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/hash.h>
13 #include <linux/rbtree.h>
14 #include <linux/ioprio.h>
19 static const int cfq_quantum = 4; /* max queue in one round of service */
20 static const int cfq_fifo_expire[2] = { HZ / 4, HZ / 8 };
21 static const int cfq_back_max = 16 * 1024; /* maximum backwards seek, in KiB */
22 static const int cfq_back_penalty = 2; /* penalty of a backwards seek */
24 static const int cfq_slice_sync = HZ / 10;
25 static int cfq_slice_async = HZ / 25;
26 static const int cfq_slice_async_rq = 2;
27 static int cfq_slice_idle = HZ / 125;
30 * grace period before allowing idle class to get disk access
32 #define CFQ_IDLE_GRACE (HZ / 10)
35 * below this threshold, we consider thinktime immediate
37 #define CFQ_MIN_TT (2)
39 #define CFQ_SLICE_SCALE (5)
41 #define CFQ_KEY_ASYNC (0)
44 * for the hash of cfqq inside the cfqd
46 #define CFQ_QHASH_SHIFT 6
47 #define CFQ_QHASH_ENTRIES (1 << CFQ_QHASH_SHIFT)
49 #define RQ_CIC(rq) ((struct cfq_io_context*)(rq)->elevator_private)
50 #define RQ_CFQQ(rq) ((rq)->elevator_private2)
52 static struct kmem_cache *cfq_pool;
53 static struct kmem_cache *cfq_ioc_pool;
55 static DEFINE_PER_CPU(unsigned long, ioc_count);
56 static struct completion *ioc_gone;
58 #define CFQ_PRIO_LISTS IOPRIO_BE_NR
59 #define cfq_class_idle(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
60 #define cfq_class_rt(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
65 #define cfq_cfqq_sync(cfqq) ((cfqq)->key != CFQ_KEY_ASYNC)
67 #define sample_valid(samples) ((samples) > 80)
70 * Most of our rbtree usage is for sorting with min extraction, so
71 * if we cache the leftmost node we don't have to walk down the tree
72 * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should
73 * move this into the elevator for the rq sorting as well.
79 #define CFQ_RB_ROOT (struct cfq_rb_root) { RB_ROOT, NULL, }
82 * Per block device queue structure
85 request_queue_t *queue;
88 * rr list of queues with requests and the count of them
90 struct cfq_rb_root service_tree;
91 unsigned int busy_queues;
96 struct hlist_head *cfq_hash;
102 * idle window management
104 struct timer_list idle_slice_timer;
105 struct work_struct unplug_work;
107 struct cfq_queue *active_queue;
108 struct cfq_io_context *active_cic;
109 unsigned int dispatch_slice;
111 struct timer_list idle_class_timer;
113 sector_t last_position;
114 unsigned long last_end_request;
117 * tunables, see top of file
119 unsigned int cfq_quantum;
120 unsigned int cfq_fifo_expire[2];
121 unsigned int cfq_back_penalty;
122 unsigned int cfq_back_max;
123 unsigned int cfq_slice[2];
124 unsigned int cfq_slice_async_rq;
125 unsigned int cfq_slice_idle;
127 struct list_head cic_list;
129 sector_t new_seek_mean;
134 * Per process-grouping structure
137 /* reference count */
139 /* parent cfq_data */
140 struct cfq_data *cfqd;
141 /* cfqq lookup hash */
142 struct hlist_node cfq_hash;
145 /* service_tree member */
146 struct rb_node rb_node;
147 /* service_tree key */
148 unsigned long rb_key;
149 /* sorted list of pending requests */
150 struct rb_root sort_list;
151 /* if fifo isn't expired, next request to serve */
152 struct request *next_rq;
153 /* requests queued in sort_list */
155 /* currently allocated requests */
157 /* pending metadata requests */
159 /* fifo list of requests in sort_list */
160 struct list_head fifo;
162 unsigned long slice_end;
165 /* number of requests that are on the dispatch list or inside driver */
168 /* io prio of this group */
169 unsigned short ioprio, org_ioprio;
170 unsigned short ioprio_class, org_ioprio_class;
172 /* various state flags, see below */
175 sector_t last_request_pos;
178 enum cfqq_state_flags {
179 CFQ_CFQQ_FLAG_on_rr = 0, /* on round-robin busy list */
180 CFQ_CFQQ_FLAG_wait_request, /* waiting for a request */
181 CFQ_CFQQ_FLAG_must_alloc, /* must be allowed rq alloc */
182 CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
183 CFQ_CFQQ_FLAG_must_dispatch, /* must dispatch, even if expired */
184 CFQ_CFQQ_FLAG_fifo_expire, /* FIFO checked in this slice */
185 CFQ_CFQQ_FLAG_idle_window, /* slice idling enabled */
186 CFQ_CFQQ_FLAG_prio_changed, /* task priority has changed */
187 CFQ_CFQQ_FLAG_queue_new, /* queue never been serviced */
188 CFQ_CFQQ_FLAG_slice_new, /* no requests dispatched in slice */
191 #define CFQ_CFQQ_FNS(name) \
192 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \
194 cfqq->flags |= (1 << CFQ_CFQQ_FLAG_##name); \
196 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \
198 cfqq->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \
200 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \
202 return (cfqq->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \
206 CFQ_CFQQ_FNS(wait_request);
207 CFQ_CFQQ_FNS(must_alloc);
208 CFQ_CFQQ_FNS(must_alloc_slice);
209 CFQ_CFQQ_FNS(must_dispatch);
210 CFQ_CFQQ_FNS(fifo_expire);
211 CFQ_CFQQ_FNS(idle_window);
212 CFQ_CFQQ_FNS(prio_changed);
213 CFQ_CFQQ_FNS(queue_new);
214 CFQ_CFQQ_FNS(slice_new);
217 static struct cfq_queue *cfq_find_cfq_hash(struct cfq_data *, unsigned int, unsigned short);
218 static void cfq_dispatch_insert(request_queue_t *, struct request *);
219 static struct cfq_queue *cfq_get_queue(struct cfq_data *, unsigned int, struct task_struct *, gfp_t);
222 * scheduler run of queue, if there are requests pending and no one in the
223 * driver that will restart queueing
225 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
227 if (cfqd->busy_queues)
228 kblockd_schedule_work(&cfqd->unplug_work);
231 static int cfq_queue_empty(request_queue_t *q)
233 struct cfq_data *cfqd = q->elevator->elevator_data;
235 return !cfqd->busy_queues;
238 static inline pid_t cfq_queue_pid(struct task_struct *task, int rw, int is_sync)
241 * Use the per-process queue, for read requests and syncronous writes
243 if (!(rw & REQ_RW) || is_sync)
246 return CFQ_KEY_ASYNC;
250 * Scale schedule slice based on io priority. Use the sync time slice only
251 * if a queue is marked sync and has sync io queued. A sync queue with async
252 * io only, should not get full sync slice length.
254 static inline int cfq_prio_slice(struct cfq_data *cfqd, int sync,
257 const int base_slice = cfqd->cfq_slice[sync];
259 WARN_ON(prio >= IOPRIO_BE_NR);
261 return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
265 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
267 return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
271 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
273 cfqq->slice_end = cfq_prio_to_slice(cfqd, cfqq) + jiffies;
277 * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
278 * isn't valid until the first request from the dispatch is activated
279 * and the slice time set.
281 static inline int cfq_slice_used(struct cfq_queue *cfqq)
283 if (cfq_cfqq_slice_new(cfqq))
285 if (time_before(jiffies, cfqq->slice_end))
292 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
293 * We choose the request that is closest to the head right now. Distance
294 * behind the head is penalized and only allowed to a certain extent.
296 static struct request *
297 cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2)
299 sector_t last, s1, s2, d1 = 0, d2 = 0;
300 unsigned long back_max;
301 #define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */
302 #define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */
303 unsigned wrap = 0; /* bit mask: requests behind the disk head? */
305 if (rq1 == NULL || rq1 == rq2)
310 if (rq_is_sync(rq1) && !rq_is_sync(rq2))
312 else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
314 if (rq_is_meta(rq1) && !rq_is_meta(rq2))
316 else if (rq_is_meta(rq2) && !rq_is_meta(rq1))
322 last = cfqd->last_position;
325 * by definition, 1KiB is 2 sectors
327 back_max = cfqd->cfq_back_max * 2;
330 * Strict one way elevator _except_ in the case where we allow
331 * short backward seeks which are biased as twice the cost of a
332 * similar forward seek.
336 else if (s1 + back_max >= last)
337 d1 = (last - s1) * cfqd->cfq_back_penalty;
339 wrap |= CFQ_RQ1_WRAP;
343 else if (s2 + back_max >= last)
344 d2 = (last - s2) * cfqd->cfq_back_penalty;
346 wrap |= CFQ_RQ2_WRAP;
348 /* Found required data */
351 * By doing switch() on the bit mask "wrap" we avoid having to
352 * check two variables for all permutations: --> faster!
355 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
371 case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
374 * Since both rqs are wrapped,
375 * start with the one that's further behind head
376 * (--> only *one* back seek required),
377 * since back seek takes more time than forward.
387 * The below is leftmost cache rbtree addon
389 static struct rb_node *cfq_rb_first(struct cfq_rb_root *root)
392 root->left = rb_first(&root->rb);
397 static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
402 rb_erase(n, &root->rb);
407 * would be nice to take fifo expire time into account as well
409 static struct request *
410 cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
411 struct request *last)
413 struct rb_node *rbnext = rb_next(&last->rb_node);
414 struct rb_node *rbprev = rb_prev(&last->rb_node);
415 struct request *next = NULL, *prev = NULL;
417 BUG_ON(RB_EMPTY_NODE(&last->rb_node));
420 prev = rb_entry_rq(rbprev);
423 next = rb_entry_rq(rbnext);
425 rbnext = rb_first(&cfqq->sort_list);
426 if (rbnext && rbnext != &last->rb_node)
427 next = rb_entry_rq(rbnext);
430 return cfq_choose_req(cfqd, next, prev);
433 static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
434 struct cfq_queue *cfqq)
437 * just an approximation, should be ok.
439 return (cfqd->busy_queues - 1) * (cfq_prio_slice(cfqd, 1, 0) -
440 cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
444 * The cfqd->service_tree holds all pending cfq_queue's that have
445 * requests waiting to be processed. It is sorted in the order that
446 * we will service the queues.
448 static void cfq_service_tree_add(struct cfq_data *cfqd,
449 struct cfq_queue *cfqq, int add_front)
451 struct rb_node **p = &cfqd->service_tree.rb.rb_node;
452 struct rb_node *parent = NULL;
453 unsigned long rb_key;
457 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
458 rb_key += cfqq->slice_resid;
459 cfqq->slice_resid = 0;
463 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
465 * same position, nothing more to do
467 if (rb_key == cfqq->rb_key)
470 cfq_rb_erase(&cfqq->rb_node, &cfqd->service_tree);
475 struct cfq_queue *__cfqq;
479 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
482 * sort RT queues first, we always want to give
483 * preference to them. IDLE queues goes to the back.
484 * after that, sort on the next service time.
486 if (cfq_class_rt(cfqq) > cfq_class_rt(__cfqq))
488 else if (cfq_class_rt(cfqq) < cfq_class_rt(__cfqq))
490 else if (cfq_class_idle(cfqq) < cfq_class_idle(__cfqq))
492 else if (cfq_class_idle(cfqq) > cfq_class_idle(__cfqq))
494 else if (rb_key < __cfqq->rb_key)
499 if (n == &(*p)->rb_right)
506 cfqd->service_tree.left = &cfqq->rb_node;
508 cfqq->rb_key = rb_key;
509 rb_link_node(&cfqq->rb_node, parent, p);
510 rb_insert_color(&cfqq->rb_node, &cfqd->service_tree.rb);
514 * Update cfqq's position in the service tree.
516 static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
519 * Resorting requires the cfqq to be on the RR list already.
521 if (cfq_cfqq_on_rr(cfqq))
522 cfq_service_tree_add(cfqd, cfqq, 0);
526 * add to busy list of queues for service, trying to be fair in ordering
527 * the pending list according to last request service
530 cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
532 BUG_ON(cfq_cfqq_on_rr(cfqq));
533 cfq_mark_cfqq_on_rr(cfqq);
536 cfq_resort_rr_list(cfqd, cfqq);
540 * Called when the cfqq no longer has requests pending, remove it from
544 cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
546 BUG_ON(!cfq_cfqq_on_rr(cfqq));
547 cfq_clear_cfqq_on_rr(cfqq);
549 if (!RB_EMPTY_NODE(&cfqq->rb_node))
550 cfq_rb_erase(&cfqq->rb_node, &cfqd->service_tree);
552 BUG_ON(!cfqd->busy_queues);
557 * rb tree support functions
559 static inline void cfq_del_rq_rb(struct request *rq)
561 struct cfq_queue *cfqq = RQ_CFQQ(rq);
562 struct cfq_data *cfqd = cfqq->cfqd;
563 const int sync = rq_is_sync(rq);
565 BUG_ON(!cfqq->queued[sync]);
566 cfqq->queued[sync]--;
568 elv_rb_del(&cfqq->sort_list, rq);
570 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
571 cfq_del_cfqq_rr(cfqd, cfqq);
574 static void cfq_add_rq_rb(struct request *rq)
576 struct cfq_queue *cfqq = RQ_CFQQ(rq);
577 struct cfq_data *cfqd = cfqq->cfqd;
578 struct request *__alias;
580 cfqq->queued[rq_is_sync(rq)]++;
583 * looks a little odd, but the first insert might return an alias.
584 * if that happens, put the alias on the dispatch list
586 while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL)
587 cfq_dispatch_insert(cfqd->queue, __alias);
589 if (!cfq_cfqq_on_rr(cfqq))
590 cfq_add_cfqq_rr(cfqd, cfqq);
593 * check if this request is a better next-serve candidate
595 cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq);
596 BUG_ON(!cfqq->next_rq);
600 cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
602 elv_rb_del(&cfqq->sort_list, rq);
603 cfqq->queued[rq_is_sync(rq)]--;
607 static struct request *
608 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
610 struct task_struct *tsk = current;
611 pid_t key = cfq_queue_pid(tsk, bio_data_dir(bio), bio_sync(bio));
612 struct cfq_queue *cfqq;
614 cfqq = cfq_find_cfq_hash(cfqd, key, tsk->ioprio);
616 sector_t sector = bio->bi_sector + bio_sectors(bio);
618 return elv_rb_find(&cfqq->sort_list, sector);
624 static void cfq_activate_request(request_queue_t *q, struct request *rq)
626 struct cfq_data *cfqd = q->elevator->elevator_data;
628 cfqd->rq_in_driver++;
631 * If the depth is larger 1, it really could be queueing. But lets
632 * make the mark a little higher - idling could still be good for
633 * low queueing, and a low queueing number could also just indicate
634 * a SCSI mid layer like behaviour where limit+1 is often seen.
636 if (!cfqd->hw_tag && cfqd->rq_in_driver > 4)
639 cfqd->last_position = rq->hard_sector + rq->hard_nr_sectors;
642 static void cfq_deactivate_request(request_queue_t *q, struct request *rq)
644 struct cfq_data *cfqd = q->elevator->elevator_data;
646 WARN_ON(!cfqd->rq_in_driver);
647 cfqd->rq_in_driver--;
650 static void cfq_remove_request(struct request *rq)
652 struct cfq_queue *cfqq = RQ_CFQQ(rq);
654 if (cfqq->next_rq == rq)
655 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
657 list_del_init(&rq->queuelist);
660 if (rq_is_meta(rq)) {
661 WARN_ON(!cfqq->meta_pending);
662 cfqq->meta_pending--;
666 static int cfq_merge(request_queue_t *q, struct request **req, struct bio *bio)
668 struct cfq_data *cfqd = q->elevator->elevator_data;
669 struct request *__rq;
671 __rq = cfq_find_rq_fmerge(cfqd, bio);
672 if (__rq && elv_rq_merge_ok(__rq, bio)) {
674 return ELEVATOR_FRONT_MERGE;
677 return ELEVATOR_NO_MERGE;
680 static void cfq_merged_request(request_queue_t *q, struct request *req,
683 if (type == ELEVATOR_FRONT_MERGE) {
684 struct cfq_queue *cfqq = RQ_CFQQ(req);
686 cfq_reposition_rq_rb(cfqq, req);
691 cfq_merged_requests(request_queue_t *q, struct request *rq,
692 struct request *next)
695 * reposition in fifo if next is older than rq
697 if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
698 time_before(next->start_time, rq->start_time))
699 list_move(&rq->queuelist, &next->queuelist);
701 cfq_remove_request(next);
704 static int cfq_allow_merge(request_queue_t *q, struct request *rq,
707 struct cfq_data *cfqd = q->elevator->elevator_data;
708 const int rw = bio_data_dir(bio);
709 struct cfq_queue *cfqq;
713 * Disallow merge of a sync bio into an async request.
715 if ((bio_data_dir(bio) == READ || bio_sync(bio)) && !rq_is_sync(rq))
719 * Lookup the cfqq that this bio will be queued with. Allow
720 * merge only if rq is queued there.
722 key = cfq_queue_pid(current, rw, bio_sync(bio));
723 cfqq = cfq_find_cfq_hash(cfqd, key, current->ioprio);
725 if (cfqq == RQ_CFQQ(rq))
732 __cfq_set_active_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
736 * stop potential idle class queues waiting service
738 del_timer(&cfqd->idle_class_timer);
741 cfq_clear_cfqq_must_alloc_slice(cfqq);
742 cfq_clear_cfqq_fifo_expire(cfqq);
743 cfq_mark_cfqq_slice_new(cfqq);
744 cfq_clear_cfqq_queue_new(cfqq);
747 cfqd->active_queue = cfqq;
751 * current cfqq expired its slice (or was too idle), select new one
754 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
755 int preempted, int timed_out)
757 if (cfq_cfqq_wait_request(cfqq))
758 del_timer(&cfqd->idle_slice_timer);
760 cfq_clear_cfqq_must_dispatch(cfqq);
761 cfq_clear_cfqq_wait_request(cfqq);
764 * store what was left of this slice, if the queue idled out
767 if (timed_out && !cfq_cfqq_slice_new(cfqq))
768 cfqq->slice_resid = cfqq->slice_end - jiffies;
770 cfq_resort_rr_list(cfqd, cfqq);
772 if (cfqq == cfqd->active_queue)
773 cfqd->active_queue = NULL;
775 if (cfqd->active_cic) {
776 put_io_context(cfqd->active_cic->ioc);
777 cfqd->active_cic = NULL;
780 cfqd->dispatch_slice = 0;
783 static inline void cfq_slice_expired(struct cfq_data *cfqd, int preempted,
786 struct cfq_queue *cfqq = cfqd->active_queue;
789 __cfq_slice_expired(cfqd, cfqq, preempted, timed_out);
793 * Get next queue for service. Unless we have a queue preemption,
794 * we'll simply select the first cfqq in the service tree.
796 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
798 struct cfq_queue *cfqq;
801 if (RB_EMPTY_ROOT(&cfqd->service_tree.rb))
804 n = cfq_rb_first(&cfqd->service_tree);
805 cfqq = rb_entry(n, struct cfq_queue, rb_node);
807 if (cfq_class_idle(cfqq)) {
811 * if we have idle queues and no rt or be queues had
812 * pending requests, either allow immediate service if
813 * the grace period has passed or arm the idle grace
816 end = cfqd->last_end_request + CFQ_IDLE_GRACE;
817 if (time_before(jiffies, end)) {
818 mod_timer(&cfqd->idle_class_timer, end);
827 * Get and set a new active queue for service.
829 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd)
831 struct cfq_queue *cfqq;
833 cfqq = cfq_get_next_queue(cfqd);
834 __cfq_set_active_queue(cfqd, cfqq);
838 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
841 if (rq->sector >= cfqd->last_position)
842 return rq->sector - cfqd->last_position;
844 return cfqd->last_position - rq->sector;
847 static inline int cfq_rq_close(struct cfq_data *cfqd, struct request *rq)
849 struct cfq_io_context *cic = cfqd->active_cic;
851 if (!sample_valid(cic->seek_samples))
854 return cfq_dist_from_last(cfqd, rq) <= cic->seek_mean;
857 static int cfq_close_cooperator(struct cfq_data *cfq_data,
858 struct cfq_queue *cfqq)
861 * We should notice if some of the queues are cooperating, eg
862 * working closely on the same area of the disk. In that case,
863 * we can group them together and don't waste time idling.
868 #define CIC_SEEKY(cic) ((cic)->seek_mean > (8 * 1024))
870 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
872 struct cfq_queue *cfqq = cfqd->active_queue;
873 struct cfq_io_context *cic;
876 WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
877 WARN_ON(cfq_cfqq_slice_new(cfqq));
880 * idle is disabled, either manually or by past process history
882 if (!cfqd->cfq_slice_idle || !cfq_cfqq_idle_window(cfqq))
886 * task has exited, don't wait
888 cic = cfqd->active_cic;
889 if (!cic || !cic->ioc->task)
893 * See if this prio level has a good candidate
895 if (cfq_close_cooperator(cfqd, cfqq) &&
896 (sample_valid(cic->ttime_samples) && cic->ttime_mean > 2))
899 cfq_mark_cfqq_must_dispatch(cfqq);
900 cfq_mark_cfqq_wait_request(cfqq);
903 * we don't want to idle for seeks, but we do want to allow
904 * fair distribution of slice time for a process doing back-to-back
905 * seeks. so allow a little bit of time for him to submit a new rq
907 sl = cfqd->cfq_slice_idle;
908 if (sample_valid(cic->seek_samples) && CIC_SEEKY(cic))
909 sl = min(sl, msecs_to_jiffies(CFQ_MIN_TT));
911 mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
915 * Move request from internal lists to the request queue dispatch list.
917 static void cfq_dispatch_insert(request_queue_t *q, struct request *rq)
919 struct cfq_queue *cfqq = RQ_CFQQ(rq);
921 cfq_remove_request(rq);
923 elv_dispatch_sort(q, rq);
927 * return expired entry, or NULL to just start from scratch in rbtree
929 static inline struct request *cfq_check_fifo(struct cfq_queue *cfqq)
931 struct cfq_data *cfqd = cfqq->cfqd;
935 if (cfq_cfqq_fifo_expire(cfqq))
938 cfq_mark_cfqq_fifo_expire(cfqq);
940 if (list_empty(&cfqq->fifo))
943 fifo = cfq_cfqq_sync(cfqq);
944 rq = rq_entry_fifo(cfqq->fifo.next);
946 if (time_before(jiffies, rq->start_time + cfqd->cfq_fifo_expire[fifo]))
953 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
955 const int base_rq = cfqd->cfq_slice_async_rq;
957 WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
959 return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
963 * Select a queue for service. If we have a current active queue,
964 * check whether to continue servicing it, or retrieve and set a new one.
966 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
968 struct cfq_queue *cfqq;
970 cfqq = cfqd->active_queue;
975 * The active queue has run out of time, expire it and select new.
977 if (cfq_slice_used(cfqq))
981 * The active queue has requests and isn't expired, allow it to
984 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
988 * No requests pending. If the active queue still has requests in
989 * flight or is idling for a new request, allow either of these
990 * conditions to happen (or time out) before selecting a new queue.
992 if (cfqq->dispatched || timer_pending(&cfqd->idle_slice_timer)) {
998 cfq_slice_expired(cfqd, 0, 0);
1000 cfqq = cfq_set_active_queue(cfqd);
1006 * Dispatch some requests from cfqq, moving them to the request queue
1010 __cfq_dispatch_requests(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1015 BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
1021 * follow expired path, else get first next available
1023 if ((rq = cfq_check_fifo(cfqq)) == NULL)
1027 * finally, insert request into driver dispatch list
1029 cfq_dispatch_insert(cfqd->queue, rq);
1031 cfqd->dispatch_slice++;
1034 if (!cfqd->active_cic) {
1035 atomic_inc(&RQ_CIC(rq)->ioc->refcount);
1036 cfqd->active_cic = RQ_CIC(rq);
1039 if (RB_EMPTY_ROOT(&cfqq->sort_list))
1042 } while (dispatched < max_dispatch);
1045 * expire an async queue immediately if it has used up its slice. idle
1046 * queue always expire after 1 dispatch round.
1048 if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
1049 cfqd->dispatch_slice >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
1050 cfq_class_idle(cfqq))) {
1051 cfqq->slice_end = jiffies + 1;
1052 cfq_slice_expired(cfqd, 0, 0);
1058 static inline int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
1062 while (cfqq->next_rq) {
1063 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
1067 BUG_ON(!list_empty(&cfqq->fifo));
1072 * Drain our current requests. Used for barriers and when switching
1073 * io schedulers on-the-fly.
1075 static int cfq_forced_dispatch(struct cfq_data *cfqd)
1080 while ((n = cfq_rb_first(&cfqd->service_tree)) != NULL) {
1081 struct cfq_queue *cfqq = rb_entry(n, struct cfq_queue, rb_node);
1083 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
1086 cfq_slice_expired(cfqd, 0, 0);
1088 BUG_ON(cfqd->busy_queues);
1093 static int cfq_dispatch_requests(request_queue_t *q, int force)
1095 struct cfq_data *cfqd = q->elevator->elevator_data;
1096 struct cfq_queue *cfqq;
1099 if (!cfqd->busy_queues)
1102 if (unlikely(force))
1103 return cfq_forced_dispatch(cfqd);
1106 while ((cfqq = cfq_select_queue(cfqd)) != NULL) {
1109 if (cfqd->busy_queues > 1) {
1111 * So we have dispatched before in this round, if the
1112 * next queue has idling enabled (must be sync), don't
1113 * allow it service until the previous have completed.
1115 if (cfqd->rq_in_driver && cfq_cfqq_idle_window(cfqq) &&
1118 if (cfqq->dispatched >= cfqd->cfq_quantum)
1122 cfq_clear_cfqq_must_dispatch(cfqq);
1123 cfq_clear_cfqq_wait_request(cfqq);
1124 del_timer(&cfqd->idle_slice_timer);
1126 max_dispatch = cfqd->cfq_quantum;
1127 if (cfq_class_idle(cfqq))
1130 dispatched += __cfq_dispatch_requests(cfqd, cfqq, max_dispatch);
1137 * task holds one reference to the queue, dropped when task exits. each rq
1138 * in-flight on this queue also holds a reference, dropped when rq is freed.
1140 * queue lock must be held here.
1142 static void cfq_put_queue(struct cfq_queue *cfqq)
1144 struct cfq_data *cfqd = cfqq->cfqd;
1146 BUG_ON(atomic_read(&cfqq->ref) <= 0);
1148 if (!atomic_dec_and_test(&cfqq->ref))
1151 BUG_ON(rb_first(&cfqq->sort_list));
1152 BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
1153 BUG_ON(cfq_cfqq_on_rr(cfqq));
1155 if (unlikely(cfqd->active_queue == cfqq)) {
1156 __cfq_slice_expired(cfqd, cfqq, 0, 0);
1157 cfq_schedule_dispatch(cfqd);
1161 * it's on the empty list and still hashed
1163 hlist_del(&cfqq->cfq_hash);
1164 kmem_cache_free(cfq_pool, cfqq);
1167 static struct cfq_queue *
1168 __cfq_find_cfq_hash(struct cfq_data *cfqd, unsigned int key, unsigned int prio,
1171 struct hlist_head *hash_list = &cfqd->cfq_hash[hashval];
1172 struct hlist_node *entry;
1173 struct cfq_queue *__cfqq;
1175 hlist_for_each_entry(__cfqq, entry, hash_list, cfq_hash) {
1176 const unsigned short __p = IOPRIO_PRIO_VALUE(__cfqq->org_ioprio_class, __cfqq->org_ioprio);
1178 if (__cfqq->key == key && (__p == prio || !prio))
1185 static struct cfq_queue *
1186 cfq_find_cfq_hash(struct cfq_data *cfqd, unsigned int key, unsigned short prio)
1188 return __cfq_find_cfq_hash(cfqd, key, prio, hash_long(key, CFQ_QHASH_SHIFT));
1191 static void cfq_free_io_context(struct io_context *ioc)
1193 struct cfq_io_context *__cic;
1197 while ((n = rb_first(&ioc->cic_root)) != NULL) {
1198 __cic = rb_entry(n, struct cfq_io_context, rb_node);
1199 rb_erase(&__cic->rb_node, &ioc->cic_root);
1200 kmem_cache_free(cfq_ioc_pool, __cic);
1204 elv_ioc_count_mod(ioc_count, -freed);
1206 if (ioc_gone && !elv_ioc_count_read(ioc_count))
1210 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1212 if (unlikely(cfqq == cfqd->active_queue)) {
1213 __cfq_slice_expired(cfqd, cfqq, 0, 0);
1214 cfq_schedule_dispatch(cfqd);
1217 cfq_put_queue(cfqq);
1220 static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
1221 struct cfq_io_context *cic)
1223 list_del_init(&cic->queue_list);
1227 if (cic->cfqq[ASYNC]) {
1228 cfq_exit_cfqq(cfqd, cic->cfqq[ASYNC]);
1229 cic->cfqq[ASYNC] = NULL;
1232 if (cic->cfqq[SYNC]) {
1233 cfq_exit_cfqq(cfqd, cic->cfqq[SYNC]);
1234 cic->cfqq[SYNC] = NULL;
1238 static void cfq_exit_single_io_context(struct cfq_io_context *cic)
1240 struct cfq_data *cfqd = cic->key;
1243 request_queue_t *q = cfqd->queue;
1245 spin_lock_irq(q->queue_lock);
1246 __cfq_exit_single_io_context(cfqd, cic);
1247 spin_unlock_irq(q->queue_lock);
1252 * The process that ioc belongs to has exited, we need to clean up
1253 * and put the internal structures we have that belongs to that process.
1255 static void cfq_exit_io_context(struct io_context *ioc)
1257 struct cfq_io_context *__cic;
1261 * put the reference this task is holding to the various queues
1264 n = rb_first(&ioc->cic_root);
1266 __cic = rb_entry(n, struct cfq_io_context, rb_node);
1268 cfq_exit_single_io_context(__cic);
1273 static struct cfq_io_context *
1274 cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
1276 struct cfq_io_context *cic;
1278 cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask, cfqd->queue->node);
1280 memset(cic, 0, sizeof(*cic));
1281 cic->last_end_request = jiffies;
1282 INIT_LIST_HEAD(&cic->queue_list);
1283 cic->dtor = cfq_free_io_context;
1284 cic->exit = cfq_exit_io_context;
1285 elv_ioc_count_inc(ioc_count);
1291 static void cfq_init_prio_data(struct cfq_queue *cfqq)
1293 struct task_struct *tsk = current;
1296 if (!cfq_cfqq_prio_changed(cfqq))
1299 ioprio_class = IOPRIO_PRIO_CLASS(tsk->ioprio);
1300 switch (ioprio_class) {
1302 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
1303 case IOPRIO_CLASS_NONE:
1305 * no prio set, place us in the middle of the BE classes
1307 cfqq->ioprio = task_nice_ioprio(tsk);
1308 cfqq->ioprio_class = IOPRIO_CLASS_BE;
1310 case IOPRIO_CLASS_RT:
1311 cfqq->ioprio = task_ioprio(tsk);
1312 cfqq->ioprio_class = IOPRIO_CLASS_RT;
1314 case IOPRIO_CLASS_BE:
1315 cfqq->ioprio = task_ioprio(tsk);
1316 cfqq->ioprio_class = IOPRIO_CLASS_BE;
1318 case IOPRIO_CLASS_IDLE:
1319 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
1321 cfq_clear_cfqq_idle_window(cfqq);
1326 * keep track of original prio settings in case we have to temporarily
1327 * elevate the priority of this queue
1329 cfqq->org_ioprio = cfqq->ioprio;
1330 cfqq->org_ioprio_class = cfqq->ioprio_class;
1331 cfq_clear_cfqq_prio_changed(cfqq);
1334 static inline void changed_ioprio(struct cfq_io_context *cic)
1336 struct cfq_data *cfqd = cic->key;
1337 struct cfq_queue *cfqq;
1338 unsigned long flags;
1340 if (unlikely(!cfqd))
1343 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1345 cfqq = cic->cfqq[ASYNC];
1347 struct cfq_queue *new_cfqq;
1348 new_cfqq = cfq_get_queue(cfqd, CFQ_KEY_ASYNC, cic->ioc->task,
1351 cic->cfqq[ASYNC] = new_cfqq;
1352 cfq_put_queue(cfqq);
1356 cfqq = cic->cfqq[SYNC];
1358 cfq_mark_cfqq_prio_changed(cfqq);
1360 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1363 static void cfq_ioc_set_ioprio(struct io_context *ioc)
1365 struct cfq_io_context *cic;
1368 ioc->ioprio_changed = 0;
1370 n = rb_first(&ioc->cic_root);
1372 cic = rb_entry(n, struct cfq_io_context, rb_node);
1374 changed_ioprio(cic);
1379 static struct cfq_queue *
1380 cfq_get_queue(struct cfq_data *cfqd, unsigned int key, struct task_struct *tsk,
1383 const int hashval = hash_long(key, CFQ_QHASH_SHIFT);
1384 struct cfq_queue *cfqq, *new_cfqq = NULL;
1385 unsigned short ioprio;
1388 ioprio = tsk->ioprio;
1389 cfqq = __cfq_find_cfq_hash(cfqd, key, ioprio, hashval);
1395 } else if (gfp_mask & __GFP_WAIT) {
1397 * Inform the allocator of the fact that we will
1398 * just repeat this allocation if it fails, to allow
1399 * the allocator to do whatever it needs to attempt to
1402 spin_unlock_irq(cfqd->queue->queue_lock);
1403 new_cfqq = kmem_cache_alloc_node(cfq_pool, gfp_mask|__GFP_NOFAIL, cfqd->queue->node);
1404 spin_lock_irq(cfqd->queue->queue_lock);
1407 cfqq = kmem_cache_alloc_node(cfq_pool, gfp_mask, cfqd->queue->node);
1412 memset(cfqq, 0, sizeof(*cfqq));
1414 INIT_HLIST_NODE(&cfqq->cfq_hash);
1415 RB_CLEAR_NODE(&cfqq->rb_node);
1416 INIT_LIST_HEAD(&cfqq->fifo);
1419 hlist_add_head(&cfqq->cfq_hash, &cfqd->cfq_hash[hashval]);
1420 atomic_set(&cfqq->ref, 0);
1423 if (key != CFQ_KEY_ASYNC)
1424 cfq_mark_cfqq_idle_window(cfqq);
1426 cfq_mark_cfqq_prio_changed(cfqq);
1427 cfq_mark_cfqq_queue_new(cfqq);
1428 cfq_init_prio_data(cfqq);
1432 kmem_cache_free(cfq_pool, new_cfqq);
1434 atomic_inc(&cfqq->ref);
1436 WARN_ON((gfp_mask & __GFP_WAIT) && !cfqq);
1441 * We drop cfq io contexts lazily, so we may find a dead one.
1444 cfq_drop_dead_cic(struct io_context *ioc, struct cfq_io_context *cic)
1446 WARN_ON(!list_empty(&cic->queue_list));
1447 rb_erase(&cic->rb_node, &ioc->cic_root);
1448 kmem_cache_free(cfq_ioc_pool, cic);
1449 elv_ioc_count_dec(ioc_count);
1452 static struct cfq_io_context *
1453 cfq_cic_rb_lookup(struct cfq_data *cfqd, struct io_context *ioc)
1456 struct cfq_io_context *cic;
1457 void *k, *key = cfqd;
1460 n = ioc->cic_root.rb_node;
1462 cic = rb_entry(n, struct cfq_io_context, rb_node);
1463 /* ->key must be copied to avoid race with cfq_exit_queue() */
1466 cfq_drop_dead_cic(ioc, cic);
1482 cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
1483 struct cfq_io_context *cic)
1486 struct rb_node *parent;
1487 struct cfq_io_context *__cic;
1488 unsigned long flags;
1496 p = &ioc->cic_root.rb_node;
1499 __cic = rb_entry(parent, struct cfq_io_context, rb_node);
1500 /* ->key must be copied to avoid race with cfq_exit_queue() */
1503 cfq_drop_dead_cic(ioc, __cic);
1509 else if (cic->key > k)
1510 p = &(*p)->rb_right;
1515 rb_link_node(&cic->rb_node, parent, p);
1516 rb_insert_color(&cic->rb_node, &ioc->cic_root);
1518 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1519 list_add(&cic->queue_list, &cfqd->cic_list);
1520 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1524 * Setup general io context and cfq io context. There can be several cfq
1525 * io contexts per general io context, if this process is doing io to more
1526 * than one device managed by cfq.
1528 static struct cfq_io_context *
1529 cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
1531 struct io_context *ioc = NULL;
1532 struct cfq_io_context *cic;
1534 might_sleep_if(gfp_mask & __GFP_WAIT);
1536 ioc = get_io_context(gfp_mask, cfqd->queue->node);
1540 cic = cfq_cic_rb_lookup(cfqd, ioc);
1544 cic = cfq_alloc_io_context(cfqd, gfp_mask);
1548 cfq_cic_link(cfqd, ioc, cic);
1550 smp_read_barrier_depends();
1551 if (unlikely(ioc->ioprio_changed))
1552 cfq_ioc_set_ioprio(ioc);
1556 put_io_context(ioc);
1561 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
1563 unsigned long elapsed = jiffies - cic->last_end_request;
1564 unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
1566 cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
1567 cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
1568 cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
1572 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_io_context *cic,
1578 if (cic->last_request_pos < rq->sector)
1579 sdist = rq->sector - cic->last_request_pos;
1581 sdist = cic->last_request_pos - rq->sector;
1583 if (!cic->seek_samples) {
1584 cfqd->new_seek_total = (7*cic->seek_total + (u64)256*sdist) / 8;
1585 cfqd->new_seek_mean = cfqd->new_seek_total / 256;
1589 * Don't allow the seek distance to get too large from the
1590 * odd fragment, pagein, etc
1592 if (cic->seek_samples <= 60) /* second&third seek */
1593 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*1024);
1595 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*64);
1597 cic->seek_samples = (7*cic->seek_samples + 256) / 8;
1598 cic->seek_total = (7*cic->seek_total + (u64)256*sdist) / 8;
1599 total = cic->seek_total + (cic->seek_samples/2);
1600 do_div(total, cic->seek_samples);
1601 cic->seek_mean = (sector_t)total;
1605 * Disable idle window if the process thinks too long or seeks so much that
1609 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1610 struct cfq_io_context *cic)
1612 int enable_idle = cfq_cfqq_idle_window(cfqq);
1614 if (!cic->ioc->task || !cfqd->cfq_slice_idle ||
1615 (cfqd->hw_tag && CIC_SEEKY(cic)))
1617 else if (sample_valid(cic->ttime_samples)) {
1618 if (cic->ttime_mean > cfqd->cfq_slice_idle)
1625 cfq_mark_cfqq_idle_window(cfqq);
1627 cfq_clear_cfqq_idle_window(cfqq);
1631 * Check if new_cfqq should preempt the currently active queue. Return 0 for
1632 * no or if we aren't sure, a 1 will cause a preempt.
1635 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
1638 struct cfq_queue *cfqq;
1640 cfqq = cfqd->active_queue;
1644 if (cfq_slice_used(cfqq))
1647 if (cfq_class_idle(new_cfqq))
1650 if (cfq_class_idle(cfqq))
1654 * if the new request is sync, but the currently running queue is
1655 * not, let the sync request have priority.
1657 if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
1661 * So both queues are sync. Let the new request get disk time if
1662 * it's a metadata request and the current queue is doing regular IO.
1664 if (rq_is_meta(rq) && !cfqq->meta_pending)
1667 if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
1671 * if this request is as-good as one we would expect from the
1672 * current cfqq, let it preempt
1674 if (cfq_rq_close(cfqd, rq))
1681 * cfqq preempts the active queue. if we allowed preempt with no slice left,
1682 * let it have half of its nominal slice.
1684 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1686 cfq_slice_expired(cfqd, 1, 1);
1689 * Put the new queue at the front of the of the current list,
1690 * so we know that it will be selected next.
1692 BUG_ON(!cfq_cfqq_on_rr(cfqq));
1694 cfq_service_tree_add(cfqd, cfqq, 1);
1696 cfqq->slice_end = 0;
1697 cfq_mark_cfqq_slice_new(cfqq);
1701 * Called when a new fs request (rq) is added (to cfqq). Check if there's
1702 * something we should do about it
1705 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1708 struct cfq_io_context *cic = RQ_CIC(rq);
1711 cfqq->meta_pending++;
1713 cfq_update_io_thinktime(cfqd, cic);
1714 cfq_update_io_seektime(cfqd, cic, rq);
1715 cfq_update_idle_window(cfqd, cfqq, cic);
1717 cic->last_request_pos = rq->sector + rq->nr_sectors;
1718 cfqq->last_request_pos = cic->last_request_pos;
1720 if (cfqq == cfqd->active_queue) {
1722 * if we are waiting for a request for this queue, let it rip
1723 * immediately and flag that we must not expire this queue
1726 if (cfq_cfqq_wait_request(cfqq)) {
1727 cfq_mark_cfqq_must_dispatch(cfqq);
1728 del_timer(&cfqd->idle_slice_timer);
1729 blk_start_queueing(cfqd->queue);
1731 } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
1733 * not the active queue - expire current slice if it is
1734 * idle and has expired it's mean thinktime or this new queue
1735 * has some old slice time left and is of higher priority
1737 cfq_preempt_queue(cfqd, cfqq);
1738 cfq_mark_cfqq_must_dispatch(cfqq);
1739 blk_start_queueing(cfqd->queue);
1743 static void cfq_insert_request(request_queue_t *q, struct request *rq)
1745 struct cfq_data *cfqd = q->elevator->elevator_data;
1746 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1748 cfq_init_prio_data(cfqq);
1752 list_add_tail(&rq->queuelist, &cfqq->fifo);
1754 cfq_rq_enqueued(cfqd, cfqq, rq);
1757 static void cfq_completed_request(request_queue_t *q, struct request *rq)
1759 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1760 struct cfq_data *cfqd = cfqq->cfqd;
1761 const int sync = rq_is_sync(rq);
1766 WARN_ON(!cfqd->rq_in_driver);
1767 WARN_ON(!cfqq->dispatched);
1768 cfqd->rq_in_driver--;
1771 if (!cfq_class_idle(cfqq))
1772 cfqd->last_end_request = now;
1775 RQ_CIC(rq)->last_end_request = now;
1778 * If this is the active queue, check if it needs to be expired,
1779 * or if we want to idle in case it has no pending requests.
1781 if (cfqd->active_queue == cfqq) {
1782 if (cfq_cfqq_slice_new(cfqq)) {
1783 cfq_set_prio_slice(cfqd, cfqq);
1784 cfq_clear_cfqq_slice_new(cfqq);
1786 if (cfq_slice_used(cfqq))
1787 cfq_slice_expired(cfqd, 0, 1);
1788 else if (sync && RB_EMPTY_ROOT(&cfqq->sort_list))
1789 cfq_arm_slice_timer(cfqd);
1792 if (!cfqd->rq_in_driver)
1793 cfq_schedule_dispatch(cfqd);
1797 * we temporarily boost lower priority queues if they are holding fs exclusive
1798 * resources. they are boosted to normal prio (CLASS_BE/4)
1800 static void cfq_prio_boost(struct cfq_queue *cfqq)
1802 if (has_fs_excl()) {
1804 * boost idle prio on transactions that would lock out other
1805 * users of the filesystem
1807 if (cfq_class_idle(cfqq))
1808 cfqq->ioprio_class = IOPRIO_CLASS_BE;
1809 if (cfqq->ioprio > IOPRIO_NORM)
1810 cfqq->ioprio = IOPRIO_NORM;
1813 * check if we need to unboost the queue
1815 if (cfqq->ioprio_class != cfqq->org_ioprio_class)
1816 cfqq->ioprio_class = cfqq->org_ioprio_class;
1817 if (cfqq->ioprio != cfqq->org_ioprio)
1818 cfqq->ioprio = cfqq->org_ioprio;
1822 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
1824 if ((cfq_cfqq_wait_request(cfqq) || cfq_cfqq_must_alloc(cfqq)) &&
1825 !cfq_cfqq_must_alloc_slice(cfqq)) {
1826 cfq_mark_cfqq_must_alloc_slice(cfqq);
1827 return ELV_MQUEUE_MUST;
1830 return ELV_MQUEUE_MAY;
1833 static int cfq_may_queue(request_queue_t *q, int rw)
1835 struct cfq_data *cfqd = q->elevator->elevator_data;
1836 struct task_struct *tsk = current;
1837 struct cfq_queue *cfqq;
1840 key = cfq_queue_pid(tsk, rw, rw & REQ_RW_SYNC);
1843 * don't force setup of a queue from here, as a call to may_queue
1844 * does not necessarily imply that a request actually will be queued.
1845 * so just lookup a possibly existing queue, or return 'may queue'
1848 cfqq = cfq_find_cfq_hash(cfqd, key, tsk->ioprio);
1850 cfq_init_prio_data(cfqq);
1851 cfq_prio_boost(cfqq);
1853 return __cfq_may_queue(cfqq);
1856 return ELV_MQUEUE_MAY;
1860 * queue lock held here
1862 static void cfq_put_request(struct request *rq)
1864 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1867 const int rw = rq_data_dir(rq);
1869 BUG_ON(!cfqq->allocated[rw]);
1870 cfqq->allocated[rw]--;
1872 put_io_context(RQ_CIC(rq)->ioc);
1874 rq->elevator_private = NULL;
1875 rq->elevator_private2 = NULL;
1877 cfq_put_queue(cfqq);
1882 * Allocate cfq data structures associated with this request.
1885 cfq_set_request(request_queue_t *q, struct request *rq, gfp_t gfp_mask)
1887 struct cfq_data *cfqd = q->elevator->elevator_data;
1888 struct task_struct *tsk = current;
1889 struct cfq_io_context *cic;
1890 const int rw = rq_data_dir(rq);
1891 const int is_sync = rq_is_sync(rq);
1892 pid_t key = cfq_queue_pid(tsk, rw, is_sync);
1893 struct cfq_queue *cfqq;
1894 unsigned long flags;
1896 might_sleep_if(gfp_mask & __GFP_WAIT);
1898 cic = cfq_get_io_context(cfqd, gfp_mask);
1900 spin_lock_irqsave(q->queue_lock, flags);
1905 if (!cic->cfqq[is_sync]) {
1906 cfqq = cfq_get_queue(cfqd, key, tsk, gfp_mask);
1910 cic->cfqq[is_sync] = cfqq;
1912 cfqq = cic->cfqq[is_sync];
1914 cfqq->allocated[rw]++;
1915 cfq_clear_cfqq_must_alloc(cfqq);
1916 atomic_inc(&cfqq->ref);
1918 spin_unlock_irqrestore(q->queue_lock, flags);
1920 rq->elevator_private = cic;
1921 rq->elevator_private2 = cfqq;
1926 put_io_context(cic->ioc);
1928 cfq_schedule_dispatch(cfqd);
1929 spin_unlock_irqrestore(q->queue_lock, flags);
1933 static void cfq_kick_queue(struct work_struct *work)
1935 struct cfq_data *cfqd =
1936 container_of(work, struct cfq_data, unplug_work);
1937 request_queue_t *q = cfqd->queue;
1938 unsigned long flags;
1940 spin_lock_irqsave(q->queue_lock, flags);
1941 blk_start_queueing(q);
1942 spin_unlock_irqrestore(q->queue_lock, flags);
1946 * Timer running if the active_queue is currently idling inside its time slice
1948 static void cfq_idle_slice_timer(unsigned long data)
1950 struct cfq_data *cfqd = (struct cfq_data *) data;
1951 struct cfq_queue *cfqq;
1952 unsigned long flags;
1955 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1957 if ((cfqq = cfqd->active_queue) != NULL) {
1963 if (cfq_slice_used(cfqq))
1967 * only expire and reinvoke request handler, if there are
1968 * other queues with pending requests
1970 if (!cfqd->busy_queues)
1974 * not expired and it has a request pending, let it dispatch
1976 if (!RB_EMPTY_ROOT(&cfqq->sort_list)) {
1977 cfq_mark_cfqq_must_dispatch(cfqq);
1982 cfq_slice_expired(cfqd, 0, timed_out);
1984 cfq_schedule_dispatch(cfqd);
1986 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1990 * Timer running if an idle class queue is waiting for service
1992 static void cfq_idle_class_timer(unsigned long data)
1994 struct cfq_data *cfqd = (struct cfq_data *) data;
1995 unsigned long flags, end;
1997 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2000 * race with a non-idle queue, reset timer
2002 end = cfqd->last_end_request + CFQ_IDLE_GRACE;
2003 if (!time_after_eq(jiffies, end))
2004 mod_timer(&cfqd->idle_class_timer, end);
2006 cfq_schedule_dispatch(cfqd);
2008 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2011 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
2013 del_timer_sync(&cfqd->idle_slice_timer);
2014 del_timer_sync(&cfqd->idle_class_timer);
2015 blk_sync_queue(cfqd->queue);
2018 static void cfq_exit_queue(elevator_t *e)
2020 struct cfq_data *cfqd = e->elevator_data;
2021 request_queue_t *q = cfqd->queue;
2023 cfq_shutdown_timer_wq(cfqd);
2025 spin_lock_irq(q->queue_lock);
2027 if (cfqd->active_queue)
2028 __cfq_slice_expired(cfqd, cfqd->active_queue, 0, 0);
2030 while (!list_empty(&cfqd->cic_list)) {
2031 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
2032 struct cfq_io_context,
2035 __cfq_exit_single_io_context(cfqd, cic);
2038 spin_unlock_irq(q->queue_lock);
2040 cfq_shutdown_timer_wq(cfqd);
2042 kfree(cfqd->cfq_hash);
2046 static void *cfq_init_queue(request_queue_t *q)
2048 struct cfq_data *cfqd;
2051 cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL, q->node);
2055 memset(cfqd, 0, sizeof(*cfqd));
2057 cfqd->service_tree = CFQ_RB_ROOT;
2058 INIT_LIST_HEAD(&cfqd->cic_list);
2060 cfqd->cfq_hash = kmalloc_node(sizeof(struct hlist_head) * CFQ_QHASH_ENTRIES, GFP_KERNEL, q->node);
2061 if (!cfqd->cfq_hash)
2064 for (i = 0; i < CFQ_QHASH_ENTRIES; i++)
2065 INIT_HLIST_HEAD(&cfqd->cfq_hash[i]);
2069 init_timer(&cfqd->idle_slice_timer);
2070 cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
2071 cfqd->idle_slice_timer.data = (unsigned long) cfqd;
2073 init_timer(&cfqd->idle_class_timer);
2074 cfqd->idle_class_timer.function = cfq_idle_class_timer;
2075 cfqd->idle_class_timer.data = (unsigned long) cfqd;
2077 INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
2079 cfqd->cfq_quantum = cfq_quantum;
2080 cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
2081 cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
2082 cfqd->cfq_back_max = cfq_back_max;
2083 cfqd->cfq_back_penalty = cfq_back_penalty;
2084 cfqd->cfq_slice[0] = cfq_slice_async;
2085 cfqd->cfq_slice[1] = cfq_slice_sync;
2086 cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
2087 cfqd->cfq_slice_idle = cfq_slice_idle;
2095 static void cfq_slab_kill(void)
2098 kmem_cache_destroy(cfq_pool);
2100 kmem_cache_destroy(cfq_ioc_pool);
2103 static int __init cfq_slab_setup(void)
2105 cfq_pool = kmem_cache_create("cfq_pool", sizeof(struct cfq_queue), 0, 0,
2110 cfq_ioc_pool = kmem_cache_create("cfq_ioc_pool",
2111 sizeof(struct cfq_io_context), 0, 0, NULL, NULL);
2122 * sysfs parts below -->
2125 cfq_var_show(unsigned int var, char *page)
2127 return sprintf(page, "%d\n", var);
2131 cfq_var_store(unsigned int *var, const char *page, size_t count)
2133 char *p = (char *) page;
2135 *var = simple_strtoul(p, &p, 10);
2139 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
2140 static ssize_t __FUNC(elevator_t *e, char *page) \
2142 struct cfq_data *cfqd = e->elevator_data; \
2143 unsigned int __data = __VAR; \
2145 __data = jiffies_to_msecs(__data); \
2146 return cfq_var_show(__data, (page)); \
2148 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
2149 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
2150 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
2151 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
2152 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
2153 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
2154 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
2155 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
2156 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
2157 #undef SHOW_FUNCTION
2159 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
2160 static ssize_t __FUNC(elevator_t *e, const char *page, size_t count) \
2162 struct cfq_data *cfqd = e->elevator_data; \
2163 unsigned int __data; \
2164 int ret = cfq_var_store(&__data, (page), count); \
2165 if (__data < (MIN)) \
2167 else if (__data > (MAX)) \
2170 *(__PTR) = msecs_to_jiffies(__data); \
2172 *(__PTR) = __data; \
2175 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
2176 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1, UINT_MAX, 1);
2177 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1, UINT_MAX, 1);
2178 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
2179 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1, UINT_MAX, 0);
2180 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
2181 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
2182 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
2183 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1, UINT_MAX, 0);
2184 #undef STORE_FUNCTION
2186 #define CFQ_ATTR(name) \
2187 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
2189 static struct elv_fs_entry cfq_attrs[] = {
2191 CFQ_ATTR(fifo_expire_sync),
2192 CFQ_ATTR(fifo_expire_async),
2193 CFQ_ATTR(back_seek_max),
2194 CFQ_ATTR(back_seek_penalty),
2195 CFQ_ATTR(slice_sync),
2196 CFQ_ATTR(slice_async),
2197 CFQ_ATTR(slice_async_rq),
2198 CFQ_ATTR(slice_idle),
2202 static struct elevator_type iosched_cfq = {
2204 .elevator_merge_fn = cfq_merge,
2205 .elevator_merged_fn = cfq_merged_request,
2206 .elevator_merge_req_fn = cfq_merged_requests,
2207 .elevator_allow_merge_fn = cfq_allow_merge,
2208 .elevator_dispatch_fn = cfq_dispatch_requests,
2209 .elevator_add_req_fn = cfq_insert_request,
2210 .elevator_activate_req_fn = cfq_activate_request,
2211 .elevator_deactivate_req_fn = cfq_deactivate_request,
2212 .elevator_queue_empty_fn = cfq_queue_empty,
2213 .elevator_completed_req_fn = cfq_completed_request,
2214 .elevator_former_req_fn = elv_rb_former_request,
2215 .elevator_latter_req_fn = elv_rb_latter_request,
2216 .elevator_set_req_fn = cfq_set_request,
2217 .elevator_put_req_fn = cfq_put_request,
2218 .elevator_may_queue_fn = cfq_may_queue,
2219 .elevator_init_fn = cfq_init_queue,
2220 .elevator_exit_fn = cfq_exit_queue,
2221 .trim = cfq_free_io_context,
2223 .elevator_attrs = cfq_attrs,
2224 .elevator_name = "cfq",
2225 .elevator_owner = THIS_MODULE,
2228 static int __init cfq_init(void)
2233 * could be 0 on HZ < 1000 setups
2235 if (!cfq_slice_async)
2236 cfq_slice_async = 1;
2237 if (!cfq_slice_idle)
2240 if (cfq_slab_setup())
2243 ret = elv_register(&iosched_cfq);
2250 static void __exit cfq_exit(void)
2252 DECLARE_COMPLETION_ONSTACK(all_gone);
2253 elv_unregister(&iosched_cfq);
2254 ioc_gone = &all_gone;
2255 /* ioc_gone's update must be visible before reading ioc_count */
2257 if (elv_ioc_count_read(ioc_count))
2258 wait_for_completion(ioc_gone);
2263 module_init(cfq_init);
2264 module_exit(cfq_exit);
2266 MODULE_AUTHOR("Jens Axboe");
2267 MODULE_LICENSE("GPL");
2268 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");