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;
110 struct timer_list idle_class_timer;
112 sector_t last_position;
113 unsigned long last_end_request;
116 * tunables, see top of file
118 unsigned int cfq_quantum;
119 unsigned int cfq_fifo_expire[2];
120 unsigned int cfq_back_penalty;
121 unsigned int cfq_back_max;
122 unsigned int cfq_slice[2];
123 unsigned int cfq_slice_async_rq;
124 unsigned int cfq_slice_idle;
126 struct list_head cic_list;
128 sector_t new_seek_mean;
133 * Per process-grouping structure
136 /* reference count */
138 /* parent cfq_data */
139 struct cfq_data *cfqd;
140 /* cfqq lookup hash */
141 struct hlist_node cfq_hash;
144 /* service_tree member */
145 struct rb_node rb_node;
146 /* service_tree key */
147 unsigned long rb_key;
148 /* sorted list of pending requests */
149 struct rb_root sort_list;
150 /* if fifo isn't expired, next request to serve */
151 struct request *next_rq;
152 /* requests queued in sort_list */
154 /* currently allocated requests */
156 /* pending metadata requests */
158 /* fifo list of requests in sort_list */
159 struct list_head fifo;
161 unsigned long slice_end;
164 /* number of requests that are on the dispatch list or inside driver */
167 /* io prio of this group */
168 unsigned short ioprio, org_ioprio;
169 unsigned short ioprio_class, org_ioprio_class;
171 /* various state flags, see below */
174 sector_t last_request_pos;
177 enum cfqq_state_flags {
178 CFQ_CFQQ_FLAG_on_rr = 0, /* on round-robin busy list */
179 CFQ_CFQQ_FLAG_wait_request, /* waiting for a request */
180 CFQ_CFQQ_FLAG_must_alloc, /* must be allowed rq alloc */
181 CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
182 CFQ_CFQQ_FLAG_must_dispatch, /* must dispatch, even if expired */
183 CFQ_CFQQ_FLAG_fifo_expire, /* FIFO checked in this slice */
184 CFQ_CFQQ_FLAG_idle_window, /* slice idling enabled */
185 CFQ_CFQQ_FLAG_prio_changed, /* task priority has changed */
186 CFQ_CFQQ_FLAG_queue_new, /* queue never been serviced */
187 CFQ_CFQQ_FLAG_slice_new, /* no requests dispatched in slice */
190 #define CFQ_CFQQ_FNS(name) \
191 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \
193 cfqq->flags |= (1 << CFQ_CFQQ_FLAG_##name); \
195 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \
197 cfqq->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \
199 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \
201 return (cfqq->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \
205 CFQ_CFQQ_FNS(wait_request);
206 CFQ_CFQQ_FNS(must_alloc);
207 CFQ_CFQQ_FNS(must_alloc_slice);
208 CFQ_CFQQ_FNS(must_dispatch);
209 CFQ_CFQQ_FNS(fifo_expire);
210 CFQ_CFQQ_FNS(idle_window);
211 CFQ_CFQQ_FNS(prio_changed);
212 CFQ_CFQQ_FNS(queue_new);
213 CFQ_CFQQ_FNS(slice_new);
216 static struct cfq_queue *cfq_find_cfq_hash(struct cfq_data *, unsigned int, unsigned short);
217 static void cfq_dispatch_insert(request_queue_t *, struct request *);
218 static struct cfq_queue *cfq_get_queue(struct cfq_data *, unsigned int, struct task_struct *, gfp_t);
221 * scheduler run of queue, if there are requests pending and no one in the
222 * driver that will restart queueing
224 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
226 if (cfqd->busy_queues)
227 kblockd_schedule_work(&cfqd->unplug_work);
230 static int cfq_queue_empty(request_queue_t *q)
232 struct cfq_data *cfqd = q->elevator->elevator_data;
234 return !cfqd->busy_queues;
237 static inline pid_t cfq_queue_pid(struct task_struct *task, int rw, int is_sync)
240 * Use the per-process queue, for read requests and syncronous writes
242 if (!(rw & REQ_RW) || is_sync)
245 return CFQ_KEY_ASYNC;
249 * Scale schedule slice based on io priority. Use the sync time slice only
250 * if a queue is marked sync and has sync io queued. A sync queue with async
251 * io only, should not get full sync slice length.
253 static inline int cfq_prio_slice(struct cfq_data *cfqd, int sync,
256 const int base_slice = cfqd->cfq_slice[sync];
258 WARN_ON(prio >= IOPRIO_BE_NR);
260 return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
264 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
266 return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
270 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
272 cfqq->slice_end = cfq_prio_to_slice(cfqd, cfqq) + jiffies;
276 * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
277 * isn't valid until the first request from the dispatch is activated
278 * and the slice time set.
280 static inline int cfq_slice_used(struct cfq_queue *cfqq)
282 if (cfq_cfqq_slice_new(cfqq))
284 if (time_before(jiffies, cfqq->slice_end))
291 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
292 * We choose the request that is closest to the head right now. Distance
293 * behind the head is penalized and only allowed to a certain extent.
295 static struct request *
296 cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2)
298 sector_t last, s1, s2, d1 = 0, d2 = 0;
299 unsigned long back_max;
300 #define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */
301 #define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */
302 unsigned wrap = 0; /* bit mask: requests behind the disk head? */
304 if (rq1 == NULL || rq1 == rq2)
309 if (rq_is_sync(rq1) && !rq_is_sync(rq2))
311 else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
313 if (rq_is_meta(rq1) && !rq_is_meta(rq2))
315 else if (rq_is_meta(rq2) && !rq_is_meta(rq1))
321 last = cfqd->last_position;
324 * by definition, 1KiB is 2 sectors
326 back_max = cfqd->cfq_back_max * 2;
329 * Strict one way elevator _except_ in the case where we allow
330 * short backward seeks which are biased as twice the cost of a
331 * similar forward seek.
335 else if (s1 + back_max >= last)
336 d1 = (last - s1) * cfqd->cfq_back_penalty;
338 wrap |= CFQ_RQ1_WRAP;
342 else if (s2 + back_max >= last)
343 d2 = (last - s2) * cfqd->cfq_back_penalty;
345 wrap |= CFQ_RQ2_WRAP;
347 /* Found required data */
350 * By doing switch() on the bit mask "wrap" we avoid having to
351 * check two variables for all permutations: --> faster!
354 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
370 case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
373 * Since both rqs are wrapped,
374 * start with the one that's further behind head
375 * (--> only *one* back seek required),
376 * since back seek takes more time than forward.
386 * The below is leftmost cache rbtree addon
388 static struct rb_node *cfq_rb_first(struct cfq_rb_root *root)
391 root->left = rb_first(&root->rb);
396 static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
401 rb_erase(n, &root->rb);
406 * would be nice to take fifo expire time into account as well
408 static struct request *
409 cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
410 struct request *last)
412 struct rb_node *rbnext = rb_next(&last->rb_node);
413 struct rb_node *rbprev = rb_prev(&last->rb_node);
414 struct request *next = NULL, *prev = NULL;
416 BUG_ON(RB_EMPTY_NODE(&last->rb_node));
419 prev = rb_entry_rq(rbprev);
422 next = rb_entry_rq(rbnext);
424 rbnext = rb_first(&cfqq->sort_list);
425 if (rbnext && rbnext != &last->rb_node)
426 next = rb_entry_rq(rbnext);
429 return cfq_choose_req(cfqd, next, prev);
432 static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
433 struct cfq_queue *cfqq)
436 * just an approximation, should be ok.
438 return (cfqd->busy_queues - 1) * (cfq_prio_slice(cfqd, 1, 0) -
439 cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
443 * The cfqd->service_tree holds all pending cfq_queue's that have
444 * requests waiting to be processed. It is sorted in the order that
445 * we will service the queues.
447 static void cfq_service_tree_add(struct cfq_data *cfqd,
448 struct cfq_queue *cfqq, int add_front)
450 struct rb_node **p = &cfqd->service_tree.rb.rb_node;
451 struct rb_node *parent = NULL;
452 unsigned long rb_key;
456 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
457 rb_key += cfqq->slice_resid;
458 cfqq->slice_resid = 0;
462 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
464 * same position, nothing more to do
466 if (rb_key == cfqq->rb_key)
469 cfq_rb_erase(&cfqq->rb_node, &cfqd->service_tree);
474 struct cfq_queue *__cfqq;
478 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
481 * sort RT queues first, we always want to give
482 * preference to them. IDLE queues goes to the back.
483 * after that, sort on the next service time.
485 if (cfq_class_rt(cfqq) > cfq_class_rt(__cfqq))
487 else if (cfq_class_rt(cfqq) < cfq_class_rt(__cfqq))
489 else if (cfq_class_idle(cfqq) < cfq_class_idle(__cfqq))
491 else if (cfq_class_idle(cfqq) > cfq_class_idle(__cfqq))
493 else if (rb_key < __cfqq->rb_key)
498 if (n == &(*p)->rb_right)
505 cfqd->service_tree.left = &cfqq->rb_node;
507 cfqq->rb_key = rb_key;
508 rb_link_node(&cfqq->rb_node, parent, p);
509 rb_insert_color(&cfqq->rb_node, &cfqd->service_tree.rb);
513 * Update cfqq's position in the service tree.
515 static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
518 * Resorting requires the cfqq to be on the RR list already.
520 if (cfq_cfqq_on_rr(cfqq))
521 cfq_service_tree_add(cfqd, cfqq, 0);
525 * add to busy list of queues for service, trying to be fair in ordering
526 * the pending list according to last request service
529 cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
531 BUG_ON(cfq_cfqq_on_rr(cfqq));
532 cfq_mark_cfqq_on_rr(cfqq);
535 cfq_resort_rr_list(cfqd, cfqq);
539 * Called when the cfqq no longer has requests pending, remove it from
543 cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
545 BUG_ON(!cfq_cfqq_on_rr(cfqq));
546 cfq_clear_cfqq_on_rr(cfqq);
548 if (!RB_EMPTY_NODE(&cfqq->rb_node))
549 cfq_rb_erase(&cfqq->rb_node, &cfqd->service_tree);
551 BUG_ON(!cfqd->busy_queues);
556 * rb tree support functions
558 static inline void cfq_del_rq_rb(struct request *rq)
560 struct cfq_queue *cfqq = RQ_CFQQ(rq);
561 struct cfq_data *cfqd = cfqq->cfqd;
562 const int sync = rq_is_sync(rq);
564 BUG_ON(!cfqq->queued[sync]);
565 cfqq->queued[sync]--;
567 elv_rb_del(&cfqq->sort_list, rq);
569 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
570 cfq_del_cfqq_rr(cfqd, cfqq);
573 static void cfq_add_rq_rb(struct request *rq)
575 struct cfq_queue *cfqq = RQ_CFQQ(rq);
576 struct cfq_data *cfqd = cfqq->cfqd;
577 struct request *__alias;
579 cfqq->queued[rq_is_sync(rq)]++;
582 * looks a little odd, but the first insert might return an alias.
583 * if that happens, put the alias on the dispatch list
585 while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL)
586 cfq_dispatch_insert(cfqd->queue, __alias);
588 if (!cfq_cfqq_on_rr(cfqq))
589 cfq_add_cfqq_rr(cfqd, cfqq);
592 * check if this request is a better next-serve candidate
594 cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq);
595 BUG_ON(!cfqq->next_rq);
599 cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
601 elv_rb_del(&cfqq->sort_list, rq);
602 cfqq->queued[rq_is_sync(rq)]--;
606 static struct request *
607 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
609 struct task_struct *tsk = current;
610 pid_t key = cfq_queue_pid(tsk, bio_data_dir(bio), bio_sync(bio));
611 struct cfq_queue *cfqq;
613 cfqq = cfq_find_cfq_hash(cfqd, key, tsk->ioprio);
615 sector_t sector = bio->bi_sector + bio_sectors(bio);
617 return elv_rb_find(&cfqq->sort_list, sector);
623 static void cfq_activate_request(request_queue_t *q, struct request *rq)
625 struct cfq_data *cfqd = q->elevator->elevator_data;
627 cfqd->rq_in_driver++;
630 * If the depth is larger 1, it really could be queueing. But lets
631 * make the mark a little higher - idling could still be good for
632 * low queueing, and a low queueing number could also just indicate
633 * a SCSI mid layer like behaviour where limit+1 is often seen.
635 if (!cfqd->hw_tag && cfqd->rq_in_driver > 4)
638 cfqd->last_position = rq->hard_sector + rq->hard_nr_sectors;
641 static void cfq_deactivate_request(request_queue_t *q, struct request *rq)
643 struct cfq_data *cfqd = q->elevator->elevator_data;
645 WARN_ON(!cfqd->rq_in_driver);
646 cfqd->rq_in_driver--;
649 static void cfq_remove_request(struct request *rq)
651 struct cfq_queue *cfqq = RQ_CFQQ(rq);
653 if (cfqq->next_rq == rq)
654 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
656 list_del_init(&rq->queuelist);
659 if (rq_is_meta(rq)) {
660 WARN_ON(!cfqq->meta_pending);
661 cfqq->meta_pending--;
665 static int cfq_merge(request_queue_t *q, struct request **req, struct bio *bio)
667 struct cfq_data *cfqd = q->elevator->elevator_data;
668 struct request *__rq;
670 __rq = cfq_find_rq_fmerge(cfqd, bio);
671 if (__rq && elv_rq_merge_ok(__rq, bio)) {
673 return ELEVATOR_FRONT_MERGE;
676 return ELEVATOR_NO_MERGE;
679 static void cfq_merged_request(request_queue_t *q, struct request *req,
682 if (type == ELEVATOR_FRONT_MERGE) {
683 struct cfq_queue *cfqq = RQ_CFQQ(req);
685 cfq_reposition_rq_rb(cfqq, req);
690 cfq_merged_requests(request_queue_t *q, struct request *rq,
691 struct request *next)
694 * reposition in fifo if next is older than rq
696 if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
697 time_before(next->start_time, rq->start_time))
698 list_move(&rq->queuelist, &next->queuelist);
700 cfq_remove_request(next);
703 static int cfq_allow_merge(request_queue_t *q, struct request *rq,
706 struct cfq_data *cfqd = q->elevator->elevator_data;
707 const int rw = bio_data_dir(bio);
708 struct cfq_queue *cfqq;
712 * Disallow merge of a sync bio into an async request.
714 if ((bio_data_dir(bio) == READ || bio_sync(bio)) && !rq_is_sync(rq))
718 * Lookup the cfqq that this bio will be queued with. Allow
719 * merge only if rq is queued there.
721 key = cfq_queue_pid(current, rw, bio_sync(bio));
722 cfqq = cfq_find_cfq_hash(cfqd, key, current->ioprio);
724 if (cfqq == RQ_CFQQ(rq))
731 __cfq_set_active_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
735 * stop potential idle class queues waiting service
737 del_timer(&cfqd->idle_class_timer);
740 cfq_clear_cfqq_must_alloc_slice(cfqq);
741 cfq_clear_cfqq_fifo_expire(cfqq);
742 cfq_mark_cfqq_slice_new(cfqq);
743 cfq_clear_cfqq_queue_new(cfqq);
746 cfqd->active_queue = cfqq;
750 * current cfqq expired its slice (or was too idle), select new one
753 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
756 if (cfq_cfqq_wait_request(cfqq))
757 del_timer(&cfqd->idle_slice_timer);
759 cfq_clear_cfqq_must_dispatch(cfqq);
760 cfq_clear_cfqq_wait_request(cfqq);
763 * store what was left of this slice, if the queue idled/timed out
765 if (timed_out && !cfq_cfqq_slice_new(cfqq))
766 cfqq->slice_resid = cfqq->slice_end - jiffies;
768 cfq_resort_rr_list(cfqd, cfqq);
770 if (cfqq == cfqd->active_queue)
771 cfqd->active_queue = NULL;
773 if (cfqd->active_cic) {
774 put_io_context(cfqd->active_cic->ioc);
775 cfqd->active_cic = NULL;
779 static inline void cfq_slice_expired(struct cfq_data *cfqd, int timed_out)
781 struct cfq_queue *cfqq = cfqd->active_queue;
784 __cfq_slice_expired(cfqd, cfqq, timed_out);
788 * Get next queue for service. Unless we have a queue preemption,
789 * we'll simply select the first cfqq in the service tree.
791 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
793 struct cfq_queue *cfqq;
796 if (RB_EMPTY_ROOT(&cfqd->service_tree.rb))
799 n = cfq_rb_first(&cfqd->service_tree);
800 cfqq = rb_entry(n, struct cfq_queue, rb_node);
802 if (cfq_class_idle(cfqq)) {
806 * if we have idle queues and no rt or be queues had
807 * pending requests, either allow immediate service if
808 * the grace period has passed or arm the idle grace
811 end = cfqd->last_end_request + CFQ_IDLE_GRACE;
812 if (time_before(jiffies, end)) {
813 mod_timer(&cfqd->idle_class_timer, end);
822 * Get and set a new active queue for service.
824 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd)
826 struct cfq_queue *cfqq;
828 cfqq = cfq_get_next_queue(cfqd);
829 __cfq_set_active_queue(cfqd, cfqq);
833 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
836 if (rq->sector >= cfqd->last_position)
837 return rq->sector - cfqd->last_position;
839 return cfqd->last_position - rq->sector;
842 static inline int cfq_rq_close(struct cfq_data *cfqd, struct request *rq)
844 struct cfq_io_context *cic = cfqd->active_cic;
846 if (!sample_valid(cic->seek_samples))
849 return cfq_dist_from_last(cfqd, rq) <= cic->seek_mean;
852 static int cfq_close_cooperator(struct cfq_data *cfq_data,
853 struct cfq_queue *cfqq)
856 * We should notice if some of the queues are cooperating, eg
857 * working closely on the same area of the disk. In that case,
858 * we can group them together and don't waste time idling.
863 #define CIC_SEEKY(cic) ((cic)->seek_mean > (8 * 1024))
865 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
867 struct cfq_queue *cfqq = cfqd->active_queue;
868 struct cfq_io_context *cic;
871 WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
872 WARN_ON(cfq_cfqq_slice_new(cfqq));
875 * idle is disabled, either manually or by past process history
877 if (!cfqd->cfq_slice_idle || !cfq_cfqq_idle_window(cfqq))
881 * task has exited, don't wait
883 cic = cfqd->active_cic;
884 if (!cic || !cic->ioc->task)
888 * See if this prio level has a good candidate
890 if (cfq_close_cooperator(cfqd, cfqq) &&
891 (sample_valid(cic->ttime_samples) && cic->ttime_mean > 2))
894 cfq_mark_cfqq_must_dispatch(cfqq);
895 cfq_mark_cfqq_wait_request(cfqq);
898 * we don't want to idle for seeks, but we do want to allow
899 * fair distribution of slice time for a process doing back-to-back
900 * seeks. so allow a little bit of time for him to submit a new rq
902 sl = cfqd->cfq_slice_idle;
903 if (sample_valid(cic->seek_samples) && CIC_SEEKY(cic))
904 sl = min(sl, msecs_to_jiffies(CFQ_MIN_TT));
906 mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
910 * Move request from internal lists to the request queue dispatch list.
912 static void cfq_dispatch_insert(request_queue_t *q, struct request *rq)
914 struct cfq_queue *cfqq = RQ_CFQQ(rq);
916 cfq_remove_request(rq);
918 elv_dispatch_sort(q, rq);
922 * return expired entry, or NULL to just start from scratch in rbtree
924 static inline struct request *cfq_check_fifo(struct cfq_queue *cfqq)
926 struct cfq_data *cfqd = cfqq->cfqd;
930 if (cfq_cfqq_fifo_expire(cfqq))
933 cfq_mark_cfqq_fifo_expire(cfqq);
935 if (list_empty(&cfqq->fifo))
938 fifo = cfq_cfqq_sync(cfqq);
939 rq = rq_entry_fifo(cfqq->fifo.next);
941 if (time_before(jiffies, rq->start_time + cfqd->cfq_fifo_expire[fifo]))
948 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
950 const int base_rq = cfqd->cfq_slice_async_rq;
952 WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
954 return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
958 * Select a queue for service. If we have a current active queue,
959 * check whether to continue servicing it, or retrieve and set a new one.
961 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
963 struct cfq_queue *cfqq;
965 cfqq = cfqd->active_queue;
970 * The active queue has run out of time, expire it and select new.
972 if (cfq_slice_used(cfqq))
976 * The active queue has requests and isn't expired, allow it to
979 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
983 * No requests pending. If the active queue still has requests in
984 * flight or is idling for a new request, allow either of these
985 * conditions to happen (or time out) before selecting a new queue.
987 if (cfqq->dispatched || timer_pending(&cfqd->idle_slice_timer)) {
993 cfq_slice_expired(cfqd, 0);
995 cfqq = cfq_set_active_queue(cfqd);
1001 * Dispatch some requests from cfqq, moving them to the request queue
1005 __cfq_dispatch_requests(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1010 BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
1016 * follow expired path, else get first next available
1018 if ((rq = cfq_check_fifo(cfqq)) == NULL)
1022 * finally, insert request into driver dispatch list
1024 cfq_dispatch_insert(cfqd->queue, rq);
1028 if (!cfqd->active_cic) {
1029 atomic_inc(&RQ_CIC(rq)->ioc->refcount);
1030 cfqd->active_cic = RQ_CIC(rq);
1033 if (RB_EMPTY_ROOT(&cfqq->sort_list))
1036 } while (dispatched < max_dispatch);
1039 * expire an async queue immediately if it has used up its slice. idle
1040 * queue always expire after 1 dispatch round.
1042 if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
1043 dispatched >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
1044 cfq_class_idle(cfqq))) {
1045 cfqq->slice_end = jiffies + 1;
1046 cfq_slice_expired(cfqd, 0);
1052 static inline int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
1056 while (cfqq->next_rq) {
1057 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
1061 BUG_ON(!list_empty(&cfqq->fifo));
1066 * Drain our current requests. Used for barriers and when switching
1067 * io schedulers on-the-fly.
1069 static int cfq_forced_dispatch(struct cfq_data *cfqd)
1074 while ((n = cfq_rb_first(&cfqd->service_tree)) != NULL) {
1075 struct cfq_queue *cfqq = rb_entry(n, struct cfq_queue, rb_node);
1077 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
1080 cfq_slice_expired(cfqd, 0);
1082 BUG_ON(cfqd->busy_queues);
1087 static int cfq_dispatch_requests(request_queue_t *q, int force)
1089 struct cfq_data *cfqd = q->elevator->elevator_data;
1090 struct cfq_queue *cfqq;
1093 if (!cfqd->busy_queues)
1096 if (unlikely(force))
1097 return cfq_forced_dispatch(cfqd);
1100 while ((cfqq = cfq_select_queue(cfqd)) != NULL) {
1103 if (cfqd->busy_queues > 1) {
1105 * So we have dispatched before in this round, if the
1106 * next queue has idling enabled (must be sync), don't
1107 * allow it service until the previous have completed.
1109 if (cfqd->rq_in_driver && cfq_cfqq_idle_window(cfqq) &&
1112 if (cfqq->dispatched >= cfqd->cfq_quantum)
1116 cfq_clear_cfqq_must_dispatch(cfqq);
1117 cfq_clear_cfqq_wait_request(cfqq);
1118 del_timer(&cfqd->idle_slice_timer);
1120 max_dispatch = cfqd->cfq_quantum;
1121 if (cfq_class_idle(cfqq))
1124 dispatched += __cfq_dispatch_requests(cfqd, cfqq, max_dispatch);
1131 * task holds one reference to the queue, dropped when task exits. each rq
1132 * in-flight on this queue also holds a reference, dropped when rq is freed.
1134 * queue lock must be held here.
1136 static void cfq_put_queue(struct cfq_queue *cfqq)
1138 struct cfq_data *cfqd = cfqq->cfqd;
1140 BUG_ON(atomic_read(&cfqq->ref) <= 0);
1142 if (!atomic_dec_and_test(&cfqq->ref))
1145 BUG_ON(rb_first(&cfqq->sort_list));
1146 BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
1147 BUG_ON(cfq_cfqq_on_rr(cfqq));
1149 if (unlikely(cfqd->active_queue == cfqq)) {
1150 __cfq_slice_expired(cfqd, cfqq, 0);
1151 cfq_schedule_dispatch(cfqd);
1155 * it's on the empty list and still hashed
1157 hlist_del(&cfqq->cfq_hash);
1158 kmem_cache_free(cfq_pool, cfqq);
1161 static struct cfq_queue *
1162 __cfq_find_cfq_hash(struct cfq_data *cfqd, unsigned int key, unsigned int prio,
1165 struct hlist_head *hash_list = &cfqd->cfq_hash[hashval];
1166 struct hlist_node *entry;
1167 struct cfq_queue *__cfqq;
1169 hlist_for_each_entry(__cfqq, entry, hash_list, cfq_hash) {
1170 const unsigned short __p = IOPRIO_PRIO_VALUE(__cfqq->org_ioprio_class, __cfqq->org_ioprio);
1172 if (__cfqq->key == key && (__p == prio || !prio))
1179 static struct cfq_queue *
1180 cfq_find_cfq_hash(struct cfq_data *cfqd, unsigned int key, unsigned short prio)
1182 return __cfq_find_cfq_hash(cfqd, key, prio, hash_long(key, CFQ_QHASH_SHIFT));
1185 static void cfq_free_io_context(struct io_context *ioc)
1187 struct cfq_io_context *__cic;
1191 while ((n = rb_first(&ioc->cic_root)) != NULL) {
1192 __cic = rb_entry(n, struct cfq_io_context, rb_node);
1193 rb_erase(&__cic->rb_node, &ioc->cic_root);
1194 kmem_cache_free(cfq_ioc_pool, __cic);
1198 elv_ioc_count_mod(ioc_count, -freed);
1200 if (ioc_gone && !elv_ioc_count_read(ioc_count))
1204 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1206 if (unlikely(cfqq == cfqd->active_queue)) {
1207 __cfq_slice_expired(cfqd, cfqq, 0);
1208 cfq_schedule_dispatch(cfqd);
1211 cfq_put_queue(cfqq);
1214 static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
1215 struct cfq_io_context *cic)
1217 list_del_init(&cic->queue_list);
1221 if (cic->cfqq[ASYNC]) {
1222 cfq_exit_cfqq(cfqd, cic->cfqq[ASYNC]);
1223 cic->cfqq[ASYNC] = NULL;
1226 if (cic->cfqq[SYNC]) {
1227 cfq_exit_cfqq(cfqd, cic->cfqq[SYNC]);
1228 cic->cfqq[SYNC] = NULL;
1232 static void cfq_exit_single_io_context(struct cfq_io_context *cic)
1234 struct cfq_data *cfqd = cic->key;
1237 request_queue_t *q = cfqd->queue;
1239 spin_lock_irq(q->queue_lock);
1240 __cfq_exit_single_io_context(cfqd, cic);
1241 spin_unlock_irq(q->queue_lock);
1246 * The process that ioc belongs to has exited, we need to clean up
1247 * and put the internal structures we have that belongs to that process.
1249 static void cfq_exit_io_context(struct io_context *ioc)
1251 struct cfq_io_context *__cic;
1255 * put the reference this task is holding to the various queues
1258 n = rb_first(&ioc->cic_root);
1260 __cic = rb_entry(n, struct cfq_io_context, rb_node);
1262 cfq_exit_single_io_context(__cic);
1267 static struct cfq_io_context *
1268 cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
1270 struct cfq_io_context *cic;
1272 cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask, cfqd->queue->node);
1274 memset(cic, 0, sizeof(*cic));
1275 cic->last_end_request = jiffies;
1276 INIT_LIST_HEAD(&cic->queue_list);
1277 cic->dtor = cfq_free_io_context;
1278 cic->exit = cfq_exit_io_context;
1279 elv_ioc_count_inc(ioc_count);
1285 static void cfq_init_prio_data(struct cfq_queue *cfqq)
1287 struct task_struct *tsk = current;
1290 if (!cfq_cfqq_prio_changed(cfqq))
1293 ioprio_class = IOPRIO_PRIO_CLASS(tsk->ioprio);
1294 switch (ioprio_class) {
1296 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
1297 case IOPRIO_CLASS_NONE:
1299 * no prio set, place us in the middle of the BE classes
1301 cfqq->ioprio = task_nice_ioprio(tsk);
1302 cfqq->ioprio_class = IOPRIO_CLASS_BE;
1304 case IOPRIO_CLASS_RT:
1305 cfqq->ioprio = task_ioprio(tsk);
1306 cfqq->ioprio_class = IOPRIO_CLASS_RT;
1308 case IOPRIO_CLASS_BE:
1309 cfqq->ioprio = task_ioprio(tsk);
1310 cfqq->ioprio_class = IOPRIO_CLASS_BE;
1312 case IOPRIO_CLASS_IDLE:
1313 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
1315 cfq_clear_cfqq_idle_window(cfqq);
1320 * keep track of original prio settings in case we have to temporarily
1321 * elevate the priority of this queue
1323 cfqq->org_ioprio = cfqq->ioprio;
1324 cfqq->org_ioprio_class = cfqq->ioprio_class;
1325 cfq_clear_cfqq_prio_changed(cfqq);
1328 static inline void changed_ioprio(struct cfq_io_context *cic)
1330 struct cfq_data *cfqd = cic->key;
1331 struct cfq_queue *cfqq;
1332 unsigned long flags;
1334 if (unlikely(!cfqd))
1337 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1339 cfqq = cic->cfqq[ASYNC];
1341 struct cfq_queue *new_cfqq;
1342 new_cfqq = cfq_get_queue(cfqd, CFQ_KEY_ASYNC, cic->ioc->task,
1345 cic->cfqq[ASYNC] = new_cfqq;
1346 cfq_put_queue(cfqq);
1350 cfqq = cic->cfqq[SYNC];
1352 cfq_mark_cfqq_prio_changed(cfqq);
1354 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1357 static void cfq_ioc_set_ioprio(struct io_context *ioc)
1359 struct cfq_io_context *cic;
1362 ioc->ioprio_changed = 0;
1364 n = rb_first(&ioc->cic_root);
1366 cic = rb_entry(n, struct cfq_io_context, rb_node);
1368 changed_ioprio(cic);
1373 static struct cfq_queue *
1374 cfq_get_queue(struct cfq_data *cfqd, unsigned int key, struct task_struct *tsk,
1377 const int hashval = hash_long(key, CFQ_QHASH_SHIFT);
1378 struct cfq_queue *cfqq, *new_cfqq = NULL;
1379 unsigned short ioprio;
1382 ioprio = tsk->ioprio;
1383 cfqq = __cfq_find_cfq_hash(cfqd, key, ioprio, hashval);
1389 } else if (gfp_mask & __GFP_WAIT) {
1391 * Inform the allocator of the fact that we will
1392 * just repeat this allocation if it fails, to allow
1393 * the allocator to do whatever it needs to attempt to
1396 spin_unlock_irq(cfqd->queue->queue_lock);
1397 new_cfqq = kmem_cache_alloc_node(cfq_pool, gfp_mask|__GFP_NOFAIL, cfqd->queue->node);
1398 spin_lock_irq(cfqd->queue->queue_lock);
1401 cfqq = kmem_cache_alloc_node(cfq_pool, gfp_mask, cfqd->queue->node);
1406 memset(cfqq, 0, sizeof(*cfqq));
1408 INIT_HLIST_NODE(&cfqq->cfq_hash);
1409 RB_CLEAR_NODE(&cfqq->rb_node);
1410 INIT_LIST_HEAD(&cfqq->fifo);
1413 hlist_add_head(&cfqq->cfq_hash, &cfqd->cfq_hash[hashval]);
1414 atomic_set(&cfqq->ref, 0);
1417 if (key != CFQ_KEY_ASYNC)
1418 cfq_mark_cfqq_idle_window(cfqq);
1420 cfq_mark_cfqq_prio_changed(cfqq);
1421 cfq_mark_cfqq_queue_new(cfqq);
1422 cfq_init_prio_data(cfqq);
1426 kmem_cache_free(cfq_pool, new_cfqq);
1428 atomic_inc(&cfqq->ref);
1430 WARN_ON((gfp_mask & __GFP_WAIT) && !cfqq);
1435 * We drop cfq io contexts lazily, so we may find a dead one.
1438 cfq_drop_dead_cic(struct io_context *ioc, struct cfq_io_context *cic)
1440 WARN_ON(!list_empty(&cic->queue_list));
1441 rb_erase(&cic->rb_node, &ioc->cic_root);
1442 kmem_cache_free(cfq_ioc_pool, cic);
1443 elv_ioc_count_dec(ioc_count);
1446 static struct cfq_io_context *
1447 cfq_cic_rb_lookup(struct cfq_data *cfqd, struct io_context *ioc)
1450 struct cfq_io_context *cic;
1451 void *k, *key = cfqd;
1454 n = ioc->cic_root.rb_node;
1456 cic = rb_entry(n, struct cfq_io_context, rb_node);
1457 /* ->key must be copied to avoid race with cfq_exit_queue() */
1460 cfq_drop_dead_cic(ioc, cic);
1476 cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
1477 struct cfq_io_context *cic)
1480 struct rb_node *parent;
1481 struct cfq_io_context *__cic;
1482 unsigned long flags;
1490 p = &ioc->cic_root.rb_node;
1493 __cic = rb_entry(parent, struct cfq_io_context, rb_node);
1494 /* ->key must be copied to avoid race with cfq_exit_queue() */
1497 cfq_drop_dead_cic(ioc, __cic);
1503 else if (cic->key > k)
1504 p = &(*p)->rb_right;
1509 rb_link_node(&cic->rb_node, parent, p);
1510 rb_insert_color(&cic->rb_node, &ioc->cic_root);
1512 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1513 list_add(&cic->queue_list, &cfqd->cic_list);
1514 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1518 * Setup general io context and cfq io context. There can be several cfq
1519 * io contexts per general io context, if this process is doing io to more
1520 * than one device managed by cfq.
1522 static struct cfq_io_context *
1523 cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
1525 struct io_context *ioc = NULL;
1526 struct cfq_io_context *cic;
1528 might_sleep_if(gfp_mask & __GFP_WAIT);
1530 ioc = get_io_context(gfp_mask, cfqd->queue->node);
1534 cic = cfq_cic_rb_lookup(cfqd, ioc);
1538 cic = cfq_alloc_io_context(cfqd, gfp_mask);
1542 cfq_cic_link(cfqd, ioc, cic);
1544 smp_read_barrier_depends();
1545 if (unlikely(ioc->ioprio_changed))
1546 cfq_ioc_set_ioprio(ioc);
1550 put_io_context(ioc);
1555 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
1557 unsigned long elapsed = jiffies - cic->last_end_request;
1558 unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
1560 cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
1561 cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
1562 cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
1566 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_io_context *cic,
1572 if (cic->last_request_pos < rq->sector)
1573 sdist = rq->sector - cic->last_request_pos;
1575 sdist = cic->last_request_pos - rq->sector;
1577 if (!cic->seek_samples) {
1578 cfqd->new_seek_total = (7*cic->seek_total + (u64)256*sdist) / 8;
1579 cfqd->new_seek_mean = cfqd->new_seek_total / 256;
1583 * Don't allow the seek distance to get too large from the
1584 * odd fragment, pagein, etc
1586 if (cic->seek_samples <= 60) /* second&third seek */
1587 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*1024);
1589 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*64);
1591 cic->seek_samples = (7*cic->seek_samples + 256) / 8;
1592 cic->seek_total = (7*cic->seek_total + (u64)256*sdist) / 8;
1593 total = cic->seek_total + (cic->seek_samples/2);
1594 do_div(total, cic->seek_samples);
1595 cic->seek_mean = (sector_t)total;
1599 * Disable idle window if the process thinks too long or seeks so much that
1603 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1604 struct cfq_io_context *cic)
1608 if (!cfq_cfqq_sync(cfqq))
1611 enable_idle = cfq_cfqq_idle_window(cfqq);
1613 if (!cic->ioc->task || !cfqd->cfq_slice_idle ||
1614 (cfqd->hw_tag && CIC_SEEKY(cic)))
1616 else if (sample_valid(cic->ttime_samples)) {
1617 if (cic->ttime_mean > cfqd->cfq_slice_idle)
1624 cfq_mark_cfqq_idle_window(cfqq);
1626 cfq_clear_cfqq_idle_window(cfqq);
1630 * Check if new_cfqq should preempt the currently active queue. Return 0 for
1631 * no or if we aren't sure, a 1 will cause a preempt.
1634 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
1637 struct cfq_queue *cfqq;
1639 cfqq = cfqd->active_queue;
1643 if (cfq_slice_used(cfqq))
1646 if (cfq_class_idle(new_cfqq))
1649 if (cfq_class_idle(cfqq))
1653 * if the new request is sync, but the currently running queue is
1654 * not, let the sync request have priority.
1656 if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
1660 * So both queues are sync. Let the new request get disk time if
1661 * it's a metadata request and the current queue is doing regular IO.
1663 if (rq_is_meta(rq) && !cfqq->meta_pending)
1666 if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
1670 * if this request is as-good as one we would expect from the
1671 * current cfqq, let it preempt
1673 if (cfq_rq_close(cfqd, rq))
1680 * cfqq preempts the active queue. if we allowed preempt with no slice left,
1681 * let it have half of its nominal slice.
1683 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1685 cfq_slice_expired(cfqd, 1);
1688 * Put the new queue at the front of the of the current list,
1689 * so we know that it will be selected next.
1691 BUG_ON(!cfq_cfqq_on_rr(cfqq));
1693 cfq_service_tree_add(cfqd, cfqq, 1);
1695 cfqq->slice_end = 0;
1696 cfq_mark_cfqq_slice_new(cfqq);
1700 * Called when a new fs request (rq) is added (to cfqq). Check if there's
1701 * something we should do about it
1704 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1707 struct cfq_io_context *cic = RQ_CIC(rq);
1710 cfqq->meta_pending++;
1712 cfq_update_io_thinktime(cfqd, cic);
1713 cfq_update_io_seektime(cfqd, cic, rq);
1714 cfq_update_idle_window(cfqd, cfqq, cic);
1716 cic->last_request_pos = rq->sector + rq->nr_sectors;
1717 cfqq->last_request_pos = cic->last_request_pos;
1719 if (cfqq == cfqd->active_queue) {
1721 * if we are waiting for a request for this queue, let it rip
1722 * immediately and flag that we must not expire this queue
1725 if (cfq_cfqq_wait_request(cfqq)) {
1726 cfq_mark_cfqq_must_dispatch(cfqq);
1727 del_timer(&cfqd->idle_slice_timer);
1728 blk_start_queueing(cfqd->queue);
1730 } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
1732 * not the active queue - expire current slice if it is
1733 * idle and has expired it's mean thinktime or this new queue
1734 * has some old slice time left and is of higher priority
1736 cfq_preempt_queue(cfqd, cfqq);
1737 cfq_mark_cfqq_must_dispatch(cfqq);
1738 blk_start_queueing(cfqd->queue);
1742 static void cfq_insert_request(request_queue_t *q, struct request *rq)
1744 struct cfq_data *cfqd = q->elevator->elevator_data;
1745 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1747 cfq_init_prio_data(cfqq);
1751 list_add_tail(&rq->queuelist, &cfqq->fifo);
1753 cfq_rq_enqueued(cfqd, cfqq, rq);
1756 static void cfq_completed_request(request_queue_t *q, struct request *rq)
1758 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1759 struct cfq_data *cfqd = cfqq->cfqd;
1760 const int sync = rq_is_sync(rq);
1765 WARN_ON(!cfqd->rq_in_driver);
1766 WARN_ON(!cfqq->dispatched);
1767 cfqd->rq_in_driver--;
1770 if (!cfq_class_idle(cfqq))
1771 cfqd->last_end_request = now;
1774 RQ_CIC(rq)->last_end_request = now;
1777 * If this is the active queue, check if it needs to be expired,
1778 * or if we want to idle in case it has no pending requests.
1780 if (cfqd->active_queue == cfqq) {
1781 if (cfq_cfqq_slice_new(cfqq)) {
1782 cfq_set_prio_slice(cfqd, cfqq);
1783 cfq_clear_cfqq_slice_new(cfqq);
1785 if (cfq_slice_used(cfqq))
1786 cfq_slice_expired(cfqd, 1);
1787 else if (sync && RB_EMPTY_ROOT(&cfqq->sort_list))
1788 cfq_arm_slice_timer(cfqd);
1791 if (!cfqd->rq_in_driver)
1792 cfq_schedule_dispatch(cfqd);
1796 * we temporarily boost lower priority queues if they are holding fs exclusive
1797 * resources. they are boosted to normal prio (CLASS_BE/4)
1799 static void cfq_prio_boost(struct cfq_queue *cfqq)
1801 if (has_fs_excl()) {
1803 * boost idle prio on transactions that would lock out other
1804 * users of the filesystem
1806 if (cfq_class_idle(cfqq))
1807 cfqq->ioprio_class = IOPRIO_CLASS_BE;
1808 if (cfqq->ioprio > IOPRIO_NORM)
1809 cfqq->ioprio = IOPRIO_NORM;
1812 * check if we need to unboost the queue
1814 if (cfqq->ioprio_class != cfqq->org_ioprio_class)
1815 cfqq->ioprio_class = cfqq->org_ioprio_class;
1816 if (cfqq->ioprio != cfqq->org_ioprio)
1817 cfqq->ioprio = cfqq->org_ioprio;
1821 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
1823 if ((cfq_cfqq_wait_request(cfqq) || cfq_cfqq_must_alloc(cfqq)) &&
1824 !cfq_cfqq_must_alloc_slice(cfqq)) {
1825 cfq_mark_cfqq_must_alloc_slice(cfqq);
1826 return ELV_MQUEUE_MUST;
1829 return ELV_MQUEUE_MAY;
1832 static int cfq_may_queue(request_queue_t *q, int rw)
1834 struct cfq_data *cfqd = q->elevator->elevator_data;
1835 struct task_struct *tsk = current;
1836 struct cfq_queue *cfqq;
1839 key = cfq_queue_pid(tsk, rw, rw & REQ_RW_SYNC);
1842 * don't force setup of a queue from here, as a call to may_queue
1843 * does not necessarily imply that a request actually will be queued.
1844 * so just lookup a possibly existing queue, or return 'may queue'
1847 cfqq = cfq_find_cfq_hash(cfqd, key, tsk->ioprio);
1849 cfq_init_prio_data(cfqq);
1850 cfq_prio_boost(cfqq);
1852 return __cfq_may_queue(cfqq);
1855 return ELV_MQUEUE_MAY;
1859 * queue lock held here
1861 static void cfq_put_request(struct request *rq)
1863 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1866 const int rw = rq_data_dir(rq);
1868 BUG_ON(!cfqq->allocated[rw]);
1869 cfqq->allocated[rw]--;
1871 put_io_context(RQ_CIC(rq)->ioc);
1873 rq->elevator_private = NULL;
1874 rq->elevator_private2 = NULL;
1876 cfq_put_queue(cfqq);
1881 * Allocate cfq data structures associated with this request.
1884 cfq_set_request(request_queue_t *q, struct request *rq, gfp_t gfp_mask)
1886 struct cfq_data *cfqd = q->elevator->elevator_data;
1887 struct task_struct *tsk = current;
1888 struct cfq_io_context *cic;
1889 const int rw = rq_data_dir(rq);
1890 const int is_sync = rq_is_sync(rq);
1891 pid_t key = cfq_queue_pid(tsk, rw, is_sync);
1892 struct cfq_queue *cfqq;
1893 unsigned long flags;
1895 might_sleep_if(gfp_mask & __GFP_WAIT);
1897 cic = cfq_get_io_context(cfqd, gfp_mask);
1899 spin_lock_irqsave(q->queue_lock, flags);
1904 if (!cic->cfqq[is_sync]) {
1905 cfqq = cfq_get_queue(cfqd, key, tsk, gfp_mask);
1909 cic->cfqq[is_sync] = cfqq;
1911 cfqq = cic->cfqq[is_sync];
1913 cfqq->allocated[rw]++;
1914 cfq_clear_cfqq_must_alloc(cfqq);
1915 atomic_inc(&cfqq->ref);
1917 spin_unlock_irqrestore(q->queue_lock, flags);
1919 rq->elevator_private = cic;
1920 rq->elevator_private2 = cfqq;
1925 put_io_context(cic->ioc);
1927 cfq_schedule_dispatch(cfqd);
1928 spin_unlock_irqrestore(q->queue_lock, flags);
1932 static void cfq_kick_queue(struct work_struct *work)
1934 struct cfq_data *cfqd =
1935 container_of(work, struct cfq_data, unplug_work);
1936 request_queue_t *q = cfqd->queue;
1937 unsigned long flags;
1939 spin_lock_irqsave(q->queue_lock, flags);
1940 blk_start_queueing(q);
1941 spin_unlock_irqrestore(q->queue_lock, flags);
1945 * Timer running if the active_queue is currently idling inside its time slice
1947 static void cfq_idle_slice_timer(unsigned long data)
1949 struct cfq_data *cfqd = (struct cfq_data *) data;
1950 struct cfq_queue *cfqq;
1951 unsigned long flags;
1954 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1956 if ((cfqq = cfqd->active_queue) != NULL) {
1962 if (cfq_slice_used(cfqq))
1966 * only expire and reinvoke request handler, if there are
1967 * other queues with pending requests
1969 if (!cfqd->busy_queues)
1973 * not expired and it has a request pending, let it dispatch
1975 if (!RB_EMPTY_ROOT(&cfqq->sort_list)) {
1976 cfq_mark_cfqq_must_dispatch(cfqq);
1981 cfq_slice_expired(cfqd, timed_out);
1983 cfq_schedule_dispatch(cfqd);
1985 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1989 * Timer running if an idle class queue is waiting for service
1991 static void cfq_idle_class_timer(unsigned long data)
1993 struct cfq_data *cfqd = (struct cfq_data *) data;
1994 unsigned long flags, end;
1996 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1999 * race with a non-idle queue, reset timer
2001 end = cfqd->last_end_request + CFQ_IDLE_GRACE;
2002 if (!time_after_eq(jiffies, end))
2003 mod_timer(&cfqd->idle_class_timer, end);
2005 cfq_schedule_dispatch(cfqd);
2007 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2010 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
2012 del_timer_sync(&cfqd->idle_slice_timer);
2013 del_timer_sync(&cfqd->idle_class_timer);
2014 blk_sync_queue(cfqd->queue);
2017 static void cfq_exit_queue(elevator_t *e)
2019 struct cfq_data *cfqd = e->elevator_data;
2020 request_queue_t *q = cfqd->queue;
2022 cfq_shutdown_timer_wq(cfqd);
2024 spin_lock_irq(q->queue_lock);
2026 if (cfqd->active_queue)
2027 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
2029 while (!list_empty(&cfqd->cic_list)) {
2030 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
2031 struct cfq_io_context,
2034 __cfq_exit_single_io_context(cfqd, cic);
2037 spin_unlock_irq(q->queue_lock);
2039 cfq_shutdown_timer_wq(cfqd);
2041 kfree(cfqd->cfq_hash);
2045 static void *cfq_init_queue(request_queue_t *q)
2047 struct cfq_data *cfqd;
2050 cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL, q->node);
2054 memset(cfqd, 0, sizeof(*cfqd));
2056 cfqd->service_tree = CFQ_RB_ROOT;
2057 INIT_LIST_HEAD(&cfqd->cic_list);
2059 cfqd->cfq_hash = kmalloc_node(sizeof(struct hlist_head) * CFQ_QHASH_ENTRIES, GFP_KERNEL, q->node);
2060 if (!cfqd->cfq_hash)
2063 for (i = 0; i < CFQ_QHASH_ENTRIES; i++)
2064 INIT_HLIST_HEAD(&cfqd->cfq_hash[i]);
2068 init_timer(&cfqd->idle_slice_timer);
2069 cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
2070 cfqd->idle_slice_timer.data = (unsigned long) cfqd;
2072 init_timer(&cfqd->idle_class_timer);
2073 cfqd->idle_class_timer.function = cfq_idle_class_timer;
2074 cfqd->idle_class_timer.data = (unsigned long) cfqd;
2076 INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
2078 cfqd->cfq_quantum = cfq_quantum;
2079 cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
2080 cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
2081 cfqd->cfq_back_max = cfq_back_max;
2082 cfqd->cfq_back_penalty = cfq_back_penalty;
2083 cfqd->cfq_slice[0] = cfq_slice_async;
2084 cfqd->cfq_slice[1] = cfq_slice_sync;
2085 cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
2086 cfqd->cfq_slice_idle = cfq_slice_idle;
2094 static void cfq_slab_kill(void)
2097 kmem_cache_destroy(cfq_pool);
2099 kmem_cache_destroy(cfq_ioc_pool);
2102 static int __init cfq_slab_setup(void)
2104 cfq_pool = kmem_cache_create("cfq_pool", sizeof(struct cfq_queue), 0, 0,
2109 cfq_ioc_pool = kmem_cache_create("cfq_ioc_pool",
2110 sizeof(struct cfq_io_context), 0, 0, NULL, NULL);
2121 * sysfs parts below -->
2124 cfq_var_show(unsigned int var, char *page)
2126 return sprintf(page, "%d\n", var);
2130 cfq_var_store(unsigned int *var, const char *page, size_t count)
2132 char *p = (char *) page;
2134 *var = simple_strtoul(p, &p, 10);
2138 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
2139 static ssize_t __FUNC(elevator_t *e, char *page) \
2141 struct cfq_data *cfqd = e->elevator_data; \
2142 unsigned int __data = __VAR; \
2144 __data = jiffies_to_msecs(__data); \
2145 return cfq_var_show(__data, (page)); \
2147 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
2148 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
2149 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
2150 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
2151 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
2152 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
2153 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
2154 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
2155 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
2156 #undef SHOW_FUNCTION
2158 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
2159 static ssize_t __FUNC(elevator_t *e, const char *page, size_t count) \
2161 struct cfq_data *cfqd = e->elevator_data; \
2162 unsigned int __data; \
2163 int ret = cfq_var_store(&__data, (page), count); \
2164 if (__data < (MIN)) \
2166 else if (__data > (MAX)) \
2169 *(__PTR) = msecs_to_jiffies(__data); \
2171 *(__PTR) = __data; \
2174 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
2175 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1, UINT_MAX, 1);
2176 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1, UINT_MAX, 1);
2177 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
2178 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1, UINT_MAX, 0);
2179 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
2180 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
2181 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
2182 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1, UINT_MAX, 0);
2183 #undef STORE_FUNCTION
2185 #define CFQ_ATTR(name) \
2186 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
2188 static struct elv_fs_entry cfq_attrs[] = {
2190 CFQ_ATTR(fifo_expire_sync),
2191 CFQ_ATTR(fifo_expire_async),
2192 CFQ_ATTR(back_seek_max),
2193 CFQ_ATTR(back_seek_penalty),
2194 CFQ_ATTR(slice_sync),
2195 CFQ_ATTR(slice_async),
2196 CFQ_ATTR(slice_async_rq),
2197 CFQ_ATTR(slice_idle),
2201 static struct elevator_type iosched_cfq = {
2203 .elevator_merge_fn = cfq_merge,
2204 .elevator_merged_fn = cfq_merged_request,
2205 .elevator_merge_req_fn = cfq_merged_requests,
2206 .elevator_allow_merge_fn = cfq_allow_merge,
2207 .elevator_dispatch_fn = cfq_dispatch_requests,
2208 .elevator_add_req_fn = cfq_insert_request,
2209 .elevator_activate_req_fn = cfq_activate_request,
2210 .elevator_deactivate_req_fn = cfq_deactivate_request,
2211 .elevator_queue_empty_fn = cfq_queue_empty,
2212 .elevator_completed_req_fn = cfq_completed_request,
2213 .elevator_former_req_fn = elv_rb_former_request,
2214 .elevator_latter_req_fn = elv_rb_latter_request,
2215 .elevator_set_req_fn = cfq_set_request,
2216 .elevator_put_req_fn = cfq_put_request,
2217 .elevator_may_queue_fn = cfq_may_queue,
2218 .elevator_init_fn = cfq_init_queue,
2219 .elevator_exit_fn = cfq_exit_queue,
2220 .trim = cfq_free_io_context,
2222 .elevator_attrs = cfq_attrs,
2223 .elevator_name = "cfq",
2224 .elevator_owner = THIS_MODULE,
2227 static int __init cfq_init(void)
2232 * could be 0 on HZ < 1000 setups
2234 if (!cfq_slice_async)
2235 cfq_slice_async = 1;
2236 if (!cfq_slice_idle)
2239 if (cfq_slab_setup())
2242 ret = elv_register(&iosched_cfq);
2249 static void __exit cfq_exit(void)
2251 DECLARE_COMPLETION_ONSTACK(all_gone);
2252 elv_unregister(&iosched_cfq);
2253 ioc_gone = &all_gone;
2254 /* ioc_gone's update must be visible before reading ioc_count */
2256 if (elv_ioc_count_read(ioc_count))
2257 wait_for_completion(ioc_gone);
2262 module_init(cfq_init);
2263 module_exit(cfq_exit);
2265 MODULE_AUTHOR("Jens Axboe");
2266 MODULE_LICENSE("GPL");
2267 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");