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)
48 #define list_entry_qhash(entry) hlist_entry((entry), struct cfq_queue, cfq_hash)
50 #define list_entry_cfqq(ptr) list_entry((ptr), struct cfq_queue, cfq_list)
52 #define RQ_CIC(rq) ((struct cfq_io_context*)(rq)->elevator_private)
53 #define RQ_CFQQ(rq) ((rq)->elevator_private2)
55 static struct kmem_cache *cfq_pool;
56 static struct kmem_cache *cfq_ioc_pool;
58 static DEFINE_PER_CPU(unsigned long, ioc_count);
59 static struct completion *ioc_gone;
61 #define CFQ_PRIO_LISTS IOPRIO_BE_NR
62 #define cfq_class_idle(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
63 #define cfq_class_rt(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
68 #define cfq_cfqq_sync(cfqq) ((cfqq)->key != CFQ_KEY_ASYNC)
70 #define sample_valid(samples) ((samples) > 80)
73 * Most of our rbtree usage is for sorting with min extraction, so
74 * if we cache the leftmost node we don't have to walk down the tree
75 * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should
76 * move this into the elevator for the rq sorting as well.
82 #define CFQ_RB_ROOT (struct cfq_rb_root) { RB_ROOT, NULL, }
85 * Per block device queue structure
88 request_queue_t *queue;
91 * rr list of queues with requests and the count of them
93 struct cfq_rb_root service_tree;
94 struct list_head cur_rr;
95 unsigned int busy_queues;
100 struct hlist_head *cfq_hash;
106 * idle window management
108 struct timer_list idle_slice_timer;
109 struct work_struct unplug_work;
111 struct cfq_queue *active_queue;
112 struct cfq_io_context *active_cic;
113 unsigned int dispatch_slice;
115 struct timer_list idle_class_timer;
117 sector_t last_position;
118 unsigned long last_end_request;
121 * tunables, see top of file
123 unsigned int cfq_quantum;
124 unsigned int cfq_fifo_expire[2];
125 unsigned int cfq_back_penalty;
126 unsigned int cfq_back_max;
127 unsigned int cfq_slice[2];
128 unsigned int cfq_slice_async_rq;
129 unsigned int cfq_slice_idle;
131 struct list_head cic_list;
133 sector_t new_seek_mean;
138 * Per process-grouping structure
141 /* reference count */
143 /* parent cfq_data */
144 struct cfq_data *cfqd;
145 /* cfqq lookup hash */
146 struct hlist_node cfq_hash;
149 /* member of the rr/busy/cur/idle cfqd list */
150 struct list_head cfq_list;
151 /* service_tree member */
152 struct rb_node rb_node;
153 /* service_tree key */
154 unsigned long rb_key;
155 /* sorted list of pending requests */
156 struct rb_root sort_list;
157 /* if fifo isn't expired, next request to serve */
158 struct request *next_rq;
159 /* requests queued in sort_list */
161 /* currently allocated requests */
163 /* pending metadata requests */
165 /* fifo list of requests in sort_list */
166 struct list_head fifo;
168 unsigned long slice_end;
171 /* number of requests that are on the dispatch list or inside driver */
174 /* io prio of this group */
175 unsigned short ioprio, org_ioprio;
176 unsigned short ioprio_class, org_ioprio_class;
178 /* various state flags, see below */
181 sector_t last_request_pos;
184 enum cfqq_state_flags {
185 CFQ_CFQQ_FLAG_on_rr = 0, /* on round-robin busy list */
186 CFQ_CFQQ_FLAG_wait_request, /* waiting for a request */
187 CFQ_CFQQ_FLAG_must_alloc, /* must be allowed rq alloc */
188 CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
189 CFQ_CFQQ_FLAG_must_dispatch, /* must dispatch, even if expired */
190 CFQ_CFQQ_FLAG_fifo_expire, /* FIFO checked in this slice */
191 CFQ_CFQQ_FLAG_idle_window, /* slice idling enabled */
192 CFQ_CFQQ_FLAG_prio_changed, /* task priority has changed */
193 CFQ_CFQQ_FLAG_queue_new, /* queue never been serviced */
194 CFQ_CFQQ_FLAG_slice_new, /* no requests dispatched in slice */
197 #define CFQ_CFQQ_FNS(name) \
198 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \
200 cfqq->flags |= (1 << CFQ_CFQQ_FLAG_##name); \
202 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \
204 cfqq->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \
206 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \
208 return (cfqq->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \
212 CFQ_CFQQ_FNS(wait_request);
213 CFQ_CFQQ_FNS(must_alloc);
214 CFQ_CFQQ_FNS(must_alloc_slice);
215 CFQ_CFQQ_FNS(must_dispatch);
216 CFQ_CFQQ_FNS(fifo_expire);
217 CFQ_CFQQ_FNS(idle_window);
218 CFQ_CFQQ_FNS(prio_changed);
219 CFQ_CFQQ_FNS(queue_new);
220 CFQ_CFQQ_FNS(slice_new);
223 static struct cfq_queue *cfq_find_cfq_hash(struct cfq_data *, unsigned int, unsigned short);
224 static void cfq_dispatch_insert(request_queue_t *, struct request *);
225 static struct cfq_queue *cfq_get_queue(struct cfq_data *cfqd, unsigned int key, struct task_struct *tsk, gfp_t gfp_mask);
228 * scheduler run of queue, if there are requests pending and no one in the
229 * driver that will restart queueing
231 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
233 if (cfqd->busy_queues)
234 kblockd_schedule_work(&cfqd->unplug_work);
237 static int cfq_queue_empty(request_queue_t *q)
239 struct cfq_data *cfqd = q->elevator->elevator_data;
241 return !cfqd->busy_queues;
244 static inline pid_t cfq_queue_pid(struct task_struct *task, int rw, int is_sync)
247 * Use the per-process queue, for read requests and syncronous writes
249 if (!(rw & REQ_RW) || is_sync)
252 return CFQ_KEY_ASYNC;
256 * Scale schedule slice based on io priority. Use the sync time slice only
257 * if a queue is marked sync and has sync io queued. A sync queue with async
258 * io only, should not get full sync slice length.
260 static inline int cfq_prio_slice(struct cfq_data *cfqd, int sync,
263 const int base_slice = cfqd->cfq_slice[sync];
265 WARN_ON(prio >= IOPRIO_BE_NR);
267 return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
271 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
273 return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
277 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
279 cfqq->slice_end = cfq_prio_to_slice(cfqd, cfqq) + jiffies;
283 * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
284 * isn't valid until the first request from the dispatch is activated
285 * and the slice time set.
287 static inline int cfq_slice_used(struct cfq_queue *cfqq)
289 if (cfq_cfqq_slice_new(cfqq))
291 if (time_before(jiffies, cfqq->slice_end))
298 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
299 * We choose the request that is closest to the head right now. Distance
300 * behind the head is penalized and only allowed to a certain extent.
302 static struct request *
303 cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2)
305 sector_t last, s1, s2, d1 = 0, d2 = 0;
306 unsigned long back_max;
307 #define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */
308 #define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */
309 unsigned wrap = 0; /* bit mask: requests behind the disk head? */
311 if (rq1 == NULL || rq1 == rq2)
316 if (rq_is_sync(rq1) && !rq_is_sync(rq2))
318 else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
320 if (rq_is_meta(rq1) && !rq_is_meta(rq2))
322 else if (rq_is_meta(rq2) && !rq_is_meta(rq1))
328 last = cfqd->last_position;
331 * by definition, 1KiB is 2 sectors
333 back_max = cfqd->cfq_back_max * 2;
336 * Strict one way elevator _except_ in the case where we allow
337 * short backward seeks which are biased as twice the cost of a
338 * similar forward seek.
342 else if (s1 + back_max >= last)
343 d1 = (last - s1) * cfqd->cfq_back_penalty;
345 wrap |= CFQ_RQ1_WRAP;
349 else if (s2 + back_max >= last)
350 d2 = (last - s2) * cfqd->cfq_back_penalty;
352 wrap |= CFQ_RQ2_WRAP;
354 /* Found required data */
357 * By doing switch() on the bit mask "wrap" we avoid having to
358 * check two variables for all permutations: --> faster!
361 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
377 case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
380 * Since both rqs are wrapped,
381 * start with the one that's further behind head
382 * (--> only *one* back seek required),
383 * since back seek takes more time than forward.
392 static struct rb_node *cfq_rb_first(struct cfq_rb_root *root)
395 root->left = rb_first(&root->rb);
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));
445 static void cfq_service_tree_add(struct cfq_data *cfqd,
446 struct cfq_queue *cfqq)
448 struct rb_node **p = &cfqd->service_tree.rb.rb_node;
449 struct rb_node *parent = NULL;
450 unsigned long rb_key;
453 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
454 rb_key += cfqq->slice_resid;
455 cfqq->slice_resid = 0;
457 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
459 * same position, nothing more to do
461 if (rb_key == cfqq->rb_key)
464 cfq_rb_erase(&cfqq->rb_node, &cfqd->service_tree);
468 struct cfq_queue *__cfqq;
472 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
475 * sort RT queues first, we always want to give
476 * preference to them. IDLE queues goes to the back.
477 * after that, sort on the next service time.
479 if (cfq_class_rt(cfqq) > cfq_class_rt(__cfqq))
481 else if (cfq_class_rt(cfqq) < cfq_class_rt(__cfqq))
483 else if (cfq_class_idle(cfqq) < cfq_class_idle(__cfqq))
485 else if (cfq_class_idle(cfqq) > cfq_class_idle(__cfqq))
487 else if (rb_key < __cfqq->rb_key)
492 if (n == &(*p)->rb_right)
499 cfqd->service_tree.left = &cfqq->rb_node;
501 cfqq->rb_key = rb_key;
502 rb_link_node(&cfqq->rb_node, parent, p);
503 rb_insert_color(&cfqq->rb_node, &cfqd->service_tree.rb);
506 static void cfq_resort_rr_list(struct cfq_queue *cfqq, int preempted)
508 struct cfq_data *cfqd = cfqq->cfqd;
511 * Resorting requires the cfqq to be on the RR list already.
513 if (!cfq_cfqq_on_rr(cfqq))
516 cfq_service_tree_add(cfqd, cfqq);
520 * add to busy list of queues for service, trying to be fair in ordering
521 * the pending list according to last request service
524 cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
526 BUG_ON(cfq_cfqq_on_rr(cfqq));
527 cfq_mark_cfqq_on_rr(cfqq);
530 cfq_resort_rr_list(cfqq, 0);
534 cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
536 BUG_ON(!cfq_cfqq_on_rr(cfqq));
537 cfq_clear_cfqq_on_rr(cfqq);
538 list_del_init(&cfqq->cfq_list);
540 if (!RB_EMPTY_NODE(&cfqq->rb_node))
541 cfq_rb_erase(&cfqq->rb_node, &cfqd->service_tree);
543 BUG_ON(!cfqd->busy_queues);
548 * rb tree support functions
550 static inline void cfq_del_rq_rb(struct request *rq)
552 struct cfq_queue *cfqq = RQ_CFQQ(rq);
553 struct cfq_data *cfqd = cfqq->cfqd;
554 const int sync = rq_is_sync(rq);
556 BUG_ON(!cfqq->queued[sync]);
557 cfqq->queued[sync]--;
559 elv_rb_del(&cfqq->sort_list, rq);
561 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
562 cfq_del_cfqq_rr(cfqd, cfqq);
565 static void cfq_add_rq_rb(struct request *rq)
567 struct cfq_queue *cfqq = RQ_CFQQ(rq);
568 struct cfq_data *cfqd = cfqq->cfqd;
569 struct request *__alias;
571 cfqq->queued[rq_is_sync(rq)]++;
574 * looks a little odd, but the first insert might return an alias.
575 * if that happens, put the alias on the dispatch list
577 while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL)
578 cfq_dispatch_insert(cfqd->queue, __alias);
580 if (!cfq_cfqq_on_rr(cfqq))
581 cfq_add_cfqq_rr(cfqd, cfqq);
584 * check if this request is a better next-serve candidate
586 cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq);
587 BUG_ON(!cfqq->next_rq);
591 cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
593 elv_rb_del(&cfqq->sort_list, rq);
594 cfqq->queued[rq_is_sync(rq)]--;
598 static struct request *
599 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
601 struct task_struct *tsk = current;
602 pid_t key = cfq_queue_pid(tsk, bio_data_dir(bio), bio_sync(bio));
603 struct cfq_queue *cfqq;
605 cfqq = cfq_find_cfq_hash(cfqd, key, tsk->ioprio);
607 sector_t sector = bio->bi_sector + bio_sectors(bio);
609 return elv_rb_find(&cfqq->sort_list, sector);
615 static void cfq_activate_request(request_queue_t *q, struct request *rq)
617 struct cfq_data *cfqd = q->elevator->elevator_data;
619 cfqd->rq_in_driver++;
622 * If the depth is larger 1, it really could be queueing. But lets
623 * make the mark a little higher - idling could still be good for
624 * low queueing, and a low queueing number could also just indicate
625 * a SCSI mid layer like behaviour where limit+1 is often seen.
627 if (!cfqd->hw_tag && cfqd->rq_in_driver > 4)
630 cfqd->last_position = rq->hard_sector + rq->hard_nr_sectors;
633 static void cfq_deactivate_request(request_queue_t *q, struct request *rq)
635 struct cfq_data *cfqd = q->elevator->elevator_data;
637 WARN_ON(!cfqd->rq_in_driver);
638 cfqd->rq_in_driver--;
641 static void cfq_remove_request(struct request *rq)
643 struct cfq_queue *cfqq = RQ_CFQQ(rq);
645 if (cfqq->next_rq == rq)
646 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
648 list_del_init(&rq->queuelist);
651 if (rq_is_meta(rq)) {
652 WARN_ON(!cfqq->meta_pending);
653 cfqq->meta_pending--;
658 cfq_merge(request_queue_t *q, struct request **req, struct bio *bio)
660 struct cfq_data *cfqd = q->elevator->elevator_data;
661 struct request *__rq;
663 __rq = cfq_find_rq_fmerge(cfqd, bio);
664 if (__rq && elv_rq_merge_ok(__rq, bio)) {
666 return ELEVATOR_FRONT_MERGE;
669 return ELEVATOR_NO_MERGE;
672 static void cfq_merged_request(request_queue_t *q, struct request *req,
675 if (type == ELEVATOR_FRONT_MERGE) {
676 struct cfq_queue *cfqq = RQ_CFQQ(req);
678 cfq_reposition_rq_rb(cfqq, req);
683 cfq_merged_requests(request_queue_t *q, struct request *rq,
684 struct request *next)
687 * reposition in fifo if next is older than rq
689 if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
690 time_before(next->start_time, rq->start_time))
691 list_move(&rq->queuelist, &next->queuelist);
693 cfq_remove_request(next);
696 static int cfq_allow_merge(request_queue_t *q, struct request *rq,
699 struct cfq_data *cfqd = q->elevator->elevator_data;
700 const int rw = bio_data_dir(bio);
701 struct cfq_queue *cfqq;
705 * Disallow merge of a sync bio into an async request.
707 if ((bio_data_dir(bio) == READ || bio_sync(bio)) && !rq_is_sync(rq))
711 * Lookup the cfqq that this bio will be queued with. Allow
712 * merge only if rq is queued there.
714 key = cfq_queue_pid(current, rw, bio_sync(bio));
715 cfqq = cfq_find_cfq_hash(cfqd, key, current->ioprio);
717 if (cfqq == RQ_CFQQ(rq))
724 __cfq_set_active_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
728 * stop potential idle class queues waiting service
730 del_timer(&cfqd->idle_class_timer);
733 cfq_clear_cfqq_must_alloc_slice(cfqq);
734 cfq_clear_cfqq_fifo_expire(cfqq);
735 cfq_mark_cfqq_slice_new(cfqq);
736 cfq_clear_cfqq_queue_new(cfqq);
739 cfqd->active_queue = cfqq;
743 * current cfqq expired its slice (or was too idle), select new one
746 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
747 int preempted, int timed_out)
749 if (cfq_cfqq_wait_request(cfqq))
750 del_timer(&cfqd->idle_slice_timer);
752 cfq_clear_cfqq_must_dispatch(cfqq);
753 cfq_clear_cfqq_wait_request(cfqq);
756 * store what was left of this slice, if the queue idled out
759 if (timed_out && !cfq_cfqq_slice_new(cfqq))
760 cfqq->slice_resid = cfqq->slice_end - jiffies;
762 cfq_resort_rr_list(cfqq, preempted);
764 if (cfqq == cfqd->active_queue)
765 cfqd->active_queue = NULL;
767 if (cfqd->active_cic) {
768 put_io_context(cfqd->active_cic->ioc);
769 cfqd->active_cic = NULL;
772 cfqd->dispatch_slice = 0;
775 static inline void cfq_slice_expired(struct cfq_data *cfqd, int preempted,
778 struct cfq_queue *cfqq = cfqd->active_queue;
781 __cfq_slice_expired(cfqd, cfqq, preempted, timed_out);
784 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
786 struct cfq_queue *cfqq = NULL;
788 if (!list_empty(&cfqd->cur_rr)) {
790 * if current list is non-empty, grab first entry.
792 cfqq = list_entry_cfqq(cfqd->cur_rr.next);
793 } else if (!RB_EMPTY_ROOT(&cfqd->service_tree.rb)) {
794 struct rb_node *n = cfq_rb_first(&cfqd->service_tree);
797 cfqq = rb_entry(n, struct cfq_queue, rb_node);
798 if (cfq_class_idle(cfqq)) {
800 * if we have idle queues and no rt or be queues had
801 * pending requests, either allow immediate service if
802 * the grace period has passed or arm the idle grace
805 end = cfqd->last_end_request + CFQ_IDLE_GRACE;
806 if (time_before(jiffies, end)) {
807 mod_timer(&cfqd->idle_class_timer, end);
816 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd)
818 struct cfq_queue *cfqq;
820 cfqq = cfq_get_next_queue(cfqd);
821 __cfq_set_active_queue(cfqd, cfqq);
825 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
828 if (rq->sector >= cfqd->last_position)
829 return rq->sector - cfqd->last_position;
831 return cfqd->last_position - rq->sector;
834 static inline int cfq_rq_close(struct cfq_data *cfqd, struct request *rq)
836 struct cfq_io_context *cic = cfqd->active_cic;
838 if (!sample_valid(cic->seek_samples))
841 return cfq_dist_from_last(cfqd, rq) <= cic->seek_mean;
844 static int cfq_close_cooperator(struct cfq_data *cfq_data,
845 struct cfq_queue *cfqq)
848 * We should notice if some of the queues are cooperating, eg
849 * working closely on the same area of the disk. In that case,
850 * we can group them together and don't waste time idling.
855 #define CIC_SEEKY(cic) ((cic)->seek_mean > (8 * 1024))
857 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
859 struct cfq_queue *cfqq = cfqd->active_queue;
860 struct cfq_io_context *cic;
863 WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
864 WARN_ON(cfq_cfqq_slice_new(cfqq));
867 * idle is disabled, either manually or by past process history
869 if (!cfqd->cfq_slice_idle || !cfq_cfqq_idle_window(cfqq))
873 * task has exited, don't wait
875 cic = cfqd->active_cic;
876 if (!cic || !cic->ioc->task)
880 * See if this prio level has a good candidate
882 if (cfq_close_cooperator(cfqd, cfqq) &&
883 (sample_valid(cic->ttime_samples) && cic->ttime_mean > 2))
886 cfq_mark_cfqq_must_dispatch(cfqq);
887 cfq_mark_cfqq_wait_request(cfqq);
890 * we don't want to idle for seeks, but we do want to allow
891 * fair distribution of slice time for a process doing back-to-back
892 * seeks. so allow a little bit of time for him to submit a new rq
894 sl = cfqd->cfq_slice_idle;
895 if (sample_valid(cic->seek_samples) && CIC_SEEKY(cic))
896 sl = min(sl, msecs_to_jiffies(CFQ_MIN_TT));
898 mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
901 static void cfq_dispatch_insert(request_queue_t *q, struct request *rq)
903 struct cfq_queue *cfqq = RQ_CFQQ(rq);
905 cfq_remove_request(rq);
907 elv_dispatch_sort(q, rq);
911 * return expired entry, or NULL to just start from scratch in rbtree
913 static inline struct request *cfq_check_fifo(struct cfq_queue *cfqq)
915 struct cfq_data *cfqd = cfqq->cfqd;
919 if (cfq_cfqq_fifo_expire(cfqq))
922 cfq_mark_cfqq_fifo_expire(cfqq);
924 if (list_empty(&cfqq->fifo))
927 fifo = cfq_cfqq_sync(cfqq);
928 rq = rq_entry_fifo(cfqq->fifo.next);
930 if (time_before(jiffies, rq->start_time + cfqd->cfq_fifo_expire[fifo]))
937 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
939 const int base_rq = cfqd->cfq_slice_async_rq;
941 WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
943 return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
947 * get next queue for service
949 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
951 struct cfq_queue *cfqq;
953 cfqq = cfqd->active_queue;
958 * The active queue has run out of time, expire it and select new.
960 if (cfq_slice_used(cfqq))
964 * The active queue has requests and isn't expired, allow it to
967 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
971 * No requests pending. If the active queue still has requests in
972 * flight or is idling for a new request, allow either of these
973 * conditions to happen (or time out) before selecting a new queue.
975 if (cfqq->dispatched || timer_pending(&cfqd->idle_slice_timer)) {
981 cfq_slice_expired(cfqd, 0, 0);
983 cfqq = cfq_set_active_queue(cfqd);
989 __cfq_dispatch_requests(struct cfq_data *cfqd, struct cfq_queue *cfqq,
994 BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
1000 * follow expired path, else get first next available
1002 if ((rq = cfq_check_fifo(cfqq)) == NULL)
1006 * finally, insert request into driver dispatch list
1008 cfq_dispatch_insert(cfqd->queue, rq);
1010 cfqd->dispatch_slice++;
1013 if (!cfqd->active_cic) {
1014 atomic_inc(&RQ_CIC(rq)->ioc->refcount);
1015 cfqd->active_cic = RQ_CIC(rq);
1018 if (RB_EMPTY_ROOT(&cfqq->sort_list))
1021 } while (dispatched < max_dispatch);
1024 * expire an async queue immediately if it has used up its slice. idle
1025 * queue always expire after 1 dispatch round.
1027 if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
1028 cfqd->dispatch_slice >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
1029 cfq_class_idle(cfqq))) {
1030 cfqq->slice_end = jiffies + 1;
1031 cfq_slice_expired(cfqd, 0, 0);
1037 static inline int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
1041 while (cfqq->next_rq) {
1042 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
1046 BUG_ON(!list_empty(&cfqq->fifo));
1050 static int cfq_forced_dispatch_cfqqs(struct list_head *list)
1052 struct cfq_queue *cfqq, *next;
1056 list_for_each_entry_safe(cfqq, next, list, cfq_list)
1057 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
1062 static int cfq_forced_dispatch(struct cfq_data *cfqd)
1067 while ((n = cfq_rb_first(&cfqd->service_tree)) != NULL) {
1068 struct cfq_queue *cfqq = rb_entry(n, struct cfq_queue, rb_node);
1070 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
1073 dispatched += cfq_forced_dispatch_cfqqs(&cfqd->cur_rr);
1075 cfq_slice_expired(cfqd, 0, 0);
1077 BUG_ON(cfqd->busy_queues);
1082 static int cfq_dispatch_requests(request_queue_t *q, int force)
1084 struct cfq_data *cfqd = q->elevator->elevator_data;
1085 struct cfq_queue *cfqq;
1088 if (!cfqd->busy_queues)
1091 if (unlikely(force))
1092 return cfq_forced_dispatch(cfqd);
1095 while ((cfqq = cfq_select_queue(cfqd)) != NULL) {
1098 if (cfqd->busy_queues > 1) {
1100 * So we have dispatched before in this round, if the
1101 * next queue has idling enabled (must be sync), don't
1102 * allow it service until the previous have completed.
1104 if (cfqd->rq_in_driver && cfq_cfqq_idle_window(cfqq) &&
1107 if (cfqq->dispatched >= cfqd->cfq_quantum)
1111 cfq_clear_cfqq_must_dispatch(cfqq);
1112 cfq_clear_cfqq_wait_request(cfqq);
1113 del_timer(&cfqd->idle_slice_timer);
1115 max_dispatch = cfqd->cfq_quantum;
1116 if (cfq_class_idle(cfqq))
1119 dispatched += __cfq_dispatch_requests(cfqd, cfqq, max_dispatch);
1126 * task holds one reference to the queue, dropped when task exits. each rq
1127 * in-flight on this queue also holds a reference, dropped when rq is freed.
1129 * queue lock must be held here.
1131 static void cfq_put_queue(struct cfq_queue *cfqq)
1133 struct cfq_data *cfqd = cfqq->cfqd;
1135 BUG_ON(atomic_read(&cfqq->ref) <= 0);
1137 if (!atomic_dec_and_test(&cfqq->ref))
1140 BUG_ON(rb_first(&cfqq->sort_list));
1141 BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
1142 BUG_ON(cfq_cfqq_on_rr(cfqq));
1144 if (unlikely(cfqd->active_queue == cfqq)) {
1145 __cfq_slice_expired(cfqd, cfqq, 0, 0);
1146 cfq_schedule_dispatch(cfqd);
1150 * it's on the empty list and still hashed
1152 hlist_del(&cfqq->cfq_hash);
1153 kmem_cache_free(cfq_pool, cfqq);
1156 static struct cfq_queue *
1157 __cfq_find_cfq_hash(struct cfq_data *cfqd, unsigned int key, unsigned int prio,
1160 struct hlist_head *hash_list = &cfqd->cfq_hash[hashval];
1161 struct hlist_node *entry;
1162 struct cfq_queue *__cfqq;
1164 hlist_for_each_entry(__cfqq, entry, hash_list, cfq_hash) {
1165 const unsigned short __p = IOPRIO_PRIO_VALUE(__cfqq->org_ioprio_class, __cfqq->org_ioprio);
1167 if (__cfqq->key == key && (__p == prio || !prio))
1174 static struct cfq_queue *
1175 cfq_find_cfq_hash(struct cfq_data *cfqd, unsigned int key, unsigned short prio)
1177 return __cfq_find_cfq_hash(cfqd, key, prio, hash_long(key, CFQ_QHASH_SHIFT));
1180 static void cfq_free_io_context(struct io_context *ioc)
1182 struct cfq_io_context *__cic;
1186 while ((n = rb_first(&ioc->cic_root)) != NULL) {
1187 __cic = rb_entry(n, struct cfq_io_context, rb_node);
1188 rb_erase(&__cic->rb_node, &ioc->cic_root);
1189 kmem_cache_free(cfq_ioc_pool, __cic);
1193 elv_ioc_count_mod(ioc_count, -freed);
1195 if (ioc_gone && !elv_ioc_count_read(ioc_count))
1199 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1201 if (unlikely(cfqq == cfqd->active_queue)) {
1202 __cfq_slice_expired(cfqd, cfqq, 0, 0);
1203 cfq_schedule_dispatch(cfqd);
1206 cfq_put_queue(cfqq);
1209 static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
1210 struct cfq_io_context *cic)
1212 list_del_init(&cic->queue_list);
1216 if (cic->cfqq[ASYNC]) {
1217 cfq_exit_cfqq(cfqd, cic->cfqq[ASYNC]);
1218 cic->cfqq[ASYNC] = NULL;
1221 if (cic->cfqq[SYNC]) {
1222 cfq_exit_cfqq(cfqd, cic->cfqq[SYNC]);
1223 cic->cfqq[SYNC] = NULL;
1229 * Called with interrupts disabled
1231 static void cfq_exit_single_io_context(struct cfq_io_context *cic)
1233 struct cfq_data *cfqd = cic->key;
1236 request_queue_t *q = cfqd->queue;
1238 spin_lock_irq(q->queue_lock);
1239 __cfq_exit_single_io_context(cfqd, cic);
1240 spin_unlock_irq(q->queue_lock);
1244 static void cfq_exit_io_context(struct io_context *ioc)
1246 struct cfq_io_context *__cic;
1250 * put the reference this task is holding to the various queues
1253 n = rb_first(&ioc->cic_root);
1255 __cic = rb_entry(n, struct cfq_io_context, rb_node);
1257 cfq_exit_single_io_context(__cic);
1262 static struct cfq_io_context *
1263 cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
1265 struct cfq_io_context *cic;
1267 cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask, cfqd->queue->node);
1269 memset(cic, 0, sizeof(*cic));
1270 cic->last_end_request = jiffies;
1271 INIT_LIST_HEAD(&cic->queue_list);
1272 cic->dtor = cfq_free_io_context;
1273 cic->exit = cfq_exit_io_context;
1274 elv_ioc_count_inc(ioc_count);
1280 static void cfq_init_prio_data(struct cfq_queue *cfqq)
1282 struct task_struct *tsk = current;
1285 if (!cfq_cfqq_prio_changed(cfqq))
1288 ioprio_class = IOPRIO_PRIO_CLASS(tsk->ioprio);
1289 switch (ioprio_class) {
1291 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
1292 case IOPRIO_CLASS_NONE:
1294 * no prio set, place us in the middle of the BE classes
1296 cfqq->ioprio = task_nice_ioprio(tsk);
1297 cfqq->ioprio_class = IOPRIO_CLASS_BE;
1299 case IOPRIO_CLASS_RT:
1300 cfqq->ioprio = task_ioprio(tsk);
1301 cfqq->ioprio_class = IOPRIO_CLASS_RT;
1303 case IOPRIO_CLASS_BE:
1304 cfqq->ioprio = task_ioprio(tsk);
1305 cfqq->ioprio_class = IOPRIO_CLASS_BE;
1307 case IOPRIO_CLASS_IDLE:
1308 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
1310 cfq_clear_cfqq_idle_window(cfqq);
1315 * keep track of original prio settings in case we have to temporarily
1316 * elevate the priority of this queue
1318 cfqq->org_ioprio = cfqq->ioprio;
1319 cfqq->org_ioprio_class = cfqq->ioprio_class;
1320 cfq_clear_cfqq_prio_changed(cfqq);
1323 static inline void changed_ioprio(struct cfq_io_context *cic)
1325 struct cfq_data *cfqd = cic->key;
1326 struct cfq_queue *cfqq;
1327 unsigned long flags;
1329 if (unlikely(!cfqd))
1332 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1334 cfqq = cic->cfqq[ASYNC];
1336 struct cfq_queue *new_cfqq;
1337 new_cfqq = cfq_get_queue(cfqd, CFQ_KEY_ASYNC, cic->ioc->task,
1340 cic->cfqq[ASYNC] = new_cfqq;
1341 cfq_put_queue(cfqq);
1345 cfqq = cic->cfqq[SYNC];
1347 cfq_mark_cfqq_prio_changed(cfqq);
1349 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1352 static void cfq_ioc_set_ioprio(struct io_context *ioc)
1354 struct cfq_io_context *cic;
1357 ioc->ioprio_changed = 0;
1359 n = rb_first(&ioc->cic_root);
1361 cic = rb_entry(n, struct cfq_io_context, rb_node);
1363 changed_ioprio(cic);
1368 static struct cfq_queue *
1369 cfq_get_queue(struct cfq_data *cfqd, unsigned int key, struct task_struct *tsk,
1372 const int hashval = hash_long(key, CFQ_QHASH_SHIFT);
1373 struct cfq_queue *cfqq, *new_cfqq = NULL;
1374 unsigned short ioprio;
1377 ioprio = tsk->ioprio;
1378 cfqq = __cfq_find_cfq_hash(cfqd, key, ioprio, hashval);
1384 } else if (gfp_mask & __GFP_WAIT) {
1386 * Inform the allocator of the fact that we will
1387 * just repeat this allocation if it fails, to allow
1388 * the allocator to do whatever it needs to attempt to
1391 spin_unlock_irq(cfqd->queue->queue_lock);
1392 new_cfqq = kmem_cache_alloc_node(cfq_pool, gfp_mask|__GFP_NOFAIL, cfqd->queue->node);
1393 spin_lock_irq(cfqd->queue->queue_lock);
1396 cfqq = kmem_cache_alloc_node(cfq_pool, gfp_mask, cfqd->queue->node);
1401 memset(cfqq, 0, sizeof(*cfqq));
1403 INIT_HLIST_NODE(&cfqq->cfq_hash);
1404 INIT_LIST_HEAD(&cfqq->cfq_list);
1405 RB_CLEAR_NODE(&cfqq->rb_node);
1406 INIT_LIST_HEAD(&cfqq->fifo);
1409 hlist_add_head(&cfqq->cfq_hash, &cfqd->cfq_hash[hashval]);
1410 atomic_set(&cfqq->ref, 0);
1413 if (key != CFQ_KEY_ASYNC)
1414 cfq_mark_cfqq_idle_window(cfqq);
1416 cfq_mark_cfqq_prio_changed(cfqq);
1417 cfq_mark_cfqq_queue_new(cfqq);
1418 cfq_init_prio_data(cfqq);
1422 kmem_cache_free(cfq_pool, new_cfqq);
1424 atomic_inc(&cfqq->ref);
1426 WARN_ON((gfp_mask & __GFP_WAIT) && !cfqq);
1431 cfq_drop_dead_cic(struct io_context *ioc, struct cfq_io_context *cic)
1433 WARN_ON(!list_empty(&cic->queue_list));
1434 rb_erase(&cic->rb_node, &ioc->cic_root);
1435 kmem_cache_free(cfq_ioc_pool, cic);
1436 elv_ioc_count_dec(ioc_count);
1439 static struct cfq_io_context *
1440 cfq_cic_rb_lookup(struct cfq_data *cfqd, struct io_context *ioc)
1443 struct cfq_io_context *cic;
1444 void *k, *key = cfqd;
1447 n = ioc->cic_root.rb_node;
1449 cic = rb_entry(n, struct cfq_io_context, rb_node);
1450 /* ->key must be copied to avoid race with cfq_exit_queue() */
1453 cfq_drop_dead_cic(ioc, cic);
1469 cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
1470 struct cfq_io_context *cic)
1473 struct rb_node *parent;
1474 struct cfq_io_context *__cic;
1475 unsigned long flags;
1483 p = &ioc->cic_root.rb_node;
1486 __cic = rb_entry(parent, struct cfq_io_context, rb_node);
1487 /* ->key must be copied to avoid race with cfq_exit_queue() */
1490 cfq_drop_dead_cic(ioc, __cic);
1496 else if (cic->key > k)
1497 p = &(*p)->rb_right;
1502 rb_link_node(&cic->rb_node, parent, p);
1503 rb_insert_color(&cic->rb_node, &ioc->cic_root);
1505 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1506 list_add(&cic->queue_list, &cfqd->cic_list);
1507 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1511 * Setup general io context and cfq io context. There can be several cfq
1512 * io contexts per general io context, if this process is doing io to more
1513 * than one device managed by cfq.
1515 static struct cfq_io_context *
1516 cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
1518 struct io_context *ioc = NULL;
1519 struct cfq_io_context *cic;
1521 might_sleep_if(gfp_mask & __GFP_WAIT);
1523 ioc = get_io_context(gfp_mask, cfqd->queue->node);
1527 cic = cfq_cic_rb_lookup(cfqd, ioc);
1531 cic = cfq_alloc_io_context(cfqd, gfp_mask);
1535 cfq_cic_link(cfqd, ioc, cic);
1537 smp_read_barrier_depends();
1538 if (unlikely(ioc->ioprio_changed))
1539 cfq_ioc_set_ioprio(ioc);
1543 put_io_context(ioc);
1548 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
1550 unsigned long elapsed = jiffies - cic->last_end_request;
1551 unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
1553 cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
1554 cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
1555 cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
1559 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_io_context *cic,
1565 if (cic->last_request_pos < rq->sector)
1566 sdist = rq->sector - cic->last_request_pos;
1568 sdist = cic->last_request_pos - rq->sector;
1570 if (!cic->seek_samples) {
1571 cfqd->new_seek_total = (7*cic->seek_total + (u64)256*sdist) / 8;
1572 cfqd->new_seek_mean = cfqd->new_seek_total / 256;
1576 * Don't allow the seek distance to get too large from the
1577 * odd fragment, pagein, etc
1579 if (cic->seek_samples <= 60) /* second&third seek */
1580 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*1024);
1582 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*64);
1584 cic->seek_samples = (7*cic->seek_samples + 256) / 8;
1585 cic->seek_total = (7*cic->seek_total + (u64)256*sdist) / 8;
1586 total = cic->seek_total + (cic->seek_samples/2);
1587 do_div(total, cic->seek_samples);
1588 cic->seek_mean = (sector_t)total;
1592 * Disable idle window if the process thinks too long or seeks so much that
1596 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1597 struct cfq_io_context *cic)
1599 int enable_idle = cfq_cfqq_idle_window(cfqq);
1601 if (!cic->ioc->task || !cfqd->cfq_slice_idle ||
1602 (cfqd->hw_tag && CIC_SEEKY(cic)))
1604 else if (sample_valid(cic->ttime_samples)) {
1605 if (cic->ttime_mean > cfqd->cfq_slice_idle)
1612 cfq_mark_cfqq_idle_window(cfqq);
1614 cfq_clear_cfqq_idle_window(cfqq);
1618 * Check if new_cfqq should preempt the currently active queue. Return 0 for
1619 * no or if we aren't sure, a 1 will cause a preempt.
1622 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
1625 struct cfq_queue *cfqq;
1627 cfqq = cfqd->active_queue;
1631 if (cfq_slice_used(cfqq))
1634 if (cfq_class_idle(new_cfqq))
1637 if (cfq_class_idle(cfqq))
1641 * if the new request is sync, but the currently running queue is
1642 * not, let the sync request have priority.
1644 if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
1648 * So both queues are sync. Let the new request get disk time if
1649 * it's a metadata request and the current queue is doing regular IO.
1651 if (rq_is_meta(rq) && !cfqq->meta_pending)
1654 if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
1658 * if this request is as-good as one we would expect from the
1659 * current cfqq, let it preempt
1661 if (cfq_rq_close(cfqd, rq))
1668 * cfqq preempts the active queue. if we allowed preempt with no slice left,
1669 * let it have half of its nominal slice.
1671 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1673 cfq_slice_expired(cfqd, 1, 1);
1676 * Put the new queue at the front of the of the current list,
1677 * so we know that it will be selected next.
1679 BUG_ON(!cfq_cfqq_on_rr(cfqq));
1680 list_del_init(&cfqq->cfq_list);
1681 list_add(&cfqq->cfq_list, &cfqd->cur_rr);
1683 cfqq->slice_end = 0;
1684 cfq_mark_cfqq_slice_new(cfqq);
1688 * Called when a new fs request (rq) is added (to cfqq). Check if there's
1689 * something we should do about it
1692 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1695 struct cfq_io_context *cic = RQ_CIC(rq);
1698 cfqq->meta_pending++;
1700 cfq_update_io_thinktime(cfqd, cic);
1701 cfq_update_io_seektime(cfqd, cic, rq);
1702 cfq_update_idle_window(cfqd, cfqq, cic);
1704 cic->last_request_pos = rq->sector + rq->nr_sectors;
1705 cfqq->last_request_pos = cic->last_request_pos;
1707 if (cfqq == cfqd->active_queue) {
1709 * if we are waiting for a request for this queue, let it rip
1710 * immediately and flag that we must not expire this queue
1713 if (cfq_cfqq_wait_request(cfqq)) {
1714 cfq_mark_cfqq_must_dispatch(cfqq);
1715 del_timer(&cfqd->idle_slice_timer);
1716 blk_start_queueing(cfqd->queue);
1718 } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
1720 * not the active queue - expire current slice if it is
1721 * idle and has expired it's mean thinktime or this new queue
1722 * has some old slice time left and is of higher priority
1724 cfq_preempt_queue(cfqd, cfqq);
1725 cfq_mark_cfqq_must_dispatch(cfqq);
1726 blk_start_queueing(cfqd->queue);
1730 static void cfq_insert_request(request_queue_t *q, struct request *rq)
1732 struct cfq_data *cfqd = q->elevator->elevator_data;
1733 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1735 cfq_init_prio_data(cfqq);
1739 list_add_tail(&rq->queuelist, &cfqq->fifo);
1741 cfq_rq_enqueued(cfqd, cfqq, rq);
1744 static void cfq_completed_request(request_queue_t *q, struct request *rq)
1746 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1747 struct cfq_data *cfqd = cfqq->cfqd;
1748 const int sync = rq_is_sync(rq);
1753 WARN_ON(!cfqd->rq_in_driver);
1754 WARN_ON(!cfqq->dispatched);
1755 cfqd->rq_in_driver--;
1758 if (!cfq_class_idle(cfqq))
1759 cfqd->last_end_request = now;
1762 RQ_CIC(rq)->last_end_request = now;
1765 * If this is the active queue, check if it needs to be expired,
1766 * or if we want to idle in case it has no pending requests.
1768 if (cfqd->active_queue == cfqq) {
1769 if (cfq_cfqq_slice_new(cfqq)) {
1770 cfq_set_prio_slice(cfqd, cfqq);
1771 cfq_clear_cfqq_slice_new(cfqq);
1773 if (cfq_slice_used(cfqq))
1774 cfq_slice_expired(cfqd, 0, 1);
1775 else if (sync && RB_EMPTY_ROOT(&cfqq->sort_list))
1776 cfq_arm_slice_timer(cfqd);
1779 if (!cfqd->rq_in_driver)
1780 cfq_schedule_dispatch(cfqd);
1784 * we temporarily boost lower priority queues if they are holding fs exclusive
1785 * resources. they are boosted to normal prio (CLASS_BE/4)
1787 static void cfq_prio_boost(struct cfq_queue *cfqq)
1789 if (has_fs_excl()) {
1791 * boost idle prio on transactions that would lock out other
1792 * users of the filesystem
1794 if (cfq_class_idle(cfqq))
1795 cfqq->ioprio_class = IOPRIO_CLASS_BE;
1796 if (cfqq->ioprio > IOPRIO_NORM)
1797 cfqq->ioprio = IOPRIO_NORM;
1800 * check if we need to unboost the queue
1802 if (cfqq->ioprio_class != cfqq->org_ioprio_class)
1803 cfqq->ioprio_class = cfqq->org_ioprio_class;
1804 if (cfqq->ioprio != cfqq->org_ioprio)
1805 cfqq->ioprio = cfqq->org_ioprio;
1809 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
1811 if ((cfq_cfqq_wait_request(cfqq) || cfq_cfqq_must_alloc(cfqq)) &&
1812 !cfq_cfqq_must_alloc_slice(cfqq)) {
1813 cfq_mark_cfqq_must_alloc_slice(cfqq);
1814 return ELV_MQUEUE_MUST;
1817 return ELV_MQUEUE_MAY;
1820 static int cfq_may_queue(request_queue_t *q, int rw)
1822 struct cfq_data *cfqd = q->elevator->elevator_data;
1823 struct task_struct *tsk = current;
1824 struct cfq_queue *cfqq;
1827 key = cfq_queue_pid(tsk, rw, rw & REQ_RW_SYNC);
1830 * don't force setup of a queue from here, as a call to may_queue
1831 * does not necessarily imply that a request actually will be queued.
1832 * so just lookup a possibly existing queue, or return 'may queue'
1835 cfqq = cfq_find_cfq_hash(cfqd, key, tsk->ioprio);
1837 cfq_init_prio_data(cfqq);
1838 cfq_prio_boost(cfqq);
1840 return __cfq_may_queue(cfqq);
1843 return ELV_MQUEUE_MAY;
1847 * queue lock held here
1849 static void cfq_put_request(struct request *rq)
1851 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1854 const int rw = rq_data_dir(rq);
1856 BUG_ON(!cfqq->allocated[rw]);
1857 cfqq->allocated[rw]--;
1859 put_io_context(RQ_CIC(rq)->ioc);
1861 rq->elevator_private = NULL;
1862 rq->elevator_private2 = NULL;
1864 cfq_put_queue(cfqq);
1869 * Allocate cfq data structures associated with this request.
1872 cfq_set_request(request_queue_t *q, struct request *rq, gfp_t gfp_mask)
1874 struct cfq_data *cfqd = q->elevator->elevator_data;
1875 struct task_struct *tsk = current;
1876 struct cfq_io_context *cic;
1877 const int rw = rq_data_dir(rq);
1878 const int is_sync = rq_is_sync(rq);
1879 pid_t key = cfq_queue_pid(tsk, rw, is_sync);
1880 struct cfq_queue *cfqq;
1881 unsigned long flags;
1883 might_sleep_if(gfp_mask & __GFP_WAIT);
1885 cic = cfq_get_io_context(cfqd, gfp_mask);
1887 spin_lock_irqsave(q->queue_lock, flags);
1892 if (!cic->cfqq[is_sync]) {
1893 cfqq = cfq_get_queue(cfqd, key, tsk, gfp_mask);
1897 cic->cfqq[is_sync] = cfqq;
1899 cfqq = cic->cfqq[is_sync];
1901 cfqq->allocated[rw]++;
1902 cfq_clear_cfqq_must_alloc(cfqq);
1903 atomic_inc(&cfqq->ref);
1905 spin_unlock_irqrestore(q->queue_lock, flags);
1907 rq->elevator_private = cic;
1908 rq->elevator_private2 = cfqq;
1913 put_io_context(cic->ioc);
1915 cfq_schedule_dispatch(cfqd);
1916 spin_unlock_irqrestore(q->queue_lock, flags);
1920 static void cfq_kick_queue(struct work_struct *work)
1922 struct cfq_data *cfqd =
1923 container_of(work, struct cfq_data, unplug_work);
1924 request_queue_t *q = cfqd->queue;
1925 unsigned long flags;
1927 spin_lock_irqsave(q->queue_lock, flags);
1928 blk_start_queueing(q);
1929 spin_unlock_irqrestore(q->queue_lock, flags);
1933 * Timer running if the active_queue is currently idling inside its time slice
1935 static void cfq_idle_slice_timer(unsigned long data)
1937 struct cfq_data *cfqd = (struct cfq_data *) data;
1938 struct cfq_queue *cfqq;
1939 unsigned long flags;
1942 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1944 if ((cfqq = cfqd->active_queue) != NULL) {
1950 if (cfq_slice_used(cfqq))
1954 * only expire and reinvoke request handler, if there are
1955 * other queues with pending requests
1957 if (!cfqd->busy_queues)
1961 * not expired and it has a request pending, let it dispatch
1963 if (!RB_EMPTY_ROOT(&cfqq->sort_list)) {
1964 cfq_mark_cfqq_must_dispatch(cfqq);
1969 cfq_slice_expired(cfqd, 0, timed_out);
1971 cfq_schedule_dispatch(cfqd);
1973 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1977 * Timer running if an idle class queue is waiting for service
1979 static void cfq_idle_class_timer(unsigned long data)
1981 struct cfq_data *cfqd = (struct cfq_data *) data;
1982 unsigned long flags, end;
1984 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1987 * race with a non-idle queue, reset timer
1989 end = cfqd->last_end_request + CFQ_IDLE_GRACE;
1990 if (!time_after_eq(jiffies, end))
1991 mod_timer(&cfqd->idle_class_timer, end);
1993 cfq_schedule_dispatch(cfqd);
1995 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1998 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
2000 del_timer_sync(&cfqd->idle_slice_timer);
2001 del_timer_sync(&cfqd->idle_class_timer);
2002 blk_sync_queue(cfqd->queue);
2005 static void cfq_exit_queue(elevator_t *e)
2007 struct cfq_data *cfqd = e->elevator_data;
2008 request_queue_t *q = cfqd->queue;
2010 cfq_shutdown_timer_wq(cfqd);
2012 spin_lock_irq(q->queue_lock);
2014 if (cfqd->active_queue)
2015 __cfq_slice_expired(cfqd, cfqd->active_queue, 0, 0);
2017 while (!list_empty(&cfqd->cic_list)) {
2018 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
2019 struct cfq_io_context,
2022 __cfq_exit_single_io_context(cfqd, cic);
2025 spin_unlock_irq(q->queue_lock);
2027 cfq_shutdown_timer_wq(cfqd);
2029 kfree(cfqd->cfq_hash);
2033 static void *cfq_init_queue(request_queue_t *q)
2035 struct cfq_data *cfqd;
2038 cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL, q->node);
2042 memset(cfqd, 0, sizeof(*cfqd));
2044 cfqd->service_tree = CFQ_RB_ROOT;
2045 INIT_LIST_HEAD(&cfqd->cur_rr);
2046 INIT_LIST_HEAD(&cfqd->cic_list);
2048 cfqd->cfq_hash = kmalloc_node(sizeof(struct hlist_head) * CFQ_QHASH_ENTRIES, GFP_KERNEL, q->node);
2049 if (!cfqd->cfq_hash)
2052 for (i = 0; i < CFQ_QHASH_ENTRIES; i++)
2053 INIT_HLIST_HEAD(&cfqd->cfq_hash[i]);
2057 init_timer(&cfqd->idle_slice_timer);
2058 cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
2059 cfqd->idle_slice_timer.data = (unsigned long) cfqd;
2061 init_timer(&cfqd->idle_class_timer);
2062 cfqd->idle_class_timer.function = cfq_idle_class_timer;
2063 cfqd->idle_class_timer.data = (unsigned long) cfqd;
2065 INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
2067 cfqd->cfq_quantum = cfq_quantum;
2068 cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
2069 cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
2070 cfqd->cfq_back_max = cfq_back_max;
2071 cfqd->cfq_back_penalty = cfq_back_penalty;
2072 cfqd->cfq_slice[0] = cfq_slice_async;
2073 cfqd->cfq_slice[1] = cfq_slice_sync;
2074 cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
2075 cfqd->cfq_slice_idle = cfq_slice_idle;
2083 static void cfq_slab_kill(void)
2086 kmem_cache_destroy(cfq_pool);
2088 kmem_cache_destroy(cfq_ioc_pool);
2091 static int __init cfq_slab_setup(void)
2093 cfq_pool = kmem_cache_create("cfq_pool", sizeof(struct cfq_queue), 0, 0,
2098 cfq_ioc_pool = kmem_cache_create("cfq_ioc_pool",
2099 sizeof(struct cfq_io_context), 0, 0, NULL, NULL);
2110 * sysfs parts below -->
2113 cfq_var_show(unsigned int var, char *page)
2115 return sprintf(page, "%d\n", var);
2119 cfq_var_store(unsigned int *var, const char *page, size_t count)
2121 char *p = (char *) page;
2123 *var = simple_strtoul(p, &p, 10);
2127 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
2128 static ssize_t __FUNC(elevator_t *e, char *page) \
2130 struct cfq_data *cfqd = e->elevator_data; \
2131 unsigned int __data = __VAR; \
2133 __data = jiffies_to_msecs(__data); \
2134 return cfq_var_show(__data, (page)); \
2136 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
2137 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
2138 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
2139 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
2140 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
2141 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
2142 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
2143 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
2144 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
2145 #undef SHOW_FUNCTION
2147 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
2148 static ssize_t __FUNC(elevator_t *e, const char *page, size_t count) \
2150 struct cfq_data *cfqd = e->elevator_data; \
2151 unsigned int __data; \
2152 int ret = cfq_var_store(&__data, (page), count); \
2153 if (__data < (MIN)) \
2155 else if (__data > (MAX)) \
2158 *(__PTR) = msecs_to_jiffies(__data); \
2160 *(__PTR) = __data; \
2163 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
2164 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1, UINT_MAX, 1);
2165 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1, UINT_MAX, 1);
2166 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
2167 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1, UINT_MAX, 0);
2168 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
2169 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
2170 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
2171 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1, UINT_MAX, 0);
2172 #undef STORE_FUNCTION
2174 #define CFQ_ATTR(name) \
2175 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
2177 static struct elv_fs_entry cfq_attrs[] = {
2179 CFQ_ATTR(fifo_expire_sync),
2180 CFQ_ATTR(fifo_expire_async),
2181 CFQ_ATTR(back_seek_max),
2182 CFQ_ATTR(back_seek_penalty),
2183 CFQ_ATTR(slice_sync),
2184 CFQ_ATTR(slice_async),
2185 CFQ_ATTR(slice_async_rq),
2186 CFQ_ATTR(slice_idle),
2190 static struct elevator_type iosched_cfq = {
2192 .elevator_merge_fn = cfq_merge,
2193 .elevator_merged_fn = cfq_merged_request,
2194 .elevator_merge_req_fn = cfq_merged_requests,
2195 .elevator_allow_merge_fn = cfq_allow_merge,
2196 .elevator_dispatch_fn = cfq_dispatch_requests,
2197 .elevator_add_req_fn = cfq_insert_request,
2198 .elevator_activate_req_fn = cfq_activate_request,
2199 .elevator_deactivate_req_fn = cfq_deactivate_request,
2200 .elevator_queue_empty_fn = cfq_queue_empty,
2201 .elevator_completed_req_fn = cfq_completed_request,
2202 .elevator_former_req_fn = elv_rb_former_request,
2203 .elevator_latter_req_fn = elv_rb_latter_request,
2204 .elevator_set_req_fn = cfq_set_request,
2205 .elevator_put_req_fn = cfq_put_request,
2206 .elevator_may_queue_fn = cfq_may_queue,
2207 .elevator_init_fn = cfq_init_queue,
2208 .elevator_exit_fn = cfq_exit_queue,
2209 .trim = cfq_free_io_context,
2211 .elevator_attrs = cfq_attrs,
2212 .elevator_name = "cfq",
2213 .elevator_owner = THIS_MODULE,
2216 static int __init cfq_init(void)
2221 * could be 0 on HZ < 1000 setups
2223 if (!cfq_slice_async)
2224 cfq_slice_async = 1;
2225 if (!cfq_slice_idle)
2228 if (cfq_slab_setup())
2231 ret = elv_register(&iosched_cfq);
2238 static void __exit cfq_exit(void)
2240 DECLARE_COMPLETION_ONSTACK(all_gone);
2241 elv_unregister(&iosched_cfq);
2242 ioc_gone = &all_gone;
2243 /* ioc_gone's update must be visible before reading ioc_count */
2245 if (elv_ioc_count_read(ioc_count))
2246 wait_for_completion(ioc_gone);
2251 module_init(cfq_init);
2252 module_exit(cfq_exit);
2254 MODULE_AUTHOR("Jens Axboe");
2255 MODULE_LICENSE("GPL");
2256 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");