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>
18 /* max queue in one round of service */
19 static const int cfq_quantum = 4;
20 static const int cfq_fifo_expire[2] = { HZ / 4, HZ / 8 };
21 /* maximum backwards seek, in KiB */
22 static const int cfq_back_max = 16 * 1024;
23 /* penalty of a backwards seek */
24 static const int cfq_back_penalty = 2;
25 static const int cfq_slice_sync = HZ / 10;
26 static int cfq_slice_async = HZ / 25;
27 static const int cfq_slice_async_rq = 2;
28 static int cfq_slice_idle = HZ / 125;
31 * offset from end of service tree
33 #define CFQ_IDLE_DELAY (HZ / 5)
36 * below this threshold, we consider thinktime immediate
38 #define CFQ_MIN_TT (2)
40 #define CFQ_SLICE_SCALE (5)
43 ((struct cfq_io_context *) (rq)->elevator_private)
44 #define RQ_CFQQ(rq) ((rq)->elevator_private2)
46 static struct kmem_cache *cfq_pool;
47 static struct kmem_cache *cfq_ioc_pool;
49 static DEFINE_PER_CPU(unsigned long, ioc_count);
50 static struct completion *ioc_gone;
51 static DEFINE_SPINLOCK(ioc_gone_lock);
53 #define CFQ_PRIO_LISTS IOPRIO_BE_NR
54 #define cfq_class_idle(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
55 #define cfq_class_rt(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
60 #define sample_valid(samples) ((samples) > 80)
63 * Most of our rbtree usage is for sorting with min extraction, so
64 * if we cache the leftmost node we don't have to walk down the tree
65 * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should
66 * move this into the elevator for the rq sorting as well.
72 #define CFQ_RB_ROOT (struct cfq_rb_root) { RB_ROOT, NULL, }
75 * Per block device queue structure
78 struct request_queue *queue;
81 * rr list of queues with requests and the count of them
83 struct cfq_rb_root service_tree;
84 unsigned int busy_queues;
91 * idle window management
93 struct timer_list idle_slice_timer;
94 struct work_struct unplug_work;
96 struct cfq_queue *active_queue;
97 struct cfq_io_context *active_cic;
100 * async queue for each priority case
102 struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
103 struct cfq_queue *async_idle_cfqq;
105 sector_t last_position;
106 unsigned long last_end_request;
109 * tunables, see top of file
111 unsigned int cfq_quantum;
112 unsigned int cfq_fifo_expire[2];
113 unsigned int cfq_back_penalty;
114 unsigned int cfq_back_max;
115 unsigned int cfq_slice[2];
116 unsigned int cfq_slice_async_rq;
117 unsigned int cfq_slice_idle;
119 struct list_head cic_list;
123 * Per process-grouping structure
126 /* reference count */
128 /* various state flags, see below */
130 /* parent cfq_data */
131 struct cfq_data *cfqd;
132 /* service_tree member */
133 struct rb_node rb_node;
134 /* service_tree key */
135 unsigned long rb_key;
136 /* sorted list of pending requests */
137 struct rb_root sort_list;
138 /* if fifo isn't expired, next request to serve */
139 struct request *next_rq;
140 /* requests queued in sort_list */
142 /* currently allocated requests */
144 /* fifo list of requests in sort_list */
145 struct list_head fifo;
147 unsigned long slice_end;
150 /* pending metadata requests */
152 /* number of requests that are on the dispatch list or inside driver */
155 /* io prio of this group */
156 unsigned short ioprio, org_ioprio;
157 unsigned short ioprio_class, org_ioprio_class;
161 enum cfqq_state_flags {
162 CFQ_CFQQ_FLAG_on_rr = 0, /* on round-robin busy list */
163 CFQ_CFQQ_FLAG_wait_request, /* waiting for a request */
164 CFQ_CFQQ_FLAG_must_alloc, /* must be allowed rq alloc */
165 CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
166 CFQ_CFQQ_FLAG_must_dispatch, /* must dispatch, even if expired */
167 CFQ_CFQQ_FLAG_fifo_expire, /* FIFO checked in this slice */
168 CFQ_CFQQ_FLAG_idle_window, /* slice idling enabled */
169 CFQ_CFQQ_FLAG_prio_changed, /* task priority has changed */
170 CFQ_CFQQ_FLAG_queue_new, /* queue never been serviced */
171 CFQ_CFQQ_FLAG_slice_new, /* no requests dispatched in slice */
172 CFQ_CFQQ_FLAG_sync, /* synchronous queue */
175 #define CFQ_CFQQ_FNS(name) \
176 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \
178 (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name); \
180 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \
182 (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \
184 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \
186 return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \
190 CFQ_CFQQ_FNS(wait_request);
191 CFQ_CFQQ_FNS(must_alloc);
192 CFQ_CFQQ_FNS(must_alloc_slice);
193 CFQ_CFQQ_FNS(must_dispatch);
194 CFQ_CFQQ_FNS(fifo_expire);
195 CFQ_CFQQ_FNS(idle_window);
196 CFQ_CFQQ_FNS(prio_changed);
197 CFQ_CFQQ_FNS(queue_new);
198 CFQ_CFQQ_FNS(slice_new);
202 static void cfq_dispatch_insert(struct request_queue *, struct request *);
203 static struct cfq_queue *cfq_get_queue(struct cfq_data *, int,
204 struct io_context *, gfp_t);
205 static struct cfq_io_context *cfq_cic_lookup(struct cfq_data *,
206 struct io_context *);
208 static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_context *cic,
211 return cic->cfqq[!!is_sync];
214 static inline void cic_set_cfqq(struct cfq_io_context *cic,
215 struct cfq_queue *cfqq, int is_sync)
217 cic->cfqq[!!is_sync] = cfqq;
221 * We regard a request as SYNC, if it's either a read or has the SYNC bit
222 * set (in which case it could also be direct WRITE).
224 static inline int cfq_bio_sync(struct bio *bio)
226 if (bio_data_dir(bio) == READ || bio_sync(bio))
233 * scheduler run of queue, if there are requests pending and no one in the
234 * driver that will restart queueing
236 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
238 if (cfqd->busy_queues)
239 kblockd_schedule_work(&cfqd->unplug_work);
242 static int cfq_queue_empty(struct request_queue *q)
244 struct cfq_data *cfqd = q->elevator->elevator_data;
246 return !cfqd->busy_queues;
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 cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
392 root->left = rb_first(&root->rb);
395 return rb_entry(root->left, struct cfq_queue, rb_node);
400 static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
405 rb_erase(n, &root->rb);
410 * would be nice to take fifo expire time into account as well
412 static struct request *
413 cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
414 struct request *last)
416 struct rb_node *rbnext = rb_next(&last->rb_node);
417 struct rb_node *rbprev = rb_prev(&last->rb_node);
418 struct request *next = NULL, *prev = NULL;
420 BUG_ON(RB_EMPTY_NODE(&last->rb_node));
423 prev = rb_entry_rq(rbprev);
426 next = rb_entry_rq(rbnext);
428 rbnext = rb_first(&cfqq->sort_list);
429 if (rbnext && rbnext != &last->rb_node)
430 next = rb_entry_rq(rbnext);
433 return cfq_choose_req(cfqd, next, prev);
436 static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
437 struct cfq_queue *cfqq)
440 * just an approximation, should be ok.
442 return (cfqd->busy_queues - 1) * (cfq_prio_slice(cfqd, 1, 0) -
443 cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
447 * The cfqd->service_tree holds all pending cfq_queue's that have
448 * requests waiting to be processed. It is sorted in the order that
449 * we will service the queues.
451 static void cfq_service_tree_add(struct cfq_data *cfqd,
452 struct cfq_queue *cfqq, int add_front)
454 struct rb_node **p, *parent;
455 struct cfq_queue *__cfqq;
456 unsigned long rb_key;
459 if (cfq_class_idle(cfqq)) {
460 rb_key = CFQ_IDLE_DELAY;
461 parent = rb_last(&cfqd->service_tree.rb);
462 if (parent && parent != &cfqq->rb_node) {
463 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
464 rb_key += __cfqq->rb_key;
467 } else if (!add_front) {
468 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
469 rb_key += cfqq->slice_resid;
470 cfqq->slice_resid = 0;
474 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
476 * same position, nothing more to do
478 if (rb_key == cfqq->rb_key)
481 cfq_rb_erase(&cfqq->rb_node, &cfqd->service_tree);
486 p = &cfqd->service_tree.rb.rb_node;
491 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
494 * sort RT queues first, we always want to give
495 * preference to them. IDLE queues goes to the back.
496 * after that, sort on the next service time.
498 if (cfq_class_rt(cfqq) > cfq_class_rt(__cfqq))
500 else if (cfq_class_rt(cfqq) < cfq_class_rt(__cfqq))
502 else if (cfq_class_idle(cfqq) < cfq_class_idle(__cfqq))
504 else if (cfq_class_idle(cfqq) > cfq_class_idle(__cfqq))
506 else if (rb_key < __cfqq->rb_key)
511 if (n == &(*p)->rb_right)
518 cfqd->service_tree.left = &cfqq->rb_node;
520 cfqq->rb_key = rb_key;
521 rb_link_node(&cfqq->rb_node, parent, p);
522 rb_insert_color(&cfqq->rb_node, &cfqd->service_tree.rb);
526 * Update cfqq's position in the service tree.
528 static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
531 * Resorting requires the cfqq to be on the RR list already.
533 if (cfq_cfqq_on_rr(cfqq))
534 cfq_service_tree_add(cfqd, cfqq, 0);
538 * add to busy list of queues for service, trying to be fair in ordering
539 * the pending list according to last request service
541 static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
543 BUG_ON(cfq_cfqq_on_rr(cfqq));
544 cfq_mark_cfqq_on_rr(cfqq);
547 cfq_resort_rr_list(cfqd, cfqq);
551 * Called when the cfqq no longer has requests pending, remove it from
554 static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
556 BUG_ON(!cfq_cfqq_on_rr(cfqq));
557 cfq_clear_cfqq_on_rr(cfqq);
559 if (!RB_EMPTY_NODE(&cfqq->rb_node))
560 cfq_rb_erase(&cfqq->rb_node, &cfqd->service_tree);
562 BUG_ON(!cfqd->busy_queues);
567 * rb tree support functions
569 static void cfq_del_rq_rb(struct request *rq)
571 struct cfq_queue *cfqq = RQ_CFQQ(rq);
572 struct cfq_data *cfqd = cfqq->cfqd;
573 const int sync = rq_is_sync(rq);
575 BUG_ON(!cfqq->queued[sync]);
576 cfqq->queued[sync]--;
578 elv_rb_del(&cfqq->sort_list, rq);
580 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
581 cfq_del_cfqq_rr(cfqd, cfqq);
584 static void cfq_add_rq_rb(struct request *rq)
586 struct cfq_queue *cfqq = RQ_CFQQ(rq);
587 struct cfq_data *cfqd = cfqq->cfqd;
588 struct request *__alias;
590 cfqq->queued[rq_is_sync(rq)]++;
593 * looks a little odd, but the first insert might return an alias.
594 * if that happens, put the alias on the dispatch list
596 while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL)
597 cfq_dispatch_insert(cfqd->queue, __alias);
599 if (!cfq_cfqq_on_rr(cfqq))
600 cfq_add_cfqq_rr(cfqd, cfqq);
603 * check if this request is a better next-serve candidate
605 cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq);
606 BUG_ON(!cfqq->next_rq);
609 static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
611 elv_rb_del(&cfqq->sort_list, rq);
612 cfqq->queued[rq_is_sync(rq)]--;
616 static struct request *
617 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
619 struct task_struct *tsk = current;
620 struct cfq_io_context *cic;
621 struct cfq_queue *cfqq;
623 cic = cfq_cic_lookup(cfqd, tsk->io_context);
627 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
629 sector_t sector = bio->bi_sector + bio_sectors(bio);
631 return elv_rb_find(&cfqq->sort_list, sector);
637 static void cfq_activate_request(struct request_queue *q, struct request *rq)
639 struct cfq_data *cfqd = q->elevator->elevator_data;
641 cfqd->rq_in_driver++;
644 * If the depth is larger 1, it really could be queueing. But lets
645 * make the mark a little higher - idling could still be good for
646 * low queueing, and a low queueing number could also just indicate
647 * a SCSI mid layer like behaviour where limit+1 is often seen.
649 if (!cfqd->hw_tag && cfqd->rq_in_driver > 4)
652 cfqd->last_position = rq->hard_sector + rq->hard_nr_sectors;
655 static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
657 struct cfq_data *cfqd = q->elevator->elevator_data;
659 WARN_ON(!cfqd->rq_in_driver);
660 cfqd->rq_in_driver--;
663 static void cfq_remove_request(struct request *rq)
665 struct cfq_queue *cfqq = RQ_CFQQ(rq);
667 if (cfqq->next_rq == rq)
668 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
670 list_del_init(&rq->queuelist);
673 if (rq_is_meta(rq)) {
674 WARN_ON(!cfqq->meta_pending);
675 cfqq->meta_pending--;
679 static int cfq_merge(struct request_queue *q, struct request **req,
682 struct cfq_data *cfqd = q->elevator->elevator_data;
683 struct request *__rq;
685 __rq = cfq_find_rq_fmerge(cfqd, bio);
686 if (__rq && elv_rq_merge_ok(__rq, bio)) {
688 return ELEVATOR_FRONT_MERGE;
691 return ELEVATOR_NO_MERGE;
694 static void cfq_merged_request(struct request_queue *q, struct request *req,
697 if (type == ELEVATOR_FRONT_MERGE) {
698 struct cfq_queue *cfqq = RQ_CFQQ(req);
700 cfq_reposition_rq_rb(cfqq, req);
705 cfq_merged_requests(struct request_queue *q, struct request *rq,
706 struct request *next)
709 * reposition in fifo if next is older than rq
711 if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
712 time_before(next->start_time, rq->start_time))
713 list_move(&rq->queuelist, &next->queuelist);
715 cfq_remove_request(next);
718 static int cfq_allow_merge(struct request_queue *q, struct request *rq,
721 struct cfq_data *cfqd = q->elevator->elevator_data;
722 struct cfq_io_context *cic;
723 struct cfq_queue *cfqq;
726 * Disallow merge of a sync bio into an async request.
728 if (cfq_bio_sync(bio) && !rq_is_sync(rq))
732 * Lookup the cfqq that this bio will be queued with. Allow
733 * merge only if rq is queued there.
735 cic = cfq_cic_lookup(cfqd, current->io_context);
739 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
740 if (cfqq == RQ_CFQQ(rq))
746 static void __cfq_set_active_queue(struct cfq_data *cfqd,
747 struct cfq_queue *cfqq)
751 cfq_clear_cfqq_must_alloc_slice(cfqq);
752 cfq_clear_cfqq_fifo_expire(cfqq);
753 cfq_mark_cfqq_slice_new(cfqq);
754 cfq_clear_cfqq_queue_new(cfqq);
757 cfqd->active_queue = cfqq;
761 * current cfqq expired its slice (or was too idle), select new one
764 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
767 if (cfq_cfqq_wait_request(cfqq))
768 del_timer(&cfqd->idle_slice_timer);
770 cfq_clear_cfqq_must_dispatch(cfqq);
771 cfq_clear_cfqq_wait_request(cfqq);
774 * store what was left of this slice, if the queue idled/timed out
776 if (timed_out && !cfq_cfqq_slice_new(cfqq))
777 cfqq->slice_resid = cfqq->slice_end - jiffies;
779 cfq_resort_rr_list(cfqd, cfqq);
781 if (cfqq == cfqd->active_queue)
782 cfqd->active_queue = NULL;
784 if (cfqd->active_cic) {
785 put_io_context(cfqd->active_cic->ioc);
786 cfqd->active_cic = NULL;
790 static inline void cfq_slice_expired(struct cfq_data *cfqd, int timed_out)
792 struct cfq_queue *cfqq = cfqd->active_queue;
795 __cfq_slice_expired(cfqd, cfqq, timed_out);
799 * Get next queue for service. Unless we have a queue preemption,
800 * we'll simply select the first cfqq in the service tree.
802 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
804 if (RB_EMPTY_ROOT(&cfqd->service_tree.rb))
807 return cfq_rb_first(&cfqd->service_tree);
811 * Get and set a new active queue for service.
813 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd)
815 struct cfq_queue *cfqq;
817 cfqq = cfq_get_next_queue(cfqd);
818 __cfq_set_active_queue(cfqd, cfqq);
822 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
825 if (rq->sector >= cfqd->last_position)
826 return rq->sector - cfqd->last_position;
828 return cfqd->last_position - rq->sector;
831 static inline int cfq_rq_close(struct cfq_data *cfqd, struct request *rq)
833 struct cfq_io_context *cic = cfqd->active_cic;
835 if (!sample_valid(cic->seek_samples))
838 return cfq_dist_from_last(cfqd, rq) <= cic->seek_mean;
841 static int cfq_close_cooperator(struct cfq_data *cfq_data,
842 struct cfq_queue *cfqq)
845 * We should notice if some of the queues are cooperating, eg
846 * working closely on the same area of the disk. In that case,
847 * we can group them together and don't waste time idling.
852 #define CIC_SEEKY(cic) ((cic)->seek_mean > (8 * 1024))
854 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
856 struct cfq_queue *cfqq = cfqd->active_queue;
857 struct cfq_io_context *cic;
860 WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
861 WARN_ON(cfq_cfqq_slice_new(cfqq));
864 * idle is disabled, either manually or by past process history
866 if (!cfqd->cfq_slice_idle || !cfq_cfqq_idle_window(cfqq))
870 * task has exited, don't wait
872 cic = cfqd->active_cic;
873 if (!cic || !atomic_read(&cic->ioc->nr_tasks))
877 * See if this prio level has a good candidate
879 if (cfq_close_cooperator(cfqd, cfqq) &&
880 (sample_valid(cic->ttime_samples) && cic->ttime_mean > 2))
883 cfq_mark_cfqq_must_dispatch(cfqq);
884 cfq_mark_cfqq_wait_request(cfqq);
887 * we don't want to idle for seeks, but we do want to allow
888 * fair distribution of slice time for a process doing back-to-back
889 * seeks. so allow a little bit of time for him to submit a new rq
891 sl = cfqd->cfq_slice_idle;
892 if (sample_valid(cic->seek_samples) && CIC_SEEKY(cic))
893 sl = min(sl, msecs_to_jiffies(CFQ_MIN_TT));
895 mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
899 * Move request from internal lists to the request queue dispatch list.
901 static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
903 struct cfq_data *cfqd = q->elevator->elevator_data;
904 struct cfq_queue *cfqq = RQ_CFQQ(rq);
906 cfq_remove_request(rq);
908 elv_dispatch_sort(q, rq);
910 if (cfq_cfqq_sync(cfqq))
915 * return expired entry, or NULL to just start from scratch in rbtree
917 static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
919 struct cfq_data *cfqd = cfqq->cfqd;
923 if (cfq_cfqq_fifo_expire(cfqq))
926 cfq_mark_cfqq_fifo_expire(cfqq);
928 if (list_empty(&cfqq->fifo))
931 fifo = cfq_cfqq_sync(cfqq);
932 rq = rq_entry_fifo(cfqq->fifo.next);
934 if (time_before(jiffies, rq->start_time + cfqd->cfq_fifo_expire[fifo]))
941 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
943 const int base_rq = cfqd->cfq_slice_async_rq;
945 WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
947 return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
951 * Select a queue for service. If we have a current active queue,
952 * check whether to continue servicing it, or retrieve and set a new one.
954 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
956 struct cfq_queue *cfqq;
958 cfqq = cfqd->active_queue;
963 * The active queue has run out of time, expire it and select new.
965 if (cfq_slice_used(cfqq))
969 * The active queue has requests and isn't expired, allow it to
972 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
976 * No requests pending. If the active queue still has requests in
977 * flight or is idling for a new request, allow either of these
978 * conditions to happen (or time out) before selecting a new queue.
980 if (timer_pending(&cfqd->idle_slice_timer) ||
981 (cfqq->dispatched && cfq_cfqq_idle_window(cfqq))) {
987 cfq_slice_expired(cfqd, 0);
989 cfqq = cfq_set_active_queue(cfqd);
995 * Dispatch some requests from cfqq, moving them to the request queue
999 __cfq_dispatch_requests(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1004 BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
1010 * follow expired path, else get first next available
1012 rq = cfq_check_fifo(cfqq);
1017 * finally, insert request into driver dispatch list
1019 cfq_dispatch_insert(cfqd->queue, rq);
1023 if (!cfqd->active_cic) {
1024 atomic_inc(&RQ_CIC(rq)->ioc->refcount);
1025 cfqd->active_cic = RQ_CIC(rq);
1028 if (RB_EMPTY_ROOT(&cfqq->sort_list))
1031 } while (dispatched < max_dispatch);
1034 * expire an async queue immediately if it has used up its slice. idle
1035 * queue always expire after 1 dispatch round.
1037 if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
1038 dispatched >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
1039 cfq_class_idle(cfqq))) {
1040 cfqq->slice_end = jiffies + 1;
1041 cfq_slice_expired(cfqd, 0);
1047 static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
1051 while (cfqq->next_rq) {
1052 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
1056 BUG_ON(!list_empty(&cfqq->fifo));
1061 * Drain our current requests. Used for barriers and when switching
1062 * io schedulers on-the-fly.
1064 static int cfq_forced_dispatch(struct cfq_data *cfqd)
1066 struct cfq_queue *cfqq;
1069 while ((cfqq = cfq_rb_first(&cfqd->service_tree)) != NULL)
1070 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
1072 cfq_slice_expired(cfqd, 0);
1074 BUG_ON(cfqd->busy_queues);
1079 static int cfq_dispatch_requests(struct request_queue *q, int force)
1081 struct cfq_data *cfqd = q->elevator->elevator_data;
1082 struct cfq_queue *cfqq;
1085 if (!cfqd->busy_queues)
1088 if (unlikely(force))
1089 return cfq_forced_dispatch(cfqd);
1092 while ((cfqq = cfq_select_queue(cfqd)) != NULL) {
1095 max_dispatch = cfqd->cfq_quantum;
1096 if (cfq_class_idle(cfqq))
1099 if (cfqq->dispatched >= max_dispatch) {
1100 if (cfqd->busy_queues > 1)
1102 if (cfqq->dispatched >= 4 * max_dispatch)
1106 if (cfqd->sync_flight && !cfq_cfqq_sync(cfqq))
1109 cfq_clear_cfqq_must_dispatch(cfqq);
1110 cfq_clear_cfqq_wait_request(cfqq);
1111 del_timer(&cfqd->idle_slice_timer);
1113 dispatched += __cfq_dispatch_requests(cfqd, cfqq, max_dispatch);
1120 * task holds one reference to the queue, dropped when task exits. each rq
1121 * in-flight on this queue also holds a reference, dropped when rq is freed.
1123 * queue lock must be held here.
1125 static void cfq_put_queue(struct cfq_queue *cfqq)
1127 struct cfq_data *cfqd = cfqq->cfqd;
1129 BUG_ON(atomic_read(&cfqq->ref) <= 0);
1131 if (!atomic_dec_and_test(&cfqq->ref))
1134 BUG_ON(rb_first(&cfqq->sort_list));
1135 BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
1136 BUG_ON(cfq_cfqq_on_rr(cfqq));
1138 if (unlikely(cfqd->active_queue == cfqq)) {
1139 __cfq_slice_expired(cfqd, cfqq, 0);
1140 cfq_schedule_dispatch(cfqd);
1143 kmem_cache_free(cfq_pool, cfqq);
1147 * Must always be called with the rcu_read_lock() held
1150 __call_for_each_cic(struct io_context *ioc,
1151 void (*func)(struct io_context *, struct cfq_io_context *))
1153 struct cfq_io_context *cic;
1154 struct hlist_node *n;
1156 hlist_for_each_entry_rcu(cic, n, &ioc->cic_list, cic_list)
1161 * Call func for each cic attached to this ioc.
1164 call_for_each_cic(struct io_context *ioc,
1165 void (*func)(struct io_context *, struct cfq_io_context *))
1168 __call_for_each_cic(ioc, func);
1172 static void cfq_cic_free_rcu(struct rcu_head *head)
1174 struct cfq_io_context *cic;
1176 cic = container_of(head, struct cfq_io_context, rcu_head);
1178 kmem_cache_free(cfq_ioc_pool, cic);
1179 elv_ioc_count_dec(ioc_count);
1183 * CFQ scheduler is exiting, grab exit lock and check
1184 * the pending io context count. If it hits zero,
1185 * complete ioc_gone and set it back to NULL
1187 spin_lock(&ioc_gone_lock);
1188 if (ioc_gone && !elv_ioc_count_read(ioc_count)) {
1192 spin_unlock(&ioc_gone_lock);
1196 static void cfq_cic_free(struct cfq_io_context *cic)
1198 call_rcu(&cic->rcu_head, cfq_cic_free_rcu);
1201 static void cic_free_func(struct io_context *ioc, struct cfq_io_context *cic)
1203 unsigned long flags;
1205 BUG_ON(!cic->dead_key);
1207 spin_lock_irqsave(&ioc->lock, flags);
1208 radix_tree_delete(&ioc->radix_root, cic->dead_key);
1209 hlist_del_rcu(&cic->cic_list);
1210 spin_unlock_irqrestore(&ioc->lock, flags);
1216 * Must be called with rcu_read_lock() held or preemption otherwise disabled.
1217 * Only two callers of this - ->dtor() which is called with the rcu_read_lock(),
1218 * and ->trim() which is called with the task lock held
1220 static void cfq_free_io_context(struct io_context *ioc)
1223 * ioc->refcount is zero here, or we are called from elv_unregister(),
1224 * so no more cic's are allowed to be linked into this ioc. So it
1225 * should be ok to iterate over the known list, we will see all cic's
1226 * since no new ones are added.
1228 __call_for_each_cic(ioc, cic_free_func);
1231 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1233 if (unlikely(cfqq == cfqd->active_queue)) {
1234 __cfq_slice_expired(cfqd, cfqq, 0);
1235 cfq_schedule_dispatch(cfqd);
1238 cfq_put_queue(cfqq);
1241 static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
1242 struct cfq_io_context *cic)
1244 struct io_context *ioc = cic->ioc;
1246 list_del_init(&cic->queue_list);
1249 * Make sure key == NULL is seen for dead queues
1252 cic->dead_key = (unsigned long) cic->key;
1255 if (ioc->ioc_data == cic)
1256 rcu_assign_pointer(ioc->ioc_data, NULL);
1258 if (cic->cfqq[ASYNC]) {
1259 cfq_exit_cfqq(cfqd, cic->cfqq[ASYNC]);
1260 cic->cfqq[ASYNC] = NULL;
1263 if (cic->cfqq[SYNC]) {
1264 cfq_exit_cfqq(cfqd, cic->cfqq[SYNC]);
1265 cic->cfqq[SYNC] = NULL;
1269 static void cfq_exit_single_io_context(struct io_context *ioc,
1270 struct cfq_io_context *cic)
1272 struct cfq_data *cfqd = cic->key;
1275 struct request_queue *q = cfqd->queue;
1276 unsigned long flags;
1278 spin_lock_irqsave(q->queue_lock, flags);
1279 __cfq_exit_single_io_context(cfqd, cic);
1280 spin_unlock_irqrestore(q->queue_lock, flags);
1285 * The process that ioc belongs to has exited, we need to clean up
1286 * and put the internal structures we have that belongs to that process.
1288 static void cfq_exit_io_context(struct io_context *ioc)
1290 call_for_each_cic(ioc, cfq_exit_single_io_context);
1293 static struct cfq_io_context *
1294 cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
1296 struct cfq_io_context *cic;
1298 cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO,
1301 cic->last_end_request = jiffies;
1302 INIT_LIST_HEAD(&cic->queue_list);
1303 INIT_HLIST_NODE(&cic->cic_list);
1304 cic->dtor = cfq_free_io_context;
1305 cic->exit = cfq_exit_io_context;
1306 elv_ioc_count_inc(ioc_count);
1312 static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc)
1314 struct task_struct *tsk = current;
1317 if (!cfq_cfqq_prio_changed(cfqq))
1320 ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio);
1321 switch (ioprio_class) {
1323 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
1324 case IOPRIO_CLASS_NONE:
1326 * no prio set, inherit CPU scheduling settings
1328 cfqq->ioprio = task_nice_ioprio(tsk);
1329 cfqq->ioprio_class = task_nice_ioclass(tsk);
1331 case IOPRIO_CLASS_RT:
1332 cfqq->ioprio = task_ioprio(ioc);
1333 cfqq->ioprio_class = IOPRIO_CLASS_RT;
1335 case IOPRIO_CLASS_BE:
1336 cfqq->ioprio = task_ioprio(ioc);
1337 cfqq->ioprio_class = IOPRIO_CLASS_BE;
1339 case IOPRIO_CLASS_IDLE:
1340 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
1342 cfq_clear_cfqq_idle_window(cfqq);
1347 * keep track of original prio settings in case we have to temporarily
1348 * elevate the priority of this queue
1350 cfqq->org_ioprio = cfqq->ioprio;
1351 cfqq->org_ioprio_class = cfqq->ioprio_class;
1352 cfq_clear_cfqq_prio_changed(cfqq);
1355 static void changed_ioprio(struct io_context *ioc, struct cfq_io_context *cic)
1357 struct cfq_data *cfqd = cic->key;
1358 struct cfq_queue *cfqq;
1359 unsigned long flags;
1361 if (unlikely(!cfqd))
1364 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1366 cfqq = cic->cfqq[ASYNC];
1368 struct cfq_queue *new_cfqq;
1369 new_cfqq = cfq_get_queue(cfqd, ASYNC, cic->ioc, GFP_ATOMIC);
1371 cic->cfqq[ASYNC] = new_cfqq;
1372 cfq_put_queue(cfqq);
1376 cfqq = cic->cfqq[SYNC];
1378 cfq_mark_cfqq_prio_changed(cfqq);
1380 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1383 static void cfq_ioc_set_ioprio(struct io_context *ioc)
1385 call_for_each_cic(ioc, changed_ioprio);
1386 ioc->ioprio_changed = 0;
1389 static struct cfq_queue *
1390 cfq_find_alloc_queue(struct cfq_data *cfqd, int is_sync,
1391 struct io_context *ioc, gfp_t gfp_mask)
1393 struct cfq_queue *cfqq, *new_cfqq = NULL;
1394 struct cfq_io_context *cic;
1397 cic = cfq_cic_lookup(cfqd, ioc);
1398 /* cic always exists here */
1399 cfqq = cic_to_cfqq(cic, is_sync);
1405 } else if (gfp_mask & __GFP_WAIT) {
1407 * Inform the allocator of the fact that we will
1408 * just repeat this allocation if it fails, to allow
1409 * the allocator to do whatever it needs to attempt to
1412 spin_unlock_irq(cfqd->queue->queue_lock);
1413 new_cfqq = kmem_cache_alloc_node(cfq_pool,
1414 gfp_mask | __GFP_NOFAIL | __GFP_ZERO,
1416 spin_lock_irq(cfqd->queue->queue_lock);
1419 cfqq = kmem_cache_alloc_node(cfq_pool,
1420 gfp_mask | __GFP_ZERO,
1426 RB_CLEAR_NODE(&cfqq->rb_node);
1427 INIT_LIST_HEAD(&cfqq->fifo);
1429 atomic_set(&cfqq->ref, 0);
1432 cfq_mark_cfqq_prio_changed(cfqq);
1433 cfq_mark_cfqq_queue_new(cfqq);
1435 cfq_init_prio_data(cfqq, ioc);
1438 if (!cfq_class_idle(cfqq))
1439 cfq_mark_cfqq_idle_window(cfqq);
1440 cfq_mark_cfqq_sync(cfqq);
1445 kmem_cache_free(cfq_pool, new_cfqq);
1448 WARN_ON((gfp_mask & __GFP_WAIT) && !cfqq);
1452 static struct cfq_queue **
1453 cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
1455 switch (ioprio_class) {
1456 case IOPRIO_CLASS_RT:
1457 return &cfqd->async_cfqq[0][ioprio];
1458 case IOPRIO_CLASS_BE:
1459 return &cfqd->async_cfqq[1][ioprio];
1460 case IOPRIO_CLASS_IDLE:
1461 return &cfqd->async_idle_cfqq;
1467 static struct cfq_queue *
1468 cfq_get_queue(struct cfq_data *cfqd, int is_sync, struct io_context *ioc,
1471 const int ioprio = task_ioprio(ioc);
1472 const int ioprio_class = task_ioprio_class(ioc);
1473 struct cfq_queue **async_cfqq = NULL;
1474 struct cfq_queue *cfqq = NULL;
1477 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
1482 cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask);
1488 * pin the queue now that it's allocated, scheduler exit will prune it
1490 if (!is_sync && !(*async_cfqq)) {
1491 atomic_inc(&cfqq->ref);
1495 atomic_inc(&cfqq->ref);
1500 * We drop cfq io contexts lazily, so we may find a dead one.
1503 cfq_drop_dead_cic(struct cfq_data *cfqd, struct io_context *ioc,
1504 struct cfq_io_context *cic)
1506 unsigned long flags;
1508 WARN_ON(!list_empty(&cic->queue_list));
1510 spin_lock_irqsave(&ioc->lock, flags);
1512 BUG_ON(ioc->ioc_data == cic);
1514 radix_tree_delete(&ioc->radix_root, (unsigned long) cfqd);
1515 hlist_del_rcu(&cic->cic_list);
1516 spin_unlock_irqrestore(&ioc->lock, flags);
1521 static struct cfq_io_context *
1522 cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc)
1524 struct cfq_io_context *cic;
1525 unsigned long flags;
1534 * we maintain a last-hit cache, to avoid browsing over the tree
1536 cic = rcu_dereference(ioc->ioc_data);
1537 if (cic && cic->key == cfqd) {
1543 cic = radix_tree_lookup(&ioc->radix_root, (unsigned long) cfqd);
1547 /* ->key must be copied to avoid race with cfq_exit_queue() */
1550 cfq_drop_dead_cic(cfqd, ioc, cic);
1555 spin_lock_irqsave(&ioc->lock, flags);
1556 rcu_assign_pointer(ioc->ioc_data, cic);
1557 spin_unlock_irqrestore(&ioc->lock, flags);
1565 * Add cic into ioc, using cfqd as the search key. This enables us to lookup
1566 * the process specific cfq io context when entered from the block layer.
1567 * Also adds the cic to a per-cfqd list, used when this queue is removed.
1569 static int cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
1570 struct cfq_io_context *cic, gfp_t gfp_mask)
1572 unsigned long flags;
1575 ret = radix_tree_preload(gfp_mask);
1580 spin_lock_irqsave(&ioc->lock, flags);
1581 ret = radix_tree_insert(&ioc->radix_root,
1582 (unsigned long) cfqd, cic);
1584 hlist_add_head_rcu(&cic->cic_list, &ioc->cic_list);
1585 spin_unlock_irqrestore(&ioc->lock, flags);
1587 radix_tree_preload_end();
1590 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1591 list_add(&cic->queue_list, &cfqd->cic_list);
1592 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1597 printk(KERN_ERR "cfq: cic link failed!\n");
1603 * Setup general io context and cfq io context. There can be several cfq
1604 * io contexts per general io context, if this process is doing io to more
1605 * than one device managed by cfq.
1607 static struct cfq_io_context *
1608 cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
1610 struct io_context *ioc = NULL;
1611 struct cfq_io_context *cic;
1613 might_sleep_if(gfp_mask & __GFP_WAIT);
1615 ioc = get_io_context(gfp_mask, cfqd->queue->node);
1619 cic = cfq_cic_lookup(cfqd, ioc);
1623 cic = cfq_alloc_io_context(cfqd, gfp_mask);
1627 if (cfq_cic_link(cfqd, ioc, cic, gfp_mask))
1631 smp_read_barrier_depends();
1632 if (unlikely(ioc->ioprio_changed))
1633 cfq_ioc_set_ioprio(ioc);
1639 put_io_context(ioc);
1644 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
1646 unsigned long elapsed = jiffies - cic->last_end_request;
1647 unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
1649 cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
1650 cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
1651 cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
1655 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_io_context *cic,
1661 if (cic->last_request_pos < rq->sector)
1662 sdist = rq->sector - cic->last_request_pos;
1664 sdist = cic->last_request_pos - rq->sector;
1667 * Don't allow the seek distance to get too large from the
1668 * odd fragment, pagein, etc
1670 if (cic->seek_samples <= 60) /* second&third seek */
1671 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*1024);
1673 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*64);
1675 cic->seek_samples = (7*cic->seek_samples + 256) / 8;
1676 cic->seek_total = (7*cic->seek_total + (u64)256*sdist) / 8;
1677 total = cic->seek_total + (cic->seek_samples/2);
1678 do_div(total, cic->seek_samples);
1679 cic->seek_mean = (sector_t)total;
1683 * Disable idle window if the process thinks too long or seeks so much that
1687 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1688 struct cfq_io_context *cic)
1693 * Don't idle for async or idle io prio class
1695 if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
1698 enable_idle = cfq_cfqq_idle_window(cfqq);
1700 if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle ||
1701 (cfqd->hw_tag && CIC_SEEKY(cic)))
1703 else if (sample_valid(cic->ttime_samples)) {
1704 if (cic->ttime_mean > cfqd->cfq_slice_idle)
1711 cfq_mark_cfqq_idle_window(cfqq);
1713 cfq_clear_cfqq_idle_window(cfqq);
1717 * Check if new_cfqq should preempt the currently active queue. Return 0 for
1718 * no or if we aren't sure, a 1 will cause a preempt.
1721 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
1724 struct cfq_queue *cfqq;
1726 cfqq = cfqd->active_queue;
1730 if (cfq_slice_used(cfqq))
1733 if (cfq_class_idle(new_cfqq))
1736 if (cfq_class_idle(cfqq))
1740 * if the new request is sync, but the currently running queue is
1741 * not, let the sync request have priority.
1743 if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
1747 * So both queues are sync. Let the new request get disk time if
1748 * it's a metadata request and the current queue is doing regular IO.
1750 if (rq_is_meta(rq) && !cfqq->meta_pending)
1753 if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
1757 * if this request is as-good as one we would expect from the
1758 * current cfqq, let it preempt
1760 if (cfq_rq_close(cfqd, rq))
1767 * cfqq preempts the active queue. if we allowed preempt with no slice left,
1768 * let it have half of its nominal slice.
1770 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1772 cfq_slice_expired(cfqd, 1);
1775 * Put the new queue at the front of the of the current list,
1776 * so we know that it will be selected next.
1778 BUG_ON(!cfq_cfqq_on_rr(cfqq));
1780 cfq_service_tree_add(cfqd, cfqq, 1);
1782 cfqq->slice_end = 0;
1783 cfq_mark_cfqq_slice_new(cfqq);
1787 * Called when a new fs request (rq) is added (to cfqq). Check if there's
1788 * something we should do about it
1791 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1794 struct cfq_io_context *cic = RQ_CIC(rq);
1797 cfqq->meta_pending++;
1799 cfq_update_io_thinktime(cfqd, cic);
1800 cfq_update_io_seektime(cfqd, cic, rq);
1801 cfq_update_idle_window(cfqd, cfqq, cic);
1803 cic->last_request_pos = rq->sector + rq->nr_sectors;
1805 if (cfqq == cfqd->active_queue) {
1807 * if we are waiting for a request for this queue, let it rip
1808 * immediately and flag that we must not expire this queue
1811 if (cfq_cfqq_wait_request(cfqq)) {
1812 cfq_mark_cfqq_must_dispatch(cfqq);
1813 del_timer(&cfqd->idle_slice_timer);
1814 blk_start_queueing(cfqd->queue);
1816 } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
1818 * not the active queue - expire current slice if it is
1819 * idle and has expired it's mean thinktime or this new queue
1820 * has some old slice time left and is of higher priority
1822 cfq_preempt_queue(cfqd, cfqq);
1823 cfq_mark_cfqq_must_dispatch(cfqq);
1824 blk_start_queueing(cfqd->queue);
1828 static void cfq_insert_request(struct request_queue *q, struct request *rq)
1830 struct cfq_data *cfqd = q->elevator->elevator_data;
1831 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1833 cfq_init_prio_data(cfqq, RQ_CIC(rq)->ioc);
1837 list_add_tail(&rq->queuelist, &cfqq->fifo);
1839 cfq_rq_enqueued(cfqd, cfqq, rq);
1842 static void cfq_completed_request(struct request_queue *q, struct request *rq)
1844 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1845 struct cfq_data *cfqd = cfqq->cfqd;
1846 const int sync = rq_is_sync(rq);
1851 WARN_ON(!cfqd->rq_in_driver);
1852 WARN_ON(!cfqq->dispatched);
1853 cfqd->rq_in_driver--;
1856 if (cfq_cfqq_sync(cfqq))
1857 cfqd->sync_flight--;
1859 if (!cfq_class_idle(cfqq))
1860 cfqd->last_end_request = now;
1863 RQ_CIC(rq)->last_end_request = now;
1866 * If this is the active queue, check if it needs to be expired,
1867 * or if we want to idle in case it has no pending requests.
1869 if (cfqd->active_queue == cfqq) {
1870 if (cfq_cfqq_slice_new(cfqq)) {
1871 cfq_set_prio_slice(cfqd, cfqq);
1872 cfq_clear_cfqq_slice_new(cfqq);
1874 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
1875 cfq_slice_expired(cfqd, 1);
1876 else if (sync && RB_EMPTY_ROOT(&cfqq->sort_list))
1877 cfq_arm_slice_timer(cfqd);
1880 if (!cfqd->rq_in_driver)
1881 cfq_schedule_dispatch(cfqd);
1885 * we temporarily boost lower priority queues if they are holding fs exclusive
1886 * resources. they are boosted to normal prio (CLASS_BE/4)
1888 static void cfq_prio_boost(struct cfq_queue *cfqq)
1890 if (has_fs_excl()) {
1892 * boost idle prio on transactions that would lock out other
1893 * users of the filesystem
1895 if (cfq_class_idle(cfqq))
1896 cfqq->ioprio_class = IOPRIO_CLASS_BE;
1897 if (cfqq->ioprio > IOPRIO_NORM)
1898 cfqq->ioprio = IOPRIO_NORM;
1901 * check if we need to unboost the queue
1903 if (cfqq->ioprio_class != cfqq->org_ioprio_class)
1904 cfqq->ioprio_class = cfqq->org_ioprio_class;
1905 if (cfqq->ioprio != cfqq->org_ioprio)
1906 cfqq->ioprio = cfqq->org_ioprio;
1910 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
1912 if ((cfq_cfqq_wait_request(cfqq) || cfq_cfqq_must_alloc(cfqq)) &&
1913 !cfq_cfqq_must_alloc_slice(cfqq)) {
1914 cfq_mark_cfqq_must_alloc_slice(cfqq);
1915 return ELV_MQUEUE_MUST;
1918 return ELV_MQUEUE_MAY;
1921 static int cfq_may_queue(struct request_queue *q, int rw)
1923 struct cfq_data *cfqd = q->elevator->elevator_data;
1924 struct task_struct *tsk = current;
1925 struct cfq_io_context *cic;
1926 struct cfq_queue *cfqq;
1929 * don't force setup of a queue from here, as a call to may_queue
1930 * does not necessarily imply that a request actually will be queued.
1931 * so just lookup a possibly existing queue, or return 'may queue'
1934 cic = cfq_cic_lookup(cfqd, tsk->io_context);
1936 return ELV_MQUEUE_MAY;
1938 cfqq = cic_to_cfqq(cic, rw & REQ_RW_SYNC);
1940 cfq_init_prio_data(cfqq, cic->ioc);
1941 cfq_prio_boost(cfqq);
1943 return __cfq_may_queue(cfqq);
1946 return ELV_MQUEUE_MAY;
1950 * queue lock held here
1952 static void cfq_put_request(struct request *rq)
1954 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1957 const int rw = rq_data_dir(rq);
1959 BUG_ON(!cfqq->allocated[rw]);
1960 cfqq->allocated[rw]--;
1962 put_io_context(RQ_CIC(rq)->ioc);
1964 rq->elevator_private = NULL;
1965 rq->elevator_private2 = NULL;
1967 cfq_put_queue(cfqq);
1972 * Allocate cfq data structures associated with this request.
1975 cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
1977 struct cfq_data *cfqd = q->elevator->elevator_data;
1978 struct cfq_io_context *cic;
1979 const int rw = rq_data_dir(rq);
1980 const int is_sync = rq_is_sync(rq);
1981 struct cfq_queue *cfqq;
1982 unsigned long flags;
1984 might_sleep_if(gfp_mask & __GFP_WAIT);
1986 cic = cfq_get_io_context(cfqd, gfp_mask);
1988 spin_lock_irqsave(q->queue_lock, flags);
1993 cfqq = cic_to_cfqq(cic, is_sync);
1995 cfqq = cfq_get_queue(cfqd, is_sync, cic->ioc, gfp_mask);
2000 cic_set_cfqq(cic, cfqq, is_sync);
2003 cfqq->allocated[rw]++;
2004 cfq_clear_cfqq_must_alloc(cfqq);
2005 atomic_inc(&cfqq->ref);
2007 spin_unlock_irqrestore(q->queue_lock, flags);
2009 rq->elevator_private = cic;
2010 rq->elevator_private2 = cfqq;
2015 put_io_context(cic->ioc);
2017 cfq_schedule_dispatch(cfqd);
2018 spin_unlock_irqrestore(q->queue_lock, flags);
2022 static void cfq_kick_queue(struct work_struct *work)
2024 struct cfq_data *cfqd =
2025 container_of(work, struct cfq_data, unplug_work);
2026 struct request_queue *q = cfqd->queue;
2027 unsigned long flags;
2029 spin_lock_irqsave(q->queue_lock, flags);
2030 blk_start_queueing(q);
2031 spin_unlock_irqrestore(q->queue_lock, flags);
2035 * Timer running if the active_queue is currently idling inside its time slice
2037 static void cfq_idle_slice_timer(unsigned long data)
2039 struct cfq_data *cfqd = (struct cfq_data *) data;
2040 struct cfq_queue *cfqq;
2041 unsigned long flags;
2044 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2046 cfqq = cfqd->active_queue;
2053 if (cfq_slice_used(cfqq))
2057 * only expire and reinvoke request handler, if there are
2058 * other queues with pending requests
2060 if (!cfqd->busy_queues)
2064 * not expired and it has a request pending, let it dispatch
2066 if (!RB_EMPTY_ROOT(&cfqq->sort_list)) {
2067 cfq_mark_cfqq_must_dispatch(cfqq);
2072 cfq_slice_expired(cfqd, timed_out);
2074 cfq_schedule_dispatch(cfqd);
2076 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2079 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
2081 del_timer_sync(&cfqd->idle_slice_timer);
2082 kblockd_flush_work(&cfqd->unplug_work);
2085 static void cfq_put_async_queues(struct cfq_data *cfqd)
2089 for (i = 0; i < IOPRIO_BE_NR; i++) {
2090 if (cfqd->async_cfqq[0][i])
2091 cfq_put_queue(cfqd->async_cfqq[0][i]);
2092 if (cfqd->async_cfqq[1][i])
2093 cfq_put_queue(cfqd->async_cfqq[1][i]);
2096 if (cfqd->async_idle_cfqq)
2097 cfq_put_queue(cfqd->async_idle_cfqq);
2100 static void cfq_exit_queue(elevator_t *e)
2102 struct cfq_data *cfqd = e->elevator_data;
2103 struct request_queue *q = cfqd->queue;
2105 cfq_shutdown_timer_wq(cfqd);
2107 spin_lock_irq(q->queue_lock);
2109 if (cfqd->active_queue)
2110 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
2112 while (!list_empty(&cfqd->cic_list)) {
2113 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
2114 struct cfq_io_context,
2117 __cfq_exit_single_io_context(cfqd, cic);
2120 cfq_put_async_queues(cfqd);
2122 spin_unlock_irq(q->queue_lock);
2124 cfq_shutdown_timer_wq(cfqd);
2129 static void *cfq_init_queue(struct request_queue *q)
2131 struct cfq_data *cfqd;
2133 cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
2137 cfqd->service_tree = CFQ_RB_ROOT;
2138 INIT_LIST_HEAD(&cfqd->cic_list);
2142 init_timer(&cfqd->idle_slice_timer);
2143 cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
2144 cfqd->idle_slice_timer.data = (unsigned long) cfqd;
2146 INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
2148 cfqd->last_end_request = jiffies;
2149 cfqd->cfq_quantum = cfq_quantum;
2150 cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
2151 cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
2152 cfqd->cfq_back_max = cfq_back_max;
2153 cfqd->cfq_back_penalty = cfq_back_penalty;
2154 cfqd->cfq_slice[0] = cfq_slice_async;
2155 cfqd->cfq_slice[1] = cfq_slice_sync;
2156 cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
2157 cfqd->cfq_slice_idle = cfq_slice_idle;
2162 static void cfq_slab_kill(void)
2165 * Caller already ensured that pending RCU callbacks are completed,
2166 * so we should have no busy allocations at this point.
2169 kmem_cache_destroy(cfq_pool);
2171 kmem_cache_destroy(cfq_ioc_pool);
2174 static int __init cfq_slab_setup(void)
2176 cfq_pool = KMEM_CACHE(cfq_queue, 0);
2180 cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0);
2191 * sysfs parts below -->
2194 cfq_var_show(unsigned int var, char *page)
2196 return sprintf(page, "%d\n", var);
2200 cfq_var_store(unsigned int *var, const char *page, size_t count)
2202 char *p = (char *) page;
2204 *var = simple_strtoul(p, &p, 10);
2208 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
2209 static ssize_t __FUNC(elevator_t *e, char *page) \
2211 struct cfq_data *cfqd = e->elevator_data; \
2212 unsigned int __data = __VAR; \
2214 __data = jiffies_to_msecs(__data); \
2215 return cfq_var_show(__data, (page)); \
2217 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
2218 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
2219 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
2220 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
2221 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
2222 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
2223 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
2224 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
2225 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
2226 #undef SHOW_FUNCTION
2228 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
2229 static ssize_t __FUNC(elevator_t *e, const char *page, size_t count) \
2231 struct cfq_data *cfqd = e->elevator_data; \
2232 unsigned int __data; \
2233 int ret = cfq_var_store(&__data, (page), count); \
2234 if (__data < (MIN)) \
2236 else if (__data > (MAX)) \
2239 *(__PTR) = msecs_to_jiffies(__data); \
2241 *(__PTR) = __data; \
2244 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
2245 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
2247 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
2249 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
2250 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
2252 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
2253 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
2254 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
2255 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
2257 #undef STORE_FUNCTION
2259 #define CFQ_ATTR(name) \
2260 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
2262 static struct elv_fs_entry cfq_attrs[] = {
2264 CFQ_ATTR(fifo_expire_sync),
2265 CFQ_ATTR(fifo_expire_async),
2266 CFQ_ATTR(back_seek_max),
2267 CFQ_ATTR(back_seek_penalty),
2268 CFQ_ATTR(slice_sync),
2269 CFQ_ATTR(slice_async),
2270 CFQ_ATTR(slice_async_rq),
2271 CFQ_ATTR(slice_idle),
2275 static struct elevator_type iosched_cfq = {
2277 .elevator_merge_fn = cfq_merge,
2278 .elevator_merged_fn = cfq_merged_request,
2279 .elevator_merge_req_fn = cfq_merged_requests,
2280 .elevator_allow_merge_fn = cfq_allow_merge,
2281 .elevator_dispatch_fn = cfq_dispatch_requests,
2282 .elevator_add_req_fn = cfq_insert_request,
2283 .elevator_activate_req_fn = cfq_activate_request,
2284 .elevator_deactivate_req_fn = cfq_deactivate_request,
2285 .elevator_queue_empty_fn = cfq_queue_empty,
2286 .elevator_completed_req_fn = cfq_completed_request,
2287 .elevator_former_req_fn = elv_rb_former_request,
2288 .elevator_latter_req_fn = elv_rb_latter_request,
2289 .elevator_set_req_fn = cfq_set_request,
2290 .elevator_put_req_fn = cfq_put_request,
2291 .elevator_may_queue_fn = cfq_may_queue,
2292 .elevator_init_fn = cfq_init_queue,
2293 .elevator_exit_fn = cfq_exit_queue,
2294 .trim = cfq_free_io_context,
2296 .elevator_attrs = cfq_attrs,
2297 .elevator_name = "cfq",
2298 .elevator_owner = THIS_MODULE,
2301 static int __init cfq_init(void)
2304 * could be 0 on HZ < 1000 setups
2306 if (!cfq_slice_async)
2307 cfq_slice_async = 1;
2308 if (!cfq_slice_idle)
2311 if (cfq_slab_setup())
2314 elv_register(&iosched_cfq);
2319 static void __exit cfq_exit(void)
2321 DECLARE_COMPLETION_ONSTACK(all_gone);
2322 elv_unregister(&iosched_cfq);
2323 ioc_gone = &all_gone;
2324 /* ioc_gone's update must be visible before reading ioc_count */
2328 * this also protects us from entering cfq_slab_kill() with
2329 * pending RCU callbacks
2331 if (elv_ioc_count_read(ioc_count))
2332 wait_for_completion(&all_gone);
2336 module_init(cfq_init);
2337 module_exit(cfq_exit);
2339 MODULE_AUTHOR("Jens Axboe");
2340 MODULE_LICENSE("GPL");
2341 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");