2 * CFQ, or complete fairness queueing, disk scheduler.
4 * Based on ideas from a previously unfinished io
5 * scheduler (round robin per-process disk scheduling) and Andrea Arcangeli.
7 * Copyright (C) 2003 Jens Axboe <axboe@suse.de>
9 #include <linux/config.h>
10 #include <linux/module.h>
11 #include <linux/blkdev.h>
12 #include <linux/elevator.h>
13 #include <linux/hash.h>
14 #include <linux/rbtree.h>
15 #include <linux/ioprio.h>
20 static const int cfq_quantum = 4; /* max queue in one round of service */
21 static const int cfq_queued = 8; /* minimum rq allocate limit per-queue*/
22 static const int cfq_fifo_expire[2] = { HZ / 4, HZ / 8 };
23 static const int cfq_back_max = 16 * 1024; /* maximum backwards seek, in KiB */
24 static const int cfq_back_penalty = 2; /* penalty of a backwards seek */
26 static const int cfq_slice_sync = HZ / 10;
27 static int cfq_slice_async = HZ / 25;
28 static const int cfq_slice_async_rq = 2;
29 static int cfq_slice_idle = HZ / 100;
31 #define CFQ_IDLE_GRACE (HZ / 10)
32 #define CFQ_SLICE_SCALE (5)
34 #define CFQ_KEY_ASYNC (0)
35 #define CFQ_KEY_ANY (0xffff)
38 * disable queueing at the driver/hardware level
40 static const int cfq_max_depth = 2;
42 static DEFINE_RWLOCK(cfq_exit_lock);
45 * for the hash of cfqq inside the cfqd
47 #define CFQ_QHASH_SHIFT 6
48 #define CFQ_QHASH_ENTRIES (1 << CFQ_QHASH_SHIFT)
49 #define list_entry_qhash(entry) hlist_entry((entry), struct cfq_queue, cfq_hash)
52 * for the hash of crq inside the cfqq
54 #define CFQ_MHASH_SHIFT 6
55 #define CFQ_MHASH_BLOCK(sec) ((sec) >> 3)
56 #define CFQ_MHASH_ENTRIES (1 << CFQ_MHASH_SHIFT)
57 #define CFQ_MHASH_FN(sec) hash_long(CFQ_MHASH_BLOCK(sec), CFQ_MHASH_SHIFT)
58 #define rq_hash_key(rq) ((rq)->sector + (rq)->nr_sectors)
59 #define list_entry_hash(ptr) hlist_entry((ptr), struct cfq_rq, hash)
61 #define list_entry_cfqq(ptr) list_entry((ptr), struct cfq_queue, cfq_list)
62 #define list_entry_fifo(ptr) list_entry((ptr), struct request, queuelist)
64 #define RQ_DATA(rq) (rq)->elevator_private
70 #define RB_EMPTY(node) ((node)->rb_node == NULL)
71 #define RB_CLEAR_COLOR(node) (node)->rb_color = RB_NONE
72 #define RB_CLEAR(node) do { \
73 (node)->rb_parent = NULL; \
74 RB_CLEAR_COLOR((node)); \
75 (node)->rb_right = NULL; \
76 (node)->rb_left = NULL; \
78 #define RB_CLEAR_ROOT(root) ((root)->rb_node = NULL)
79 #define rb_entry_crq(node) rb_entry((node), struct cfq_rq, rb_node)
80 #define rq_rb_key(rq) (rq)->sector
82 static kmem_cache_t *crq_pool;
83 static kmem_cache_t *cfq_pool;
84 static kmem_cache_t *cfq_ioc_pool;
86 static atomic_t ioc_count = ATOMIC_INIT(0);
87 static struct completion *ioc_gone;
89 #define CFQ_PRIO_LISTS IOPRIO_BE_NR
90 #define cfq_class_idle(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
91 #define cfq_class_be(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_BE)
92 #define cfq_class_rt(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
97 #define cfq_cfqq_dispatched(cfqq) \
98 ((cfqq)->on_dispatch[ASYNC] + (cfqq)->on_dispatch[SYNC])
100 #define cfq_cfqq_class_sync(cfqq) ((cfqq)->key != CFQ_KEY_ASYNC)
102 #define cfq_cfqq_sync(cfqq) \
103 (cfq_cfqq_class_sync(cfqq) || (cfqq)->on_dispatch[SYNC])
106 * Per block device queue structure
109 request_queue_t *queue;
112 * rr list of queues with requests and the count of them
114 struct list_head rr_list[CFQ_PRIO_LISTS];
115 struct list_head busy_rr;
116 struct list_head cur_rr;
117 struct list_head idle_rr;
118 unsigned int busy_queues;
121 * non-ordered list of empty cfqq's
123 struct list_head empty_list;
128 struct hlist_head *cfq_hash;
131 * global crq hash for all queues
133 struct hlist_head *crq_hash;
135 unsigned int max_queued;
142 * schedule slice state info
145 * idle window management
147 struct timer_list idle_slice_timer;
148 struct work_struct unplug_work;
150 struct cfq_queue *active_queue;
151 struct cfq_io_context *active_cic;
152 int cur_prio, cur_end_prio;
153 unsigned int dispatch_slice;
155 struct timer_list idle_class_timer;
157 sector_t last_sector;
158 unsigned long last_end_request;
160 unsigned int rq_starved;
163 * tunables, see top of file
165 unsigned int cfq_quantum;
166 unsigned int cfq_queued;
167 unsigned int cfq_fifo_expire[2];
168 unsigned int cfq_back_penalty;
169 unsigned int cfq_back_max;
170 unsigned int cfq_slice[2];
171 unsigned int cfq_slice_async_rq;
172 unsigned int cfq_slice_idle;
173 unsigned int cfq_max_depth;
175 struct list_head cic_list;
179 * Per process-grouping structure
182 /* reference count */
184 /* parent cfq_data */
185 struct cfq_data *cfqd;
186 /* cfqq lookup hash */
187 struct hlist_node cfq_hash;
190 /* on either rr or empty list of cfqd */
191 struct list_head cfq_list;
192 /* sorted list of pending requests */
193 struct rb_root sort_list;
194 /* if fifo isn't expired, next request to serve */
195 struct cfq_rq *next_crq;
196 /* requests queued in sort_list */
198 /* currently allocated requests */
200 /* fifo list of requests in sort_list */
201 struct list_head fifo;
203 unsigned long slice_start;
204 unsigned long slice_end;
205 unsigned long slice_left;
206 unsigned long service_last;
208 /* number of requests that are on the dispatch list */
211 /* io prio of this group */
212 unsigned short ioprio, org_ioprio;
213 unsigned short ioprio_class, org_ioprio_class;
215 /* various state flags, see below */
220 struct rb_node rb_node;
222 struct request *request;
223 struct hlist_node hash;
225 struct cfq_queue *cfq_queue;
226 struct cfq_io_context *io_context;
228 unsigned int crq_flags;
231 enum cfqq_state_flags {
232 CFQ_CFQQ_FLAG_on_rr = 0,
233 CFQ_CFQQ_FLAG_wait_request,
234 CFQ_CFQQ_FLAG_must_alloc,
235 CFQ_CFQQ_FLAG_must_alloc_slice,
236 CFQ_CFQQ_FLAG_must_dispatch,
237 CFQ_CFQQ_FLAG_fifo_expire,
238 CFQ_CFQQ_FLAG_idle_window,
239 CFQ_CFQQ_FLAG_prio_changed,
242 #define CFQ_CFQQ_FNS(name) \
243 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \
245 cfqq->flags |= (1 << CFQ_CFQQ_FLAG_##name); \
247 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \
249 cfqq->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \
251 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \
253 return (cfqq->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \
257 CFQ_CFQQ_FNS(wait_request);
258 CFQ_CFQQ_FNS(must_alloc);
259 CFQ_CFQQ_FNS(must_alloc_slice);
260 CFQ_CFQQ_FNS(must_dispatch);
261 CFQ_CFQQ_FNS(fifo_expire);
262 CFQ_CFQQ_FNS(idle_window);
263 CFQ_CFQQ_FNS(prio_changed);
266 enum cfq_rq_state_flags {
267 CFQ_CRQ_FLAG_is_sync = 0,
270 #define CFQ_CRQ_FNS(name) \
271 static inline void cfq_mark_crq_##name(struct cfq_rq *crq) \
273 crq->crq_flags |= (1 << CFQ_CRQ_FLAG_##name); \
275 static inline void cfq_clear_crq_##name(struct cfq_rq *crq) \
277 crq->crq_flags &= ~(1 << CFQ_CRQ_FLAG_##name); \
279 static inline int cfq_crq_##name(const struct cfq_rq *crq) \
281 return (crq->crq_flags & (1 << CFQ_CRQ_FLAG_##name)) != 0; \
284 CFQ_CRQ_FNS(is_sync);
287 static struct cfq_queue *cfq_find_cfq_hash(struct cfq_data *, unsigned int, unsigned short);
288 static void cfq_dispatch_insert(request_queue_t *, struct cfq_rq *);
289 static struct cfq_queue *cfq_get_queue(struct cfq_data *cfqd, unsigned int key, struct task_struct *tsk, gfp_t gfp_mask);
291 #define process_sync(tsk) ((tsk)->flags & PF_SYNCWRITE)
294 * lots of deadline iosched dupes, can be abstracted later...
296 static inline void cfq_del_crq_hash(struct cfq_rq *crq)
298 hlist_del_init(&crq->hash);
301 static inline void cfq_add_crq_hash(struct cfq_data *cfqd, struct cfq_rq *crq)
303 const int hash_idx = CFQ_MHASH_FN(rq_hash_key(crq->request));
305 hlist_add_head(&crq->hash, &cfqd->crq_hash[hash_idx]);
308 static struct request *cfq_find_rq_hash(struct cfq_data *cfqd, sector_t offset)
310 struct hlist_head *hash_list = &cfqd->crq_hash[CFQ_MHASH_FN(offset)];
311 struct hlist_node *entry, *next;
313 hlist_for_each_safe(entry, next, hash_list) {
314 struct cfq_rq *crq = list_entry_hash(entry);
315 struct request *__rq = crq->request;
317 if (!rq_mergeable(__rq)) {
318 cfq_del_crq_hash(crq);
322 if (rq_hash_key(__rq) == offset)
330 * scheduler run of queue, if there are requests pending and no one in the
331 * driver that will restart queueing
333 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
335 if (cfqd->busy_queues)
336 kblockd_schedule_work(&cfqd->unplug_work);
339 static int cfq_queue_empty(request_queue_t *q)
341 struct cfq_data *cfqd = q->elevator->elevator_data;
343 return !cfqd->busy_queues;
347 * Lifted from AS - choose which of crq1 and crq2 that is best served now.
348 * We choose the request that is closest to the head right now. Distance
349 * behind the head are penalized and only allowed to a certain extent.
351 static struct cfq_rq *
352 cfq_choose_req(struct cfq_data *cfqd, struct cfq_rq *crq1, struct cfq_rq *crq2)
354 sector_t last, s1, s2, d1 = 0, d2 = 0;
355 int r1_wrap = 0, r2_wrap = 0; /* requests are behind the disk head */
356 unsigned long back_max;
358 if (crq1 == NULL || crq1 == crq2)
363 if (cfq_crq_is_sync(crq1) && !cfq_crq_is_sync(crq2))
365 else if (cfq_crq_is_sync(crq2) && !cfq_crq_is_sync(crq1))
368 s1 = crq1->request->sector;
369 s2 = crq2->request->sector;
371 last = cfqd->last_sector;
374 * by definition, 1KiB is 2 sectors
376 back_max = cfqd->cfq_back_max * 2;
379 * Strict one way elevator _except_ in the case where we allow
380 * short backward seeks which are biased as twice the cost of a
381 * similar forward seek.
385 else if (s1 + back_max >= last)
386 d1 = (last - s1) * cfqd->cfq_back_penalty;
392 else if (s2 + back_max >= last)
393 d2 = (last - s2) * cfqd->cfq_back_penalty;
397 /* Found required data */
398 if (!r1_wrap && r2_wrap)
400 else if (!r2_wrap && r1_wrap)
402 else if (r1_wrap && r2_wrap) {
403 /* both behind the head */
410 /* Both requests in front of the head */
424 * would be nice to take fifo expire time into account as well
426 static struct cfq_rq *
427 cfq_find_next_crq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
430 struct cfq_rq *crq_next = NULL, *crq_prev = NULL;
431 struct rb_node *rbnext, *rbprev;
433 if (!(rbnext = rb_next(&last->rb_node))) {
434 rbnext = rb_first(&cfqq->sort_list);
435 if (rbnext == &last->rb_node)
439 rbprev = rb_prev(&last->rb_node);
442 crq_prev = rb_entry_crq(rbprev);
444 crq_next = rb_entry_crq(rbnext);
446 return cfq_choose_req(cfqd, crq_next, crq_prev);
449 static void cfq_update_next_crq(struct cfq_rq *crq)
451 struct cfq_queue *cfqq = crq->cfq_queue;
453 if (cfqq->next_crq == crq)
454 cfqq->next_crq = cfq_find_next_crq(cfqq->cfqd, cfqq, crq);
457 static void cfq_resort_rr_list(struct cfq_queue *cfqq, int preempted)
459 struct cfq_data *cfqd = cfqq->cfqd;
460 struct list_head *list, *entry;
462 BUG_ON(!cfq_cfqq_on_rr(cfqq));
464 list_del(&cfqq->cfq_list);
466 if (cfq_class_rt(cfqq))
467 list = &cfqd->cur_rr;
468 else if (cfq_class_idle(cfqq))
469 list = &cfqd->idle_rr;
472 * if cfqq has requests in flight, don't allow it to be
473 * found in cfq_set_active_queue before it has finished them.
474 * this is done to increase fairness between a process that
475 * has lots of io pending vs one that only generates one
476 * sporadically or synchronously
478 if (cfq_cfqq_dispatched(cfqq))
479 list = &cfqd->busy_rr;
481 list = &cfqd->rr_list[cfqq->ioprio];
485 * if queue was preempted, just add to front to be fair. busy_rr
488 if (preempted || list == &cfqd->busy_rr) {
489 list_add(&cfqq->cfq_list, list);
494 * sort by when queue was last serviced
497 while ((entry = entry->prev) != list) {
498 struct cfq_queue *__cfqq = list_entry_cfqq(entry);
500 if (!__cfqq->service_last)
502 if (time_before(__cfqq->service_last, cfqq->service_last))
506 list_add(&cfqq->cfq_list, entry);
510 * add to busy list of queues for service, trying to be fair in ordering
511 * the pending list according to last request service
514 cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
516 BUG_ON(cfq_cfqq_on_rr(cfqq));
517 cfq_mark_cfqq_on_rr(cfqq);
520 cfq_resort_rr_list(cfqq, 0);
524 cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
526 BUG_ON(!cfq_cfqq_on_rr(cfqq));
527 cfq_clear_cfqq_on_rr(cfqq);
528 list_move(&cfqq->cfq_list, &cfqd->empty_list);
530 BUG_ON(!cfqd->busy_queues);
535 * rb tree support functions
537 static inline void cfq_del_crq_rb(struct cfq_rq *crq)
539 struct cfq_queue *cfqq = crq->cfq_queue;
540 struct cfq_data *cfqd = cfqq->cfqd;
541 const int sync = cfq_crq_is_sync(crq);
543 BUG_ON(!cfqq->queued[sync]);
544 cfqq->queued[sync]--;
546 cfq_update_next_crq(crq);
548 rb_erase(&crq->rb_node, &cfqq->sort_list);
549 RB_CLEAR_COLOR(&crq->rb_node);
551 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY(&cfqq->sort_list))
552 cfq_del_cfqq_rr(cfqd, cfqq);
555 static struct cfq_rq *
556 __cfq_add_crq_rb(struct cfq_rq *crq)
558 struct rb_node **p = &crq->cfq_queue->sort_list.rb_node;
559 struct rb_node *parent = NULL;
560 struct cfq_rq *__crq;
564 __crq = rb_entry_crq(parent);
566 if (crq->rb_key < __crq->rb_key)
568 else if (crq->rb_key > __crq->rb_key)
574 rb_link_node(&crq->rb_node, parent, p);
578 static void cfq_add_crq_rb(struct cfq_rq *crq)
580 struct cfq_queue *cfqq = crq->cfq_queue;
581 struct cfq_data *cfqd = cfqq->cfqd;
582 struct request *rq = crq->request;
583 struct cfq_rq *__alias;
585 crq->rb_key = rq_rb_key(rq);
586 cfqq->queued[cfq_crq_is_sync(crq)]++;
589 * looks a little odd, but the first insert might return an alias.
590 * if that happens, put the alias on the dispatch list
592 while ((__alias = __cfq_add_crq_rb(crq)) != NULL)
593 cfq_dispatch_insert(cfqd->queue, __alias);
595 rb_insert_color(&crq->rb_node, &cfqq->sort_list);
597 if (!cfq_cfqq_on_rr(cfqq))
598 cfq_add_cfqq_rr(cfqd, cfqq);
601 * check if this request is a better next-serve candidate
603 cfqq->next_crq = cfq_choose_req(cfqd, cfqq->next_crq, crq);
607 cfq_reposition_crq_rb(struct cfq_queue *cfqq, struct cfq_rq *crq)
609 rb_erase(&crq->rb_node, &cfqq->sort_list);
610 cfqq->queued[cfq_crq_is_sync(crq)]--;
615 static struct request *cfq_find_rq_rb(struct cfq_data *cfqd, sector_t sector)
618 struct cfq_queue *cfqq = cfq_find_cfq_hash(cfqd, current->pid, CFQ_KEY_ANY);
624 n = cfqq->sort_list.rb_node;
626 struct cfq_rq *crq = rb_entry_crq(n);
628 if (sector < crq->rb_key)
630 else if (sector > crq->rb_key)
640 static void cfq_activate_request(request_queue_t *q, struct request *rq)
642 struct cfq_data *cfqd = q->elevator->elevator_data;
644 cfqd->rq_in_driver++;
647 static void cfq_deactivate_request(request_queue_t *q, struct request *rq)
649 struct cfq_data *cfqd = q->elevator->elevator_data;
651 WARN_ON(!cfqd->rq_in_driver);
652 cfqd->rq_in_driver--;
655 static void cfq_remove_request(struct request *rq)
657 struct cfq_rq *crq = RQ_DATA(rq);
659 list_del_init(&rq->queuelist);
661 cfq_del_crq_hash(crq);
665 cfq_merge(request_queue_t *q, struct request **req, struct bio *bio)
667 struct cfq_data *cfqd = q->elevator->elevator_data;
668 struct request *__rq;
671 __rq = cfq_find_rq_hash(cfqd, bio->bi_sector);
672 if (__rq && elv_rq_merge_ok(__rq, bio)) {
673 ret = ELEVATOR_BACK_MERGE;
677 __rq = cfq_find_rq_rb(cfqd, bio->bi_sector + bio_sectors(bio));
678 if (__rq && elv_rq_merge_ok(__rq, bio)) {
679 ret = ELEVATOR_FRONT_MERGE;
683 return ELEVATOR_NO_MERGE;
689 static void cfq_merged_request(request_queue_t *q, struct request *req)
691 struct cfq_data *cfqd = q->elevator->elevator_data;
692 struct cfq_rq *crq = RQ_DATA(req);
694 cfq_del_crq_hash(crq);
695 cfq_add_crq_hash(cfqd, crq);
697 if (rq_rb_key(req) != crq->rb_key) {
698 struct cfq_queue *cfqq = crq->cfq_queue;
700 cfq_update_next_crq(crq);
701 cfq_reposition_crq_rb(cfqq, crq);
706 cfq_merged_requests(request_queue_t *q, struct request *rq,
707 struct request *next)
709 cfq_merged_request(q, rq);
712 * reposition in fifo if next is older than rq
714 if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
715 time_before(next->start_time, rq->start_time))
716 list_move(&rq->queuelist, &next->queuelist);
718 cfq_remove_request(next);
722 __cfq_set_active_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
726 * stop potential idle class queues waiting service
728 del_timer(&cfqd->idle_class_timer);
730 cfqq->slice_start = jiffies;
732 cfqq->slice_left = 0;
733 cfq_clear_cfqq_must_alloc_slice(cfqq);
734 cfq_clear_cfqq_fifo_expire(cfqq);
737 cfqd->active_queue = cfqq;
741 * current cfqq expired its slice (or was too idle), select new one
744 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
747 unsigned long now = jiffies;
749 if (cfq_cfqq_wait_request(cfqq))
750 del_timer(&cfqd->idle_slice_timer);
752 if (!preempted && !cfq_cfqq_dispatched(cfqq)) {
753 cfqq->service_last = now;
754 cfq_schedule_dispatch(cfqd);
757 cfq_clear_cfqq_must_dispatch(cfqq);
758 cfq_clear_cfqq_wait_request(cfqq);
761 * store what was left of this slice, if the queue idled out
764 if (time_after(cfqq->slice_end, now))
765 cfqq->slice_left = cfqq->slice_end - now;
767 cfqq->slice_left = 0;
769 if (cfq_cfqq_on_rr(cfqq))
770 cfq_resort_rr_list(cfqq, preempted);
772 if (cfqq == cfqd->active_queue)
773 cfqd->active_queue = NULL;
775 if (cfqd->active_cic) {
776 put_io_context(cfqd->active_cic->ioc);
777 cfqd->active_cic = NULL;
780 cfqd->dispatch_slice = 0;
783 static inline void cfq_slice_expired(struct cfq_data *cfqd, int preempted)
785 struct cfq_queue *cfqq = cfqd->active_queue;
788 __cfq_slice_expired(cfqd, cfqq, preempted);
801 static int cfq_get_next_prio_level(struct cfq_data *cfqd)
810 for (p = cfqd->cur_prio; p <= cfqd->cur_end_prio; p++) {
811 if (!list_empty(&cfqd->rr_list[p])) {
820 if (++cfqd->cur_end_prio == CFQ_PRIO_LISTS) {
821 cfqd->cur_end_prio = 0;
828 if (unlikely(prio == -1))
831 BUG_ON(prio >= CFQ_PRIO_LISTS);
833 list_splice_init(&cfqd->rr_list[prio], &cfqd->cur_rr);
835 cfqd->cur_prio = prio + 1;
836 if (cfqd->cur_prio > cfqd->cur_end_prio) {
837 cfqd->cur_end_prio = cfqd->cur_prio;
840 if (cfqd->cur_end_prio == CFQ_PRIO_LISTS) {
842 cfqd->cur_end_prio = 0;
848 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd)
850 struct cfq_queue *cfqq = NULL;
853 * if current list is non-empty, grab first entry. if it is empty,
854 * get next prio level and grab first entry then if any are spliced
856 if (!list_empty(&cfqd->cur_rr) || cfq_get_next_prio_level(cfqd) != -1)
857 cfqq = list_entry_cfqq(cfqd->cur_rr.next);
860 * if we have idle queues and no rt or be queues had pending
861 * requests, either allow immediate service if the grace period
862 * has passed or arm the idle grace timer
864 if (!cfqq && !list_empty(&cfqd->idle_rr)) {
865 unsigned long end = cfqd->last_end_request + CFQ_IDLE_GRACE;
867 if (time_after_eq(jiffies, end))
868 cfqq = list_entry_cfqq(cfqd->idle_rr.next);
870 mod_timer(&cfqd->idle_class_timer, end);
873 __cfq_set_active_queue(cfqd, cfqq);
877 static int cfq_arm_slice_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq)
882 WARN_ON(!RB_EMPTY(&cfqq->sort_list));
883 WARN_ON(cfqq != cfqd->active_queue);
886 * idle is disabled, either manually or by past process history
888 if (!cfqd->cfq_slice_idle)
890 if (!cfq_cfqq_idle_window(cfqq))
893 * task has exited, don't wait
895 if (cfqd->active_cic && !cfqd->active_cic->ioc->task)
898 cfq_mark_cfqq_must_dispatch(cfqq);
899 cfq_mark_cfqq_wait_request(cfqq);
901 sl = min(cfqq->slice_end - 1, (unsigned long) cfqd->cfq_slice_idle);
902 mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
906 static void cfq_dispatch_insert(request_queue_t *q, struct cfq_rq *crq)
908 struct cfq_data *cfqd = q->elevator->elevator_data;
909 struct cfq_queue *cfqq = crq->cfq_queue;
911 cfqq->next_crq = cfq_find_next_crq(cfqd, cfqq, crq);
912 cfq_remove_request(crq->request);
913 cfqq->on_dispatch[cfq_crq_is_sync(crq)]++;
914 elv_dispatch_sort(q, crq->request);
918 * return expired entry, or NULL to just start from scratch in rbtree
920 static inline struct cfq_rq *cfq_check_fifo(struct cfq_queue *cfqq)
922 struct cfq_data *cfqd = cfqq->cfqd;
926 if (cfq_cfqq_fifo_expire(cfqq))
929 if (!list_empty(&cfqq->fifo)) {
930 int fifo = cfq_cfqq_class_sync(cfqq);
932 crq = RQ_DATA(list_entry_fifo(cfqq->fifo.next));
934 if (time_after(jiffies, rq->start_time + cfqd->cfq_fifo_expire[fifo])) {
935 cfq_mark_cfqq_fifo_expire(cfqq);
944 * Scale schedule slice based on io priority. Use the sync time slice only
945 * if a queue is marked sync and has sync io queued. A sync queue with async
946 * io only, should not get full sync slice length.
949 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
951 const int base_slice = cfqd->cfq_slice[cfq_cfqq_sync(cfqq)];
953 WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
955 return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - cfqq->ioprio));
959 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
961 cfqq->slice_end = cfq_prio_to_slice(cfqd, cfqq) + jiffies;
965 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
967 const int base_rq = cfqd->cfq_slice_async_rq;
969 WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
971 return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
975 * get next queue for service
977 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
979 unsigned long now = jiffies;
980 struct cfq_queue *cfqq;
982 cfqq = cfqd->active_queue;
989 if (!cfq_cfqq_must_dispatch(cfqq) && time_after(now, cfqq->slice_end))
993 * if queue has requests, dispatch one. if not, check if
994 * enough slice is left to wait for one
996 if (!RB_EMPTY(&cfqq->sort_list))
998 else if (cfq_cfqq_class_sync(cfqq) &&
999 time_before(now, cfqq->slice_end)) {
1000 if (cfq_arm_slice_timer(cfqd, cfqq))
1005 cfq_slice_expired(cfqd, 0);
1007 cfqq = cfq_set_active_queue(cfqd);
1013 __cfq_dispatch_requests(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1018 BUG_ON(RB_EMPTY(&cfqq->sort_list));
1024 * follow expired path, else get first next available
1026 if ((crq = cfq_check_fifo(cfqq)) == NULL)
1027 crq = cfqq->next_crq;
1030 * finally, insert request into driver dispatch list
1032 cfq_dispatch_insert(cfqd->queue, crq);
1034 cfqd->dispatch_slice++;
1037 if (!cfqd->active_cic) {
1038 atomic_inc(&crq->io_context->ioc->refcount);
1039 cfqd->active_cic = crq->io_context;
1042 if (RB_EMPTY(&cfqq->sort_list))
1045 } while (dispatched < max_dispatch);
1048 * if slice end isn't set yet, set it. if at least one request was
1049 * sync, use the sync time slice value
1051 if (!cfqq->slice_end)
1052 cfq_set_prio_slice(cfqd, cfqq);
1055 * expire an async queue immediately if it has used up its slice. idle
1056 * queue always expire after 1 dispatch round.
1058 if ((!cfq_cfqq_sync(cfqq) &&
1059 cfqd->dispatch_slice >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
1060 cfq_class_idle(cfqq))
1061 cfq_slice_expired(cfqd, 0);
1067 cfq_forced_dispatch_cfqqs(struct list_head *list)
1070 struct cfq_queue *cfqq, *next;
1073 list_for_each_entry_safe(cfqq, next, list, cfq_list) {
1074 while ((crq = cfqq->next_crq)) {
1075 cfq_dispatch_insert(cfqq->cfqd->queue, crq);
1078 BUG_ON(!list_empty(&cfqq->fifo));
1084 cfq_forced_dispatch(struct cfq_data *cfqd)
1086 int i, dispatched = 0;
1088 for (i = 0; i < CFQ_PRIO_LISTS; i++)
1089 dispatched += cfq_forced_dispatch_cfqqs(&cfqd->rr_list[i]);
1091 dispatched += cfq_forced_dispatch_cfqqs(&cfqd->busy_rr);
1092 dispatched += cfq_forced_dispatch_cfqqs(&cfqd->cur_rr);
1093 dispatched += cfq_forced_dispatch_cfqqs(&cfqd->idle_rr);
1095 cfq_slice_expired(cfqd, 0);
1097 BUG_ON(cfqd->busy_queues);
1103 cfq_dispatch_requests(request_queue_t *q, int force)
1105 struct cfq_data *cfqd = q->elevator->elevator_data;
1106 struct cfq_queue *cfqq;
1108 if (!cfqd->busy_queues)
1111 if (unlikely(force))
1112 return cfq_forced_dispatch(cfqd);
1114 cfqq = cfq_select_queue(cfqd);
1119 * if idle window is disabled, allow queue buildup
1121 if (!cfq_cfqq_idle_window(cfqq) &&
1122 cfqd->rq_in_driver >= cfqd->cfq_max_depth)
1125 cfq_clear_cfqq_must_dispatch(cfqq);
1126 cfq_clear_cfqq_wait_request(cfqq);
1127 del_timer(&cfqd->idle_slice_timer);
1129 max_dispatch = cfqd->cfq_quantum;
1130 if (cfq_class_idle(cfqq))
1133 return __cfq_dispatch_requests(cfqd, cfqq, max_dispatch);
1140 * task holds one reference to the queue, dropped when task exits. each crq
1141 * in-flight on this queue also holds a reference, dropped when crq is freed.
1143 * queue lock must be held here.
1145 static void cfq_put_queue(struct cfq_queue *cfqq)
1147 struct cfq_data *cfqd = cfqq->cfqd;
1149 BUG_ON(atomic_read(&cfqq->ref) <= 0);
1151 if (!atomic_dec_and_test(&cfqq->ref))
1154 BUG_ON(rb_first(&cfqq->sort_list));
1155 BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
1156 BUG_ON(cfq_cfqq_on_rr(cfqq));
1158 if (unlikely(cfqd->active_queue == cfqq))
1159 __cfq_slice_expired(cfqd, cfqq, 0);
1162 * it's on the empty list and still hashed
1164 list_del(&cfqq->cfq_list);
1165 hlist_del(&cfqq->cfq_hash);
1166 kmem_cache_free(cfq_pool, cfqq);
1169 static inline struct cfq_queue *
1170 __cfq_find_cfq_hash(struct cfq_data *cfqd, unsigned int key, unsigned int prio,
1173 struct hlist_head *hash_list = &cfqd->cfq_hash[hashval];
1174 struct hlist_node *entry, *next;
1176 hlist_for_each_safe(entry, next, hash_list) {
1177 struct cfq_queue *__cfqq = list_entry_qhash(entry);
1178 const unsigned short __p = IOPRIO_PRIO_VALUE(__cfqq->org_ioprio_class, __cfqq->org_ioprio);
1180 if (__cfqq->key == key && (__p == prio || prio == CFQ_KEY_ANY))
1187 static struct cfq_queue *
1188 cfq_find_cfq_hash(struct cfq_data *cfqd, unsigned int key, unsigned short prio)
1190 return __cfq_find_cfq_hash(cfqd, key, prio, hash_long(key, CFQ_QHASH_SHIFT));
1193 static void cfq_free_io_context(struct io_context *ioc)
1195 struct cfq_io_context *__cic;
1199 while ((n = rb_first(&ioc->cic_root)) != NULL) {
1200 __cic = rb_entry(n, struct cfq_io_context, rb_node);
1201 rb_erase(&__cic->rb_node, &ioc->cic_root);
1202 kmem_cache_free(cfq_ioc_pool, __cic);
1206 if (atomic_sub_and_test(freed, &ioc_count) && ioc_gone)
1210 static void cfq_trim(struct io_context *ioc)
1212 ioc->set_ioprio = NULL;
1213 cfq_free_io_context(ioc);
1217 * Called with interrupts disabled
1219 static void cfq_exit_single_io_context(struct cfq_io_context *cic)
1221 struct cfq_data *cfqd = cic->key;
1229 WARN_ON(!irqs_disabled());
1231 spin_lock(q->queue_lock);
1233 if (cic->cfqq[ASYNC]) {
1234 if (unlikely(cic->cfqq[ASYNC] == cfqd->active_queue))
1235 __cfq_slice_expired(cfqd, cic->cfqq[ASYNC], 0);
1236 cfq_put_queue(cic->cfqq[ASYNC]);
1237 cic->cfqq[ASYNC] = NULL;
1240 if (cic->cfqq[SYNC]) {
1241 if (unlikely(cic->cfqq[SYNC] == cfqd->active_queue))
1242 __cfq_slice_expired(cfqd, cic->cfqq[SYNC], 0);
1243 cfq_put_queue(cic->cfqq[SYNC]);
1244 cic->cfqq[SYNC] = NULL;
1248 list_del_init(&cic->queue_list);
1249 spin_unlock(q->queue_lock);
1252 static void cfq_exit_io_context(struct io_context *ioc)
1254 struct cfq_io_context *__cic;
1255 unsigned long flags;
1259 * put the reference this task is holding to the various queues
1261 read_lock_irqsave(&cfq_exit_lock, flags);
1263 n = rb_first(&ioc->cic_root);
1265 __cic = rb_entry(n, struct cfq_io_context, rb_node);
1267 cfq_exit_single_io_context(__cic);
1271 read_unlock_irqrestore(&cfq_exit_lock, flags);
1274 static struct cfq_io_context *
1275 cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
1277 struct cfq_io_context *cic = kmem_cache_alloc(cfq_ioc_pool, gfp_mask);
1280 RB_CLEAR(&cic->rb_node);
1282 cic->cfqq[ASYNC] = NULL;
1283 cic->cfqq[SYNC] = NULL;
1284 cic->last_end_request = jiffies;
1285 cic->ttime_total = 0;
1286 cic->ttime_samples = 0;
1287 cic->ttime_mean = 0;
1288 cic->dtor = cfq_free_io_context;
1289 cic->exit = cfq_exit_io_context;
1290 INIT_LIST_HEAD(&cic->queue_list);
1291 atomic_inc(&ioc_count);
1297 static void cfq_init_prio_data(struct cfq_queue *cfqq)
1299 struct task_struct *tsk = current;
1302 if (!cfq_cfqq_prio_changed(cfqq))
1305 ioprio_class = IOPRIO_PRIO_CLASS(tsk->ioprio);
1306 switch (ioprio_class) {
1308 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
1309 case IOPRIO_CLASS_NONE:
1311 * no prio set, place us in the middle of the BE classes
1313 cfqq->ioprio = task_nice_ioprio(tsk);
1314 cfqq->ioprio_class = IOPRIO_CLASS_BE;
1316 case IOPRIO_CLASS_RT:
1317 cfqq->ioprio = task_ioprio(tsk);
1318 cfqq->ioprio_class = IOPRIO_CLASS_RT;
1320 case IOPRIO_CLASS_BE:
1321 cfqq->ioprio = task_ioprio(tsk);
1322 cfqq->ioprio_class = IOPRIO_CLASS_BE;
1324 case IOPRIO_CLASS_IDLE:
1325 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
1327 cfq_clear_cfqq_idle_window(cfqq);
1332 * keep track of original prio settings in case we have to temporarily
1333 * elevate the priority of this queue
1335 cfqq->org_ioprio = cfqq->ioprio;
1336 cfqq->org_ioprio_class = cfqq->ioprio_class;
1338 if (cfq_cfqq_on_rr(cfqq))
1339 cfq_resort_rr_list(cfqq, 0);
1341 cfq_clear_cfqq_prio_changed(cfqq);
1344 static inline void changed_ioprio(struct cfq_io_context *cic)
1346 struct cfq_data *cfqd = cic->key;
1347 struct cfq_queue *cfqq;
1349 spin_lock(cfqd->queue->queue_lock);
1350 cfqq = cic->cfqq[ASYNC];
1352 struct cfq_queue *new_cfqq;
1353 new_cfqq = cfq_get_queue(cfqd, CFQ_KEY_ASYNC,
1354 cic->ioc->task, GFP_ATOMIC);
1356 cic->cfqq[ASYNC] = new_cfqq;
1357 cfq_put_queue(cfqq);
1360 cfqq = cic->cfqq[SYNC];
1362 cfq_mark_cfqq_prio_changed(cfqq);
1363 cfq_init_prio_data(cfqq);
1365 spin_unlock(cfqd->queue->queue_lock);
1370 * callback from sys_ioprio_set, irqs are disabled
1372 static int cfq_ioc_set_ioprio(struct io_context *ioc, unsigned int ioprio)
1374 struct cfq_io_context *cic;
1377 write_lock(&cfq_exit_lock);
1379 n = rb_first(&ioc->cic_root);
1381 cic = rb_entry(n, struct cfq_io_context, rb_node);
1383 changed_ioprio(cic);
1387 write_unlock(&cfq_exit_lock);
1392 static struct cfq_queue *
1393 cfq_get_queue(struct cfq_data *cfqd, unsigned int key, struct task_struct *tsk,
1396 const int hashval = hash_long(key, CFQ_QHASH_SHIFT);
1397 struct cfq_queue *cfqq, *new_cfqq = NULL;
1398 unsigned short ioprio;
1401 ioprio = tsk->ioprio;
1402 cfqq = __cfq_find_cfq_hash(cfqd, key, ioprio, hashval);
1408 } else if (gfp_mask & __GFP_WAIT) {
1409 spin_unlock_irq(cfqd->queue->queue_lock);
1410 new_cfqq = kmem_cache_alloc(cfq_pool, gfp_mask);
1411 spin_lock_irq(cfqd->queue->queue_lock);
1414 cfqq = kmem_cache_alloc(cfq_pool, gfp_mask);
1419 memset(cfqq, 0, sizeof(*cfqq));
1421 INIT_HLIST_NODE(&cfqq->cfq_hash);
1422 INIT_LIST_HEAD(&cfqq->cfq_list);
1423 RB_CLEAR_ROOT(&cfqq->sort_list);
1424 INIT_LIST_HEAD(&cfqq->fifo);
1427 hlist_add_head(&cfqq->cfq_hash, &cfqd->cfq_hash[hashval]);
1428 atomic_set(&cfqq->ref, 0);
1430 cfqq->service_last = 0;
1432 * set ->slice_left to allow preemption for a new process
1434 cfqq->slice_left = 2 * cfqd->cfq_slice_idle;
1435 cfq_mark_cfqq_idle_window(cfqq);
1436 cfq_mark_cfqq_prio_changed(cfqq);
1437 cfq_init_prio_data(cfqq);
1441 kmem_cache_free(cfq_pool, new_cfqq);
1443 atomic_inc(&cfqq->ref);
1445 WARN_ON((gfp_mask & __GFP_WAIT) && !cfqq);
1449 static struct cfq_io_context *
1450 cfq_cic_rb_lookup(struct cfq_data *cfqd, struct io_context *ioc)
1452 struct rb_node *n = ioc->cic_root.rb_node;
1453 struct cfq_io_context *cic;
1457 cic = rb_entry(n, struct cfq_io_context, rb_node);
1461 else if (key > cic->key)
1471 cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
1472 struct cfq_io_context *cic)
1474 struct rb_node **p = &ioc->cic_root.rb_node;
1475 struct rb_node *parent = NULL;
1476 struct cfq_io_context *__cic;
1478 read_lock(&cfq_exit_lock);
1483 ioc->set_ioprio = cfq_ioc_set_ioprio;
1487 __cic = rb_entry(parent, struct cfq_io_context, rb_node);
1489 if (cic->key < __cic->key)
1491 else if (cic->key > __cic->key)
1492 p = &(*p)->rb_right;
1497 rb_link_node(&cic->rb_node, parent, p);
1498 rb_insert_color(&cic->rb_node, &ioc->cic_root);
1499 list_add(&cic->queue_list, &cfqd->cic_list);
1500 read_unlock(&cfq_exit_lock);
1504 * Setup general io context and cfq io context. There can be several cfq
1505 * io contexts per general io context, if this process is doing io to more
1506 * than one device managed by cfq.
1508 static struct cfq_io_context *
1509 cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
1511 struct io_context *ioc = NULL;
1512 struct cfq_io_context *cic;
1514 might_sleep_if(gfp_mask & __GFP_WAIT);
1516 ioc = get_io_context(gfp_mask);
1520 cic = cfq_cic_rb_lookup(cfqd, ioc);
1524 cic = cfq_alloc_io_context(cfqd, gfp_mask);
1528 cfq_cic_link(cfqd, ioc, cic);
1532 put_io_context(ioc);
1537 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
1539 unsigned long elapsed, ttime;
1542 * if this context already has stuff queued, thinktime is from
1543 * last queue not last end
1546 if (time_after(cic->last_end_request, cic->last_queue))
1547 elapsed = jiffies - cic->last_end_request;
1549 elapsed = jiffies - cic->last_queue;
1551 elapsed = jiffies - cic->last_end_request;
1554 ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
1556 cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
1557 cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
1558 cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
1561 #define sample_valid(samples) ((samples) > 80)
1564 * Disable idle window if the process thinks too long or seeks so much that
1568 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1569 struct cfq_io_context *cic)
1571 int enable_idle = cfq_cfqq_idle_window(cfqq);
1573 if (!cic->ioc->task || !cfqd->cfq_slice_idle)
1575 else if (sample_valid(cic->ttime_samples)) {
1576 if (cic->ttime_mean > cfqd->cfq_slice_idle)
1583 cfq_mark_cfqq_idle_window(cfqq);
1585 cfq_clear_cfqq_idle_window(cfqq);
1590 * Check if new_cfqq should preempt the currently active queue. Return 0 for
1591 * no or if we aren't sure, a 1 will cause a preempt.
1594 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
1597 struct cfq_queue *cfqq = cfqd->active_queue;
1599 if (cfq_class_idle(new_cfqq))
1605 if (cfq_class_idle(cfqq))
1607 if (!cfq_cfqq_wait_request(new_cfqq))
1610 * if it doesn't have slice left, forget it
1612 if (new_cfqq->slice_left < cfqd->cfq_slice_idle)
1614 if (cfq_crq_is_sync(crq) && !cfq_cfqq_sync(cfqq))
1621 * cfqq preempts the active queue. if we allowed preempt with no slice left,
1622 * let it have half of its nominal slice.
1624 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1626 struct cfq_queue *__cfqq, *next;
1628 list_for_each_entry_safe(__cfqq, next, &cfqd->cur_rr, cfq_list)
1629 cfq_resort_rr_list(__cfqq, 1);
1631 if (!cfqq->slice_left)
1632 cfqq->slice_left = cfq_prio_to_slice(cfqd, cfqq) / 2;
1634 cfqq->slice_end = cfqq->slice_left + jiffies;
1635 __cfq_slice_expired(cfqd, cfqq, 1);
1636 __cfq_set_active_queue(cfqd, cfqq);
1640 * should really be a ll_rw_blk.c helper
1642 static void cfq_start_queueing(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1644 request_queue_t *q = cfqd->queue;
1646 if (!blk_queue_plugged(q))
1649 __generic_unplug_device(q);
1653 * Called when a new fs request (crq) is added (to cfqq). Check if there's
1654 * something we should do about it
1657 cfq_crq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1660 struct cfq_io_context *cic;
1662 cfqq->next_crq = cfq_choose_req(cfqd, cfqq->next_crq, crq);
1665 * we never wait for an async request and we don't allow preemption
1666 * of an async request. so just return early
1668 if (!cfq_crq_is_sync(crq))
1671 cic = crq->io_context;
1673 cfq_update_io_thinktime(cfqd, cic);
1674 cfq_update_idle_window(cfqd, cfqq, cic);
1676 cic->last_queue = jiffies;
1678 if (cfqq == cfqd->active_queue) {
1680 * if we are waiting for a request for this queue, let it rip
1681 * immediately and flag that we must not expire this queue
1684 if (cfq_cfqq_wait_request(cfqq)) {
1685 cfq_mark_cfqq_must_dispatch(cfqq);
1686 del_timer(&cfqd->idle_slice_timer);
1687 cfq_start_queueing(cfqd, cfqq);
1689 } else if (cfq_should_preempt(cfqd, cfqq, crq)) {
1691 * not the active queue - expire current slice if it is
1692 * idle and has expired it's mean thinktime or this new queue
1693 * has some old slice time left and is of higher priority
1695 cfq_preempt_queue(cfqd, cfqq);
1696 cfq_mark_cfqq_must_dispatch(cfqq);
1697 cfq_start_queueing(cfqd, cfqq);
1701 static void cfq_insert_request(request_queue_t *q, struct request *rq)
1703 struct cfq_data *cfqd = q->elevator->elevator_data;
1704 struct cfq_rq *crq = RQ_DATA(rq);
1705 struct cfq_queue *cfqq = crq->cfq_queue;
1707 cfq_init_prio_data(cfqq);
1709 cfq_add_crq_rb(crq);
1711 list_add_tail(&rq->queuelist, &cfqq->fifo);
1713 if (rq_mergeable(rq))
1714 cfq_add_crq_hash(cfqd, crq);
1716 cfq_crq_enqueued(cfqd, cfqq, crq);
1719 static void cfq_completed_request(request_queue_t *q, struct request *rq)
1721 struct cfq_rq *crq = RQ_DATA(rq);
1722 struct cfq_queue *cfqq = crq->cfq_queue;
1723 struct cfq_data *cfqd = cfqq->cfqd;
1724 const int sync = cfq_crq_is_sync(crq);
1729 WARN_ON(!cfqd->rq_in_driver);
1730 WARN_ON(!cfqq->on_dispatch[sync]);
1731 cfqd->rq_in_driver--;
1732 cfqq->on_dispatch[sync]--;
1734 if (!cfq_class_idle(cfqq))
1735 cfqd->last_end_request = now;
1737 if (!cfq_cfqq_dispatched(cfqq)) {
1738 if (cfq_cfqq_on_rr(cfqq)) {
1739 cfqq->service_last = now;
1740 cfq_resort_rr_list(cfqq, 0);
1742 cfq_schedule_dispatch(cfqd);
1745 if (cfq_crq_is_sync(crq))
1746 crq->io_context->last_end_request = now;
1749 static struct request *
1750 cfq_former_request(request_queue_t *q, struct request *rq)
1752 struct cfq_rq *crq = RQ_DATA(rq);
1753 struct rb_node *rbprev = rb_prev(&crq->rb_node);
1756 return rb_entry_crq(rbprev)->request;
1761 static struct request *
1762 cfq_latter_request(request_queue_t *q, struct request *rq)
1764 struct cfq_rq *crq = RQ_DATA(rq);
1765 struct rb_node *rbnext = rb_next(&crq->rb_node);
1768 return rb_entry_crq(rbnext)->request;
1774 * we temporarily boost lower priority queues if they are holding fs exclusive
1775 * resources. they are boosted to normal prio (CLASS_BE/4)
1777 static void cfq_prio_boost(struct cfq_queue *cfqq)
1779 const int ioprio_class = cfqq->ioprio_class;
1780 const int ioprio = cfqq->ioprio;
1782 if (has_fs_excl()) {
1784 * boost idle prio on transactions that would lock out other
1785 * users of the filesystem
1787 if (cfq_class_idle(cfqq))
1788 cfqq->ioprio_class = IOPRIO_CLASS_BE;
1789 if (cfqq->ioprio > IOPRIO_NORM)
1790 cfqq->ioprio = IOPRIO_NORM;
1793 * check if we need to unboost the queue
1795 if (cfqq->ioprio_class != cfqq->org_ioprio_class)
1796 cfqq->ioprio_class = cfqq->org_ioprio_class;
1797 if (cfqq->ioprio != cfqq->org_ioprio)
1798 cfqq->ioprio = cfqq->org_ioprio;
1802 * refile between round-robin lists if we moved the priority class
1804 if ((ioprio_class != cfqq->ioprio_class || ioprio != cfqq->ioprio) &&
1805 cfq_cfqq_on_rr(cfqq))
1806 cfq_resort_rr_list(cfqq, 0);
1809 static inline pid_t cfq_queue_pid(struct task_struct *task, int rw)
1811 if (rw == READ || process_sync(task))
1814 return CFQ_KEY_ASYNC;
1818 __cfq_may_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1819 struct task_struct *task, int rw)
1822 if ((cfq_cfqq_wait_request(cfqq) || cfq_cfqq_must_alloc(cfqq)) &&
1823 !cfq_cfqq_must_alloc_slice(cfqq)) {
1824 cfq_mark_cfqq_must_alloc_slice(cfqq);
1825 return ELV_MQUEUE_MUST;
1828 return ELV_MQUEUE_MAY;
1830 if (!cfqq || task->flags & PF_MEMALLOC)
1831 return ELV_MQUEUE_MAY;
1832 if (!cfqq->allocated[rw] || cfq_cfqq_must_alloc(cfqq)) {
1833 if (cfq_cfqq_wait_request(cfqq))
1834 return ELV_MQUEUE_MUST;
1837 * only allow 1 ELV_MQUEUE_MUST per slice, otherwise we
1838 * can quickly flood the queue with writes from a single task
1840 if (rw == READ || !cfq_cfqq_must_alloc_slice(cfqq)) {
1841 cfq_mark_cfqq_must_alloc_slice(cfqq);
1842 return ELV_MQUEUE_MUST;
1845 return ELV_MQUEUE_MAY;
1847 if (cfq_class_idle(cfqq))
1848 return ELV_MQUEUE_NO;
1849 if (cfqq->allocated[rw] >= cfqd->max_queued) {
1850 struct io_context *ioc = get_io_context(GFP_ATOMIC);
1851 int ret = ELV_MQUEUE_NO;
1853 if (ioc && ioc->nr_batch_requests)
1854 ret = ELV_MQUEUE_MAY;
1856 put_io_context(ioc);
1860 return ELV_MQUEUE_MAY;
1864 static int cfq_may_queue(request_queue_t *q, int rw, struct bio *bio)
1866 struct cfq_data *cfqd = q->elevator->elevator_data;
1867 struct task_struct *tsk = current;
1868 struct cfq_queue *cfqq;
1871 * don't force setup of a queue from here, as a call to may_queue
1872 * does not necessarily imply that a request actually will be queued.
1873 * so just lookup a possibly existing queue, or return 'may queue'
1876 cfqq = cfq_find_cfq_hash(cfqd, cfq_queue_pid(tsk, rw), tsk->ioprio);
1878 cfq_init_prio_data(cfqq);
1879 cfq_prio_boost(cfqq);
1881 return __cfq_may_queue(cfqd, cfqq, tsk, rw);
1884 return ELV_MQUEUE_MAY;
1887 static void cfq_check_waiters(request_queue_t *q, struct cfq_queue *cfqq)
1889 struct cfq_data *cfqd = q->elevator->elevator_data;
1890 struct request_list *rl = &q->rq;
1892 if (cfqq->allocated[READ] <= cfqd->max_queued || cfqd->rq_starved) {
1894 if (waitqueue_active(&rl->wait[READ]))
1895 wake_up(&rl->wait[READ]);
1898 if (cfqq->allocated[WRITE] <= cfqd->max_queued || cfqd->rq_starved) {
1900 if (waitqueue_active(&rl->wait[WRITE]))
1901 wake_up(&rl->wait[WRITE]);
1906 * queue lock held here
1908 static void cfq_put_request(request_queue_t *q, struct request *rq)
1910 struct cfq_data *cfqd = q->elevator->elevator_data;
1911 struct cfq_rq *crq = RQ_DATA(rq);
1914 struct cfq_queue *cfqq = crq->cfq_queue;
1915 const int rw = rq_data_dir(rq);
1917 BUG_ON(!cfqq->allocated[rw]);
1918 cfqq->allocated[rw]--;
1920 put_io_context(crq->io_context->ioc);
1922 mempool_free(crq, cfqd->crq_pool);
1923 rq->elevator_private = NULL;
1925 cfq_check_waiters(q, cfqq);
1926 cfq_put_queue(cfqq);
1931 * Allocate cfq data structures associated with this request.
1934 cfq_set_request(request_queue_t *q, struct request *rq, struct bio *bio,
1937 struct cfq_data *cfqd = q->elevator->elevator_data;
1938 struct task_struct *tsk = current;
1939 struct cfq_io_context *cic;
1940 const int rw = rq_data_dir(rq);
1941 pid_t key = cfq_queue_pid(tsk, rw);
1942 struct cfq_queue *cfqq;
1944 unsigned long flags;
1945 int is_sync = key != CFQ_KEY_ASYNC;
1947 might_sleep_if(gfp_mask & __GFP_WAIT);
1949 cic = cfq_get_io_context(cfqd, gfp_mask);
1951 spin_lock_irqsave(q->queue_lock, flags);
1956 if (!cic->cfqq[is_sync]) {
1957 cfqq = cfq_get_queue(cfqd, key, tsk, gfp_mask);
1961 cic->cfqq[is_sync] = cfqq;
1963 cfqq = cic->cfqq[is_sync];
1965 cfqq->allocated[rw]++;
1966 cfq_clear_cfqq_must_alloc(cfqq);
1967 cfqd->rq_starved = 0;
1968 atomic_inc(&cfqq->ref);
1969 spin_unlock_irqrestore(q->queue_lock, flags);
1971 crq = mempool_alloc(cfqd->crq_pool, gfp_mask);
1973 RB_CLEAR(&crq->rb_node);
1976 INIT_HLIST_NODE(&crq->hash);
1977 crq->cfq_queue = cfqq;
1978 crq->io_context = cic;
1981 cfq_mark_crq_is_sync(crq);
1983 cfq_clear_crq_is_sync(crq);
1985 rq->elevator_private = crq;
1989 spin_lock_irqsave(q->queue_lock, flags);
1990 cfqq->allocated[rw]--;
1991 if (!(cfqq->allocated[0] + cfqq->allocated[1]))
1992 cfq_mark_cfqq_must_alloc(cfqq);
1993 cfq_put_queue(cfqq);
1996 put_io_context(cic->ioc);
1998 * mark us rq allocation starved. we need to kickstart the process
1999 * ourselves if there are no pending requests that can do it for us.
2000 * that would be an extremely rare OOM situation
2002 cfqd->rq_starved = 1;
2003 cfq_schedule_dispatch(cfqd);
2004 spin_unlock_irqrestore(q->queue_lock, flags);
2008 static void cfq_kick_queue(void *data)
2010 request_queue_t *q = data;
2011 struct cfq_data *cfqd = q->elevator->elevator_data;
2012 unsigned long flags;
2014 spin_lock_irqsave(q->queue_lock, flags);
2016 if (cfqd->rq_starved) {
2017 struct request_list *rl = &q->rq;
2020 * we aren't guaranteed to get a request after this, but we
2021 * have to be opportunistic
2024 if (waitqueue_active(&rl->wait[READ]))
2025 wake_up(&rl->wait[READ]);
2026 if (waitqueue_active(&rl->wait[WRITE]))
2027 wake_up(&rl->wait[WRITE]);
2032 spin_unlock_irqrestore(q->queue_lock, flags);
2036 * Timer running if the active_queue is currently idling inside its time slice
2038 static void cfq_idle_slice_timer(unsigned long data)
2040 struct cfq_data *cfqd = (struct cfq_data *) data;
2041 struct cfq_queue *cfqq;
2042 unsigned long flags;
2044 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2046 if ((cfqq = cfqd->active_queue) != NULL) {
2047 unsigned long now = jiffies;
2052 if (time_after(now, cfqq->slice_end))
2056 * only expire and reinvoke request handler, if there are
2057 * other queues with pending requests
2059 if (!cfqd->busy_queues) {
2060 cfqd->idle_slice_timer.expires = min(now + cfqd->cfq_slice_idle, cfqq->slice_end);
2061 add_timer(&cfqd->idle_slice_timer);
2066 * not expired and it has a request pending, let it dispatch
2068 if (!RB_EMPTY(&cfqq->sort_list)) {
2069 cfq_mark_cfqq_must_dispatch(cfqq);
2074 cfq_slice_expired(cfqd, 0);
2076 cfq_schedule_dispatch(cfqd);
2078 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2082 * Timer running if an idle class queue is waiting for service
2084 static void cfq_idle_class_timer(unsigned long data)
2086 struct cfq_data *cfqd = (struct cfq_data *) data;
2087 unsigned long flags, end;
2089 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2092 * race with a non-idle queue, reset timer
2094 end = cfqd->last_end_request + CFQ_IDLE_GRACE;
2095 if (!time_after_eq(jiffies, end)) {
2096 cfqd->idle_class_timer.expires = end;
2097 add_timer(&cfqd->idle_class_timer);
2099 cfq_schedule_dispatch(cfqd);
2101 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2104 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
2106 del_timer_sync(&cfqd->idle_slice_timer);
2107 del_timer_sync(&cfqd->idle_class_timer);
2108 blk_sync_queue(cfqd->queue);
2111 static void cfq_exit_queue(elevator_t *e)
2113 struct cfq_data *cfqd = e->elevator_data;
2114 request_queue_t *q = cfqd->queue;
2116 cfq_shutdown_timer_wq(cfqd);
2118 write_lock(&cfq_exit_lock);
2119 spin_lock_irq(q->queue_lock);
2121 if (cfqd->active_queue)
2122 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
2124 while (!list_empty(&cfqd->cic_list)) {
2125 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
2126 struct cfq_io_context,
2128 if (cic->cfqq[ASYNC]) {
2129 cfq_put_queue(cic->cfqq[ASYNC]);
2130 cic->cfqq[ASYNC] = NULL;
2132 if (cic->cfqq[SYNC]) {
2133 cfq_put_queue(cic->cfqq[SYNC]);
2134 cic->cfqq[SYNC] = NULL;
2137 list_del_init(&cic->queue_list);
2140 spin_unlock_irq(q->queue_lock);
2141 write_unlock(&cfq_exit_lock);
2143 cfq_shutdown_timer_wq(cfqd);
2145 mempool_destroy(cfqd->crq_pool);
2146 kfree(cfqd->crq_hash);
2147 kfree(cfqd->cfq_hash);
2151 static int cfq_init_queue(request_queue_t *q, elevator_t *e)
2153 struct cfq_data *cfqd;
2156 cfqd = kmalloc(sizeof(*cfqd), GFP_KERNEL);
2160 memset(cfqd, 0, sizeof(*cfqd));
2162 for (i = 0; i < CFQ_PRIO_LISTS; i++)
2163 INIT_LIST_HEAD(&cfqd->rr_list[i]);
2165 INIT_LIST_HEAD(&cfqd->busy_rr);
2166 INIT_LIST_HEAD(&cfqd->cur_rr);
2167 INIT_LIST_HEAD(&cfqd->idle_rr);
2168 INIT_LIST_HEAD(&cfqd->empty_list);
2169 INIT_LIST_HEAD(&cfqd->cic_list);
2171 cfqd->crq_hash = kmalloc(sizeof(struct hlist_head) * CFQ_MHASH_ENTRIES, GFP_KERNEL);
2172 if (!cfqd->crq_hash)
2175 cfqd->cfq_hash = kmalloc(sizeof(struct hlist_head) * CFQ_QHASH_ENTRIES, GFP_KERNEL);
2176 if (!cfqd->cfq_hash)
2179 cfqd->crq_pool = mempool_create_slab_pool(BLKDEV_MIN_RQ, crq_pool);
2180 if (!cfqd->crq_pool)
2183 for (i = 0; i < CFQ_MHASH_ENTRIES; i++)
2184 INIT_HLIST_HEAD(&cfqd->crq_hash[i]);
2185 for (i = 0; i < CFQ_QHASH_ENTRIES; i++)
2186 INIT_HLIST_HEAD(&cfqd->cfq_hash[i]);
2188 e->elevator_data = cfqd;
2192 cfqd->max_queued = q->nr_requests / 4;
2193 q->nr_batching = cfq_queued;
2195 init_timer(&cfqd->idle_slice_timer);
2196 cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
2197 cfqd->idle_slice_timer.data = (unsigned long) cfqd;
2199 init_timer(&cfqd->idle_class_timer);
2200 cfqd->idle_class_timer.function = cfq_idle_class_timer;
2201 cfqd->idle_class_timer.data = (unsigned long) cfqd;
2203 INIT_WORK(&cfqd->unplug_work, cfq_kick_queue, q);
2205 cfqd->cfq_queued = cfq_queued;
2206 cfqd->cfq_quantum = cfq_quantum;
2207 cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
2208 cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
2209 cfqd->cfq_back_max = cfq_back_max;
2210 cfqd->cfq_back_penalty = cfq_back_penalty;
2211 cfqd->cfq_slice[0] = cfq_slice_async;
2212 cfqd->cfq_slice[1] = cfq_slice_sync;
2213 cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
2214 cfqd->cfq_slice_idle = cfq_slice_idle;
2215 cfqd->cfq_max_depth = cfq_max_depth;
2219 kfree(cfqd->cfq_hash);
2221 kfree(cfqd->crq_hash);
2227 static void cfq_slab_kill(void)
2230 kmem_cache_destroy(crq_pool);
2232 kmem_cache_destroy(cfq_pool);
2234 kmem_cache_destroy(cfq_ioc_pool);
2237 static int __init cfq_slab_setup(void)
2239 crq_pool = kmem_cache_create("crq_pool", sizeof(struct cfq_rq), 0, 0,
2244 cfq_pool = kmem_cache_create("cfq_pool", sizeof(struct cfq_queue), 0, 0,
2249 cfq_ioc_pool = kmem_cache_create("cfq_ioc_pool",
2250 sizeof(struct cfq_io_context), 0, 0, NULL, NULL);
2261 * sysfs parts below -->
2265 cfq_var_show(unsigned int var, char *page)
2267 return sprintf(page, "%d\n", var);
2271 cfq_var_store(unsigned int *var, const char *page, size_t count)
2273 char *p = (char *) page;
2275 *var = simple_strtoul(p, &p, 10);
2279 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
2280 static ssize_t __FUNC(elevator_t *e, char *page) \
2282 struct cfq_data *cfqd = e->elevator_data; \
2283 unsigned int __data = __VAR; \
2285 __data = jiffies_to_msecs(__data); \
2286 return cfq_var_show(__data, (page)); \
2288 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
2289 SHOW_FUNCTION(cfq_queued_show, cfqd->cfq_queued, 0);
2290 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
2291 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
2292 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
2293 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
2294 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
2295 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
2296 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
2297 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
2298 SHOW_FUNCTION(cfq_max_depth_show, cfqd->cfq_max_depth, 0);
2299 #undef SHOW_FUNCTION
2301 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
2302 static ssize_t __FUNC(elevator_t *e, const char *page, size_t count) \
2304 struct cfq_data *cfqd = e->elevator_data; \
2305 unsigned int __data; \
2306 int ret = cfq_var_store(&__data, (page), count); \
2307 if (__data < (MIN)) \
2309 else if (__data > (MAX)) \
2312 *(__PTR) = msecs_to_jiffies(__data); \
2314 *(__PTR) = __data; \
2317 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
2318 STORE_FUNCTION(cfq_queued_store, &cfqd->cfq_queued, 1, UINT_MAX, 0);
2319 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1, UINT_MAX, 1);
2320 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1, UINT_MAX, 1);
2321 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
2322 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1, UINT_MAX, 0);
2323 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
2324 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
2325 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
2326 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1, UINT_MAX, 0);
2327 STORE_FUNCTION(cfq_max_depth_store, &cfqd->cfq_max_depth, 1, UINT_MAX, 0);
2328 #undef STORE_FUNCTION
2330 #define CFQ_ATTR(name) \
2331 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
2333 static struct elv_fs_entry cfq_attrs[] = {
2336 CFQ_ATTR(fifo_expire_sync),
2337 CFQ_ATTR(fifo_expire_async),
2338 CFQ_ATTR(back_seek_max),
2339 CFQ_ATTR(back_seek_penalty),
2340 CFQ_ATTR(slice_sync),
2341 CFQ_ATTR(slice_async),
2342 CFQ_ATTR(slice_async_rq),
2343 CFQ_ATTR(slice_idle),
2344 CFQ_ATTR(max_depth),
2348 static struct elevator_type iosched_cfq = {
2350 .elevator_merge_fn = cfq_merge,
2351 .elevator_merged_fn = cfq_merged_request,
2352 .elevator_merge_req_fn = cfq_merged_requests,
2353 .elevator_dispatch_fn = cfq_dispatch_requests,
2354 .elevator_add_req_fn = cfq_insert_request,
2355 .elevator_activate_req_fn = cfq_activate_request,
2356 .elevator_deactivate_req_fn = cfq_deactivate_request,
2357 .elevator_queue_empty_fn = cfq_queue_empty,
2358 .elevator_completed_req_fn = cfq_completed_request,
2359 .elevator_former_req_fn = cfq_former_request,
2360 .elevator_latter_req_fn = cfq_latter_request,
2361 .elevator_set_req_fn = cfq_set_request,
2362 .elevator_put_req_fn = cfq_put_request,
2363 .elevator_may_queue_fn = cfq_may_queue,
2364 .elevator_init_fn = cfq_init_queue,
2365 .elevator_exit_fn = cfq_exit_queue,
2368 .elevator_attrs = cfq_attrs,
2369 .elevator_name = "cfq",
2370 .elevator_owner = THIS_MODULE,
2373 static int __init cfq_init(void)
2378 * could be 0 on HZ < 1000 setups
2380 if (!cfq_slice_async)
2381 cfq_slice_async = 1;
2382 if (!cfq_slice_idle)
2385 if (cfq_slab_setup())
2388 ret = elv_register(&iosched_cfq);
2395 static void __exit cfq_exit(void)
2397 DECLARE_COMPLETION(all_gone);
2398 elv_unregister(&iosched_cfq);
2399 ioc_gone = &all_gone;
2401 if (atomic_read(&ioc_count))
2407 module_init(cfq_init);
2408 module_exit(cfq_exit);
2410 MODULE_AUTHOR("Jens Axboe");
2411 MODULE_LICENSE("GPL");
2412 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");