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);
538 cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
540 elv_rb_del(&cfqq->sort_list, rq);
541 cfqq->queued[rq_is_sync(rq)]--;
545 static struct request *
546 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
548 struct task_struct *tsk = current;
549 pid_t key = cfq_queue_pid(tsk, bio_data_dir(bio), bio_sync(bio));
550 struct cfq_queue *cfqq;
552 cfqq = cfq_find_cfq_hash(cfqd, key, tsk->ioprio);
554 sector_t sector = bio->bi_sector + bio_sectors(bio);
556 return elv_rb_find(&cfqq->sort_list, sector);
562 static void cfq_activate_request(request_queue_t *q, struct request *rq)
564 struct cfq_data *cfqd = q->elevator->elevator_data;
566 cfqd->rq_in_driver++;
569 * If the depth is larger 1, it really could be queueing. But lets
570 * make the mark a little higher - idling could still be good for
571 * low queueing, and a low queueing number could also just indicate
572 * a SCSI mid layer like behaviour where limit+1 is often seen.
574 if (!cfqd->hw_tag && cfqd->rq_in_driver > 4)
578 static void cfq_deactivate_request(request_queue_t *q, struct request *rq)
580 struct cfq_data *cfqd = q->elevator->elevator_data;
582 WARN_ON(!cfqd->rq_in_driver);
583 cfqd->rq_in_driver--;
586 static void cfq_remove_request(struct request *rq)
588 struct cfq_queue *cfqq = RQ_CFQQ(rq);
590 if (cfqq->next_rq == rq)
591 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
593 list_del_init(&rq->queuelist);
596 if (rq_is_meta(rq)) {
597 WARN_ON(!cfqq->meta_pending);
598 cfqq->meta_pending--;
603 cfq_merge(request_queue_t *q, struct request **req, struct bio *bio)
605 struct cfq_data *cfqd = q->elevator->elevator_data;
606 struct request *__rq;
608 __rq = cfq_find_rq_fmerge(cfqd, bio);
609 if (__rq && elv_rq_merge_ok(__rq, bio)) {
611 return ELEVATOR_FRONT_MERGE;
614 return ELEVATOR_NO_MERGE;
617 static void cfq_merged_request(request_queue_t *q, struct request *req,
620 if (type == ELEVATOR_FRONT_MERGE) {
621 struct cfq_queue *cfqq = RQ_CFQQ(req);
623 cfq_reposition_rq_rb(cfqq, req);
628 cfq_merged_requests(request_queue_t *q, struct request *rq,
629 struct request *next)
632 * reposition in fifo if next is older than rq
634 if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
635 time_before(next->start_time, rq->start_time))
636 list_move(&rq->queuelist, &next->queuelist);
638 cfq_remove_request(next);
641 static int cfq_allow_merge(request_queue_t *q, struct request *rq,
644 struct cfq_data *cfqd = q->elevator->elevator_data;
645 const int rw = bio_data_dir(bio);
646 struct cfq_queue *cfqq;
650 * Disallow merge of a sync bio into an async request.
652 if ((bio_data_dir(bio) == READ || bio_sync(bio)) && !rq_is_sync(rq))
656 * Lookup the cfqq that this bio will be queued with. Allow
657 * merge only if rq is queued there.
659 key = cfq_queue_pid(current, rw, bio_sync(bio));
660 cfqq = cfq_find_cfq_hash(cfqd, key, current->ioprio);
662 if (cfqq == RQ_CFQQ(rq))
669 __cfq_set_active_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
673 * stop potential idle class queues waiting service
675 del_timer(&cfqd->idle_class_timer);
678 cfq_clear_cfqq_must_alloc_slice(cfqq);
679 cfq_clear_cfqq_fifo_expire(cfqq);
680 cfq_mark_cfqq_slice_new(cfqq);
683 cfqd->active_queue = cfqq;
687 * current cfqq expired its slice (or was too idle), select new one
690 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
693 if (cfq_cfqq_wait_request(cfqq))
694 del_timer(&cfqd->idle_slice_timer);
696 if (!preempted && !cfq_cfqq_dispatched(cfqq))
697 cfq_schedule_dispatch(cfqd);
699 cfq_clear_cfqq_must_dispatch(cfqq);
700 cfq_clear_cfqq_wait_request(cfqq);
701 cfq_clear_cfqq_queue_new(cfqq);
704 * store what was left of this slice, if the queue idled out
707 if (!cfq_cfqq_slice_new(cfqq))
708 cfqq->slice_resid = cfqq->slice_end - jiffies;
710 cfq_resort_rr_list(cfqq, preempted);
712 if (cfqq == cfqd->active_queue)
713 cfqd->active_queue = NULL;
715 if (cfqd->active_cic) {
716 put_io_context(cfqd->active_cic->ioc);
717 cfqd->active_cic = NULL;
720 cfqd->dispatch_slice = 0;
723 static inline void cfq_slice_expired(struct cfq_data *cfqd, int preempted)
725 struct cfq_queue *cfqq = cfqd->active_queue;
728 __cfq_slice_expired(cfqd, cfqq, preempted);
741 static int cfq_get_next_prio_level(struct cfq_data *cfqd)
750 for (p = cfqd->cur_prio; p <= cfqd->cur_end_prio; p++) {
751 if (!list_empty(&cfqd->rr_list[p])) {
760 if (++cfqd->cur_end_prio == CFQ_PRIO_LISTS) {
761 cfqd->cur_end_prio = 0;
768 if (unlikely(prio == -1))
771 BUG_ON(prio >= CFQ_PRIO_LISTS);
773 list_splice_init(&cfqd->rr_list[prio], &cfqd->cur_rr);
775 cfqd->cur_prio = prio + 1;
776 if (cfqd->cur_prio > cfqd->cur_end_prio) {
777 cfqd->cur_end_prio = cfqd->cur_prio;
780 if (cfqd->cur_end_prio == CFQ_PRIO_LISTS) {
782 cfqd->cur_end_prio = 0;
788 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd)
790 struct cfq_queue *cfqq = NULL;
792 if (!list_empty(&cfqd->cur_rr) || cfq_get_next_prio_level(cfqd) != -1) {
794 * if current list is non-empty, grab first entry. if it is
795 * empty, get next prio level and grab first entry then if any
798 cfqq = list_entry_cfqq(cfqd->cur_rr.next);
799 } else if (!list_empty(&cfqd->busy_rr)) {
801 * If no new queues are available, check if the busy list has
802 * some before falling back to idle io.
804 cfqq = list_entry_cfqq(cfqd->busy_rr.next);
805 } else if (!list_empty(&cfqd->idle_rr)) {
807 * if we have idle queues and no rt or be queues had pending
808 * requests, either allow immediate service if the grace period
809 * has passed or arm the idle grace timer
811 unsigned long end = cfqd->last_end_request + CFQ_IDLE_GRACE;
813 if (time_after_eq(jiffies, end))
814 cfqq = list_entry_cfqq(cfqd->idle_rr.next);
816 mod_timer(&cfqd->idle_class_timer, end);
819 __cfq_set_active_queue(cfqd, cfqq);
823 #define CIC_SEEKY(cic) ((cic)->seek_mean > (128 * 1024))
825 static int cfq_arm_slice_timer(struct cfq_data *cfqd)
827 struct cfq_queue *cfqq = cfqd->active_queue;
828 struct cfq_io_context *cic;
831 WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
834 * idle is disabled, either manually or by past process history
836 if (!cfqd->cfq_slice_idle)
838 if (!cfq_cfqq_idle_window(cfqq))
841 * task has exited, don't wait
843 cic = cfqd->active_cic;
844 if (!cic || !cic->ioc->task)
847 cfq_mark_cfqq_must_dispatch(cfqq);
848 cfq_mark_cfqq_wait_request(cfqq);
850 sl = min(cfqq->slice_end - 1, (unsigned long) cfqd->cfq_slice_idle);
853 * we don't want to idle for seeks, but we do want to allow
854 * fair distribution of slice time for a process doing back-to-back
855 * seeks. so allow a little bit of time for him to submit a new rq
857 if (sample_valid(cic->seek_samples) && CIC_SEEKY(cic))
858 sl = min(sl, msecs_to_jiffies(2));
860 mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
864 static void cfq_dispatch_insert(request_queue_t *q, struct request *rq)
866 struct cfq_data *cfqd = q->elevator->elevator_data;
867 struct cfq_queue *cfqq = RQ_CFQQ(rq);
869 cfq_remove_request(rq);
870 cfqq->on_dispatch[rq_is_sync(rq)]++;
871 elv_dispatch_sort(q, rq);
873 rq = list_entry(q->queue_head.prev, struct request, queuelist);
874 cfqd->last_sector = rq->sector + rq->nr_sectors;
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);
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 ((!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);
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);
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) {
1057 * Don't repeat dispatch from the previous queue.
1059 if (prev_cfqq == cfqq)
1062 cfq_clear_cfqq_must_dispatch(cfqq);
1063 cfq_clear_cfqq_wait_request(cfqq);
1064 del_timer(&cfqd->idle_slice_timer);
1066 max_dispatch = cfqd->cfq_quantum;
1067 if (cfq_class_idle(cfqq))
1070 dispatched += __cfq_dispatch_requests(cfqd, cfqq, max_dispatch);
1073 * If the dispatch cfqq has idling enabled and is still
1074 * the active queue, break out.
1076 if (cfq_cfqq_idle_window(cfqq) && cfqd->active_queue)
1086 * task holds one reference to the queue, dropped when task exits. each rq
1087 * in-flight on this queue also holds a reference, dropped when rq is freed.
1089 * queue lock must be held here.
1091 static void cfq_put_queue(struct cfq_queue *cfqq)
1093 struct cfq_data *cfqd = cfqq->cfqd;
1095 BUG_ON(atomic_read(&cfqq->ref) <= 0);
1097 if (!atomic_dec_and_test(&cfqq->ref))
1100 BUG_ON(rb_first(&cfqq->sort_list));
1101 BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
1102 BUG_ON(cfq_cfqq_on_rr(cfqq));
1104 if (unlikely(cfqd->active_queue == cfqq))
1105 __cfq_slice_expired(cfqd, cfqq, 0);
1108 * it's on the empty list and still hashed
1110 list_del(&cfqq->cfq_list);
1111 hlist_del(&cfqq->cfq_hash);
1112 kmem_cache_free(cfq_pool, cfqq);
1115 static struct cfq_queue *
1116 __cfq_find_cfq_hash(struct cfq_data *cfqd, unsigned int key, unsigned int prio,
1119 struct hlist_head *hash_list = &cfqd->cfq_hash[hashval];
1120 struct hlist_node *entry;
1121 struct cfq_queue *__cfqq;
1123 hlist_for_each_entry(__cfqq, entry, hash_list, cfq_hash) {
1124 const unsigned short __p = IOPRIO_PRIO_VALUE(__cfqq->org_ioprio_class, __cfqq->org_ioprio);
1126 if (__cfqq->key == key && (__p == prio || !prio))
1133 static struct cfq_queue *
1134 cfq_find_cfq_hash(struct cfq_data *cfqd, unsigned int key, unsigned short prio)
1136 return __cfq_find_cfq_hash(cfqd, key, prio, hash_long(key, CFQ_QHASH_SHIFT));
1139 static void cfq_free_io_context(struct io_context *ioc)
1141 struct cfq_io_context *__cic;
1145 while ((n = rb_first(&ioc->cic_root)) != NULL) {
1146 __cic = rb_entry(n, struct cfq_io_context, rb_node);
1147 rb_erase(&__cic->rb_node, &ioc->cic_root);
1148 kmem_cache_free(cfq_ioc_pool, __cic);
1152 elv_ioc_count_mod(ioc_count, -freed);
1154 if (ioc_gone && !elv_ioc_count_read(ioc_count))
1158 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1160 if (unlikely(cfqq == cfqd->active_queue))
1161 __cfq_slice_expired(cfqd, cfqq, 0);
1163 cfq_put_queue(cfqq);
1166 static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
1167 struct cfq_io_context *cic)
1169 list_del_init(&cic->queue_list);
1173 if (cic->cfqq[ASYNC]) {
1174 cfq_exit_cfqq(cfqd, cic->cfqq[ASYNC]);
1175 cic->cfqq[ASYNC] = NULL;
1178 if (cic->cfqq[SYNC]) {
1179 cfq_exit_cfqq(cfqd, cic->cfqq[SYNC]);
1180 cic->cfqq[SYNC] = NULL;
1186 * Called with interrupts disabled
1188 static void cfq_exit_single_io_context(struct cfq_io_context *cic)
1190 struct cfq_data *cfqd = cic->key;
1193 request_queue_t *q = cfqd->queue;
1195 spin_lock_irq(q->queue_lock);
1196 __cfq_exit_single_io_context(cfqd, cic);
1197 spin_unlock_irq(q->queue_lock);
1201 static void cfq_exit_io_context(struct io_context *ioc)
1203 struct cfq_io_context *__cic;
1207 * put the reference this task is holding to the various queues
1210 n = rb_first(&ioc->cic_root);
1212 __cic = rb_entry(n, struct cfq_io_context, rb_node);
1214 cfq_exit_single_io_context(__cic);
1219 static struct cfq_io_context *
1220 cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
1222 struct cfq_io_context *cic;
1224 cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask, cfqd->queue->node);
1226 memset(cic, 0, sizeof(*cic));
1227 cic->last_end_request = jiffies;
1228 INIT_LIST_HEAD(&cic->queue_list);
1229 cic->dtor = cfq_free_io_context;
1230 cic->exit = cfq_exit_io_context;
1231 elv_ioc_count_inc(ioc_count);
1237 static void cfq_init_prio_data(struct cfq_queue *cfqq)
1239 struct task_struct *tsk = current;
1242 if (!cfq_cfqq_prio_changed(cfqq))
1245 ioprio_class = IOPRIO_PRIO_CLASS(tsk->ioprio);
1246 switch (ioprio_class) {
1248 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
1249 case IOPRIO_CLASS_NONE:
1251 * no prio set, place us in the middle of the BE classes
1253 cfqq->ioprio = task_nice_ioprio(tsk);
1254 cfqq->ioprio_class = IOPRIO_CLASS_BE;
1256 case IOPRIO_CLASS_RT:
1257 cfqq->ioprio = task_ioprio(tsk);
1258 cfqq->ioprio_class = IOPRIO_CLASS_RT;
1260 case IOPRIO_CLASS_BE:
1261 cfqq->ioprio = task_ioprio(tsk);
1262 cfqq->ioprio_class = IOPRIO_CLASS_BE;
1264 case IOPRIO_CLASS_IDLE:
1265 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
1267 cfq_clear_cfqq_idle_window(cfqq);
1272 * keep track of original prio settings in case we have to temporarily
1273 * elevate the priority of this queue
1275 cfqq->org_ioprio = cfqq->ioprio;
1276 cfqq->org_ioprio_class = cfqq->ioprio_class;
1278 cfq_resort_rr_list(cfqq, 0);
1279 cfq_clear_cfqq_prio_changed(cfqq);
1282 static inline void changed_ioprio(struct cfq_io_context *cic)
1284 struct cfq_data *cfqd = cic->key;
1285 struct cfq_queue *cfqq;
1286 unsigned long flags;
1288 if (unlikely(!cfqd))
1291 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1293 cfqq = cic->cfqq[ASYNC];
1295 struct cfq_queue *new_cfqq;
1296 new_cfqq = cfq_get_queue(cfqd, CFQ_KEY_ASYNC, cic->ioc->task,
1299 cic->cfqq[ASYNC] = new_cfqq;
1300 cfq_put_queue(cfqq);
1304 cfqq = cic->cfqq[SYNC];
1306 cfq_mark_cfqq_prio_changed(cfqq);
1308 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1311 static void cfq_ioc_set_ioprio(struct io_context *ioc)
1313 struct cfq_io_context *cic;
1316 ioc->ioprio_changed = 0;
1318 n = rb_first(&ioc->cic_root);
1320 cic = rb_entry(n, struct cfq_io_context, rb_node);
1322 changed_ioprio(cic);
1327 static struct cfq_queue *
1328 cfq_get_queue(struct cfq_data *cfqd, unsigned int key, struct task_struct *tsk,
1331 const int hashval = hash_long(key, CFQ_QHASH_SHIFT);
1332 struct cfq_queue *cfqq, *new_cfqq = NULL;
1333 unsigned short ioprio;
1336 ioprio = tsk->ioprio;
1337 cfqq = __cfq_find_cfq_hash(cfqd, key, ioprio, hashval);
1343 } else if (gfp_mask & __GFP_WAIT) {
1345 * Inform the allocator of the fact that we will
1346 * just repeat this allocation if it fails, to allow
1347 * the allocator to do whatever it needs to attempt to
1350 spin_unlock_irq(cfqd->queue->queue_lock);
1351 new_cfqq = kmem_cache_alloc_node(cfq_pool, gfp_mask|__GFP_NOFAIL, cfqd->queue->node);
1352 spin_lock_irq(cfqd->queue->queue_lock);
1355 cfqq = kmem_cache_alloc_node(cfq_pool, gfp_mask, cfqd->queue->node);
1360 memset(cfqq, 0, sizeof(*cfqq));
1362 INIT_HLIST_NODE(&cfqq->cfq_hash);
1363 INIT_LIST_HEAD(&cfqq->cfq_list);
1364 INIT_LIST_HEAD(&cfqq->fifo);
1367 hlist_add_head(&cfqq->cfq_hash, &cfqd->cfq_hash[hashval]);
1368 atomic_set(&cfqq->ref, 0);
1371 cfq_mark_cfqq_idle_window(cfqq);
1372 cfq_mark_cfqq_prio_changed(cfqq);
1373 cfq_mark_cfqq_queue_new(cfqq);
1374 cfq_init_prio_data(cfqq);
1378 kmem_cache_free(cfq_pool, new_cfqq);
1380 atomic_inc(&cfqq->ref);
1382 WARN_ON((gfp_mask & __GFP_WAIT) && !cfqq);
1387 cfq_drop_dead_cic(struct io_context *ioc, struct cfq_io_context *cic)
1389 WARN_ON(!list_empty(&cic->queue_list));
1390 rb_erase(&cic->rb_node, &ioc->cic_root);
1391 kmem_cache_free(cfq_ioc_pool, cic);
1392 elv_ioc_count_dec(ioc_count);
1395 static struct cfq_io_context *
1396 cfq_cic_rb_lookup(struct cfq_data *cfqd, struct io_context *ioc)
1399 struct cfq_io_context *cic;
1400 void *k, *key = cfqd;
1403 n = ioc->cic_root.rb_node;
1405 cic = rb_entry(n, struct cfq_io_context, rb_node);
1406 /* ->key must be copied to avoid race with cfq_exit_queue() */
1409 cfq_drop_dead_cic(ioc, cic);
1425 cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
1426 struct cfq_io_context *cic)
1429 struct rb_node *parent;
1430 struct cfq_io_context *__cic;
1431 unsigned long flags;
1439 p = &ioc->cic_root.rb_node;
1442 __cic = rb_entry(parent, struct cfq_io_context, rb_node);
1443 /* ->key must be copied to avoid race with cfq_exit_queue() */
1446 cfq_drop_dead_cic(ioc, __cic);
1452 else if (cic->key > k)
1453 p = &(*p)->rb_right;
1458 rb_link_node(&cic->rb_node, parent, p);
1459 rb_insert_color(&cic->rb_node, &ioc->cic_root);
1461 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1462 list_add(&cic->queue_list, &cfqd->cic_list);
1463 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1467 * Setup general io context and cfq io context. There can be several cfq
1468 * io contexts per general io context, if this process is doing io to more
1469 * than one device managed by cfq.
1471 static struct cfq_io_context *
1472 cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
1474 struct io_context *ioc = NULL;
1475 struct cfq_io_context *cic;
1477 might_sleep_if(gfp_mask & __GFP_WAIT);
1479 ioc = get_io_context(gfp_mask, cfqd->queue->node);
1483 cic = cfq_cic_rb_lookup(cfqd, ioc);
1487 cic = cfq_alloc_io_context(cfqd, gfp_mask);
1491 cfq_cic_link(cfqd, ioc, cic);
1493 smp_read_barrier_depends();
1494 if (unlikely(ioc->ioprio_changed))
1495 cfq_ioc_set_ioprio(ioc);
1499 put_io_context(ioc);
1504 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
1506 unsigned long elapsed = jiffies - cic->last_end_request;
1507 unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
1509 cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
1510 cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
1511 cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
1515 cfq_update_io_seektime(struct cfq_io_context *cic, struct request *rq)
1520 if (cic->last_request_pos < rq->sector)
1521 sdist = rq->sector - cic->last_request_pos;
1523 sdist = cic->last_request_pos - rq->sector;
1526 * Don't allow the seek distance to get too large from the
1527 * odd fragment, pagein, etc
1529 if (cic->seek_samples <= 60) /* second&third seek */
1530 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*1024);
1532 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*64);
1534 cic->seek_samples = (7*cic->seek_samples + 256) / 8;
1535 cic->seek_total = (7*cic->seek_total + (u64)256*sdist) / 8;
1536 total = cic->seek_total + (cic->seek_samples/2);
1537 do_div(total, cic->seek_samples);
1538 cic->seek_mean = (sector_t)total;
1542 * Disable idle window if the process thinks too long or seeks so much that
1546 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1547 struct cfq_io_context *cic)
1549 int enable_idle = cfq_cfqq_idle_window(cfqq);
1551 if (!cic->ioc->task || !cfqd->cfq_slice_idle ||
1552 (cfqd->hw_tag && CIC_SEEKY(cic)))
1554 else if (sample_valid(cic->ttime_samples)) {
1555 if (cic->ttime_mean > cfqd->cfq_slice_idle)
1562 cfq_mark_cfqq_idle_window(cfqq);
1564 cfq_clear_cfqq_idle_window(cfqq);
1569 * Check if new_cfqq should preempt the currently active queue. Return 0 for
1570 * no or if we aren't sure, a 1 will cause a preempt.
1573 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
1576 struct cfq_queue *cfqq = cfqd->active_queue;
1578 if (cfq_class_idle(new_cfqq))
1584 if (cfq_class_idle(cfqq))
1586 if (!cfq_cfqq_wait_request(new_cfqq))
1589 * if the new request is sync, but the currently running queue is
1590 * not, let the sync request have priority.
1592 if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
1595 * So both queues are sync. Let the new request get disk time if
1596 * it's a metadata request and the current queue is doing regular IO.
1598 if (rq_is_meta(rq) && !cfqq->meta_pending)
1605 * cfqq preempts the active queue. if we allowed preempt with no slice left,
1606 * let it have half of its nominal slice.
1608 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1610 cfq_slice_expired(cfqd, 1);
1613 * Put the new queue at the front of the of the current list,
1614 * so we know that it will be selected next.
1616 BUG_ON(!cfq_cfqq_on_rr(cfqq));
1617 list_move(&cfqq->cfq_list, &cfqd->cur_rr);
1619 cfqq->slice_end = 0;
1620 cfq_mark_cfqq_slice_new(cfqq);
1624 * Called when a new fs request (rq) is added (to cfqq). Check if there's
1625 * something we should do about it
1628 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1631 struct cfq_io_context *cic = RQ_CIC(rq);
1634 cfqq->meta_pending++;
1637 * check if this request is a better next-serve candidate)) {
1639 cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq);
1640 BUG_ON(!cfqq->next_rq);
1643 * we never wait for an async request and we don't allow preemption
1644 * of an async request. so just return early
1646 if (!rq_is_sync(rq)) {
1648 * sync process issued an async request, if it's waiting
1649 * then expire it and kick rq handling.
1651 if (cic == cfqd->active_cic &&
1652 del_timer(&cfqd->idle_slice_timer)) {
1653 cfq_slice_expired(cfqd, 0);
1654 blk_start_queueing(cfqd->queue);
1659 cfq_update_io_thinktime(cfqd, cic);
1660 cfq_update_io_seektime(cic, rq);
1661 cfq_update_idle_window(cfqd, cfqq, cic);
1663 cic->last_request_pos = rq->sector + rq->nr_sectors;
1665 if (cfqq == cfqd->active_queue) {
1667 * if we are waiting for a request for this queue, let it rip
1668 * immediately and flag that we must not expire this queue
1671 if (cfq_cfqq_wait_request(cfqq)) {
1672 cfq_mark_cfqq_must_dispatch(cfqq);
1673 del_timer(&cfqd->idle_slice_timer);
1674 blk_start_queueing(cfqd->queue);
1676 } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
1678 * not the active queue - expire current slice if it is
1679 * idle and has expired it's mean thinktime or this new queue
1680 * has some old slice time left and is of higher priority
1682 cfq_preempt_queue(cfqd, cfqq);
1683 cfq_mark_cfqq_must_dispatch(cfqq);
1684 blk_start_queueing(cfqd->queue);
1688 static void cfq_insert_request(request_queue_t *q, struct request *rq)
1690 struct cfq_data *cfqd = q->elevator->elevator_data;
1691 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1693 cfq_init_prio_data(cfqq);
1697 list_add_tail(&rq->queuelist, &cfqq->fifo);
1699 cfq_rq_enqueued(cfqd, cfqq, rq);
1702 static void cfq_completed_request(request_queue_t *q, struct request *rq)
1704 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1705 struct cfq_data *cfqd = cfqq->cfqd;
1706 const int sync = rq_is_sync(rq);
1711 WARN_ON(!cfqd->rq_in_driver);
1712 WARN_ON(!cfqq->on_dispatch[sync]);
1713 cfqd->rq_in_driver--;
1714 cfqq->on_dispatch[sync]--;
1715 cfqq->service_last = now;
1717 if (!cfq_class_idle(cfqq))
1718 cfqd->last_end_request = now;
1720 cfq_resort_rr_list(cfqq, 0);
1723 RQ_CIC(rq)->last_end_request = now;
1726 * If this is the active queue, check if it needs to be expired,
1727 * or if we want to idle in case it has no pending requests.
1729 if (cfqd->active_queue == cfqq) {
1730 if (cfq_cfqq_slice_new(cfqq)) {
1731 cfq_set_prio_slice(cfqd, cfqq);
1732 cfq_clear_cfqq_slice_new(cfqq);
1734 if (cfq_slice_used(cfqq))
1735 cfq_slice_expired(cfqd, 0);
1736 else if (sync && RB_EMPTY_ROOT(&cfqq->sort_list)) {
1737 if (!cfq_arm_slice_timer(cfqd))
1738 cfq_schedule_dispatch(cfqd);
1744 * we temporarily boost lower priority queues if they are holding fs exclusive
1745 * resources. they are boosted to normal prio (CLASS_BE/4)
1747 static void cfq_prio_boost(struct cfq_queue *cfqq)
1749 const int ioprio_class = cfqq->ioprio_class;
1750 const int ioprio = cfqq->ioprio;
1752 if (has_fs_excl()) {
1754 * boost idle prio on transactions that would lock out other
1755 * users of the filesystem
1757 if (cfq_class_idle(cfqq))
1758 cfqq->ioprio_class = IOPRIO_CLASS_BE;
1759 if (cfqq->ioprio > IOPRIO_NORM)
1760 cfqq->ioprio = IOPRIO_NORM;
1763 * check if we need to unboost the queue
1765 if (cfqq->ioprio_class != cfqq->org_ioprio_class)
1766 cfqq->ioprio_class = cfqq->org_ioprio_class;
1767 if (cfqq->ioprio != cfqq->org_ioprio)
1768 cfqq->ioprio = cfqq->org_ioprio;
1772 * refile between round-robin lists if we moved the priority class
1774 if ((ioprio_class != cfqq->ioprio_class || ioprio != cfqq->ioprio))
1775 cfq_resort_rr_list(cfqq, 0);
1778 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
1780 if ((cfq_cfqq_wait_request(cfqq) || cfq_cfqq_must_alloc(cfqq)) &&
1781 !cfq_cfqq_must_alloc_slice(cfqq)) {
1782 cfq_mark_cfqq_must_alloc_slice(cfqq);
1783 return ELV_MQUEUE_MUST;
1786 return ELV_MQUEUE_MAY;
1789 static int cfq_may_queue(request_queue_t *q, int rw)
1791 struct cfq_data *cfqd = q->elevator->elevator_data;
1792 struct task_struct *tsk = current;
1793 struct cfq_queue *cfqq;
1796 key = cfq_queue_pid(tsk, rw, rw & REQ_RW_SYNC);
1799 * don't force setup of a queue from here, as a call to may_queue
1800 * does not necessarily imply that a request actually will be queued.
1801 * so just lookup a possibly existing queue, or return 'may queue'
1804 cfqq = cfq_find_cfq_hash(cfqd, key, tsk->ioprio);
1806 cfq_init_prio_data(cfqq);
1807 cfq_prio_boost(cfqq);
1809 return __cfq_may_queue(cfqq);
1812 return ELV_MQUEUE_MAY;
1816 * queue lock held here
1818 static void cfq_put_request(struct request *rq)
1820 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1823 const int rw = rq_data_dir(rq);
1825 BUG_ON(!cfqq->allocated[rw]);
1826 cfqq->allocated[rw]--;
1828 put_io_context(RQ_CIC(rq)->ioc);
1830 rq->elevator_private = NULL;
1831 rq->elevator_private2 = NULL;
1833 cfq_put_queue(cfqq);
1838 * Allocate cfq data structures associated with this request.
1841 cfq_set_request(request_queue_t *q, struct request *rq, gfp_t gfp_mask)
1843 struct cfq_data *cfqd = q->elevator->elevator_data;
1844 struct task_struct *tsk = current;
1845 struct cfq_io_context *cic;
1846 const int rw = rq_data_dir(rq);
1847 const int is_sync = rq_is_sync(rq);
1848 pid_t key = cfq_queue_pid(tsk, rw, is_sync);
1849 struct cfq_queue *cfqq;
1850 unsigned long flags;
1852 might_sleep_if(gfp_mask & __GFP_WAIT);
1854 cic = cfq_get_io_context(cfqd, gfp_mask);
1856 spin_lock_irqsave(q->queue_lock, flags);
1861 if (!cic->cfqq[is_sync]) {
1862 cfqq = cfq_get_queue(cfqd, key, tsk, gfp_mask);
1866 cic->cfqq[is_sync] = cfqq;
1868 cfqq = cic->cfqq[is_sync];
1870 cfqq->allocated[rw]++;
1871 cfq_clear_cfqq_must_alloc(cfqq);
1872 atomic_inc(&cfqq->ref);
1874 spin_unlock_irqrestore(q->queue_lock, flags);
1876 rq->elevator_private = cic;
1877 rq->elevator_private2 = cfqq;
1882 put_io_context(cic->ioc);
1884 cfq_schedule_dispatch(cfqd);
1885 spin_unlock_irqrestore(q->queue_lock, flags);
1889 static void cfq_kick_queue(struct work_struct *work)
1891 struct cfq_data *cfqd =
1892 container_of(work, struct cfq_data, unplug_work);
1893 request_queue_t *q = cfqd->queue;
1894 unsigned long flags;
1896 spin_lock_irqsave(q->queue_lock, flags);
1897 blk_start_queueing(q);
1898 spin_unlock_irqrestore(q->queue_lock, flags);
1902 * Timer running if the active_queue is currently idling inside its time slice
1904 static void cfq_idle_slice_timer(unsigned long data)
1906 struct cfq_data *cfqd = (struct cfq_data *) data;
1907 struct cfq_queue *cfqq;
1908 unsigned long flags;
1910 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1912 if ((cfqq = cfqd->active_queue) != NULL) {
1916 if (cfq_slice_used(cfqq))
1920 * only expire and reinvoke request handler, if there are
1921 * other queues with pending requests
1923 if (!cfqd->busy_queues)
1927 * not expired and it has a request pending, let it dispatch
1929 if (!RB_EMPTY_ROOT(&cfqq->sort_list)) {
1930 cfq_mark_cfqq_must_dispatch(cfqq);
1935 cfq_slice_expired(cfqd, 0);
1937 cfq_schedule_dispatch(cfqd);
1939 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1943 * Timer running if an idle class queue is waiting for service
1945 static void cfq_idle_class_timer(unsigned long data)
1947 struct cfq_data *cfqd = (struct cfq_data *) data;
1948 unsigned long flags, end;
1950 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1953 * race with a non-idle queue, reset timer
1955 end = cfqd->last_end_request + CFQ_IDLE_GRACE;
1956 if (!time_after_eq(jiffies, end))
1957 mod_timer(&cfqd->idle_class_timer, end);
1959 cfq_schedule_dispatch(cfqd);
1961 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1964 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
1966 del_timer_sync(&cfqd->idle_slice_timer);
1967 del_timer_sync(&cfqd->idle_class_timer);
1968 blk_sync_queue(cfqd->queue);
1971 static void cfq_exit_queue(elevator_t *e)
1973 struct cfq_data *cfqd = e->elevator_data;
1974 request_queue_t *q = cfqd->queue;
1976 cfq_shutdown_timer_wq(cfqd);
1978 spin_lock_irq(q->queue_lock);
1980 if (cfqd->active_queue)
1981 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
1983 while (!list_empty(&cfqd->cic_list)) {
1984 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
1985 struct cfq_io_context,
1988 __cfq_exit_single_io_context(cfqd, cic);
1991 spin_unlock_irq(q->queue_lock);
1993 cfq_shutdown_timer_wq(cfqd);
1995 kfree(cfqd->cfq_hash);
1999 static void *cfq_init_queue(request_queue_t *q)
2001 struct cfq_data *cfqd;
2004 cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL, q->node);
2008 memset(cfqd, 0, sizeof(*cfqd));
2010 for (i = 0; i < CFQ_PRIO_LISTS; i++)
2011 INIT_LIST_HEAD(&cfqd->rr_list[i]);
2013 INIT_LIST_HEAD(&cfqd->busy_rr);
2014 INIT_LIST_HEAD(&cfqd->cur_rr);
2015 INIT_LIST_HEAD(&cfqd->idle_rr);
2016 INIT_LIST_HEAD(&cfqd->cic_list);
2018 cfqd->cfq_hash = kmalloc_node(sizeof(struct hlist_head) * CFQ_QHASH_ENTRIES, GFP_KERNEL, q->node);
2019 if (!cfqd->cfq_hash)
2022 for (i = 0; i < CFQ_QHASH_ENTRIES; i++)
2023 INIT_HLIST_HEAD(&cfqd->cfq_hash[i]);
2027 init_timer(&cfqd->idle_slice_timer);
2028 cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
2029 cfqd->idle_slice_timer.data = (unsigned long) cfqd;
2031 init_timer(&cfqd->idle_class_timer);
2032 cfqd->idle_class_timer.function = cfq_idle_class_timer;
2033 cfqd->idle_class_timer.data = (unsigned long) cfqd;
2035 INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
2037 cfqd->cfq_quantum = cfq_quantum;
2038 cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
2039 cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
2040 cfqd->cfq_back_max = cfq_back_max;
2041 cfqd->cfq_back_penalty = cfq_back_penalty;
2042 cfqd->cfq_slice[0] = cfq_slice_async;
2043 cfqd->cfq_slice[1] = cfq_slice_sync;
2044 cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
2045 cfqd->cfq_slice_idle = cfq_slice_idle;
2053 static void cfq_slab_kill(void)
2056 kmem_cache_destroy(cfq_pool);
2058 kmem_cache_destroy(cfq_ioc_pool);
2061 static int __init cfq_slab_setup(void)
2063 cfq_pool = kmem_cache_create("cfq_pool", sizeof(struct cfq_queue), 0, 0,
2068 cfq_ioc_pool = kmem_cache_create("cfq_ioc_pool",
2069 sizeof(struct cfq_io_context), 0, 0, NULL, NULL);
2080 * sysfs parts below -->
2084 cfq_var_show(unsigned int var, char *page)
2086 return sprintf(page, "%d\n", var);
2090 cfq_var_store(unsigned int *var, const char *page, size_t count)
2092 char *p = (char *) page;
2094 *var = simple_strtoul(p, &p, 10);
2098 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
2099 static ssize_t __FUNC(elevator_t *e, char *page) \
2101 struct cfq_data *cfqd = e->elevator_data; \
2102 unsigned int __data = __VAR; \
2104 __data = jiffies_to_msecs(__data); \
2105 return cfq_var_show(__data, (page)); \
2107 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
2108 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
2109 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
2110 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
2111 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
2112 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
2113 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
2114 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
2115 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
2116 #undef SHOW_FUNCTION
2118 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
2119 static ssize_t __FUNC(elevator_t *e, const char *page, size_t count) \
2121 struct cfq_data *cfqd = e->elevator_data; \
2122 unsigned int __data; \
2123 int ret = cfq_var_store(&__data, (page), count); \
2124 if (__data < (MIN)) \
2126 else if (__data > (MAX)) \
2129 *(__PTR) = msecs_to_jiffies(__data); \
2131 *(__PTR) = __data; \
2134 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
2135 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1, UINT_MAX, 1);
2136 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1, UINT_MAX, 1);
2137 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
2138 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1, UINT_MAX, 0);
2139 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
2140 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
2141 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
2142 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1, UINT_MAX, 0);
2143 #undef STORE_FUNCTION
2145 #define CFQ_ATTR(name) \
2146 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
2148 static struct elv_fs_entry cfq_attrs[] = {
2150 CFQ_ATTR(fifo_expire_sync),
2151 CFQ_ATTR(fifo_expire_async),
2152 CFQ_ATTR(back_seek_max),
2153 CFQ_ATTR(back_seek_penalty),
2154 CFQ_ATTR(slice_sync),
2155 CFQ_ATTR(slice_async),
2156 CFQ_ATTR(slice_async_rq),
2157 CFQ_ATTR(slice_idle),
2161 static struct elevator_type iosched_cfq = {
2163 .elevator_merge_fn = cfq_merge,
2164 .elevator_merged_fn = cfq_merged_request,
2165 .elevator_merge_req_fn = cfq_merged_requests,
2166 .elevator_allow_merge_fn = cfq_allow_merge,
2167 .elevator_dispatch_fn = cfq_dispatch_requests,
2168 .elevator_add_req_fn = cfq_insert_request,
2169 .elevator_activate_req_fn = cfq_activate_request,
2170 .elevator_deactivate_req_fn = cfq_deactivate_request,
2171 .elevator_queue_empty_fn = cfq_queue_empty,
2172 .elevator_completed_req_fn = cfq_completed_request,
2173 .elevator_former_req_fn = elv_rb_former_request,
2174 .elevator_latter_req_fn = elv_rb_latter_request,
2175 .elevator_set_req_fn = cfq_set_request,
2176 .elevator_put_req_fn = cfq_put_request,
2177 .elevator_may_queue_fn = cfq_may_queue,
2178 .elevator_init_fn = cfq_init_queue,
2179 .elevator_exit_fn = cfq_exit_queue,
2180 .trim = cfq_free_io_context,
2182 .elevator_attrs = cfq_attrs,
2183 .elevator_name = "cfq",
2184 .elevator_owner = THIS_MODULE,
2187 static int __init cfq_init(void)
2192 * could be 0 on HZ < 1000 setups
2194 if (!cfq_slice_async)
2195 cfq_slice_async = 1;
2196 if (!cfq_slice_idle)
2199 if (cfq_slab_setup())
2202 ret = elv_register(&iosched_cfq);
2209 static void __exit cfq_exit(void)
2211 DECLARE_COMPLETION_ONSTACK(all_gone);
2212 elv_unregister(&iosched_cfq);
2213 ioc_gone = &all_gone;
2214 /* ioc_gone's update must be visible before reading ioc_count */
2216 if (elv_ioc_count_read(ioc_count))
2217 wait_for_completion(ioc_gone);
2222 module_init(cfq_init);
2223 module_exit(cfq_exit);
2225 MODULE_AUTHOR("Jens Axboe");
2226 MODULE_LICENSE("GPL");
2227 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");