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 slice_left;
151 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;
257 * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
258 * isn't valid until the first request from the dispatch is activated
259 * and the slice time set.
261 static inline int cfq_slice_used(struct cfq_queue *cfqq)
263 if (cfq_cfqq_slice_new(cfqq))
265 if (time_before(jiffies, cfqq->slice_end))
272 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
273 * We choose the request that is closest to the head right now. Distance
274 * behind the head is penalized and only allowed to a certain extent.
276 static struct request *
277 cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2)
279 sector_t last, s1, s2, d1 = 0, d2 = 0;
280 unsigned long back_max;
281 #define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */
282 #define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */
283 unsigned wrap = 0; /* bit mask: requests behind the disk head? */
285 if (rq1 == NULL || rq1 == rq2)
290 if (rq_is_sync(rq1) && !rq_is_sync(rq2))
292 else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
294 if (rq_is_meta(rq1) && !rq_is_meta(rq2))
296 else if (rq_is_meta(rq2) && !rq_is_meta(rq1))
302 last = cfqd->last_sector;
305 * by definition, 1KiB is 2 sectors
307 back_max = cfqd->cfq_back_max * 2;
310 * Strict one way elevator _except_ in the case where we allow
311 * short backward seeks which are biased as twice the cost of a
312 * similar forward seek.
316 else if (s1 + back_max >= last)
317 d1 = (last - s1) * cfqd->cfq_back_penalty;
319 wrap |= CFQ_RQ1_WRAP;
323 else if (s2 + back_max >= last)
324 d2 = (last - s2) * cfqd->cfq_back_penalty;
326 wrap |= CFQ_RQ2_WRAP;
328 /* Found required data */
331 * By doing switch() on the bit mask "wrap" we avoid having to
332 * check two variables for all permutations: --> faster!
335 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
351 case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
354 * Since both rqs are wrapped,
355 * start with the one that's further behind head
356 * (--> only *one* back seek required),
357 * since back seek takes more time than forward.
367 * would be nice to take fifo expire time into account as well
369 static struct request *
370 cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
371 struct request *last)
373 struct rb_node *rbnext = rb_next(&last->rb_node);
374 struct rb_node *rbprev = rb_prev(&last->rb_node);
375 struct request *next = NULL, *prev = NULL;
377 BUG_ON(RB_EMPTY_NODE(&last->rb_node));
380 prev = rb_entry_rq(rbprev);
383 next = rb_entry_rq(rbnext);
385 rbnext = rb_first(&cfqq->sort_list);
386 if (rbnext && rbnext != &last->rb_node)
387 next = rb_entry_rq(rbnext);
390 return cfq_choose_req(cfqd, next, prev);
393 static void cfq_resort_rr_list(struct cfq_queue *cfqq, int preempted)
395 struct cfq_data *cfqd = cfqq->cfqd;
396 struct list_head *list, *n;
397 struct cfq_queue *__cfqq;
400 * Resorting requires the cfqq to be on the RR list already.
402 if (!cfq_cfqq_on_rr(cfqq))
405 list_del(&cfqq->cfq_list);
407 if (cfq_class_rt(cfqq))
408 list = &cfqd->cur_rr;
409 else if (cfq_class_idle(cfqq))
410 list = &cfqd->idle_rr;
413 * if cfqq has requests in flight, don't allow it to be
414 * found in cfq_set_active_queue before it has finished them.
415 * this is done to increase fairness between a process that
416 * has lots of io pending vs one that only generates one
417 * sporadically or synchronously
419 if (cfq_cfqq_dispatched(cfqq))
420 list = &cfqd->busy_rr;
422 list = &cfqd->rr_list[cfqq->ioprio];
425 if (preempted || cfq_cfqq_queue_new(cfqq)) {
427 * If this queue was preempted or is new (never been serviced),
428 * let it be added first for fairness but beind other new
432 while (n->next != list) {
433 __cfqq = list_entry_cfqq(n->next);
434 if (!cfq_cfqq_queue_new(__cfqq))
439 list_add_tail(&cfqq->cfq_list, n);
440 } else if (!cfq_cfqq_class_sync(cfqq)) {
442 * async queue always goes to the end. this wont be overly
443 * unfair to writes, as the sort of the sync queue wont be
444 * allowed to pass the async queue again.
446 list_add_tail(&cfqq->cfq_list, list);
449 * sort by last service, but don't cross a new or async
450 * queue. we don't cross a new queue because it hasn't been
451 * service before, and we don't cross an async queue because
452 * it gets added to the end on expire.
455 while ((n = n->prev) != list) {
456 struct cfq_queue *__cfqq = list_entry_cfqq(n);
458 if (!cfq_cfqq_class_sync(cfqq) || !__cfqq->service_last)
460 if (time_before(__cfqq->service_last, cfqq->service_last))
463 list_add(&cfqq->cfq_list, n);
468 * add to busy list of queues for service, trying to be fair in ordering
469 * the pending list according to last request service
472 cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
474 BUG_ON(cfq_cfqq_on_rr(cfqq));
475 cfq_mark_cfqq_on_rr(cfqq);
478 cfq_resort_rr_list(cfqq, 0);
482 cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
484 BUG_ON(!cfq_cfqq_on_rr(cfqq));
485 cfq_clear_cfqq_on_rr(cfqq);
486 list_del_init(&cfqq->cfq_list);
488 BUG_ON(!cfqd->busy_queues);
493 * rb tree support functions
495 static inline void cfq_del_rq_rb(struct request *rq)
497 struct cfq_queue *cfqq = RQ_CFQQ(rq);
498 struct cfq_data *cfqd = cfqq->cfqd;
499 const int sync = rq_is_sync(rq);
501 BUG_ON(!cfqq->queued[sync]);
502 cfqq->queued[sync]--;
504 elv_rb_del(&cfqq->sort_list, rq);
506 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
507 cfq_del_cfqq_rr(cfqd, cfqq);
510 static void cfq_add_rq_rb(struct request *rq)
512 struct cfq_queue *cfqq = RQ_CFQQ(rq);
513 struct cfq_data *cfqd = cfqq->cfqd;
514 struct request *__alias;
516 cfqq->queued[rq_is_sync(rq)]++;
519 * looks a little odd, but the first insert might return an alias.
520 * if that happens, put the alias on the dispatch list
522 while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL)
523 cfq_dispatch_insert(cfqd->queue, __alias);
525 if (!cfq_cfqq_on_rr(cfqq))
526 cfq_add_cfqq_rr(cfqd, cfqq);
530 cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
532 elv_rb_del(&cfqq->sort_list, rq);
533 cfqq->queued[rq_is_sync(rq)]--;
537 static struct request *
538 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
540 struct task_struct *tsk = current;
541 pid_t key = cfq_queue_pid(tsk, bio_data_dir(bio), bio_sync(bio));
542 struct cfq_queue *cfqq;
544 cfqq = cfq_find_cfq_hash(cfqd, key, tsk->ioprio);
546 sector_t sector = bio->bi_sector + bio_sectors(bio);
548 return elv_rb_find(&cfqq->sort_list, sector);
554 static void cfq_activate_request(request_queue_t *q, struct request *rq)
556 struct cfq_data *cfqd = q->elevator->elevator_data;
558 cfqd->rq_in_driver++;
561 * If the depth is larger 1, it really could be queueing. But lets
562 * make the mark a little higher - idling could still be good for
563 * low queueing, and a low queueing number could also just indicate
564 * a SCSI mid layer like behaviour where limit+1 is often seen.
566 if (!cfqd->hw_tag && cfqd->rq_in_driver > 4)
570 static void cfq_deactivate_request(request_queue_t *q, struct request *rq)
572 struct cfq_data *cfqd = q->elevator->elevator_data;
574 WARN_ON(!cfqd->rq_in_driver);
575 cfqd->rq_in_driver--;
578 static void cfq_remove_request(struct request *rq)
580 struct cfq_queue *cfqq = RQ_CFQQ(rq);
582 if (cfqq->next_rq == rq)
583 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
585 list_del_init(&rq->queuelist);
588 if (rq_is_meta(rq)) {
589 WARN_ON(!cfqq->meta_pending);
590 cfqq->meta_pending--;
595 cfq_merge(request_queue_t *q, struct request **req, struct bio *bio)
597 struct cfq_data *cfqd = q->elevator->elevator_data;
598 struct request *__rq;
600 __rq = cfq_find_rq_fmerge(cfqd, bio);
601 if (__rq && elv_rq_merge_ok(__rq, bio)) {
603 return ELEVATOR_FRONT_MERGE;
606 return ELEVATOR_NO_MERGE;
609 static void cfq_merged_request(request_queue_t *q, struct request *req,
612 if (type == ELEVATOR_FRONT_MERGE) {
613 struct cfq_queue *cfqq = RQ_CFQQ(req);
615 cfq_reposition_rq_rb(cfqq, req);
620 cfq_merged_requests(request_queue_t *q, struct request *rq,
621 struct request *next)
624 * reposition in fifo if next is older than rq
626 if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
627 time_before(next->start_time, rq->start_time))
628 list_move(&rq->queuelist, &next->queuelist);
630 cfq_remove_request(next);
633 static int cfq_allow_merge(request_queue_t *q, struct request *rq,
636 struct cfq_data *cfqd = q->elevator->elevator_data;
637 const int rw = bio_data_dir(bio);
638 struct cfq_queue *cfqq;
642 * Disallow merge of a sync bio into an async request.
644 if ((bio_data_dir(bio) == READ || bio_sync(bio)) && !rq_is_sync(rq))
648 * Lookup the cfqq that this bio will be queued with. Allow
649 * merge only if rq is queued there.
651 key = cfq_queue_pid(current, rw, bio_sync(bio));
652 cfqq = cfq_find_cfq_hash(cfqd, key, current->ioprio);
654 if (cfqq == RQ_CFQQ(rq))
661 __cfq_set_active_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
665 * stop potential idle class queues waiting service
667 del_timer(&cfqd->idle_class_timer);
670 cfqq->slice_left = 0;
671 cfq_clear_cfqq_must_alloc_slice(cfqq);
672 cfq_clear_cfqq_fifo_expire(cfqq);
673 cfq_mark_cfqq_slice_new(cfqq);
676 cfqd->active_queue = cfqq;
680 * current cfqq expired its slice (or was too idle), select new one
683 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
686 unsigned long now = jiffies;
688 if (cfq_cfqq_wait_request(cfqq))
689 del_timer(&cfqd->idle_slice_timer);
691 if (!preempted && !cfq_cfqq_dispatched(cfqq))
692 cfq_schedule_dispatch(cfqd);
694 cfq_clear_cfqq_must_dispatch(cfqq);
695 cfq_clear_cfqq_wait_request(cfqq);
696 cfq_clear_cfqq_queue_new(cfqq);
699 * store what was left of this slice, if the queue idled out
702 if (cfq_slice_used(cfqq))
703 cfqq->slice_left = cfqq->slice_end - now;
705 cfqq->slice_left = 0;
707 cfq_resort_rr_list(cfqq, preempted);
709 if (cfqq == cfqd->active_queue)
710 cfqd->active_queue = NULL;
712 if (cfqd->active_cic) {
713 put_io_context(cfqd->active_cic->ioc);
714 cfqd->active_cic = NULL;
717 cfqd->dispatch_slice = 0;
720 static inline void cfq_slice_expired(struct cfq_data *cfqd, int preempted)
722 struct cfq_queue *cfqq = cfqd->active_queue;
725 __cfq_slice_expired(cfqd, cfqq, preempted);
738 static int cfq_get_next_prio_level(struct cfq_data *cfqd)
747 for (p = cfqd->cur_prio; p <= cfqd->cur_end_prio; p++) {
748 if (!list_empty(&cfqd->rr_list[p])) {
757 if (++cfqd->cur_end_prio == CFQ_PRIO_LISTS) {
758 cfqd->cur_end_prio = 0;
765 if (unlikely(prio == -1))
768 BUG_ON(prio >= CFQ_PRIO_LISTS);
770 list_splice_init(&cfqd->rr_list[prio], &cfqd->cur_rr);
772 cfqd->cur_prio = prio + 1;
773 if (cfqd->cur_prio > cfqd->cur_end_prio) {
774 cfqd->cur_end_prio = cfqd->cur_prio;
777 if (cfqd->cur_end_prio == CFQ_PRIO_LISTS) {
779 cfqd->cur_end_prio = 0;
785 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd)
787 struct cfq_queue *cfqq = NULL;
789 if (!list_empty(&cfqd->cur_rr) || cfq_get_next_prio_level(cfqd) != -1) {
791 * if current list is non-empty, grab first entry. if it is
792 * empty, get next prio level and grab first entry then if any
795 cfqq = list_entry_cfqq(cfqd->cur_rr.next);
796 } else if (!list_empty(&cfqd->busy_rr)) {
798 * If no new queues are available, check if the busy list has
799 * some before falling back to idle io.
801 cfqq = list_entry_cfqq(cfqd->busy_rr.next);
802 } else if (!list_empty(&cfqd->idle_rr)) {
804 * if we have idle queues and no rt or be queues had pending
805 * requests, either allow immediate service if the grace period
806 * has passed or arm the idle grace timer
808 unsigned long end = cfqd->last_end_request + CFQ_IDLE_GRACE;
810 if (time_after_eq(jiffies, end))
811 cfqq = list_entry_cfqq(cfqd->idle_rr.next);
813 mod_timer(&cfqd->idle_class_timer, end);
816 __cfq_set_active_queue(cfqd, cfqq);
820 #define CIC_SEEKY(cic) ((cic)->seek_mean > (128 * 1024))
822 static int cfq_arm_slice_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq)
825 struct cfq_io_context *cic;
828 WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
829 WARN_ON(cfqq != cfqd->active_queue);
832 * idle is disabled, either manually or by past process history
834 if (!cfqd->cfq_slice_idle)
836 if (!cfq_cfqq_idle_window(cfqq))
839 * task has exited, don't wait
841 cic = cfqd->active_cic;
842 if (!cic || !cic->ioc->task)
845 cfq_mark_cfqq_must_dispatch(cfqq);
846 cfq_mark_cfqq_wait_request(cfqq);
848 sl = min(cfqq->slice_end - 1, (unsigned long) cfqd->cfq_slice_idle);
851 * we don't want to idle for seeks, but we do want to allow
852 * fair distribution of slice time for a process doing back-to-back
853 * seeks. so allow a little bit of time for him to submit a new rq
855 if (sample_valid(cic->seek_samples) && CIC_SEEKY(cic))
856 sl = min(sl, msecs_to_jiffies(2));
858 mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
862 static void cfq_dispatch_insert(request_queue_t *q, struct request *rq)
864 struct cfq_data *cfqd = q->elevator->elevator_data;
865 struct cfq_queue *cfqq = RQ_CFQQ(rq);
867 cfq_remove_request(rq);
868 cfqq->on_dispatch[rq_is_sync(rq)]++;
869 elv_dispatch_sort(q, rq);
871 rq = list_entry(q->queue_head.prev, struct request, queuelist);
872 cfqd->last_sector = rq->sector + rq->nr_sectors;
876 * return expired entry, or NULL to just start from scratch in rbtree
878 static inline struct request *cfq_check_fifo(struct cfq_queue *cfqq)
880 struct cfq_data *cfqd = cfqq->cfqd;
884 if (cfq_cfqq_fifo_expire(cfqq))
886 if (list_empty(&cfqq->fifo))
889 fifo = cfq_cfqq_class_sync(cfqq);
890 rq = rq_entry_fifo(cfqq->fifo.next);
892 if (time_after(jiffies, rq->start_time + cfqd->cfq_fifo_expire[fifo])) {
893 cfq_mark_cfqq_fifo_expire(cfqq);
901 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
903 const int base_rq = cfqd->cfq_slice_async_rq;
905 WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
907 return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
911 * get next queue for service
913 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
915 struct cfq_queue *cfqq;
917 cfqq = cfqd->active_queue;
924 if (!cfq_cfqq_must_dispatch(cfqq) && cfq_slice_used(cfqq))
928 * if queue has requests, dispatch one. if not, check if
929 * enough slice is left to wait for one
931 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
933 else if (cfq_cfqq_slice_new(cfqq) || cfq_cfqq_dispatched(cfqq)) {
936 } else if (cfq_cfqq_class_sync(cfqq)) {
937 if (cfq_arm_slice_timer(cfqd, cfqq))
942 cfq_slice_expired(cfqd, 0);
944 cfqq = cfq_set_active_queue(cfqd);
950 __cfq_dispatch_requests(struct cfq_data *cfqd, struct cfq_queue *cfqq,
955 BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
961 * follow expired path, else get first next available
963 if ((rq = cfq_check_fifo(cfqq)) == NULL)
967 * finally, insert request into driver dispatch list
969 cfq_dispatch_insert(cfqd->queue, rq);
971 cfqd->dispatch_slice++;
974 if (!cfqd->active_cic) {
975 atomic_inc(&RQ_CIC(rq)->ioc->refcount);
976 cfqd->active_cic = RQ_CIC(rq);
979 if (RB_EMPTY_ROOT(&cfqq->sort_list))
982 } while (dispatched < max_dispatch);
985 * expire an async queue immediately if it has used up its slice. idle
986 * queue always expire after 1 dispatch round.
988 if ((!cfq_cfqq_sync(cfqq) &&
989 cfqd->dispatch_slice >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
990 cfq_class_idle(cfqq)) {
991 cfqq->slice_end = jiffies + 1;
992 cfq_slice_expired(cfqd, 0);
999 cfq_forced_dispatch_cfqqs(struct list_head *list)
1001 struct cfq_queue *cfqq, *next;
1005 list_for_each_entry_safe(cfqq, next, list, cfq_list) {
1006 while (cfqq->next_rq) {
1007 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
1010 BUG_ON(!list_empty(&cfqq->fifo));
1017 cfq_forced_dispatch(struct cfq_data *cfqd)
1019 int i, dispatched = 0;
1021 for (i = 0; i < CFQ_PRIO_LISTS; i++)
1022 dispatched += cfq_forced_dispatch_cfqqs(&cfqd->rr_list[i]);
1024 dispatched += cfq_forced_dispatch_cfqqs(&cfqd->busy_rr);
1025 dispatched += cfq_forced_dispatch_cfqqs(&cfqd->cur_rr);
1026 dispatched += cfq_forced_dispatch_cfqqs(&cfqd->idle_rr);
1028 cfq_slice_expired(cfqd, 0);
1030 BUG_ON(cfqd->busy_queues);
1036 cfq_dispatch_requests(request_queue_t *q, int force)
1038 struct cfq_data *cfqd = q->elevator->elevator_data;
1039 struct cfq_queue *cfqq, *prev_cfqq;
1042 if (!cfqd->busy_queues)
1045 if (unlikely(force))
1046 return cfq_forced_dispatch(cfqd);
1050 while ((cfqq = cfq_select_queue(cfqd)) != NULL) {
1054 * Don't repeat dispatch from the previous queue.
1056 if (prev_cfqq == cfqq)
1059 cfq_clear_cfqq_must_dispatch(cfqq);
1060 cfq_clear_cfqq_wait_request(cfqq);
1061 del_timer(&cfqd->idle_slice_timer);
1063 max_dispatch = cfqd->cfq_quantum;
1064 if (cfq_class_idle(cfqq))
1067 dispatched += __cfq_dispatch_requests(cfqd, cfqq, max_dispatch);
1070 * If the dispatch cfqq has idling enabled and is still
1071 * the active queue, break out.
1073 if (cfq_cfqq_idle_window(cfqq) && cfqd->active_queue)
1083 * task holds one reference to the queue, dropped when task exits. each rq
1084 * in-flight on this queue also holds a reference, dropped when rq is freed.
1086 * queue lock must be held here.
1088 static void cfq_put_queue(struct cfq_queue *cfqq)
1090 struct cfq_data *cfqd = cfqq->cfqd;
1092 BUG_ON(atomic_read(&cfqq->ref) <= 0);
1094 if (!atomic_dec_and_test(&cfqq->ref))
1097 BUG_ON(rb_first(&cfqq->sort_list));
1098 BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
1099 BUG_ON(cfq_cfqq_on_rr(cfqq));
1101 if (unlikely(cfqd->active_queue == cfqq))
1102 __cfq_slice_expired(cfqd, cfqq, 0);
1105 * it's on the empty list and still hashed
1107 list_del(&cfqq->cfq_list);
1108 hlist_del(&cfqq->cfq_hash);
1109 kmem_cache_free(cfq_pool, cfqq);
1112 static struct cfq_queue *
1113 __cfq_find_cfq_hash(struct cfq_data *cfqd, unsigned int key, unsigned int prio,
1116 struct hlist_head *hash_list = &cfqd->cfq_hash[hashval];
1117 struct hlist_node *entry;
1118 struct cfq_queue *__cfqq;
1120 hlist_for_each_entry(__cfqq, entry, hash_list, cfq_hash) {
1121 const unsigned short __p = IOPRIO_PRIO_VALUE(__cfqq->org_ioprio_class, __cfqq->org_ioprio);
1123 if (__cfqq->key == key && (__p == prio || !prio))
1130 static struct cfq_queue *
1131 cfq_find_cfq_hash(struct cfq_data *cfqd, unsigned int key, unsigned short prio)
1133 return __cfq_find_cfq_hash(cfqd, key, prio, hash_long(key, CFQ_QHASH_SHIFT));
1136 static void cfq_free_io_context(struct io_context *ioc)
1138 struct cfq_io_context *__cic;
1142 while ((n = rb_first(&ioc->cic_root)) != NULL) {
1143 __cic = rb_entry(n, struct cfq_io_context, rb_node);
1144 rb_erase(&__cic->rb_node, &ioc->cic_root);
1145 kmem_cache_free(cfq_ioc_pool, __cic);
1149 elv_ioc_count_mod(ioc_count, -freed);
1151 if (ioc_gone && !elv_ioc_count_read(ioc_count))
1155 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1157 if (unlikely(cfqq == cfqd->active_queue))
1158 __cfq_slice_expired(cfqd, cfqq, 0);
1160 cfq_put_queue(cfqq);
1163 static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
1164 struct cfq_io_context *cic)
1166 list_del_init(&cic->queue_list);
1170 if (cic->cfqq[ASYNC]) {
1171 cfq_exit_cfqq(cfqd, cic->cfqq[ASYNC]);
1172 cic->cfqq[ASYNC] = NULL;
1175 if (cic->cfqq[SYNC]) {
1176 cfq_exit_cfqq(cfqd, cic->cfqq[SYNC]);
1177 cic->cfqq[SYNC] = NULL;
1183 * Called with interrupts disabled
1185 static void cfq_exit_single_io_context(struct cfq_io_context *cic)
1187 struct cfq_data *cfqd = cic->key;
1190 request_queue_t *q = cfqd->queue;
1192 spin_lock_irq(q->queue_lock);
1193 __cfq_exit_single_io_context(cfqd, cic);
1194 spin_unlock_irq(q->queue_lock);
1198 static void cfq_exit_io_context(struct io_context *ioc)
1200 struct cfq_io_context *__cic;
1204 * put the reference this task is holding to the various queues
1207 n = rb_first(&ioc->cic_root);
1209 __cic = rb_entry(n, struct cfq_io_context, rb_node);
1211 cfq_exit_single_io_context(__cic);
1216 static struct cfq_io_context *
1217 cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
1219 struct cfq_io_context *cic;
1221 cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask, cfqd->queue->node);
1223 memset(cic, 0, sizeof(*cic));
1224 cic->last_end_request = jiffies;
1225 INIT_LIST_HEAD(&cic->queue_list);
1226 cic->dtor = cfq_free_io_context;
1227 cic->exit = cfq_exit_io_context;
1228 elv_ioc_count_inc(ioc_count);
1234 static void cfq_init_prio_data(struct cfq_queue *cfqq)
1236 struct task_struct *tsk = current;
1239 if (!cfq_cfqq_prio_changed(cfqq))
1242 ioprio_class = IOPRIO_PRIO_CLASS(tsk->ioprio);
1243 switch (ioprio_class) {
1245 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
1246 case IOPRIO_CLASS_NONE:
1248 * no prio set, place us in the middle of the BE classes
1250 cfqq->ioprio = task_nice_ioprio(tsk);
1251 cfqq->ioprio_class = IOPRIO_CLASS_BE;
1253 case IOPRIO_CLASS_RT:
1254 cfqq->ioprio = task_ioprio(tsk);
1255 cfqq->ioprio_class = IOPRIO_CLASS_RT;
1257 case IOPRIO_CLASS_BE:
1258 cfqq->ioprio = task_ioprio(tsk);
1259 cfqq->ioprio_class = IOPRIO_CLASS_BE;
1261 case IOPRIO_CLASS_IDLE:
1262 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
1264 cfq_clear_cfqq_idle_window(cfqq);
1269 * keep track of original prio settings in case we have to temporarily
1270 * elevate the priority of this queue
1272 cfqq->org_ioprio = cfqq->ioprio;
1273 cfqq->org_ioprio_class = cfqq->ioprio_class;
1275 cfq_resort_rr_list(cfqq, 0);
1276 cfq_clear_cfqq_prio_changed(cfqq);
1279 static inline void changed_ioprio(struct cfq_io_context *cic)
1281 struct cfq_data *cfqd = cic->key;
1282 struct cfq_queue *cfqq;
1283 unsigned long flags;
1285 if (unlikely(!cfqd))
1288 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1290 cfqq = cic->cfqq[ASYNC];
1292 struct cfq_queue *new_cfqq;
1293 new_cfqq = cfq_get_queue(cfqd, CFQ_KEY_ASYNC, cic->ioc->task,
1296 cic->cfqq[ASYNC] = new_cfqq;
1297 cfq_put_queue(cfqq);
1301 cfqq = cic->cfqq[SYNC];
1303 cfq_mark_cfqq_prio_changed(cfqq);
1305 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1308 static void cfq_ioc_set_ioprio(struct io_context *ioc)
1310 struct cfq_io_context *cic;
1313 ioc->ioprio_changed = 0;
1315 n = rb_first(&ioc->cic_root);
1317 cic = rb_entry(n, struct cfq_io_context, rb_node);
1319 changed_ioprio(cic);
1324 static struct cfq_queue *
1325 cfq_get_queue(struct cfq_data *cfqd, unsigned int key, struct task_struct *tsk,
1328 const int hashval = hash_long(key, CFQ_QHASH_SHIFT);
1329 struct cfq_queue *cfqq, *new_cfqq = NULL;
1330 unsigned short ioprio;
1333 ioprio = tsk->ioprio;
1334 cfqq = __cfq_find_cfq_hash(cfqd, key, ioprio, hashval);
1340 } else if (gfp_mask & __GFP_WAIT) {
1342 * Inform the allocator of the fact that we will
1343 * just repeat this allocation if it fails, to allow
1344 * the allocator to do whatever it needs to attempt to
1347 spin_unlock_irq(cfqd->queue->queue_lock);
1348 new_cfqq = kmem_cache_alloc_node(cfq_pool, gfp_mask|__GFP_NOFAIL, cfqd->queue->node);
1349 spin_lock_irq(cfqd->queue->queue_lock);
1352 cfqq = kmem_cache_alloc_node(cfq_pool, gfp_mask, cfqd->queue->node);
1357 memset(cfqq, 0, sizeof(*cfqq));
1359 INIT_HLIST_NODE(&cfqq->cfq_hash);
1360 INIT_LIST_HEAD(&cfqq->cfq_list);
1361 INIT_LIST_HEAD(&cfqq->fifo);
1364 hlist_add_head(&cfqq->cfq_hash, &cfqd->cfq_hash[hashval]);
1365 atomic_set(&cfqq->ref, 0);
1368 * set ->slice_left to allow preemption for a new process
1370 cfqq->slice_left = 2 * cfqd->cfq_slice_idle;
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 it doesn't have slice left, forget it
1591 if (new_cfqq->slice_left < cfqd->cfq_slice_idle)
1594 * if the new request is sync, but the currently running queue is
1595 * not, let the sync request have priority.
1597 if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
1600 * So both queues are sync. Let the new request get disk time if
1601 * it's a metadata request and the current queue is doing regular IO.
1603 if (rq_is_meta(rq) && !cfqq->meta_pending)
1610 * cfqq preempts the active queue. if we allowed preempt with no slice left,
1611 * let it have half of its nominal slice.
1613 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1615 cfq_slice_expired(cfqd, 1);
1617 if (!cfqq->slice_left)
1618 cfqq->slice_left = cfq_prio_to_slice(cfqd, cfqq) / 2;
1621 * Put the new queue at the front of the of the current list,
1622 * so we know that it will be selected next.
1624 BUG_ON(!cfq_cfqq_on_rr(cfqq));
1625 list_move(&cfqq->cfq_list, &cfqd->cur_rr);
1627 cfqq->slice_end = 0;
1628 cfq_mark_cfqq_slice_new(cfqq);
1632 * Called when a new fs request (rq) is added (to cfqq). Check if there's
1633 * something we should do about it
1636 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1639 struct cfq_io_context *cic = RQ_CIC(rq);
1642 cfqq->meta_pending++;
1645 * check if this request is a better next-serve candidate)) {
1647 cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq);
1648 BUG_ON(!cfqq->next_rq);
1651 * we never wait for an async request and we don't allow preemption
1652 * of an async request. so just return early
1654 if (!rq_is_sync(rq)) {
1656 * sync process issued an async request, if it's waiting
1657 * then expire it and kick rq handling.
1659 if (cic == cfqd->active_cic &&
1660 del_timer(&cfqd->idle_slice_timer)) {
1661 cfq_slice_expired(cfqd, 0);
1662 blk_start_queueing(cfqd->queue);
1667 cfq_update_io_thinktime(cfqd, cic);
1668 cfq_update_io_seektime(cic, rq);
1669 cfq_update_idle_window(cfqd, cfqq, cic);
1671 cic->last_request_pos = rq->sector + rq->nr_sectors;
1673 if (cfqq == cfqd->active_queue) {
1675 * if we are waiting for a request for this queue, let it rip
1676 * immediately and flag that we must not expire this queue
1679 if (cfq_cfqq_wait_request(cfqq)) {
1680 cfq_mark_cfqq_must_dispatch(cfqq);
1681 del_timer(&cfqd->idle_slice_timer);
1682 blk_start_queueing(cfqd->queue);
1684 } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
1686 * not the active queue - expire current slice if it is
1687 * idle and has expired it's mean thinktime or this new queue
1688 * has some old slice time left and is of higher priority
1690 cfq_preempt_queue(cfqd, cfqq);
1691 cfq_mark_cfqq_must_dispatch(cfqq);
1692 blk_start_queueing(cfqd->queue);
1696 static void cfq_insert_request(request_queue_t *q, struct request *rq)
1698 struct cfq_data *cfqd = q->elevator->elevator_data;
1699 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1701 cfq_init_prio_data(cfqq);
1705 list_add_tail(&rq->queuelist, &cfqq->fifo);
1707 cfq_rq_enqueued(cfqd, cfqq, rq);
1710 static void cfq_completed_request(request_queue_t *q, struct request *rq)
1712 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1713 struct cfq_data *cfqd = cfqq->cfqd;
1714 const int sync = rq_is_sync(rq);
1719 WARN_ON(!cfqd->rq_in_driver);
1720 WARN_ON(!cfqq->on_dispatch[sync]);
1721 cfqd->rq_in_driver--;
1722 cfqq->on_dispatch[sync]--;
1723 cfqq->service_last = now;
1725 if (!cfq_class_idle(cfqq))
1726 cfqd->last_end_request = now;
1728 cfq_resort_rr_list(cfqq, 0);
1731 RQ_CIC(rq)->last_end_request = now;
1734 * If this is the active queue, check if it needs to be expired,
1735 * or if we want to idle in case it has no pending requests.
1737 if (cfqd->active_queue == cfqq) {
1738 if (cfq_cfqq_slice_new(cfqq)) {
1739 cfq_set_prio_slice(cfqd, cfqq);
1740 cfq_clear_cfqq_slice_new(cfqq);
1742 if (cfq_slice_used(cfqq))
1743 cfq_slice_expired(cfqd, 0);
1744 else if (sync && RB_EMPTY_ROOT(&cfqq->sort_list)) {
1745 if (!cfq_arm_slice_timer(cfqd, cfqq))
1746 cfq_schedule_dispatch(cfqd);
1752 * we temporarily boost lower priority queues if they are holding fs exclusive
1753 * resources. they are boosted to normal prio (CLASS_BE/4)
1755 static void cfq_prio_boost(struct cfq_queue *cfqq)
1757 const int ioprio_class = cfqq->ioprio_class;
1758 const int ioprio = cfqq->ioprio;
1760 if (has_fs_excl()) {
1762 * boost idle prio on transactions that would lock out other
1763 * users of the filesystem
1765 if (cfq_class_idle(cfqq))
1766 cfqq->ioprio_class = IOPRIO_CLASS_BE;
1767 if (cfqq->ioprio > IOPRIO_NORM)
1768 cfqq->ioprio = IOPRIO_NORM;
1771 * check if we need to unboost the queue
1773 if (cfqq->ioprio_class != cfqq->org_ioprio_class)
1774 cfqq->ioprio_class = cfqq->org_ioprio_class;
1775 if (cfqq->ioprio != cfqq->org_ioprio)
1776 cfqq->ioprio = cfqq->org_ioprio;
1780 * refile between round-robin lists if we moved the priority class
1782 if ((ioprio_class != cfqq->ioprio_class || ioprio != cfqq->ioprio))
1783 cfq_resort_rr_list(cfqq, 0);
1786 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
1788 if ((cfq_cfqq_wait_request(cfqq) || cfq_cfqq_must_alloc(cfqq)) &&
1789 !cfq_cfqq_must_alloc_slice(cfqq)) {
1790 cfq_mark_cfqq_must_alloc_slice(cfqq);
1791 return ELV_MQUEUE_MUST;
1794 return ELV_MQUEUE_MAY;
1797 static int cfq_may_queue(request_queue_t *q, int rw)
1799 struct cfq_data *cfqd = q->elevator->elevator_data;
1800 struct task_struct *tsk = current;
1801 struct cfq_queue *cfqq;
1804 key = cfq_queue_pid(tsk, rw, rw & REQ_RW_SYNC);
1807 * don't force setup of a queue from here, as a call to may_queue
1808 * does not necessarily imply that a request actually will be queued.
1809 * so just lookup a possibly existing queue, or return 'may queue'
1812 cfqq = cfq_find_cfq_hash(cfqd, key, tsk->ioprio);
1814 cfq_init_prio_data(cfqq);
1815 cfq_prio_boost(cfqq);
1817 return __cfq_may_queue(cfqq);
1820 return ELV_MQUEUE_MAY;
1824 * queue lock held here
1826 static void cfq_put_request(struct request *rq)
1828 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1831 const int rw = rq_data_dir(rq);
1833 BUG_ON(!cfqq->allocated[rw]);
1834 cfqq->allocated[rw]--;
1836 put_io_context(RQ_CIC(rq)->ioc);
1838 rq->elevator_private = NULL;
1839 rq->elevator_private2 = NULL;
1841 cfq_put_queue(cfqq);
1846 * Allocate cfq data structures associated with this request.
1849 cfq_set_request(request_queue_t *q, struct request *rq, gfp_t gfp_mask)
1851 struct cfq_data *cfqd = q->elevator->elevator_data;
1852 struct task_struct *tsk = current;
1853 struct cfq_io_context *cic;
1854 const int rw = rq_data_dir(rq);
1855 const int is_sync = rq_is_sync(rq);
1856 pid_t key = cfq_queue_pid(tsk, rw, is_sync);
1857 struct cfq_queue *cfqq;
1858 unsigned long flags;
1860 might_sleep_if(gfp_mask & __GFP_WAIT);
1862 cic = cfq_get_io_context(cfqd, gfp_mask);
1864 spin_lock_irqsave(q->queue_lock, flags);
1869 if (!cic->cfqq[is_sync]) {
1870 cfqq = cfq_get_queue(cfqd, key, tsk, gfp_mask);
1874 cic->cfqq[is_sync] = cfqq;
1876 cfqq = cic->cfqq[is_sync];
1878 cfqq->allocated[rw]++;
1879 cfq_clear_cfqq_must_alloc(cfqq);
1880 atomic_inc(&cfqq->ref);
1882 spin_unlock_irqrestore(q->queue_lock, flags);
1884 rq->elevator_private = cic;
1885 rq->elevator_private2 = cfqq;
1890 put_io_context(cic->ioc);
1892 cfq_schedule_dispatch(cfqd);
1893 spin_unlock_irqrestore(q->queue_lock, flags);
1897 static void cfq_kick_queue(struct work_struct *work)
1899 struct cfq_data *cfqd =
1900 container_of(work, struct cfq_data, unplug_work);
1901 request_queue_t *q = cfqd->queue;
1902 unsigned long flags;
1904 spin_lock_irqsave(q->queue_lock, flags);
1905 blk_start_queueing(q);
1906 spin_unlock_irqrestore(q->queue_lock, flags);
1910 * Timer running if the active_queue is currently idling inside its time slice
1912 static void cfq_idle_slice_timer(unsigned long data)
1914 struct cfq_data *cfqd = (struct cfq_data *) data;
1915 struct cfq_queue *cfqq;
1916 unsigned long flags;
1918 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1920 if ((cfqq = cfqd->active_queue) != NULL) {
1924 if (cfq_slice_used(cfqq))
1928 * only expire and reinvoke request handler, if there are
1929 * other queues with pending requests
1931 if (!cfqd->busy_queues)
1935 * not expired and it has a request pending, let it dispatch
1937 if (!RB_EMPTY_ROOT(&cfqq->sort_list)) {
1938 cfq_mark_cfqq_must_dispatch(cfqq);
1943 cfq_slice_expired(cfqd, 0);
1945 cfq_schedule_dispatch(cfqd);
1947 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1951 * Timer running if an idle class queue is waiting for service
1953 static void cfq_idle_class_timer(unsigned long data)
1955 struct cfq_data *cfqd = (struct cfq_data *) data;
1956 unsigned long flags, end;
1958 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1961 * race with a non-idle queue, reset timer
1963 end = cfqd->last_end_request + CFQ_IDLE_GRACE;
1964 if (!time_after_eq(jiffies, end))
1965 mod_timer(&cfqd->idle_class_timer, end);
1967 cfq_schedule_dispatch(cfqd);
1969 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1972 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
1974 del_timer_sync(&cfqd->idle_slice_timer);
1975 del_timer_sync(&cfqd->idle_class_timer);
1976 blk_sync_queue(cfqd->queue);
1979 static void cfq_exit_queue(elevator_t *e)
1981 struct cfq_data *cfqd = e->elevator_data;
1982 request_queue_t *q = cfqd->queue;
1984 cfq_shutdown_timer_wq(cfqd);
1986 spin_lock_irq(q->queue_lock);
1988 if (cfqd->active_queue)
1989 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
1991 while (!list_empty(&cfqd->cic_list)) {
1992 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
1993 struct cfq_io_context,
1996 __cfq_exit_single_io_context(cfqd, cic);
1999 spin_unlock_irq(q->queue_lock);
2001 cfq_shutdown_timer_wq(cfqd);
2003 kfree(cfqd->cfq_hash);
2007 static void *cfq_init_queue(request_queue_t *q)
2009 struct cfq_data *cfqd;
2012 cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL, q->node);
2016 memset(cfqd, 0, sizeof(*cfqd));
2018 for (i = 0; i < CFQ_PRIO_LISTS; i++)
2019 INIT_LIST_HEAD(&cfqd->rr_list[i]);
2021 INIT_LIST_HEAD(&cfqd->busy_rr);
2022 INIT_LIST_HEAD(&cfqd->cur_rr);
2023 INIT_LIST_HEAD(&cfqd->idle_rr);
2024 INIT_LIST_HEAD(&cfqd->cic_list);
2026 cfqd->cfq_hash = kmalloc_node(sizeof(struct hlist_head) * CFQ_QHASH_ENTRIES, GFP_KERNEL, q->node);
2027 if (!cfqd->cfq_hash)
2030 for (i = 0; i < CFQ_QHASH_ENTRIES; i++)
2031 INIT_HLIST_HEAD(&cfqd->cfq_hash[i]);
2035 init_timer(&cfqd->idle_slice_timer);
2036 cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
2037 cfqd->idle_slice_timer.data = (unsigned long) cfqd;
2039 init_timer(&cfqd->idle_class_timer);
2040 cfqd->idle_class_timer.function = cfq_idle_class_timer;
2041 cfqd->idle_class_timer.data = (unsigned long) cfqd;
2043 INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
2045 cfqd->cfq_quantum = cfq_quantum;
2046 cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
2047 cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
2048 cfqd->cfq_back_max = cfq_back_max;
2049 cfqd->cfq_back_penalty = cfq_back_penalty;
2050 cfqd->cfq_slice[0] = cfq_slice_async;
2051 cfqd->cfq_slice[1] = cfq_slice_sync;
2052 cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
2053 cfqd->cfq_slice_idle = cfq_slice_idle;
2061 static void cfq_slab_kill(void)
2064 kmem_cache_destroy(cfq_pool);
2066 kmem_cache_destroy(cfq_ioc_pool);
2069 static int __init cfq_slab_setup(void)
2071 cfq_pool = kmem_cache_create("cfq_pool", sizeof(struct cfq_queue), 0, 0,
2076 cfq_ioc_pool = kmem_cache_create("cfq_ioc_pool",
2077 sizeof(struct cfq_io_context), 0, 0, NULL, NULL);
2088 * sysfs parts below -->
2092 cfq_var_show(unsigned int var, char *page)
2094 return sprintf(page, "%d\n", var);
2098 cfq_var_store(unsigned int *var, const char *page, size_t count)
2100 char *p = (char *) page;
2102 *var = simple_strtoul(p, &p, 10);
2106 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
2107 static ssize_t __FUNC(elevator_t *e, char *page) \
2109 struct cfq_data *cfqd = e->elevator_data; \
2110 unsigned int __data = __VAR; \
2112 __data = jiffies_to_msecs(__data); \
2113 return cfq_var_show(__data, (page)); \
2115 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
2116 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
2117 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
2118 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
2119 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
2120 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
2121 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
2122 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
2123 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
2124 #undef SHOW_FUNCTION
2126 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
2127 static ssize_t __FUNC(elevator_t *e, const char *page, size_t count) \
2129 struct cfq_data *cfqd = e->elevator_data; \
2130 unsigned int __data; \
2131 int ret = cfq_var_store(&__data, (page), count); \
2132 if (__data < (MIN)) \
2134 else if (__data > (MAX)) \
2137 *(__PTR) = msecs_to_jiffies(__data); \
2139 *(__PTR) = __data; \
2142 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
2143 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1, UINT_MAX, 1);
2144 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1, UINT_MAX, 1);
2145 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
2146 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1, UINT_MAX, 0);
2147 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
2148 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
2149 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
2150 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1, UINT_MAX, 0);
2151 #undef STORE_FUNCTION
2153 #define CFQ_ATTR(name) \
2154 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
2156 static struct elv_fs_entry cfq_attrs[] = {
2158 CFQ_ATTR(fifo_expire_sync),
2159 CFQ_ATTR(fifo_expire_async),
2160 CFQ_ATTR(back_seek_max),
2161 CFQ_ATTR(back_seek_penalty),
2162 CFQ_ATTR(slice_sync),
2163 CFQ_ATTR(slice_async),
2164 CFQ_ATTR(slice_async_rq),
2165 CFQ_ATTR(slice_idle),
2169 static struct elevator_type iosched_cfq = {
2171 .elevator_merge_fn = cfq_merge,
2172 .elevator_merged_fn = cfq_merged_request,
2173 .elevator_merge_req_fn = cfq_merged_requests,
2174 .elevator_allow_merge_fn = cfq_allow_merge,
2175 .elevator_dispatch_fn = cfq_dispatch_requests,
2176 .elevator_add_req_fn = cfq_insert_request,
2177 .elevator_activate_req_fn = cfq_activate_request,
2178 .elevator_deactivate_req_fn = cfq_deactivate_request,
2179 .elevator_queue_empty_fn = cfq_queue_empty,
2180 .elevator_completed_req_fn = cfq_completed_request,
2181 .elevator_former_req_fn = elv_rb_former_request,
2182 .elevator_latter_req_fn = elv_rb_latter_request,
2183 .elevator_set_req_fn = cfq_set_request,
2184 .elevator_put_req_fn = cfq_put_request,
2185 .elevator_may_queue_fn = cfq_may_queue,
2186 .elevator_init_fn = cfq_init_queue,
2187 .elevator_exit_fn = cfq_exit_queue,
2188 .trim = cfq_free_io_context,
2190 .elevator_attrs = cfq_attrs,
2191 .elevator_name = "cfq",
2192 .elevator_owner = THIS_MODULE,
2195 static int __init cfq_init(void)
2200 * could be 0 on HZ < 1000 setups
2202 if (!cfq_slice_async)
2203 cfq_slice_async = 1;
2204 if (!cfq_slice_idle)
2207 if (cfq_slab_setup())
2210 ret = elv_register(&iosched_cfq);
2217 static void __exit cfq_exit(void)
2219 DECLARE_COMPLETION_ONSTACK(all_gone);
2220 elv_unregister(&iosched_cfq);
2221 ioc_gone = &all_gone;
2222 /* ioc_gone's update must be visible before reading ioc_count */
2224 if (elv_ioc_count_read(ioc_count))
2225 wait_for_completion(ioc_gone);
2230 module_init(cfq_init);
2231 module_exit(cfq_exit);
2233 MODULE_AUTHOR("Jens Axboe");
2234 MODULE_LICENSE("GPL");
2235 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");