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@suse.de>
9 #include <linux/config.h>
10 #include <linux/module.h>
11 #include <linux/blkdev.h>
12 #include <linux/elevator.h>
13 #include <linux/hash.h>
14 #include <linux/rbtree.h>
15 #include <linux/ioprio.h>
20 static const int cfq_quantum = 4; /* max queue in one round of service */
21 static const int cfq_queued = 8; /* minimum rq allocate limit per-queue*/
22 static const int cfq_fifo_expire[2] = { HZ / 4, HZ / 8 };
23 static const int cfq_back_max = 16 * 1024; /* maximum backwards seek, in KiB */
24 static const int cfq_back_penalty = 2; /* penalty of a backwards seek */
26 static const int cfq_slice_sync = HZ / 10;
27 static int cfq_slice_async = HZ / 25;
28 static const int cfq_slice_async_rq = 2;
29 static int cfq_slice_idle = HZ / 70;
31 #define CFQ_IDLE_GRACE (HZ / 10)
32 #define CFQ_SLICE_SCALE (5)
34 #define CFQ_KEY_ASYNC (0)
36 static DEFINE_RWLOCK(cfq_exit_lock);
39 * for the hash of cfqq inside the cfqd
41 #define CFQ_QHASH_SHIFT 6
42 #define CFQ_QHASH_ENTRIES (1 << CFQ_QHASH_SHIFT)
43 #define list_entry_qhash(entry) hlist_entry((entry), struct cfq_queue, cfq_hash)
46 * for the hash of crq inside the cfqq
48 #define CFQ_MHASH_SHIFT 6
49 #define CFQ_MHASH_BLOCK(sec) ((sec) >> 3)
50 #define CFQ_MHASH_ENTRIES (1 << CFQ_MHASH_SHIFT)
51 #define CFQ_MHASH_FN(sec) hash_long(CFQ_MHASH_BLOCK(sec), CFQ_MHASH_SHIFT)
52 #define rq_hash_key(rq) ((rq)->sector + (rq)->nr_sectors)
53 #define list_entry_hash(ptr) hlist_entry((ptr), struct cfq_rq, hash)
55 #define list_entry_cfqq(ptr) list_entry((ptr), struct cfq_queue, cfq_list)
56 #define list_entry_fifo(ptr) list_entry((ptr), struct request, queuelist)
58 #define RQ_DATA(rq) (rq)->elevator_private
63 #define RB_EMPTY(node) ((node)->rb_node == NULL)
64 #define RB_CLEAR(node) do { \
65 memset(node, 0, sizeof(*node)); \
67 #define RB_CLEAR_ROOT(root) ((root)->rb_node = NULL)
68 #define rb_entry_crq(node) rb_entry((node), struct cfq_rq, rb_node)
69 #define rq_rb_key(rq) (rq)->sector
71 static kmem_cache_t *crq_pool;
72 static kmem_cache_t *cfq_pool;
73 static kmem_cache_t *cfq_ioc_pool;
75 static atomic_t ioc_count = ATOMIC_INIT(0);
76 static struct completion *ioc_gone;
78 #define CFQ_PRIO_LISTS IOPRIO_BE_NR
79 #define cfq_class_idle(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
80 #define cfq_class_be(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_BE)
81 #define cfq_class_rt(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
86 #define cfq_cfqq_dispatched(cfqq) \
87 ((cfqq)->on_dispatch[ASYNC] + (cfqq)->on_dispatch[SYNC])
89 #define cfq_cfqq_class_sync(cfqq) ((cfqq)->key != CFQ_KEY_ASYNC)
91 #define cfq_cfqq_sync(cfqq) \
92 (cfq_cfqq_class_sync(cfqq) || (cfqq)->on_dispatch[SYNC])
94 #define sample_valid(samples) ((samples) > 80)
97 * Per block device queue structure
100 request_queue_t *queue;
103 * rr list of queues with requests and the count of them
105 struct list_head rr_list[CFQ_PRIO_LISTS];
106 struct list_head busy_rr;
107 struct list_head cur_rr;
108 struct list_head idle_rr;
109 unsigned int busy_queues;
112 * non-ordered list of empty cfqq's
114 struct list_head empty_list;
119 struct hlist_head *cfq_hash;
122 * global crq hash for all queues
124 struct hlist_head *crq_hash;
126 unsigned int max_queued;
133 * schedule slice state info
136 * idle window management
138 struct timer_list idle_slice_timer;
139 struct work_struct unplug_work;
141 struct cfq_queue *active_queue;
142 struct cfq_io_context *active_cic;
143 int cur_prio, cur_end_prio;
144 unsigned int dispatch_slice;
146 struct timer_list idle_class_timer;
148 sector_t last_sector;
149 unsigned long last_end_request;
151 unsigned int rq_starved;
154 * tunables, see top of file
156 unsigned int cfq_quantum;
157 unsigned int cfq_queued;
158 unsigned int cfq_fifo_expire[2];
159 unsigned int cfq_back_penalty;
160 unsigned int cfq_back_max;
161 unsigned int cfq_slice[2];
162 unsigned int cfq_slice_async_rq;
163 unsigned int cfq_slice_idle;
165 struct list_head cic_list;
169 * Per process-grouping structure
172 /* reference count */
174 /* parent cfq_data */
175 struct cfq_data *cfqd;
176 /* cfqq lookup hash */
177 struct hlist_node cfq_hash;
180 /* on either rr or empty list of cfqd */
181 struct list_head cfq_list;
182 /* sorted list of pending requests */
183 struct rb_root sort_list;
184 /* if fifo isn't expired, next request to serve */
185 struct cfq_rq *next_crq;
186 /* requests queued in sort_list */
188 /* currently allocated requests */
190 /* fifo list of requests in sort_list */
191 struct list_head fifo;
193 unsigned long slice_start;
194 unsigned long slice_end;
195 unsigned long slice_left;
196 unsigned long service_last;
198 /* number of requests that are on the dispatch list */
201 /* io prio of this group */
202 unsigned short ioprio, org_ioprio;
203 unsigned short ioprio_class, org_ioprio_class;
205 /* various state flags, see below */
210 struct rb_node rb_node;
212 struct request *request;
213 struct hlist_node hash;
215 struct cfq_queue *cfq_queue;
216 struct cfq_io_context *io_context;
218 unsigned int crq_flags;
221 enum cfqq_state_flags {
222 CFQ_CFQQ_FLAG_on_rr = 0,
223 CFQ_CFQQ_FLAG_wait_request,
224 CFQ_CFQQ_FLAG_must_alloc,
225 CFQ_CFQQ_FLAG_must_alloc_slice,
226 CFQ_CFQQ_FLAG_must_dispatch,
227 CFQ_CFQQ_FLAG_fifo_expire,
228 CFQ_CFQQ_FLAG_idle_window,
229 CFQ_CFQQ_FLAG_prio_changed,
232 #define CFQ_CFQQ_FNS(name) \
233 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \
235 cfqq->flags |= (1 << CFQ_CFQQ_FLAG_##name); \
237 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \
239 cfqq->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \
241 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \
243 return (cfqq->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \
247 CFQ_CFQQ_FNS(wait_request);
248 CFQ_CFQQ_FNS(must_alloc);
249 CFQ_CFQQ_FNS(must_alloc_slice);
250 CFQ_CFQQ_FNS(must_dispatch);
251 CFQ_CFQQ_FNS(fifo_expire);
252 CFQ_CFQQ_FNS(idle_window);
253 CFQ_CFQQ_FNS(prio_changed);
256 enum cfq_rq_state_flags {
257 CFQ_CRQ_FLAG_is_sync = 0,
260 #define CFQ_CRQ_FNS(name) \
261 static inline void cfq_mark_crq_##name(struct cfq_rq *crq) \
263 crq->crq_flags |= (1 << CFQ_CRQ_FLAG_##name); \
265 static inline void cfq_clear_crq_##name(struct cfq_rq *crq) \
267 crq->crq_flags &= ~(1 << CFQ_CRQ_FLAG_##name); \
269 static inline int cfq_crq_##name(const struct cfq_rq *crq) \
271 return (crq->crq_flags & (1 << CFQ_CRQ_FLAG_##name)) != 0; \
274 CFQ_CRQ_FNS(is_sync);
277 static struct cfq_queue *cfq_find_cfq_hash(struct cfq_data *, unsigned int, unsigned short);
278 static void cfq_dispatch_insert(request_queue_t *, struct cfq_rq *);
279 static struct cfq_queue *cfq_get_queue(struct cfq_data *cfqd, unsigned int key, struct task_struct *tsk, gfp_t gfp_mask);
281 #define process_sync(tsk) ((tsk)->flags & PF_SYNCWRITE)
284 * lots of deadline iosched dupes, can be abstracted later...
286 static inline void cfq_del_crq_hash(struct cfq_rq *crq)
288 hlist_del_init(&crq->hash);
291 static inline void cfq_add_crq_hash(struct cfq_data *cfqd, struct cfq_rq *crq)
293 const int hash_idx = CFQ_MHASH_FN(rq_hash_key(crq->request));
295 hlist_add_head(&crq->hash, &cfqd->crq_hash[hash_idx]);
298 static struct request *cfq_find_rq_hash(struct cfq_data *cfqd, sector_t offset)
300 struct hlist_head *hash_list = &cfqd->crq_hash[CFQ_MHASH_FN(offset)];
301 struct hlist_node *entry, *next;
303 hlist_for_each_safe(entry, next, hash_list) {
304 struct cfq_rq *crq = list_entry_hash(entry);
305 struct request *__rq = crq->request;
307 if (!rq_mergeable(__rq)) {
308 cfq_del_crq_hash(crq);
312 if (rq_hash_key(__rq) == offset)
320 * scheduler run of queue, if there are requests pending and no one in the
321 * driver that will restart queueing
323 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
325 if (cfqd->busy_queues)
326 kblockd_schedule_work(&cfqd->unplug_work);
329 static int cfq_queue_empty(request_queue_t *q)
331 struct cfq_data *cfqd = q->elevator->elevator_data;
333 return !cfqd->busy_queues;
336 static inline pid_t cfq_queue_pid(struct task_struct *task, int rw)
338 if (rw == READ || process_sync(task))
341 return CFQ_KEY_ASYNC;
345 * Lifted from AS - choose which of crq1 and crq2 that is best served now.
346 * We choose the request that is closest to the head right now. Distance
347 * behind the head is penalized and only allowed to a certain extent.
349 static struct cfq_rq *
350 cfq_choose_req(struct cfq_data *cfqd, struct cfq_rq *crq1, struct cfq_rq *crq2)
352 sector_t last, s1, s2, d1 = 0, d2 = 0;
353 unsigned long back_max;
354 #define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */
355 #define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */
356 unsigned wrap = 0; /* bit mask: requests behind the disk head? */
358 if (crq1 == NULL || crq1 == crq2)
363 if (cfq_crq_is_sync(crq1) && !cfq_crq_is_sync(crq2))
365 else if (cfq_crq_is_sync(crq2) && !cfq_crq_is_sync(crq1))
368 s1 = crq1->request->sector;
369 s2 = crq2->request->sector;
371 last = cfqd->last_sector;
374 * by definition, 1KiB is 2 sectors
376 back_max = cfqd->cfq_back_max * 2;
379 * Strict one way elevator _except_ in the case where we allow
380 * short backward seeks which are biased as twice the cost of a
381 * similar forward seek.
385 else if (s1 + back_max >= last)
386 d1 = (last - s1) * cfqd->cfq_back_penalty;
388 wrap |= CFQ_RQ1_WRAP;
392 else if (s2 + back_max >= last)
393 d2 = (last - s2) * cfqd->cfq_back_penalty;
395 wrap |= CFQ_RQ2_WRAP;
397 /* Found required data */
400 * By doing switch() on the bit mask "wrap" we avoid having to
401 * check two variables for all permutations: --> faster!
404 case 0: /* common case for CFQ: crq1 and crq2 not wrapped */
420 case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both crqs wrapped */
423 * Since both rqs are wrapped,
424 * start with the one that's further behind head
425 * (--> only *one* back seek required),
426 * since back seek takes more time than forward.
436 * would be nice to take fifo expire time into account as well
438 static struct cfq_rq *
439 cfq_find_next_crq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
442 struct cfq_rq *crq_next = NULL, *crq_prev = NULL;
443 struct rb_node *rbnext, *rbprev;
445 if (!(rbnext = rb_next(&last->rb_node))) {
446 rbnext = rb_first(&cfqq->sort_list);
447 if (rbnext == &last->rb_node)
451 rbprev = rb_prev(&last->rb_node);
454 crq_prev = rb_entry_crq(rbprev);
456 crq_next = rb_entry_crq(rbnext);
458 return cfq_choose_req(cfqd, crq_next, crq_prev);
461 static void cfq_update_next_crq(struct cfq_rq *crq)
463 struct cfq_queue *cfqq = crq->cfq_queue;
465 if (cfqq->next_crq == crq)
466 cfqq->next_crq = cfq_find_next_crq(cfqq->cfqd, cfqq, crq);
469 static void cfq_resort_rr_list(struct cfq_queue *cfqq, int preempted)
471 struct cfq_data *cfqd = cfqq->cfqd;
472 struct list_head *list, *entry;
474 BUG_ON(!cfq_cfqq_on_rr(cfqq));
476 list_del(&cfqq->cfq_list);
478 if (cfq_class_rt(cfqq))
479 list = &cfqd->cur_rr;
480 else if (cfq_class_idle(cfqq))
481 list = &cfqd->idle_rr;
484 * if cfqq has requests in flight, don't allow it to be
485 * found in cfq_set_active_queue before it has finished them.
486 * this is done to increase fairness between a process that
487 * has lots of io pending vs one that only generates one
488 * sporadically or synchronously
490 if (cfq_cfqq_dispatched(cfqq))
491 list = &cfqd->busy_rr;
493 list = &cfqd->rr_list[cfqq->ioprio];
497 * if queue was preempted, just add to front to be fair. busy_rr
500 if (preempted || list == &cfqd->busy_rr) {
501 list_add(&cfqq->cfq_list, list);
506 * sort by when queue was last serviced
509 while ((entry = entry->prev) != list) {
510 struct cfq_queue *__cfqq = list_entry_cfqq(entry);
512 if (!__cfqq->service_last)
514 if (time_before(__cfqq->service_last, cfqq->service_last))
518 list_add(&cfqq->cfq_list, entry);
522 * add to busy list of queues for service, trying to be fair in ordering
523 * the pending list according to last request service
526 cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
528 BUG_ON(cfq_cfqq_on_rr(cfqq));
529 cfq_mark_cfqq_on_rr(cfqq);
532 cfq_resort_rr_list(cfqq, 0);
536 cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
538 BUG_ON(!cfq_cfqq_on_rr(cfqq));
539 cfq_clear_cfqq_on_rr(cfqq);
540 list_move(&cfqq->cfq_list, &cfqd->empty_list);
542 BUG_ON(!cfqd->busy_queues);
547 * rb tree support functions
549 static inline void cfq_del_crq_rb(struct cfq_rq *crq)
551 struct cfq_queue *cfqq = crq->cfq_queue;
552 struct cfq_data *cfqd = cfqq->cfqd;
553 const int sync = cfq_crq_is_sync(crq);
555 BUG_ON(!cfqq->queued[sync]);
556 cfqq->queued[sync]--;
558 cfq_update_next_crq(crq);
560 rb_erase(&crq->rb_node, &cfqq->sort_list);
562 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY(&cfqq->sort_list))
563 cfq_del_cfqq_rr(cfqd, cfqq);
566 static struct cfq_rq *
567 __cfq_add_crq_rb(struct cfq_rq *crq)
569 struct rb_node **p = &crq->cfq_queue->sort_list.rb_node;
570 struct rb_node *parent = NULL;
571 struct cfq_rq *__crq;
575 __crq = rb_entry_crq(parent);
577 if (crq->rb_key < __crq->rb_key)
579 else if (crq->rb_key > __crq->rb_key)
585 rb_link_node(&crq->rb_node, parent, p);
589 static void cfq_add_crq_rb(struct cfq_rq *crq)
591 struct cfq_queue *cfqq = crq->cfq_queue;
592 struct cfq_data *cfqd = cfqq->cfqd;
593 struct request *rq = crq->request;
594 struct cfq_rq *__alias;
596 crq->rb_key = rq_rb_key(rq);
597 cfqq->queued[cfq_crq_is_sync(crq)]++;
600 * looks a little odd, but the first insert might return an alias.
601 * if that happens, put the alias on the dispatch list
603 while ((__alias = __cfq_add_crq_rb(crq)) != NULL)
604 cfq_dispatch_insert(cfqd->queue, __alias);
606 rb_insert_color(&crq->rb_node, &cfqq->sort_list);
608 if (!cfq_cfqq_on_rr(cfqq))
609 cfq_add_cfqq_rr(cfqd, cfqq);
612 * check if this request is a better next-serve candidate
614 cfqq->next_crq = cfq_choose_req(cfqd, cfqq->next_crq, crq);
618 cfq_reposition_crq_rb(struct cfq_queue *cfqq, struct cfq_rq *crq)
620 rb_erase(&crq->rb_node, &cfqq->sort_list);
621 cfqq->queued[cfq_crq_is_sync(crq)]--;
626 static struct request *
627 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
629 struct task_struct *tsk = current;
630 pid_t key = cfq_queue_pid(tsk, bio_data_dir(bio));
631 struct cfq_queue *cfqq;
635 cfqq = cfq_find_cfq_hash(cfqd, key, tsk->ioprio);
639 sector = bio->bi_sector + bio_sectors(bio);
640 n = cfqq->sort_list.rb_node;
642 struct cfq_rq *crq = rb_entry_crq(n);
644 if (sector < crq->rb_key)
646 else if (sector > crq->rb_key)
656 static void cfq_activate_request(request_queue_t *q, struct request *rq)
658 struct cfq_data *cfqd = q->elevator->elevator_data;
660 cfqd->rq_in_driver++;
663 static void cfq_deactivate_request(request_queue_t *q, struct request *rq)
665 struct cfq_data *cfqd = q->elevator->elevator_data;
667 WARN_ON(!cfqd->rq_in_driver);
668 cfqd->rq_in_driver--;
671 static void cfq_remove_request(struct request *rq)
673 struct cfq_rq *crq = RQ_DATA(rq);
675 list_del_init(&rq->queuelist);
677 cfq_del_crq_hash(crq);
681 cfq_merge(request_queue_t *q, struct request **req, struct bio *bio)
683 struct cfq_data *cfqd = q->elevator->elevator_data;
684 struct request *__rq;
687 __rq = cfq_find_rq_hash(cfqd, bio->bi_sector);
688 if (__rq && elv_rq_merge_ok(__rq, bio)) {
689 ret = ELEVATOR_BACK_MERGE;
693 __rq = cfq_find_rq_fmerge(cfqd, bio);
694 if (__rq && elv_rq_merge_ok(__rq, bio)) {
695 ret = ELEVATOR_FRONT_MERGE;
699 return ELEVATOR_NO_MERGE;
705 static void cfq_merged_request(request_queue_t *q, struct request *req)
707 struct cfq_data *cfqd = q->elevator->elevator_data;
708 struct cfq_rq *crq = RQ_DATA(req);
710 cfq_del_crq_hash(crq);
711 cfq_add_crq_hash(cfqd, crq);
713 if (rq_rb_key(req) != crq->rb_key) {
714 struct cfq_queue *cfqq = crq->cfq_queue;
716 cfq_update_next_crq(crq);
717 cfq_reposition_crq_rb(cfqq, crq);
722 cfq_merged_requests(request_queue_t *q, struct request *rq,
723 struct request *next)
725 cfq_merged_request(q, rq);
728 * reposition in fifo if next is older than rq
730 if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
731 time_before(next->start_time, rq->start_time))
732 list_move(&rq->queuelist, &next->queuelist);
734 cfq_remove_request(next);
738 __cfq_set_active_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
742 * stop potential idle class queues waiting service
744 del_timer(&cfqd->idle_class_timer);
746 cfqq->slice_start = jiffies;
748 cfqq->slice_left = 0;
749 cfq_clear_cfqq_must_alloc_slice(cfqq);
750 cfq_clear_cfqq_fifo_expire(cfqq);
753 cfqd->active_queue = cfqq;
757 * current cfqq expired its slice (or was too idle), select new one
760 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
763 unsigned long now = jiffies;
765 if (cfq_cfqq_wait_request(cfqq))
766 del_timer(&cfqd->idle_slice_timer);
768 if (!preempted && !cfq_cfqq_dispatched(cfqq)) {
769 cfqq->service_last = now;
770 cfq_schedule_dispatch(cfqd);
773 cfq_clear_cfqq_must_dispatch(cfqq);
774 cfq_clear_cfqq_wait_request(cfqq);
777 * store what was left of this slice, if the queue idled out
780 if (time_after(cfqq->slice_end, now))
781 cfqq->slice_left = cfqq->slice_end - now;
783 cfqq->slice_left = 0;
785 if (cfq_cfqq_on_rr(cfqq))
786 cfq_resort_rr_list(cfqq, preempted);
788 if (cfqq == cfqd->active_queue)
789 cfqd->active_queue = NULL;
791 if (cfqd->active_cic) {
792 put_io_context(cfqd->active_cic->ioc);
793 cfqd->active_cic = NULL;
796 cfqd->dispatch_slice = 0;
799 static inline void cfq_slice_expired(struct cfq_data *cfqd, int preempted)
801 struct cfq_queue *cfqq = cfqd->active_queue;
804 __cfq_slice_expired(cfqd, cfqq, preempted);
817 static int cfq_get_next_prio_level(struct cfq_data *cfqd)
826 for (p = cfqd->cur_prio; p <= cfqd->cur_end_prio; p++) {
827 if (!list_empty(&cfqd->rr_list[p])) {
836 if (++cfqd->cur_end_prio == CFQ_PRIO_LISTS) {
837 cfqd->cur_end_prio = 0;
844 if (unlikely(prio == -1))
847 BUG_ON(prio >= CFQ_PRIO_LISTS);
849 list_splice_init(&cfqd->rr_list[prio], &cfqd->cur_rr);
851 cfqd->cur_prio = prio + 1;
852 if (cfqd->cur_prio > cfqd->cur_end_prio) {
853 cfqd->cur_end_prio = cfqd->cur_prio;
856 if (cfqd->cur_end_prio == CFQ_PRIO_LISTS) {
858 cfqd->cur_end_prio = 0;
864 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd)
866 struct cfq_queue *cfqq = NULL;
869 * if current list is non-empty, grab first entry. if it is empty,
870 * get next prio level and grab first entry then if any are spliced
872 if (!list_empty(&cfqd->cur_rr) || cfq_get_next_prio_level(cfqd) != -1)
873 cfqq = list_entry_cfqq(cfqd->cur_rr.next);
876 * if we have idle queues and no rt or be queues had pending
877 * requests, either allow immediate service if the grace period
878 * has passed or arm the idle grace timer
880 if (!cfqq && !list_empty(&cfqd->idle_rr)) {
881 unsigned long end = cfqd->last_end_request + CFQ_IDLE_GRACE;
883 if (time_after_eq(jiffies, end))
884 cfqq = list_entry_cfqq(cfqd->idle_rr.next);
886 mod_timer(&cfqd->idle_class_timer, end);
889 __cfq_set_active_queue(cfqd, cfqq);
893 static int cfq_arm_slice_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq)
896 struct cfq_io_context *cic;
899 WARN_ON(!RB_EMPTY(&cfqq->sort_list));
900 WARN_ON(cfqq != cfqd->active_queue);
903 * idle is disabled, either manually or by past process history
905 if (!cfqd->cfq_slice_idle)
907 if (!cfq_cfqq_idle_window(cfqq))
910 * task has exited, don't wait
912 cic = cfqd->active_cic;
913 if (!cic || !cic->ioc->task)
916 cfq_mark_cfqq_must_dispatch(cfqq);
917 cfq_mark_cfqq_wait_request(cfqq);
919 sl = min(cfqq->slice_end - 1, (unsigned long) cfqd->cfq_slice_idle);
922 * we don't want to idle for seeks, but we do want to allow
923 * fair distribution of slice time for a process doing back-to-back
924 * seeks. so allow a little bit of time for him to submit a new rq
926 if (sample_valid(cic->seek_samples) && cic->seek_mean > 131072)
929 mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
933 static void cfq_dispatch_insert(request_queue_t *q, struct cfq_rq *crq)
935 struct cfq_data *cfqd = q->elevator->elevator_data;
936 struct cfq_queue *cfqq = crq->cfq_queue;
938 cfqq->next_crq = cfq_find_next_crq(cfqd, cfqq, crq);
939 cfq_remove_request(crq->request);
940 cfqq->on_dispatch[cfq_crq_is_sync(crq)]++;
941 elv_dispatch_sort(q, crq->request);
945 * return expired entry, or NULL to just start from scratch in rbtree
947 static inline struct cfq_rq *cfq_check_fifo(struct cfq_queue *cfqq)
949 struct cfq_data *cfqd = cfqq->cfqd;
953 if (cfq_cfqq_fifo_expire(cfqq))
956 if (!list_empty(&cfqq->fifo)) {
957 int fifo = cfq_cfqq_class_sync(cfqq);
959 crq = RQ_DATA(list_entry_fifo(cfqq->fifo.next));
961 if (time_after(jiffies, rq->start_time + cfqd->cfq_fifo_expire[fifo])) {
962 cfq_mark_cfqq_fifo_expire(cfqq);
971 * Scale schedule slice based on io priority. Use the sync time slice only
972 * if a queue is marked sync and has sync io queued. A sync queue with async
973 * io only, should not get full sync slice length.
976 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
978 const int base_slice = cfqd->cfq_slice[cfq_cfqq_sync(cfqq)];
980 WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
982 return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - cfqq->ioprio));
986 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
988 cfqq->slice_end = cfq_prio_to_slice(cfqd, cfqq) + jiffies;
992 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
994 const int base_rq = cfqd->cfq_slice_async_rq;
996 WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
998 return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
1002 * get next queue for service
1004 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
1006 unsigned long now = jiffies;
1007 struct cfq_queue *cfqq;
1009 cfqq = cfqd->active_queue;
1016 if (!cfq_cfqq_must_dispatch(cfqq) && time_after(now, cfqq->slice_end))
1020 * if queue has requests, dispatch one. if not, check if
1021 * enough slice is left to wait for one
1023 if (!RB_EMPTY(&cfqq->sort_list))
1025 else if (cfq_cfqq_class_sync(cfqq) &&
1026 time_before(now, cfqq->slice_end)) {
1027 if (cfq_arm_slice_timer(cfqd, cfqq))
1032 cfq_slice_expired(cfqd, 0);
1034 cfqq = cfq_set_active_queue(cfqd);
1040 __cfq_dispatch_requests(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1045 BUG_ON(RB_EMPTY(&cfqq->sort_list));
1051 * follow expired path, else get first next available
1053 if ((crq = cfq_check_fifo(cfqq)) == NULL)
1054 crq = cfqq->next_crq;
1057 * finally, insert request into driver dispatch list
1059 cfq_dispatch_insert(cfqd->queue, crq);
1061 cfqd->dispatch_slice++;
1064 if (!cfqd->active_cic) {
1065 atomic_inc(&crq->io_context->ioc->refcount);
1066 cfqd->active_cic = crq->io_context;
1069 if (RB_EMPTY(&cfqq->sort_list))
1072 } while (dispatched < max_dispatch);
1075 * if slice end isn't set yet, set it. if at least one request was
1076 * sync, use the sync time slice value
1078 if (!cfqq->slice_end)
1079 cfq_set_prio_slice(cfqd, cfqq);
1082 * expire an async queue immediately if it has used up its slice. idle
1083 * queue always expire after 1 dispatch round.
1085 if ((!cfq_cfqq_sync(cfqq) &&
1086 cfqd->dispatch_slice >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
1087 cfq_class_idle(cfqq))
1088 cfq_slice_expired(cfqd, 0);
1094 cfq_forced_dispatch_cfqqs(struct list_head *list)
1097 struct cfq_queue *cfqq, *next;
1100 list_for_each_entry_safe(cfqq, next, list, cfq_list) {
1101 while ((crq = cfqq->next_crq)) {
1102 cfq_dispatch_insert(cfqq->cfqd->queue, crq);
1105 BUG_ON(!list_empty(&cfqq->fifo));
1111 cfq_forced_dispatch(struct cfq_data *cfqd)
1113 int i, dispatched = 0;
1115 for (i = 0; i < CFQ_PRIO_LISTS; i++)
1116 dispatched += cfq_forced_dispatch_cfqqs(&cfqd->rr_list[i]);
1118 dispatched += cfq_forced_dispatch_cfqqs(&cfqd->busy_rr);
1119 dispatched += cfq_forced_dispatch_cfqqs(&cfqd->cur_rr);
1120 dispatched += cfq_forced_dispatch_cfqqs(&cfqd->idle_rr);
1122 cfq_slice_expired(cfqd, 0);
1124 BUG_ON(cfqd->busy_queues);
1130 cfq_dispatch_requests(request_queue_t *q, int force)
1132 struct cfq_data *cfqd = q->elevator->elevator_data;
1133 struct cfq_queue *cfqq;
1135 if (!cfqd->busy_queues)
1138 if (unlikely(force))
1139 return cfq_forced_dispatch(cfqd);
1141 cfqq = cfq_select_queue(cfqd);
1145 cfq_clear_cfqq_must_dispatch(cfqq);
1146 cfq_clear_cfqq_wait_request(cfqq);
1147 del_timer(&cfqd->idle_slice_timer);
1149 max_dispatch = cfqd->cfq_quantum;
1150 if (cfq_class_idle(cfqq))
1153 return __cfq_dispatch_requests(cfqd, cfqq, max_dispatch);
1160 * task holds one reference to the queue, dropped when task exits. each crq
1161 * in-flight on this queue also holds a reference, dropped when crq is freed.
1163 * queue lock must be held here.
1165 static void cfq_put_queue(struct cfq_queue *cfqq)
1167 struct cfq_data *cfqd = cfqq->cfqd;
1169 BUG_ON(atomic_read(&cfqq->ref) <= 0);
1171 if (!atomic_dec_and_test(&cfqq->ref))
1174 BUG_ON(rb_first(&cfqq->sort_list));
1175 BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
1176 BUG_ON(cfq_cfqq_on_rr(cfqq));
1178 if (unlikely(cfqd->active_queue == cfqq))
1179 __cfq_slice_expired(cfqd, cfqq, 0);
1182 * it's on the empty list and still hashed
1184 list_del(&cfqq->cfq_list);
1185 hlist_del(&cfqq->cfq_hash);
1186 kmem_cache_free(cfq_pool, cfqq);
1189 static inline struct cfq_queue *
1190 __cfq_find_cfq_hash(struct cfq_data *cfqd, unsigned int key, unsigned int prio,
1193 struct hlist_head *hash_list = &cfqd->cfq_hash[hashval];
1194 struct hlist_node *entry;
1195 struct cfq_queue *__cfqq;
1197 hlist_for_each_entry(__cfqq, entry, hash_list, cfq_hash) {
1198 const unsigned short __p = IOPRIO_PRIO_VALUE(__cfqq->org_ioprio_class, __cfqq->org_ioprio);
1200 if (__cfqq->key == key && (__p == prio || !prio))
1207 static struct cfq_queue *
1208 cfq_find_cfq_hash(struct cfq_data *cfqd, unsigned int key, unsigned short prio)
1210 return __cfq_find_cfq_hash(cfqd, key, prio, hash_long(key, CFQ_QHASH_SHIFT));
1213 static void cfq_free_io_context(struct io_context *ioc)
1215 struct cfq_io_context *__cic;
1219 while ((n = rb_first(&ioc->cic_root)) != NULL) {
1220 __cic = rb_entry(n, struct cfq_io_context, rb_node);
1221 rb_erase(&__cic->rb_node, &ioc->cic_root);
1222 kmem_cache_free(cfq_ioc_pool, __cic);
1226 if (atomic_sub_and_test(freed, &ioc_count) && ioc_gone)
1230 static void cfq_trim(struct io_context *ioc)
1232 ioc->set_ioprio = NULL;
1233 cfq_free_io_context(ioc);
1237 * Called with interrupts disabled
1239 static void cfq_exit_single_io_context(struct cfq_io_context *cic)
1241 struct cfq_data *cfqd = cic->key;
1249 WARN_ON(!irqs_disabled());
1251 spin_lock(q->queue_lock);
1253 if (cic->cfqq[ASYNC]) {
1254 if (unlikely(cic->cfqq[ASYNC] == cfqd->active_queue))
1255 __cfq_slice_expired(cfqd, cic->cfqq[ASYNC], 0);
1256 cfq_put_queue(cic->cfqq[ASYNC]);
1257 cic->cfqq[ASYNC] = NULL;
1260 if (cic->cfqq[SYNC]) {
1261 if (unlikely(cic->cfqq[SYNC] == cfqd->active_queue))
1262 __cfq_slice_expired(cfqd, cic->cfqq[SYNC], 0);
1263 cfq_put_queue(cic->cfqq[SYNC]);
1264 cic->cfqq[SYNC] = NULL;
1268 list_del_init(&cic->queue_list);
1269 spin_unlock(q->queue_lock);
1272 static void cfq_exit_io_context(struct io_context *ioc)
1274 struct cfq_io_context *__cic;
1275 unsigned long flags;
1279 * put the reference this task is holding to the various queues
1281 read_lock_irqsave(&cfq_exit_lock, flags);
1283 n = rb_first(&ioc->cic_root);
1285 __cic = rb_entry(n, struct cfq_io_context, rb_node);
1287 cfq_exit_single_io_context(__cic);
1291 read_unlock_irqrestore(&cfq_exit_lock, flags);
1294 static struct cfq_io_context *
1295 cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
1297 struct cfq_io_context *cic = kmem_cache_alloc(cfq_ioc_pool, gfp_mask);
1300 RB_CLEAR(&cic->rb_node);
1302 cic->cfqq[ASYNC] = NULL;
1303 cic->cfqq[SYNC] = NULL;
1304 cic->last_end_request = jiffies;
1305 cic->ttime_total = 0;
1306 cic->ttime_samples = 0;
1307 cic->ttime_mean = 0;
1308 cic->dtor = cfq_free_io_context;
1309 cic->exit = cfq_exit_io_context;
1310 INIT_LIST_HEAD(&cic->queue_list);
1311 atomic_inc(&ioc_count);
1317 static void cfq_init_prio_data(struct cfq_queue *cfqq)
1319 struct task_struct *tsk = current;
1322 if (!cfq_cfqq_prio_changed(cfqq))
1325 ioprio_class = IOPRIO_PRIO_CLASS(tsk->ioprio);
1326 switch (ioprio_class) {
1328 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
1329 case IOPRIO_CLASS_NONE:
1331 * no prio set, place us in the middle of the BE classes
1333 cfqq->ioprio = task_nice_ioprio(tsk);
1334 cfqq->ioprio_class = IOPRIO_CLASS_BE;
1336 case IOPRIO_CLASS_RT:
1337 cfqq->ioprio = task_ioprio(tsk);
1338 cfqq->ioprio_class = IOPRIO_CLASS_RT;
1340 case IOPRIO_CLASS_BE:
1341 cfqq->ioprio = task_ioprio(tsk);
1342 cfqq->ioprio_class = IOPRIO_CLASS_BE;
1344 case IOPRIO_CLASS_IDLE:
1345 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
1347 cfq_clear_cfqq_idle_window(cfqq);
1352 * keep track of original prio settings in case we have to temporarily
1353 * elevate the priority of this queue
1355 cfqq->org_ioprio = cfqq->ioprio;
1356 cfqq->org_ioprio_class = cfqq->ioprio_class;
1358 if (cfq_cfqq_on_rr(cfqq))
1359 cfq_resort_rr_list(cfqq, 0);
1361 cfq_clear_cfqq_prio_changed(cfqq);
1364 static inline void changed_ioprio(struct cfq_io_context *cic)
1366 struct cfq_data *cfqd = cic->key;
1367 struct cfq_queue *cfqq;
1369 spin_lock(cfqd->queue->queue_lock);
1370 cfqq = cic->cfqq[ASYNC];
1372 struct cfq_queue *new_cfqq;
1373 new_cfqq = cfq_get_queue(cfqd, CFQ_KEY_ASYNC,
1374 cic->ioc->task, GFP_ATOMIC);
1376 cic->cfqq[ASYNC] = new_cfqq;
1377 cfq_put_queue(cfqq);
1380 cfqq = cic->cfqq[SYNC];
1382 cfq_mark_cfqq_prio_changed(cfqq);
1383 cfq_init_prio_data(cfqq);
1385 spin_unlock(cfqd->queue->queue_lock);
1390 * callback from sys_ioprio_set, irqs are disabled
1392 static int cfq_ioc_set_ioprio(struct io_context *ioc, unsigned int ioprio)
1394 struct cfq_io_context *cic;
1397 write_lock(&cfq_exit_lock);
1399 n = rb_first(&ioc->cic_root);
1401 cic = rb_entry(n, struct cfq_io_context, rb_node);
1403 changed_ioprio(cic);
1407 write_unlock(&cfq_exit_lock);
1412 static struct cfq_queue *
1413 cfq_get_queue(struct cfq_data *cfqd, unsigned int key, struct task_struct *tsk,
1416 const int hashval = hash_long(key, CFQ_QHASH_SHIFT);
1417 struct cfq_queue *cfqq, *new_cfqq = NULL;
1418 unsigned short ioprio;
1421 ioprio = tsk->ioprio;
1422 cfqq = __cfq_find_cfq_hash(cfqd, key, ioprio, hashval);
1428 } else if (gfp_mask & __GFP_WAIT) {
1429 spin_unlock_irq(cfqd->queue->queue_lock);
1430 new_cfqq = kmem_cache_alloc(cfq_pool, gfp_mask);
1431 spin_lock_irq(cfqd->queue->queue_lock);
1434 cfqq = kmem_cache_alloc(cfq_pool, gfp_mask);
1439 memset(cfqq, 0, sizeof(*cfqq));
1441 INIT_HLIST_NODE(&cfqq->cfq_hash);
1442 INIT_LIST_HEAD(&cfqq->cfq_list);
1443 RB_CLEAR_ROOT(&cfqq->sort_list);
1444 INIT_LIST_HEAD(&cfqq->fifo);
1447 hlist_add_head(&cfqq->cfq_hash, &cfqd->cfq_hash[hashval]);
1448 atomic_set(&cfqq->ref, 0);
1450 cfqq->service_last = 0;
1452 * set ->slice_left to allow preemption for a new process
1454 cfqq->slice_left = 2 * cfqd->cfq_slice_idle;
1455 cfq_mark_cfqq_idle_window(cfqq);
1456 cfq_mark_cfqq_prio_changed(cfqq);
1457 cfq_init_prio_data(cfqq);
1461 kmem_cache_free(cfq_pool, new_cfqq);
1463 atomic_inc(&cfqq->ref);
1465 WARN_ON((gfp_mask & __GFP_WAIT) && !cfqq);
1470 cfq_drop_dead_cic(struct io_context *ioc, struct cfq_io_context *cic)
1472 read_lock(&cfq_exit_lock);
1473 rb_erase(&cic->rb_node, &ioc->cic_root);
1474 read_unlock(&cfq_exit_lock);
1475 kmem_cache_free(cfq_ioc_pool, cic);
1476 atomic_dec(&ioc_count);
1479 static struct cfq_io_context *
1480 cfq_cic_rb_lookup(struct cfq_data *cfqd, struct io_context *ioc)
1483 struct cfq_io_context *cic;
1484 void *k, *key = cfqd;
1487 n = ioc->cic_root.rb_node;
1489 cic = rb_entry(n, struct cfq_io_context, rb_node);
1490 /* ->key must be copied to avoid race with cfq_exit_queue() */
1493 cfq_drop_dead_cic(ioc, cic);
1509 cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
1510 struct cfq_io_context *cic)
1513 struct rb_node *parent;
1514 struct cfq_io_context *__cic;
1520 ioc->set_ioprio = cfq_ioc_set_ioprio;
1523 p = &ioc->cic_root.rb_node;
1526 __cic = rb_entry(parent, struct cfq_io_context, rb_node);
1527 /* ->key must be copied to avoid race with cfq_exit_queue() */
1530 cfq_drop_dead_cic(ioc, cic);
1536 else if (cic->key > k)
1537 p = &(*p)->rb_right;
1542 read_lock(&cfq_exit_lock);
1543 rb_link_node(&cic->rb_node, parent, p);
1544 rb_insert_color(&cic->rb_node, &ioc->cic_root);
1545 list_add(&cic->queue_list, &cfqd->cic_list);
1546 read_unlock(&cfq_exit_lock);
1550 * Setup general io context and cfq io context. There can be several cfq
1551 * io contexts per general io context, if this process is doing io to more
1552 * than one device managed by cfq.
1554 static struct cfq_io_context *
1555 cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
1557 struct io_context *ioc = NULL;
1558 struct cfq_io_context *cic;
1560 might_sleep_if(gfp_mask & __GFP_WAIT);
1562 ioc = get_io_context(gfp_mask);
1566 cic = cfq_cic_rb_lookup(cfqd, ioc);
1570 cic = cfq_alloc_io_context(cfqd, gfp_mask);
1574 cfq_cic_link(cfqd, ioc, cic);
1578 put_io_context(ioc);
1583 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
1585 unsigned long elapsed, ttime;
1588 * if this context already has stuff queued, thinktime is from
1589 * last queue not last end
1592 if (time_after(cic->last_end_request, cic->last_queue))
1593 elapsed = jiffies - cic->last_end_request;
1595 elapsed = jiffies - cic->last_queue;
1597 elapsed = jiffies - cic->last_end_request;
1600 ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
1602 cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
1603 cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
1604 cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
1608 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_io_context *cic,
1614 if (cic->last_request_pos < crq->request->sector)
1615 sdist = crq->request->sector - cic->last_request_pos;
1617 sdist = cic->last_request_pos - crq->request->sector;
1620 * Don't allow the seek distance to get too large from the
1621 * odd fragment, pagein, etc
1623 if (cic->seek_samples <= 60) /* second&third seek */
1624 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*1024);
1626 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*64);
1628 cic->seek_samples = (7*cic->seek_samples + 256) / 8;
1629 cic->seek_total = (7*cic->seek_total + (u64)256*sdist) / 8;
1630 total = cic->seek_total + (cic->seek_samples/2);
1631 do_div(total, cic->seek_samples);
1632 cic->seek_mean = (sector_t)total;
1636 * Disable idle window if the process thinks too long or seeks so much that
1640 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1641 struct cfq_io_context *cic)
1643 int enable_idle = cfq_cfqq_idle_window(cfqq);
1645 if (!cic->ioc->task || !cfqd->cfq_slice_idle)
1647 else if (sample_valid(cic->ttime_samples)) {
1648 if (cic->ttime_mean > cfqd->cfq_slice_idle)
1655 cfq_mark_cfqq_idle_window(cfqq);
1657 cfq_clear_cfqq_idle_window(cfqq);
1662 * Check if new_cfqq should preempt the currently active queue. Return 0 for
1663 * no or if we aren't sure, a 1 will cause a preempt.
1666 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
1669 struct cfq_queue *cfqq = cfqd->active_queue;
1671 if (cfq_class_idle(new_cfqq))
1677 if (cfq_class_idle(cfqq))
1679 if (!cfq_cfqq_wait_request(new_cfqq))
1682 * if it doesn't have slice left, forget it
1684 if (new_cfqq->slice_left < cfqd->cfq_slice_idle)
1686 if (cfq_crq_is_sync(crq) && !cfq_cfqq_sync(cfqq))
1693 * cfqq preempts the active queue. if we allowed preempt with no slice left,
1694 * let it have half of its nominal slice.
1696 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1698 struct cfq_queue *__cfqq, *next;
1700 list_for_each_entry_safe(__cfqq, next, &cfqd->cur_rr, cfq_list)
1701 cfq_resort_rr_list(__cfqq, 1);
1703 if (!cfqq->slice_left)
1704 cfqq->slice_left = cfq_prio_to_slice(cfqd, cfqq) / 2;
1706 cfqq->slice_end = cfqq->slice_left + jiffies;
1707 __cfq_slice_expired(cfqd, cfqq, 1);
1708 __cfq_set_active_queue(cfqd, cfqq);
1712 * should really be a ll_rw_blk.c helper
1714 static void cfq_start_queueing(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1716 request_queue_t *q = cfqd->queue;
1718 if (!blk_queue_plugged(q))
1721 __generic_unplug_device(q);
1725 * Called when a new fs request (crq) is added (to cfqq). Check if there's
1726 * something we should do about it
1729 cfq_crq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1732 struct cfq_io_context *cic;
1734 cfqq->next_crq = cfq_choose_req(cfqd, cfqq->next_crq, crq);
1737 * we never wait for an async request and we don't allow preemption
1738 * of an async request. so just return early
1740 if (!cfq_crq_is_sync(crq))
1743 cic = crq->io_context;
1745 cfq_update_io_thinktime(cfqd, cic);
1746 cfq_update_io_seektime(cfqd, cic, crq);
1747 cfq_update_idle_window(cfqd, cfqq, cic);
1749 cic->last_queue = jiffies;
1750 cic->last_request_pos = crq->request->sector + crq->request->nr_sectors;
1752 if (cfqq == cfqd->active_queue) {
1754 * if we are waiting for a request for this queue, let it rip
1755 * immediately and flag that we must not expire this queue
1758 if (cfq_cfqq_wait_request(cfqq)) {
1759 cfq_mark_cfqq_must_dispatch(cfqq);
1760 del_timer(&cfqd->idle_slice_timer);
1761 cfq_start_queueing(cfqd, cfqq);
1763 } else if (cfq_should_preempt(cfqd, cfqq, crq)) {
1765 * not the active queue - expire current slice if it is
1766 * idle and has expired it's mean thinktime or this new queue
1767 * has some old slice time left and is of higher priority
1769 cfq_preempt_queue(cfqd, cfqq);
1770 cfq_mark_cfqq_must_dispatch(cfqq);
1771 cfq_start_queueing(cfqd, cfqq);
1775 static void cfq_insert_request(request_queue_t *q, struct request *rq)
1777 struct cfq_data *cfqd = q->elevator->elevator_data;
1778 struct cfq_rq *crq = RQ_DATA(rq);
1779 struct cfq_queue *cfqq = crq->cfq_queue;
1781 cfq_init_prio_data(cfqq);
1783 cfq_add_crq_rb(crq);
1785 list_add_tail(&rq->queuelist, &cfqq->fifo);
1787 if (rq_mergeable(rq))
1788 cfq_add_crq_hash(cfqd, crq);
1790 cfq_crq_enqueued(cfqd, cfqq, crq);
1793 static void cfq_completed_request(request_queue_t *q, struct request *rq)
1795 struct cfq_rq *crq = RQ_DATA(rq);
1796 struct cfq_queue *cfqq = crq->cfq_queue;
1797 struct cfq_data *cfqd = cfqq->cfqd;
1798 const int sync = cfq_crq_is_sync(crq);
1803 WARN_ON(!cfqd->rq_in_driver);
1804 WARN_ON(!cfqq->on_dispatch[sync]);
1805 cfqd->rq_in_driver--;
1806 cfqq->on_dispatch[sync]--;
1808 if (!cfq_class_idle(cfqq))
1809 cfqd->last_end_request = now;
1811 if (!cfq_cfqq_dispatched(cfqq)) {
1812 if (cfq_cfqq_on_rr(cfqq)) {
1813 cfqq->service_last = now;
1814 cfq_resort_rr_list(cfqq, 0);
1816 cfq_schedule_dispatch(cfqd);
1819 if (cfq_crq_is_sync(crq))
1820 crq->io_context->last_end_request = now;
1823 static struct request *
1824 cfq_former_request(request_queue_t *q, struct request *rq)
1826 struct cfq_rq *crq = RQ_DATA(rq);
1827 struct rb_node *rbprev = rb_prev(&crq->rb_node);
1830 return rb_entry_crq(rbprev)->request;
1835 static struct request *
1836 cfq_latter_request(request_queue_t *q, struct request *rq)
1838 struct cfq_rq *crq = RQ_DATA(rq);
1839 struct rb_node *rbnext = rb_next(&crq->rb_node);
1842 return rb_entry_crq(rbnext)->request;
1848 * we temporarily boost lower priority queues if they are holding fs exclusive
1849 * resources. they are boosted to normal prio (CLASS_BE/4)
1851 static void cfq_prio_boost(struct cfq_queue *cfqq)
1853 const int ioprio_class = cfqq->ioprio_class;
1854 const int ioprio = cfqq->ioprio;
1856 if (has_fs_excl()) {
1858 * boost idle prio on transactions that would lock out other
1859 * users of the filesystem
1861 if (cfq_class_idle(cfqq))
1862 cfqq->ioprio_class = IOPRIO_CLASS_BE;
1863 if (cfqq->ioprio > IOPRIO_NORM)
1864 cfqq->ioprio = IOPRIO_NORM;
1867 * check if we need to unboost the queue
1869 if (cfqq->ioprio_class != cfqq->org_ioprio_class)
1870 cfqq->ioprio_class = cfqq->org_ioprio_class;
1871 if (cfqq->ioprio != cfqq->org_ioprio)
1872 cfqq->ioprio = cfqq->org_ioprio;
1876 * refile between round-robin lists if we moved the priority class
1878 if ((ioprio_class != cfqq->ioprio_class || ioprio != cfqq->ioprio) &&
1879 cfq_cfqq_on_rr(cfqq))
1880 cfq_resort_rr_list(cfqq, 0);
1884 __cfq_may_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1885 struct task_struct *task, int rw)
1888 if ((cfq_cfqq_wait_request(cfqq) || cfq_cfqq_must_alloc(cfqq)) &&
1889 !cfq_cfqq_must_alloc_slice(cfqq)) {
1890 cfq_mark_cfqq_must_alloc_slice(cfqq);
1891 return ELV_MQUEUE_MUST;
1894 return ELV_MQUEUE_MAY;
1896 if (!cfqq || task->flags & PF_MEMALLOC)
1897 return ELV_MQUEUE_MAY;
1898 if (!cfqq->allocated[rw] || cfq_cfqq_must_alloc(cfqq)) {
1899 if (cfq_cfqq_wait_request(cfqq))
1900 return ELV_MQUEUE_MUST;
1903 * only allow 1 ELV_MQUEUE_MUST per slice, otherwise we
1904 * can quickly flood the queue with writes from a single task
1906 if (rw == READ || !cfq_cfqq_must_alloc_slice(cfqq)) {
1907 cfq_mark_cfqq_must_alloc_slice(cfqq);
1908 return ELV_MQUEUE_MUST;
1911 return ELV_MQUEUE_MAY;
1913 if (cfq_class_idle(cfqq))
1914 return ELV_MQUEUE_NO;
1915 if (cfqq->allocated[rw] >= cfqd->max_queued) {
1916 struct io_context *ioc = get_io_context(GFP_ATOMIC);
1917 int ret = ELV_MQUEUE_NO;
1919 if (ioc && ioc->nr_batch_requests)
1920 ret = ELV_MQUEUE_MAY;
1922 put_io_context(ioc);
1926 return ELV_MQUEUE_MAY;
1930 static int cfq_may_queue(request_queue_t *q, int rw, struct bio *bio)
1932 struct cfq_data *cfqd = q->elevator->elevator_data;
1933 struct task_struct *tsk = current;
1934 struct cfq_queue *cfqq;
1937 * don't force setup of a queue from here, as a call to may_queue
1938 * does not necessarily imply that a request actually will be queued.
1939 * so just lookup a possibly existing queue, or return 'may queue'
1942 cfqq = cfq_find_cfq_hash(cfqd, cfq_queue_pid(tsk, rw), tsk->ioprio);
1944 cfq_init_prio_data(cfqq);
1945 cfq_prio_boost(cfqq);
1947 return __cfq_may_queue(cfqd, cfqq, tsk, rw);
1950 return ELV_MQUEUE_MAY;
1953 static void cfq_check_waiters(request_queue_t *q, struct cfq_queue *cfqq)
1955 struct cfq_data *cfqd = q->elevator->elevator_data;
1956 struct request_list *rl = &q->rq;
1958 if (cfqq->allocated[READ] <= cfqd->max_queued || cfqd->rq_starved) {
1960 if (waitqueue_active(&rl->wait[READ]))
1961 wake_up(&rl->wait[READ]);
1964 if (cfqq->allocated[WRITE] <= cfqd->max_queued || cfqd->rq_starved) {
1966 if (waitqueue_active(&rl->wait[WRITE]))
1967 wake_up(&rl->wait[WRITE]);
1972 * queue lock held here
1974 static void cfq_put_request(request_queue_t *q, struct request *rq)
1976 struct cfq_data *cfqd = q->elevator->elevator_data;
1977 struct cfq_rq *crq = RQ_DATA(rq);
1980 struct cfq_queue *cfqq = crq->cfq_queue;
1981 const int rw = rq_data_dir(rq);
1983 BUG_ON(!cfqq->allocated[rw]);
1984 cfqq->allocated[rw]--;
1986 put_io_context(crq->io_context->ioc);
1988 mempool_free(crq, cfqd->crq_pool);
1989 rq->elevator_private = NULL;
1991 cfq_check_waiters(q, cfqq);
1992 cfq_put_queue(cfqq);
1997 * Allocate cfq data structures associated with this request.
2000 cfq_set_request(request_queue_t *q, struct request *rq, struct bio *bio,
2003 struct cfq_data *cfqd = q->elevator->elevator_data;
2004 struct task_struct *tsk = current;
2005 struct cfq_io_context *cic;
2006 const int rw = rq_data_dir(rq);
2007 pid_t key = cfq_queue_pid(tsk, rw);
2008 struct cfq_queue *cfqq;
2010 unsigned long flags;
2011 int is_sync = key != CFQ_KEY_ASYNC;
2013 might_sleep_if(gfp_mask & __GFP_WAIT);
2015 cic = cfq_get_io_context(cfqd, gfp_mask);
2017 spin_lock_irqsave(q->queue_lock, flags);
2022 if (!cic->cfqq[is_sync]) {
2023 cfqq = cfq_get_queue(cfqd, key, tsk, gfp_mask);
2027 cic->cfqq[is_sync] = cfqq;
2029 cfqq = cic->cfqq[is_sync];
2031 cfqq->allocated[rw]++;
2032 cfq_clear_cfqq_must_alloc(cfqq);
2033 cfqd->rq_starved = 0;
2034 atomic_inc(&cfqq->ref);
2035 spin_unlock_irqrestore(q->queue_lock, flags);
2037 crq = mempool_alloc(cfqd->crq_pool, gfp_mask);
2039 RB_CLEAR(&crq->rb_node);
2042 INIT_HLIST_NODE(&crq->hash);
2043 crq->cfq_queue = cfqq;
2044 crq->io_context = cic;
2047 cfq_mark_crq_is_sync(crq);
2049 cfq_clear_crq_is_sync(crq);
2051 rq->elevator_private = crq;
2055 spin_lock_irqsave(q->queue_lock, flags);
2056 cfqq->allocated[rw]--;
2057 if (!(cfqq->allocated[0] + cfqq->allocated[1]))
2058 cfq_mark_cfqq_must_alloc(cfqq);
2059 cfq_put_queue(cfqq);
2062 put_io_context(cic->ioc);
2064 * mark us rq allocation starved. we need to kickstart the process
2065 * ourselves if there are no pending requests that can do it for us.
2066 * that would be an extremely rare OOM situation
2068 cfqd->rq_starved = 1;
2069 cfq_schedule_dispatch(cfqd);
2070 spin_unlock_irqrestore(q->queue_lock, flags);
2074 static void cfq_kick_queue(void *data)
2076 request_queue_t *q = data;
2077 struct cfq_data *cfqd = q->elevator->elevator_data;
2078 unsigned long flags;
2080 spin_lock_irqsave(q->queue_lock, flags);
2082 if (cfqd->rq_starved) {
2083 struct request_list *rl = &q->rq;
2086 * we aren't guaranteed to get a request after this, but we
2087 * have to be opportunistic
2090 if (waitqueue_active(&rl->wait[READ]))
2091 wake_up(&rl->wait[READ]);
2092 if (waitqueue_active(&rl->wait[WRITE]))
2093 wake_up(&rl->wait[WRITE]);
2098 spin_unlock_irqrestore(q->queue_lock, flags);
2102 * Timer running if the active_queue is currently idling inside its time slice
2104 static void cfq_idle_slice_timer(unsigned long data)
2106 struct cfq_data *cfqd = (struct cfq_data *) data;
2107 struct cfq_queue *cfqq;
2108 unsigned long flags;
2110 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2112 if ((cfqq = cfqd->active_queue) != NULL) {
2113 unsigned long now = jiffies;
2118 if (time_after(now, cfqq->slice_end))
2122 * only expire and reinvoke request handler, if there are
2123 * other queues with pending requests
2125 if (!cfqd->busy_queues) {
2126 cfqd->idle_slice_timer.expires = min(now + cfqd->cfq_slice_idle, cfqq->slice_end);
2127 add_timer(&cfqd->idle_slice_timer);
2132 * not expired and it has a request pending, let it dispatch
2134 if (!RB_EMPTY(&cfqq->sort_list)) {
2135 cfq_mark_cfqq_must_dispatch(cfqq);
2140 cfq_slice_expired(cfqd, 0);
2142 cfq_schedule_dispatch(cfqd);
2144 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2148 * Timer running if an idle class queue is waiting for service
2150 static void cfq_idle_class_timer(unsigned long data)
2152 struct cfq_data *cfqd = (struct cfq_data *) data;
2153 unsigned long flags, end;
2155 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2158 * race with a non-idle queue, reset timer
2160 end = cfqd->last_end_request + CFQ_IDLE_GRACE;
2161 if (!time_after_eq(jiffies, end)) {
2162 cfqd->idle_class_timer.expires = end;
2163 add_timer(&cfqd->idle_class_timer);
2165 cfq_schedule_dispatch(cfqd);
2167 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2170 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
2172 del_timer_sync(&cfqd->idle_slice_timer);
2173 del_timer_sync(&cfqd->idle_class_timer);
2174 blk_sync_queue(cfqd->queue);
2177 static void cfq_exit_queue(elevator_t *e)
2179 struct cfq_data *cfqd = e->elevator_data;
2180 request_queue_t *q = cfqd->queue;
2182 cfq_shutdown_timer_wq(cfqd);
2184 write_lock(&cfq_exit_lock);
2185 spin_lock_irq(q->queue_lock);
2187 if (cfqd->active_queue)
2188 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
2190 while (!list_empty(&cfqd->cic_list)) {
2191 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
2192 struct cfq_io_context,
2194 if (cic->cfqq[ASYNC]) {
2195 cfq_put_queue(cic->cfqq[ASYNC]);
2196 cic->cfqq[ASYNC] = NULL;
2198 if (cic->cfqq[SYNC]) {
2199 cfq_put_queue(cic->cfqq[SYNC]);
2200 cic->cfqq[SYNC] = NULL;
2203 list_del_init(&cic->queue_list);
2206 spin_unlock_irq(q->queue_lock);
2207 write_unlock(&cfq_exit_lock);
2209 cfq_shutdown_timer_wq(cfqd);
2211 mempool_destroy(cfqd->crq_pool);
2212 kfree(cfqd->crq_hash);
2213 kfree(cfqd->cfq_hash);
2217 static int cfq_init_queue(request_queue_t *q, elevator_t *e)
2219 struct cfq_data *cfqd;
2222 cfqd = kmalloc(sizeof(*cfqd), GFP_KERNEL);
2226 memset(cfqd, 0, sizeof(*cfqd));
2228 for (i = 0; i < CFQ_PRIO_LISTS; i++)
2229 INIT_LIST_HEAD(&cfqd->rr_list[i]);
2231 INIT_LIST_HEAD(&cfqd->busy_rr);
2232 INIT_LIST_HEAD(&cfqd->cur_rr);
2233 INIT_LIST_HEAD(&cfqd->idle_rr);
2234 INIT_LIST_HEAD(&cfqd->empty_list);
2235 INIT_LIST_HEAD(&cfqd->cic_list);
2237 cfqd->crq_hash = kmalloc(sizeof(struct hlist_head) * CFQ_MHASH_ENTRIES, GFP_KERNEL);
2238 if (!cfqd->crq_hash)
2241 cfqd->cfq_hash = kmalloc(sizeof(struct hlist_head) * CFQ_QHASH_ENTRIES, GFP_KERNEL);
2242 if (!cfqd->cfq_hash)
2245 cfqd->crq_pool = mempool_create_slab_pool(BLKDEV_MIN_RQ, crq_pool);
2246 if (!cfqd->crq_pool)
2249 for (i = 0; i < CFQ_MHASH_ENTRIES; i++)
2250 INIT_HLIST_HEAD(&cfqd->crq_hash[i]);
2251 for (i = 0; i < CFQ_QHASH_ENTRIES; i++)
2252 INIT_HLIST_HEAD(&cfqd->cfq_hash[i]);
2254 e->elevator_data = cfqd;
2258 cfqd->max_queued = q->nr_requests / 4;
2259 q->nr_batching = cfq_queued;
2261 init_timer(&cfqd->idle_slice_timer);
2262 cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
2263 cfqd->idle_slice_timer.data = (unsigned long) cfqd;
2265 init_timer(&cfqd->idle_class_timer);
2266 cfqd->idle_class_timer.function = cfq_idle_class_timer;
2267 cfqd->idle_class_timer.data = (unsigned long) cfqd;
2269 INIT_WORK(&cfqd->unplug_work, cfq_kick_queue, q);
2271 cfqd->cfq_queued = cfq_queued;
2272 cfqd->cfq_quantum = cfq_quantum;
2273 cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
2274 cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
2275 cfqd->cfq_back_max = cfq_back_max;
2276 cfqd->cfq_back_penalty = cfq_back_penalty;
2277 cfqd->cfq_slice[0] = cfq_slice_async;
2278 cfqd->cfq_slice[1] = cfq_slice_sync;
2279 cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
2280 cfqd->cfq_slice_idle = cfq_slice_idle;
2284 kfree(cfqd->cfq_hash);
2286 kfree(cfqd->crq_hash);
2292 static void cfq_slab_kill(void)
2295 kmem_cache_destroy(crq_pool);
2297 kmem_cache_destroy(cfq_pool);
2299 kmem_cache_destroy(cfq_ioc_pool);
2302 static int __init cfq_slab_setup(void)
2304 crq_pool = kmem_cache_create("crq_pool", sizeof(struct cfq_rq), 0, 0,
2309 cfq_pool = kmem_cache_create("cfq_pool", sizeof(struct cfq_queue), 0, 0,
2314 cfq_ioc_pool = kmem_cache_create("cfq_ioc_pool",
2315 sizeof(struct cfq_io_context), 0, 0, NULL, NULL);
2326 * sysfs parts below -->
2330 cfq_var_show(unsigned int var, char *page)
2332 return sprintf(page, "%d\n", var);
2336 cfq_var_store(unsigned int *var, const char *page, size_t count)
2338 char *p = (char *) page;
2340 *var = simple_strtoul(p, &p, 10);
2344 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
2345 static ssize_t __FUNC(elevator_t *e, char *page) \
2347 struct cfq_data *cfqd = e->elevator_data; \
2348 unsigned int __data = __VAR; \
2350 __data = jiffies_to_msecs(__data); \
2351 return cfq_var_show(__data, (page)); \
2353 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
2354 SHOW_FUNCTION(cfq_queued_show, cfqd->cfq_queued, 0);
2355 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
2356 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
2357 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
2358 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
2359 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
2360 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
2361 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
2362 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
2363 #undef SHOW_FUNCTION
2365 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
2366 static ssize_t __FUNC(elevator_t *e, const char *page, size_t count) \
2368 struct cfq_data *cfqd = e->elevator_data; \
2369 unsigned int __data; \
2370 int ret = cfq_var_store(&__data, (page), count); \
2371 if (__data < (MIN)) \
2373 else if (__data > (MAX)) \
2376 *(__PTR) = msecs_to_jiffies(__data); \
2378 *(__PTR) = __data; \
2381 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
2382 STORE_FUNCTION(cfq_queued_store, &cfqd->cfq_queued, 1, UINT_MAX, 0);
2383 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1, UINT_MAX, 1);
2384 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1, UINT_MAX, 1);
2385 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
2386 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1, UINT_MAX, 0);
2387 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
2388 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
2389 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
2390 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1, UINT_MAX, 0);
2391 #undef STORE_FUNCTION
2393 #define CFQ_ATTR(name) \
2394 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
2396 static struct elv_fs_entry cfq_attrs[] = {
2399 CFQ_ATTR(fifo_expire_sync),
2400 CFQ_ATTR(fifo_expire_async),
2401 CFQ_ATTR(back_seek_max),
2402 CFQ_ATTR(back_seek_penalty),
2403 CFQ_ATTR(slice_sync),
2404 CFQ_ATTR(slice_async),
2405 CFQ_ATTR(slice_async_rq),
2406 CFQ_ATTR(slice_idle),
2410 static struct elevator_type iosched_cfq = {
2412 .elevator_merge_fn = cfq_merge,
2413 .elevator_merged_fn = cfq_merged_request,
2414 .elevator_merge_req_fn = cfq_merged_requests,
2415 .elevator_dispatch_fn = cfq_dispatch_requests,
2416 .elevator_add_req_fn = cfq_insert_request,
2417 .elevator_activate_req_fn = cfq_activate_request,
2418 .elevator_deactivate_req_fn = cfq_deactivate_request,
2419 .elevator_queue_empty_fn = cfq_queue_empty,
2420 .elevator_completed_req_fn = cfq_completed_request,
2421 .elevator_former_req_fn = cfq_former_request,
2422 .elevator_latter_req_fn = cfq_latter_request,
2423 .elevator_set_req_fn = cfq_set_request,
2424 .elevator_put_req_fn = cfq_put_request,
2425 .elevator_may_queue_fn = cfq_may_queue,
2426 .elevator_init_fn = cfq_init_queue,
2427 .elevator_exit_fn = cfq_exit_queue,
2430 .elevator_attrs = cfq_attrs,
2431 .elevator_name = "cfq",
2432 .elevator_owner = THIS_MODULE,
2435 static int __init cfq_init(void)
2440 * could be 0 on HZ < 1000 setups
2442 if (!cfq_slice_async)
2443 cfq_slice_async = 1;
2444 if (!cfq_slice_idle)
2447 if (cfq_slab_setup())
2450 ret = elv_register(&iosched_cfq);
2457 static void __exit cfq_exit(void)
2459 DECLARE_COMPLETION(all_gone);
2460 elv_unregister(&iosched_cfq);
2461 ioc_gone = &all_gone;
2462 /* ioc_gone's update must be visible before reading ioc_count */
2464 if (atomic_read(&ioc_count))
2465 wait_for_completion(ioc_gone);
2470 module_init(cfq_init);
2471 module_exit(cfq_exit);
2473 MODULE_AUTHOR("Jens Axboe");
2474 MODULE_LICENSE("GPL");
2475 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");