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/kernel.h>
11 #include <linux/blkdev.h>
12 #include <linux/elevator.h>
13 #include <linux/bio.h>
14 #include <linux/config.h>
15 #include <linux/module.h>
16 #include <linux/slab.h>
17 #include <linux/init.h>
18 #include <linux/compiler.h>
19 #include <linux/hash.h>
20 #include <linux/rbtree.h>
21 #include <linux/mempool.h>
22 #include <linux/ioprio.h>
23 #include <linux/writeback.h>
28 static const int cfq_quantum = 4; /* max queue in one round of service */
29 static const int cfq_queued = 8; /* minimum rq allocate limit per-queue*/
30 static const int cfq_fifo_expire[2] = { HZ / 4, HZ / 8 };
31 static const int cfq_back_max = 16 * 1024; /* maximum backwards seek, in KiB */
32 static const int cfq_back_penalty = 2; /* penalty of a backwards seek */
34 static const int cfq_slice_sync = HZ / 10;
35 static int cfq_slice_async = HZ / 25;
36 static const int cfq_slice_async_rq = 2;
37 static int cfq_slice_idle = HZ / 100;
39 #define CFQ_IDLE_GRACE (HZ / 10)
40 #define CFQ_SLICE_SCALE (5)
42 #define CFQ_KEY_ASYNC (0)
43 #define CFQ_KEY_ANY (0xffff)
46 * disable queueing at the driver/hardware level
48 static const int cfq_max_depth = 2;
50 static DEFINE_RWLOCK(cfq_exit_lock);
53 * for the hash of cfqq inside the cfqd
55 #define CFQ_QHASH_SHIFT 6
56 #define CFQ_QHASH_ENTRIES (1 << CFQ_QHASH_SHIFT)
57 #define list_entry_qhash(entry) hlist_entry((entry), struct cfq_queue, cfq_hash)
60 * for the hash of crq inside the cfqq
62 #define CFQ_MHASH_SHIFT 6
63 #define CFQ_MHASH_BLOCK(sec) ((sec) >> 3)
64 #define CFQ_MHASH_ENTRIES (1 << CFQ_MHASH_SHIFT)
65 #define CFQ_MHASH_FN(sec) hash_long(CFQ_MHASH_BLOCK(sec), CFQ_MHASH_SHIFT)
66 #define rq_hash_key(rq) ((rq)->sector + (rq)->nr_sectors)
67 #define list_entry_hash(ptr) hlist_entry((ptr), struct cfq_rq, hash)
69 #define list_entry_cfqq(ptr) list_entry((ptr), struct cfq_queue, cfq_list)
70 #define list_entry_fifo(ptr) list_entry((ptr), struct request, queuelist)
72 #define RQ_DATA(rq) (rq)->elevator_private
78 #define RB_EMPTY(node) ((node)->rb_node == NULL)
79 #define RB_CLEAR_COLOR(node) (node)->rb_color = RB_NONE
80 #define RB_CLEAR(node) do { \
81 (node)->rb_parent = NULL; \
82 RB_CLEAR_COLOR((node)); \
83 (node)->rb_right = NULL; \
84 (node)->rb_left = NULL; \
86 #define RB_CLEAR_ROOT(root) ((root)->rb_node = NULL)
87 #define rb_entry_crq(node) rb_entry((node), struct cfq_rq, rb_node)
88 #define rq_rb_key(rq) (rq)->sector
90 static kmem_cache_t *crq_pool;
91 static kmem_cache_t *cfq_pool;
92 static kmem_cache_t *cfq_ioc_pool;
94 static atomic_t ioc_count = ATOMIC_INIT(0);
95 static struct completion *ioc_gone;
97 #define CFQ_PRIO_LISTS IOPRIO_BE_NR
98 #define cfq_class_idle(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
99 #define cfq_class_be(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_BE)
100 #define cfq_class_rt(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
105 #define cfq_cfqq_dispatched(cfqq) \
106 ((cfqq)->on_dispatch[ASYNC] + (cfqq)->on_dispatch[SYNC])
108 #define cfq_cfqq_class_sync(cfqq) ((cfqq)->key != CFQ_KEY_ASYNC)
110 #define cfq_cfqq_sync(cfqq) \
111 (cfq_cfqq_class_sync(cfqq) || (cfqq)->on_dispatch[SYNC])
114 * Per block device queue structure
117 request_queue_t *queue;
120 * rr list of queues with requests and the count of them
122 struct list_head rr_list[CFQ_PRIO_LISTS];
123 struct list_head busy_rr;
124 struct list_head cur_rr;
125 struct list_head idle_rr;
126 unsigned int busy_queues;
129 * non-ordered list of empty cfqq's
131 struct list_head empty_list;
136 struct hlist_head *cfq_hash;
139 * global crq hash for all queues
141 struct hlist_head *crq_hash;
143 unsigned int max_queued;
150 * schedule slice state info
153 * idle window management
155 struct timer_list idle_slice_timer;
156 struct work_struct unplug_work;
158 struct cfq_queue *active_queue;
159 struct cfq_io_context *active_cic;
160 int cur_prio, cur_end_prio;
161 unsigned int dispatch_slice;
163 struct timer_list idle_class_timer;
165 sector_t last_sector;
166 unsigned long last_end_request;
168 unsigned int rq_starved;
171 * tunables, see top of file
173 unsigned int cfq_quantum;
174 unsigned int cfq_queued;
175 unsigned int cfq_fifo_expire[2];
176 unsigned int cfq_back_penalty;
177 unsigned int cfq_back_max;
178 unsigned int cfq_slice[2];
179 unsigned int cfq_slice_async_rq;
180 unsigned int cfq_slice_idle;
181 unsigned int cfq_max_depth;
183 struct list_head cic_list;
187 * Per process-grouping structure
190 /* reference count */
192 /* parent cfq_data */
193 struct cfq_data *cfqd;
194 /* cfqq lookup hash */
195 struct hlist_node cfq_hash;
198 /* on either rr or empty list of cfqd */
199 struct list_head cfq_list;
200 /* sorted list of pending requests */
201 struct rb_root sort_list;
202 /* if fifo isn't expired, next request to serve */
203 struct cfq_rq *next_crq;
204 /* requests queued in sort_list */
206 /* currently allocated requests */
208 /* fifo list of requests in sort_list */
209 struct list_head fifo;
211 unsigned long slice_start;
212 unsigned long slice_end;
213 unsigned long slice_left;
214 unsigned long service_last;
216 /* number of requests that are on the dispatch list */
219 /* io prio of this group */
220 unsigned short ioprio, org_ioprio;
221 unsigned short ioprio_class, org_ioprio_class;
223 /* various state flags, see below */
228 struct rb_node rb_node;
230 struct request *request;
231 struct hlist_node hash;
233 struct cfq_queue *cfq_queue;
234 struct cfq_io_context *io_context;
236 unsigned int crq_flags;
239 enum cfqq_state_flags {
240 CFQ_CFQQ_FLAG_on_rr = 0,
241 CFQ_CFQQ_FLAG_wait_request,
242 CFQ_CFQQ_FLAG_must_alloc,
243 CFQ_CFQQ_FLAG_must_alloc_slice,
244 CFQ_CFQQ_FLAG_must_dispatch,
245 CFQ_CFQQ_FLAG_fifo_expire,
246 CFQ_CFQQ_FLAG_idle_window,
247 CFQ_CFQQ_FLAG_prio_changed,
250 #define CFQ_CFQQ_FNS(name) \
251 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \
253 cfqq->flags |= (1 << CFQ_CFQQ_FLAG_##name); \
255 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \
257 cfqq->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \
259 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \
261 return (cfqq->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \
265 CFQ_CFQQ_FNS(wait_request);
266 CFQ_CFQQ_FNS(must_alloc);
267 CFQ_CFQQ_FNS(must_alloc_slice);
268 CFQ_CFQQ_FNS(must_dispatch);
269 CFQ_CFQQ_FNS(fifo_expire);
270 CFQ_CFQQ_FNS(idle_window);
271 CFQ_CFQQ_FNS(prio_changed);
274 enum cfq_rq_state_flags {
275 CFQ_CRQ_FLAG_is_sync = 0,
278 #define CFQ_CRQ_FNS(name) \
279 static inline void cfq_mark_crq_##name(struct cfq_rq *crq) \
281 crq->crq_flags |= (1 << CFQ_CRQ_FLAG_##name); \
283 static inline void cfq_clear_crq_##name(struct cfq_rq *crq) \
285 crq->crq_flags &= ~(1 << CFQ_CRQ_FLAG_##name); \
287 static inline int cfq_crq_##name(const struct cfq_rq *crq) \
289 return (crq->crq_flags & (1 << CFQ_CRQ_FLAG_##name)) != 0; \
292 CFQ_CRQ_FNS(is_sync);
295 static struct cfq_queue *cfq_find_cfq_hash(struct cfq_data *, unsigned int, unsigned short);
296 static void cfq_dispatch_insert(request_queue_t *, struct cfq_rq *);
297 static struct cfq_queue *cfq_get_queue(struct cfq_data *cfqd, unsigned int key, struct task_struct *tsk, gfp_t gfp_mask);
299 #define process_sync(tsk) ((tsk)->flags & PF_SYNCWRITE)
302 * lots of deadline iosched dupes, can be abstracted later...
304 static inline void cfq_del_crq_hash(struct cfq_rq *crq)
306 hlist_del_init(&crq->hash);
309 static inline void cfq_add_crq_hash(struct cfq_data *cfqd, struct cfq_rq *crq)
311 const int hash_idx = CFQ_MHASH_FN(rq_hash_key(crq->request));
313 hlist_add_head(&crq->hash, &cfqd->crq_hash[hash_idx]);
316 static struct request *cfq_find_rq_hash(struct cfq_data *cfqd, sector_t offset)
318 struct hlist_head *hash_list = &cfqd->crq_hash[CFQ_MHASH_FN(offset)];
319 struct hlist_node *entry, *next;
321 hlist_for_each_safe(entry, next, hash_list) {
322 struct cfq_rq *crq = list_entry_hash(entry);
323 struct request *__rq = crq->request;
325 if (!rq_mergeable(__rq)) {
326 cfq_del_crq_hash(crq);
330 if (rq_hash_key(__rq) == offset)
338 * scheduler run of queue, if there are requests pending and no one in the
339 * driver that will restart queueing
341 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
343 if (cfqd->busy_queues)
344 kblockd_schedule_work(&cfqd->unplug_work);
347 static int cfq_queue_empty(request_queue_t *q)
349 struct cfq_data *cfqd = q->elevator->elevator_data;
351 return !cfqd->busy_queues;
355 * Lifted from AS - choose which of crq1 and crq2 that is best served now.
356 * We choose the request that is closest to the head right now. Distance
357 * behind the head are penalized and only allowed to a certain extent.
359 static struct cfq_rq *
360 cfq_choose_req(struct cfq_data *cfqd, struct cfq_rq *crq1, struct cfq_rq *crq2)
362 sector_t last, s1, s2, d1 = 0, d2 = 0;
363 int r1_wrap = 0, r2_wrap = 0; /* requests are behind the disk head */
364 unsigned long back_max;
366 if (crq1 == NULL || crq1 == crq2)
371 if (cfq_crq_is_sync(crq1) && !cfq_crq_is_sync(crq2))
373 else if (cfq_crq_is_sync(crq2) && !cfq_crq_is_sync(crq1))
376 s1 = crq1->request->sector;
377 s2 = crq2->request->sector;
379 last = cfqd->last_sector;
382 * by definition, 1KiB is 2 sectors
384 back_max = cfqd->cfq_back_max * 2;
387 * Strict one way elevator _except_ in the case where we allow
388 * short backward seeks which are biased as twice the cost of a
389 * similar forward seek.
393 else if (s1 + back_max >= last)
394 d1 = (last - s1) * cfqd->cfq_back_penalty;
400 else if (s2 + back_max >= last)
401 d2 = (last - s2) * cfqd->cfq_back_penalty;
405 /* Found required data */
406 if (!r1_wrap && r2_wrap)
408 else if (!r2_wrap && r1_wrap)
410 else if (r1_wrap && r2_wrap) {
411 /* both behind the head */
418 /* Both requests in front of the head */
432 * would be nice to take fifo expire time into account as well
434 static struct cfq_rq *
435 cfq_find_next_crq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
438 struct cfq_rq *crq_next = NULL, *crq_prev = NULL;
439 struct rb_node *rbnext, *rbprev;
441 if (!(rbnext = rb_next(&last->rb_node))) {
442 rbnext = rb_first(&cfqq->sort_list);
443 if (rbnext == &last->rb_node)
447 rbprev = rb_prev(&last->rb_node);
450 crq_prev = rb_entry_crq(rbprev);
452 crq_next = rb_entry_crq(rbnext);
454 return cfq_choose_req(cfqd, crq_next, crq_prev);
457 static void cfq_update_next_crq(struct cfq_rq *crq)
459 struct cfq_queue *cfqq = crq->cfq_queue;
461 if (cfqq->next_crq == crq)
462 cfqq->next_crq = cfq_find_next_crq(cfqq->cfqd, cfqq, crq);
465 static void cfq_resort_rr_list(struct cfq_queue *cfqq, int preempted)
467 struct cfq_data *cfqd = cfqq->cfqd;
468 struct list_head *list, *entry;
470 BUG_ON(!cfq_cfqq_on_rr(cfqq));
472 list_del(&cfqq->cfq_list);
474 if (cfq_class_rt(cfqq))
475 list = &cfqd->cur_rr;
476 else if (cfq_class_idle(cfqq))
477 list = &cfqd->idle_rr;
480 * if cfqq has requests in flight, don't allow it to be
481 * found in cfq_set_active_queue before it has finished them.
482 * this is done to increase fairness between a process that
483 * has lots of io pending vs one that only generates one
484 * sporadically or synchronously
486 if (cfq_cfqq_dispatched(cfqq))
487 list = &cfqd->busy_rr;
489 list = &cfqd->rr_list[cfqq->ioprio];
493 * if queue was preempted, just add to front to be fair. busy_rr
496 if (preempted || list == &cfqd->busy_rr) {
497 list_add(&cfqq->cfq_list, list);
502 * sort by when queue was last serviced
505 while ((entry = entry->prev) != list) {
506 struct cfq_queue *__cfqq = list_entry_cfqq(entry);
508 if (!__cfqq->service_last)
510 if (time_before(__cfqq->service_last, cfqq->service_last))
514 list_add(&cfqq->cfq_list, entry);
518 * add to busy list of queues for service, trying to be fair in ordering
519 * the pending list according to last request service
522 cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
524 BUG_ON(cfq_cfqq_on_rr(cfqq));
525 cfq_mark_cfqq_on_rr(cfqq);
528 cfq_resort_rr_list(cfqq, 0);
532 cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
534 BUG_ON(!cfq_cfqq_on_rr(cfqq));
535 cfq_clear_cfqq_on_rr(cfqq);
536 list_move(&cfqq->cfq_list, &cfqd->empty_list);
538 BUG_ON(!cfqd->busy_queues);
543 * rb tree support functions
545 static inline void cfq_del_crq_rb(struct cfq_rq *crq)
547 struct cfq_queue *cfqq = crq->cfq_queue;
548 struct cfq_data *cfqd = cfqq->cfqd;
549 const int sync = cfq_crq_is_sync(crq);
551 BUG_ON(!cfqq->queued[sync]);
552 cfqq->queued[sync]--;
554 cfq_update_next_crq(crq);
556 rb_erase(&crq->rb_node, &cfqq->sort_list);
557 RB_CLEAR_COLOR(&crq->rb_node);
559 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY(&cfqq->sort_list))
560 cfq_del_cfqq_rr(cfqd, cfqq);
563 static struct cfq_rq *
564 __cfq_add_crq_rb(struct cfq_rq *crq)
566 struct rb_node **p = &crq->cfq_queue->sort_list.rb_node;
567 struct rb_node *parent = NULL;
568 struct cfq_rq *__crq;
572 __crq = rb_entry_crq(parent);
574 if (crq->rb_key < __crq->rb_key)
576 else if (crq->rb_key > __crq->rb_key)
582 rb_link_node(&crq->rb_node, parent, p);
586 static void cfq_add_crq_rb(struct cfq_rq *crq)
588 struct cfq_queue *cfqq = crq->cfq_queue;
589 struct cfq_data *cfqd = cfqq->cfqd;
590 struct request *rq = crq->request;
591 struct cfq_rq *__alias;
593 crq->rb_key = rq_rb_key(rq);
594 cfqq->queued[cfq_crq_is_sync(crq)]++;
597 * looks a little odd, but the first insert might return an alias.
598 * if that happens, put the alias on the dispatch list
600 while ((__alias = __cfq_add_crq_rb(crq)) != NULL)
601 cfq_dispatch_insert(cfqd->queue, __alias);
603 rb_insert_color(&crq->rb_node, &cfqq->sort_list);
605 if (!cfq_cfqq_on_rr(cfqq))
606 cfq_add_cfqq_rr(cfqd, cfqq);
609 * check if this request is a better next-serve candidate
611 cfqq->next_crq = cfq_choose_req(cfqd, cfqq->next_crq, crq);
615 cfq_reposition_crq_rb(struct cfq_queue *cfqq, struct cfq_rq *crq)
617 rb_erase(&crq->rb_node, &cfqq->sort_list);
618 cfqq->queued[cfq_crq_is_sync(crq)]--;
623 static struct request *cfq_find_rq_rb(struct cfq_data *cfqd, sector_t sector)
626 struct cfq_queue *cfqq = cfq_find_cfq_hash(cfqd, current->pid, CFQ_KEY_ANY);
632 n = cfqq->sort_list.rb_node;
634 struct cfq_rq *crq = rb_entry_crq(n);
636 if (sector < crq->rb_key)
638 else if (sector > crq->rb_key)
648 static void cfq_activate_request(request_queue_t *q, struct request *rq)
650 struct cfq_data *cfqd = q->elevator->elevator_data;
652 cfqd->rq_in_driver++;
655 static void cfq_deactivate_request(request_queue_t *q, struct request *rq)
657 struct cfq_data *cfqd = q->elevator->elevator_data;
659 WARN_ON(!cfqd->rq_in_driver);
660 cfqd->rq_in_driver--;
663 static void cfq_remove_request(struct request *rq)
665 struct cfq_rq *crq = RQ_DATA(rq);
667 list_del_init(&rq->queuelist);
669 cfq_del_crq_hash(crq);
673 cfq_merge(request_queue_t *q, struct request **req, struct bio *bio)
675 struct cfq_data *cfqd = q->elevator->elevator_data;
676 struct request *__rq;
679 __rq = cfq_find_rq_hash(cfqd, bio->bi_sector);
680 if (__rq && elv_rq_merge_ok(__rq, bio)) {
681 ret = ELEVATOR_BACK_MERGE;
685 __rq = cfq_find_rq_rb(cfqd, bio->bi_sector + bio_sectors(bio));
686 if (__rq && elv_rq_merge_ok(__rq, bio)) {
687 ret = ELEVATOR_FRONT_MERGE;
691 return ELEVATOR_NO_MERGE;
697 static void cfq_merged_request(request_queue_t *q, struct request *req)
699 struct cfq_data *cfqd = q->elevator->elevator_data;
700 struct cfq_rq *crq = RQ_DATA(req);
702 cfq_del_crq_hash(crq);
703 cfq_add_crq_hash(cfqd, crq);
705 if (rq_rb_key(req) != crq->rb_key) {
706 struct cfq_queue *cfqq = crq->cfq_queue;
708 cfq_update_next_crq(crq);
709 cfq_reposition_crq_rb(cfqq, crq);
714 cfq_merged_requests(request_queue_t *q, struct request *rq,
715 struct request *next)
717 cfq_merged_request(q, rq);
720 * reposition in fifo if next is older than rq
722 if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
723 time_before(next->start_time, rq->start_time))
724 list_move(&rq->queuelist, &next->queuelist);
726 cfq_remove_request(next);
730 __cfq_set_active_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
734 * stop potential idle class queues waiting service
736 del_timer(&cfqd->idle_class_timer);
738 cfqq->slice_start = jiffies;
740 cfqq->slice_left = 0;
741 cfq_clear_cfqq_must_alloc_slice(cfqq);
742 cfq_clear_cfqq_fifo_expire(cfqq);
745 cfqd->active_queue = cfqq;
749 * current cfqq expired its slice (or was too idle), select new one
752 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
755 unsigned long now = jiffies;
757 if (cfq_cfqq_wait_request(cfqq))
758 del_timer(&cfqd->idle_slice_timer);
760 if (!preempted && !cfq_cfqq_dispatched(cfqq)) {
761 cfqq->service_last = now;
762 cfq_schedule_dispatch(cfqd);
765 cfq_clear_cfqq_must_dispatch(cfqq);
766 cfq_clear_cfqq_wait_request(cfqq);
769 * store what was left of this slice, if the queue idled out
772 if (time_after(cfqq->slice_end, now))
773 cfqq->slice_left = cfqq->slice_end - now;
775 cfqq->slice_left = 0;
777 if (cfq_cfqq_on_rr(cfqq))
778 cfq_resort_rr_list(cfqq, preempted);
780 if (cfqq == cfqd->active_queue)
781 cfqd->active_queue = NULL;
783 if (cfqd->active_cic) {
784 put_io_context(cfqd->active_cic->ioc);
785 cfqd->active_cic = NULL;
788 cfqd->dispatch_slice = 0;
791 static inline void cfq_slice_expired(struct cfq_data *cfqd, int preempted)
793 struct cfq_queue *cfqq = cfqd->active_queue;
796 __cfq_slice_expired(cfqd, cfqq, preempted);
809 static int cfq_get_next_prio_level(struct cfq_data *cfqd)
818 for (p = cfqd->cur_prio; p <= cfqd->cur_end_prio; p++) {
819 if (!list_empty(&cfqd->rr_list[p])) {
828 if (++cfqd->cur_end_prio == CFQ_PRIO_LISTS) {
829 cfqd->cur_end_prio = 0;
836 if (unlikely(prio == -1))
839 BUG_ON(prio >= CFQ_PRIO_LISTS);
841 list_splice_init(&cfqd->rr_list[prio], &cfqd->cur_rr);
843 cfqd->cur_prio = prio + 1;
844 if (cfqd->cur_prio > cfqd->cur_end_prio) {
845 cfqd->cur_end_prio = cfqd->cur_prio;
848 if (cfqd->cur_end_prio == CFQ_PRIO_LISTS) {
850 cfqd->cur_end_prio = 0;
856 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd)
858 struct cfq_queue *cfqq = NULL;
861 * if current list is non-empty, grab first entry. if it is empty,
862 * get next prio level and grab first entry then if any are spliced
864 if (!list_empty(&cfqd->cur_rr) || cfq_get_next_prio_level(cfqd) != -1)
865 cfqq = list_entry_cfqq(cfqd->cur_rr.next);
868 * if we have idle queues and no rt or be queues had pending
869 * requests, either allow immediate service if the grace period
870 * has passed or arm the idle grace timer
872 if (!cfqq && !list_empty(&cfqd->idle_rr)) {
873 unsigned long end = cfqd->last_end_request + CFQ_IDLE_GRACE;
875 if (time_after_eq(jiffies, end))
876 cfqq = list_entry_cfqq(cfqd->idle_rr.next);
878 mod_timer(&cfqd->idle_class_timer, end);
881 __cfq_set_active_queue(cfqd, cfqq);
885 static int cfq_arm_slice_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq)
890 WARN_ON(!RB_EMPTY(&cfqq->sort_list));
891 WARN_ON(cfqq != cfqd->active_queue);
894 * idle is disabled, either manually or by past process history
896 if (!cfqd->cfq_slice_idle)
898 if (!cfq_cfqq_idle_window(cfqq))
901 * task has exited, don't wait
903 if (cfqd->active_cic && !cfqd->active_cic->ioc->task)
906 cfq_mark_cfqq_must_dispatch(cfqq);
907 cfq_mark_cfqq_wait_request(cfqq);
909 sl = min(cfqq->slice_end - 1, (unsigned long) cfqd->cfq_slice_idle);
910 mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
914 static void cfq_dispatch_insert(request_queue_t *q, struct cfq_rq *crq)
916 struct cfq_data *cfqd = q->elevator->elevator_data;
917 struct cfq_queue *cfqq = crq->cfq_queue;
919 cfqq->next_crq = cfq_find_next_crq(cfqd, cfqq, crq);
920 cfq_remove_request(crq->request);
921 cfqq->on_dispatch[cfq_crq_is_sync(crq)]++;
922 elv_dispatch_sort(q, crq->request);
926 * return expired entry, or NULL to just start from scratch in rbtree
928 static inline struct cfq_rq *cfq_check_fifo(struct cfq_queue *cfqq)
930 struct cfq_data *cfqd = cfqq->cfqd;
934 if (cfq_cfqq_fifo_expire(cfqq))
937 if (!list_empty(&cfqq->fifo)) {
938 int fifo = cfq_cfqq_class_sync(cfqq);
940 crq = RQ_DATA(list_entry_fifo(cfqq->fifo.next));
942 if (time_after(jiffies, rq->start_time + cfqd->cfq_fifo_expire[fifo])) {
943 cfq_mark_cfqq_fifo_expire(cfqq);
952 * Scale schedule slice based on io priority. Use the sync time slice only
953 * if a queue is marked sync and has sync io queued. A sync queue with async
954 * io only, should not get full sync slice length.
957 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
959 const int base_slice = cfqd->cfq_slice[cfq_cfqq_sync(cfqq)];
961 WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
963 return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - cfqq->ioprio));
967 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
969 cfqq->slice_end = cfq_prio_to_slice(cfqd, cfqq) + jiffies;
973 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
975 const int base_rq = cfqd->cfq_slice_async_rq;
977 WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
979 return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
983 * get next queue for service
985 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
987 unsigned long now = jiffies;
988 struct cfq_queue *cfqq;
990 cfqq = cfqd->active_queue;
997 if (!cfq_cfqq_must_dispatch(cfqq) && time_after(now, cfqq->slice_end))
1001 * if queue has requests, dispatch one. if not, check if
1002 * enough slice is left to wait for one
1004 if (!RB_EMPTY(&cfqq->sort_list))
1006 else if (cfq_cfqq_class_sync(cfqq) &&
1007 time_before(now, cfqq->slice_end)) {
1008 if (cfq_arm_slice_timer(cfqd, cfqq))
1013 cfq_slice_expired(cfqd, 0);
1015 cfqq = cfq_set_active_queue(cfqd);
1021 __cfq_dispatch_requests(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1026 BUG_ON(RB_EMPTY(&cfqq->sort_list));
1032 * follow expired path, else get first next available
1034 if ((crq = cfq_check_fifo(cfqq)) == NULL)
1035 crq = cfqq->next_crq;
1038 * finally, insert request into driver dispatch list
1040 cfq_dispatch_insert(cfqd->queue, crq);
1042 cfqd->dispatch_slice++;
1045 if (!cfqd->active_cic) {
1046 atomic_inc(&crq->io_context->ioc->refcount);
1047 cfqd->active_cic = crq->io_context;
1050 if (RB_EMPTY(&cfqq->sort_list))
1053 } while (dispatched < max_dispatch);
1056 * if slice end isn't set yet, set it. if at least one request was
1057 * sync, use the sync time slice value
1059 if (!cfqq->slice_end)
1060 cfq_set_prio_slice(cfqd, cfqq);
1063 * expire an async queue immediately if it has used up its slice. idle
1064 * queue always expire after 1 dispatch round.
1066 if ((!cfq_cfqq_sync(cfqq) &&
1067 cfqd->dispatch_slice >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
1068 cfq_class_idle(cfqq))
1069 cfq_slice_expired(cfqd, 0);
1075 cfq_forced_dispatch_cfqqs(struct list_head *list)
1078 struct cfq_queue *cfqq, *next;
1081 list_for_each_entry_safe(cfqq, next, list, cfq_list) {
1082 while ((crq = cfqq->next_crq)) {
1083 cfq_dispatch_insert(cfqq->cfqd->queue, crq);
1086 BUG_ON(!list_empty(&cfqq->fifo));
1092 cfq_forced_dispatch(struct cfq_data *cfqd)
1094 int i, dispatched = 0;
1096 for (i = 0; i < CFQ_PRIO_LISTS; i++)
1097 dispatched += cfq_forced_dispatch_cfqqs(&cfqd->rr_list[i]);
1099 dispatched += cfq_forced_dispatch_cfqqs(&cfqd->busy_rr);
1100 dispatched += cfq_forced_dispatch_cfqqs(&cfqd->cur_rr);
1101 dispatched += cfq_forced_dispatch_cfqqs(&cfqd->idle_rr);
1103 cfq_slice_expired(cfqd, 0);
1105 BUG_ON(cfqd->busy_queues);
1111 cfq_dispatch_requests(request_queue_t *q, int force)
1113 struct cfq_data *cfqd = q->elevator->elevator_data;
1114 struct cfq_queue *cfqq;
1116 if (!cfqd->busy_queues)
1119 if (unlikely(force))
1120 return cfq_forced_dispatch(cfqd);
1122 cfqq = cfq_select_queue(cfqd);
1127 * if idle window is disabled, allow queue buildup
1129 if (!cfq_cfqq_idle_window(cfqq) &&
1130 cfqd->rq_in_driver >= cfqd->cfq_max_depth)
1133 cfq_clear_cfqq_must_dispatch(cfqq);
1134 cfq_clear_cfqq_wait_request(cfqq);
1135 del_timer(&cfqd->idle_slice_timer);
1137 max_dispatch = cfqd->cfq_quantum;
1138 if (cfq_class_idle(cfqq))
1141 return __cfq_dispatch_requests(cfqd, cfqq, max_dispatch);
1148 * task holds one reference to the queue, dropped when task exits. each crq
1149 * in-flight on this queue also holds a reference, dropped when crq is freed.
1151 * queue lock must be held here.
1153 static void cfq_put_queue(struct cfq_queue *cfqq)
1155 struct cfq_data *cfqd = cfqq->cfqd;
1157 BUG_ON(atomic_read(&cfqq->ref) <= 0);
1159 if (!atomic_dec_and_test(&cfqq->ref))
1162 BUG_ON(rb_first(&cfqq->sort_list));
1163 BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
1164 BUG_ON(cfq_cfqq_on_rr(cfqq));
1166 if (unlikely(cfqd->active_queue == cfqq))
1167 __cfq_slice_expired(cfqd, cfqq, 0);
1170 * it's on the empty list and still hashed
1172 list_del(&cfqq->cfq_list);
1173 hlist_del(&cfqq->cfq_hash);
1174 kmem_cache_free(cfq_pool, cfqq);
1177 static inline struct cfq_queue *
1178 __cfq_find_cfq_hash(struct cfq_data *cfqd, unsigned int key, unsigned int prio,
1181 struct hlist_head *hash_list = &cfqd->cfq_hash[hashval];
1182 struct hlist_node *entry, *next;
1184 hlist_for_each_safe(entry, next, hash_list) {
1185 struct cfq_queue *__cfqq = list_entry_qhash(entry);
1186 const unsigned short __p = IOPRIO_PRIO_VALUE(__cfqq->org_ioprio_class, __cfqq->org_ioprio);
1188 if (__cfqq->key == key && (__p == prio || prio == CFQ_KEY_ANY))
1195 static struct cfq_queue *
1196 cfq_find_cfq_hash(struct cfq_data *cfqd, unsigned int key, unsigned short prio)
1198 return __cfq_find_cfq_hash(cfqd, key, prio, hash_long(key, CFQ_QHASH_SHIFT));
1201 static void cfq_free_io_context(struct cfq_io_context *cic)
1203 struct cfq_io_context *__cic;
1204 struct list_head *entry, *next;
1207 list_for_each_safe(entry, next, &cic->list) {
1208 __cic = list_entry(entry, struct cfq_io_context, list);
1209 kmem_cache_free(cfq_ioc_pool, __cic);
1213 kmem_cache_free(cfq_ioc_pool, cic);
1214 if (atomic_sub_and_test(freed, &ioc_count) && ioc_gone)
1218 static void cfq_trim(struct io_context *ioc)
1220 ioc->set_ioprio = NULL;
1222 cfq_free_io_context(ioc->cic);
1226 * Called with interrupts disabled
1228 static void cfq_exit_single_io_context(struct cfq_io_context *cic)
1230 struct cfq_data *cfqd = cic->key;
1238 WARN_ON(!irqs_disabled());
1240 spin_lock(q->queue_lock);
1242 if (cic->cfqq[ASYNC]) {
1243 if (unlikely(cic->cfqq[ASYNC] == cfqd->active_queue))
1244 __cfq_slice_expired(cfqd, cic->cfqq[ASYNC], 0);
1245 cfq_put_queue(cic->cfqq[ASYNC]);
1246 cic->cfqq[ASYNC] = NULL;
1249 if (cic->cfqq[SYNC]) {
1250 if (unlikely(cic->cfqq[SYNC] == cfqd->active_queue))
1251 __cfq_slice_expired(cfqd, cic->cfqq[SYNC], 0);
1252 cfq_put_queue(cic->cfqq[SYNC]);
1253 cic->cfqq[SYNC] = NULL;
1257 list_del_init(&cic->queue_list);
1258 spin_unlock(q->queue_lock);
1262 * Another task may update the task cic list, if it is doing a queue lookup
1263 * on its behalf. cfq_cic_lock excludes such concurrent updates
1265 static void cfq_exit_io_context(struct cfq_io_context *cic)
1267 struct cfq_io_context *__cic;
1268 struct list_head *entry;
1269 unsigned long flags;
1271 local_irq_save(flags);
1274 * put the reference this task is holding to the various queues
1276 read_lock(&cfq_exit_lock);
1277 list_for_each(entry, &cic->list) {
1278 __cic = list_entry(entry, struct cfq_io_context, list);
1279 cfq_exit_single_io_context(__cic);
1282 cfq_exit_single_io_context(cic);
1283 read_unlock(&cfq_exit_lock);
1284 local_irq_restore(flags);
1287 static struct cfq_io_context *
1288 cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
1290 struct cfq_io_context *cic = kmem_cache_alloc(cfq_ioc_pool, gfp_mask);
1293 INIT_LIST_HEAD(&cic->list);
1294 cic->cfqq[ASYNC] = NULL;
1295 cic->cfqq[SYNC] = NULL;
1297 cic->last_end_request = jiffies;
1298 cic->ttime_total = 0;
1299 cic->ttime_samples = 0;
1300 cic->ttime_mean = 0;
1301 cic->dtor = cfq_free_io_context;
1302 cic->exit = cfq_exit_io_context;
1303 INIT_LIST_HEAD(&cic->queue_list);
1304 atomic_inc(&ioc_count);
1310 static void cfq_init_prio_data(struct cfq_queue *cfqq)
1312 struct task_struct *tsk = current;
1315 if (!cfq_cfqq_prio_changed(cfqq))
1318 ioprio_class = IOPRIO_PRIO_CLASS(tsk->ioprio);
1319 switch (ioprio_class) {
1321 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
1322 case IOPRIO_CLASS_NONE:
1324 * no prio set, place us in the middle of the BE classes
1326 cfqq->ioprio = task_nice_ioprio(tsk);
1327 cfqq->ioprio_class = IOPRIO_CLASS_BE;
1329 case IOPRIO_CLASS_RT:
1330 cfqq->ioprio = task_ioprio(tsk);
1331 cfqq->ioprio_class = IOPRIO_CLASS_RT;
1333 case IOPRIO_CLASS_BE:
1334 cfqq->ioprio = task_ioprio(tsk);
1335 cfqq->ioprio_class = IOPRIO_CLASS_BE;
1337 case IOPRIO_CLASS_IDLE:
1338 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
1340 cfq_clear_cfqq_idle_window(cfqq);
1345 * keep track of original prio settings in case we have to temporarily
1346 * elevate the priority of this queue
1348 cfqq->org_ioprio = cfqq->ioprio;
1349 cfqq->org_ioprio_class = cfqq->ioprio_class;
1351 if (cfq_cfqq_on_rr(cfqq))
1352 cfq_resort_rr_list(cfqq, 0);
1354 cfq_clear_cfqq_prio_changed(cfqq);
1357 static inline void changed_ioprio(struct cfq_io_context *cic)
1359 struct cfq_data *cfqd = cic->key;
1360 struct cfq_queue *cfqq;
1362 spin_lock(cfqd->queue->queue_lock);
1363 cfqq = cic->cfqq[ASYNC];
1365 struct cfq_queue *new_cfqq;
1366 new_cfqq = cfq_get_queue(cfqd, CFQ_KEY_ASYNC,
1367 cic->ioc->task, GFP_ATOMIC);
1369 cic->cfqq[ASYNC] = new_cfqq;
1370 cfq_put_queue(cfqq);
1373 cfqq = cic->cfqq[SYNC];
1375 cfq_mark_cfqq_prio_changed(cfqq);
1376 cfq_init_prio_data(cfqq);
1378 spin_unlock(cfqd->queue->queue_lock);
1383 * callback from sys_ioprio_set, irqs are disabled
1385 static int cfq_ioc_set_ioprio(struct io_context *ioc, unsigned int ioprio)
1387 struct cfq_io_context *cic;
1389 write_lock(&cfq_exit_lock);
1393 changed_ioprio(cic);
1395 list_for_each_entry(cic, &cic->list, list)
1396 changed_ioprio(cic);
1398 write_unlock(&cfq_exit_lock);
1403 static struct cfq_queue *
1404 cfq_get_queue(struct cfq_data *cfqd, unsigned int key, struct task_struct *tsk,
1407 const int hashval = hash_long(key, CFQ_QHASH_SHIFT);
1408 struct cfq_queue *cfqq, *new_cfqq = NULL;
1409 unsigned short ioprio;
1412 ioprio = tsk->ioprio;
1413 cfqq = __cfq_find_cfq_hash(cfqd, key, ioprio, hashval);
1419 } else if (gfp_mask & __GFP_WAIT) {
1420 spin_unlock_irq(cfqd->queue->queue_lock);
1421 new_cfqq = kmem_cache_alloc(cfq_pool, gfp_mask);
1422 spin_lock_irq(cfqd->queue->queue_lock);
1425 cfqq = kmem_cache_alloc(cfq_pool, gfp_mask);
1430 memset(cfqq, 0, sizeof(*cfqq));
1432 INIT_HLIST_NODE(&cfqq->cfq_hash);
1433 INIT_LIST_HEAD(&cfqq->cfq_list);
1434 RB_CLEAR_ROOT(&cfqq->sort_list);
1435 INIT_LIST_HEAD(&cfqq->fifo);
1438 hlist_add_head(&cfqq->cfq_hash, &cfqd->cfq_hash[hashval]);
1439 atomic_set(&cfqq->ref, 0);
1441 cfqq->service_last = 0;
1443 * set ->slice_left to allow preemption for a new process
1445 cfqq->slice_left = 2 * cfqd->cfq_slice_idle;
1446 cfq_mark_cfqq_idle_window(cfqq);
1447 cfq_mark_cfqq_prio_changed(cfqq);
1448 cfq_init_prio_data(cfqq);
1452 kmem_cache_free(cfq_pool, new_cfqq);
1454 atomic_inc(&cfqq->ref);
1456 WARN_ON((gfp_mask & __GFP_WAIT) && !cfqq);
1461 * Setup general io context and cfq io context. There can be several cfq
1462 * io contexts per general io context, if this process is doing io to more
1463 * than one device managed by cfq. Note that caller is holding a reference to
1464 * cfqq, so we don't need to worry about it disappearing
1466 static struct cfq_io_context *
1467 cfq_get_io_context(struct cfq_data *cfqd, pid_t pid, gfp_t gfp_mask)
1469 struct io_context *ioc = NULL;
1470 struct cfq_io_context *cic;
1472 might_sleep_if(gfp_mask & __GFP_WAIT);
1474 ioc = get_io_context(gfp_mask);
1479 if ((cic = ioc->cic) == NULL) {
1480 cic = cfq_alloc_io_context(cfqd, gfp_mask);
1486 * manually increment generic io_context usage count, it
1487 * cannot go away since we are already holding one ref to it
1491 read_lock(&cfq_exit_lock);
1492 ioc->set_ioprio = cfq_ioc_set_ioprio;
1494 list_add(&cic->queue_list, &cfqd->cic_list);
1495 read_unlock(&cfq_exit_lock);
1497 struct cfq_io_context *__cic;
1500 * the first cic on the list is actually the head itself
1502 if (cic->key == cfqd)
1505 if (unlikely(!cic->key)) {
1506 read_lock(&cfq_exit_lock);
1507 if (list_empty(&cic->list))
1510 ioc->cic = list_entry(cic->list.next,
1511 struct cfq_io_context,
1513 read_unlock(&cfq_exit_lock);
1514 kmem_cache_free(cfq_ioc_pool, cic);
1515 atomic_dec(&ioc_count);
1520 * cic exists, check if we already are there. linear search
1521 * should be ok here, the list will usually not be more than
1522 * 1 or a few entries long
1524 list_for_each_entry(__cic, &cic->list, list) {
1526 * this process is already holding a reference to
1527 * this queue, so no need to get one more
1529 if (__cic->key == cfqd) {
1533 if (unlikely(!__cic->key)) {
1534 read_lock(&cfq_exit_lock);
1535 list_del(&__cic->list);
1536 read_unlock(&cfq_exit_lock);
1537 kmem_cache_free(cfq_ioc_pool, __cic);
1538 atomic_dec(&ioc_count);
1544 * nope, process doesn't have a cic assoicated with this
1545 * cfqq yet. get a new one and add to list
1547 __cic = cfq_alloc_io_context(cfqd, gfp_mask);
1553 read_lock(&cfq_exit_lock);
1554 list_add(&__cic->list, &cic->list);
1555 list_add(&__cic->queue_list, &cfqd->cic_list);
1556 read_unlock(&cfq_exit_lock);
1563 put_io_context(ioc);
1568 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
1570 unsigned long elapsed, ttime;
1573 * if this context already has stuff queued, thinktime is from
1574 * last queue not last end
1577 if (time_after(cic->last_end_request, cic->last_queue))
1578 elapsed = jiffies - cic->last_end_request;
1580 elapsed = jiffies - cic->last_queue;
1582 elapsed = jiffies - cic->last_end_request;
1585 ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
1587 cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
1588 cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
1589 cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
1592 #define sample_valid(samples) ((samples) > 80)
1595 * Disable idle window if the process thinks too long or seeks so much that
1599 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1600 struct cfq_io_context *cic)
1602 int enable_idle = cfq_cfqq_idle_window(cfqq);
1604 if (!cic->ioc->task || !cfqd->cfq_slice_idle)
1606 else if (sample_valid(cic->ttime_samples)) {
1607 if (cic->ttime_mean > cfqd->cfq_slice_idle)
1614 cfq_mark_cfqq_idle_window(cfqq);
1616 cfq_clear_cfqq_idle_window(cfqq);
1621 * Check if new_cfqq should preempt the currently active queue. Return 0 for
1622 * no or if we aren't sure, a 1 will cause a preempt.
1625 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
1628 struct cfq_queue *cfqq = cfqd->active_queue;
1630 if (cfq_class_idle(new_cfqq))
1636 if (cfq_class_idle(cfqq))
1638 if (!cfq_cfqq_wait_request(new_cfqq))
1641 * if it doesn't have slice left, forget it
1643 if (new_cfqq->slice_left < cfqd->cfq_slice_idle)
1645 if (cfq_crq_is_sync(crq) && !cfq_cfqq_sync(cfqq))
1652 * cfqq preempts the active queue. if we allowed preempt with no slice left,
1653 * let it have half of its nominal slice.
1655 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1657 struct cfq_queue *__cfqq, *next;
1659 list_for_each_entry_safe(__cfqq, next, &cfqd->cur_rr, cfq_list)
1660 cfq_resort_rr_list(__cfqq, 1);
1662 if (!cfqq->slice_left)
1663 cfqq->slice_left = cfq_prio_to_slice(cfqd, cfqq) / 2;
1665 cfqq->slice_end = cfqq->slice_left + jiffies;
1666 __cfq_slice_expired(cfqd, cfqq, 1);
1667 __cfq_set_active_queue(cfqd, cfqq);
1671 * should really be a ll_rw_blk.c helper
1673 static void cfq_start_queueing(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1675 request_queue_t *q = cfqd->queue;
1677 if (!blk_queue_plugged(q))
1680 __generic_unplug_device(q);
1684 * Called when a new fs request (crq) is added (to cfqq). Check if there's
1685 * something we should do about it
1688 cfq_crq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1691 struct cfq_io_context *cic;
1693 cfqq->next_crq = cfq_choose_req(cfqd, cfqq->next_crq, crq);
1696 * we never wait for an async request and we don't allow preemption
1697 * of an async request. so just return early
1699 if (!cfq_crq_is_sync(crq))
1702 cic = crq->io_context;
1704 cfq_update_io_thinktime(cfqd, cic);
1705 cfq_update_idle_window(cfqd, cfqq, cic);
1707 cic->last_queue = jiffies;
1709 if (cfqq == cfqd->active_queue) {
1711 * if we are waiting for a request for this queue, let it rip
1712 * immediately and flag that we must not expire this queue
1715 if (cfq_cfqq_wait_request(cfqq)) {
1716 cfq_mark_cfqq_must_dispatch(cfqq);
1717 del_timer(&cfqd->idle_slice_timer);
1718 cfq_start_queueing(cfqd, cfqq);
1720 } else if (cfq_should_preempt(cfqd, cfqq, crq)) {
1722 * not the active queue - expire current slice if it is
1723 * idle and has expired it's mean thinktime or this new queue
1724 * has some old slice time left and is of higher priority
1726 cfq_preempt_queue(cfqd, cfqq);
1727 cfq_mark_cfqq_must_dispatch(cfqq);
1728 cfq_start_queueing(cfqd, cfqq);
1732 static void cfq_insert_request(request_queue_t *q, struct request *rq)
1734 struct cfq_data *cfqd = q->elevator->elevator_data;
1735 struct cfq_rq *crq = RQ_DATA(rq);
1736 struct cfq_queue *cfqq = crq->cfq_queue;
1738 cfq_init_prio_data(cfqq);
1740 cfq_add_crq_rb(crq);
1742 list_add_tail(&rq->queuelist, &cfqq->fifo);
1744 if (rq_mergeable(rq))
1745 cfq_add_crq_hash(cfqd, crq);
1747 cfq_crq_enqueued(cfqd, cfqq, crq);
1750 static void cfq_completed_request(request_queue_t *q, struct request *rq)
1752 struct cfq_rq *crq = RQ_DATA(rq);
1753 struct cfq_queue *cfqq = crq->cfq_queue;
1754 struct cfq_data *cfqd = cfqq->cfqd;
1755 const int sync = cfq_crq_is_sync(crq);
1760 WARN_ON(!cfqd->rq_in_driver);
1761 WARN_ON(!cfqq->on_dispatch[sync]);
1762 cfqd->rq_in_driver--;
1763 cfqq->on_dispatch[sync]--;
1765 if (!cfq_class_idle(cfqq))
1766 cfqd->last_end_request = now;
1768 if (!cfq_cfqq_dispatched(cfqq)) {
1769 if (cfq_cfqq_on_rr(cfqq)) {
1770 cfqq->service_last = now;
1771 cfq_resort_rr_list(cfqq, 0);
1773 cfq_schedule_dispatch(cfqd);
1776 if (cfq_crq_is_sync(crq))
1777 crq->io_context->last_end_request = now;
1780 static struct request *
1781 cfq_former_request(request_queue_t *q, struct request *rq)
1783 struct cfq_rq *crq = RQ_DATA(rq);
1784 struct rb_node *rbprev = rb_prev(&crq->rb_node);
1787 return rb_entry_crq(rbprev)->request;
1792 static struct request *
1793 cfq_latter_request(request_queue_t *q, struct request *rq)
1795 struct cfq_rq *crq = RQ_DATA(rq);
1796 struct rb_node *rbnext = rb_next(&crq->rb_node);
1799 return rb_entry_crq(rbnext)->request;
1805 * we temporarily boost lower priority queues if they are holding fs exclusive
1806 * resources. they are boosted to normal prio (CLASS_BE/4)
1808 static void cfq_prio_boost(struct cfq_queue *cfqq)
1810 const int ioprio_class = cfqq->ioprio_class;
1811 const int ioprio = cfqq->ioprio;
1813 if (has_fs_excl()) {
1815 * boost idle prio on transactions that would lock out other
1816 * users of the filesystem
1818 if (cfq_class_idle(cfqq))
1819 cfqq->ioprio_class = IOPRIO_CLASS_BE;
1820 if (cfqq->ioprio > IOPRIO_NORM)
1821 cfqq->ioprio = IOPRIO_NORM;
1824 * check if we need to unboost the queue
1826 if (cfqq->ioprio_class != cfqq->org_ioprio_class)
1827 cfqq->ioprio_class = cfqq->org_ioprio_class;
1828 if (cfqq->ioprio != cfqq->org_ioprio)
1829 cfqq->ioprio = cfqq->org_ioprio;
1833 * refile between round-robin lists if we moved the priority class
1835 if ((ioprio_class != cfqq->ioprio_class || ioprio != cfqq->ioprio) &&
1836 cfq_cfqq_on_rr(cfqq))
1837 cfq_resort_rr_list(cfqq, 0);
1840 static inline pid_t cfq_queue_pid(struct task_struct *task, int rw)
1842 if (rw == READ || process_sync(task))
1845 return CFQ_KEY_ASYNC;
1849 __cfq_may_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1850 struct task_struct *task, int rw)
1853 if ((cfq_cfqq_wait_request(cfqq) || cfq_cfqq_must_alloc(cfqq)) &&
1854 !cfq_cfqq_must_alloc_slice(cfqq)) {
1855 cfq_mark_cfqq_must_alloc_slice(cfqq);
1856 return ELV_MQUEUE_MUST;
1859 return ELV_MQUEUE_MAY;
1861 if (!cfqq || task->flags & PF_MEMALLOC)
1862 return ELV_MQUEUE_MAY;
1863 if (!cfqq->allocated[rw] || cfq_cfqq_must_alloc(cfqq)) {
1864 if (cfq_cfqq_wait_request(cfqq))
1865 return ELV_MQUEUE_MUST;
1868 * only allow 1 ELV_MQUEUE_MUST per slice, otherwise we
1869 * can quickly flood the queue with writes from a single task
1871 if (rw == READ || !cfq_cfqq_must_alloc_slice(cfqq)) {
1872 cfq_mark_cfqq_must_alloc_slice(cfqq);
1873 return ELV_MQUEUE_MUST;
1876 return ELV_MQUEUE_MAY;
1878 if (cfq_class_idle(cfqq))
1879 return ELV_MQUEUE_NO;
1880 if (cfqq->allocated[rw] >= cfqd->max_queued) {
1881 struct io_context *ioc = get_io_context(GFP_ATOMIC);
1882 int ret = ELV_MQUEUE_NO;
1884 if (ioc && ioc->nr_batch_requests)
1885 ret = ELV_MQUEUE_MAY;
1887 put_io_context(ioc);
1891 return ELV_MQUEUE_MAY;
1895 static int cfq_may_queue(request_queue_t *q, int rw, struct bio *bio)
1897 struct cfq_data *cfqd = q->elevator->elevator_data;
1898 struct task_struct *tsk = current;
1899 struct cfq_queue *cfqq;
1902 * don't force setup of a queue from here, as a call to may_queue
1903 * does not necessarily imply that a request actually will be queued.
1904 * so just lookup a possibly existing queue, or return 'may queue'
1907 cfqq = cfq_find_cfq_hash(cfqd, cfq_queue_pid(tsk, rw), tsk->ioprio);
1909 cfq_init_prio_data(cfqq);
1910 cfq_prio_boost(cfqq);
1912 return __cfq_may_queue(cfqd, cfqq, tsk, rw);
1915 return ELV_MQUEUE_MAY;
1918 static void cfq_check_waiters(request_queue_t *q, struct cfq_queue *cfqq)
1920 struct cfq_data *cfqd = q->elevator->elevator_data;
1921 struct request_list *rl = &q->rq;
1923 if (cfqq->allocated[READ] <= cfqd->max_queued || cfqd->rq_starved) {
1925 if (waitqueue_active(&rl->wait[READ]))
1926 wake_up(&rl->wait[READ]);
1929 if (cfqq->allocated[WRITE] <= cfqd->max_queued || cfqd->rq_starved) {
1931 if (waitqueue_active(&rl->wait[WRITE]))
1932 wake_up(&rl->wait[WRITE]);
1937 * queue lock held here
1939 static void cfq_put_request(request_queue_t *q, struct request *rq)
1941 struct cfq_data *cfqd = q->elevator->elevator_data;
1942 struct cfq_rq *crq = RQ_DATA(rq);
1945 struct cfq_queue *cfqq = crq->cfq_queue;
1946 const int rw = rq_data_dir(rq);
1948 BUG_ON(!cfqq->allocated[rw]);
1949 cfqq->allocated[rw]--;
1951 put_io_context(crq->io_context->ioc);
1953 mempool_free(crq, cfqd->crq_pool);
1954 rq->elevator_private = NULL;
1956 cfq_check_waiters(q, cfqq);
1957 cfq_put_queue(cfqq);
1962 * Allocate cfq data structures associated with this request.
1965 cfq_set_request(request_queue_t *q, struct request *rq, struct bio *bio,
1968 struct cfq_data *cfqd = q->elevator->elevator_data;
1969 struct task_struct *tsk = current;
1970 struct cfq_io_context *cic;
1971 const int rw = rq_data_dir(rq);
1972 pid_t key = cfq_queue_pid(tsk, rw);
1973 struct cfq_queue *cfqq;
1975 unsigned long flags;
1976 int is_sync = key != CFQ_KEY_ASYNC;
1978 might_sleep_if(gfp_mask & __GFP_WAIT);
1980 cic = cfq_get_io_context(cfqd, key, gfp_mask);
1982 spin_lock_irqsave(q->queue_lock, flags);
1987 if (!cic->cfqq[is_sync]) {
1988 cfqq = cfq_get_queue(cfqd, key, tsk, gfp_mask);
1992 cic->cfqq[is_sync] = cfqq;
1994 cfqq = cic->cfqq[is_sync];
1996 cfqq->allocated[rw]++;
1997 cfq_clear_cfqq_must_alloc(cfqq);
1998 cfqd->rq_starved = 0;
1999 atomic_inc(&cfqq->ref);
2000 spin_unlock_irqrestore(q->queue_lock, flags);
2002 crq = mempool_alloc(cfqd->crq_pool, gfp_mask);
2004 RB_CLEAR(&crq->rb_node);
2007 INIT_HLIST_NODE(&crq->hash);
2008 crq->cfq_queue = cfqq;
2009 crq->io_context = cic;
2012 cfq_mark_crq_is_sync(crq);
2014 cfq_clear_crq_is_sync(crq);
2016 rq->elevator_private = crq;
2020 spin_lock_irqsave(q->queue_lock, flags);
2021 cfqq->allocated[rw]--;
2022 if (!(cfqq->allocated[0] + cfqq->allocated[1]))
2023 cfq_mark_cfqq_must_alloc(cfqq);
2024 cfq_put_queue(cfqq);
2027 put_io_context(cic->ioc);
2029 * mark us rq allocation starved. we need to kickstart the process
2030 * ourselves if there are no pending requests that can do it for us.
2031 * that would be an extremely rare OOM situation
2033 cfqd->rq_starved = 1;
2034 cfq_schedule_dispatch(cfqd);
2035 spin_unlock_irqrestore(q->queue_lock, flags);
2039 static void cfq_kick_queue(void *data)
2041 request_queue_t *q = data;
2042 struct cfq_data *cfqd = q->elevator->elevator_data;
2043 unsigned long flags;
2045 spin_lock_irqsave(q->queue_lock, flags);
2047 if (cfqd->rq_starved) {
2048 struct request_list *rl = &q->rq;
2051 * we aren't guaranteed to get a request after this, but we
2052 * have to be opportunistic
2055 if (waitqueue_active(&rl->wait[READ]))
2056 wake_up(&rl->wait[READ]);
2057 if (waitqueue_active(&rl->wait[WRITE]))
2058 wake_up(&rl->wait[WRITE]);
2063 spin_unlock_irqrestore(q->queue_lock, flags);
2067 * Timer running if the active_queue is currently idling inside its time slice
2069 static void cfq_idle_slice_timer(unsigned long data)
2071 struct cfq_data *cfqd = (struct cfq_data *) data;
2072 struct cfq_queue *cfqq;
2073 unsigned long flags;
2075 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2077 if ((cfqq = cfqd->active_queue) != NULL) {
2078 unsigned long now = jiffies;
2083 if (time_after(now, cfqq->slice_end))
2087 * only expire and reinvoke request handler, if there are
2088 * other queues with pending requests
2090 if (!cfqd->busy_queues) {
2091 cfqd->idle_slice_timer.expires = min(now + cfqd->cfq_slice_idle, cfqq->slice_end);
2092 add_timer(&cfqd->idle_slice_timer);
2097 * not expired and it has a request pending, let it dispatch
2099 if (!RB_EMPTY(&cfqq->sort_list)) {
2100 cfq_mark_cfqq_must_dispatch(cfqq);
2105 cfq_slice_expired(cfqd, 0);
2107 cfq_schedule_dispatch(cfqd);
2109 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2113 * Timer running if an idle class queue is waiting for service
2115 static void cfq_idle_class_timer(unsigned long data)
2117 struct cfq_data *cfqd = (struct cfq_data *) data;
2118 unsigned long flags, end;
2120 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2123 * race with a non-idle queue, reset timer
2125 end = cfqd->last_end_request + CFQ_IDLE_GRACE;
2126 if (!time_after_eq(jiffies, end)) {
2127 cfqd->idle_class_timer.expires = end;
2128 add_timer(&cfqd->idle_class_timer);
2130 cfq_schedule_dispatch(cfqd);
2132 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2135 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
2137 del_timer_sync(&cfqd->idle_slice_timer);
2138 del_timer_sync(&cfqd->idle_class_timer);
2139 blk_sync_queue(cfqd->queue);
2142 static void cfq_exit_queue(elevator_t *e)
2144 struct cfq_data *cfqd = e->elevator_data;
2145 request_queue_t *q = cfqd->queue;
2147 cfq_shutdown_timer_wq(cfqd);
2148 write_lock(&cfq_exit_lock);
2149 spin_lock_irq(q->queue_lock);
2150 if (cfqd->active_queue)
2151 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
2152 while(!list_empty(&cfqd->cic_list)) {
2153 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
2154 struct cfq_io_context,
2156 if (cic->cfqq[ASYNC]) {
2157 cfq_put_queue(cic->cfqq[ASYNC]);
2158 cic->cfqq[ASYNC] = NULL;
2160 if (cic->cfqq[SYNC]) {
2161 cfq_put_queue(cic->cfqq[SYNC]);
2162 cic->cfqq[SYNC] = NULL;
2165 list_del_init(&cic->queue_list);
2167 spin_unlock_irq(q->queue_lock);
2168 write_unlock(&cfq_exit_lock);
2170 cfq_shutdown_timer_wq(cfqd);
2172 mempool_destroy(cfqd->crq_pool);
2173 kfree(cfqd->crq_hash);
2174 kfree(cfqd->cfq_hash);
2178 static int cfq_init_queue(request_queue_t *q, elevator_t *e)
2180 struct cfq_data *cfqd;
2183 cfqd = kmalloc(sizeof(*cfqd), GFP_KERNEL);
2187 memset(cfqd, 0, sizeof(*cfqd));
2189 for (i = 0; i < CFQ_PRIO_LISTS; i++)
2190 INIT_LIST_HEAD(&cfqd->rr_list[i]);
2192 INIT_LIST_HEAD(&cfqd->busy_rr);
2193 INIT_LIST_HEAD(&cfqd->cur_rr);
2194 INIT_LIST_HEAD(&cfqd->idle_rr);
2195 INIT_LIST_HEAD(&cfqd->empty_list);
2196 INIT_LIST_HEAD(&cfqd->cic_list);
2198 cfqd->crq_hash = kmalloc(sizeof(struct hlist_head) * CFQ_MHASH_ENTRIES, GFP_KERNEL);
2199 if (!cfqd->crq_hash)
2202 cfqd->cfq_hash = kmalloc(sizeof(struct hlist_head) * CFQ_QHASH_ENTRIES, GFP_KERNEL);
2203 if (!cfqd->cfq_hash)
2206 cfqd->crq_pool = mempool_create(BLKDEV_MIN_RQ, mempool_alloc_slab, mempool_free_slab, crq_pool);
2207 if (!cfqd->crq_pool)
2210 for (i = 0; i < CFQ_MHASH_ENTRIES; i++)
2211 INIT_HLIST_HEAD(&cfqd->crq_hash[i]);
2212 for (i = 0; i < CFQ_QHASH_ENTRIES; i++)
2213 INIT_HLIST_HEAD(&cfqd->cfq_hash[i]);
2215 e->elevator_data = cfqd;
2219 cfqd->max_queued = q->nr_requests / 4;
2220 q->nr_batching = cfq_queued;
2222 init_timer(&cfqd->idle_slice_timer);
2223 cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
2224 cfqd->idle_slice_timer.data = (unsigned long) cfqd;
2226 init_timer(&cfqd->idle_class_timer);
2227 cfqd->idle_class_timer.function = cfq_idle_class_timer;
2228 cfqd->idle_class_timer.data = (unsigned long) cfqd;
2230 INIT_WORK(&cfqd->unplug_work, cfq_kick_queue, q);
2232 cfqd->cfq_queued = cfq_queued;
2233 cfqd->cfq_quantum = cfq_quantum;
2234 cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
2235 cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
2236 cfqd->cfq_back_max = cfq_back_max;
2237 cfqd->cfq_back_penalty = cfq_back_penalty;
2238 cfqd->cfq_slice[0] = cfq_slice_async;
2239 cfqd->cfq_slice[1] = cfq_slice_sync;
2240 cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
2241 cfqd->cfq_slice_idle = cfq_slice_idle;
2242 cfqd->cfq_max_depth = cfq_max_depth;
2246 kfree(cfqd->cfq_hash);
2248 kfree(cfqd->crq_hash);
2254 static void cfq_slab_kill(void)
2257 kmem_cache_destroy(crq_pool);
2259 kmem_cache_destroy(cfq_pool);
2261 kmem_cache_destroy(cfq_ioc_pool);
2264 static int __init cfq_slab_setup(void)
2266 crq_pool = kmem_cache_create("crq_pool", sizeof(struct cfq_rq), 0, 0,
2271 cfq_pool = kmem_cache_create("cfq_pool", sizeof(struct cfq_queue), 0, 0,
2276 cfq_ioc_pool = kmem_cache_create("cfq_ioc_pool",
2277 sizeof(struct cfq_io_context), 0, 0, NULL, NULL);
2288 * sysfs parts below -->
2290 struct cfq_fs_entry {
2291 struct attribute attr;
2292 ssize_t (*show)(struct cfq_data *, char *);
2293 ssize_t (*store)(struct cfq_data *, const char *, size_t);
2297 cfq_var_show(unsigned int var, char *page)
2299 return sprintf(page, "%d\n", var);
2303 cfq_var_store(unsigned int *var, const char *page, size_t count)
2305 char *p = (char *) page;
2307 *var = simple_strtoul(p, &p, 10);
2311 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
2312 static ssize_t __FUNC(struct cfq_data *cfqd, char *page) \
2314 unsigned int __data = __VAR; \
2316 __data = jiffies_to_msecs(__data); \
2317 return cfq_var_show(__data, (page)); \
2319 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
2320 SHOW_FUNCTION(cfq_queued_show, cfqd->cfq_queued, 0);
2321 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
2322 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
2323 SHOW_FUNCTION(cfq_back_max_show, cfqd->cfq_back_max, 0);
2324 SHOW_FUNCTION(cfq_back_penalty_show, cfqd->cfq_back_penalty, 0);
2325 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
2326 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
2327 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
2328 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
2329 SHOW_FUNCTION(cfq_max_depth_show, cfqd->cfq_max_depth, 0);
2330 #undef SHOW_FUNCTION
2332 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
2333 static ssize_t __FUNC(struct cfq_data *cfqd, const char *page, size_t count) \
2335 unsigned int __data; \
2336 int ret = cfq_var_store(&__data, (page), count); \
2337 if (__data < (MIN)) \
2339 else if (__data > (MAX)) \
2342 *(__PTR) = msecs_to_jiffies(__data); \
2344 *(__PTR) = __data; \
2347 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
2348 STORE_FUNCTION(cfq_queued_store, &cfqd->cfq_queued, 1, UINT_MAX, 0);
2349 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1, UINT_MAX, 1);
2350 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1, UINT_MAX, 1);
2351 STORE_FUNCTION(cfq_back_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
2352 STORE_FUNCTION(cfq_back_penalty_store, &cfqd->cfq_back_penalty, 1, UINT_MAX, 0);
2353 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
2354 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
2355 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
2356 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1, UINT_MAX, 0);
2357 STORE_FUNCTION(cfq_max_depth_store, &cfqd->cfq_max_depth, 1, UINT_MAX, 0);
2358 #undef STORE_FUNCTION
2360 static struct cfq_fs_entry cfq_quantum_entry = {
2361 .attr = {.name = "quantum", .mode = S_IRUGO | S_IWUSR },
2362 .show = cfq_quantum_show,
2363 .store = cfq_quantum_store,
2365 static struct cfq_fs_entry cfq_queued_entry = {
2366 .attr = {.name = "queued", .mode = S_IRUGO | S_IWUSR },
2367 .show = cfq_queued_show,
2368 .store = cfq_queued_store,
2370 static struct cfq_fs_entry cfq_fifo_expire_sync_entry = {
2371 .attr = {.name = "fifo_expire_sync", .mode = S_IRUGO | S_IWUSR },
2372 .show = cfq_fifo_expire_sync_show,
2373 .store = cfq_fifo_expire_sync_store,
2375 static struct cfq_fs_entry cfq_fifo_expire_async_entry = {
2376 .attr = {.name = "fifo_expire_async", .mode = S_IRUGO | S_IWUSR },
2377 .show = cfq_fifo_expire_async_show,
2378 .store = cfq_fifo_expire_async_store,
2380 static struct cfq_fs_entry cfq_back_max_entry = {
2381 .attr = {.name = "back_seek_max", .mode = S_IRUGO | S_IWUSR },
2382 .show = cfq_back_max_show,
2383 .store = cfq_back_max_store,
2385 static struct cfq_fs_entry cfq_back_penalty_entry = {
2386 .attr = {.name = "back_seek_penalty", .mode = S_IRUGO | S_IWUSR },
2387 .show = cfq_back_penalty_show,
2388 .store = cfq_back_penalty_store,
2390 static struct cfq_fs_entry cfq_slice_sync_entry = {
2391 .attr = {.name = "slice_sync", .mode = S_IRUGO | S_IWUSR },
2392 .show = cfq_slice_sync_show,
2393 .store = cfq_slice_sync_store,
2395 static struct cfq_fs_entry cfq_slice_async_entry = {
2396 .attr = {.name = "slice_async", .mode = S_IRUGO | S_IWUSR },
2397 .show = cfq_slice_async_show,
2398 .store = cfq_slice_async_store,
2400 static struct cfq_fs_entry cfq_slice_async_rq_entry = {
2401 .attr = {.name = "slice_async_rq", .mode = S_IRUGO | S_IWUSR },
2402 .show = cfq_slice_async_rq_show,
2403 .store = cfq_slice_async_rq_store,
2405 static struct cfq_fs_entry cfq_slice_idle_entry = {
2406 .attr = {.name = "slice_idle", .mode = S_IRUGO | S_IWUSR },
2407 .show = cfq_slice_idle_show,
2408 .store = cfq_slice_idle_store,
2410 static struct cfq_fs_entry cfq_max_depth_entry = {
2411 .attr = {.name = "max_depth", .mode = S_IRUGO | S_IWUSR },
2412 .show = cfq_max_depth_show,
2413 .store = cfq_max_depth_store,
2416 static struct attribute *default_attrs[] = {
2417 &cfq_quantum_entry.attr,
2418 &cfq_queued_entry.attr,
2419 &cfq_fifo_expire_sync_entry.attr,
2420 &cfq_fifo_expire_async_entry.attr,
2421 &cfq_back_max_entry.attr,
2422 &cfq_back_penalty_entry.attr,
2423 &cfq_slice_sync_entry.attr,
2424 &cfq_slice_async_entry.attr,
2425 &cfq_slice_async_rq_entry.attr,
2426 &cfq_slice_idle_entry.attr,
2427 &cfq_max_depth_entry.attr,
2431 #define to_cfq(atr) container_of((atr), struct cfq_fs_entry, attr)
2434 cfq_attr_show(struct kobject *kobj, struct attribute *attr, char *page)
2436 elevator_t *e = container_of(kobj, elevator_t, kobj);
2437 struct cfq_fs_entry *entry = to_cfq(attr);
2442 return entry->show(e->elevator_data, page);
2446 cfq_attr_store(struct kobject *kobj, struct attribute *attr,
2447 const char *page, size_t length)
2449 elevator_t *e = container_of(kobj, elevator_t, kobj);
2450 struct cfq_fs_entry *entry = to_cfq(attr);
2455 return entry->store(e->elevator_data, page, length);
2458 static struct sysfs_ops cfq_sysfs_ops = {
2459 .show = cfq_attr_show,
2460 .store = cfq_attr_store,
2463 static struct kobj_type cfq_ktype = {
2464 .sysfs_ops = &cfq_sysfs_ops,
2465 .default_attrs = default_attrs,
2468 static struct elevator_type iosched_cfq = {
2470 .elevator_merge_fn = cfq_merge,
2471 .elevator_merged_fn = cfq_merged_request,
2472 .elevator_merge_req_fn = cfq_merged_requests,
2473 .elevator_dispatch_fn = cfq_dispatch_requests,
2474 .elevator_add_req_fn = cfq_insert_request,
2475 .elevator_activate_req_fn = cfq_activate_request,
2476 .elevator_deactivate_req_fn = cfq_deactivate_request,
2477 .elevator_queue_empty_fn = cfq_queue_empty,
2478 .elevator_completed_req_fn = cfq_completed_request,
2479 .elevator_former_req_fn = cfq_former_request,
2480 .elevator_latter_req_fn = cfq_latter_request,
2481 .elevator_set_req_fn = cfq_set_request,
2482 .elevator_put_req_fn = cfq_put_request,
2483 .elevator_may_queue_fn = cfq_may_queue,
2484 .elevator_init_fn = cfq_init_queue,
2485 .elevator_exit_fn = cfq_exit_queue,
2488 .elevator_ktype = &cfq_ktype,
2489 .elevator_name = "cfq",
2490 .elevator_owner = THIS_MODULE,
2493 static int __init cfq_init(void)
2498 * could be 0 on HZ < 1000 setups
2500 if (!cfq_slice_async)
2501 cfq_slice_async = 1;
2502 if (!cfq_slice_idle)
2505 if (cfq_slab_setup())
2508 ret = elv_register(&iosched_cfq);
2515 static void __exit cfq_exit(void)
2517 DECLARE_COMPLETION(all_gone);
2518 elv_unregister(&iosched_cfq);
2519 ioc_gone = &all_gone;
2521 if (atomic_read(&ioc_count))
2527 module_init(cfq_init);
2528 module_exit(cfq_exit);
2530 MODULE_AUTHOR("Jens Axboe");
2531 MODULE_LICENSE("GPL");
2532 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");