A valid cfq_io_context is necessary to compare requests against
1037 * the seek_mean of the current cfqq.
1039 if (!cfqd->active_cic)
1043 * We should notice if some of the queues are cooperating, eg
1044 * working closely on the same area of the disk. In that case,
1045 * we can group them together and don't waste time idling.
1047 cfqq = cfqq_close(cfqd, cur_cfqq);
1051 if (cfq_cfqq_coop(cfqq))
1055 cfq_mark_cfqq_coop(cfqq);
1059 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
1061 struct cfq_queue *cfqq = cfqd->active_queue;
1062 struct cfq_io_context *cic;
1066 * SSD device without seek penalty, disable idling. But only do so
1067 * for devices that support queuing, otherwise we still have a problem
1068 * with sync vs async workloads.
1070 if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
1073 WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
1074 WARN_ON(cfq_cfqq_slice_new(cfqq));
1077 * idle is disabled, either manually or by past process history
1079 if (!cfqd->cfq_slice_idle || !cfq_cfqq_idle_window(cfqq))
1083 * still requests with the driver, don't idle
1085 if (rq_in_driver(cfqd))
1089 * task has exited, don't wait
1091 cic = cfqd->active_cic;
1092 if (!cic || !atomic_read(&cic->ioc->nr_tasks))
1095 cfq_mark_cfqq_wait_request(cfqq);
1098 * we don't want to idle for seeks, but we do want to allow
1099 * fair distribution of slice time for a process doing back-to-back
1100 * seeks. so allow a little bit of time for him to submit a new rq
1102 sl = cfqd->cfq_slice_idle;
1103 if (sample_valid(cic->seek_samples) && CIC_SEEKY(cic))
1104 sl = min(sl, msecs_to_jiffies(CFQ_MIN_TT));
1106 mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
1107 cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu", sl);
1111 * Move request from internal lists to the request queue dispatch list.
1113 static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
1115 struct cfq_data *cfqd = q->elevator->elevator_data;
1116 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1118 cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
1120 cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
1121 cfq_remove_request(rq);
1123 elv_dispatch_sort(q, rq);
1125 if (cfq_cfqq_sync(cfqq))
1126 cfqd->sync_flight++;
1130 * return expired entry, or NULL to just start from scratch in rbtree
1132 static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
1134 struct request *rq = NULL;
1136 if (cfq_cfqq_fifo_expire(cfqq))
1139 cfq_mark_cfqq_fifo_expire(cfqq);
1141 if (list_empty(&cfqq->fifo))
1144 rq = rq_entry_fifo(cfqq->fifo.next);
1145 if (time_before(jiffies, rq_fifo_time(rq)))
1148 cfq_log_cfqq(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) {
1320 * idle queue must always only have a single IO in flight
1322 if (cfq_class_idle(cfqq))
1326 * We have other queues, don't allow more IO from this one
1328 if (cfqd->busy_queues > 1)
1332 * Sole queue user, allow bigger slice
1338 * Async queues must wait a bit before being allowed dispatch.
1339 * We also ramp up the dispatch depth gradually for async IO,
1340 * based on the last sync IO we serviced
1342 if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
1343 unsigned long last_sync = jiffies - cfqd->last_end_sync_rq;
1346 depth = last_sync / cfqd->cfq_slice[1];
1347 if (!depth && !cfqq->dispatched)
1349 if (depth < max_dispatch)
1350 max_dispatch = depth;
1353 if (cfqq->dispatched >= max_dispatch)
1357 * Dispatch a request from this cfqq
1359 cfq_dispatch_request(cfqd, cfqq);
1360 cfqq->slice_dispatch++;
1361 cfq_clear_cfqq_must_dispatch(cfqq);
1364 * expire an async queue immediately if it has used up its slice. idle
1365 * queue always expire after 1 dispatch round.
1367 if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
1368 cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
1369 cfq_class_idle(cfqq))) {
1370 cfqq->slice_end = jiffies + 1;
1371 cfq_slice_expired(cfqd, 0);
1374 cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
1379 * task holds one reference to the queue, dropped when task exits. each rq
1380 * in-flight on this queue also holds a reference, dropped when rq is freed.
1382 * queue lock must be held here.
1384 static void cfq_put_queue(struct cfq_queue *cfqq)
1386 struct cfq_data *cfqd = cfqq->cfqd;
1388 BUG_ON(atomic_read(&cfqq->ref) <= 0);
1390 if (!atomic_dec_and_test(&cfqq->ref))
1393 cfq_log_cfqq(cfqd, cfqq, "put_queue");
1394 BUG_ON(rb_first(&cfqq->sort_list));
1395 BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
1396 BUG_ON(cfq_cfqq_on_rr(cfqq));
1398 if (unlikely(cfqd->active_queue == cfqq)) {
1399 __cfq_slice_expired(cfqd, cfqq, 0);
1400 cfq_schedule_dispatch(cfqd, 0);
1403 kmem_cache_free(cfq_pool, cfqq);
1407 * Must always be called with the rcu_read_lock() held
1410 __call_for_each_cic(struct io_context *ioc,
1411 void (*func)(struct io_context *, struct cfq_io_context *))
1413 struct cfq_io_context *cic;
1414 struct hlist_node *n;
1416 hlist_for_each_entry_rcu(cic, n, &ioc->cic_list, cic_list)
1421 * Call func for each cic attached to this ioc.
1424 call_for_each_cic(struct io_context *ioc,
1425 void (*func)(struct io_context *, struct cfq_io_context *))
1428 __call_for_each_cic(ioc, func);
1432 static void cfq_cic_free_rcu(struct rcu_head *head)
1434 struct cfq_io_context *cic;
1436 cic = container_of(head, struct cfq_io_context, rcu_head);
1438 kmem_cache_free(cfq_ioc_pool, cic);
1439 elv_ioc_count_dec(cfq_ioc_count);
1443 * CFQ scheduler is exiting, grab exit lock and check
1444 * the pending io context count. If it hits zero,
1445 * complete ioc_gone and set it back to NULL
1447 spin_lock(&ioc_gone_lock);
1448 if (ioc_gone && !elv_ioc_count_read(cfq_ioc_count)) {
1452 spin_unlock(&ioc_gone_lock);
1456 static void cfq_cic_free(struct cfq_io_context *cic)
1458 call_rcu(&cic->rcu_head, cfq_cic_free_rcu);
1461 static void cic_free_func(struct io_context *ioc, struct cfq_io_context *cic)
1463 unsigned long flags;
1465 BUG_ON(!cic->dead_key);
1467 spin_lock_irqsave(&ioc->lock, flags);
1468 radix_tree_delete(&ioc->radix_root, cic->dead_key);
1469 hlist_del_rcu(&cic->cic_list);
1470 spin_unlock_irqrestore(&ioc->lock, flags);
1476 * Must be called with rcu_read_lock() held or preemption otherwise disabled.
1477 * Only two callers of this - ->dtor() which is called with the rcu_read_lock(),
1478 * and ->trim() which is called with the task lock held
1480 static void cfq_free_io_context(struct io_context *ioc)
1483 * ioc->refcount is zero here, or we are called from elv_unregister(),
1484 * so no more cic's are allowed to be linked into this ioc. So it
1485 * should be ok to iterate over the known list, we will see all cic's
1486 * since no new ones are added.
1488 __call_for_each_cic(ioc, cic_free_func);
1491 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1493 if (unlikely(cfqq == cfqd->active_queue)) {
1494 __cfq_slice_expired(cfqd, cfqq, 0);
1495 cfq_schedule_dispatch(cfqd, 0);
1498 cfq_put_queue(cfqq);
1501 static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
1502 struct cfq_io_context *cic)
1504 struct io_context *ioc = cic->ioc;
1506 list_del_init(&cic->queue_list);
1509 * Make sure key == NULL is seen for dead queues
1512 cic->dead_key = (unsigned long) cic->key;
1515 if (ioc->ioc_data == cic)
1516 rcu_assign_pointer(ioc->ioc_data, NULL);
1518 if (cic->cfqq[BLK_RW_ASYNC]) {
1519 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]);
1520 cic->cfqq[BLK_RW_ASYNC] = NULL;
1523 if (cic->cfqq[BLK_RW_SYNC]) {
1524 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]);
1525 cic->cfqq[BLK_RW_SYNC] = NULL;
1529 static void cfq_exit_single_io_context(struct io_context *ioc,
1530 struct cfq_io_context *cic)
1532 struct cfq_data *cfqd = cic->key;
1535 struct request_queue *q = cfqd->queue;
1536 unsigned long flags;
1538 spin_lock_irqsave(q->queue_lock, flags);
1541 * Ensure we get a fresh copy of the ->key to prevent
1542 * race between exiting task and queue
1544 smp_read_barrier_depends();
1546 __cfq_exit_single_io_context(cfqd, cic);
1548 spin_unlock_irqrestore(q->queue_lock, flags);
1553 * The process that ioc belongs to has exited, we need to clean up
1554 * and put the internal structures we have that belongs to that process.
1556 static void cfq_exit_io_context(struct io_context *ioc)
1558 call_for_each_cic(ioc, cfq_exit_single_io_context);
1561 static struct cfq_io_context *
1562 cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
1564 struct cfq_io_context *cic;
1566 cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO,
1569 cic->last_end_request = jiffies;
1570 INIT_LIST_HEAD(&cic->queue_list);
1571 INIT_HLIST_NODE(&cic->cic_list);
1572 cic->dtor = cfq_free_io_context;
1573 cic->exit = cfq_exit_io_context;
1574 elv_ioc_count_inc(cfq_ioc_count);
1580 static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc)
1582 struct task_struct *tsk = current;
1585 if (!cfq_cfqq_prio_changed(cfqq))
1588 ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio);
1589 switch (ioprio_class) {
1591 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
1592 case IOPRIO_CLASS_NONE:
1594 * no prio set, inherit CPU scheduling settings
1596 cfqq->ioprio = task_nice_ioprio(tsk);
1597 cfqq->ioprio_class = task_nice_ioclass(tsk);
1599 case IOPRIO_CLASS_RT:
1600 cfqq->ioprio = task_ioprio(ioc);
1601 cfqq->ioprio_class = IOPRIO_CLASS_RT;
1603 case IOPRIO_CLASS_BE:
1604 cfqq->ioprio = task_ioprio(ioc);
1605 cfqq->ioprio_class = IOPRIO_CLASS_BE;
1607 case IOPRIO_CLASS_IDLE:
1608 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
1610 cfq_clear_cfqq_idle_window(cfqq);
1615 * keep track of original prio settings in case we have to temporarily
1616 * elevate the priority of this queue
1618 cfqq->org_ioprio = cfqq->ioprio;
1619 cfqq->org_ioprio_class = cfqq->ioprio_class;
1620 cfq_clear_cfqq_prio_changed(cfqq);
1623 static void changed_ioprio(struct io_context *ioc, struct cfq_io_context *cic)
1625 struct cfq_data *cfqd = cic->key;
1626 struct cfq_queue *cfqq;
1627 unsigned long flags;
1629 if (unlikely(!cfqd))
1632 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1634 cfqq = cic->cfqq[BLK_RW_ASYNC];
1636 struct cfq_queue *new_cfqq;
1637 new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic->ioc,
1640 cic->cfqq[BLK_RW_ASYNC] = new_cfqq;
1641 cfq_put_queue(cfqq);
1645 cfqq = cic->cfqq[BLK_RW_SYNC];
1647 cfq_mark_cfqq_prio_changed(cfqq);
1649 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1652 static void cfq_ioc_set_ioprio(struct io_context *ioc)
1654 call_for_each_cic(ioc, changed_ioprio);
1655 ioc->ioprio_changed = 0;
1658 static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1659 pid_t pid, int is_sync)
1661 RB_CLEAR_NODE(&cfqq->rb_node);
1662 RB_CLEAR_NODE(&cfqq->p_node);
1663 INIT_LIST_HEAD(&cfqq->fifo);
1665 atomic_set(&cfqq->ref, 0);
1668 cfq_mark_cfqq_prio_changed(cfqq);
1671 if (!cfq_class_idle(cfqq))
1672 cfq_mark_cfqq_idle_window(cfqq);
1673 cfq_mark_cfqq_sync(cfqq);
1678 static struct cfq_queue *
1679 cfq_find_alloc_queue(struct cfq_data *cfqd, int is_sync,
1680 struct io_context *ioc, gfp_t gfp_mask)
1682 struct cfq_queue *cfqq, *new_cfqq = NULL;
1683 struct cfq_io_context *cic;
1686 cic = cfq_cic_lookup(cfqd, ioc);
1687 /* cic always exists here */
1688 cfqq = cic_to_cfqq(cic, is_sync);
1691 * Always try a new alloc if we fell back to the OOM cfqq
1692 * originally, since it should just be a temporary situation.
1694 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
1699 } else if (gfp_mask & __GFP_WAIT) {
1700 spin_unlock_irq(cfqd->queue->queue_lock);
1701 new_cfqq = kmem_cache_alloc_node(cfq_pool,
1702 gfp_mask | __GFP_ZERO,
1704 spin_lock_irq(cfqd->queue->queue_lock);
1708 cfqq = kmem_cache_alloc_node(cfq_pool,
1709 gfp_mask | __GFP_ZERO,
1714 cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
1715 cfq_init_prio_data(cfqq, ioc);
1716 cfq_log_cfqq(cfqd, cfqq, "alloced");
1718 cfqq = &cfqd->oom_cfqq;
1722 kmem_cache_free(cfq_pool, new_cfqq);
1727 static struct cfq_queue **
1728 cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
1730 switch (ioprio_class) {
1731 case IOPRIO_CLASS_RT:
1732 return &cfqd->async_cfqq[0][ioprio];
1733 case IOPRIO_CLASS_BE:
1734 return &cfqd->async_cfqq[1][ioprio];
1735 case IOPRIO_CLASS_IDLE:
1736 return &cfqd->async_idle_cfqq;
1742 static struct cfq_queue *
1743 cfq_get_queue(struct cfq_data *cfqd, int is_sync, struct io_context *ioc,
1746 const int ioprio = task_ioprio(ioc);
1747 const int ioprio_class = task_ioprio_class(ioc);
1748 struct cfq_queue **async_cfqq = NULL;
1749 struct cfq_queue *cfqq = NULL;
1752 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
1757 cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask);
1760 * pin the queue now that it's allocated, scheduler exit will prune it
1762 if (!is_sync && !(*async_cfqq)) {
1763 atomic_inc(&cfqq->ref);
1767 atomic_inc(&cfqq->ref);
1772 * We drop cfq io contexts lazily, so we may find a dead one.
1775 cfq_drop_dead_cic(struct cfq_data *cfqd, struct io_context *ioc,
1776 struct cfq_io_context *cic)
1778 unsigned long flags;
1780 WARN_ON(!list_empty(&cic->queue_list));
1782 spin_lock_irqsave(&ioc->lock, flags);
1784 BUG_ON(ioc->ioc_data == cic);
1786 radix_tree_delete(&ioc->radix_root, (unsigned long) cfqd);
1787 hlist_del_rcu(&cic->cic_list);
1788 spin_unlock_irqrestore(&ioc->lock, flags);
1793 static struct cfq_io_context *
1794 cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc)
1796 struct cfq_io_context *cic;
1797 unsigned long flags;
1806 * we maintain a last-hit cache, to avoid browsing over the tree
1808 cic = rcu_dereference(ioc->ioc_data);
1809 if (cic && cic->key == cfqd) {
1815 cic = radix_tree_lookup(&ioc->radix_root, (unsigned long) cfqd);
1819 /* ->key must be copied to avoid race with cfq_exit_queue() */
1822 cfq_drop_dead_cic(cfqd, ioc, cic);
1827 spin_lock_irqsave(&ioc->lock, flags);
1828 rcu_assign_pointer(ioc->ioc_data, cic);
1829 spin_unlock_irqrestore(&ioc->lock, flags);
1837 * Add cic into ioc, using cfqd as the search key. This enables us to lookup
1838 * the process specific cfq io context when entered from the block layer.
1839 * Also adds the cic to a per-cfqd list, used when this queue is removed.
1841 static int cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
1842 struct cfq_io_context *cic, gfp_t gfp_mask)
1844 unsigned long flags;
1847 ret = radix_tree_preload(gfp_mask);
1852 spin_lock_irqsave(&ioc->lock, flags);
1853 ret = radix_tree_insert(&ioc->radix_root,
1854 (unsigned long) cfqd, cic);
1856 hlist_add_head_rcu(&cic->cic_list, &ioc->cic_list);
1857 spin_unlock_irqrestore(&ioc->lock, flags);
1859 radix_tree_preload_end();
1862 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1863 list_add(&cic->queue_list, &cfqd->cic_list);
1864 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1869 printk(KERN_ERR "cfq: cic link failed!\n");
1875 * Setup general io context and cfq io context. There can be several cfq
1876 * io contexts per general io context, if this process is doing io to more
1877 * than one device managed by cfq.
1879 static struct cfq_io_context *
1880 cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
1882 struct io_context *ioc = NULL;
1883 struct cfq_io_context *cic;
1885 might_sleep_if(gfp_mask & __GFP_WAIT);
1887 ioc = get_io_context(gfp_mask, cfqd->queue->node);
1891 cic = cfq_cic_lookup(cfqd, ioc);
1895 cic = cfq_alloc_io_context(cfqd, gfp_mask);
1899 if (cfq_cic_link(cfqd, ioc, cic, gfp_mask))
1903 smp_read_barrier_depends();
1904 if (unlikely(ioc->ioprio_changed))
1905 cfq_ioc_set_ioprio(ioc);
1911 put_io_context(ioc);
1916 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
1918 unsigned long elapsed = jiffies - cic->last_end_request;
1919 unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
1921 cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
1922 cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
1923 cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
1927 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_io_context *cic,
1933 if (!cic->last_request_pos)
1935 else if (cic->last_request_pos < blk_rq_pos(rq))
1936 sdist = blk_rq_pos(rq) - cic->last_request_pos;
1938 sdist = cic->last_request_pos - blk_rq_pos(rq);
1941 * Don't allow the seek distance to get too large from the
1942 * odd fragment, pagein, etc
1944 if (cic->seek_samples <= 60) /* second&third seek */
1945 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*1024);
1947 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*64);
1949 cic->seek_samples = (7*cic->seek_samples + 256) / 8;
1950 cic->seek_total = (7*cic->seek_total + (u64)256*sdist) / 8;
1951 total = cic->seek_total + (cic->seek_samples/2);
1952 do_div(total, cic->seek_samples);
1953 cic->seek_mean = (sector_t)total;
1957 * Disable idle window if the process thinks too long or seeks so much that
1961 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1962 struct cfq_io_context *cic)
1964 int old_idle, enable_idle;
1967 * Don't idle for async or idle io prio class
1969 if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
1972 enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
1974 if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle ||
1975 (!cfqd->cfq_latency && cfqd->hw_tag && CIC_SEEKY(cic)))
1977 else if (sample_valid(cic->ttime_samples)) {
1978 if (cic->ttime_mean > cfqd->cfq_slice_idle)
1984 if (old_idle != enable_idle) {
1985 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
1987 cfq_mark_cfqq_idle_window(cfqq);
1989 cfq_clear_cfqq_idle_window(cfqq);
1994 * Check if new_cfqq should preempt the currently active queue. Return 0 for
1995 * no or if we aren't sure, a 1 will cause a preempt.
1998 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
2001 struct cfq_queue *cfqq;
2003 cfqq = cfqd->active_queue;
2007 if (cfq_slice_used(cfqq))
2010 if (cfq_class_idle(new_cfqq))
2013 if (cfq_class_idle(cfqq))
2017 * if the new request is sync, but the currently running queue is
2018 * not, let the sync request have priority.
2020 if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
2024 * So both queues are sync. Let the new request get disk time if
2025 * it's a metadata request and the current queue is doing regular IO.
2027 if (rq_is_meta(rq) && !cfqq->meta_pending)
2031 * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
2033 if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
2036 if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
2040 * if this request is as-good as one we would expect from the
2041 * current cfqq, let it preempt
2043 if (cfq_rq_close(cfqd, rq))
2050 * cfqq preempts the active queue. if we allowed preempt with no slice left,
2051 * let it have half of its nominal slice.
2053 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2055 cfq_log_cfqq(cfqd, cfqq, "preempt");
2056 cfq_slice_expired(cfqd, 1);
2059 * Put the new queue at the front of the of the current list,
2060 * so we know that it will be selected next.
2062 BUG_ON(!cfq_cfqq_on_rr(cfqq));
2064 cfq_service_tree_add(cfqd, cfqq, 1);
2066 cfqq->slice_end = 0;
2067 cfq_mark_cfqq_slice_new(cfqq);
2071 * Called when a new fs request (rq) is added (to cfqq). Check if there's
2072 * something we should do about it
2075 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2078 struct cfq_io_context *cic = RQ_CIC(rq);
2082 cfqq->meta_pending++;
2084 cfq_update_io_thinktime(cfqd, cic);
2085 cfq_update_io_seektime(cfqd, cic, rq);
2086 cfq_update_idle_window(cfqd, cfqq, cic);
2088 cic->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
2090 if (cfqq == cfqd->active_queue) {
2092 * Remember that we saw a request from this process, but
2093 * don't start queuing just yet. Otherwise we risk seeing lots
2094 * of tiny requests, because we disrupt the normal plugging
2095 * and merging. If the request is already larger than a single
2096 * page, let it rip immediately. For that case we assume that
2097 * merging is already done. Ditto for a busy system that
2098 * has other work pending, don't risk delaying until the
2099 * idle timer unplug to continue working.
2101 if (cfq_cfqq_wait_request(cfqq)) {
2102 if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
2103 cfqd->busy_queues > 1) {
2104 del_timer(&cfqd->idle_slice_timer);
2105 __blk_run_queue(cfqd->queue);
2107 cfq_mark_cfqq_must_dispatch(cfqq);
2109 } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
2111 * not the active queue - expire current slice if it is
2112 * idle and has expired it's mean thinktime or this new queue
2113 * has some old slice time left and is of higher priority or
2114 * this new queue is RT and the current one is BE
2116 cfq_preempt_queue(cfqd, cfqq);
2117 __blk_run_queue(cfqd->queue);
2121 static void cfq_insert_request(struct request_queue *q, struct request *rq)
2123 struct cfq_data *cfqd = q->elevator->elevator_data;
2124 struct cfq_queue *cfqq = RQ_CFQQ(rq);
2126 cfq_log_cfqq(cfqd, cfqq, "insert_request");
2127 cfq_init_prio_data(cfqq, RQ_CIC(rq)->ioc);
2131 rq_set_fifo_time(rq, jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)]);
2132 list_add_tail(&rq->queuelist, &cfqq->fifo);
2134 cfq_rq_enqueued(cfqd, cfqq, rq);
2138 * Update hw_tag based on peak queue depth over 50 samples under
2141 static void cfq_update_hw_tag(struct cfq_data *cfqd)
2143 if (rq_in_driver(cfqd) > cfqd->rq_in_driver_peak)
2144 cfqd->rq_in_driver_peak = rq_in_driver(cfqd);
2146 if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
2147 rq_in_driver(cfqd) <= CFQ_HW_QUEUE_MIN)
2150 if (cfqd->hw_tag_samples++ < 50)
2153 if (cfqd->rq_in_driver_peak >= CFQ_HW_QUEUE_MIN)
2158 cfqd->hw_tag_samples = 0;
2159 cfqd->rq_in_driver_peak = 0;
2162 static void cfq_completed_request(struct request_queue *q, struct request *rq)
2164 struct cfq_queue *cfqq = RQ_CFQQ(rq);
2165 struct cfq_data *cfqd = cfqq->cfqd;
2166 const int sync = rq_is_sync(rq);
2170 cfq_log_cfqq(cfqd, cfqq, "complete");
2172 cfq_update_hw_tag(cfqd);
2174 WARN_ON(!cfqd->rq_in_driver[sync]);
2175 WARN_ON(!cfqq->dispatched);
2176 cfqd->rq_in_driver[sync]--;
2179 if (cfq_cfqq_sync(cfqq))
2180 cfqd->sync_flight--;
2183 RQ_CIC(rq)->last_end_request = now;
2184 cfqd->last_end_sync_rq = now;
2188 * If this is the active queue, check if it needs to be expired,
2189 * or if we want to idle in case it has no pending requests.
2191 if (cfqd->active_queue == cfqq) {
2192 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
2194 if (cfq_cfqq_slice_new(cfqq)) {
2195 cfq_set_prio_slice(cfqd, cfqq);
2196 cfq_clear_cfqq_slice_new(cfqq);
2199 * If there are no requests waiting in this queue, and
2200 * there are other queues ready to issue requests, AND
2201 * those other queues are issuing requests within our
2202 * mean seek distance, give them a chance to run instead
2205 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
2206 cfq_slice_expired(cfqd, 1);
2207 else if (cfqq_empty && !cfq_close_cooperator(cfqd, cfqq, 1) &&
2208 sync && !rq_noidle(rq))
2209 cfq_arm_slice_timer(cfqd);
2212 if (!rq_in_driver(cfqd))
2213 cfq_schedule_dispatch(cfqd, 0);
2217 * we temporarily boost lower priority queues if they are holding fs exclusive
2218 * resources. they are boosted to normal prio (CLASS_BE/4)
2220 static void cfq_prio_boost(struct cfq_queue *cfqq)
2222 if (has_fs_excl()) {
2224 * boost idle prio on transactions that would lock out other
2225 * users of the filesystem
2227 if (cfq_class_idle(cfqq))
2228 cfqq->ioprio_class = IOPRIO_CLASS_BE;
2229 if (cfqq->ioprio > IOPRIO_NORM)
2230 cfqq->ioprio = IOPRIO_NORM;
2233 * check if we need to unboost the queue
2235 if (cfqq->ioprio_class != cfqq->org_ioprio_class)
2236 cfqq->ioprio_class = cfqq->org_ioprio_class;
2237 if (cfqq->ioprio != cfqq->org_ioprio)
2238 cfqq->ioprio = cfqq->org_ioprio;
2242 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
2244 if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
2245 cfq_mark_cfqq_must_alloc_slice(cfqq);
2246 return ELV_MQUEUE_MUST;
2249 return ELV_MQUEUE_MAY;
2252 static int cfq_may_queue(struct request_queue *q, int rw)
2254 struct cfq_data *cfqd = q->elevator->elevator_data;
2255 struct task_struct *tsk = current;
2256 struct cfq_io_context *cic;
2257 struct cfq_queue *cfqq;
2260 * don't force setup of a queue from here, as a call to may_queue
2261 * does not necessarily imply that a request actually will be queued.
2262 * so just lookup a possibly existing queue, or return 'may queue'
2265 cic = cfq_cic_lookup(cfqd, tsk->io_context);
2267 return ELV_MQUEUE_MAY;
2269 cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
2271 cfq_init_prio_data(cfqq, cic->ioc);
2272 cfq_prio_boost(cfqq);
2274 return __cfq_may_queue(cfqq);
2277 return ELV_MQUEUE_MAY;
2281 * queue lock held here
2283 static void cfq_put_request(struct request *rq)
2285 struct cfq_queue *cfqq = RQ_CFQQ(rq);
2288 const int rw = rq_data_dir(rq);
2290 BUG_ON(!cfqq->allocated[rw]);
2291 cfqq->allocated[rw]--;
2293 put_io_context(RQ_CIC(rq)->ioc);
2295 rq->elevator_private = NULL;
2296 rq->elevator_private2 = NULL;
2298 cfq_put_queue(cfqq);
2303 * Allocate cfq data structures associated with this request.
2306 cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
2308 struct cfq_data *cfqd = q->elevator->elevator_data;
2309 struct cfq_io_context *cic;
2310 const int rw = rq_data_dir(rq);
2311 const int is_sync = rq_is_sync(rq);
2312 struct cfq_queue *cfqq;
2313 unsigned long flags;
2315 might_sleep_if(gfp_mask & __GFP_WAIT);
2317 cic = cfq_get_io_context(cfqd, gfp_mask);
2319 spin_lock_irqsave(q->queue_lock, flags);
2324 cfqq = cic_to_cfqq(cic, is_sync);
2325 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
2326 cfqq = cfq_get_queue(cfqd, is_sync, cic->ioc, gfp_mask);
2327 cic_set_cfqq(cic, cfqq, is_sync);
2330 cfqq->allocated[rw]++;
2331 atomic_inc(&cfqq->ref);
2333 spin_unlock_irqrestore(q->queue_lock, flags);
2335 rq->elevator_private = cic;
2336 rq->elevator_private2 = cfqq;
2341 put_io_context(cic->ioc);
2343 cfq_schedule_dispatch(cfqd, 0);
2344 spin_unlock_irqrestore(q->queue_lock, flags);
2345 cfq_log(cfqd, "set_request fail");
2349 static void cfq_kick_queue(struct work_struct *work)
2351 struct cfq_data *cfqd =
2352 container_of(work, struct cfq_data, unplug_work.work);
2353 struct request_queue *q = cfqd->queue;
2355 spin_lock_irq(q->queue_lock);
2356 __blk_run_queue(cfqd->queue);
2357 spin_unlock_irq(q->queue_lock);
2361 * Timer running if the active_queue is currently idling inside its time slice
2363 static void cfq_idle_slice_timer(unsigned long data)
2365 struct cfq_data *cfqd = (struct cfq_data *) data;
2366 struct cfq_queue *cfqq;
2367 unsigned long flags;
2370 cfq_log(cfqd, "idle timer fired");
2372 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2374 cfqq = cfqd->active_queue;
2379 * We saw a request before the queue expired, let it through
2381 if (cfq_cfqq_must_dispatch(cfqq))
2387 if (cfq_slice_used(cfqq))
2391 * only expire and reinvoke request handler, if there are
2392 * other queues with pending requests
2394 if (!cfqd->busy_queues)
2398 * not expired and it has a request pending, let it dispatch
2400 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
2404 cfq_slice_expired(cfqd, timed_out);
2406 cfq_schedule_dispatch(cfqd, 0);
2408 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2411 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
2413 del_timer_sync(&cfqd->idle_slice_timer);
2414 cancel_delayed_work_sync(&cfqd->unplug_work);
2417 static void cfq_put_async_queues(struct cfq_data *cfqd)
2421 for (i = 0; i < IOPRIO_BE_NR; i++) {
2422 if (cfqd->async_cfqq[0][i])
2423 cfq_put_queue(cfqd->async_cfqq[0][i]);
2424 if (cfqd->async_cfqq[1][i])
2425 cfq_put_queue(cfqd->async_cfqq[1][i]);
2428 if (cfqd->async_idle_cfqq)
2429 cfq_put_queue(cfqd->async_idle_cfqq);
2432 static void cfq_exit_queue(struct elevator_queue *e)
2434 struct cfq_data *cfqd = e->elevator_data;
2435 struct request_queue *q = cfqd->queue;
2437 cfq_shutdown_timer_wq(cfqd);
2439 spin_lock_irq(q->queue_lock);
2441 if (cfqd->active_queue)
2442 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
2444 while (!list_empty(&cfqd->cic_list)) {
2445 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
2446 struct cfq_io_context,
2449 __cfq_exit_single_io_context(cfqd, cic);
2452 cfq_put_async_queues(cfqd);
2454 spin_unlock_irq(q->queue_lock);
2456 cfq_shutdown_timer_wq(cfqd);
2461 static void *cfq_init_queue(struct request_queue *q)
2463 struct cfq_data *cfqd;
2466 cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
2470 cfqd->service_tree = CFQ_RB_ROOT;
2473 * Not strictly needed (since RB_ROOT just clears the node and we
2474 * zeroed cfqd on alloc), but better be safe in case someone decides
2475 * to add magic to the rb code
2477 for (i = 0; i < CFQ_PRIO_LISTS; i++)
2478 cfqd->prio_trees[i] = RB_ROOT;
2481 * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
2482 * Grab a permanent reference to it, so that the normal code flow
2483 * will not attempt to free it.
2485 cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
2486 atomic_inc(&cfqd->oom_cfqq.ref);
2488 INIT_LIST_HEAD(&cfqd->cic_list);
2492 init_timer(&cfqd->idle_slice_timer);
2493 cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
2494 cfqd->idle_slice_timer.data = (unsigned long) cfqd;
2496 INIT_DELAYED_WORK(&cfqd->unplug_work, cfq_kick_queue);
2498 cfqd->cfq_quantum = cfq_quantum;
2499 cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
2500 cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
2501 cfqd->cfq_back_max = cfq_back_max;
2502 cfqd->cfq_back_penalty = cfq_back_penalty;
2503 cfqd->cfq_slice[0] = cfq_slice_async;
2504 cfqd->cfq_slice[1] = cfq_slice_sync;
2505 cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
2506 cfqd->cfq_slice_idle = cfq_slice_idle;
2507 cfqd->cfq_latency = 1;
2509 cfqd->last_end_sync_rq = jiffies;
2513 static void cfq_slab_kill(void)
2516 * Caller already ensured that pending RCU callbacks are completed,
2517 * so we should have no busy allocations at this point.
2520 kmem_cache_destroy(cfq_pool);
2522 kmem_cache_destroy(cfq_ioc_pool);
2525 static int __init cfq_slab_setup(void)
2527 cfq_pool = KMEM_CACHE(cfq_queue, 0);
2531 cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0);
2542 * sysfs parts below -->
2545 cfq_var_show(unsigned int var, char *page)
2547 return sprintf(page, "%d\n", var);
2551 cfq_var_store(unsigned int *var, const char *page, size_t count)
2553 char *p = (char *) page;
2555 *var = simple_strtoul(p, &p, 10);
2559 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
2560 static ssize_t __FUNC(struct elevator_queue *e, char *page) \
2562 struct cfq_data *cfqd = e->elevator_data; \
2563 unsigned int __data = __VAR; \
2565 __data = jiffies_to_msecs(__data); \
2566 return cfq_var_show(__data, (page)); \
2568 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
2569 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
2570 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
2571 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
2572 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
2573 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
2574 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
2575 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
2576 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
2577 SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
2578 #undef SHOW_FUNCTION
2580 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
2581 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
2583 struct cfq_data *cfqd = e->elevator_data; \
2584 unsigned int __data; \
2585 int ret = cfq_var_store(&__data, (page), count); \
2586 if (__data < (MIN)) \
2588 else if (__data > (MAX)) \
2591 *(__PTR) = msecs_to_jiffies(__data); \
2593 *(__PTR) = __data; \
2596 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
2597 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
2599 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
2601 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
2602 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
2604 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
2605 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
2606 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
2607 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
2609 STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
2610 #undef STORE_FUNCTION
2612 #define CFQ_ATTR(name) \
2613 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
2615 static struct elv_fs_entry cfq_attrs[] = {
2617 CFQ_ATTR(fifo_expire_sync),
2618 CFQ_ATTR(fifo_expire_async),
2619 CFQ_ATTR(back_seek_max),
2620 CFQ_ATTR(back_seek_penalty),
2621 CFQ_ATTR(slice_sync),
2622 CFQ_ATTR(slice_async),
2623 CFQ_ATTR(slice_async_rq),
2624 CFQ_ATTR(slice_idle),
2625 CFQ_ATTR(low_latency),
2629 static struct elevator_type iosched_cfq = {
2631 .elevator_merge_fn = cfq_merge,
2632 .elevator_merged_fn = cfq_merged_request,
2633 .elevator_merge_req_fn = cfq_merged_requests,
2634 .elevator_allow_merge_fn = cfq_allow_merge,
2635 .elevator_dispatch_fn = cfq_dispatch_requests,
2636 .elevator_add_req_fn = cfq_insert_request,
2637 .elevator_activate_req_fn = cfq_activate_request,
2638 .elevator_deactivate_req_fn = cfq_deactivate_request,
2639 .elevator_queue_empty_fn = cfq_queue_empty,
2640 .elevator_completed_req_fn = cfq_completed_request,
2641 .elevator_former_req_fn = elv_rb_former_request,
2642 .elevator_latter_req_fn = elv_rb_latter_request,
2643 .elevator_set_req_fn = cfq_set_request,
2644 .elevator_put_req_fn = cfq_put_request,
2645 .elevator_may_queue_fn = cfq_may_queue,
2646 .elevator_init_fn = cfq_init_queue,
2647 .elevator_exit_fn = cfq_exit_queue,
2648 .trim = cfq_free_io_context,
2650 .elevator_attrs = cfq_attrs,
2651 .elevator_name = "cfq",
2652 .elevator_owner = THIS_MODULE,
2655 static int __init cfq_init(void)
2658 * could be 0 on HZ < 1000 setups
2660 if (!cfq_slice_async)
2661 cfq_slice_async = 1;
2662 if (!cfq_slice_idle)
2665 if (cfq_slab_setup())
2668 elv_register(&iosched_cfq);
2673 static void __exit cfq_exit(void)
2675 DECLARE_COMPLETION_ONSTACK(all_gone);
2676 elv_unregister(&iosched_cfq);
2677 ioc_gone = &all_gone;
2678 /* ioc_gone's update must be visible before reading ioc_count */
2682 * this also protects us from entering cfq_slab_kill() with
2683 * pending RCU callbacks
2685 if (elv_ioc_count_read(cfq_ioc_count))
2686 wait_for_completion(&all_gone);
2690 module_init(cfq_init);
2691 module_exit(cfq_exit);
2693 MODULE_AUTHOR("Jens Axboe");
2694 MODULE_LICENSE("GPL");
2695 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");