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@kernel.dk>
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 struct kmem_cache *cfq_pool;
47 static struct kmem_cache *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;
87 struct hlist_head *cfq_hash;
93 * idle window management
95 struct timer_list idle_slice_timer;
96 struct work_struct unplug_work;
98 struct cfq_queue *active_queue;
99 struct cfq_io_context *active_cic;
100 int cur_prio, cur_end_prio;
101 unsigned int dispatch_slice;
103 struct timer_list idle_class_timer;
105 sector_t last_sector;
106 unsigned long last_end_request;
109 * tunables, see top of file
111 unsigned int cfq_quantum;
112 unsigned int cfq_fifo_expire[2];
113 unsigned int cfq_back_penalty;
114 unsigned int cfq_back_max;
115 unsigned int cfq_slice[2];
116 unsigned int cfq_slice_async_rq;
117 unsigned int cfq_slice_idle;
119 struct list_head cic_list;
123 * Per process-grouping structure
126 /* reference count */
128 /* parent cfq_data */
129 struct cfq_data *cfqd;
130 /* cfqq lookup hash */
131 struct hlist_node cfq_hash;
134 /* member of the rr/busy/cur/idle cfqd list */
135 struct list_head cfq_list;
136 /* sorted list of pending requests */
137 struct rb_root sort_list;
138 /* if fifo isn't expired, next request to serve */
139 struct request *next_rq;
140 /* requests queued in sort_list */
142 /* currently allocated requests */
144 /* pending metadata requests */
146 /* fifo list of requests in sort_list */
147 struct list_head fifo;
149 unsigned long slice_end;
150 unsigned long service_last;
153 /* number of requests that are on the dispatch list */
156 /* io prio of this group */
157 unsigned short ioprio, org_ioprio;
158 unsigned short ioprio_class, org_ioprio_class;
160 /* various state flags, see below */
164 enum cfqq_state_flags {
165 CFQ_CFQQ_FLAG_on_rr = 0, /* on round-robin busy list */
166 CFQ_CFQQ_FLAG_wait_request, /* waiting for a request */
167 CFQ_CFQQ_FLAG_must_alloc, /* must be allowed rq alloc */
168 CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
169 CFQ_CFQQ_FLAG_must_dispatch, /* must dispatch, even if expired */
170 CFQ_CFQQ_FLAG_fifo_expire, /* FIFO checked in this slice */
171 CFQ_CFQQ_FLAG_idle_window, /* slice idling enabled */
172 CFQ_CFQQ_FLAG_prio_changed, /* task priority has changed */
173 CFQ_CFQQ_FLAG_queue_new, /* queue never been serviced */
174 CFQ_CFQQ_FLAG_slice_new, /* no requests dispatched in slice */
177 #define CFQ_CFQQ_FNS(name) \
178 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \
180 cfqq->flags |= (1 << CFQ_CFQQ_FLAG_##name); \
182 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \
184 cfqq->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \
186 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \
188 return (cfqq->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \
192 CFQ_CFQQ_FNS(wait_request);
193 CFQ_CFQQ_FNS(must_alloc);
194 CFQ_CFQQ_FNS(must_alloc_slice);
195 CFQ_CFQQ_FNS(must_dispatch);
196 CFQ_CFQQ_FNS(fifo_expire);
197 CFQ_CFQQ_FNS(idle_window);
198 CFQ_CFQQ_FNS(prio_changed);
199 CFQ_CFQQ_FNS(queue_new);
200 CFQ_CFQQ_FNS(slice_new);
203 static struct cfq_queue *cfq_find_cfq_hash(struct cfq_data *, unsigned int, unsigned short);
204 static void cfq_dispatch_insert(request_queue_t *, struct request *);
205 static struct cfq_queue *cfq_get_queue(struct cfq_data *cfqd, unsigned int key, struct task_struct *tsk, gfp_t gfp_mask);
208 * scheduler run of queue, if there are requests pending and no one in the
209 * driver that will restart queueing
211 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
213 if (cfqd->busy_queues)
214 kblockd_schedule_work(&cfqd->unplug_work);
217 static int cfq_queue_empty(request_queue_t *q)
219 struct cfq_data *cfqd = q->elevator->elevator_data;
221 return !cfqd->busy_queues;
224 static inline pid_t cfq_queue_pid(struct task_struct *task, int rw, int is_sync)
227 * Use the per-process queue, for read requests and syncronous writes
229 if (!(rw & REQ_RW) || is_sync)
232 return CFQ_KEY_ASYNC;
236 * Scale schedule slice based on io priority. Use the sync time slice only
237 * if a queue is marked sync and has sync io queued. A sync queue with async
238 * io only, should not get full sync slice length.
241 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
243 const int base_slice = cfqd->cfq_slice[cfq_cfqq_sync(cfqq)];
245 WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
247 return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - cfqq->ioprio));
251 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
253 cfqq->slice_end = cfq_prio_to_slice(cfqd, cfqq) + jiffies;
254 cfqq->slice_end += cfqq->slice_resid;
257 * Don't carry over residual for more than one slice, we only want
258 * to slightly correct the fairness. Carrying over forever would
259 * easily introduce oscillations.
261 cfqq->slice_resid = 0;
265 * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
266 * isn't valid until the first request from the dispatch is activated
267 * and the slice time set.
269 static inline int cfq_slice_used(struct cfq_queue *cfqq)
271 if (cfq_cfqq_slice_new(cfqq))
273 if (time_before(jiffies, cfqq->slice_end))
280 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
281 * We choose the request that is closest to the head right now. Distance
282 * behind the head is penalized and only allowed to a certain extent.
284 static struct request *
285 cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2)
287 sector_t last, s1, s2, d1 = 0, d2 = 0;
288 unsigned long back_max;
289 #define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */
290 #define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */
291 unsigned wrap = 0; /* bit mask: requests behind the disk head? */
293 if (rq1 == NULL || rq1 == rq2)
298 if (rq_is_sync(rq1) && !rq_is_sync(rq2))
300 else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
302 if (rq_is_meta(rq1) && !rq_is_meta(rq2))
304 else if (rq_is_meta(rq2) && !rq_is_meta(rq1))
310 last = cfqd->last_sector;
313 * by definition, 1KiB is 2 sectors
315 back_max = cfqd->cfq_back_max * 2;
318 * Strict one way elevator _except_ in the case where we allow
319 * short backward seeks which are biased as twice the cost of a
320 * similar forward seek.
324 else if (s1 + back_max >= last)
325 d1 = (last - s1) * cfqd->cfq_back_penalty;
327 wrap |= CFQ_RQ1_WRAP;
331 else if (s2 + back_max >= last)
332 d2 = (last - s2) * cfqd->cfq_back_penalty;
334 wrap |= CFQ_RQ2_WRAP;
336 /* Found required data */
339 * By doing switch() on the bit mask "wrap" we avoid having to
340 * check two variables for all permutations: --> faster!
343 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
359 case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
362 * Since both rqs are wrapped,
363 * start with the one that's further behind head
364 * (--> only *one* back seek required),
365 * since back seek takes more time than forward.
375 * would be nice to take fifo expire time into account as well
377 static struct request *
378 cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
379 struct request *last)
381 struct rb_node *rbnext = rb_next(&last->rb_node);
382 struct rb_node *rbprev = rb_prev(&last->rb_node);
383 struct request *next = NULL, *prev = NULL;
385 BUG_ON(RB_EMPTY_NODE(&last->rb_node));
388 prev = rb_entry_rq(rbprev);
391 next = rb_entry_rq(rbnext);
393 rbnext = rb_first(&cfqq->sort_list);
394 if (rbnext && rbnext != &last->rb_node)
395 next = rb_entry_rq(rbnext);
398 return cfq_choose_req(cfqd, next, prev);
401 static void cfq_resort_rr_list(struct cfq_queue *cfqq, int preempted)
403 struct cfq_data *cfqd = cfqq->cfqd;
404 struct list_head *list, *n;
405 struct cfq_queue *__cfqq;
408 * Resorting requires the cfqq to be on the RR list already.
410 if (!cfq_cfqq_on_rr(cfqq))
413 list_del(&cfqq->cfq_list);
415 if (cfq_class_rt(cfqq))
416 list = &cfqd->cur_rr;
417 else if (cfq_class_idle(cfqq))
418 list = &cfqd->idle_rr;
421 * if cfqq has requests in flight, don't allow it to be
422 * found in cfq_set_active_queue before it has finished them.
423 * this is done to increase fairness between a process that
424 * has lots of io pending vs one that only generates one
425 * sporadically or synchronously
427 if (cfq_cfqq_dispatched(cfqq))
428 list = &cfqd->busy_rr;
430 list = &cfqd->rr_list[cfqq->ioprio];
433 if (preempted || cfq_cfqq_queue_new(cfqq)) {
435 * If this queue was preempted or is new (never been serviced),
436 * let it be added first for fairness but beind other new
440 while (n->next != list) {
441 __cfqq = list_entry_cfqq(n->next);
442 if (!cfq_cfqq_queue_new(__cfqq))
447 list_add_tail(&cfqq->cfq_list, n);
448 } else if (!cfq_cfqq_class_sync(cfqq)) {
450 * async queue always goes to the end. this wont be overly
451 * unfair to writes, as the sort of the sync queue wont be
452 * allowed to pass the async queue again.
454 list_add_tail(&cfqq->cfq_list, list);
457 * sort by last service, but don't cross a new or async
458 * queue. we don't cross a new queue because it hasn't been
459 * service before, and we don't cross an async queue because
460 * it gets added to the end on expire.
463 while ((n = n->prev) != list) {
464 struct cfq_queue *__cfqq = list_entry_cfqq(n);
466 if (!cfq_cfqq_class_sync(cfqq) || !__cfqq->service_last)
468 if (time_before(__cfqq->service_last, cfqq->service_last))
471 list_add(&cfqq->cfq_list, n);
476 * add to busy list of queues for service, trying to be fair in ordering
477 * the pending list according to last request service
480 cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
482 BUG_ON(cfq_cfqq_on_rr(cfqq));
483 cfq_mark_cfqq_on_rr(cfqq);
486 cfq_resort_rr_list(cfqq, 0);
490 cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
492 BUG_ON(!cfq_cfqq_on_rr(cfqq));
493 cfq_clear_cfqq_on_rr(cfqq);
494 list_del_init(&cfqq->cfq_list);
496 BUG_ON(!cfqd->busy_queues);
501 * rb tree support functions
503 static inline void cfq_del_rq_rb(struct request *rq)
505 struct cfq_queue *cfqq = RQ_CFQQ(rq);
506 struct cfq_data *cfqd = cfqq->cfqd;
507 const int sync = rq_is_sync(rq);
509 BUG_ON(!cfqq->queued[sync]);
510 cfqq->queued[sync]--;
512 elv_rb_del(&cfqq->sort_list, rq);
514 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
515 cfq_del_cfqq_rr(cfqd, cfqq);
518 static void cfq_add_rq_rb(struct request *rq)
520 struct cfq_queue *cfqq = RQ_CFQQ(rq);
521 struct cfq_data *cfqd = cfqq->cfqd;
522 struct request *__alias;
524 cfqq->queued[rq_is_sync(rq)]++;
527 * looks a little odd, but the first insert might return an alias.
528 * if that happens, put the alias on the dispatch list
530 while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL)
531 cfq_dispatch_insert(cfqd->queue, __alias);
533 if (!cfq_cfqq_on_rr(cfqq))
534 cfq_add_cfqq_rr(cfqd, cfqq);
537 * check if this request is a better next-serve candidate
539 cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq);
540 BUG_ON(!cfqq->next_rq);
544 cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
546 elv_rb_del(&cfqq->sort_list, rq);
547 cfqq->queued[rq_is_sync(rq)]--;
551 static struct request *
552 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
554 struct task_struct *tsk = current;
555 pid_t key = cfq_queue_pid(tsk, bio_data_dir(bio), bio_sync(bio));
556 struct cfq_queue *cfqq;
558 cfqq = cfq_find_cfq_hash(cfqd, key, tsk->ioprio);
560 sector_t sector = bio->bi_sector + bio_sectors(bio);
562 return elv_rb_find(&cfqq->sort_list, sector);
568 static void cfq_activate_request(request_queue_t *q, struct request *rq)
570 struct cfq_data *cfqd = q->elevator->elevator_data;
572 cfqd->rq_in_driver++;
575 * If the depth is larger 1, it really could be queueing. But lets
576 * make the mark a little higher - idling could still be good for
577 * low queueing, and a low queueing number could also just indicate
578 * a SCSI mid layer like behaviour where limit+1 is often seen.
580 if (!cfqd->hw_tag && cfqd->rq_in_driver > 4)
584 static void cfq_deactivate_request(request_queue_t *q, struct request *rq)
586 struct cfq_data *cfqd = q->elevator->elevator_data;
588 WARN_ON(!cfqd->rq_in_driver);
589 cfqd->rq_in_driver--;
592 static void cfq_remove_request(struct request *rq)
594 struct cfq_queue *cfqq = RQ_CFQQ(rq);
596 if (cfqq->next_rq == rq)
597 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
599 list_del_init(&rq->queuelist);
602 if (rq_is_meta(rq)) {
603 WARN_ON(!cfqq->meta_pending);
604 cfqq->meta_pending--;
609 cfq_merge(request_queue_t *q, struct request **req, struct bio *bio)
611 struct cfq_data *cfqd = q->elevator->elevator_data;
612 struct request *__rq;
614 __rq = cfq_find_rq_fmerge(cfqd, bio);
615 if (__rq && elv_rq_merge_ok(__rq, bio)) {
617 return ELEVATOR_FRONT_MERGE;
620 return ELEVATOR_NO_MERGE;
623 static void cfq_merged_request(request_queue_t *q, struct request *req,
626 if (type == ELEVATOR_FRONT_MERGE) {
627 struct cfq_queue *cfqq = RQ_CFQQ(req);
629 cfq_reposition_rq_rb(cfqq, req);
634 cfq_merged_requests(request_queue_t *q, struct request *rq,
635 struct request *next)
638 * reposition in fifo if next is older than rq
640 if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
641 time_before(next->start_time, rq->start_time))
642 list_move(&rq->queuelist, &next->queuelist);
644 cfq_remove_request(next);
647 static int cfq_allow_merge(request_queue_t *q, struct request *rq,
650 struct cfq_data *cfqd = q->elevator->elevator_data;
651 const int rw = bio_data_dir(bio);
652 struct cfq_queue *cfqq;
656 * Disallow merge of a sync bio into an async request.
658 if ((bio_data_dir(bio) == READ || bio_sync(bio)) && !rq_is_sync(rq))
662 * Lookup the cfqq that this bio will be queued with. Allow
663 * merge only if rq is queued there.
665 key = cfq_queue_pid(current, rw, bio_sync(bio));
666 cfqq = cfq_find_cfq_hash(cfqd, key, current->ioprio);
668 if (cfqq == RQ_CFQQ(rq))
675 __cfq_set_active_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
679 * stop potential idle class queues waiting service
681 del_timer(&cfqd->idle_class_timer);
684 cfq_clear_cfqq_must_alloc_slice(cfqq);
685 cfq_clear_cfqq_fifo_expire(cfqq);
686 cfq_mark_cfqq_slice_new(cfqq);
689 cfqd->active_queue = cfqq;
693 * current cfqq expired its slice (or was too idle), select new one
696 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
697 int preempted, int timed_out)
699 if (cfq_cfqq_wait_request(cfqq))
700 del_timer(&cfqd->idle_slice_timer);
702 cfq_clear_cfqq_must_dispatch(cfqq);
703 cfq_clear_cfqq_wait_request(cfqq);
704 cfq_clear_cfqq_queue_new(cfqq);
707 * store what was left of this slice, if the queue idled out
710 if (timed_out && !cfq_cfqq_slice_new(cfqq))
711 cfqq->slice_resid = cfqq->slice_end - jiffies;
713 cfq_resort_rr_list(cfqq, preempted);
715 if (cfqq == cfqd->active_queue)
716 cfqd->active_queue = NULL;
718 if (cfqd->active_cic) {
719 put_io_context(cfqd->active_cic->ioc);
720 cfqd->active_cic = NULL;
723 cfqd->dispatch_slice = 0;
726 static inline void cfq_slice_expired(struct cfq_data *cfqd, int preempted,
729 struct cfq_queue *cfqq = cfqd->active_queue;
732 __cfq_slice_expired(cfqd, cfqq, preempted, timed_out);
745 static int cfq_get_next_prio_level(struct cfq_data *cfqd)
754 for (p = cfqd->cur_prio; p <= cfqd->cur_end_prio; p++) {
755 if (!list_empty(&cfqd->rr_list[p])) {
764 if (++cfqd->cur_end_prio == CFQ_PRIO_LISTS) {
765 cfqd->cur_end_prio = 0;
772 if (unlikely(prio == -1))
775 BUG_ON(prio >= CFQ_PRIO_LISTS);
777 list_splice_init(&cfqd->rr_list[prio], &cfqd->cur_rr);
779 cfqd->cur_prio = prio + 1;
780 if (cfqd->cur_prio > cfqd->cur_end_prio) {
781 cfqd->cur_end_prio = cfqd->cur_prio;
784 if (cfqd->cur_end_prio == CFQ_PRIO_LISTS) {
786 cfqd->cur_end_prio = 0;
792 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd)
794 struct cfq_queue *cfqq = NULL;
796 if (!list_empty(&cfqd->cur_rr) || cfq_get_next_prio_level(cfqd) != -1) {
798 * if current list is non-empty, grab first entry. if it is
799 * empty, get next prio level and grab first entry then if any
802 cfqq = list_entry_cfqq(cfqd->cur_rr.next);
803 } else if (!list_empty(&cfqd->busy_rr)) {
805 * If no new queues are available, check if the busy list has
806 * some before falling back to idle io.
808 cfqq = list_entry_cfqq(cfqd->busy_rr.next);
809 } else if (!list_empty(&cfqd->idle_rr)) {
811 * if we have idle queues and no rt or be queues had pending
812 * requests, either allow immediate service if the grace period
813 * has passed or arm the idle grace timer
815 unsigned long end = cfqd->last_end_request + CFQ_IDLE_GRACE;
817 if (time_after_eq(jiffies, end))
818 cfqq = list_entry_cfqq(cfqd->idle_rr.next);
820 mod_timer(&cfqd->idle_class_timer, end);
823 __cfq_set_active_queue(cfqd, cfqq);
827 #define CIC_SEEKY(cic) ((cic)->seek_mean > (128 * 1024))
829 static int cfq_arm_slice_timer(struct cfq_data *cfqd)
831 struct cfq_queue *cfqq = cfqd->active_queue;
832 struct cfq_io_context *cic;
835 WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
838 * idle is disabled, either manually or by past process history
840 if (!cfqd->cfq_slice_idle)
842 if (!cfq_cfqq_idle_window(cfqq))
845 * task has exited, don't wait
847 cic = cfqd->active_cic;
848 if (!cic || !cic->ioc->task)
851 cfq_mark_cfqq_must_dispatch(cfqq);
852 cfq_mark_cfqq_wait_request(cfqq);
854 sl = min(cfqq->slice_end - 1, (unsigned long) cfqd->cfq_slice_idle);
857 * we don't want to idle for seeks, but we do want to allow
858 * fair distribution of slice time for a process doing back-to-back
859 * seeks. so allow a little bit of time for him to submit a new rq
861 if (sample_valid(cic->seek_samples) && CIC_SEEKY(cic))
862 sl = min(sl, msecs_to_jiffies(2));
864 mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
868 static void cfq_dispatch_insert(request_queue_t *q, struct request *rq)
870 struct cfq_queue *cfqq = RQ_CFQQ(rq);
872 cfq_remove_request(rq);
873 cfqq->on_dispatch[rq_is_sync(rq)]++;
874 elv_dispatch_sort(q, rq);
878 * return expired entry, or NULL to just start from scratch in rbtree
880 static inline struct request *cfq_check_fifo(struct cfq_queue *cfqq)
882 struct cfq_data *cfqd = cfqq->cfqd;
886 if (cfq_cfqq_fifo_expire(cfqq))
889 cfq_mark_cfqq_fifo_expire(cfqq);
891 if (list_empty(&cfqq->fifo))
894 fifo = cfq_cfqq_class_sync(cfqq);
895 rq = rq_entry_fifo(cfqq->fifo.next);
897 if (time_after(jiffies, rq->start_time + cfqd->cfq_fifo_expire[fifo]))
904 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
906 const int base_rq = cfqd->cfq_slice_async_rq;
908 WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
910 return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
914 * get next queue for service
916 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
918 struct cfq_queue *cfqq;
920 cfqq = cfqd->active_queue;
927 if (!cfq_cfqq_must_dispatch(cfqq) && cfq_slice_used(cfqq))
931 * if queue has requests, dispatch one. if not, check if
932 * enough slice is left to wait for one
934 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
936 else if (cfq_cfqq_slice_new(cfqq) || cfq_cfqq_dispatched(cfqq)) {
939 } else if (cfq_cfqq_class_sync(cfqq)) {
940 if (cfq_arm_slice_timer(cfqd))
945 cfq_slice_expired(cfqd, 0, 0);
947 cfqq = cfq_set_active_queue(cfqd);
953 __cfq_dispatch_requests(struct cfq_data *cfqd, struct cfq_queue *cfqq,
958 BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
964 * follow expired path, else get first next available
966 if ((rq = cfq_check_fifo(cfqq)) == NULL)
970 * finally, insert request into driver dispatch list
972 cfq_dispatch_insert(cfqd->queue, rq);
974 cfqd->dispatch_slice++;
977 if (!cfqd->active_cic) {
978 atomic_inc(&RQ_CIC(rq)->ioc->refcount);
979 cfqd->active_cic = RQ_CIC(rq);
982 if (RB_EMPTY_ROOT(&cfqq->sort_list))
985 } while (dispatched < max_dispatch);
988 * expire an async queue immediately if it has used up its slice. idle
989 * queue always expire after 1 dispatch round.
991 if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
992 cfqd->dispatch_slice >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
993 cfq_class_idle(cfqq))) {
994 cfqq->slice_end = jiffies + 1;
995 cfq_slice_expired(cfqd, 0, 0);
1002 cfq_forced_dispatch_cfqqs(struct list_head *list)
1004 struct cfq_queue *cfqq, *next;
1008 list_for_each_entry_safe(cfqq, next, list, cfq_list) {
1009 while (cfqq->next_rq) {
1010 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
1013 BUG_ON(!list_empty(&cfqq->fifo));
1020 cfq_forced_dispatch(struct cfq_data *cfqd)
1022 int i, dispatched = 0;
1024 for (i = 0; i < CFQ_PRIO_LISTS; i++)
1025 dispatched += cfq_forced_dispatch_cfqqs(&cfqd->rr_list[i]);
1027 dispatched += cfq_forced_dispatch_cfqqs(&cfqd->busy_rr);
1028 dispatched += cfq_forced_dispatch_cfqqs(&cfqd->cur_rr);
1029 dispatched += cfq_forced_dispatch_cfqqs(&cfqd->idle_rr);
1031 cfq_slice_expired(cfqd, 0, 0);
1033 BUG_ON(cfqd->busy_queues);
1039 cfq_dispatch_requests(request_queue_t *q, int force)
1041 struct cfq_data *cfqd = q->elevator->elevator_data;
1042 struct cfq_queue *cfqq, *prev_cfqq;
1045 if (!cfqd->busy_queues)
1048 if (unlikely(force))
1049 return cfq_forced_dispatch(cfqd);
1053 while ((cfqq = cfq_select_queue(cfqd)) != NULL) {
1056 if (cfqd->busy_queues > 1) {
1058 * Don't repeat dispatch from the previous queue.
1060 if (prev_cfqq == cfqq)
1064 * So we have dispatched before in this round, if the
1065 * next queue has idling enabled (must be sync), don't
1066 * allow it service until the previous have continued.
1068 if (cfqd->rq_in_driver && cfq_cfqq_idle_window(cfqq))
1072 cfq_clear_cfqq_must_dispatch(cfqq);
1073 cfq_clear_cfqq_wait_request(cfqq);
1074 del_timer(&cfqd->idle_slice_timer);
1076 max_dispatch = cfqd->cfq_quantum;
1077 if (cfq_class_idle(cfqq))
1080 dispatched += __cfq_dispatch_requests(cfqd, cfqq, max_dispatch);
1088 * task holds one reference to the queue, dropped when task exits. each rq
1089 * in-flight on this queue also holds a reference, dropped when rq is freed.
1091 * queue lock must be held here.
1093 static void cfq_put_queue(struct cfq_queue *cfqq)
1095 struct cfq_data *cfqd = cfqq->cfqd;
1097 BUG_ON(atomic_read(&cfqq->ref) <= 0);
1099 if (!atomic_dec_and_test(&cfqq->ref))
1102 BUG_ON(rb_first(&cfqq->sort_list));
1103 BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
1104 BUG_ON(cfq_cfqq_on_rr(cfqq));
1106 if (unlikely(cfqd->active_queue == cfqq)) {
1107 __cfq_slice_expired(cfqd, cfqq, 0, 0);
1108 cfq_schedule_dispatch(cfqd);
1112 * it's on the empty list and still hashed
1114 list_del(&cfqq->cfq_list);
1115 hlist_del(&cfqq->cfq_hash);
1116 kmem_cache_free(cfq_pool, cfqq);
1119 static struct cfq_queue *
1120 __cfq_find_cfq_hash(struct cfq_data *cfqd, unsigned int key, unsigned int prio,
1123 struct hlist_head *hash_list = &cfqd->cfq_hash[hashval];
1124 struct hlist_node *entry;
1125 struct cfq_queue *__cfqq;
1127 hlist_for_each_entry(__cfqq, entry, hash_list, cfq_hash) {
1128 const unsigned short __p = IOPRIO_PRIO_VALUE(__cfqq->org_ioprio_class, __cfqq->org_ioprio);
1130 if (__cfqq->key == key && (__p == prio || !prio))
1137 static struct cfq_queue *
1138 cfq_find_cfq_hash(struct cfq_data *cfqd, unsigned int key, unsigned short prio)
1140 return __cfq_find_cfq_hash(cfqd, key, prio, hash_long(key, CFQ_QHASH_SHIFT));
1143 static void cfq_free_io_context(struct io_context *ioc)
1145 struct cfq_io_context *__cic;
1149 while ((n = rb_first(&ioc->cic_root)) != NULL) {
1150 __cic = rb_entry(n, struct cfq_io_context, rb_node);
1151 rb_erase(&__cic->rb_node, &ioc->cic_root);
1152 kmem_cache_free(cfq_ioc_pool, __cic);
1156 elv_ioc_count_mod(ioc_count, -freed);
1158 if (ioc_gone && !elv_ioc_count_read(ioc_count))
1162 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1164 if (unlikely(cfqq == cfqd->active_queue)) {
1165 __cfq_slice_expired(cfqd, cfqq, 0, 0);
1166 cfq_schedule_dispatch(cfqd);
1169 cfq_put_queue(cfqq);
1172 static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
1173 struct cfq_io_context *cic)
1175 list_del_init(&cic->queue_list);
1179 if (cic->cfqq[ASYNC]) {
1180 cfq_exit_cfqq(cfqd, cic->cfqq[ASYNC]);
1181 cic->cfqq[ASYNC] = NULL;
1184 if (cic->cfqq[SYNC]) {
1185 cfq_exit_cfqq(cfqd, cic->cfqq[SYNC]);
1186 cic->cfqq[SYNC] = NULL;
1192 * Called with interrupts disabled
1194 static void cfq_exit_single_io_context(struct cfq_io_context *cic)
1196 struct cfq_data *cfqd = cic->key;
1199 request_queue_t *q = cfqd->queue;
1201 spin_lock_irq(q->queue_lock);
1202 __cfq_exit_single_io_context(cfqd, cic);
1203 spin_unlock_irq(q->queue_lock);
1207 static void cfq_exit_io_context(struct io_context *ioc)
1209 struct cfq_io_context *__cic;
1213 * put the reference this task is holding to the various queues
1216 n = rb_first(&ioc->cic_root);
1218 __cic = rb_entry(n, struct cfq_io_context, rb_node);
1220 cfq_exit_single_io_context(__cic);
1225 static struct cfq_io_context *
1226 cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
1228 struct cfq_io_context *cic;
1230 cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask, cfqd->queue->node);
1232 memset(cic, 0, sizeof(*cic));
1233 cic->last_end_request = jiffies;
1234 INIT_LIST_HEAD(&cic->queue_list);
1235 cic->dtor = cfq_free_io_context;
1236 cic->exit = cfq_exit_io_context;
1237 elv_ioc_count_inc(ioc_count);
1243 static void cfq_init_prio_data(struct cfq_queue *cfqq)
1245 struct task_struct *tsk = current;
1248 if (!cfq_cfqq_prio_changed(cfqq))
1251 ioprio_class = IOPRIO_PRIO_CLASS(tsk->ioprio);
1252 switch (ioprio_class) {
1254 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
1255 case IOPRIO_CLASS_NONE:
1257 * no prio set, place us in the middle of the BE classes
1259 cfqq->ioprio = task_nice_ioprio(tsk);
1260 cfqq->ioprio_class = IOPRIO_CLASS_BE;
1262 case IOPRIO_CLASS_RT:
1263 cfqq->ioprio = task_ioprio(tsk);
1264 cfqq->ioprio_class = IOPRIO_CLASS_RT;
1266 case IOPRIO_CLASS_BE:
1267 cfqq->ioprio = task_ioprio(tsk);
1268 cfqq->ioprio_class = IOPRIO_CLASS_BE;
1270 case IOPRIO_CLASS_IDLE:
1271 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
1273 cfq_clear_cfqq_idle_window(cfqq);
1278 * keep track of original prio settings in case we have to temporarily
1279 * elevate the priority of this queue
1281 cfqq->org_ioprio = cfqq->ioprio;
1282 cfqq->org_ioprio_class = cfqq->ioprio_class;
1284 cfq_resort_rr_list(cfqq, 0);
1285 cfq_clear_cfqq_prio_changed(cfqq);
1288 static inline void changed_ioprio(struct cfq_io_context *cic)
1290 struct cfq_data *cfqd = cic->key;
1291 struct cfq_queue *cfqq;
1292 unsigned long flags;
1294 if (unlikely(!cfqd))
1297 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1299 cfqq = cic->cfqq[ASYNC];
1301 struct cfq_queue *new_cfqq;
1302 new_cfqq = cfq_get_queue(cfqd, CFQ_KEY_ASYNC, cic->ioc->task,
1305 cic->cfqq[ASYNC] = new_cfqq;
1306 cfq_put_queue(cfqq);
1310 cfqq = cic->cfqq[SYNC];
1312 cfq_mark_cfqq_prio_changed(cfqq);
1314 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1317 static void cfq_ioc_set_ioprio(struct io_context *ioc)
1319 struct cfq_io_context *cic;
1322 ioc->ioprio_changed = 0;
1324 n = rb_first(&ioc->cic_root);
1326 cic = rb_entry(n, struct cfq_io_context, rb_node);
1328 changed_ioprio(cic);
1333 static struct cfq_queue *
1334 cfq_get_queue(struct cfq_data *cfqd, unsigned int key, struct task_struct *tsk,
1337 const int hashval = hash_long(key, CFQ_QHASH_SHIFT);
1338 struct cfq_queue *cfqq, *new_cfqq = NULL;
1339 unsigned short ioprio;
1342 ioprio = tsk->ioprio;
1343 cfqq = __cfq_find_cfq_hash(cfqd, key, ioprio, hashval);
1349 } else if (gfp_mask & __GFP_WAIT) {
1351 * Inform the allocator of the fact that we will
1352 * just repeat this allocation if it fails, to allow
1353 * the allocator to do whatever it needs to attempt to
1356 spin_unlock_irq(cfqd->queue->queue_lock);
1357 new_cfqq = kmem_cache_alloc_node(cfq_pool, gfp_mask|__GFP_NOFAIL, cfqd->queue->node);
1358 spin_lock_irq(cfqd->queue->queue_lock);
1361 cfqq = kmem_cache_alloc_node(cfq_pool, gfp_mask, cfqd->queue->node);
1366 memset(cfqq, 0, sizeof(*cfqq));
1368 INIT_HLIST_NODE(&cfqq->cfq_hash);
1369 INIT_LIST_HEAD(&cfqq->cfq_list);
1370 INIT_LIST_HEAD(&cfqq->fifo);
1373 hlist_add_head(&cfqq->cfq_hash, &cfqd->cfq_hash[hashval]);
1374 atomic_set(&cfqq->ref, 0);
1377 if (key != CFQ_KEY_ASYNC)
1378 cfq_mark_cfqq_idle_window(cfqq);
1380 cfq_mark_cfqq_prio_changed(cfqq);
1381 cfq_mark_cfqq_queue_new(cfqq);
1382 cfq_init_prio_data(cfqq);
1386 kmem_cache_free(cfq_pool, new_cfqq);
1388 atomic_inc(&cfqq->ref);
1390 WARN_ON((gfp_mask & __GFP_WAIT) && !cfqq);
1395 cfq_drop_dead_cic(struct io_context *ioc, struct cfq_io_context *cic)
1397 WARN_ON(!list_empty(&cic->queue_list));
1398 rb_erase(&cic->rb_node, &ioc->cic_root);
1399 kmem_cache_free(cfq_ioc_pool, cic);
1400 elv_ioc_count_dec(ioc_count);
1403 static struct cfq_io_context *
1404 cfq_cic_rb_lookup(struct cfq_data *cfqd, struct io_context *ioc)
1407 struct cfq_io_context *cic;
1408 void *k, *key = cfqd;
1411 n = ioc->cic_root.rb_node;
1413 cic = rb_entry(n, struct cfq_io_context, rb_node);
1414 /* ->key must be copied to avoid race with cfq_exit_queue() */
1417 cfq_drop_dead_cic(ioc, cic);
1433 cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
1434 struct cfq_io_context *cic)
1437 struct rb_node *parent;
1438 struct cfq_io_context *__cic;
1439 unsigned long flags;
1447 p = &ioc->cic_root.rb_node;
1450 __cic = rb_entry(parent, struct cfq_io_context, rb_node);
1451 /* ->key must be copied to avoid race with cfq_exit_queue() */
1454 cfq_drop_dead_cic(ioc, __cic);
1460 else if (cic->key > k)
1461 p = &(*p)->rb_right;
1466 rb_link_node(&cic->rb_node, parent, p);
1467 rb_insert_color(&cic->rb_node, &ioc->cic_root);
1469 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1470 list_add(&cic->queue_list, &cfqd->cic_list);
1471 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1475 * Setup general io context and cfq io context. There can be several cfq
1476 * io contexts per general io context, if this process is doing io to more
1477 * than one device managed by cfq.
1479 static struct cfq_io_context *
1480 cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
1482 struct io_context *ioc = NULL;
1483 struct cfq_io_context *cic;
1485 might_sleep_if(gfp_mask & __GFP_WAIT);
1487 ioc = get_io_context(gfp_mask, cfqd->queue->node);
1491 cic = cfq_cic_rb_lookup(cfqd, ioc);
1495 cic = cfq_alloc_io_context(cfqd, gfp_mask);
1499 cfq_cic_link(cfqd, ioc, cic);
1501 smp_read_barrier_depends();
1502 if (unlikely(ioc->ioprio_changed))
1503 cfq_ioc_set_ioprio(ioc);
1507 put_io_context(ioc);
1512 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
1514 unsigned long elapsed = jiffies - cic->last_end_request;
1515 unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
1517 cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
1518 cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
1519 cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
1523 cfq_update_io_seektime(struct cfq_io_context *cic, struct request *rq)
1528 if (cic->last_request_pos < rq->sector)
1529 sdist = rq->sector - cic->last_request_pos;
1531 sdist = cic->last_request_pos - rq->sector;
1534 * Don't allow the seek distance to get too large from the
1535 * odd fragment, pagein, etc
1537 if (cic->seek_samples <= 60) /* second&third seek */
1538 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*1024);
1540 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*64);
1542 cic->seek_samples = (7*cic->seek_samples + 256) / 8;
1543 cic->seek_total = (7*cic->seek_total + (u64)256*sdist) / 8;
1544 total = cic->seek_total + (cic->seek_samples/2);
1545 do_div(total, cic->seek_samples);
1546 cic->seek_mean = (sector_t)total;
1550 * Disable idle window if the process thinks too long or seeks so much that
1554 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1555 struct cfq_io_context *cic)
1557 int enable_idle = cfq_cfqq_idle_window(cfqq);
1559 if (!cic->ioc->task || !cfqd->cfq_slice_idle ||
1560 (cfqd->hw_tag && CIC_SEEKY(cic)))
1562 else if (sample_valid(cic->ttime_samples)) {
1563 if (cic->ttime_mean > cfqd->cfq_slice_idle)
1570 cfq_mark_cfqq_idle_window(cfqq);
1572 cfq_clear_cfqq_idle_window(cfqq);
1576 * Check if new_cfqq should preempt the currently active queue. Return 0 for
1577 * no or if we aren't sure, a 1 will cause a preempt.
1580 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
1583 struct cfq_queue *cfqq = cfqd->active_queue;
1586 if (cfq_class_idle(new_cfqq))
1592 if (cfq_class_idle(cfqq))
1596 * if the new request is sync, but the currently running queue is
1597 * not, let the sync request have priority.
1599 if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
1603 * So both queues are sync. Let the new request get disk time if
1604 * it's a metadata request and the current queue is doing regular IO.
1606 if (rq_is_meta(rq) && !cfqq->meta_pending)
1609 if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
1613 * if this request is as-good as one we would expect from the
1614 * current cfqq, let it preempt
1616 if (rq->sector > cfqd->last_sector)
1617 dist = rq->sector - cfqd->last_sector;
1619 dist = cfqd->last_sector - rq->sector;
1621 if (dist <= cfqd->active_cic->seek_mean)
1628 * cfqq preempts the active queue. if we allowed preempt with no slice left,
1629 * let it have half of its nominal slice.
1631 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1633 cfq_slice_expired(cfqd, 1, 1);
1636 * Put the new queue at the front of the of the current list,
1637 * so we know that it will be selected next.
1639 BUG_ON(!cfq_cfqq_on_rr(cfqq));
1640 list_move(&cfqq->cfq_list, &cfqd->cur_rr);
1642 cfqq->slice_end = 0;
1643 cfq_mark_cfqq_slice_new(cfqq);
1647 * Called when a new fs request (rq) is added (to cfqq). Check if there's
1648 * something we should do about it
1651 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1654 struct cfq_io_context *cic = RQ_CIC(rq);
1657 cfqq->meta_pending++;
1660 * we never wait for an async request and we don't allow preemption
1661 * of an async request. so just return early
1663 if (!rq_is_sync(rq)) {
1665 * sync process issued an async request, if it's waiting
1666 * then expire it and kick rq handling.
1668 if (cic == cfqd->active_cic &&
1669 del_timer(&cfqd->idle_slice_timer)) {
1670 cfq_slice_expired(cfqd, 0, 0);
1671 blk_start_queueing(cfqd->queue);
1676 cfq_update_io_thinktime(cfqd, cic);
1677 cfq_update_io_seektime(cic, rq);
1678 cfq_update_idle_window(cfqd, cfqq, cic);
1680 cic->last_request_pos = rq->sector + rq->nr_sectors;
1682 if (cfqq == cfqd->active_queue) {
1684 * if we are waiting for a request for this queue, let it rip
1685 * immediately and flag that we must not expire this queue
1688 if (cfq_cfqq_wait_request(cfqq)) {
1689 cfq_mark_cfqq_must_dispatch(cfqq);
1690 del_timer(&cfqd->idle_slice_timer);
1691 blk_start_queueing(cfqd->queue);
1693 } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
1695 * not the active queue - expire current slice if it is
1696 * idle and has expired it's mean thinktime or this new queue
1697 * has some old slice time left and is of higher priority
1699 cfq_preempt_queue(cfqd, cfqq);
1700 cfq_mark_cfqq_must_dispatch(cfqq);
1701 blk_start_queueing(cfqd->queue);
1705 static void cfq_insert_request(request_queue_t *q, struct request *rq)
1707 struct cfq_data *cfqd = q->elevator->elevator_data;
1708 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1710 cfq_init_prio_data(cfqq);
1714 list_add_tail(&rq->queuelist, &cfqq->fifo);
1716 cfq_rq_enqueued(cfqd, cfqq, rq);
1719 static void cfq_completed_request(request_queue_t *q, struct request *rq)
1721 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1722 struct cfq_data *cfqd = cfqq->cfqd;
1723 const int sync = rq_is_sync(rq);
1728 WARN_ON(!cfqd->rq_in_driver);
1729 WARN_ON(!cfqq->on_dispatch[sync]);
1730 cfqd->rq_in_driver--;
1731 cfqq->on_dispatch[sync]--;
1732 cfqq->service_last = now;
1734 cfqd->last_sector = rq->hard_sector + rq->hard_nr_sectors;
1736 if (!cfq_class_idle(cfqq))
1737 cfqd->last_end_request = now;
1739 cfq_resort_rr_list(cfqq, 0);
1742 RQ_CIC(rq)->last_end_request = now;
1745 * If this is the active queue, check if it needs to be expired,
1746 * or if we want to idle in case it has no pending requests.
1748 if (cfqd->active_queue == cfqq) {
1749 if (cfq_cfqq_slice_new(cfqq)) {
1750 cfq_set_prio_slice(cfqd, cfqq);
1751 cfq_clear_cfqq_slice_new(cfqq);
1753 if (cfq_slice_used(cfqq))
1754 cfq_slice_expired(cfqd, 0, 1);
1755 else if (sync && RB_EMPTY_ROOT(&cfqq->sort_list)) {
1756 if (!cfq_arm_slice_timer(cfqd))
1757 cfq_schedule_dispatch(cfqd);
1763 * we temporarily boost lower priority queues if they are holding fs exclusive
1764 * resources. they are boosted to normal prio (CLASS_BE/4)
1766 static void cfq_prio_boost(struct cfq_queue *cfqq)
1768 const int ioprio_class = cfqq->ioprio_class;
1769 const int ioprio = cfqq->ioprio;
1771 if (has_fs_excl()) {
1773 * boost idle prio on transactions that would lock out other
1774 * users of the filesystem
1776 if (cfq_class_idle(cfqq))
1777 cfqq->ioprio_class = IOPRIO_CLASS_BE;
1778 if (cfqq->ioprio > IOPRIO_NORM)
1779 cfqq->ioprio = IOPRIO_NORM;
1782 * check if we need to unboost the queue
1784 if (cfqq->ioprio_class != cfqq->org_ioprio_class)
1785 cfqq->ioprio_class = cfqq->org_ioprio_class;
1786 if (cfqq->ioprio != cfqq->org_ioprio)
1787 cfqq->ioprio = cfqq->org_ioprio;
1791 * refile between round-robin lists if we moved the priority class
1793 if ((ioprio_class != cfqq->ioprio_class || ioprio != cfqq->ioprio))
1794 cfq_resort_rr_list(cfqq, 0);
1797 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
1799 if ((cfq_cfqq_wait_request(cfqq) || cfq_cfqq_must_alloc(cfqq)) &&
1800 !cfq_cfqq_must_alloc_slice(cfqq)) {
1801 cfq_mark_cfqq_must_alloc_slice(cfqq);
1802 return ELV_MQUEUE_MUST;
1805 return ELV_MQUEUE_MAY;
1808 static int cfq_may_queue(request_queue_t *q, int rw)
1810 struct cfq_data *cfqd = q->elevator->elevator_data;
1811 struct task_struct *tsk = current;
1812 struct cfq_queue *cfqq;
1815 key = cfq_queue_pid(tsk, rw, rw & REQ_RW_SYNC);
1818 * don't force setup of a queue from here, as a call to may_queue
1819 * does not necessarily imply that a request actually will be queued.
1820 * so just lookup a possibly existing queue, or return 'may queue'
1823 cfqq = cfq_find_cfq_hash(cfqd, key, tsk->ioprio);
1825 cfq_init_prio_data(cfqq);
1826 cfq_prio_boost(cfqq);
1828 return __cfq_may_queue(cfqq);
1831 return ELV_MQUEUE_MAY;
1835 * queue lock held here
1837 static void cfq_put_request(struct request *rq)
1839 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1842 const int rw = rq_data_dir(rq);
1844 BUG_ON(!cfqq->allocated[rw]);
1845 cfqq->allocated[rw]--;
1847 put_io_context(RQ_CIC(rq)->ioc);
1849 rq->elevator_private = NULL;
1850 rq->elevator_private2 = NULL;
1852 cfq_put_queue(cfqq);
1857 * Allocate cfq data structures associated with this request.
1860 cfq_set_request(request_queue_t *q, struct request *rq, gfp_t gfp_mask)
1862 struct cfq_data *cfqd = q->elevator->elevator_data;
1863 struct task_struct *tsk = current;
1864 struct cfq_io_context *cic;
1865 const int rw = rq_data_dir(rq);
1866 const int is_sync = rq_is_sync(rq);
1867 pid_t key = cfq_queue_pid(tsk, rw, is_sync);
1868 struct cfq_queue *cfqq;
1869 unsigned long flags;
1871 might_sleep_if(gfp_mask & __GFP_WAIT);
1873 cic = cfq_get_io_context(cfqd, gfp_mask);
1875 spin_lock_irqsave(q->queue_lock, flags);
1880 if (!cic->cfqq[is_sync]) {
1881 cfqq = cfq_get_queue(cfqd, key, tsk, gfp_mask);
1885 cic->cfqq[is_sync] = cfqq;
1887 cfqq = cic->cfqq[is_sync];
1889 cfqq->allocated[rw]++;
1890 cfq_clear_cfqq_must_alloc(cfqq);
1891 atomic_inc(&cfqq->ref);
1893 spin_unlock_irqrestore(q->queue_lock, flags);
1895 rq->elevator_private = cic;
1896 rq->elevator_private2 = cfqq;
1901 put_io_context(cic->ioc);
1903 cfq_schedule_dispatch(cfqd);
1904 spin_unlock_irqrestore(q->queue_lock, flags);
1908 static void cfq_kick_queue(struct work_struct *work)
1910 struct cfq_data *cfqd =
1911 container_of(work, struct cfq_data, unplug_work);
1912 request_queue_t *q = cfqd->queue;
1913 unsigned long flags;
1915 spin_lock_irqsave(q->queue_lock, flags);
1916 blk_start_queueing(q);
1917 spin_unlock_irqrestore(q->queue_lock, flags);
1921 * Timer running if the active_queue is currently idling inside its time slice
1923 static void cfq_idle_slice_timer(unsigned long data)
1925 struct cfq_data *cfqd = (struct cfq_data *) data;
1926 struct cfq_queue *cfqq;
1927 unsigned long flags;
1930 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1932 if ((cfqq = cfqd->active_queue) != NULL) {
1938 if (cfq_slice_used(cfqq))
1942 * only expire and reinvoke request handler, if there are
1943 * other queues with pending requests
1945 if (!cfqd->busy_queues)
1949 * not expired and it has a request pending, let it dispatch
1951 if (!RB_EMPTY_ROOT(&cfqq->sort_list)) {
1952 cfq_mark_cfqq_must_dispatch(cfqq);
1957 cfq_slice_expired(cfqd, 0, timed_out);
1959 cfq_schedule_dispatch(cfqd);
1961 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1965 * Timer running if an idle class queue is waiting for service
1967 static void cfq_idle_class_timer(unsigned long data)
1969 struct cfq_data *cfqd = (struct cfq_data *) data;
1970 unsigned long flags, end;
1972 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1975 * race with a non-idle queue, reset timer
1977 end = cfqd->last_end_request + CFQ_IDLE_GRACE;
1978 if (!time_after_eq(jiffies, end))
1979 mod_timer(&cfqd->idle_class_timer, end);
1981 cfq_schedule_dispatch(cfqd);
1983 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1986 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
1988 del_timer_sync(&cfqd->idle_slice_timer);
1989 del_timer_sync(&cfqd->idle_class_timer);
1990 blk_sync_queue(cfqd->queue);
1993 static void cfq_exit_queue(elevator_t *e)
1995 struct cfq_data *cfqd = e->elevator_data;
1996 request_queue_t *q = cfqd->queue;
1998 cfq_shutdown_timer_wq(cfqd);
2000 spin_lock_irq(q->queue_lock);
2002 if (cfqd->active_queue)
2003 __cfq_slice_expired(cfqd, cfqd->active_queue, 0, 0);
2005 while (!list_empty(&cfqd->cic_list)) {
2006 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
2007 struct cfq_io_context,
2010 __cfq_exit_single_io_context(cfqd, cic);
2013 spin_unlock_irq(q->queue_lock);
2015 cfq_shutdown_timer_wq(cfqd);
2017 kfree(cfqd->cfq_hash);
2021 static void *cfq_init_queue(request_queue_t *q)
2023 struct cfq_data *cfqd;
2026 cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL, q->node);
2030 memset(cfqd, 0, sizeof(*cfqd));
2032 for (i = 0; i < CFQ_PRIO_LISTS; i++)
2033 INIT_LIST_HEAD(&cfqd->rr_list[i]);
2035 INIT_LIST_HEAD(&cfqd->busy_rr);
2036 INIT_LIST_HEAD(&cfqd->cur_rr);
2037 INIT_LIST_HEAD(&cfqd->idle_rr);
2038 INIT_LIST_HEAD(&cfqd->cic_list);
2040 cfqd->cfq_hash = kmalloc_node(sizeof(struct hlist_head) * CFQ_QHASH_ENTRIES, GFP_KERNEL, q->node);
2041 if (!cfqd->cfq_hash)
2044 for (i = 0; i < CFQ_QHASH_ENTRIES; i++)
2045 INIT_HLIST_HEAD(&cfqd->cfq_hash[i]);
2049 init_timer(&cfqd->idle_slice_timer);
2050 cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
2051 cfqd->idle_slice_timer.data = (unsigned long) cfqd;
2053 init_timer(&cfqd->idle_class_timer);
2054 cfqd->idle_class_timer.function = cfq_idle_class_timer;
2055 cfqd->idle_class_timer.data = (unsigned long) cfqd;
2057 INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
2059 cfqd->cfq_quantum = cfq_quantum;
2060 cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
2061 cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
2062 cfqd->cfq_back_max = cfq_back_max;
2063 cfqd->cfq_back_penalty = cfq_back_penalty;
2064 cfqd->cfq_slice[0] = cfq_slice_async;
2065 cfqd->cfq_slice[1] = cfq_slice_sync;
2066 cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
2067 cfqd->cfq_slice_idle = cfq_slice_idle;
2075 static void cfq_slab_kill(void)
2078 kmem_cache_destroy(cfq_pool);
2080 kmem_cache_destroy(cfq_ioc_pool);
2083 static int __init cfq_slab_setup(void)
2085 cfq_pool = kmem_cache_create("cfq_pool", sizeof(struct cfq_queue), 0, 0,
2090 cfq_ioc_pool = kmem_cache_create("cfq_ioc_pool",
2091 sizeof(struct cfq_io_context), 0, 0, NULL, NULL);
2102 * sysfs parts below -->
2106 cfq_var_show(unsigned int var, char *page)
2108 return sprintf(page, "%d\n", var);
2112 cfq_var_store(unsigned int *var, const char *page, size_t count)
2114 char *p = (char *) page;
2116 *var = simple_strtoul(p, &p, 10);
2120 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
2121 static ssize_t __FUNC(elevator_t *e, char *page) \
2123 struct cfq_data *cfqd = e->elevator_data; \
2124 unsigned int __data = __VAR; \
2126 __data = jiffies_to_msecs(__data); \
2127 return cfq_var_show(__data, (page)); \
2129 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
2130 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
2131 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
2132 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
2133 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
2134 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
2135 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
2136 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
2137 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
2138 #undef SHOW_FUNCTION
2140 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
2141 static ssize_t __FUNC(elevator_t *e, const char *page, size_t count) \
2143 struct cfq_data *cfqd = e->elevator_data; \
2144 unsigned int __data; \
2145 int ret = cfq_var_store(&__data, (page), count); \
2146 if (__data < (MIN)) \
2148 else if (__data > (MAX)) \
2151 *(__PTR) = msecs_to_jiffies(__data); \
2153 *(__PTR) = __data; \
2156 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
2157 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1, UINT_MAX, 1);
2158 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1, UINT_MAX, 1);
2159 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
2160 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1, UINT_MAX, 0);
2161 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
2162 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
2163 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
2164 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1, UINT_MAX, 0);
2165 #undef STORE_FUNCTION
2167 #define CFQ_ATTR(name) \
2168 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
2170 static struct elv_fs_entry cfq_attrs[] = {
2172 CFQ_ATTR(fifo_expire_sync),
2173 CFQ_ATTR(fifo_expire_async),
2174 CFQ_ATTR(back_seek_max),
2175 CFQ_ATTR(back_seek_penalty),
2176 CFQ_ATTR(slice_sync),
2177 CFQ_ATTR(slice_async),
2178 CFQ_ATTR(slice_async_rq),
2179 CFQ_ATTR(slice_idle),
2183 static struct elevator_type iosched_cfq = {
2185 .elevator_merge_fn = cfq_merge,
2186 .elevator_merged_fn = cfq_merged_request,
2187 .elevator_merge_req_fn = cfq_merged_requests,
2188 .elevator_allow_merge_fn = cfq_allow_merge,
2189 .elevator_dispatch_fn = cfq_dispatch_requests,
2190 .elevator_add_req_fn = cfq_insert_request,
2191 .elevator_activate_req_fn = cfq_activate_request,
2192 .elevator_deactivate_req_fn = cfq_deactivate_request,
2193 .elevator_queue_empty_fn = cfq_queue_empty,
2194 .elevator_completed_req_fn = cfq_completed_request,
2195 .elevator_former_req_fn = elv_rb_former_request,
2196 .elevator_latter_req_fn = elv_rb_latter_request,
2197 .elevator_set_req_fn = cfq_set_request,
2198 .elevator_put_req_fn = cfq_put_request,
2199 .elevator_may_queue_fn = cfq_may_queue,
2200 .elevator_init_fn = cfq_init_queue,
2201 .elevator_exit_fn = cfq_exit_queue,
2202 .trim = cfq_free_io_context,
2204 .elevator_attrs = cfq_attrs,
2205 .elevator_name = "cfq",
2206 .elevator_owner = THIS_MODULE,
2209 static int __init cfq_init(void)
2214 * could be 0 on HZ < 1000 setups
2216 if (!cfq_slice_async)
2217 cfq_slice_async = 1;
2218 if (!cfq_slice_idle)
2221 if (cfq_slab_setup())
2224 ret = elv_register(&iosched_cfq);
2231 static void __exit cfq_exit(void)
2233 DECLARE_COMPLETION_ONSTACK(all_gone);
2234 elv_unregister(&iosched_cfq);
2235 ioc_gone = &all_gone;
2236 /* ioc_gone's update must be visible before reading ioc_count */
2238 if (elv_ioc_count_read(ioc_count))
2239 wait_for_completion(ioc_gone);
2244 module_init(cfq_init);
2245 module_exit(cfq_exit);
2247 MODULE_AUTHOR("Jens Axboe");
2248 MODULE_LICENSE("GPL");
2249 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");