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/module.h>
10 #include <linux/blkdev.h>
11 #include <linux/elevator.h>
12 #include <linux/hash.h>
13 #include <linux/rbtree.h>
14 #include <linux/ioprio.h>
19 static const int cfq_quantum = 4; /* max queue in one round of service */
20 static const int cfq_queued = 8; /* minimum rq allocate limit per-queue*/
21 static const int cfq_fifo_expire[2] = { HZ / 4, HZ / 8 };
22 static const int cfq_back_max = 16 * 1024; /* maximum backwards seek, in KiB */
23 static const int cfq_back_penalty = 2; /* penalty of a backwards seek */
25 static const int cfq_slice_sync = HZ / 10;
26 static int cfq_slice_async = HZ / 25;
27 static const int cfq_slice_async_rq = 2;
28 static int cfq_slice_idle = HZ / 125;
30 #define CFQ_IDLE_GRACE (HZ / 10)
31 #define CFQ_SLICE_SCALE (5)
33 #define CFQ_KEY_ASYNC (0)
35 static DEFINE_SPINLOCK(cfq_exit_lock);
38 * for the hash of cfqq inside the cfqd
40 #define CFQ_QHASH_SHIFT 6
41 #define CFQ_QHASH_ENTRIES (1 << CFQ_QHASH_SHIFT)
42 #define list_entry_qhash(entry) hlist_entry((entry), struct cfq_queue, cfq_hash)
44 #define list_entry_cfqq(ptr) list_entry((ptr), struct cfq_queue, cfq_list)
46 #define RQ_DATA(rq) (rq)->elevator_private
48 static kmem_cache_t *crq_pool;
49 static kmem_cache_t *cfq_pool;
50 static kmem_cache_t *cfq_ioc_pool;
52 static atomic_t ioc_count = ATOMIC_INIT(0);
53 static struct completion *ioc_gone;
55 #define CFQ_PRIO_LISTS IOPRIO_BE_NR
56 #define cfq_class_idle(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
57 #define cfq_class_be(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_BE)
58 #define cfq_class_rt(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
63 #define cfq_cfqq_dispatched(cfqq) \
64 ((cfqq)->on_dispatch[ASYNC] + (cfqq)->on_dispatch[SYNC])
66 #define cfq_cfqq_class_sync(cfqq) ((cfqq)->key != CFQ_KEY_ASYNC)
68 #define cfq_cfqq_sync(cfqq) \
69 (cfq_cfqq_class_sync(cfqq) || (cfqq)->on_dispatch[SYNC])
71 #define sample_valid(samples) ((samples) > 80)
74 * Per block device queue structure
77 request_queue_t *queue;
80 * rr list of queues with requests and the count of them
82 struct list_head rr_list[CFQ_PRIO_LISTS];
83 struct list_head busy_rr;
84 struct list_head cur_rr;
85 struct list_head idle_rr;
86 unsigned int busy_queues;
89 * non-ordered list of empty cfqq's
91 struct list_head empty_list;
96 struct hlist_head *cfq_hash;
104 * schedule slice state info
107 * idle window management
109 struct timer_list idle_slice_timer;
110 struct work_struct unplug_work;
112 struct cfq_queue *active_queue;
113 struct cfq_io_context *active_cic;
114 int cur_prio, cur_end_prio;
115 unsigned int dispatch_slice;
117 struct timer_list idle_class_timer;
119 sector_t last_sector;
120 unsigned long last_end_request;
122 unsigned int rq_starved;
125 * tunables, see top of file
127 unsigned int cfq_quantum;
128 unsigned int cfq_queued;
129 unsigned int cfq_fifo_expire[2];
130 unsigned int cfq_back_penalty;
131 unsigned int cfq_back_max;
132 unsigned int cfq_slice[2];
133 unsigned int cfq_slice_async_rq;
134 unsigned int cfq_slice_idle;
136 struct list_head cic_list;
140 * Per process-grouping structure
143 /* reference count */
145 /* parent cfq_data */
146 struct cfq_data *cfqd;
147 /* cfqq lookup hash */
148 struct hlist_node cfq_hash;
151 /* on either rr or empty list of cfqd */
152 struct list_head cfq_list;
153 /* sorted list of pending requests */
154 struct rb_root sort_list;
155 /* if fifo isn't expired, next request to serve */
156 struct cfq_rq *next_crq;
157 /* requests queued in sort_list */
159 /* currently allocated requests */
161 /* fifo list of requests in sort_list */
162 struct list_head fifo;
164 unsigned long slice_start;
165 unsigned long slice_end;
166 unsigned long slice_left;
167 unsigned long service_last;
169 /* number of requests that are on the dispatch list */
172 /* io prio of this group */
173 unsigned short ioprio, org_ioprio;
174 unsigned short ioprio_class, org_ioprio_class;
176 /* various state flags, see below */
181 struct request *request;
183 struct cfq_queue *cfq_queue;
184 struct cfq_io_context *io_context;
186 unsigned int crq_flags;
189 enum cfqq_state_flags {
190 CFQ_CFQQ_FLAG_on_rr = 0,
191 CFQ_CFQQ_FLAG_wait_request,
192 CFQ_CFQQ_FLAG_must_alloc,
193 CFQ_CFQQ_FLAG_must_alloc_slice,
194 CFQ_CFQQ_FLAG_must_dispatch,
195 CFQ_CFQQ_FLAG_fifo_expire,
196 CFQ_CFQQ_FLAG_idle_window,
197 CFQ_CFQQ_FLAG_prio_changed,
200 #define CFQ_CFQQ_FNS(name) \
201 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \
203 cfqq->flags |= (1 << CFQ_CFQQ_FLAG_##name); \
205 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \
207 cfqq->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \
209 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \
211 return (cfqq->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \
215 CFQ_CFQQ_FNS(wait_request);
216 CFQ_CFQQ_FNS(must_alloc);
217 CFQ_CFQQ_FNS(must_alloc_slice);
218 CFQ_CFQQ_FNS(must_dispatch);
219 CFQ_CFQQ_FNS(fifo_expire);
220 CFQ_CFQQ_FNS(idle_window);
221 CFQ_CFQQ_FNS(prio_changed);
224 enum cfq_rq_state_flags {
225 CFQ_CRQ_FLAG_is_sync = 0,
228 #define CFQ_CRQ_FNS(name) \
229 static inline void cfq_mark_crq_##name(struct cfq_rq *crq) \
231 crq->crq_flags |= (1 << CFQ_CRQ_FLAG_##name); \
233 static inline void cfq_clear_crq_##name(struct cfq_rq *crq) \
235 crq->crq_flags &= ~(1 << CFQ_CRQ_FLAG_##name); \
237 static inline int cfq_crq_##name(const struct cfq_rq *crq) \
239 return (crq->crq_flags & (1 << CFQ_CRQ_FLAG_##name)) != 0; \
242 CFQ_CRQ_FNS(is_sync);
245 static struct cfq_queue *cfq_find_cfq_hash(struct cfq_data *, unsigned int, unsigned short);
246 static void cfq_dispatch_insert(request_queue_t *, struct cfq_rq *);
247 static struct cfq_queue *cfq_get_queue(struct cfq_data *cfqd, unsigned int key, struct task_struct *tsk, gfp_t gfp_mask);
250 * scheduler run of queue, if there are requests pending and no one in the
251 * driver that will restart queueing
253 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
255 if (cfqd->busy_queues)
256 kblockd_schedule_work(&cfqd->unplug_work);
259 static int cfq_queue_empty(request_queue_t *q)
261 struct cfq_data *cfqd = q->elevator->elevator_data;
263 return !cfqd->busy_queues;
266 static inline pid_t cfq_queue_pid(struct task_struct *task, int rw)
268 if (rw == READ || rw == WRITE_SYNC)
271 return CFQ_KEY_ASYNC;
275 * Lifted from AS - choose which of crq1 and crq2 that is best served now.
276 * We choose the request that is closest to the head right now. Distance
277 * behind the head is penalized and only allowed to a certain extent.
279 static struct cfq_rq *
280 cfq_choose_req(struct cfq_data *cfqd, struct cfq_rq *crq1, struct cfq_rq *crq2)
282 sector_t last, s1, s2, d1 = 0, d2 = 0;
283 unsigned long back_max;
284 #define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */
285 #define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */
286 unsigned wrap = 0; /* bit mask: requests behind the disk head? */
288 if (crq1 == NULL || crq1 == crq2)
293 if (cfq_crq_is_sync(crq1) && !cfq_crq_is_sync(crq2))
295 else if (cfq_crq_is_sync(crq2) && !cfq_crq_is_sync(crq1))
298 s1 = crq1->request->sector;
299 s2 = crq2->request->sector;
301 last = cfqd->last_sector;
304 * by definition, 1KiB is 2 sectors
306 back_max = cfqd->cfq_back_max * 2;
309 * Strict one way elevator _except_ in the case where we allow
310 * short backward seeks which are biased as twice the cost of a
311 * similar forward seek.
315 else if (s1 + back_max >= last)
316 d1 = (last - s1) * cfqd->cfq_back_penalty;
318 wrap |= CFQ_RQ1_WRAP;
322 else if (s2 + back_max >= last)
323 d2 = (last - s2) * cfqd->cfq_back_penalty;
325 wrap |= CFQ_RQ2_WRAP;
327 /* Found required data */
330 * By doing switch() on the bit mask "wrap" we avoid having to
331 * check two variables for all permutations: --> faster!
334 case 0: /* common case for CFQ: crq1 and crq2 not wrapped */
350 case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both crqs wrapped */
353 * Since both rqs are wrapped,
354 * start with the one that's further behind head
355 * (--> only *one* back seek required),
356 * since back seek takes more time than forward.
366 * would be nice to take fifo expire time into account as well
368 static struct cfq_rq *
369 cfq_find_next_crq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
370 struct cfq_rq *last_crq)
372 struct request *last = last_crq->request;
373 struct rb_node *rbnext = rb_next(&last->rb_node);
374 struct rb_node *rbprev = rb_prev(&last->rb_node);
375 struct cfq_rq *next = NULL, *prev = NULL;
377 BUG_ON(RB_EMPTY_NODE(&last->rb_node));
380 prev = RQ_DATA(rb_entry_rq(rbprev));
383 next = RQ_DATA(rb_entry_rq(rbnext));
385 rbnext = rb_first(&cfqq->sort_list);
386 if (rbnext && rbnext != &last->rb_node)
387 next = RQ_DATA(rb_entry_rq(rbnext));
390 return cfq_choose_req(cfqd, next, prev);
393 static void cfq_resort_rr_list(struct cfq_queue *cfqq, int preempted)
395 struct cfq_data *cfqd = cfqq->cfqd;
396 struct list_head *list, *entry;
398 BUG_ON(!cfq_cfqq_on_rr(cfqq));
400 list_del(&cfqq->cfq_list);
402 if (cfq_class_rt(cfqq))
403 list = &cfqd->cur_rr;
404 else if (cfq_class_idle(cfqq))
405 list = &cfqd->idle_rr;
408 * if cfqq has requests in flight, don't allow it to be
409 * found in cfq_set_active_queue before it has finished them.
410 * this is done to increase fairness between a process that
411 * has lots of io pending vs one that only generates one
412 * sporadically or synchronously
414 if (cfq_cfqq_dispatched(cfqq))
415 list = &cfqd->busy_rr;
417 list = &cfqd->rr_list[cfqq->ioprio];
421 * if queue was preempted, just add to front to be fair. busy_rr
422 * isn't sorted, but insert at the back for fairness.
424 if (preempted || list == &cfqd->busy_rr) {
428 list_add_tail(&cfqq->cfq_list, list);
433 * sort by when queue was last serviced
436 while ((entry = entry->prev) != list) {
437 struct cfq_queue *__cfqq = list_entry_cfqq(entry);
439 if (!__cfqq->service_last)
441 if (time_before(__cfqq->service_last, cfqq->service_last))
445 list_add(&cfqq->cfq_list, entry);
449 * add to busy list of queues for service, trying to be fair in ordering
450 * the pending list according to last request service
453 cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
455 BUG_ON(cfq_cfqq_on_rr(cfqq));
456 cfq_mark_cfqq_on_rr(cfqq);
459 cfq_resort_rr_list(cfqq, 0);
463 cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
465 BUG_ON(!cfq_cfqq_on_rr(cfqq));
466 cfq_clear_cfqq_on_rr(cfqq);
467 list_move(&cfqq->cfq_list, &cfqd->empty_list);
469 BUG_ON(!cfqd->busy_queues);
474 * rb tree support functions
476 static inline void cfq_del_crq_rb(struct cfq_rq *crq)
478 struct cfq_queue *cfqq = crq->cfq_queue;
479 struct cfq_data *cfqd = cfqq->cfqd;
480 const int sync = cfq_crq_is_sync(crq);
482 BUG_ON(!cfqq->queued[sync]);
483 cfqq->queued[sync]--;
485 elv_rb_del(&cfqq->sort_list, crq->request);
487 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
488 cfq_del_cfqq_rr(cfqd, cfqq);
491 static void cfq_add_crq_rb(struct cfq_rq *crq)
493 struct cfq_queue *cfqq = crq->cfq_queue;
494 struct cfq_data *cfqd = cfqq->cfqd;
495 struct request *rq = crq->request;
496 struct request *__alias;
498 cfqq->queued[cfq_crq_is_sync(crq)]++;
501 * looks a little odd, but the first insert might return an alias.
502 * if that happens, put the alias on the dispatch list
504 while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL)
505 cfq_dispatch_insert(cfqd->queue, RQ_DATA(__alias));
509 cfq_reposition_crq_rb(struct cfq_queue *cfqq, struct cfq_rq *crq)
511 elv_rb_del(&cfqq->sort_list, crq->request);
512 cfqq->queued[cfq_crq_is_sync(crq)]--;
516 static struct request *
517 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
519 struct task_struct *tsk = current;
520 pid_t key = cfq_queue_pid(tsk, bio_data_dir(bio));
521 sector_t sector = bio->bi_sector + bio_sectors(bio);
522 struct cfq_queue *cfqq;
524 cfqq = cfq_find_cfq_hash(cfqd, key, tsk->ioprio);
526 return elv_rb_find(&cfqq->sort_list, sector);
531 static void cfq_activate_request(request_queue_t *q, struct request *rq)
533 struct cfq_data *cfqd = q->elevator->elevator_data;
535 cfqd->rq_in_driver++;
538 * If the depth is larger 1, it really could be queueing. But lets
539 * make the mark a little higher - idling could still be good for
540 * low queueing, and a low queueing number could also just indicate
541 * a SCSI mid layer like behaviour where limit+1 is often seen.
543 if (!cfqd->hw_tag && cfqd->rq_in_driver > 4)
547 static void cfq_deactivate_request(request_queue_t *q, struct request *rq)
549 struct cfq_data *cfqd = q->elevator->elevator_data;
551 WARN_ON(!cfqd->rq_in_driver);
552 cfqd->rq_in_driver--;
555 static void cfq_remove_request(struct request *rq)
557 struct cfq_rq *crq = RQ_DATA(rq);
558 struct cfq_queue *cfqq = crq->cfq_queue;
560 if (cfqq->next_crq == crq)
561 cfqq->next_crq = cfq_find_next_crq(cfqq->cfqd, cfqq, crq);
563 list_del_init(&rq->queuelist);
568 cfq_merge(request_queue_t *q, struct request **req, struct bio *bio)
570 struct cfq_data *cfqd = q->elevator->elevator_data;
571 struct request *__rq;
573 __rq = cfq_find_rq_fmerge(cfqd, bio);
574 if (__rq && elv_rq_merge_ok(__rq, bio)) {
576 return ELEVATOR_FRONT_MERGE;
579 return ELEVATOR_NO_MERGE;
582 static void cfq_merged_request(request_queue_t *q, struct request *req,
585 struct cfq_rq *crq = RQ_DATA(req);
587 if (type == ELEVATOR_FRONT_MERGE) {
588 struct cfq_queue *cfqq = crq->cfq_queue;
590 cfq_reposition_crq_rb(cfqq, crq);
595 cfq_merged_requests(request_queue_t *q, struct request *rq,
596 struct request *next)
599 * reposition in fifo if next is older than rq
601 if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
602 time_before(next->start_time, rq->start_time))
603 list_move(&rq->queuelist, &next->queuelist);
605 cfq_remove_request(next);
609 __cfq_set_active_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
613 * stop potential idle class queues waiting service
615 del_timer(&cfqd->idle_class_timer);
617 cfqq->slice_start = jiffies;
619 cfqq->slice_left = 0;
620 cfq_clear_cfqq_must_alloc_slice(cfqq);
621 cfq_clear_cfqq_fifo_expire(cfqq);
624 cfqd->active_queue = cfqq;
628 * current cfqq expired its slice (or was too idle), select new one
631 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
634 unsigned long now = jiffies;
636 if (cfq_cfqq_wait_request(cfqq))
637 del_timer(&cfqd->idle_slice_timer);
639 if (!preempted && !cfq_cfqq_dispatched(cfqq)) {
640 cfqq->service_last = now;
641 cfq_schedule_dispatch(cfqd);
644 cfq_clear_cfqq_must_dispatch(cfqq);
645 cfq_clear_cfqq_wait_request(cfqq);
648 * store what was left of this slice, if the queue idled out
651 if (time_after(cfqq->slice_end, now))
652 cfqq->slice_left = cfqq->slice_end - now;
654 cfqq->slice_left = 0;
656 if (cfq_cfqq_on_rr(cfqq))
657 cfq_resort_rr_list(cfqq, preempted);
659 if (cfqq == cfqd->active_queue)
660 cfqd->active_queue = NULL;
662 if (cfqd->active_cic) {
663 put_io_context(cfqd->active_cic->ioc);
664 cfqd->active_cic = NULL;
667 cfqd->dispatch_slice = 0;
670 static inline void cfq_slice_expired(struct cfq_data *cfqd, int preempted)
672 struct cfq_queue *cfqq = cfqd->active_queue;
675 __cfq_slice_expired(cfqd, cfqq, preempted);
688 static int cfq_get_next_prio_level(struct cfq_data *cfqd)
697 for (p = cfqd->cur_prio; p <= cfqd->cur_end_prio; p++) {
698 if (!list_empty(&cfqd->rr_list[p])) {
707 if (++cfqd->cur_end_prio == CFQ_PRIO_LISTS) {
708 cfqd->cur_end_prio = 0;
715 if (unlikely(prio == -1))
718 BUG_ON(prio >= CFQ_PRIO_LISTS);
720 list_splice_init(&cfqd->rr_list[prio], &cfqd->cur_rr);
722 cfqd->cur_prio = prio + 1;
723 if (cfqd->cur_prio > cfqd->cur_end_prio) {
724 cfqd->cur_end_prio = cfqd->cur_prio;
727 if (cfqd->cur_end_prio == CFQ_PRIO_LISTS) {
729 cfqd->cur_end_prio = 0;
735 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd)
737 struct cfq_queue *cfqq = NULL;
740 * if current list is non-empty, grab first entry. if it is empty,
741 * get next prio level and grab first entry then if any are spliced
743 if (!list_empty(&cfqd->cur_rr) || cfq_get_next_prio_level(cfqd) != -1)
744 cfqq = list_entry_cfqq(cfqd->cur_rr.next);
747 * If no new queues are available, check if the busy list has some
748 * before falling back to idle io.
750 if (!cfqq && !list_empty(&cfqd->busy_rr))
751 cfqq = list_entry_cfqq(cfqd->busy_rr.next);
754 * if we have idle queues and no rt or be queues had pending
755 * requests, either allow immediate service if the grace period
756 * has passed or arm the idle grace timer
758 if (!cfqq && !list_empty(&cfqd->idle_rr)) {
759 unsigned long end = cfqd->last_end_request + CFQ_IDLE_GRACE;
761 if (time_after_eq(jiffies, end))
762 cfqq = list_entry_cfqq(cfqd->idle_rr.next);
764 mod_timer(&cfqd->idle_class_timer, end);
767 __cfq_set_active_queue(cfqd, cfqq);
771 #define CIC_SEEKY(cic) ((cic)->seek_mean > (128 * 1024))
773 static int cfq_arm_slice_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq)
776 struct cfq_io_context *cic;
779 WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
780 WARN_ON(cfqq != cfqd->active_queue);
783 * idle is disabled, either manually or by past process history
785 if (!cfqd->cfq_slice_idle)
787 if (!cfq_cfqq_idle_window(cfqq))
790 * task has exited, don't wait
792 cic = cfqd->active_cic;
793 if (!cic || !cic->ioc->task)
796 cfq_mark_cfqq_must_dispatch(cfqq);
797 cfq_mark_cfqq_wait_request(cfqq);
799 sl = min(cfqq->slice_end - 1, (unsigned long) cfqd->cfq_slice_idle);
802 * we don't want to idle for seeks, but we do want to allow
803 * fair distribution of slice time for a process doing back-to-back
804 * seeks. so allow a little bit of time for him to submit a new rq
806 if (sample_valid(cic->seek_samples) && CIC_SEEKY(cic))
807 sl = min(sl, msecs_to_jiffies(2));
809 mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
813 static void cfq_dispatch_insert(request_queue_t *q, struct cfq_rq *crq)
815 struct cfq_data *cfqd = q->elevator->elevator_data;
816 struct cfq_queue *cfqq = crq->cfq_queue;
819 cfq_remove_request(crq->request);
820 cfqq->on_dispatch[cfq_crq_is_sync(crq)]++;
821 elv_dispatch_sort(q, crq->request);
823 rq = list_entry(q->queue_head.prev, struct request, queuelist);
824 cfqd->last_sector = rq->sector + rq->nr_sectors;
828 * return expired entry, or NULL to just start from scratch in rbtree
830 static inline struct cfq_rq *cfq_check_fifo(struct cfq_queue *cfqq)
832 struct cfq_data *cfqd = cfqq->cfqd;
836 if (cfq_cfqq_fifo_expire(cfqq))
839 if (!list_empty(&cfqq->fifo)) {
840 int fifo = cfq_cfqq_class_sync(cfqq);
842 crq = RQ_DATA(rq_entry_fifo(cfqq->fifo.next));
844 if (time_after(jiffies, rq->start_time + cfqd->cfq_fifo_expire[fifo])) {
845 cfq_mark_cfqq_fifo_expire(cfqq);
854 * Scale schedule slice based on io priority. Use the sync time slice only
855 * if a queue is marked sync and has sync io queued. A sync queue with async
856 * io only, should not get full sync slice length.
859 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
861 const int base_slice = cfqd->cfq_slice[cfq_cfqq_sync(cfqq)];
863 WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
865 return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - cfqq->ioprio));
869 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
871 cfqq->slice_end = cfq_prio_to_slice(cfqd, cfqq) + jiffies;
875 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
877 const int base_rq = cfqd->cfq_slice_async_rq;
879 WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
881 return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
885 * get next queue for service
887 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
889 unsigned long now = jiffies;
890 struct cfq_queue *cfqq;
892 cfqq = cfqd->active_queue;
899 if (!cfq_cfqq_must_dispatch(cfqq) && time_after(now, cfqq->slice_end))
903 * if queue has requests, dispatch one. if not, check if
904 * enough slice is left to wait for one
906 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
908 else if (cfq_cfqq_dispatched(cfqq)) {
911 } else if (cfq_cfqq_class_sync(cfqq)) {
912 if (cfq_arm_slice_timer(cfqd, cfqq))
917 cfq_slice_expired(cfqd, 0);
919 cfqq = cfq_set_active_queue(cfqd);
925 __cfq_dispatch_requests(struct cfq_data *cfqd, struct cfq_queue *cfqq,
930 BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
936 * follow expired path, else get first next available
938 if ((crq = cfq_check_fifo(cfqq)) == NULL)
939 crq = cfqq->next_crq;
942 * finally, insert request into driver dispatch list
944 cfq_dispatch_insert(cfqd->queue, crq);
946 cfqd->dispatch_slice++;
949 if (!cfqd->active_cic) {
950 atomic_inc(&crq->io_context->ioc->refcount);
951 cfqd->active_cic = crq->io_context;
954 if (RB_EMPTY_ROOT(&cfqq->sort_list))
957 } while (dispatched < max_dispatch);
960 * if slice end isn't set yet, set it.
962 if (!cfqq->slice_end)
963 cfq_set_prio_slice(cfqd, cfqq);
966 * expire an async queue immediately if it has used up its slice. idle
967 * queue always expire after 1 dispatch round.
969 if ((!cfq_cfqq_sync(cfqq) &&
970 cfqd->dispatch_slice >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
971 cfq_class_idle(cfqq) ||
972 !cfq_cfqq_idle_window(cfqq))
973 cfq_slice_expired(cfqd, 0);
979 cfq_forced_dispatch_cfqqs(struct list_head *list)
981 struct cfq_queue *cfqq, *next;
986 list_for_each_entry_safe(cfqq, next, list, cfq_list) {
987 while ((crq = cfqq->next_crq)) {
988 cfq_dispatch_insert(cfqq->cfqd->queue, crq);
991 BUG_ON(!list_empty(&cfqq->fifo));
998 cfq_forced_dispatch(struct cfq_data *cfqd)
1000 int i, dispatched = 0;
1002 for (i = 0; i < CFQ_PRIO_LISTS; i++)
1003 dispatched += cfq_forced_dispatch_cfqqs(&cfqd->rr_list[i]);
1005 dispatched += cfq_forced_dispatch_cfqqs(&cfqd->busy_rr);
1006 dispatched += cfq_forced_dispatch_cfqqs(&cfqd->cur_rr);
1007 dispatched += cfq_forced_dispatch_cfqqs(&cfqd->idle_rr);
1009 cfq_slice_expired(cfqd, 0);
1011 BUG_ON(cfqd->busy_queues);
1017 cfq_dispatch_requests(request_queue_t *q, int force)
1019 struct cfq_data *cfqd = q->elevator->elevator_data;
1020 struct cfq_queue *cfqq, *prev_cfqq;
1023 if (!cfqd->busy_queues)
1026 if (unlikely(force))
1027 return cfq_forced_dispatch(cfqd);
1031 while ((cfqq = cfq_select_queue(cfqd)) != NULL) {
1035 * Don't repeat dispatch from the previous queue.
1037 if (prev_cfqq == cfqq)
1040 cfq_clear_cfqq_must_dispatch(cfqq);
1041 cfq_clear_cfqq_wait_request(cfqq);
1042 del_timer(&cfqd->idle_slice_timer);
1044 max_dispatch = cfqd->cfq_quantum;
1045 if (cfq_class_idle(cfqq))
1048 dispatched += __cfq_dispatch_requests(cfqd, cfqq, max_dispatch);
1051 * If the dispatch cfqq has idling enabled and is still
1052 * the active queue, break out.
1054 if (cfq_cfqq_idle_window(cfqq) && cfqd->active_queue)
1064 * task holds one reference to the queue, dropped when task exits. each crq
1065 * in-flight on this queue also holds a reference, dropped when crq is freed.
1067 * queue lock must be held here.
1069 static void cfq_put_queue(struct cfq_queue *cfqq)
1071 struct cfq_data *cfqd = cfqq->cfqd;
1073 BUG_ON(atomic_read(&cfqq->ref) <= 0);
1075 if (!atomic_dec_and_test(&cfqq->ref))
1078 BUG_ON(rb_first(&cfqq->sort_list));
1079 BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
1080 BUG_ON(cfq_cfqq_on_rr(cfqq));
1082 if (unlikely(cfqd->active_queue == cfqq))
1083 __cfq_slice_expired(cfqd, cfqq, 0);
1086 * it's on the empty list and still hashed
1088 list_del(&cfqq->cfq_list);
1089 hlist_del(&cfqq->cfq_hash);
1090 kmem_cache_free(cfq_pool, cfqq);
1093 static inline struct cfq_queue *
1094 __cfq_find_cfq_hash(struct cfq_data *cfqd, unsigned int key, unsigned int prio,
1097 struct hlist_head *hash_list = &cfqd->cfq_hash[hashval];
1098 struct hlist_node *entry;
1099 struct cfq_queue *__cfqq;
1101 hlist_for_each_entry(__cfqq, entry, hash_list, cfq_hash) {
1102 const unsigned short __p = IOPRIO_PRIO_VALUE(__cfqq->org_ioprio_class, __cfqq->org_ioprio);
1104 if (__cfqq->key == key && (__p == prio || !prio))
1111 static struct cfq_queue *
1112 cfq_find_cfq_hash(struct cfq_data *cfqd, unsigned int key, unsigned short prio)
1114 return __cfq_find_cfq_hash(cfqd, key, prio, hash_long(key, CFQ_QHASH_SHIFT));
1117 static void cfq_free_io_context(struct io_context *ioc)
1119 struct cfq_io_context *__cic;
1123 while ((n = rb_first(&ioc->cic_root)) != NULL) {
1124 __cic = rb_entry(n, struct cfq_io_context, rb_node);
1125 rb_erase(&__cic->rb_node, &ioc->cic_root);
1126 kmem_cache_free(cfq_ioc_pool, __cic);
1130 if (atomic_sub_and_test(freed, &ioc_count) && ioc_gone)
1134 static void cfq_trim(struct io_context *ioc)
1136 ioc->set_ioprio = NULL;
1137 cfq_free_io_context(ioc);
1141 * Called with interrupts disabled
1143 static void cfq_exit_single_io_context(struct cfq_io_context *cic)
1145 struct cfq_data *cfqd = cic->key;
1153 WARN_ON(!irqs_disabled());
1155 spin_lock(q->queue_lock);
1157 if (cic->cfqq[ASYNC]) {
1158 if (unlikely(cic->cfqq[ASYNC] == cfqd->active_queue))
1159 __cfq_slice_expired(cfqd, cic->cfqq[ASYNC], 0);
1160 cfq_put_queue(cic->cfqq[ASYNC]);
1161 cic->cfqq[ASYNC] = NULL;
1164 if (cic->cfqq[SYNC]) {
1165 if (unlikely(cic->cfqq[SYNC] == cfqd->active_queue))
1166 __cfq_slice_expired(cfqd, cic->cfqq[SYNC], 0);
1167 cfq_put_queue(cic->cfqq[SYNC]);
1168 cic->cfqq[SYNC] = NULL;
1172 list_del_init(&cic->queue_list);
1173 spin_unlock(q->queue_lock);
1176 static void cfq_exit_io_context(struct io_context *ioc)
1178 struct cfq_io_context *__cic;
1179 unsigned long flags;
1183 * put the reference this task is holding to the various queues
1185 spin_lock_irqsave(&cfq_exit_lock, flags);
1187 n = rb_first(&ioc->cic_root);
1189 __cic = rb_entry(n, struct cfq_io_context, rb_node);
1191 cfq_exit_single_io_context(__cic);
1195 spin_unlock_irqrestore(&cfq_exit_lock, flags);
1198 static struct cfq_io_context *
1199 cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
1201 struct cfq_io_context *cic = kmem_cache_alloc(cfq_ioc_pool, gfp_mask);
1204 memset(cic, 0, sizeof(*cic));
1205 cic->last_end_request = jiffies;
1206 INIT_LIST_HEAD(&cic->queue_list);
1207 cic->dtor = cfq_free_io_context;
1208 cic->exit = cfq_exit_io_context;
1209 atomic_inc(&ioc_count);
1215 static void cfq_init_prio_data(struct cfq_queue *cfqq)
1217 struct task_struct *tsk = current;
1220 if (!cfq_cfqq_prio_changed(cfqq))
1223 ioprio_class = IOPRIO_PRIO_CLASS(tsk->ioprio);
1224 switch (ioprio_class) {
1226 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
1227 case IOPRIO_CLASS_NONE:
1229 * no prio set, place us in the middle of the BE classes
1231 cfqq->ioprio = task_nice_ioprio(tsk);
1232 cfqq->ioprio_class = IOPRIO_CLASS_BE;
1234 case IOPRIO_CLASS_RT:
1235 cfqq->ioprio = task_ioprio(tsk);
1236 cfqq->ioprio_class = IOPRIO_CLASS_RT;
1238 case IOPRIO_CLASS_BE:
1239 cfqq->ioprio = task_ioprio(tsk);
1240 cfqq->ioprio_class = IOPRIO_CLASS_BE;
1242 case IOPRIO_CLASS_IDLE:
1243 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
1245 cfq_clear_cfqq_idle_window(cfqq);
1250 * keep track of original prio settings in case we have to temporarily
1251 * elevate the priority of this queue
1253 cfqq->org_ioprio = cfqq->ioprio;
1254 cfqq->org_ioprio_class = cfqq->ioprio_class;
1256 if (cfq_cfqq_on_rr(cfqq))
1257 cfq_resort_rr_list(cfqq, 0);
1259 cfq_clear_cfqq_prio_changed(cfqq);
1262 static inline void changed_ioprio(struct cfq_io_context *cic)
1264 struct cfq_data *cfqd = cic->key;
1265 struct cfq_queue *cfqq;
1267 if (unlikely(!cfqd))
1270 spin_lock(cfqd->queue->queue_lock);
1272 cfqq = cic->cfqq[ASYNC];
1274 struct cfq_queue *new_cfqq;
1275 new_cfqq = cfq_get_queue(cfqd, CFQ_KEY_ASYNC, cic->ioc->task,
1278 cic->cfqq[ASYNC] = new_cfqq;
1279 cfq_put_queue(cfqq);
1283 cfqq = cic->cfqq[SYNC];
1285 cfq_mark_cfqq_prio_changed(cfqq);
1287 spin_unlock(cfqd->queue->queue_lock);
1291 * callback from sys_ioprio_set, irqs are disabled
1293 static int cfq_ioc_set_ioprio(struct io_context *ioc, unsigned int ioprio)
1295 struct cfq_io_context *cic;
1298 spin_lock(&cfq_exit_lock);
1300 n = rb_first(&ioc->cic_root);
1302 cic = rb_entry(n, struct cfq_io_context, rb_node);
1304 changed_ioprio(cic);
1308 spin_unlock(&cfq_exit_lock);
1313 static struct cfq_queue *
1314 cfq_get_queue(struct cfq_data *cfqd, unsigned int key, struct task_struct *tsk,
1317 const int hashval = hash_long(key, CFQ_QHASH_SHIFT);
1318 struct cfq_queue *cfqq, *new_cfqq = NULL;
1319 unsigned short ioprio;
1322 ioprio = tsk->ioprio;
1323 cfqq = __cfq_find_cfq_hash(cfqd, key, ioprio, hashval);
1329 } else if (gfp_mask & __GFP_WAIT) {
1330 spin_unlock_irq(cfqd->queue->queue_lock);
1331 new_cfqq = kmem_cache_alloc(cfq_pool, gfp_mask);
1332 spin_lock_irq(cfqd->queue->queue_lock);
1335 cfqq = kmem_cache_alloc(cfq_pool, gfp_mask);
1340 memset(cfqq, 0, sizeof(*cfqq));
1342 INIT_HLIST_NODE(&cfqq->cfq_hash);
1343 INIT_LIST_HEAD(&cfqq->cfq_list);
1344 INIT_LIST_HEAD(&cfqq->fifo);
1347 hlist_add_head(&cfqq->cfq_hash, &cfqd->cfq_hash[hashval]);
1348 atomic_set(&cfqq->ref, 0);
1350 cfqq->service_last = 0;
1352 * set ->slice_left to allow preemption for a new process
1354 cfqq->slice_left = 2 * cfqd->cfq_slice_idle;
1355 cfq_mark_cfqq_idle_window(cfqq);
1356 cfq_mark_cfqq_prio_changed(cfqq);
1357 cfq_init_prio_data(cfqq);
1361 kmem_cache_free(cfq_pool, new_cfqq);
1363 atomic_inc(&cfqq->ref);
1365 WARN_ON((gfp_mask & __GFP_WAIT) && !cfqq);
1370 cfq_drop_dead_cic(struct io_context *ioc, struct cfq_io_context *cic)
1372 spin_lock(&cfq_exit_lock);
1373 rb_erase(&cic->rb_node, &ioc->cic_root);
1374 list_del_init(&cic->queue_list);
1375 spin_unlock(&cfq_exit_lock);
1376 kmem_cache_free(cfq_ioc_pool, cic);
1377 atomic_dec(&ioc_count);
1380 static struct cfq_io_context *
1381 cfq_cic_rb_lookup(struct cfq_data *cfqd, struct io_context *ioc)
1384 struct cfq_io_context *cic;
1385 void *k, *key = cfqd;
1388 n = ioc->cic_root.rb_node;
1390 cic = rb_entry(n, struct cfq_io_context, rb_node);
1391 /* ->key must be copied to avoid race with cfq_exit_queue() */
1394 cfq_drop_dead_cic(ioc, cic);
1410 cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
1411 struct cfq_io_context *cic)
1414 struct rb_node *parent;
1415 struct cfq_io_context *__cic;
1421 ioc->set_ioprio = cfq_ioc_set_ioprio;
1424 p = &ioc->cic_root.rb_node;
1427 __cic = rb_entry(parent, struct cfq_io_context, rb_node);
1428 /* ->key must be copied to avoid race with cfq_exit_queue() */
1431 cfq_drop_dead_cic(ioc, __cic);
1437 else if (cic->key > k)
1438 p = &(*p)->rb_right;
1443 spin_lock(&cfq_exit_lock);
1444 rb_link_node(&cic->rb_node, parent, p);
1445 rb_insert_color(&cic->rb_node, &ioc->cic_root);
1446 list_add(&cic->queue_list, &cfqd->cic_list);
1447 spin_unlock(&cfq_exit_lock);
1451 * Setup general io context and cfq io context. There can be several cfq
1452 * io contexts per general io context, if this process is doing io to more
1453 * than one device managed by cfq.
1455 static struct cfq_io_context *
1456 cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
1458 struct io_context *ioc = NULL;
1459 struct cfq_io_context *cic;
1461 might_sleep_if(gfp_mask & __GFP_WAIT);
1463 ioc = get_io_context(gfp_mask);
1467 cic = cfq_cic_rb_lookup(cfqd, ioc);
1471 cic = cfq_alloc_io_context(cfqd, gfp_mask);
1475 cfq_cic_link(cfqd, ioc, cic);
1479 put_io_context(ioc);
1484 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
1486 unsigned long elapsed, ttime;
1489 * if this context already has stuff queued, thinktime is from
1490 * last queue not last end
1493 if (time_after(cic->last_end_request, cic->last_queue))
1494 elapsed = jiffies - cic->last_end_request;
1496 elapsed = jiffies - cic->last_queue;
1498 elapsed = jiffies - cic->last_end_request;
1501 ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
1503 cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
1504 cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
1505 cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
1509 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_io_context *cic,
1515 if (cic->last_request_pos < crq->request->sector)
1516 sdist = crq->request->sector - cic->last_request_pos;
1518 sdist = cic->last_request_pos - crq->request->sector;
1521 * Don't allow the seek distance to get too large from the
1522 * odd fragment, pagein, etc
1524 if (cic->seek_samples <= 60) /* second&third seek */
1525 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*1024);
1527 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*64);
1529 cic->seek_samples = (7*cic->seek_samples + 256) / 8;
1530 cic->seek_total = (7*cic->seek_total + (u64)256*sdist) / 8;
1531 total = cic->seek_total + (cic->seek_samples/2);
1532 do_div(total, cic->seek_samples);
1533 cic->seek_mean = (sector_t)total;
1537 * Disable idle window if the process thinks too long or seeks so much that
1541 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1542 struct cfq_io_context *cic)
1544 int enable_idle = cfq_cfqq_idle_window(cfqq);
1546 if (!cic->ioc->task || !cfqd->cfq_slice_idle ||
1547 (cfqd->hw_tag && CIC_SEEKY(cic)))
1549 else if (sample_valid(cic->ttime_samples)) {
1550 if (cic->ttime_mean > cfqd->cfq_slice_idle)
1557 cfq_mark_cfqq_idle_window(cfqq);
1559 cfq_clear_cfqq_idle_window(cfqq);
1564 * Check if new_cfqq should preempt the currently active queue. Return 0 for
1565 * no or if we aren't sure, a 1 will cause a preempt.
1568 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
1571 struct cfq_queue *cfqq = cfqd->active_queue;
1573 if (cfq_class_idle(new_cfqq))
1579 if (cfq_class_idle(cfqq))
1581 if (!cfq_cfqq_wait_request(new_cfqq))
1584 * if it doesn't have slice left, forget it
1586 if (new_cfqq->slice_left < cfqd->cfq_slice_idle)
1588 if (cfq_crq_is_sync(crq) && !cfq_cfqq_sync(cfqq))
1595 * cfqq preempts the active queue. if we allowed preempt with no slice left,
1596 * let it have half of its nominal slice.
1598 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1600 struct cfq_queue *__cfqq, *next;
1602 list_for_each_entry_safe(__cfqq, next, &cfqd->cur_rr, cfq_list)
1603 cfq_resort_rr_list(__cfqq, 1);
1605 if (!cfqq->slice_left)
1606 cfqq->slice_left = cfq_prio_to_slice(cfqd, cfqq) / 2;
1608 cfqq->slice_end = cfqq->slice_left + jiffies;
1609 cfq_slice_expired(cfqd, 1);
1610 __cfq_set_active_queue(cfqd, cfqq);
1614 * should really be a ll_rw_blk.c helper
1616 static void cfq_start_queueing(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1618 request_queue_t *q = cfqd->queue;
1620 if (!blk_queue_plugged(q))
1623 __generic_unplug_device(q);
1627 * Called when a new fs request (crq) is added (to cfqq). Check if there's
1628 * something we should do about it
1631 cfq_crq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1634 struct cfq_io_context *cic = crq->io_context;
1637 * check if this request is a better next-serve candidate
1639 cfqq->next_crq = cfq_choose_req(cfqd, cfqq->next_crq, crq);
1640 BUG_ON(!cfqq->next_crq);
1643 * we never wait for an async request and we don't allow preemption
1644 * of an async request. so just return early
1646 if (!cfq_crq_is_sync(crq)) {
1648 * sync process issued an async request, if it's waiting
1649 * then expire it and kick rq handling.
1651 if (cic == cfqd->active_cic &&
1652 del_timer(&cfqd->idle_slice_timer)) {
1653 cfq_slice_expired(cfqd, 0);
1654 cfq_start_queueing(cfqd, cfqq);
1659 cfq_update_io_thinktime(cfqd, cic);
1660 cfq_update_io_seektime(cfqd, cic, crq);
1661 cfq_update_idle_window(cfqd, cfqq, cic);
1663 cic->last_queue = jiffies;
1664 cic->last_request_pos = crq->request->sector + crq->request->nr_sectors;
1666 if (cfqq == cfqd->active_queue) {
1668 * if we are waiting for a request for this queue, let it rip
1669 * immediately and flag that we must not expire this queue
1672 if (cfq_cfqq_wait_request(cfqq)) {
1673 cfq_mark_cfqq_must_dispatch(cfqq);
1674 del_timer(&cfqd->idle_slice_timer);
1675 cfq_start_queueing(cfqd, cfqq);
1677 } else if (cfq_should_preempt(cfqd, cfqq, crq)) {
1679 * not the active queue - expire current slice if it is
1680 * idle and has expired it's mean thinktime or this new queue
1681 * has some old slice time left and is of higher priority
1683 cfq_preempt_queue(cfqd, cfqq);
1684 cfq_mark_cfqq_must_dispatch(cfqq);
1685 cfq_start_queueing(cfqd, cfqq);
1689 static void cfq_insert_request(request_queue_t *q, struct request *rq)
1691 struct cfq_data *cfqd = q->elevator->elevator_data;
1692 struct cfq_rq *crq = RQ_DATA(rq);
1693 struct cfq_queue *cfqq = crq->cfq_queue;
1695 cfq_init_prio_data(cfqq);
1697 cfq_add_crq_rb(crq);
1699 if (!cfq_cfqq_on_rr(cfqq))
1700 cfq_add_cfqq_rr(cfqd, cfqq);
1702 list_add_tail(&rq->queuelist, &cfqq->fifo);
1704 cfq_crq_enqueued(cfqd, cfqq, crq);
1707 static void cfq_completed_request(request_queue_t *q, struct request *rq)
1709 struct cfq_rq *crq = RQ_DATA(rq);
1710 struct cfq_queue *cfqq = crq->cfq_queue;
1711 struct cfq_data *cfqd = cfqq->cfqd;
1712 const int sync = cfq_crq_is_sync(crq);
1717 WARN_ON(!cfqd->rq_in_driver);
1718 WARN_ON(!cfqq->on_dispatch[sync]);
1719 cfqd->rq_in_driver--;
1720 cfqq->on_dispatch[sync]--;
1722 if (!cfq_class_idle(cfqq))
1723 cfqd->last_end_request = now;
1725 if (!cfq_cfqq_dispatched(cfqq)) {
1726 if (cfq_cfqq_on_rr(cfqq)) {
1727 cfqq->service_last = now;
1728 cfq_resort_rr_list(cfqq, 0);
1733 crq->io_context->last_end_request = now;
1736 * If this is the active queue, check if it needs to be expired,
1737 * or if we want to idle in case it has no pending requests.
1739 if (cfqd->active_queue == cfqq) {
1740 if (time_after(now, cfqq->slice_end))
1741 cfq_slice_expired(cfqd, 0);
1742 else if (sync && RB_EMPTY_ROOT(&cfqq->sort_list)) {
1743 if (!cfq_arm_slice_timer(cfqd, cfqq))
1744 cfq_schedule_dispatch(cfqd);
1750 * we temporarily boost lower priority queues if they are holding fs exclusive
1751 * resources. they are boosted to normal prio (CLASS_BE/4)
1753 static void cfq_prio_boost(struct cfq_queue *cfqq)
1755 const int ioprio_class = cfqq->ioprio_class;
1756 const int ioprio = cfqq->ioprio;
1758 if (has_fs_excl()) {
1760 * boost idle prio on transactions that would lock out other
1761 * users of the filesystem
1763 if (cfq_class_idle(cfqq))
1764 cfqq->ioprio_class = IOPRIO_CLASS_BE;
1765 if (cfqq->ioprio > IOPRIO_NORM)
1766 cfqq->ioprio = IOPRIO_NORM;
1769 * check if we need to unboost the queue
1771 if (cfqq->ioprio_class != cfqq->org_ioprio_class)
1772 cfqq->ioprio_class = cfqq->org_ioprio_class;
1773 if (cfqq->ioprio != cfqq->org_ioprio)
1774 cfqq->ioprio = cfqq->org_ioprio;
1778 * refile between round-robin lists if we moved the priority class
1780 if ((ioprio_class != cfqq->ioprio_class || ioprio != cfqq->ioprio) &&
1781 cfq_cfqq_on_rr(cfqq))
1782 cfq_resort_rr_list(cfqq, 0);
1786 __cfq_may_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1787 struct task_struct *task, int rw)
1789 if ((cfq_cfqq_wait_request(cfqq) || cfq_cfqq_must_alloc(cfqq)) &&
1790 !cfq_cfqq_must_alloc_slice(cfqq)) {
1791 cfq_mark_cfqq_must_alloc_slice(cfqq);
1792 return ELV_MQUEUE_MUST;
1795 return ELV_MQUEUE_MAY;
1798 static int cfq_may_queue(request_queue_t *q, int rw, struct bio *bio)
1800 struct cfq_data *cfqd = q->elevator->elevator_data;
1801 struct task_struct *tsk = current;
1802 struct cfq_queue *cfqq;
1805 * don't force setup of a queue from here, as a call to may_queue
1806 * does not necessarily imply that a request actually will be queued.
1807 * so just lookup a possibly existing queue, or return 'may queue'
1810 cfqq = cfq_find_cfq_hash(cfqd, cfq_queue_pid(tsk, rw), tsk->ioprio);
1812 cfq_init_prio_data(cfqq);
1813 cfq_prio_boost(cfqq);
1815 return __cfq_may_queue(cfqd, cfqq, tsk, rw);
1818 return ELV_MQUEUE_MAY;
1821 static void cfq_check_waiters(request_queue_t *q, struct cfq_queue *cfqq)
1823 struct cfq_data *cfqd = q->elevator->elevator_data;
1825 if (unlikely(cfqd->rq_starved)) {
1826 struct request_list *rl = &q->rq;
1829 if (waitqueue_active(&rl->wait[READ]))
1830 wake_up(&rl->wait[READ]);
1831 if (waitqueue_active(&rl->wait[WRITE]))
1832 wake_up(&rl->wait[WRITE]);
1837 * queue lock held here
1839 static void cfq_put_request(request_queue_t *q, struct request *rq)
1841 struct cfq_data *cfqd = q->elevator->elevator_data;
1842 struct cfq_rq *crq = RQ_DATA(rq);
1845 struct cfq_queue *cfqq = crq->cfq_queue;
1846 const int rw = rq_data_dir(rq);
1848 BUG_ON(!cfqq->allocated[rw]);
1849 cfqq->allocated[rw]--;
1851 put_io_context(crq->io_context->ioc);
1853 mempool_free(crq, cfqd->crq_pool);
1854 rq->elevator_private = NULL;
1856 cfq_check_waiters(q, cfqq);
1857 cfq_put_queue(cfqq);
1862 * Allocate cfq data structures associated with this request.
1865 cfq_set_request(request_queue_t *q, struct request *rq, struct bio *bio,
1868 struct cfq_data *cfqd = q->elevator->elevator_data;
1869 struct task_struct *tsk = current;
1870 struct cfq_io_context *cic;
1871 const int rw = rq_data_dir(rq);
1872 pid_t key = cfq_queue_pid(tsk, rw);
1873 struct cfq_queue *cfqq;
1875 unsigned long flags;
1876 int is_sync = key != CFQ_KEY_ASYNC;
1878 might_sleep_if(gfp_mask & __GFP_WAIT);
1880 cic = cfq_get_io_context(cfqd, gfp_mask);
1882 spin_lock_irqsave(q->queue_lock, flags);
1887 if (!cic->cfqq[is_sync]) {
1888 cfqq = cfq_get_queue(cfqd, key, tsk, gfp_mask);
1892 cic->cfqq[is_sync] = cfqq;
1894 cfqq = cic->cfqq[is_sync];
1896 cfqq->allocated[rw]++;
1897 cfq_clear_cfqq_must_alloc(cfqq);
1898 cfqd->rq_starved = 0;
1899 atomic_inc(&cfqq->ref);
1900 spin_unlock_irqrestore(q->queue_lock, flags);
1902 crq = mempool_alloc(cfqd->crq_pool, gfp_mask);
1905 crq->cfq_queue = cfqq;
1906 crq->io_context = cic;
1909 cfq_mark_crq_is_sync(crq);
1911 cfq_clear_crq_is_sync(crq);
1913 rq->elevator_private = crq;
1917 spin_lock_irqsave(q->queue_lock, flags);
1918 cfqq->allocated[rw]--;
1919 if (!(cfqq->allocated[0] + cfqq->allocated[1]))
1920 cfq_mark_cfqq_must_alloc(cfqq);
1921 cfq_put_queue(cfqq);
1924 put_io_context(cic->ioc);
1926 * mark us rq allocation starved. we need to kickstart the process
1927 * ourselves if there are no pending requests that can do it for us.
1928 * that would be an extremely rare OOM situation
1930 cfqd->rq_starved = 1;
1931 cfq_schedule_dispatch(cfqd);
1932 spin_unlock_irqrestore(q->queue_lock, flags);
1936 static void cfq_kick_queue(void *data)
1938 request_queue_t *q = data;
1939 struct cfq_data *cfqd = q->elevator->elevator_data;
1940 unsigned long flags;
1942 spin_lock_irqsave(q->queue_lock, flags);
1944 if (cfqd->rq_starved) {
1945 struct request_list *rl = &q->rq;
1948 * we aren't guaranteed to get a request after this, but we
1949 * have to be opportunistic
1952 if (waitqueue_active(&rl->wait[READ]))
1953 wake_up(&rl->wait[READ]);
1954 if (waitqueue_active(&rl->wait[WRITE]))
1955 wake_up(&rl->wait[WRITE]);
1960 spin_unlock_irqrestore(q->queue_lock, flags);
1964 * Timer running if the active_queue is currently idling inside its time slice
1966 static void cfq_idle_slice_timer(unsigned long data)
1968 struct cfq_data *cfqd = (struct cfq_data *) data;
1969 struct cfq_queue *cfqq;
1970 unsigned long flags;
1972 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1974 if ((cfqq = cfqd->active_queue) != NULL) {
1975 unsigned long now = jiffies;
1980 if (time_after(now, cfqq->slice_end))
1984 * only expire and reinvoke request handler, if there are
1985 * other queues with pending requests
1987 if (!cfqd->busy_queues)
1991 * not expired and it has a request pending, let it dispatch
1993 if (!RB_EMPTY_ROOT(&cfqq->sort_list)) {
1994 cfq_mark_cfqq_must_dispatch(cfqq);
1999 cfq_slice_expired(cfqd, 0);
2001 cfq_schedule_dispatch(cfqd);
2003 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2007 * Timer running if an idle class queue is waiting for service
2009 static void cfq_idle_class_timer(unsigned long data)
2011 struct cfq_data *cfqd = (struct cfq_data *) data;
2012 unsigned long flags, end;
2014 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2017 * race with a non-idle queue, reset timer
2019 end = cfqd->last_end_request + CFQ_IDLE_GRACE;
2020 if (!time_after_eq(jiffies, end))
2021 mod_timer(&cfqd->idle_class_timer, end);
2023 cfq_schedule_dispatch(cfqd);
2025 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2028 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
2030 del_timer_sync(&cfqd->idle_slice_timer);
2031 del_timer_sync(&cfqd->idle_class_timer);
2032 blk_sync_queue(cfqd->queue);
2035 static void cfq_exit_queue(elevator_t *e)
2037 struct cfq_data *cfqd = e->elevator_data;
2038 request_queue_t *q = cfqd->queue;
2040 cfq_shutdown_timer_wq(cfqd);
2042 spin_lock(&cfq_exit_lock);
2043 spin_lock_irq(q->queue_lock);
2045 if (cfqd->active_queue)
2046 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
2048 while (!list_empty(&cfqd->cic_list)) {
2049 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
2050 struct cfq_io_context,
2052 if (cic->cfqq[ASYNC]) {
2053 cfq_put_queue(cic->cfqq[ASYNC]);
2054 cic->cfqq[ASYNC] = NULL;
2056 if (cic->cfqq[SYNC]) {
2057 cfq_put_queue(cic->cfqq[SYNC]);
2058 cic->cfqq[SYNC] = NULL;
2061 list_del_init(&cic->queue_list);
2064 spin_unlock_irq(q->queue_lock);
2065 spin_unlock(&cfq_exit_lock);
2067 cfq_shutdown_timer_wq(cfqd);
2069 mempool_destroy(cfqd->crq_pool);
2070 kfree(cfqd->cfq_hash);
2074 static void *cfq_init_queue(request_queue_t *q, elevator_t *e)
2076 struct cfq_data *cfqd;
2079 cfqd = kmalloc(sizeof(*cfqd), GFP_KERNEL);
2083 memset(cfqd, 0, sizeof(*cfqd));
2085 for (i = 0; i < CFQ_PRIO_LISTS; i++)
2086 INIT_LIST_HEAD(&cfqd->rr_list[i]);
2088 INIT_LIST_HEAD(&cfqd->busy_rr);
2089 INIT_LIST_HEAD(&cfqd->cur_rr);
2090 INIT_LIST_HEAD(&cfqd->idle_rr);
2091 INIT_LIST_HEAD(&cfqd->empty_list);
2092 INIT_LIST_HEAD(&cfqd->cic_list);
2094 cfqd->cfq_hash = kmalloc(sizeof(struct hlist_head) * CFQ_QHASH_ENTRIES, GFP_KERNEL);
2095 if (!cfqd->cfq_hash)
2098 cfqd->crq_pool = mempool_create_slab_pool(BLKDEV_MIN_RQ, crq_pool);
2099 if (!cfqd->crq_pool)
2102 for (i = 0; i < CFQ_QHASH_ENTRIES; i++)
2103 INIT_HLIST_HEAD(&cfqd->cfq_hash[i]);
2107 init_timer(&cfqd->idle_slice_timer);
2108 cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
2109 cfqd->idle_slice_timer.data = (unsigned long) cfqd;
2111 init_timer(&cfqd->idle_class_timer);
2112 cfqd->idle_class_timer.function = cfq_idle_class_timer;
2113 cfqd->idle_class_timer.data = (unsigned long) cfqd;
2115 INIT_WORK(&cfqd->unplug_work, cfq_kick_queue, q);
2117 cfqd->cfq_queued = cfq_queued;
2118 cfqd->cfq_quantum = cfq_quantum;
2119 cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
2120 cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
2121 cfqd->cfq_back_max = cfq_back_max;
2122 cfqd->cfq_back_penalty = cfq_back_penalty;
2123 cfqd->cfq_slice[0] = cfq_slice_async;
2124 cfqd->cfq_slice[1] = cfq_slice_sync;
2125 cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
2126 cfqd->cfq_slice_idle = cfq_slice_idle;
2130 kfree(cfqd->cfq_hash);
2136 static void cfq_slab_kill(void)
2139 kmem_cache_destroy(crq_pool);
2141 kmem_cache_destroy(cfq_pool);
2143 kmem_cache_destroy(cfq_ioc_pool);
2146 static int __init cfq_slab_setup(void)
2148 crq_pool = kmem_cache_create("crq_pool", sizeof(struct cfq_rq), 0, 0,
2153 cfq_pool = kmem_cache_create("cfq_pool", sizeof(struct cfq_queue), 0, 0,
2158 cfq_ioc_pool = kmem_cache_create("cfq_ioc_pool",
2159 sizeof(struct cfq_io_context), 0, 0, NULL, NULL);
2170 * sysfs parts below -->
2174 cfq_var_show(unsigned int var, char *page)
2176 return sprintf(page, "%d\n", var);
2180 cfq_var_store(unsigned int *var, const char *page, size_t count)
2182 char *p = (char *) page;
2184 *var = simple_strtoul(p, &p, 10);
2188 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
2189 static ssize_t __FUNC(elevator_t *e, char *page) \
2191 struct cfq_data *cfqd = e->elevator_data; \
2192 unsigned int __data = __VAR; \
2194 __data = jiffies_to_msecs(__data); \
2195 return cfq_var_show(__data, (page)); \
2197 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
2198 SHOW_FUNCTION(cfq_queued_show, cfqd->cfq_queued, 0);
2199 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
2200 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
2201 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
2202 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
2203 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
2204 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
2205 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
2206 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
2207 #undef SHOW_FUNCTION
2209 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
2210 static ssize_t __FUNC(elevator_t *e, const char *page, size_t count) \
2212 struct cfq_data *cfqd = e->elevator_data; \
2213 unsigned int __data; \
2214 int ret = cfq_var_store(&__data, (page), count); \
2215 if (__data < (MIN)) \
2217 else if (__data > (MAX)) \
2220 *(__PTR) = msecs_to_jiffies(__data); \
2222 *(__PTR) = __data; \
2225 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
2226 STORE_FUNCTION(cfq_queued_store, &cfqd->cfq_queued, 1, UINT_MAX, 0);
2227 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1, UINT_MAX, 1);
2228 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1, UINT_MAX, 1);
2229 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
2230 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1, UINT_MAX, 0);
2231 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
2232 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
2233 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
2234 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1, UINT_MAX, 0);
2235 #undef STORE_FUNCTION
2237 #define CFQ_ATTR(name) \
2238 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
2240 static struct elv_fs_entry cfq_attrs[] = {
2243 CFQ_ATTR(fifo_expire_sync),
2244 CFQ_ATTR(fifo_expire_async),
2245 CFQ_ATTR(back_seek_max),
2246 CFQ_ATTR(back_seek_penalty),
2247 CFQ_ATTR(slice_sync),
2248 CFQ_ATTR(slice_async),
2249 CFQ_ATTR(slice_async_rq),
2250 CFQ_ATTR(slice_idle),
2254 static struct elevator_type iosched_cfq = {
2256 .elevator_merge_fn = cfq_merge,
2257 .elevator_merged_fn = cfq_merged_request,
2258 .elevator_merge_req_fn = cfq_merged_requests,
2259 .elevator_dispatch_fn = cfq_dispatch_requests,
2260 .elevator_add_req_fn = cfq_insert_request,
2261 .elevator_activate_req_fn = cfq_activate_request,
2262 .elevator_deactivate_req_fn = cfq_deactivate_request,
2263 .elevator_queue_empty_fn = cfq_queue_empty,
2264 .elevator_completed_req_fn = cfq_completed_request,
2265 .elevator_former_req_fn = elv_rb_former_request,
2266 .elevator_latter_req_fn = elv_rb_latter_request,
2267 .elevator_set_req_fn = cfq_set_request,
2268 .elevator_put_req_fn = cfq_put_request,
2269 .elevator_may_queue_fn = cfq_may_queue,
2270 .elevator_init_fn = cfq_init_queue,
2271 .elevator_exit_fn = cfq_exit_queue,
2274 .elevator_attrs = cfq_attrs,
2275 .elevator_name = "cfq",
2276 .elevator_owner = THIS_MODULE,
2279 static int __init cfq_init(void)
2284 * could be 0 on HZ < 1000 setups
2286 if (!cfq_slice_async)
2287 cfq_slice_async = 1;
2288 if (!cfq_slice_idle)
2291 if (cfq_slab_setup())
2294 ret = elv_register(&iosched_cfq);
2301 static void __exit cfq_exit(void)
2303 DECLARE_COMPLETION(all_gone);
2304 elv_unregister(&iosched_cfq);
2305 ioc_gone = &all_gone;
2306 /* ioc_gone's update must be visible before reading ioc_count */
2308 if (atomic_read(&ioc_count))
2309 wait_for_completion(ioc_gone);
2314 module_init(cfq_init);
2315 module_exit(cfq_exit);
2317 MODULE_AUTHOR("Jens Axboe");
2318 MODULE_LICENSE("GPL");
2319 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");