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;
51 * for the hash of cfqq inside the cfqd
53 #define CFQ_QHASH_SHIFT 6
54 #define CFQ_QHASH_ENTRIES (1 << CFQ_QHASH_SHIFT)
55 #define list_entry_qhash(entry) hlist_entry((entry), struct cfq_queue, cfq_hash)
58 * for the hash of crq inside the cfqq
60 #define CFQ_MHASH_SHIFT 6
61 #define CFQ_MHASH_BLOCK(sec) ((sec) >> 3)
62 #define CFQ_MHASH_ENTRIES (1 << CFQ_MHASH_SHIFT)
63 #define CFQ_MHASH_FN(sec) hash_long(CFQ_MHASH_BLOCK(sec), CFQ_MHASH_SHIFT)
64 #define rq_hash_key(rq) ((rq)->sector + (rq)->nr_sectors)
65 #define list_entry_hash(ptr) hlist_entry((ptr), struct cfq_rq, hash)
67 #define list_entry_cfqq(ptr) list_entry((ptr), struct cfq_queue, cfq_list)
68 #define list_entry_fifo(ptr) list_entry((ptr), struct request, queuelist)
70 #define RQ_DATA(rq) (rq)->elevator_private
76 #define RB_EMPTY(node) ((node)->rb_node == NULL)
77 #define RB_CLEAR_COLOR(node) (node)->rb_color = RB_NONE
78 #define RB_CLEAR(node) do { \
79 (node)->rb_parent = NULL; \
80 RB_CLEAR_COLOR((node)); \
81 (node)->rb_right = NULL; \
82 (node)->rb_left = NULL; \
84 #define RB_CLEAR_ROOT(root) ((root)->rb_node = NULL)
85 #define rb_entry_crq(node) rb_entry((node), struct cfq_rq, rb_node)
86 #define rq_rb_key(rq) (rq)->sector
88 static kmem_cache_t *crq_pool;
89 static kmem_cache_t *cfq_pool;
90 static kmem_cache_t *cfq_ioc_pool;
92 #define CFQ_PRIO_LISTS IOPRIO_BE_NR
93 #define cfq_class_idle(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
94 #define cfq_class_be(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_BE)
95 #define cfq_class_rt(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
100 #define cfq_cfqq_dispatched(cfqq) \
101 ((cfqq)->on_dispatch[ASYNC] + (cfqq)->on_dispatch[SYNC])
103 #define cfq_cfqq_class_sync(cfqq) ((cfqq)->key != CFQ_KEY_ASYNC)
105 #define cfq_cfqq_sync(cfqq) \
106 (cfq_cfqq_class_sync(cfqq) || (cfqq)->on_dispatch[SYNC])
109 * Per block device queue structure
113 request_queue_t *queue;
116 * rr list of queues with requests and the count of them
118 struct list_head rr_list[CFQ_PRIO_LISTS];
119 struct list_head busy_rr;
120 struct list_head cur_rr;
121 struct list_head idle_rr;
122 unsigned int busy_queues;
125 * non-ordered list of empty cfqq's
127 struct list_head empty_list;
132 struct hlist_head *cfq_hash;
135 * global crq hash for all queues
137 struct hlist_head *crq_hash;
139 unsigned int max_queued;
146 * schedule slice state info
149 * idle window management
151 struct timer_list idle_slice_timer;
152 struct work_struct unplug_work;
154 struct cfq_queue *active_queue;
155 struct cfq_io_context *active_cic;
156 int cur_prio, cur_end_prio;
157 unsigned int dispatch_slice;
159 struct timer_list idle_class_timer;
161 sector_t last_sector;
162 unsigned long last_end_request;
164 unsigned int rq_starved;
167 * tunables, see top of file
169 unsigned int cfq_quantum;
170 unsigned int cfq_queued;
171 unsigned int cfq_fifo_expire[2];
172 unsigned int cfq_back_penalty;
173 unsigned int cfq_back_max;
174 unsigned int cfq_slice[2];
175 unsigned int cfq_slice_async_rq;
176 unsigned int cfq_slice_idle;
177 unsigned int cfq_max_depth;
181 * Per process-grouping structure
184 /* reference count */
186 /* parent cfq_data */
187 struct cfq_data *cfqd;
188 /* cfqq lookup hash */
189 struct hlist_node cfq_hash;
192 /* on either rr or empty list of cfqd */
193 struct list_head cfq_list;
194 /* sorted list of pending requests */
195 struct rb_root sort_list;
196 /* if fifo isn't expired, next request to serve */
197 struct cfq_rq *next_crq;
198 /* requests queued in sort_list */
200 /* currently allocated requests */
202 /* fifo list of requests in sort_list */
203 struct list_head fifo;
205 unsigned long slice_start;
206 unsigned long slice_end;
207 unsigned long slice_left;
208 unsigned long service_last;
210 /* number of requests that are on the dispatch list */
213 /* io prio of this group */
214 unsigned short ioprio, org_ioprio;
215 unsigned short ioprio_class, org_ioprio_class;
217 /* various state flags, see below */
222 struct rb_node rb_node;
224 struct request *request;
225 struct hlist_node hash;
227 struct cfq_queue *cfq_queue;
228 struct cfq_io_context *io_context;
230 unsigned int crq_flags;
233 enum cfqq_state_flags {
234 CFQ_CFQQ_FLAG_on_rr = 0,
235 CFQ_CFQQ_FLAG_wait_request,
236 CFQ_CFQQ_FLAG_must_alloc,
237 CFQ_CFQQ_FLAG_must_alloc_slice,
238 CFQ_CFQQ_FLAG_must_dispatch,
239 CFQ_CFQQ_FLAG_fifo_expire,
240 CFQ_CFQQ_FLAG_idle_window,
241 CFQ_CFQQ_FLAG_prio_changed,
244 #define CFQ_CFQQ_FNS(name) \
245 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \
247 cfqq->flags |= (1 << CFQ_CFQQ_FLAG_##name); \
249 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \
251 cfqq->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \
253 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \
255 return (cfqq->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \
259 CFQ_CFQQ_FNS(wait_request);
260 CFQ_CFQQ_FNS(must_alloc);
261 CFQ_CFQQ_FNS(must_alloc_slice);
262 CFQ_CFQQ_FNS(must_dispatch);
263 CFQ_CFQQ_FNS(fifo_expire);
264 CFQ_CFQQ_FNS(idle_window);
265 CFQ_CFQQ_FNS(prio_changed);
268 enum cfq_rq_state_flags {
269 CFQ_CRQ_FLAG_is_sync = 0,
272 #define CFQ_CRQ_FNS(name) \
273 static inline void cfq_mark_crq_##name(struct cfq_rq *crq) \
275 crq->crq_flags |= (1 << CFQ_CRQ_FLAG_##name); \
277 static inline void cfq_clear_crq_##name(struct cfq_rq *crq) \
279 crq->crq_flags &= ~(1 << CFQ_CRQ_FLAG_##name); \
281 static inline int cfq_crq_##name(const struct cfq_rq *crq) \
283 return (crq->crq_flags & (1 << CFQ_CRQ_FLAG_##name)) != 0; \
286 CFQ_CRQ_FNS(is_sync);
289 static struct cfq_queue *cfq_find_cfq_hash(struct cfq_data *, unsigned int, unsigned short);
290 static void cfq_dispatch_insert(request_queue_t *, struct cfq_rq *);
291 static void cfq_put_cfqd(struct cfq_data *cfqd);
293 #define process_sync(tsk) ((tsk)->flags & PF_SYNCWRITE)
296 * lots of deadline iosched dupes, can be abstracted later...
298 static inline void cfq_del_crq_hash(struct cfq_rq *crq)
300 hlist_del_init(&crq->hash);
303 static inline void cfq_add_crq_hash(struct cfq_data *cfqd, struct cfq_rq *crq)
305 const int hash_idx = CFQ_MHASH_FN(rq_hash_key(crq->request));
307 hlist_add_head(&crq->hash, &cfqd->crq_hash[hash_idx]);
310 static struct request *cfq_find_rq_hash(struct cfq_data *cfqd, sector_t offset)
312 struct hlist_head *hash_list = &cfqd->crq_hash[CFQ_MHASH_FN(offset)];
313 struct hlist_node *entry, *next;
315 hlist_for_each_safe(entry, next, hash_list) {
316 struct cfq_rq *crq = list_entry_hash(entry);
317 struct request *__rq = crq->request;
319 if (!rq_mergeable(__rq)) {
320 cfq_del_crq_hash(crq);
324 if (rq_hash_key(__rq) == offset)
332 * scheduler run of queue, if there are requests pending and no one in the
333 * driver that will restart queueing
335 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
337 if (cfqd->busy_queues)
338 kblockd_schedule_work(&cfqd->unplug_work);
341 static int cfq_queue_empty(request_queue_t *q)
343 struct cfq_data *cfqd = q->elevator->elevator_data;
345 return !cfqd->busy_queues;
349 * Lifted from AS - choose which of crq1 and crq2 that is best served now.
350 * We choose the request that is closest to the head right now. Distance
351 * behind the head are penalized and only allowed to a certain extent.
353 static struct cfq_rq *
354 cfq_choose_req(struct cfq_data *cfqd, struct cfq_rq *crq1, struct cfq_rq *crq2)
356 sector_t last, s1, s2, d1 = 0, d2 = 0;
357 int r1_wrap = 0, r2_wrap = 0; /* requests are behind the disk head */
358 unsigned long back_max;
360 if (crq1 == NULL || crq1 == crq2)
365 if (cfq_crq_is_sync(crq1) && !cfq_crq_is_sync(crq2))
367 else if (cfq_crq_is_sync(crq2) && !cfq_crq_is_sync(crq1))
370 s1 = crq1->request->sector;
371 s2 = crq2->request->sector;
373 last = cfqd->last_sector;
376 * by definition, 1KiB is 2 sectors
378 back_max = cfqd->cfq_back_max * 2;
381 * Strict one way elevator _except_ in the case where we allow
382 * short backward seeks which are biased as twice the cost of a
383 * similar forward seek.
387 else if (s1 + back_max >= last)
388 d1 = (last - s1) * cfqd->cfq_back_penalty;
394 else if (s2 + back_max >= last)
395 d2 = (last - s2) * cfqd->cfq_back_penalty;
399 /* Found required data */
400 if (!r1_wrap && r2_wrap)
402 else if (!r2_wrap && r1_wrap)
404 else if (r1_wrap && r2_wrap) {
405 /* both behind the head */
412 /* Both requests in front of the head */
426 * would be nice to take fifo expire time into account as well
428 static struct cfq_rq *
429 cfq_find_next_crq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
432 struct cfq_rq *crq_next = NULL, *crq_prev = NULL;
433 struct rb_node *rbnext, *rbprev;
435 if (!(rbnext = rb_next(&last->rb_node))) {
436 rbnext = rb_first(&cfqq->sort_list);
437 if (rbnext == &last->rb_node)
441 rbprev = rb_prev(&last->rb_node);
444 crq_prev = rb_entry_crq(rbprev);
446 crq_next = rb_entry_crq(rbnext);
448 return cfq_choose_req(cfqd, crq_next, crq_prev);
451 static void cfq_update_next_crq(struct cfq_rq *crq)
453 struct cfq_queue *cfqq = crq->cfq_queue;
455 if (cfqq->next_crq == crq)
456 cfqq->next_crq = cfq_find_next_crq(cfqq->cfqd, cfqq, crq);
459 static void cfq_resort_rr_list(struct cfq_queue *cfqq, int preempted)
461 struct cfq_data *cfqd = cfqq->cfqd;
462 struct list_head *list, *entry;
464 BUG_ON(!cfq_cfqq_on_rr(cfqq));
466 list_del(&cfqq->cfq_list);
468 if (cfq_class_rt(cfqq))
469 list = &cfqd->cur_rr;
470 else if (cfq_class_idle(cfqq))
471 list = &cfqd->idle_rr;
474 * if cfqq has requests in flight, don't allow it to be
475 * found in cfq_set_active_queue before it has finished them.
476 * this is done to increase fairness between a process that
477 * has lots of io pending vs one that only generates one
478 * sporadically or synchronously
480 if (cfq_cfqq_dispatched(cfqq))
481 list = &cfqd->busy_rr;
483 list = &cfqd->rr_list[cfqq->ioprio];
487 * if queue was preempted, just add to front to be fair. busy_rr
490 if (preempted || list == &cfqd->busy_rr) {
491 list_add(&cfqq->cfq_list, list);
496 * sort by when queue was last serviced
499 while ((entry = entry->prev) != list) {
500 struct cfq_queue *__cfqq = list_entry_cfqq(entry);
502 if (!__cfqq->service_last)
504 if (time_before(__cfqq->service_last, cfqq->service_last))
508 list_add(&cfqq->cfq_list, entry);
512 * add to busy list of queues for service, trying to be fair in ordering
513 * the pending list according to last request service
516 cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
518 BUG_ON(cfq_cfqq_on_rr(cfqq));
519 cfq_mark_cfqq_on_rr(cfqq);
522 cfq_resort_rr_list(cfqq, 0);
526 cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
528 BUG_ON(!cfq_cfqq_on_rr(cfqq));
529 cfq_clear_cfqq_on_rr(cfqq);
530 list_move(&cfqq->cfq_list, &cfqd->empty_list);
532 BUG_ON(!cfqd->busy_queues);
537 * rb tree support functions
539 static inline void cfq_del_crq_rb(struct cfq_rq *crq)
541 struct cfq_queue *cfqq = crq->cfq_queue;
542 struct cfq_data *cfqd = cfqq->cfqd;
543 const int sync = cfq_crq_is_sync(crq);
545 BUG_ON(!cfqq->queued[sync]);
546 cfqq->queued[sync]--;
548 cfq_update_next_crq(crq);
550 rb_erase(&crq->rb_node, &cfqq->sort_list);
551 RB_CLEAR_COLOR(&crq->rb_node);
553 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY(&cfqq->sort_list))
554 cfq_del_cfqq_rr(cfqd, cfqq);
557 static struct cfq_rq *
558 __cfq_add_crq_rb(struct cfq_rq *crq)
560 struct rb_node **p = &crq->cfq_queue->sort_list.rb_node;
561 struct rb_node *parent = NULL;
562 struct cfq_rq *__crq;
566 __crq = rb_entry_crq(parent);
568 if (crq->rb_key < __crq->rb_key)
570 else if (crq->rb_key > __crq->rb_key)
576 rb_link_node(&crq->rb_node, parent, p);
580 static void cfq_add_crq_rb(struct cfq_rq *crq)
582 struct cfq_queue *cfqq = crq->cfq_queue;
583 struct cfq_data *cfqd = cfqq->cfqd;
584 struct request *rq = crq->request;
585 struct cfq_rq *__alias;
587 crq->rb_key = rq_rb_key(rq);
588 cfqq->queued[cfq_crq_is_sync(crq)]++;
591 * looks a little odd, but the first insert might return an alias.
592 * if that happens, put the alias on the dispatch list
594 while ((__alias = __cfq_add_crq_rb(crq)) != NULL)
595 cfq_dispatch_insert(cfqd->queue, __alias);
597 rb_insert_color(&crq->rb_node, &cfqq->sort_list);
599 if (!cfq_cfqq_on_rr(cfqq))
600 cfq_add_cfqq_rr(cfqd, cfqq);
603 * check if this request is a better next-serve candidate
605 cfqq->next_crq = cfq_choose_req(cfqd, cfqq->next_crq, crq);
609 cfq_reposition_crq_rb(struct cfq_queue *cfqq, struct cfq_rq *crq)
611 rb_erase(&crq->rb_node, &cfqq->sort_list);
612 cfqq->queued[cfq_crq_is_sync(crq)]--;
617 static struct request *cfq_find_rq_rb(struct cfq_data *cfqd, sector_t sector)
620 struct cfq_queue *cfqq = cfq_find_cfq_hash(cfqd, current->pid, CFQ_KEY_ANY);
626 n = cfqq->sort_list.rb_node;
628 struct cfq_rq *crq = rb_entry_crq(n);
630 if (sector < crq->rb_key)
632 else if (sector > crq->rb_key)
642 static void cfq_activate_request(request_queue_t *q, struct request *rq)
644 struct cfq_data *cfqd = q->elevator->elevator_data;
646 cfqd->rq_in_driver++;
649 static void cfq_deactivate_request(request_queue_t *q, struct request *rq)
651 struct cfq_data *cfqd = q->elevator->elevator_data;
653 WARN_ON(!cfqd->rq_in_driver);
654 cfqd->rq_in_driver--;
657 static void cfq_remove_request(struct request *rq)
659 struct cfq_rq *crq = RQ_DATA(rq);
661 list_del_init(&rq->queuelist);
663 cfq_del_crq_hash(crq);
667 cfq_merge(request_queue_t *q, struct request **req, struct bio *bio)
669 struct cfq_data *cfqd = q->elevator->elevator_data;
670 struct request *__rq;
673 __rq = cfq_find_rq_hash(cfqd, bio->bi_sector);
674 if (__rq && elv_rq_merge_ok(__rq, bio)) {
675 ret = ELEVATOR_BACK_MERGE;
679 __rq = cfq_find_rq_rb(cfqd, bio->bi_sector + bio_sectors(bio));
680 if (__rq && elv_rq_merge_ok(__rq, bio)) {
681 ret = ELEVATOR_FRONT_MERGE;
685 return ELEVATOR_NO_MERGE;
691 static void cfq_merged_request(request_queue_t *q, struct request *req)
693 struct cfq_data *cfqd = q->elevator->elevator_data;
694 struct cfq_rq *crq = RQ_DATA(req);
696 cfq_del_crq_hash(crq);
697 cfq_add_crq_hash(cfqd, crq);
699 if (rq_rb_key(req) != crq->rb_key) {
700 struct cfq_queue *cfqq = crq->cfq_queue;
702 cfq_update_next_crq(crq);
703 cfq_reposition_crq_rb(cfqq, crq);
708 cfq_merged_requests(request_queue_t *q, struct request *rq,
709 struct request *next)
711 cfq_merged_request(q, rq);
714 * reposition in fifo if next is older than rq
716 if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
717 time_before(next->start_time, rq->start_time))
718 list_move(&rq->queuelist, &next->queuelist);
720 cfq_remove_request(next);
724 __cfq_set_active_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
728 * stop potential idle class queues waiting service
730 del_timer(&cfqd->idle_class_timer);
732 cfqq->slice_start = jiffies;
734 cfqq->slice_left = 0;
735 cfq_clear_cfqq_must_alloc_slice(cfqq);
736 cfq_clear_cfqq_fifo_expire(cfqq);
739 cfqd->active_queue = cfqq;
743 * current cfqq expired its slice (or was too idle), select new one
746 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
749 unsigned long now = jiffies;
751 if (cfq_cfqq_wait_request(cfqq))
752 del_timer(&cfqd->idle_slice_timer);
754 if (!preempted && !cfq_cfqq_dispatched(cfqq)) {
755 cfqq->service_last = now;
756 cfq_schedule_dispatch(cfqd);
759 cfq_clear_cfqq_must_dispatch(cfqq);
760 cfq_clear_cfqq_wait_request(cfqq);
763 * store what was left of this slice, if the queue idled out
766 if (time_after(cfqq->slice_end, now))
767 cfqq->slice_left = cfqq->slice_end - now;
769 cfqq->slice_left = 0;
771 if (cfq_cfqq_on_rr(cfqq))
772 cfq_resort_rr_list(cfqq, preempted);
774 if (cfqq == cfqd->active_queue)
775 cfqd->active_queue = NULL;
777 if (cfqd->active_cic) {
778 put_io_context(cfqd->active_cic->ioc);
779 cfqd->active_cic = NULL;
782 cfqd->dispatch_slice = 0;
785 static inline void cfq_slice_expired(struct cfq_data *cfqd, int preempted)
787 struct cfq_queue *cfqq = cfqd->active_queue;
790 __cfq_slice_expired(cfqd, cfqq, preempted);
803 static int cfq_get_next_prio_level(struct cfq_data *cfqd)
812 for (p = cfqd->cur_prio; p <= cfqd->cur_end_prio; p++) {
813 if (!list_empty(&cfqd->rr_list[p])) {
822 if (++cfqd->cur_end_prio == CFQ_PRIO_LISTS) {
823 cfqd->cur_end_prio = 0;
830 if (unlikely(prio == -1))
833 BUG_ON(prio >= CFQ_PRIO_LISTS);
835 list_splice_init(&cfqd->rr_list[prio], &cfqd->cur_rr);
837 cfqd->cur_prio = prio + 1;
838 if (cfqd->cur_prio > cfqd->cur_end_prio) {
839 cfqd->cur_end_prio = cfqd->cur_prio;
842 if (cfqd->cur_end_prio == CFQ_PRIO_LISTS) {
844 cfqd->cur_end_prio = 0;
850 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd)
852 struct cfq_queue *cfqq = NULL;
855 * if current list is non-empty, grab first entry. if it is empty,
856 * get next prio level and grab first entry then if any are spliced
858 if (!list_empty(&cfqd->cur_rr) || cfq_get_next_prio_level(cfqd) != -1)
859 cfqq = list_entry_cfqq(cfqd->cur_rr.next);
862 * if we have idle queues and no rt or be queues had pending
863 * requests, either allow immediate service if the grace period
864 * has passed or arm the idle grace timer
866 if (!cfqq && !list_empty(&cfqd->idle_rr)) {
867 unsigned long end = cfqd->last_end_request + CFQ_IDLE_GRACE;
869 if (time_after_eq(jiffies, end))
870 cfqq = list_entry_cfqq(cfqd->idle_rr.next);
872 mod_timer(&cfqd->idle_class_timer, end);
875 __cfq_set_active_queue(cfqd, cfqq);
879 static int cfq_arm_slice_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq)
884 WARN_ON(!RB_EMPTY(&cfqq->sort_list));
885 WARN_ON(cfqq != cfqd->active_queue);
888 * idle is disabled, either manually or by past process history
890 if (!cfqd->cfq_slice_idle)
892 if (!cfq_cfqq_idle_window(cfqq))
895 * task has exited, don't wait
897 if (cfqd->active_cic && !cfqd->active_cic->ioc->task)
900 cfq_mark_cfqq_must_dispatch(cfqq);
901 cfq_mark_cfqq_wait_request(cfqq);
903 sl = min(cfqq->slice_end - 1, (unsigned long) cfqd->cfq_slice_idle);
904 mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
908 static void cfq_dispatch_insert(request_queue_t *q, struct cfq_rq *crq)
910 struct cfq_data *cfqd = q->elevator->elevator_data;
911 struct cfq_queue *cfqq = crq->cfq_queue;
913 cfqq->next_crq = cfq_find_next_crq(cfqd, cfqq, crq);
914 cfq_remove_request(crq->request);
915 cfqq->on_dispatch[cfq_crq_is_sync(crq)]++;
916 elv_dispatch_sort(q, crq->request);
920 * return expired entry, or NULL to just start from scratch in rbtree
922 static inline struct cfq_rq *cfq_check_fifo(struct cfq_queue *cfqq)
924 struct cfq_data *cfqd = cfqq->cfqd;
928 if (cfq_cfqq_fifo_expire(cfqq))
931 if (!list_empty(&cfqq->fifo)) {
932 int fifo = cfq_cfqq_class_sync(cfqq);
934 crq = RQ_DATA(list_entry_fifo(cfqq->fifo.next));
936 if (time_after(jiffies, rq->start_time + cfqd->cfq_fifo_expire[fifo])) {
937 cfq_mark_cfqq_fifo_expire(cfqq);
946 * Scale schedule slice based on io priority. Use the sync time slice only
947 * if a queue is marked sync and has sync io queued. A sync queue with async
948 * io only, should not get full sync slice length.
951 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
953 const int base_slice = cfqd->cfq_slice[cfq_cfqq_sync(cfqq)];
955 WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
957 return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - cfqq->ioprio));
961 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
963 cfqq->slice_end = cfq_prio_to_slice(cfqd, cfqq) + jiffies;
967 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
969 const int base_rq = cfqd->cfq_slice_async_rq;
971 WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
973 return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
977 * get next queue for service
979 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
981 unsigned long now = jiffies;
982 struct cfq_queue *cfqq;
984 cfqq = cfqd->active_queue;
991 if (!cfq_cfqq_must_dispatch(cfqq) && time_after(now, cfqq->slice_end))
995 * if queue has requests, dispatch one. if not, check if
996 * enough slice is left to wait for one
998 if (!RB_EMPTY(&cfqq->sort_list))
1000 else if (cfq_cfqq_class_sync(cfqq) &&
1001 time_before(now, cfqq->slice_end)) {
1002 if (cfq_arm_slice_timer(cfqd, cfqq))
1007 cfq_slice_expired(cfqd, 0);
1009 cfqq = cfq_set_active_queue(cfqd);
1015 __cfq_dispatch_requests(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1020 BUG_ON(RB_EMPTY(&cfqq->sort_list));
1026 * follow expired path, else get first next available
1028 if ((crq = cfq_check_fifo(cfqq)) == NULL)
1029 crq = cfqq->next_crq;
1032 * finally, insert request into driver dispatch list
1034 cfq_dispatch_insert(cfqd->queue, crq);
1036 cfqd->dispatch_slice++;
1039 if (!cfqd->active_cic) {
1040 atomic_inc(&crq->io_context->ioc->refcount);
1041 cfqd->active_cic = crq->io_context;
1044 if (RB_EMPTY(&cfqq->sort_list))
1047 } while (dispatched < max_dispatch);
1050 * if slice end isn't set yet, set it. if at least one request was
1051 * sync, use the sync time slice value
1053 if (!cfqq->slice_end)
1054 cfq_set_prio_slice(cfqd, cfqq);
1057 * expire an async queue immediately if it has used up its slice. idle
1058 * queue always expire after 1 dispatch round.
1060 if ((!cfq_cfqq_sync(cfqq) &&
1061 cfqd->dispatch_slice >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
1062 cfq_class_idle(cfqq))
1063 cfq_slice_expired(cfqd, 0);
1069 cfq_forced_dispatch_cfqqs(struct list_head *list)
1072 struct cfq_queue *cfqq, *next;
1075 list_for_each_entry_safe(cfqq, next, list, cfq_list) {
1076 while ((crq = cfqq->next_crq)) {
1077 cfq_dispatch_insert(cfqq->cfqd->queue, crq);
1080 BUG_ON(!list_empty(&cfqq->fifo));
1086 cfq_forced_dispatch(struct cfq_data *cfqd)
1088 int i, dispatched = 0;
1090 for (i = 0; i < CFQ_PRIO_LISTS; i++)
1091 dispatched += cfq_forced_dispatch_cfqqs(&cfqd->rr_list[i]);
1093 dispatched += cfq_forced_dispatch_cfqqs(&cfqd->busy_rr);
1094 dispatched += cfq_forced_dispatch_cfqqs(&cfqd->cur_rr);
1095 dispatched += cfq_forced_dispatch_cfqqs(&cfqd->idle_rr);
1097 cfq_slice_expired(cfqd, 0);
1099 BUG_ON(cfqd->busy_queues);
1105 cfq_dispatch_requests(request_queue_t *q, int force)
1107 struct cfq_data *cfqd = q->elevator->elevator_data;
1108 struct cfq_queue *cfqq;
1110 if (!cfqd->busy_queues)
1113 if (unlikely(force))
1114 return cfq_forced_dispatch(cfqd);
1116 cfqq = cfq_select_queue(cfqd);
1121 * if idle window is disabled, allow queue buildup
1123 if (!cfq_cfqq_idle_window(cfqq) &&
1124 cfqd->rq_in_driver >= cfqd->cfq_max_depth)
1127 cfq_clear_cfqq_must_dispatch(cfqq);
1128 cfq_clear_cfqq_wait_request(cfqq);
1129 del_timer(&cfqd->idle_slice_timer);
1131 max_dispatch = cfqd->cfq_quantum;
1132 if (cfq_class_idle(cfqq))
1135 return __cfq_dispatch_requests(cfqd, cfqq, max_dispatch);
1142 * task holds one reference to the queue, dropped when task exits. each crq
1143 * in-flight on this queue also holds a reference, dropped when crq is freed.
1145 * queue lock must be held here.
1147 static void cfq_put_queue(struct cfq_queue *cfqq)
1149 struct cfq_data *cfqd = cfqq->cfqd;
1151 BUG_ON(atomic_read(&cfqq->ref) <= 0);
1153 if (!atomic_dec_and_test(&cfqq->ref))
1156 BUG_ON(rb_first(&cfqq->sort_list));
1157 BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
1158 BUG_ON(cfq_cfqq_on_rr(cfqq));
1160 if (unlikely(cfqd->active_queue == cfqq))
1161 __cfq_slice_expired(cfqd, cfqq, 0);
1163 cfq_put_cfqd(cfqq->cfqd);
1166 * it's on the empty list and still hashed
1168 list_del(&cfqq->cfq_list);
1169 hlist_del(&cfqq->cfq_hash);
1170 kmem_cache_free(cfq_pool, cfqq);
1173 static inline struct cfq_queue *
1174 __cfq_find_cfq_hash(struct cfq_data *cfqd, unsigned int key, unsigned int prio,
1177 struct hlist_head *hash_list = &cfqd->cfq_hash[hashval];
1178 struct hlist_node *entry, *next;
1180 hlist_for_each_safe(entry, next, hash_list) {
1181 struct cfq_queue *__cfqq = list_entry_qhash(entry);
1182 const unsigned short __p = IOPRIO_PRIO_VALUE(__cfqq->org_ioprio_class, __cfqq->org_ioprio);
1184 if (__cfqq->key == key && (__p == prio || prio == CFQ_KEY_ANY))
1191 static struct cfq_queue *
1192 cfq_find_cfq_hash(struct cfq_data *cfqd, unsigned int key, unsigned short prio)
1194 return __cfq_find_cfq_hash(cfqd, key, prio, hash_long(key, CFQ_QHASH_SHIFT));
1197 static void cfq_free_io_context(struct cfq_io_context *cic)
1199 struct cfq_io_context *__cic;
1200 struct list_head *entry, *next;
1202 list_for_each_safe(entry, next, &cic->list) {
1203 __cic = list_entry(entry, struct cfq_io_context, list);
1204 kmem_cache_free(cfq_ioc_pool, __cic);
1207 kmem_cache_free(cfq_ioc_pool, cic);
1211 * Called with interrupts disabled
1213 static void cfq_exit_single_io_context(struct cfq_io_context *cic)
1215 struct cfq_data *cfqd = cic->key;
1216 request_queue_t *q = cfqd->queue;
1218 WARN_ON(!irqs_disabled());
1220 spin_lock(q->queue_lock);
1222 if (unlikely(cic->cfqq == cfqd->active_queue))
1223 __cfq_slice_expired(cfqd, cic->cfqq, 0);
1225 cfq_put_queue(cic->cfqq);
1228 spin_unlock(q->queue_lock);
1232 * Another task may update the task cic list, if it is doing a queue lookup
1233 * on its behalf. cfq_cic_lock excludes such concurrent updates
1235 static void cfq_exit_io_context(struct cfq_io_context *cic)
1237 struct cfq_io_context *__cic;
1238 struct list_head *entry;
1239 unsigned long flags;
1241 local_irq_save(flags);
1244 * put the reference this task is holding to the various queues
1246 list_for_each(entry, &cic->list) {
1247 __cic = list_entry(entry, struct cfq_io_context, list);
1248 cfq_exit_single_io_context(__cic);
1251 cfq_exit_single_io_context(cic);
1252 local_irq_restore(flags);
1255 static struct cfq_io_context *
1256 cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
1258 struct cfq_io_context *cic = kmem_cache_alloc(cfq_ioc_pool, gfp_mask);
1261 INIT_LIST_HEAD(&cic->list);
1264 cic->last_end_request = jiffies;
1265 cic->ttime_total = 0;
1266 cic->ttime_samples = 0;
1267 cic->ttime_mean = 0;
1268 cic->dtor = cfq_free_io_context;
1269 cic->exit = cfq_exit_io_context;
1275 static void cfq_init_prio_data(struct cfq_queue *cfqq)
1277 struct task_struct *tsk = current;
1280 if (!cfq_cfqq_prio_changed(cfqq))
1283 ioprio_class = IOPRIO_PRIO_CLASS(tsk->ioprio);
1284 switch (ioprio_class) {
1286 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
1287 case IOPRIO_CLASS_NONE:
1289 * no prio set, place us in the middle of the BE classes
1291 cfqq->ioprio = task_nice_ioprio(tsk);
1292 cfqq->ioprio_class = IOPRIO_CLASS_BE;
1294 case IOPRIO_CLASS_RT:
1295 cfqq->ioprio = task_ioprio(tsk);
1296 cfqq->ioprio_class = IOPRIO_CLASS_RT;
1298 case IOPRIO_CLASS_BE:
1299 cfqq->ioprio = task_ioprio(tsk);
1300 cfqq->ioprio_class = IOPRIO_CLASS_BE;
1302 case IOPRIO_CLASS_IDLE:
1303 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
1305 cfq_clear_cfqq_idle_window(cfqq);
1310 * keep track of original prio settings in case we have to temporarily
1311 * elevate the priority of this queue
1313 cfqq->org_ioprio = cfqq->ioprio;
1314 cfqq->org_ioprio_class = cfqq->ioprio_class;
1316 if (cfq_cfqq_on_rr(cfqq))
1317 cfq_resort_rr_list(cfqq, 0);
1319 cfq_clear_cfqq_prio_changed(cfqq);
1322 static inline void changed_ioprio(struct cfq_io_context *cic)
1324 struct cfq_data *cfqd = cic->key;
1325 struct cfq_queue *cfqq;
1327 spin_lock(cfqd->queue->queue_lock);
1330 cfq_mark_cfqq_prio_changed(cfqq);
1331 cfq_init_prio_data(cfqq);
1333 spin_unlock(cfqd->queue->queue_lock);
1338 * callback from sys_ioprio_set, irqs are disabled
1340 static int cfq_ioc_set_ioprio(struct io_context *ioc, unsigned int ioprio)
1342 struct cfq_io_context *cic = ioc->cic;
1344 changed_ioprio(cic);
1346 list_for_each_entry(cic, &cic->list, list)
1347 changed_ioprio(cic);
1352 static struct cfq_queue *
1353 cfq_get_queue(struct cfq_data *cfqd, unsigned int key, unsigned short ioprio,
1356 const int hashval = hash_long(key, CFQ_QHASH_SHIFT);
1357 struct cfq_queue *cfqq, *new_cfqq = NULL;
1360 cfqq = __cfq_find_cfq_hash(cfqd, key, ioprio, hashval);
1366 } else if (gfp_mask & __GFP_WAIT) {
1367 spin_unlock_irq(cfqd->queue->queue_lock);
1368 new_cfqq = kmem_cache_alloc(cfq_pool, gfp_mask);
1369 spin_lock_irq(cfqd->queue->queue_lock);
1372 cfqq = kmem_cache_alloc(cfq_pool, gfp_mask);
1377 memset(cfqq, 0, sizeof(*cfqq));
1379 INIT_HLIST_NODE(&cfqq->cfq_hash);
1380 INIT_LIST_HEAD(&cfqq->cfq_list);
1381 RB_CLEAR_ROOT(&cfqq->sort_list);
1382 INIT_LIST_HEAD(&cfqq->fifo);
1385 hlist_add_head(&cfqq->cfq_hash, &cfqd->cfq_hash[hashval]);
1386 atomic_set(&cfqq->ref, 0);
1388 atomic_inc(&cfqd->ref);
1389 cfqq->service_last = 0;
1391 * set ->slice_left to allow preemption for a new process
1393 cfqq->slice_left = 2 * cfqd->cfq_slice_idle;
1394 cfq_mark_cfqq_idle_window(cfqq);
1395 cfq_mark_cfqq_prio_changed(cfqq);
1396 cfq_init_prio_data(cfqq);
1400 kmem_cache_free(cfq_pool, new_cfqq);
1402 atomic_inc(&cfqq->ref);
1404 WARN_ON((gfp_mask & __GFP_WAIT) && !cfqq);
1409 * Setup general io context and cfq io context. There can be several cfq
1410 * io contexts per general io context, if this process is doing io to more
1411 * than one device managed by cfq. Note that caller is holding a reference to
1412 * cfqq, so we don't need to worry about it disappearing
1414 static struct cfq_io_context *
1415 cfq_get_io_context(struct cfq_data *cfqd, pid_t pid, gfp_t gfp_mask)
1417 struct io_context *ioc = NULL;
1418 struct cfq_io_context *cic;
1420 might_sleep_if(gfp_mask & __GFP_WAIT);
1422 ioc = get_io_context(gfp_mask);
1426 if ((cic = ioc->cic) == NULL) {
1427 cic = cfq_alloc_io_context(cfqd, gfp_mask);
1433 * manually increment generic io_context usage count, it
1434 * cannot go away since we are already holding one ref to it
1438 ioc->set_ioprio = cfq_ioc_set_ioprio;
1441 struct cfq_io_context *__cic;
1444 * the first cic on the list is actually the head itself
1446 if (cic->key == cfqd)
1450 * cic exists, check if we already are there. linear search
1451 * should be ok here, the list will usually not be more than
1452 * 1 or a few entries long
1454 list_for_each_entry(__cic, &cic->list, list) {
1456 * this process is already holding a reference to
1457 * this queue, so no need to get one more
1459 if (__cic->key == cfqd) {
1466 * nope, process doesn't have a cic assoicated with this
1467 * cfqq yet. get a new one and add to list
1469 __cic = cfq_alloc_io_context(cfqd, gfp_mask);
1475 list_add(&__cic->list, &cic->list);
1482 put_io_context(ioc);
1487 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
1489 unsigned long elapsed, ttime;
1492 * if this context already has stuff queued, thinktime is from
1493 * last queue not last end
1496 if (time_after(cic->last_end_request, cic->last_queue))
1497 elapsed = jiffies - cic->last_end_request;
1499 elapsed = jiffies - cic->last_queue;
1501 elapsed = jiffies - cic->last_end_request;
1504 ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
1506 cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
1507 cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
1508 cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
1511 #define sample_valid(samples) ((samples) > 80)
1514 * Disable idle window if the process thinks too long or seeks so much that
1518 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1519 struct cfq_io_context *cic)
1521 int enable_idle = cfq_cfqq_idle_window(cfqq);
1523 if (!cic->ioc->task || !cfqd->cfq_slice_idle)
1525 else if (sample_valid(cic->ttime_samples)) {
1526 if (cic->ttime_mean > cfqd->cfq_slice_idle)
1533 cfq_mark_cfqq_idle_window(cfqq);
1535 cfq_clear_cfqq_idle_window(cfqq);
1540 * Check if new_cfqq should preempt the currently active queue. Return 0 for
1541 * no or if we aren't sure, a 1 will cause a preempt.
1544 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
1547 struct cfq_queue *cfqq = cfqd->active_queue;
1549 if (cfq_class_idle(new_cfqq))
1555 if (cfq_class_idle(cfqq))
1557 if (!cfq_cfqq_wait_request(new_cfqq))
1560 * if it doesn't have slice left, forget it
1562 if (new_cfqq->slice_left < cfqd->cfq_slice_idle)
1564 if (cfq_crq_is_sync(crq) && !cfq_cfqq_sync(cfqq))
1571 * cfqq preempts the active queue. if we allowed preempt with no slice left,
1572 * let it have half of its nominal slice.
1574 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1576 struct cfq_queue *__cfqq, *next;
1578 list_for_each_entry_safe(__cfqq, next, &cfqd->cur_rr, cfq_list)
1579 cfq_resort_rr_list(__cfqq, 1);
1581 if (!cfqq->slice_left)
1582 cfqq->slice_left = cfq_prio_to_slice(cfqd, cfqq) / 2;
1584 cfqq->slice_end = cfqq->slice_left + jiffies;
1585 __cfq_slice_expired(cfqd, cfqq, 1);
1586 __cfq_set_active_queue(cfqd, cfqq);
1590 * should really be a ll_rw_blk.c helper
1592 static void cfq_start_queueing(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1594 request_queue_t *q = cfqd->queue;
1596 if (!blk_queue_plugged(q))
1599 __generic_unplug_device(q);
1603 * Called when a new fs request (crq) is added (to cfqq). Check if there's
1604 * something we should do about it
1607 cfq_crq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1610 struct cfq_io_context *cic;
1612 cfqq->next_crq = cfq_choose_req(cfqd, cfqq->next_crq, crq);
1615 * we never wait for an async request and we don't allow preemption
1616 * of an async request. so just return early
1618 if (!cfq_crq_is_sync(crq))
1621 cic = crq->io_context;
1623 cfq_update_io_thinktime(cfqd, cic);
1624 cfq_update_idle_window(cfqd, cfqq, cic);
1626 cic->last_queue = jiffies;
1628 if (cfqq == cfqd->active_queue) {
1630 * if we are waiting for a request for this queue, let it rip
1631 * immediately and flag that we must not expire this queue
1634 if (cfq_cfqq_wait_request(cfqq)) {
1635 cfq_mark_cfqq_must_dispatch(cfqq);
1636 del_timer(&cfqd->idle_slice_timer);
1637 cfq_start_queueing(cfqd, cfqq);
1639 } else if (cfq_should_preempt(cfqd, cfqq, crq)) {
1641 * not the active queue - expire current slice if it is
1642 * idle and has expired it's mean thinktime or this new queue
1643 * has some old slice time left and is of higher priority
1645 cfq_preempt_queue(cfqd, cfqq);
1646 cfq_mark_cfqq_must_dispatch(cfqq);
1647 cfq_start_queueing(cfqd, cfqq);
1651 static void cfq_insert_request(request_queue_t *q, struct request *rq)
1653 struct cfq_data *cfqd = q->elevator->elevator_data;
1654 struct cfq_rq *crq = RQ_DATA(rq);
1655 struct cfq_queue *cfqq = crq->cfq_queue;
1657 cfq_init_prio_data(cfqq);
1659 cfq_add_crq_rb(crq);
1661 list_add_tail(&rq->queuelist, &cfqq->fifo);
1663 if (rq_mergeable(rq))
1664 cfq_add_crq_hash(cfqd, crq);
1666 cfq_crq_enqueued(cfqd, cfqq, crq);
1669 static void cfq_completed_request(request_queue_t *q, struct request *rq)
1671 struct cfq_rq *crq = RQ_DATA(rq);
1672 struct cfq_queue *cfqq = crq->cfq_queue;
1673 struct cfq_data *cfqd = cfqq->cfqd;
1674 const int sync = cfq_crq_is_sync(crq);
1679 WARN_ON(!cfqd->rq_in_driver);
1680 WARN_ON(!cfqq->on_dispatch[sync]);
1681 cfqd->rq_in_driver--;
1682 cfqq->on_dispatch[sync]--;
1684 if (!cfq_class_idle(cfqq))
1685 cfqd->last_end_request = now;
1687 if (!cfq_cfqq_dispatched(cfqq)) {
1688 if (cfq_cfqq_on_rr(cfqq)) {
1689 cfqq->service_last = now;
1690 cfq_resort_rr_list(cfqq, 0);
1692 cfq_schedule_dispatch(cfqd);
1695 if (cfq_crq_is_sync(crq))
1696 crq->io_context->last_end_request = now;
1699 static struct request *
1700 cfq_former_request(request_queue_t *q, struct request *rq)
1702 struct cfq_rq *crq = RQ_DATA(rq);
1703 struct rb_node *rbprev = rb_prev(&crq->rb_node);
1706 return rb_entry_crq(rbprev)->request;
1711 static struct request *
1712 cfq_latter_request(request_queue_t *q, struct request *rq)
1714 struct cfq_rq *crq = RQ_DATA(rq);
1715 struct rb_node *rbnext = rb_next(&crq->rb_node);
1718 return rb_entry_crq(rbnext)->request;
1724 * we temporarily boost lower priority queues if they are holding fs exclusive
1725 * resources. they are boosted to normal prio (CLASS_BE/4)
1727 static void cfq_prio_boost(struct cfq_queue *cfqq)
1729 const int ioprio_class = cfqq->ioprio_class;
1730 const int ioprio = cfqq->ioprio;
1732 if (has_fs_excl()) {
1734 * boost idle prio on transactions that would lock out other
1735 * users of the filesystem
1737 if (cfq_class_idle(cfqq))
1738 cfqq->ioprio_class = IOPRIO_CLASS_BE;
1739 if (cfqq->ioprio > IOPRIO_NORM)
1740 cfqq->ioprio = IOPRIO_NORM;
1743 * check if we need to unboost the queue
1745 if (cfqq->ioprio_class != cfqq->org_ioprio_class)
1746 cfqq->ioprio_class = cfqq->org_ioprio_class;
1747 if (cfqq->ioprio != cfqq->org_ioprio)
1748 cfqq->ioprio = cfqq->org_ioprio;
1752 * refile between round-robin lists if we moved the priority class
1754 if ((ioprio_class != cfqq->ioprio_class || ioprio != cfqq->ioprio) &&
1755 cfq_cfqq_on_rr(cfqq))
1756 cfq_resort_rr_list(cfqq, 0);
1759 static inline pid_t cfq_queue_pid(struct task_struct *task, int rw)
1761 if (rw == READ || process_sync(task))
1764 return CFQ_KEY_ASYNC;
1768 __cfq_may_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1769 struct task_struct *task, int rw)
1772 if ((cfq_cfqq_wait_request(cfqq) || cfq_cfqq_must_alloc(cfqq)) &&
1773 !cfq_cfqq_must_alloc_slice(cfqq)) {
1774 cfq_mark_cfqq_must_alloc_slice(cfqq);
1775 return ELV_MQUEUE_MUST;
1778 return ELV_MQUEUE_MAY;
1780 if (!cfqq || task->flags & PF_MEMALLOC)
1781 return ELV_MQUEUE_MAY;
1782 if (!cfqq->allocated[rw] || cfq_cfqq_must_alloc(cfqq)) {
1783 if (cfq_cfqq_wait_request(cfqq))
1784 return ELV_MQUEUE_MUST;
1787 * only allow 1 ELV_MQUEUE_MUST per slice, otherwise we
1788 * can quickly flood the queue with writes from a single task
1790 if (rw == READ || !cfq_cfqq_must_alloc_slice(cfqq)) {
1791 cfq_mark_cfqq_must_alloc_slice(cfqq);
1792 return ELV_MQUEUE_MUST;
1795 return ELV_MQUEUE_MAY;
1797 if (cfq_class_idle(cfqq))
1798 return ELV_MQUEUE_NO;
1799 if (cfqq->allocated[rw] >= cfqd->max_queued) {
1800 struct io_context *ioc = get_io_context(GFP_ATOMIC);
1801 int ret = ELV_MQUEUE_NO;
1803 if (ioc && ioc->nr_batch_requests)
1804 ret = ELV_MQUEUE_MAY;
1806 put_io_context(ioc);
1810 return ELV_MQUEUE_MAY;
1814 static int cfq_may_queue(request_queue_t *q, int rw, struct bio *bio)
1816 struct cfq_data *cfqd = q->elevator->elevator_data;
1817 struct task_struct *tsk = current;
1818 struct cfq_queue *cfqq;
1821 * don't force setup of a queue from here, as a call to may_queue
1822 * does not necessarily imply that a request actually will be queued.
1823 * so just lookup a possibly existing queue, or return 'may queue'
1826 cfqq = cfq_find_cfq_hash(cfqd, cfq_queue_pid(tsk, rw), tsk->ioprio);
1828 cfq_init_prio_data(cfqq);
1829 cfq_prio_boost(cfqq);
1831 return __cfq_may_queue(cfqd, cfqq, tsk, rw);
1834 return ELV_MQUEUE_MAY;
1837 static void cfq_check_waiters(request_queue_t *q, struct cfq_queue *cfqq)
1839 struct cfq_data *cfqd = q->elevator->elevator_data;
1840 struct request_list *rl = &q->rq;
1842 if (cfqq->allocated[READ] <= cfqd->max_queued || cfqd->rq_starved) {
1844 if (waitqueue_active(&rl->wait[READ]))
1845 wake_up(&rl->wait[READ]);
1848 if (cfqq->allocated[WRITE] <= cfqd->max_queued || cfqd->rq_starved) {
1850 if (waitqueue_active(&rl->wait[WRITE]))
1851 wake_up(&rl->wait[WRITE]);
1856 * queue lock held here
1858 static void cfq_put_request(request_queue_t *q, struct request *rq)
1860 struct cfq_data *cfqd = q->elevator->elevator_data;
1861 struct cfq_rq *crq = RQ_DATA(rq);
1864 struct cfq_queue *cfqq = crq->cfq_queue;
1865 const int rw = rq_data_dir(rq);
1867 BUG_ON(!cfqq->allocated[rw]);
1868 cfqq->allocated[rw]--;
1870 put_io_context(crq->io_context->ioc);
1872 mempool_free(crq, cfqd->crq_pool);
1873 rq->elevator_private = NULL;
1875 cfq_check_waiters(q, cfqq);
1876 cfq_put_queue(cfqq);
1881 * Allocate cfq data structures associated with this request.
1884 cfq_set_request(request_queue_t *q, struct request *rq, struct bio *bio,
1887 struct cfq_data *cfqd = q->elevator->elevator_data;
1888 struct task_struct *tsk = current;
1889 struct cfq_io_context *cic;
1890 const int rw = rq_data_dir(rq);
1891 pid_t key = cfq_queue_pid(tsk, rw);
1892 struct cfq_queue *cfqq;
1894 unsigned long flags;
1896 might_sleep_if(gfp_mask & __GFP_WAIT);
1898 cic = cfq_get_io_context(cfqd, key, gfp_mask);
1900 spin_lock_irqsave(q->queue_lock, flags);
1906 cfqq = cfq_get_queue(cfqd, key, tsk->ioprio, gfp_mask);
1914 cfqq->allocated[rw]++;
1915 cfq_clear_cfqq_must_alloc(cfqq);
1916 cfqd->rq_starved = 0;
1917 atomic_inc(&cfqq->ref);
1918 spin_unlock_irqrestore(q->queue_lock, flags);
1920 crq = mempool_alloc(cfqd->crq_pool, gfp_mask);
1922 RB_CLEAR(&crq->rb_node);
1925 INIT_HLIST_NODE(&crq->hash);
1926 crq->cfq_queue = cfqq;
1927 crq->io_context = cic;
1929 if (rw == READ || process_sync(tsk))
1930 cfq_mark_crq_is_sync(crq);
1932 cfq_clear_crq_is_sync(crq);
1934 rq->elevator_private = crq;
1938 spin_lock_irqsave(q->queue_lock, flags);
1939 cfqq->allocated[rw]--;
1940 if (!(cfqq->allocated[0] + cfqq->allocated[1]))
1941 cfq_mark_cfqq_must_alloc(cfqq);
1942 cfq_put_queue(cfqq);
1945 put_io_context(cic->ioc);
1947 * mark us rq allocation starved. we need to kickstart the process
1948 * ourselves if there are no pending requests that can do it for us.
1949 * that would be an extremely rare OOM situation
1951 cfqd->rq_starved = 1;
1952 cfq_schedule_dispatch(cfqd);
1953 spin_unlock_irqrestore(q->queue_lock, flags);
1957 static void cfq_kick_queue(void *data)
1959 request_queue_t *q = data;
1960 struct cfq_data *cfqd = q->elevator->elevator_data;
1961 unsigned long flags;
1963 spin_lock_irqsave(q->queue_lock, flags);
1965 if (cfqd->rq_starved) {
1966 struct request_list *rl = &q->rq;
1969 * we aren't guaranteed to get a request after this, but we
1970 * have to be opportunistic
1973 if (waitqueue_active(&rl->wait[READ]))
1974 wake_up(&rl->wait[READ]);
1975 if (waitqueue_active(&rl->wait[WRITE]))
1976 wake_up(&rl->wait[WRITE]);
1981 spin_unlock_irqrestore(q->queue_lock, flags);
1985 * Timer running if the active_queue is currently idling inside its time slice
1987 static void cfq_idle_slice_timer(unsigned long data)
1989 struct cfq_data *cfqd = (struct cfq_data *) data;
1990 struct cfq_queue *cfqq;
1991 unsigned long flags;
1993 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1995 if ((cfqq = cfqd->active_queue) != NULL) {
1996 unsigned long now = jiffies;
2001 if (time_after(now, cfqq->slice_end))
2005 * only expire and reinvoke request handler, if there are
2006 * other queues with pending requests
2008 if (!cfqd->busy_queues) {
2009 cfqd->idle_slice_timer.expires = min(now + cfqd->cfq_slice_idle, cfqq->slice_end);
2010 add_timer(&cfqd->idle_slice_timer);
2015 * not expired and it has a request pending, let it dispatch
2017 if (!RB_EMPTY(&cfqq->sort_list)) {
2018 cfq_mark_cfqq_must_dispatch(cfqq);
2023 cfq_slice_expired(cfqd, 0);
2025 cfq_schedule_dispatch(cfqd);
2027 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2031 * Timer running if an idle class queue is waiting for service
2033 static void cfq_idle_class_timer(unsigned long data)
2035 struct cfq_data *cfqd = (struct cfq_data *) data;
2036 unsigned long flags, end;
2038 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2041 * race with a non-idle queue, reset timer
2043 end = cfqd->last_end_request + CFQ_IDLE_GRACE;
2044 if (!time_after_eq(jiffies, end)) {
2045 cfqd->idle_class_timer.expires = end;
2046 add_timer(&cfqd->idle_class_timer);
2048 cfq_schedule_dispatch(cfqd);
2050 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2053 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
2055 del_timer_sync(&cfqd->idle_slice_timer);
2056 del_timer_sync(&cfqd->idle_class_timer);
2057 blk_sync_queue(cfqd->queue);
2060 static void cfq_put_cfqd(struct cfq_data *cfqd)
2062 request_queue_t *q = cfqd->queue;
2064 if (!atomic_dec_and_test(&cfqd->ref))
2067 cfq_shutdown_timer_wq(cfqd);
2070 mempool_destroy(cfqd->crq_pool);
2071 kfree(cfqd->crq_hash);
2072 kfree(cfqd->cfq_hash);
2076 static void cfq_exit_queue(elevator_t *e)
2078 struct cfq_data *cfqd = e->elevator_data;
2080 cfq_shutdown_timer_wq(cfqd);
2084 static int cfq_init_queue(request_queue_t *q, elevator_t *e)
2086 struct cfq_data *cfqd;
2089 cfqd = kmalloc(sizeof(*cfqd), GFP_KERNEL);
2093 memset(cfqd, 0, sizeof(*cfqd));
2095 for (i = 0; i < CFQ_PRIO_LISTS; i++)
2096 INIT_LIST_HEAD(&cfqd->rr_list[i]);
2098 INIT_LIST_HEAD(&cfqd->busy_rr);
2099 INIT_LIST_HEAD(&cfqd->cur_rr);
2100 INIT_LIST_HEAD(&cfqd->idle_rr);
2101 INIT_LIST_HEAD(&cfqd->empty_list);
2103 cfqd->crq_hash = kmalloc(sizeof(struct hlist_head) * CFQ_MHASH_ENTRIES, GFP_KERNEL);
2104 if (!cfqd->crq_hash)
2107 cfqd->cfq_hash = kmalloc(sizeof(struct hlist_head) * CFQ_QHASH_ENTRIES, GFP_KERNEL);
2108 if (!cfqd->cfq_hash)
2111 cfqd->crq_pool = mempool_create(BLKDEV_MIN_RQ, mempool_alloc_slab, mempool_free_slab, crq_pool);
2112 if (!cfqd->crq_pool)
2115 for (i = 0; i < CFQ_MHASH_ENTRIES; i++)
2116 INIT_HLIST_HEAD(&cfqd->crq_hash[i]);
2117 for (i = 0; i < CFQ_QHASH_ENTRIES; i++)
2118 INIT_HLIST_HEAD(&cfqd->cfq_hash[i]);
2120 e->elevator_data = cfqd;
2123 atomic_inc(&q->refcnt);
2125 cfqd->max_queued = q->nr_requests / 4;
2126 q->nr_batching = cfq_queued;
2128 init_timer(&cfqd->idle_slice_timer);
2129 cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
2130 cfqd->idle_slice_timer.data = (unsigned long) cfqd;
2132 init_timer(&cfqd->idle_class_timer);
2133 cfqd->idle_class_timer.function = cfq_idle_class_timer;
2134 cfqd->idle_class_timer.data = (unsigned long) cfqd;
2136 INIT_WORK(&cfqd->unplug_work, cfq_kick_queue, q);
2138 atomic_set(&cfqd->ref, 1);
2140 cfqd->cfq_queued = cfq_queued;
2141 cfqd->cfq_quantum = cfq_quantum;
2142 cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
2143 cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
2144 cfqd->cfq_back_max = cfq_back_max;
2145 cfqd->cfq_back_penalty = cfq_back_penalty;
2146 cfqd->cfq_slice[0] = cfq_slice_async;
2147 cfqd->cfq_slice[1] = cfq_slice_sync;
2148 cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
2149 cfqd->cfq_slice_idle = cfq_slice_idle;
2150 cfqd->cfq_max_depth = cfq_max_depth;
2154 kfree(cfqd->cfq_hash);
2156 kfree(cfqd->crq_hash);
2162 static void cfq_slab_kill(void)
2165 kmem_cache_destroy(crq_pool);
2167 kmem_cache_destroy(cfq_pool);
2169 kmem_cache_destroy(cfq_ioc_pool);
2172 static int __init cfq_slab_setup(void)
2174 crq_pool = kmem_cache_create("crq_pool", sizeof(struct cfq_rq), 0, 0,
2179 cfq_pool = kmem_cache_create("cfq_pool", sizeof(struct cfq_queue), 0, 0,
2184 cfq_ioc_pool = kmem_cache_create("cfq_ioc_pool",
2185 sizeof(struct cfq_io_context), 0, 0, NULL, NULL);
2196 * sysfs parts below -->
2198 struct cfq_fs_entry {
2199 struct attribute attr;
2200 ssize_t (*show)(struct cfq_data *, char *);
2201 ssize_t (*store)(struct cfq_data *, const char *, size_t);
2205 cfq_var_show(unsigned int var, char *page)
2207 return sprintf(page, "%d\n", var);
2211 cfq_var_store(unsigned int *var, const char *page, size_t count)
2213 char *p = (char *) page;
2215 *var = simple_strtoul(p, &p, 10);
2219 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
2220 static ssize_t __FUNC(struct cfq_data *cfqd, char *page) \
2222 unsigned int __data = __VAR; \
2224 __data = jiffies_to_msecs(__data); \
2225 return cfq_var_show(__data, (page)); \
2227 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
2228 SHOW_FUNCTION(cfq_queued_show, cfqd->cfq_queued, 0);
2229 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
2230 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
2231 SHOW_FUNCTION(cfq_back_max_show, cfqd->cfq_back_max, 0);
2232 SHOW_FUNCTION(cfq_back_penalty_show, cfqd->cfq_back_penalty, 0);
2233 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
2234 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
2235 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
2236 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
2237 SHOW_FUNCTION(cfq_max_depth_show, cfqd->cfq_max_depth, 0);
2238 #undef SHOW_FUNCTION
2240 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
2241 static ssize_t __FUNC(struct cfq_data *cfqd, const char *page, size_t count) \
2243 unsigned int __data; \
2244 int ret = cfq_var_store(&__data, (page), count); \
2245 if (__data < (MIN)) \
2247 else if (__data > (MAX)) \
2250 *(__PTR) = msecs_to_jiffies(__data); \
2252 *(__PTR) = __data; \
2255 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
2256 STORE_FUNCTION(cfq_queued_store, &cfqd->cfq_queued, 1, UINT_MAX, 0);
2257 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1, UINT_MAX, 1);
2258 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1, UINT_MAX, 1);
2259 STORE_FUNCTION(cfq_back_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
2260 STORE_FUNCTION(cfq_back_penalty_store, &cfqd->cfq_back_penalty, 1, UINT_MAX, 0);
2261 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
2262 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
2263 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
2264 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1, UINT_MAX, 0);
2265 STORE_FUNCTION(cfq_max_depth_store, &cfqd->cfq_max_depth, 1, UINT_MAX, 0);
2266 #undef STORE_FUNCTION
2268 static struct cfq_fs_entry cfq_quantum_entry = {
2269 .attr = {.name = "quantum", .mode = S_IRUGO | S_IWUSR },
2270 .show = cfq_quantum_show,
2271 .store = cfq_quantum_store,
2273 static struct cfq_fs_entry cfq_queued_entry = {
2274 .attr = {.name = "queued", .mode = S_IRUGO | S_IWUSR },
2275 .show = cfq_queued_show,
2276 .store = cfq_queued_store,
2278 static struct cfq_fs_entry cfq_fifo_expire_sync_entry = {
2279 .attr = {.name = "fifo_expire_sync", .mode = S_IRUGO | S_IWUSR },
2280 .show = cfq_fifo_expire_sync_show,
2281 .store = cfq_fifo_expire_sync_store,
2283 static struct cfq_fs_entry cfq_fifo_expire_async_entry = {
2284 .attr = {.name = "fifo_expire_async", .mode = S_IRUGO | S_IWUSR },
2285 .show = cfq_fifo_expire_async_show,
2286 .store = cfq_fifo_expire_async_store,
2288 static struct cfq_fs_entry cfq_back_max_entry = {
2289 .attr = {.name = "back_seek_max", .mode = S_IRUGO | S_IWUSR },
2290 .show = cfq_back_max_show,
2291 .store = cfq_back_max_store,
2293 static struct cfq_fs_entry cfq_back_penalty_entry = {
2294 .attr = {.name = "back_seek_penalty", .mode = S_IRUGO | S_IWUSR },
2295 .show = cfq_back_penalty_show,
2296 .store = cfq_back_penalty_store,
2298 static struct cfq_fs_entry cfq_slice_sync_entry = {
2299 .attr = {.name = "slice_sync", .mode = S_IRUGO | S_IWUSR },
2300 .show = cfq_slice_sync_show,
2301 .store = cfq_slice_sync_store,
2303 static struct cfq_fs_entry cfq_slice_async_entry = {
2304 .attr = {.name = "slice_async", .mode = S_IRUGO | S_IWUSR },
2305 .show = cfq_slice_async_show,
2306 .store = cfq_slice_async_store,
2308 static struct cfq_fs_entry cfq_slice_async_rq_entry = {
2309 .attr = {.name = "slice_async_rq", .mode = S_IRUGO | S_IWUSR },
2310 .show = cfq_slice_async_rq_show,
2311 .store = cfq_slice_async_rq_store,
2313 static struct cfq_fs_entry cfq_slice_idle_entry = {
2314 .attr = {.name = "slice_idle", .mode = S_IRUGO | S_IWUSR },
2315 .show = cfq_slice_idle_show,
2316 .store = cfq_slice_idle_store,
2318 static struct cfq_fs_entry cfq_max_depth_entry = {
2319 .attr = {.name = "max_depth", .mode = S_IRUGO | S_IWUSR },
2320 .show = cfq_max_depth_show,
2321 .store = cfq_max_depth_store,
2324 static struct attribute *default_attrs[] = {
2325 &cfq_quantum_entry.attr,
2326 &cfq_queued_entry.attr,
2327 &cfq_fifo_expire_sync_entry.attr,
2328 &cfq_fifo_expire_async_entry.attr,
2329 &cfq_back_max_entry.attr,
2330 &cfq_back_penalty_entry.attr,
2331 &cfq_slice_sync_entry.attr,
2332 &cfq_slice_async_entry.attr,
2333 &cfq_slice_async_rq_entry.attr,
2334 &cfq_slice_idle_entry.attr,
2335 &cfq_max_depth_entry.attr,
2339 #define to_cfq(atr) container_of((atr), struct cfq_fs_entry, attr)
2342 cfq_attr_show(struct kobject *kobj, struct attribute *attr, char *page)
2344 elevator_t *e = container_of(kobj, elevator_t, kobj);
2345 struct cfq_fs_entry *entry = to_cfq(attr);
2350 return entry->show(e->elevator_data, page);
2354 cfq_attr_store(struct kobject *kobj, struct attribute *attr,
2355 const char *page, size_t length)
2357 elevator_t *e = container_of(kobj, elevator_t, kobj);
2358 struct cfq_fs_entry *entry = to_cfq(attr);
2363 return entry->store(e->elevator_data, page, length);
2366 static struct sysfs_ops cfq_sysfs_ops = {
2367 .show = cfq_attr_show,
2368 .store = cfq_attr_store,
2371 static struct kobj_type cfq_ktype = {
2372 .sysfs_ops = &cfq_sysfs_ops,
2373 .default_attrs = default_attrs,
2376 static struct elevator_type iosched_cfq = {
2378 .elevator_merge_fn = cfq_merge,
2379 .elevator_merged_fn = cfq_merged_request,
2380 .elevator_merge_req_fn = cfq_merged_requests,
2381 .elevator_dispatch_fn = cfq_dispatch_requests,
2382 .elevator_add_req_fn = cfq_insert_request,
2383 .elevator_activate_req_fn = cfq_activate_request,
2384 .elevator_deactivate_req_fn = cfq_deactivate_request,
2385 .elevator_queue_empty_fn = cfq_queue_empty,
2386 .elevator_completed_req_fn = cfq_completed_request,
2387 .elevator_former_req_fn = cfq_former_request,
2388 .elevator_latter_req_fn = cfq_latter_request,
2389 .elevator_set_req_fn = cfq_set_request,
2390 .elevator_put_req_fn = cfq_put_request,
2391 .elevator_may_queue_fn = cfq_may_queue,
2392 .elevator_init_fn = cfq_init_queue,
2393 .elevator_exit_fn = cfq_exit_queue,
2395 .elevator_ktype = &cfq_ktype,
2396 .elevator_name = "cfq",
2397 .elevator_owner = THIS_MODULE,
2400 static int __init cfq_init(void)
2405 * could be 0 on HZ < 1000 setups
2407 if (!cfq_slice_async)
2408 cfq_slice_async = 1;
2409 if (!cfq_slice_idle)
2412 if (cfq_slab_setup())
2415 ret = elv_register(&iosched_cfq);
2422 static void __exit cfq_exit(void)
2424 elv_unregister(&iosched_cfq);
2428 module_init(cfq_init);
2429 module_exit(cfq_exit);
2431 MODULE_AUTHOR("Jens Axboe");
2432 MODULE_LICENSE("GPL");
2433 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");