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_fifo_expire[2] = { HZ / 4, HZ / 8 };
21 static const int cfq_back_max = 16 * 1024; /* maximum backwards seek, in KiB */
22 static const int cfq_back_penalty = 2; /* penalty of a backwards seek */
24 static const int cfq_slice_sync = HZ / 10;
25 static int cfq_slice_async = HZ / 25;
26 static const int cfq_slice_async_rq = 2;
27 static int cfq_slice_idle = HZ / 125;
29 #define CFQ_IDLE_GRACE (HZ / 10)
30 #define CFQ_SLICE_SCALE (5)
32 #define CFQ_KEY_ASYNC (0)
35 * for the hash of cfqq inside the cfqd
37 #define CFQ_QHASH_SHIFT 6
38 #define CFQ_QHASH_ENTRIES (1 << CFQ_QHASH_SHIFT)
39 #define list_entry_qhash(entry) hlist_entry((entry), struct cfq_queue, cfq_hash)
41 #define list_entry_cfqq(ptr) list_entry((ptr), struct cfq_queue, cfq_list)
43 #define RQ_CIC(rq) ((struct cfq_io_context*)(rq)->elevator_private)
44 #define RQ_CFQQ(rq) ((rq)->elevator_private2)
46 static kmem_cache_t *cfq_pool;
47 static kmem_cache_t *cfq_ioc_pool;
49 static DEFINE_PER_CPU(unsigned long, ioc_count);
50 static struct completion *ioc_gone;
52 #define CFQ_PRIO_LISTS IOPRIO_BE_NR
53 #define cfq_class_idle(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
54 #define cfq_class_rt(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
59 #define cfq_cfqq_dispatched(cfqq) \
60 ((cfqq)->on_dispatch[ASYNC] + (cfqq)->on_dispatch[SYNC])
62 #define cfq_cfqq_class_sync(cfqq) ((cfqq)->key != CFQ_KEY_ASYNC)
64 #define cfq_cfqq_sync(cfqq) \
65 (cfq_cfqq_class_sync(cfqq) || (cfqq)->on_dispatch[SYNC])
67 #define sample_valid(samples) ((samples) > 80)
70 * Per block device queue structure
73 request_queue_t *queue;
76 * rr list of queues with requests and the count of them
78 struct list_head rr_list[CFQ_PRIO_LISTS];
79 struct list_head busy_rr;
80 struct list_head cur_rr;
81 struct list_head idle_rr;
82 unsigned int busy_queues;
85 * non-ordered list of empty cfqq's
87 struct list_head empty_list;
92 struct hlist_head *cfq_hash;
98 * idle window management
100 struct timer_list idle_slice_timer;
101 struct work_struct unplug_work;
103 struct cfq_queue *active_queue;
104 struct cfq_io_context *active_cic;
105 int cur_prio, cur_end_prio;
106 unsigned int dispatch_slice;
108 struct timer_list idle_class_timer;
110 sector_t last_sector;
111 unsigned long last_end_request;
114 * tunables, see top of file
116 unsigned int cfq_quantum;
117 unsigned int cfq_fifo_expire[2];
118 unsigned int cfq_back_penalty;
119 unsigned int cfq_back_max;
120 unsigned int cfq_slice[2];
121 unsigned int cfq_slice_async_rq;
122 unsigned int cfq_slice_idle;
124 struct list_head cic_list;
128 * Per process-grouping structure
131 /* reference count */
133 /* parent cfq_data */
134 struct cfq_data *cfqd;
135 /* cfqq lookup hash */
136 struct hlist_node cfq_hash;
139 /* on either rr or empty list of cfqd */
140 struct list_head cfq_list;
141 /* sorted list of pending requests */
142 struct rb_root sort_list;
143 /* if fifo isn't expired, next request to serve */
144 struct request *next_rq;
145 /* requests queued in sort_list */
147 /* currently allocated requests */
149 /* fifo list of requests in sort_list */
150 struct list_head fifo;
152 unsigned long slice_start;
153 unsigned long slice_end;
154 unsigned long slice_left;
156 /* number of requests that are on the dispatch list */
159 /* io prio of this group */
160 unsigned short ioprio, org_ioprio;
161 unsigned short ioprio_class, org_ioprio_class;
163 /* various state flags, see below */
167 enum cfqq_state_flags {
168 CFQ_CFQQ_FLAG_on_rr = 0,
169 CFQ_CFQQ_FLAG_wait_request,
170 CFQ_CFQQ_FLAG_must_alloc,
171 CFQ_CFQQ_FLAG_must_alloc_slice,
172 CFQ_CFQQ_FLAG_must_dispatch,
173 CFQ_CFQQ_FLAG_fifo_expire,
174 CFQ_CFQQ_FLAG_idle_window,
175 CFQ_CFQQ_FLAG_prio_changed,
176 CFQ_CFQQ_FLAG_queue_new,
179 #define CFQ_CFQQ_FNS(name) \
180 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \
182 cfqq->flags |= (1 << CFQ_CFQQ_FLAG_##name); \
184 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \
186 cfqq->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \
188 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \
190 return (cfqq->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \
194 CFQ_CFQQ_FNS(wait_request);
195 CFQ_CFQQ_FNS(must_alloc);
196 CFQ_CFQQ_FNS(must_alloc_slice);
197 CFQ_CFQQ_FNS(must_dispatch);
198 CFQ_CFQQ_FNS(fifo_expire);
199 CFQ_CFQQ_FNS(idle_window);
200 CFQ_CFQQ_FNS(prio_changed);
201 CFQ_CFQQ_FNS(queue_new);
204 static struct cfq_queue *cfq_find_cfq_hash(struct cfq_data *, unsigned int, unsigned short);
205 static void cfq_dispatch_insert(request_queue_t *, struct request *);
206 static struct cfq_queue *cfq_get_queue(struct cfq_data *cfqd, unsigned int key, struct task_struct *tsk, gfp_t gfp_mask);
209 * scheduler run of queue, if there are requests pending and no one in the
210 * driver that will restart queueing
212 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
214 if (cfqd->busy_queues)
215 kblockd_schedule_work(&cfqd->unplug_work);
218 static int cfq_queue_empty(request_queue_t *q)
220 struct cfq_data *cfqd = q->elevator->elevator_data;
222 return !cfqd->busy_queues;
225 static inline pid_t cfq_queue_pid(struct task_struct *task, int rw)
227 if (rw == READ || rw == WRITE_SYNC)
230 return CFQ_KEY_ASYNC;
234 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
235 * We choose the request that is closest to the head right now. Distance
236 * behind the head is penalized and only allowed to a certain extent.
238 static struct request *
239 cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2)
241 sector_t last, s1, s2, d1 = 0, d2 = 0;
242 unsigned long back_max;
243 #define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */
244 #define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */
245 unsigned wrap = 0; /* bit mask: requests behind the disk head? */
247 if (rq1 == NULL || rq1 == rq2)
252 if (rq_is_sync(rq1) && !rq_is_sync(rq2))
254 else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
260 last = cfqd->last_sector;
263 * by definition, 1KiB is 2 sectors
265 back_max = cfqd->cfq_back_max * 2;
268 * Strict one way elevator _except_ in the case where we allow
269 * short backward seeks which are biased as twice the cost of a
270 * similar forward seek.
274 else if (s1 + back_max >= last)
275 d1 = (last - s1) * cfqd->cfq_back_penalty;
277 wrap |= CFQ_RQ1_WRAP;
281 else if (s2 + back_max >= last)
282 d2 = (last - s2) * cfqd->cfq_back_penalty;
284 wrap |= CFQ_RQ2_WRAP;
286 /* Found required data */
289 * By doing switch() on the bit mask "wrap" we avoid having to
290 * check two variables for all permutations: --> faster!
293 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
309 case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
312 * Since both rqs are wrapped,
313 * start with the one that's further behind head
314 * (--> only *one* back seek required),
315 * since back seek takes more time than forward.
325 * would be nice to take fifo expire time into account as well
327 static struct request *
328 cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
329 struct request *last)
331 struct rb_node *rbnext = rb_next(&last->rb_node);
332 struct rb_node *rbprev = rb_prev(&last->rb_node);
333 struct request *next = NULL, *prev = NULL;
335 BUG_ON(RB_EMPTY_NODE(&last->rb_node));
338 prev = rb_entry_rq(rbprev);
341 next = rb_entry_rq(rbnext);
343 rbnext = rb_first(&cfqq->sort_list);
344 if (rbnext && rbnext != &last->rb_node)
345 next = rb_entry_rq(rbnext);
348 return cfq_choose_req(cfqd, next, prev);
351 static void cfq_resort_rr_list(struct cfq_queue *cfqq, int preempted)
353 struct cfq_data *cfqd = cfqq->cfqd;
354 struct list_head *list;
356 BUG_ON(!cfq_cfqq_on_rr(cfqq));
358 list_del(&cfqq->cfq_list);
360 if (cfq_class_rt(cfqq))
361 list = &cfqd->cur_rr;
362 else if (cfq_class_idle(cfqq))
363 list = &cfqd->idle_rr;
366 * if cfqq has requests in flight, don't allow it to be
367 * found in cfq_set_active_queue before it has finished them.
368 * this is done to increase fairness between a process that
369 * has lots of io pending vs one that only generates one
370 * sporadically or synchronously
372 if (cfq_cfqq_dispatched(cfqq))
373 list = &cfqd->busy_rr;
375 list = &cfqd->rr_list[cfqq->ioprio];
379 * If this queue was preempted or is new (never been serviced), let
380 * it be added first for fairness but beind other new queues.
381 * Otherwise, just add to the back of the list.
383 if (preempted || cfq_cfqq_queue_new(cfqq)) {
384 struct list_head *n = list;
385 struct cfq_queue *__cfqq;
387 while (n->next != list) {
388 __cfqq = list_entry_cfqq(n->next);
389 if (!cfq_cfqq_queue_new(__cfqq))
398 list_add_tail(&cfqq->cfq_list, list);
402 * add to busy list of queues for service, trying to be fair in ordering
403 * the pending list according to last request service
406 cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
408 BUG_ON(cfq_cfqq_on_rr(cfqq));
409 cfq_mark_cfqq_on_rr(cfqq);
412 cfq_resort_rr_list(cfqq, 0);
416 cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
418 BUG_ON(!cfq_cfqq_on_rr(cfqq));
419 cfq_clear_cfqq_on_rr(cfqq);
420 list_move(&cfqq->cfq_list, &cfqd->empty_list);
422 BUG_ON(!cfqd->busy_queues);
427 * rb tree support functions
429 static inline void cfq_del_rq_rb(struct request *rq)
431 struct cfq_queue *cfqq = RQ_CFQQ(rq);
432 struct cfq_data *cfqd = cfqq->cfqd;
433 const int sync = rq_is_sync(rq);
435 BUG_ON(!cfqq->queued[sync]);
436 cfqq->queued[sync]--;
438 elv_rb_del(&cfqq->sort_list, rq);
440 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
441 cfq_del_cfqq_rr(cfqd, cfqq);
444 static void cfq_add_rq_rb(struct request *rq)
446 struct cfq_queue *cfqq = RQ_CFQQ(rq);
447 struct cfq_data *cfqd = cfqq->cfqd;
448 struct request *__alias;
450 cfqq->queued[rq_is_sync(rq)]++;
453 * looks a little odd, but the first insert might return an alias.
454 * if that happens, put the alias on the dispatch list
456 while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL)
457 cfq_dispatch_insert(cfqd->queue, __alias);
461 cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
463 elv_rb_del(&cfqq->sort_list, rq);
464 cfqq->queued[rq_is_sync(rq)]--;
468 static struct request *
469 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
471 struct task_struct *tsk = current;
472 pid_t key = cfq_queue_pid(tsk, bio_data_dir(bio));
473 struct cfq_queue *cfqq;
475 cfqq = cfq_find_cfq_hash(cfqd, key, tsk->ioprio);
477 sector_t sector = bio->bi_sector + bio_sectors(bio);
479 return elv_rb_find(&cfqq->sort_list, sector);
485 static void cfq_activate_request(request_queue_t *q, struct request *rq)
487 struct cfq_data *cfqd = q->elevator->elevator_data;
489 cfqd->rq_in_driver++;
492 * If the depth is larger 1, it really could be queueing. But lets
493 * make the mark a little higher - idling could still be good for
494 * low queueing, and a low queueing number could also just indicate
495 * a SCSI mid layer like behaviour where limit+1 is often seen.
497 if (!cfqd->hw_tag && cfqd->rq_in_driver > 4)
501 static void cfq_deactivate_request(request_queue_t *q, struct request *rq)
503 struct cfq_data *cfqd = q->elevator->elevator_data;
505 WARN_ON(!cfqd->rq_in_driver);
506 cfqd->rq_in_driver--;
509 static void cfq_remove_request(struct request *rq)
511 struct cfq_queue *cfqq = RQ_CFQQ(rq);
513 if (cfqq->next_rq == rq)
514 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
516 list_del_init(&rq->queuelist);
521 cfq_merge(request_queue_t *q, struct request **req, struct bio *bio)
523 struct cfq_data *cfqd = q->elevator->elevator_data;
524 struct request *__rq;
526 __rq = cfq_find_rq_fmerge(cfqd, bio);
527 if (__rq && elv_rq_merge_ok(__rq, bio)) {
529 return ELEVATOR_FRONT_MERGE;
532 return ELEVATOR_NO_MERGE;
535 static void cfq_merged_request(request_queue_t *q, struct request *req,
538 if (type == ELEVATOR_FRONT_MERGE) {
539 struct cfq_queue *cfqq = RQ_CFQQ(req);
541 cfq_reposition_rq_rb(cfqq, req);
546 cfq_merged_requests(request_queue_t *q, struct request *rq,
547 struct request *next)
550 * reposition in fifo if next is older than rq
552 if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
553 time_before(next->start_time, rq->start_time))
554 list_move(&rq->queuelist, &next->queuelist);
556 cfq_remove_request(next);
560 __cfq_set_active_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
564 * stop potential idle class queues waiting service
566 del_timer(&cfqd->idle_class_timer);
568 cfqq->slice_start = jiffies;
570 cfqq->slice_left = 0;
571 cfq_clear_cfqq_must_alloc_slice(cfqq);
572 cfq_clear_cfqq_fifo_expire(cfqq);
575 cfqd->active_queue = cfqq;
579 * current cfqq expired its slice (or was too idle), select new one
582 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
585 unsigned long now = jiffies;
587 if (cfq_cfqq_wait_request(cfqq))
588 del_timer(&cfqd->idle_slice_timer);
590 if (!preempted && !cfq_cfqq_dispatched(cfqq))
591 cfq_schedule_dispatch(cfqd);
593 cfq_clear_cfqq_must_dispatch(cfqq);
594 cfq_clear_cfqq_wait_request(cfqq);
595 cfq_clear_cfqq_queue_new(cfqq);
598 * store what was left of this slice, if the queue idled out
601 if (time_after(cfqq->slice_end, now))
602 cfqq->slice_left = cfqq->slice_end - now;
604 cfqq->slice_left = 0;
606 if (cfq_cfqq_on_rr(cfqq))
607 cfq_resort_rr_list(cfqq, preempted);
609 if (cfqq == cfqd->active_queue)
610 cfqd->active_queue = NULL;
612 if (cfqd->active_cic) {
613 put_io_context(cfqd->active_cic->ioc);
614 cfqd->active_cic = NULL;
617 cfqd->dispatch_slice = 0;
620 static inline void cfq_slice_expired(struct cfq_data *cfqd, int preempted)
622 struct cfq_queue *cfqq = cfqd->active_queue;
625 __cfq_slice_expired(cfqd, cfqq, preempted);
638 static int cfq_get_next_prio_level(struct cfq_data *cfqd)
647 for (p = cfqd->cur_prio; p <= cfqd->cur_end_prio; p++) {
648 if (!list_empty(&cfqd->rr_list[p])) {
657 if (++cfqd->cur_end_prio == CFQ_PRIO_LISTS) {
658 cfqd->cur_end_prio = 0;
665 if (unlikely(prio == -1))
668 BUG_ON(prio >= CFQ_PRIO_LISTS);
670 list_splice_init(&cfqd->rr_list[prio], &cfqd->cur_rr);
672 cfqd->cur_prio = prio + 1;
673 if (cfqd->cur_prio > cfqd->cur_end_prio) {
674 cfqd->cur_end_prio = cfqd->cur_prio;
677 if (cfqd->cur_end_prio == CFQ_PRIO_LISTS) {
679 cfqd->cur_end_prio = 0;
685 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd)
687 struct cfq_queue *cfqq = NULL;
689 if (!list_empty(&cfqd->cur_rr) || cfq_get_next_prio_level(cfqd) != -1) {
691 * if current list is non-empty, grab first entry. if it is
692 * empty, get next prio level and grab first entry then if any
695 cfqq = list_entry_cfqq(cfqd->cur_rr.next);
696 } else if (!list_empty(&cfqd->busy_rr)) {
698 * If no new queues are available, check if the busy list has
699 * some before falling back to idle io.
701 cfqq = list_entry_cfqq(cfqd->busy_rr.next);
702 } else if (!list_empty(&cfqd->idle_rr)) {
704 * if we have idle queues and no rt or be queues had pending
705 * requests, either allow immediate service if the grace period
706 * has passed or arm the idle grace timer
708 unsigned long end = cfqd->last_end_request + CFQ_IDLE_GRACE;
710 if (time_after_eq(jiffies, end))
711 cfqq = list_entry_cfqq(cfqd->idle_rr.next);
713 mod_timer(&cfqd->idle_class_timer, end);
716 __cfq_set_active_queue(cfqd, cfqq);
720 #define CIC_SEEKY(cic) ((cic)->seek_mean > (128 * 1024))
722 static int cfq_arm_slice_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq)
725 struct cfq_io_context *cic;
728 WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
729 WARN_ON(cfqq != cfqd->active_queue);
732 * idle is disabled, either manually or by past process history
734 if (!cfqd->cfq_slice_idle)
736 if (!cfq_cfqq_idle_window(cfqq))
739 * task has exited, don't wait
741 cic = cfqd->active_cic;
742 if (!cic || !cic->ioc->task)
745 cfq_mark_cfqq_must_dispatch(cfqq);
746 cfq_mark_cfqq_wait_request(cfqq);
748 sl = min(cfqq->slice_end - 1, (unsigned long) cfqd->cfq_slice_idle);
751 * we don't want to idle for seeks, but we do want to allow
752 * fair distribution of slice time for a process doing back-to-back
753 * seeks. so allow a little bit of time for him to submit a new rq
755 if (sample_valid(cic->seek_samples) && CIC_SEEKY(cic))
756 sl = min(sl, msecs_to_jiffies(2));
758 mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
762 static void cfq_dispatch_insert(request_queue_t *q, struct request *rq)
764 struct cfq_data *cfqd = q->elevator->elevator_data;
765 struct cfq_queue *cfqq = RQ_CFQQ(rq);
767 cfq_remove_request(rq);
768 cfqq->on_dispatch[rq_is_sync(rq)]++;
769 elv_dispatch_sort(q, rq);
771 rq = list_entry(q->queue_head.prev, struct request, queuelist);
772 cfqd->last_sector = rq->sector + rq->nr_sectors;
776 * return expired entry, or NULL to just start from scratch in rbtree
778 static inline struct request *cfq_check_fifo(struct cfq_queue *cfqq)
780 struct cfq_data *cfqd = cfqq->cfqd;
784 if (cfq_cfqq_fifo_expire(cfqq))
786 if (list_empty(&cfqq->fifo))
789 fifo = cfq_cfqq_class_sync(cfqq);
790 rq = rq_entry_fifo(cfqq->fifo.next);
792 if (time_after(jiffies, rq->start_time + cfqd->cfq_fifo_expire[fifo])) {
793 cfq_mark_cfqq_fifo_expire(cfqq);
801 * Scale schedule slice based on io priority. Use the sync time slice only
802 * if a queue is marked sync and has sync io queued. A sync queue with async
803 * io only, should not get full sync slice length.
806 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
808 const int base_slice = cfqd->cfq_slice[cfq_cfqq_sync(cfqq)];
810 WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
812 return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - cfqq->ioprio));
816 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
818 cfqq->slice_end = cfq_prio_to_slice(cfqd, cfqq) + jiffies;
822 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
824 const int base_rq = cfqd->cfq_slice_async_rq;
826 WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
828 return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
832 * get next queue for service
834 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
836 unsigned long now = jiffies;
837 struct cfq_queue *cfqq;
839 cfqq = cfqd->active_queue;
846 if (!cfq_cfqq_must_dispatch(cfqq) && time_after(now, cfqq->slice_end))
850 * if queue has requests, dispatch one. if not, check if
851 * enough slice is left to wait for one
853 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
855 else if (cfq_cfqq_dispatched(cfqq)) {
858 } else if (cfq_cfqq_class_sync(cfqq)) {
859 if (cfq_arm_slice_timer(cfqd, cfqq))
864 cfq_slice_expired(cfqd, 0);
866 cfqq = cfq_set_active_queue(cfqd);
872 __cfq_dispatch_requests(struct cfq_data *cfqd, struct cfq_queue *cfqq,
877 BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
883 * follow expired path, else get first next available
885 if ((rq = cfq_check_fifo(cfqq)) == NULL)
889 * finally, insert request into driver dispatch list
891 cfq_dispatch_insert(cfqd->queue, rq);
893 cfqd->dispatch_slice++;
896 if (!cfqd->active_cic) {
897 atomic_inc(&RQ_CIC(rq)->ioc->refcount);
898 cfqd->active_cic = RQ_CIC(rq);
901 if (RB_EMPTY_ROOT(&cfqq->sort_list))
904 } while (dispatched < max_dispatch);
907 * if slice end isn't set yet, set it.
909 if (!cfqq->slice_end)
910 cfq_set_prio_slice(cfqd, cfqq);
913 * expire an async queue immediately if it has used up its slice. idle
914 * queue always expire after 1 dispatch round.
916 if ((!cfq_cfqq_sync(cfqq) &&
917 cfqd->dispatch_slice >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
918 cfq_class_idle(cfqq) ||
919 !cfq_cfqq_idle_window(cfqq))
920 cfq_slice_expired(cfqd, 0);
926 cfq_forced_dispatch_cfqqs(struct list_head *list)
928 struct cfq_queue *cfqq, *next;
932 list_for_each_entry_safe(cfqq, next, list, cfq_list) {
933 while (cfqq->next_rq) {
934 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
937 BUG_ON(!list_empty(&cfqq->fifo));
944 cfq_forced_dispatch(struct cfq_data *cfqd)
946 int i, dispatched = 0;
948 for (i = 0; i < CFQ_PRIO_LISTS; i++)
949 dispatched += cfq_forced_dispatch_cfqqs(&cfqd->rr_list[i]);
951 dispatched += cfq_forced_dispatch_cfqqs(&cfqd->busy_rr);
952 dispatched += cfq_forced_dispatch_cfqqs(&cfqd->cur_rr);
953 dispatched += cfq_forced_dispatch_cfqqs(&cfqd->idle_rr);
955 cfq_slice_expired(cfqd, 0);
957 BUG_ON(cfqd->busy_queues);
963 cfq_dispatch_requests(request_queue_t *q, int force)
965 struct cfq_data *cfqd = q->elevator->elevator_data;
966 struct cfq_queue *cfqq, *prev_cfqq;
969 if (!cfqd->busy_queues)
973 return cfq_forced_dispatch(cfqd);
977 while ((cfqq = cfq_select_queue(cfqd)) != NULL) {
981 * Don't repeat dispatch from the previous queue.
983 if (prev_cfqq == cfqq)
986 cfq_clear_cfqq_must_dispatch(cfqq);
987 cfq_clear_cfqq_wait_request(cfqq);
988 del_timer(&cfqd->idle_slice_timer);
990 max_dispatch = cfqd->cfq_quantum;
991 if (cfq_class_idle(cfqq))
994 dispatched += __cfq_dispatch_requests(cfqd, cfqq, max_dispatch);
997 * If the dispatch cfqq has idling enabled and is still
998 * the active queue, break out.
1000 if (cfq_cfqq_idle_window(cfqq) && cfqd->active_queue)
1010 * task holds one reference to the queue, dropped when task exits. each rq
1011 * in-flight on this queue also holds a reference, dropped when rq is freed.
1013 * queue lock must be held here.
1015 static void cfq_put_queue(struct cfq_queue *cfqq)
1017 struct cfq_data *cfqd = cfqq->cfqd;
1019 BUG_ON(atomic_read(&cfqq->ref) <= 0);
1021 if (!atomic_dec_and_test(&cfqq->ref))
1024 BUG_ON(rb_first(&cfqq->sort_list));
1025 BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
1026 BUG_ON(cfq_cfqq_on_rr(cfqq));
1028 if (unlikely(cfqd->active_queue == cfqq))
1029 __cfq_slice_expired(cfqd, cfqq, 0);
1032 * it's on the empty list and still hashed
1034 list_del(&cfqq->cfq_list);
1035 hlist_del(&cfqq->cfq_hash);
1036 kmem_cache_free(cfq_pool, cfqq);
1039 static struct cfq_queue *
1040 __cfq_find_cfq_hash(struct cfq_data *cfqd, unsigned int key, unsigned int prio,
1043 struct hlist_head *hash_list = &cfqd->cfq_hash[hashval];
1044 struct hlist_node *entry;
1045 struct cfq_queue *__cfqq;
1047 hlist_for_each_entry(__cfqq, entry, hash_list, cfq_hash) {
1048 const unsigned short __p = IOPRIO_PRIO_VALUE(__cfqq->org_ioprio_class, __cfqq->org_ioprio);
1050 if (__cfqq->key == key && (__p == prio || !prio))
1057 static struct cfq_queue *
1058 cfq_find_cfq_hash(struct cfq_data *cfqd, unsigned int key, unsigned short prio)
1060 return __cfq_find_cfq_hash(cfqd, key, prio, hash_long(key, CFQ_QHASH_SHIFT));
1063 static void cfq_free_io_context(struct io_context *ioc)
1065 struct cfq_io_context *__cic;
1069 while ((n = rb_first(&ioc->cic_root)) != NULL) {
1070 __cic = rb_entry(n, struct cfq_io_context, rb_node);
1071 rb_erase(&__cic->rb_node, &ioc->cic_root);
1072 kmem_cache_free(cfq_ioc_pool, __cic);
1076 elv_ioc_count_mod(ioc_count, -freed);
1078 if (ioc_gone && !elv_ioc_count_read(ioc_count))
1082 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1084 if (unlikely(cfqq == cfqd->active_queue))
1085 __cfq_slice_expired(cfqd, cfqq, 0);
1087 cfq_put_queue(cfqq);
1090 static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
1091 struct cfq_io_context *cic)
1093 list_del_init(&cic->queue_list);
1097 if (cic->cfqq[ASYNC]) {
1098 cfq_exit_cfqq(cfqd, cic->cfqq[ASYNC]);
1099 cic->cfqq[ASYNC] = NULL;
1102 if (cic->cfqq[SYNC]) {
1103 cfq_exit_cfqq(cfqd, cic->cfqq[SYNC]);
1104 cic->cfqq[SYNC] = NULL;
1110 * Called with interrupts disabled
1112 static void cfq_exit_single_io_context(struct cfq_io_context *cic)
1114 struct cfq_data *cfqd = cic->key;
1117 request_queue_t *q = cfqd->queue;
1119 spin_lock_irq(q->queue_lock);
1120 __cfq_exit_single_io_context(cfqd, cic);
1121 spin_unlock_irq(q->queue_lock);
1125 static void cfq_exit_io_context(struct io_context *ioc)
1127 struct cfq_io_context *__cic;
1131 * put the reference this task is holding to the various queues
1134 n = rb_first(&ioc->cic_root);
1136 __cic = rb_entry(n, struct cfq_io_context, rb_node);
1138 cfq_exit_single_io_context(__cic);
1143 static struct cfq_io_context *
1144 cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
1146 struct cfq_io_context *cic;
1148 cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask, cfqd->queue->node);
1150 memset(cic, 0, sizeof(*cic));
1151 cic->last_end_request = jiffies;
1152 INIT_LIST_HEAD(&cic->queue_list);
1153 cic->dtor = cfq_free_io_context;
1154 cic->exit = cfq_exit_io_context;
1155 elv_ioc_count_inc(ioc_count);
1161 static void cfq_init_prio_data(struct cfq_queue *cfqq)
1163 struct task_struct *tsk = current;
1166 if (!cfq_cfqq_prio_changed(cfqq))
1169 ioprio_class = IOPRIO_PRIO_CLASS(tsk->ioprio);
1170 switch (ioprio_class) {
1172 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
1173 case IOPRIO_CLASS_NONE:
1175 * no prio set, place us in the middle of the BE classes
1177 cfqq->ioprio = task_nice_ioprio(tsk);
1178 cfqq->ioprio_class = IOPRIO_CLASS_BE;
1180 case IOPRIO_CLASS_RT:
1181 cfqq->ioprio = task_ioprio(tsk);
1182 cfqq->ioprio_class = IOPRIO_CLASS_RT;
1184 case IOPRIO_CLASS_BE:
1185 cfqq->ioprio = task_ioprio(tsk);
1186 cfqq->ioprio_class = IOPRIO_CLASS_BE;
1188 case IOPRIO_CLASS_IDLE:
1189 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
1191 cfq_clear_cfqq_idle_window(cfqq);
1196 * keep track of original prio settings in case we have to temporarily
1197 * elevate the priority of this queue
1199 cfqq->org_ioprio = cfqq->ioprio;
1200 cfqq->org_ioprio_class = cfqq->ioprio_class;
1202 if (cfq_cfqq_on_rr(cfqq))
1203 cfq_resort_rr_list(cfqq, 0);
1205 cfq_clear_cfqq_prio_changed(cfqq);
1208 static inline void changed_ioprio(struct cfq_io_context *cic)
1210 struct cfq_data *cfqd = cic->key;
1211 struct cfq_queue *cfqq;
1213 if (unlikely(!cfqd))
1216 spin_lock(cfqd->queue->queue_lock);
1218 cfqq = cic->cfqq[ASYNC];
1220 struct cfq_queue *new_cfqq;
1221 new_cfqq = cfq_get_queue(cfqd, CFQ_KEY_ASYNC, cic->ioc->task,
1224 cic->cfqq[ASYNC] = new_cfqq;
1225 cfq_put_queue(cfqq);
1229 cfqq = cic->cfqq[SYNC];
1231 cfq_mark_cfqq_prio_changed(cfqq);
1233 spin_unlock(cfqd->queue->queue_lock);
1236 static void cfq_ioc_set_ioprio(struct io_context *ioc)
1238 struct cfq_io_context *cic;
1241 ioc->ioprio_changed = 0;
1243 n = rb_first(&ioc->cic_root);
1245 cic = rb_entry(n, struct cfq_io_context, rb_node);
1247 changed_ioprio(cic);
1252 static struct cfq_queue *
1253 cfq_get_queue(struct cfq_data *cfqd, unsigned int key, struct task_struct *tsk,
1256 const int hashval = hash_long(key, CFQ_QHASH_SHIFT);
1257 struct cfq_queue *cfqq, *new_cfqq = NULL;
1258 unsigned short ioprio;
1261 ioprio = tsk->ioprio;
1262 cfqq = __cfq_find_cfq_hash(cfqd, key, ioprio, hashval);
1268 } else if (gfp_mask & __GFP_WAIT) {
1270 * Inform the allocator of the fact that we will
1271 * just repeat this allocation if it fails, to allow
1272 * the allocator to do whatever it needs to attempt to
1275 spin_unlock_irq(cfqd->queue->queue_lock);
1276 new_cfqq = kmem_cache_alloc_node(cfq_pool, gfp_mask|__GFP_NOFAIL, cfqd->queue->node);
1277 spin_lock_irq(cfqd->queue->queue_lock);
1280 cfqq = kmem_cache_alloc_node(cfq_pool, gfp_mask, cfqd->queue->node);
1285 memset(cfqq, 0, sizeof(*cfqq));
1287 INIT_HLIST_NODE(&cfqq->cfq_hash);
1288 INIT_LIST_HEAD(&cfqq->cfq_list);
1289 INIT_LIST_HEAD(&cfqq->fifo);
1292 hlist_add_head(&cfqq->cfq_hash, &cfqd->cfq_hash[hashval]);
1293 atomic_set(&cfqq->ref, 0);
1296 * set ->slice_left to allow preemption for a new process
1298 cfqq->slice_left = 2 * cfqd->cfq_slice_idle;
1299 cfq_mark_cfqq_idle_window(cfqq);
1300 cfq_mark_cfqq_prio_changed(cfqq);
1301 cfq_mark_cfqq_queue_new(cfqq);
1302 cfq_init_prio_data(cfqq);
1306 kmem_cache_free(cfq_pool, new_cfqq);
1308 atomic_inc(&cfqq->ref);
1310 WARN_ON((gfp_mask & __GFP_WAIT) && !cfqq);
1315 cfq_drop_dead_cic(struct io_context *ioc, struct cfq_io_context *cic)
1317 WARN_ON(!list_empty(&cic->queue_list));
1318 rb_erase(&cic->rb_node, &ioc->cic_root);
1319 kmem_cache_free(cfq_ioc_pool, cic);
1320 elv_ioc_count_dec(ioc_count);
1323 static struct cfq_io_context *
1324 cfq_cic_rb_lookup(struct cfq_data *cfqd, struct io_context *ioc)
1327 struct cfq_io_context *cic;
1328 void *k, *key = cfqd;
1331 n = ioc->cic_root.rb_node;
1333 cic = rb_entry(n, struct cfq_io_context, rb_node);
1334 /* ->key must be copied to avoid race with cfq_exit_queue() */
1337 cfq_drop_dead_cic(ioc, cic);
1353 cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
1354 struct cfq_io_context *cic)
1357 struct rb_node *parent;
1358 struct cfq_io_context *__cic;
1366 p = &ioc->cic_root.rb_node;
1369 __cic = rb_entry(parent, struct cfq_io_context, rb_node);
1370 /* ->key must be copied to avoid race with cfq_exit_queue() */
1373 cfq_drop_dead_cic(ioc, __cic);
1379 else if (cic->key > k)
1380 p = &(*p)->rb_right;
1385 rb_link_node(&cic->rb_node, parent, p);
1386 rb_insert_color(&cic->rb_node, &ioc->cic_root);
1388 spin_lock_irq(cfqd->queue->queue_lock);
1389 list_add(&cic->queue_list, &cfqd->cic_list);
1390 spin_unlock_irq(cfqd->queue->queue_lock);
1394 * Setup general io context and cfq io context. There can be several cfq
1395 * io contexts per general io context, if this process is doing io to more
1396 * than one device managed by cfq.
1398 static struct cfq_io_context *
1399 cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
1401 struct io_context *ioc = NULL;
1402 struct cfq_io_context *cic;
1404 might_sleep_if(gfp_mask & __GFP_WAIT);
1406 ioc = get_io_context(gfp_mask, cfqd->queue->node);
1410 cic = cfq_cic_rb_lookup(cfqd, ioc);
1414 cic = cfq_alloc_io_context(cfqd, gfp_mask);
1418 cfq_cic_link(cfqd, ioc, cic);
1420 smp_read_barrier_depends();
1421 if (unlikely(ioc->ioprio_changed))
1422 cfq_ioc_set_ioprio(ioc);
1426 put_io_context(ioc);
1431 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
1433 unsigned long elapsed, ttime;
1436 * if this context already has stuff queued, thinktime is from
1437 * last queue not last end
1440 if (time_after(cic->last_end_request, cic->last_queue))
1441 elapsed = jiffies - cic->last_end_request;
1443 elapsed = jiffies - cic->last_queue;
1445 elapsed = jiffies - cic->last_end_request;
1448 ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
1450 cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
1451 cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
1452 cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
1456 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_io_context *cic,
1462 if (cic->last_request_pos < rq->sector)
1463 sdist = rq->sector - cic->last_request_pos;
1465 sdist = cic->last_request_pos - rq->sector;
1468 * Don't allow the seek distance to get too large from the
1469 * odd fragment, pagein, etc
1471 if (cic->seek_samples <= 60) /* second&third seek */
1472 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*1024);
1474 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*64);
1476 cic->seek_samples = (7*cic->seek_samples + 256) / 8;
1477 cic->seek_total = (7*cic->seek_total + (u64)256*sdist) / 8;
1478 total = cic->seek_total + (cic->seek_samples/2);
1479 do_div(total, cic->seek_samples);
1480 cic->seek_mean = (sector_t)total;
1484 * Disable idle window if the process thinks too long or seeks so much that
1488 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1489 struct cfq_io_context *cic)
1491 int enable_idle = cfq_cfqq_idle_window(cfqq);
1493 if (!cic->ioc->task || !cfqd->cfq_slice_idle ||
1494 (cfqd->hw_tag && CIC_SEEKY(cic)))
1496 else if (sample_valid(cic->ttime_samples)) {
1497 if (cic->ttime_mean > cfqd->cfq_slice_idle)
1504 cfq_mark_cfqq_idle_window(cfqq);
1506 cfq_clear_cfqq_idle_window(cfqq);
1511 * Check if new_cfqq should preempt the currently active queue. Return 0 for
1512 * no or if we aren't sure, a 1 will cause a preempt.
1515 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
1518 struct cfq_queue *cfqq = cfqd->active_queue;
1520 if (cfq_class_idle(new_cfqq))
1526 if (cfq_class_idle(cfqq))
1528 if (!cfq_cfqq_wait_request(new_cfqq))
1531 * if it doesn't have slice left, forget it
1533 if (new_cfqq->slice_left < cfqd->cfq_slice_idle)
1535 if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
1542 * cfqq preempts the active queue. if we allowed preempt with no slice left,
1543 * let it have half of its nominal slice.
1545 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1547 struct cfq_queue *__cfqq, *next;
1549 list_for_each_entry_safe(__cfqq, next, &cfqd->cur_rr, cfq_list)
1550 cfq_resort_rr_list(__cfqq, 1);
1552 if (!cfqq->slice_left)
1553 cfqq->slice_left = cfq_prio_to_slice(cfqd, cfqq) / 2;
1555 cfqq->slice_end = cfqq->slice_left + jiffies;
1556 cfq_slice_expired(cfqd, 1);
1557 __cfq_set_active_queue(cfqd, cfqq);
1561 * should really be a ll_rw_blk.c helper
1563 static void cfq_start_queueing(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1565 request_queue_t *q = cfqd->queue;
1567 if (!blk_queue_plugged(q))
1570 __generic_unplug_device(q);
1574 * Called when a new fs request (rq) is added (to cfqq). Check if there's
1575 * something we should do about it
1578 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1581 struct cfq_io_context *cic = RQ_CIC(rq);
1584 * check if this request is a better next-serve candidate)) {
1586 cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq);
1587 BUG_ON(!cfqq->next_rq);
1590 * we never wait for an async request and we don't allow preemption
1591 * of an async request. so just return early
1593 if (!rq_is_sync(rq)) {
1595 * sync process issued an async request, if it's waiting
1596 * then expire it and kick rq handling.
1598 if (cic == cfqd->active_cic &&
1599 del_timer(&cfqd->idle_slice_timer)) {
1600 cfq_slice_expired(cfqd, 0);
1601 cfq_start_queueing(cfqd, cfqq);
1606 cfq_update_io_thinktime(cfqd, cic);
1607 cfq_update_io_seektime(cfqd, cic, rq);
1608 cfq_update_idle_window(cfqd, cfqq, cic);
1610 cic->last_queue = jiffies;
1611 cic->last_request_pos = rq->sector + rq->nr_sectors;
1613 if (cfqq == cfqd->active_queue) {
1615 * if we are waiting for a request for this queue, let it rip
1616 * immediately and flag that we must not expire this queue
1619 if (cfq_cfqq_wait_request(cfqq)) {
1620 cfq_mark_cfqq_must_dispatch(cfqq);
1621 del_timer(&cfqd->idle_slice_timer);
1622 cfq_start_queueing(cfqd, cfqq);
1624 } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
1626 * not the active queue - expire current slice if it is
1627 * idle and has expired it's mean thinktime or this new queue
1628 * has some old slice time left and is of higher priority
1630 cfq_preempt_queue(cfqd, cfqq);
1631 cfq_mark_cfqq_must_dispatch(cfqq);
1632 cfq_start_queueing(cfqd, cfqq);
1636 static void cfq_insert_request(request_queue_t *q, struct request *rq)
1638 struct cfq_data *cfqd = q->elevator->elevator_data;
1639 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1641 cfq_init_prio_data(cfqq);
1645 if (!cfq_cfqq_on_rr(cfqq))
1646 cfq_add_cfqq_rr(cfqd, cfqq);
1648 list_add_tail(&rq->queuelist, &cfqq->fifo);
1650 cfq_rq_enqueued(cfqd, cfqq, rq);
1653 static void cfq_completed_request(request_queue_t *q, struct request *rq)
1655 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1656 struct cfq_data *cfqd = cfqq->cfqd;
1657 const int sync = rq_is_sync(rq);
1662 WARN_ON(!cfqd->rq_in_driver);
1663 WARN_ON(!cfqq->on_dispatch[sync]);
1664 cfqd->rq_in_driver--;
1665 cfqq->on_dispatch[sync]--;
1667 if (!cfq_class_idle(cfqq))
1668 cfqd->last_end_request = now;
1670 if (!cfq_cfqq_dispatched(cfqq) && cfq_cfqq_on_rr(cfqq))
1671 cfq_resort_rr_list(cfqq, 0);
1674 RQ_CIC(rq)->last_end_request = now;
1677 * If this is the active queue, check if it needs to be expired,
1678 * or if we want to idle in case it has no pending requests.
1680 if (cfqd->active_queue == cfqq) {
1681 if (time_after(now, cfqq->slice_end))
1682 cfq_slice_expired(cfqd, 0);
1683 else if (sync && RB_EMPTY_ROOT(&cfqq->sort_list)) {
1684 if (!cfq_arm_slice_timer(cfqd, cfqq))
1685 cfq_schedule_dispatch(cfqd);
1691 * we temporarily boost lower priority queues if they are holding fs exclusive
1692 * resources. they are boosted to normal prio (CLASS_BE/4)
1694 static void cfq_prio_boost(struct cfq_queue *cfqq)
1696 const int ioprio_class = cfqq->ioprio_class;
1697 const int ioprio = cfqq->ioprio;
1699 if (has_fs_excl()) {
1701 * boost idle prio on transactions that would lock out other
1702 * users of the filesystem
1704 if (cfq_class_idle(cfqq))
1705 cfqq->ioprio_class = IOPRIO_CLASS_BE;
1706 if (cfqq->ioprio > IOPRIO_NORM)
1707 cfqq->ioprio = IOPRIO_NORM;
1710 * check if we need to unboost the queue
1712 if (cfqq->ioprio_class != cfqq->org_ioprio_class)
1713 cfqq->ioprio_class = cfqq->org_ioprio_class;
1714 if (cfqq->ioprio != cfqq->org_ioprio)
1715 cfqq->ioprio = cfqq->org_ioprio;
1719 * refile between round-robin lists if we moved the priority class
1721 if ((ioprio_class != cfqq->ioprio_class || ioprio != cfqq->ioprio) &&
1722 cfq_cfqq_on_rr(cfqq))
1723 cfq_resort_rr_list(cfqq, 0);
1726 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
1728 if ((cfq_cfqq_wait_request(cfqq) || cfq_cfqq_must_alloc(cfqq)) &&
1729 !cfq_cfqq_must_alloc_slice(cfqq)) {
1730 cfq_mark_cfqq_must_alloc_slice(cfqq);
1731 return ELV_MQUEUE_MUST;
1734 return ELV_MQUEUE_MAY;
1737 static int cfq_may_queue(request_queue_t *q, int rw)
1739 struct cfq_data *cfqd = q->elevator->elevator_data;
1740 struct task_struct *tsk = current;
1741 struct cfq_queue *cfqq;
1744 * don't force setup of a queue from here, as a call to may_queue
1745 * does not necessarily imply that a request actually will be queued.
1746 * so just lookup a possibly existing queue, or return 'may queue'
1749 cfqq = cfq_find_cfq_hash(cfqd, cfq_queue_pid(tsk, rw), tsk->ioprio);
1751 cfq_init_prio_data(cfqq);
1752 cfq_prio_boost(cfqq);
1754 return __cfq_may_queue(cfqq);
1757 return ELV_MQUEUE_MAY;
1761 * queue lock held here
1763 static void cfq_put_request(request_queue_t *q, struct request *rq)
1765 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1768 const int rw = rq_data_dir(rq);
1770 BUG_ON(!cfqq->allocated[rw]);
1771 cfqq->allocated[rw]--;
1773 put_io_context(RQ_CIC(rq)->ioc);
1775 rq->elevator_private = NULL;
1776 rq->elevator_private2 = NULL;
1778 cfq_put_queue(cfqq);
1783 * Allocate cfq data structures associated with this request.
1786 cfq_set_request(request_queue_t *q, struct request *rq, gfp_t gfp_mask)
1788 struct cfq_data *cfqd = q->elevator->elevator_data;
1789 struct task_struct *tsk = current;
1790 struct cfq_io_context *cic;
1791 const int rw = rq_data_dir(rq);
1792 pid_t key = cfq_queue_pid(tsk, rw);
1793 struct cfq_queue *cfqq;
1794 unsigned long flags;
1795 int is_sync = key != CFQ_KEY_ASYNC;
1797 might_sleep_if(gfp_mask & __GFP_WAIT);
1799 cic = cfq_get_io_context(cfqd, gfp_mask);
1801 spin_lock_irqsave(q->queue_lock, flags);
1806 if (!cic->cfqq[is_sync]) {
1807 cfqq = cfq_get_queue(cfqd, key, tsk, gfp_mask);
1811 cic->cfqq[is_sync] = cfqq;
1813 cfqq = cic->cfqq[is_sync];
1815 cfqq->allocated[rw]++;
1816 cfq_clear_cfqq_must_alloc(cfqq);
1817 atomic_inc(&cfqq->ref);
1819 spin_unlock_irqrestore(q->queue_lock, flags);
1821 rq->elevator_private = cic;
1822 rq->elevator_private2 = cfqq;
1827 put_io_context(cic->ioc);
1829 cfq_schedule_dispatch(cfqd);
1830 spin_unlock_irqrestore(q->queue_lock, flags);
1834 static void cfq_kick_queue(void *data)
1836 request_queue_t *q = data;
1837 unsigned long flags;
1839 spin_lock_irqsave(q->queue_lock, flags);
1842 spin_unlock_irqrestore(q->queue_lock, flags);
1846 * Timer running if the active_queue is currently idling inside its time slice
1848 static void cfq_idle_slice_timer(unsigned long data)
1850 struct cfq_data *cfqd = (struct cfq_data *) data;
1851 struct cfq_queue *cfqq;
1852 unsigned long flags;
1854 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1856 if ((cfqq = cfqd->active_queue) != NULL) {
1857 unsigned long now = jiffies;
1862 if (time_after(now, cfqq->slice_end))
1866 * only expire and reinvoke request handler, if there are
1867 * other queues with pending requests
1869 if (!cfqd->busy_queues)
1873 * not expired and it has a request pending, let it dispatch
1875 if (!RB_EMPTY_ROOT(&cfqq->sort_list)) {
1876 cfq_mark_cfqq_must_dispatch(cfqq);
1881 cfq_slice_expired(cfqd, 0);
1883 cfq_schedule_dispatch(cfqd);
1885 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1889 * Timer running if an idle class queue is waiting for service
1891 static void cfq_idle_class_timer(unsigned long data)
1893 struct cfq_data *cfqd = (struct cfq_data *) data;
1894 unsigned long flags, end;
1896 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1899 * race with a non-idle queue, reset timer
1901 end = cfqd->last_end_request + CFQ_IDLE_GRACE;
1902 if (!time_after_eq(jiffies, end))
1903 mod_timer(&cfqd->idle_class_timer, end);
1905 cfq_schedule_dispatch(cfqd);
1907 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1910 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
1912 del_timer_sync(&cfqd->idle_slice_timer);
1913 del_timer_sync(&cfqd->idle_class_timer);
1914 blk_sync_queue(cfqd->queue);
1917 static void cfq_exit_queue(elevator_t *e)
1919 struct cfq_data *cfqd = e->elevator_data;
1920 request_queue_t *q = cfqd->queue;
1922 cfq_shutdown_timer_wq(cfqd);
1924 spin_lock_irq(q->queue_lock);
1926 if (cfqd->active_queue)
1927 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
1929 while (!list_empty(&cfqd->cic_list)) {
1930 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
1931 struct cfq_io_context,
1934 __cfq_exit_single_io_context(cfqd, cic);
1937 spin_unlock_irq(q->queue_lock);
1939 cfq_shutdown_timer_wq(cfqd);
1941 kfree(cfqd->cfq_hash);
1945 static void *cfq_init_queue(request_queue_t *q, elevator_t *e)
1947 struct cfq_data *cfqd;
1950 cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL, q->node);
1954 memset(cfqd, 0, sizeof(*cfqd));
1956 for (i = 0; i < CFQ_PRIO_LISTS; i++)
1957 INIT_LIST_HEAD(&cfqd->rr_list[i]);
1959 INIT_LIST_HEAD(&cfqd->busy_rr);
1960 INIT_LIST_HEAD(&cfqd->cur_rr);
1961 INIT_LIST_HEAD(&cfqd->idle_rr);
1962 INIT_LIST_HEAD(&cfqd->empty_list);
1963 INIT_LIST_HEAD(&cfqd->cic_list);
1965 cfqd->cfq_hash = kmalloc_node(sizeof(struct hlist_head) * CFQ_QHASH_ENTRIES, GFP_KERNEL, q->node);
1966 if (!cfqd->cfq_hash)
1969 for (i = 0; i < CFQ_QHASH_ENTRIES; i++)
1970 INIT_HLIST_HEAD(&cfqd->cfq_hash[i]);
1974 init_timer(&cfqd->idle_slice_timer);
1975 cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
1976 cfqd->idle_slice_timer.data = (unsigned long) cfqd;
1978 init_timer(&cfqd->idle_class_timer);
1979 cfqd->idle_class_timer.function = cfq_idle_class_timer;
1980 cfqd->idle_class_timer.data = (unsigned long) cfqd;
1982 INIT_WORK(&cfqd->unplug_work, cfq_kick_queue, q);
1984 cfqd->cfq_quantum = cfq_quantum;
1985 cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
1986 cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
1987 cfqd->cfq_back_max = cfq_back_max;
1988 cfqd->cfq_back_penalty = cfq_back_penalty;
1989 cfqd->cfq_slice[0] = cfq_slice_async;
1990 cfqd->cfq_slice[1] = cfq_slice_sync;
1991 cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
1992 cfqd->cfq_slice_idle = cfq_slice_idle;
2000 static void cfq_slab_kill(void)
2003 kmem_cache_destroy(cfq_pool);
2005 kmem_cache_destroy(cfq_ioc_pool);
2008 static int __init cfq_slab_setup(void)
2010 cfq_pool = kmem_cache_create("cfq_pool", sizeof(struct cfq_queue), 0, 0,
2015 cfq_ioc_pool = kmem_cache_create("cfq_ioc_pool",
2016 sizeof(struct cfq_io_context), 0, 0, NULL, NULL);
2027 * sysfs parts below -->
2031 cfq_var_show(unsigned int var, char *page)
2033 return sprintf(page, "%d\n", var);
2037 cfq_var_store(unsigned int *var, const char *page, size_t count)
2039 char *p = (char *) page;
2041 *var = simple_strtoul(p, &p, 10);
2045 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
2046 static ssize_t __FUNC(elevator_t *e, char *page) \
2048 struct cfq_data *cfqd = e->elevator_data; \
2049 unsigned int __data = __VAR; \
2051 __data = jiffies_to_msecs(__data); \
2052 return cfq_var_show(__data, (page)); \
2054 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
2055 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
2056 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
2057 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
2058 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
2059 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
2060 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
2061 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
2062 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
2063 #undef SHOW_FUNCTION
2065 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
2066 static ssize_t __FUNC(elevator_t *e, const char *page, size_t count) \
2068 struct cfq_data *cfqd = e->elevator_data; \
2069 unsigned int __data; \
2070 int ret = cfq_var_store(&__data, (page), count); \
2071 if (__data < (MIN)) \
2073 else if (__data > (MAX)) \
2076 *(__PTR) = msecs_to_jiffies(__data); \
2078 *(__PTR) = __data; \
2081 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
2082 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1, UINT_MAX, 1);
2083 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1, UINT_MAX, 1);
2084 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
2085 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1, UINT_MAX, 0);
2086 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
2087 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
2088 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
2089 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1, UINT_MAX, 0);
2090 #undef STORE_FUNCTION
2092 #define CFQ_ATTR(name) \
2093 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
2095 static struct elv_fs_entry cfq_attrs[] = {
2097 CFQ_ATTR(fifo_expire_sync),
2098 CFQ_ATTR(fifo_expire_async),
2099 CFQ_ATTR(back_seek_max),
2100 CFQ_ATTR(back_seek_penalty),
2101 CFQ_ATTR(slice_sync),
2102 CFQ_ATTR(slice_async),
2103 CFQ_ATTR(slice_async_rq),
2104 CFQ_ATTR(slice_idle),
2108 static struct elevator_type iosched_cfq = {
2110 .elevator_merge_fn = cfq_merge,
2111 .elevator_merged_fn = cfq_merged_request,
2112 .elevator_merge_req_fn = cfq_merged_requests,
2113 .elevator_dispatch_fn = cfq_dispatch_requests,
2114 .elevator_add_req_fn = cfq_insert_request,
2115 .elevator_activate_req_fn = cfq_activate_request,
2116 .elevator_deactivate_req_fn = cfq_deactivate_request,
2117 .elevator_queue_empty_fn = cfq_queue_empty,
2118 .elevator_completed_req_fn = cfq_completed_request,
2119 .elevator_former_req_fn = elv_rb_former_request,
2120 .elevator_latter_req_fn = elv_rb_latter_request,
2121 .elevator_set_req_fn = cfq_set_request,
2122 .elevator_put_req_fn = cfq_put_request,
2123 .elevator_may_queue_fn = cfq_may_queue,
2124 .elevator_init_fn = cfq_init_queue,
2125 .elevator_exit_fn = cfq_exit_queue,
2126 .trim = cfq_free_io_context,
2128 .elevator_attrs = cfq_attrs,
2129 .elevator_name = "cfq",
2130 .elevator_owner = THIS_MODULE,
2133 static int __init cfq_init(void)
2138 * could be 0 on HZ < 1000 setups
2140 if (!cfq_slice_async)
2141 cfq_slice_async = 1;
2142 if (!cfq_slice_idle)
2145 if (cfq_slab_setup())
2148 ret = elv_register(&iosched_cfq);
2155 static void __exit cfq_exit(void)
2157 DECLARE_COMPLETION(all_gone);
2158 elv_unregister(&iosched_cfq);
2159 ioc_gone = &all_gone;
2160 /* ioc_gone's update must be visible before reading ioc_count */
2162 if (elv_ioc_count_read(ioc_count))
2163 wait_for_completion(ioc_gone);
2168 module_init(cfq_init);
2169 module_exit(cfq_exit);
2171 MODULE_AUTHOR("Jens Axboe");
2172 MODULE_LICENSE("GPL");
2173 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");