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/rbtree.h>
13 #include <linux/ioprio.h>
14 #include <linux/blktrace_api.h>
19 /* max queue in one round of service */
20 static const int cfq_quantum = 4;
21 static const int cfq_fifo_expire[2] = { HZ / 4, HZ / 8 };
22 /* maximum backwards seek, in KiB */
23 static const int cfq_back_max = 16 * 1024;
24 /* penalty of a backwards seek */
25 static const int cfq_back_penalty = 2;
26 static const int cfq_slice_sync = HZ / 10;
27 static int cfq_slice_async = HZ / 25;
28 static const int cfq_slice_async_rq = 2;
29 static int cfq_slice_idle = HZ / 125;
32 * offset from end of service tree
34 #define CFQ_IDLE_DELAY (HZ / 5)
37 * below this threshold, we consider thinktime immediate
39 #define CFQ_MIN_TT (2)
41 #define CFQ_SLICE_SCALE (5)
42 #define CFQ_HW_QUEUE_MIN (5)
45 ((struct cfq_io_context *) (rq)->elevator_private)
46 #define RQ_CFQQ(rq) (struct cfq_queue *) ((rq)->elevator_private2)
48 static struct kmem_cache *cfq_pool;
49 static struct kmem_cache *cfq_ioc_pool;
51 static DEFINE_PER_CPU(unsigned long, ioc_count);
52 static struct completion *ioc_gone;
53 static DEFINE_SPINLOCK(ioc_gone_lock);
55 #define CFQ_PRIO_LISTS IOPRIO_BE_NR
56 #define cfq_class_idle(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
57 #define cfq_class_rt(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
62 #define sample_valid(samples) ((samples) > 80)
65 * Most of our rbtree usage is for sorting with min extraction, so
66 * if we cache the leftmost node we don't have to walk down the tree
67 * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should
68 * move this into the elevator for the rq sorting as well.
74 #define CFQ_RB_ROOT (struct cfq_rb_root) { RB_ROOT, NULL, }
77 * Per block device queue structure
80 struct request_queue *queue;
83 * rr list of queues with requests and the count of them
85 struct cfq_rb_root service_tree;
86 unsigned int busy_queues;
92 * queue-depth detection
97 int rq_in_driver_peak;
100 * idle window management
102 struct timer_list idle_slice_timer;
103 struct work_struct unplug_work;
105 struct cfq_queue *active_queue;
106 struct cfq_io_context *active_cic;
109 * async queue for each priority case
111 struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
112 struct cfq_queue *async_idle_cfqq;
114 sector_t last_position;
115 unsigned long last_end_request;
118 * tunables, see top of file
120 unsigned int cfq_quantum;
121 unsigned int cfq_fifo_expire[2];
122 unsigned int cfq_back_penalty;
123 unsigned int cfq_back_max;
124 unsigned int cfq_slice[2];
125 unsigned int cfq_slice_async_rq;
126 unsigned int cfq_slice_idle;
128 struct list_head cic_list;
132 * Per process-grouping structure
135 /* reference count */
137 /* various state flags, see below */
139 /* parent cfq_data */
140 struct cfq_data *cfqd;
141 /* service_tree member */
142 struct rb_node rb_node;
143 /* service_tree key */
144 unsigned long rb_key;
145 /* sorted list of pending requests */
146 struct rb_root sort_list;
147 /* if fifo isn't expired, next request to serve */
148 struct request *next_rq;
149 /* requests queued in sort_list */
151 /* currently allocated requests */
153 /* fifo list of requests in sort_list */
154 struct list_head fifo;
156 unsigned long slice_end;
159 /* pending metadata requests */
161 /* number of requests that are on the dispatch list or inside driver */
164 /* io prio of this group */
165 unsigned short ioprio, org_ioprio;
166 unsigned short ioprio_class, org_ioprio_class;
171 enum cfqq_state_flags {
172 CFQ_CFQQ_FLAG_on_rr = 0, /* on round-robin busy list */
173 CFQ_CFQQ_FLAG_wait_request, /* waiting for a request */
174 CFQ_CFQQ_FLAG_must_alloc, /* must be allowed rq alloc */
175 CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
176 CFQ_CFQQ_FLAG_must_dispatch, /* must dispatch, even if expired */
177 CFQ_CFQQ_FLAG_fifo_expire, /* FIFO checked in this slice */
178 CFQ_CFQQ_FLAG_idle_window, /* slice idling enabled */
179 CFQ_CFQQ_FLAG_prio_changed, /* task priority has changed */
180 CFQ_CFQQ_FLAG_queue_new, /* queue never been serviced */
181 CFQ_CFQQ_FLAG_slice_new, /* no requests dispatched in slice */
182 CFQ_CFQQ_FLAG_sync, /* synchronous queue */
185 #define CFQ_CFQQ_FNS(name) \
186 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \
188 (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name); \
190 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \
192 (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \
194 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \
196 return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \
200 CFQ_CFQQ_FNS(wait_request);
201 CFQ_CFQQ_FNS(must_alloc);
202 CFQ_CFQQ_FNS(must_alloc_slice);
203 CFQ_CFQQ_FNS(must_dispatch);
204 CFQ_CFQQ_FNS(fifo_expire);
205 CFQ_CFQQ_FNS(idle_window);
206 CFQ_CFQQ_FNS(prio_changed);
207 CFQ_CFQQ_FNS(queue_new);
208 CFQ_CFQQ_FNS(slice_new);
212 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
213 blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args)
214 #define cfq_log(cfqd, fmt, args...) \
215 blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
217 static void cfq_dispatch_insert(struct request_queue *, struct request *);
218 static struct cfq_queue *cfq_get_queue(struct cfq_data *, int,
219 struct io_context *, gfp_t);
220 static struct cfq_io_context *cfq_cic_lookup(struct cfq_data *,
221 struct io_context *);
223 static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_context *cic,
226 return cic->cfqq[!!is_sync];
229 static inline void cic_set_cfqq(struct cfq_io_context *cic,
230 struct cfq_queue *cfqq, int is_sync)
232 cic->cfqq[!!is_sync] = cfqq;
236 * We regard a request as SYNC, if it's either a read or has the SYNC bit
237 * set (in which case it could also be direct WRITE).
239 static inline int cfq_bio_sync(struct bio *bio)
241 if (bio_data_dir(bio) == READ || bio_sync(bio))
248 * scheduler run of queue, if there are requests pending and no one in the
249 * driver that will restart queueing
251 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
253 if (cfqd->busy_queues) {
254 cfq_log(cfqd, "schedule dispatch");
255 kblockd_schedule_work(cfqd->queue, &cfqd->unplug_work);
259 static int cfq_queue_empty(struct request_queue *q)
261 struct cfq_data *cfqd = q->elevator->elevator_data;
263 return !cfqd->busy_queues;
267 * Scale schedule slice based on io priority. Use the sync time slice only
268 * if a queue is marked sync and has sync io queued. A sync queue with async
269 * io only, should not get full sync slice length.
271 static inline int cfq_prio_slice(struct cfq_data *cfqd, int sync,
274 const int base_slice = cfqd->cfq_slice[sync];
276 WARN_ON(prio >= IOPRIO_BE_NR);
278 return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
282 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
284 return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
288 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
290 cfqq->slice_end = cfq_prio_to_slice(cfqd, cfqq) + jiffies;
291 cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
295 * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
296 * isn't valid until the first request from the dispatch is activated
297 * and the slice time set.
299 static inline int cfq_slice_used(struct cfq_queue *cfqq)
301 if (cfq_cfqq_slice_new(cfqq))
303 if (time_before(jiffies, cfqq->slice_end))
310 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
311 * We choose the request that is closest to the head right now. Distance
312 * behind the head is penalized and only allowed to a certain extent.
314 static struct request *
315 cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2)
317 sector_t last, s1, s2, d1 = 0, d2 = 0;
318 unsigned long back_max;
319 #define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */
320 #define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */
321 unsigned wrap = 0; /* bit mask: requests behind the disk head? */
323 if (rq1 == NULL || rq1 == rq2)
328 if (rq_is_sync(rq1) && !rq_is_sync(rq2))
330 else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
332 if (rq_is_meta(rq1) && !rq_is_meta(rq2))
334 else if (rq_is_meta(rq2) && !rq_is_meta(rq1))
340 last = cfqd->last_position;
343 * by definition, 1KiB is 2 sectors
345 back_max = cfqd->cfq_back_max * 2;
348 * Strict one way elevator _except_ in the case where we allow
349 * short backward seeks which are biased as twice the cost of a
350 * similar forward seek.
354 else if (s1 + back_max >= last)
355 d1 = (last - s1) * cfqd->cfq_back_penalty;
357 wrap |= CFQ_RQ1_WRAP;
361 else if (s2 + back_max >= last)
362 d2 = (last - s2) * cfqd->cfq_back_penalty;
364 wrap |= CFQ_RQ2_WRAP;
366 /* Found required data */
369 * By doing switch() on the bit mask "wrap" we avoid having to
370 * check two variables for all permutations: --> faster!
373 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
389 case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
392 * Since both rqs are wrapped,
393 * start with the one that's further behind head
394 * (--> only *one* back seek required),
395 * since back seek takes more time than forward.
405 * The below is leftmost cache rbtree addon
407 static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
410 root->left = rb_first(&root->rb);
413 return rb_entry(root->left, struct cfq_queue, rb_node);
418 static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
423 rb_erase(n, &root->rb);
428 * would be nice to take fifo expire time into account as well
430 static struct request *
431 cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
432 struct request *last)
434 struct rb_node *rbnext = rb_next(&last->rb_node);
435 struct rb_node *rbprev = rb_prev(&last->rb_node);
436 struct request *next = NULL, *prev = NULL;
438 BUG_ON(RB_EMPTY_NODE(&last->rb_node));
441 prev = rb_entry_rq(rbprev);
444 next = rb_entry_rq(rbnext);
446 rbnext = rb_first(&cfqq->sort_list);
447 if (rbnext && rbnext != &last->rb_node)
448 next = rb_entry_rq(rbnext);
451 return cfq_choose_req(cfqd, next, prev);
454 static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
455 struct cfq_queue *cfqq)
458 * just an approximation, should be ok.
460 return (cfqd->busy_queues - 1) * (cfq_prio_slice(cfqd, 1, 0) -
461 cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
465 * The cfqd->service_tree holds all pending cfq_queue's that have
466 * requests waiting to be processed. It is sorted in the order that
467 * we will service the queues.
469 static void cfq_service_tree_add(struct cfq_data *cfqd,
470 struct cfq_queue *cfqq, int add_front)
472 struct rb_node **p, *parent;
473 struct cfq_queue *__cfqq;
474 unsigned long rb_key;
477 if (cfq_class_idle(cfqq)) {
478 rb_key = CFQ_IDLE_DELAY;
479 parent = rb_last(&cfqd->service_tree.rb);
480 if (parent && parent != &cfqq->rb_node) {
481 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
482 rb_key += __cfqq->rb_key;
485 } else if (!add_front) {
486 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
487 rb_key += cfqq->slice_resid;
488 cfqq->slice_resid = 0;
492 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
494 * same position, nothing more to do
496 if (rb_key == cfqq->rb_key)
499 cfq_rb_erase(&cfqq->rb_node, &cfqd->service_tree);
504 p = &cfqd->service_tree.rb.rb_node;
509 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
512 * sort RT queues first, we always want to give
513 * preference to them. IDLE queues goes to the back.
514 * after that, sort on the next service time.
516 if (cfq_class_rt(cfqq) > cfq_class_rt(__cfqq))
518 else if (cfq_class_rt(cfqq) < cfq_class_rt(__cfqq))
520 else if (cfq_class_idle(cfqq) < cfq_class_idle(__cfqq))
522 else if (cfq_class_idle(cfqq) > cfq_class_idle(__cfqq))
524 else if (rb_key < __cfqq->rb_key)
529 if (n == &(*p)->rb_right)
536 cfqd->service_tree.left = &cfqq->rb_node;
538 cfqq->rb_key = rb_key;
539 rb_link_node(&cfqq->rb_node, parent, p);
540 rb_insert_color(&cfqq->rb_node, &cfqd->service_tree.rb);
544 * Update cfqq's position in the service tree.
546 static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
549 * Resorting requires the cfqq to be on the RR list already.
551 if (cfq_cfqq_on_rr(cfqq))
552 cfq_service_tree_add(cfqd, cfqq, 0);
556 * add to busy list of queues for service, trying to be fair in ordering
557 * the pending list according to last request service
559 static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
561 cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
562 BUG_ON(cfq_cfqq_on_rr(cfqq));
563 cfq_mark_cfqq_on_rr(cfqq);
566 cfq_resort_rr_list(cfqd, cfqq);
570 * Called when the cfqq no longer has requests pending, remove it from
573 static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
575 cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
576 BUG_ON(!cfq_cfqq_on_rr(cfqq));
577 cfq_clear_cfqq_on_rr(cfqq);
579 if (!RB_EMPTY_NODE(&cfqq->rb_node))
580 cfq_rb_erase(&cfqq->rb_node, &cfqd->service_tree);
582 BUG_ON(!cfqd->busy_queues);
587 * rb tree support functions
589 static void cfq_del_rq_rb(struct request *rq)
591 struct cfq_queue *cfqq = RQ_CFQQ(rq);
592 struct cfq_data *cfqd = cfqq->cfqd;
593 const int sync = rq_is_sync(rq);
595 BUG_ON(!cfqq->queued[sync]);
596 cfqq->queued[sync]--;
598 elv_rb_del(&cfqq->sort_list, rq);
600 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
601 cfq_del_cfqq_rr(cfqd, cfqq);
604 static void cfq_add_rq_rb(struct request *rq)
606 struct cfq_queue *cfqq = RQ_CFQQ(rq);
607 struct cfq_data *cfqd = cfqq->cfqd;
608 struct request *__alias;
610 cfqq->queued[rq_is_sync(rq)]++;
613 * looks a little odd, but the first insert might return an alias.
614 * if that happens, put the alias on the dispatch list
616 while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL)
617 cfq_dispatch_insert(cfqd->queue, __alias);
619 if (!cfq_cfqq_on_rr(cfqq))
620 cfq_add_cfqq_rr(cfqd, cfqq);
623 * check if this request is a better next-serve candidate
625 cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq);
626 BUG_ON(!cfqq->next_rq);
629 static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
631 elv_rb_del(&cfqq->sort_list, rq);
632 cfqq->queued[rq_is_sync(rq)]--;
636 static struct request *
637 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
639 struct task_struct *tsk = current;
640 struct cfq_io_context *cic;
641 struct cfq_queue *cfqq;
643 cic = cfq_cic_lookup(cfqd, tsk->io_context);
647 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
649 sector_t sector = bio->bi_sector + bio_sectors(bio);
651 return elv_rb_find(&cfqq->sort_list, sector);
657 static void cfq_activate_request(struct request_queue *q, struct request *rq)
659 struct cfq_data *cfqd = q->elevator->elevator_data;
661 cfqd->rq_in_driver++;
662 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
665 cfqd->last_position = rq->hard_sector + rq->hard_nr_sectors;
668 static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
670 struct cfq_data *cfqd = q->elevator->elevator_data;
672 WARN_ON(!cfqd->rq_in_driver);
673 cfqd->rq_in_driver--;
674 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
678 static void cfq_remove_request(struct request *rq)
680 struct cfq_queue *cfqq = RQ_CFQQ(rq);
682 if (cfqq->next_rq == rq)
683 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
685 list_del_init(&rq->queuelist);
688 cfqq->cfqd->rq_queued--;
689 if (rq_is_meta(rq)) {
690 WARN_ON(!cfqq->meta_pending);
691 cfqq->meta_pending--;
695 static int cfq_merge(struct request_queue *q, struct request **req,
698 struct cfq_data *cfqd = q->elevator->elevator_data;
699 struct request *__rq;
701 __rq = cfq_find_rq_fmerge(cfqd, bio);
702 if (__rq && elv_rq_merge_ok(__rq, bio)) {
704 return ELEVATOR_FRONT_MERGE;
707 return ELEVATOR_NO_MERGE;
710 static void cfq_merged_request(struct request_queue *q, struct request *req,
713 if (type == ELEVATOR_FRONT_MERGE) {
714 struct cfq_queue *cfqq = RQ_CFQQ(req);
716 cfq_reposition_rq_rb(cfqq, req);
721 cfq_merged_requests(struct request_queue *q, struct request *rq,
722 struct request *next)
725 * reposition in fifo if next is older than rq
727 if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
728 time_before(next->start_time, rq->start_time))
729 list_move(&rq->queuelist, &next->queuelist);
731 cfq_remove_request(next);
734 static int cfq_allow_merge(struct request_queue *q, struct request *rq,
737 struct cfq_data *cfqd = q->elevator->elevator_data;
738 struct cfq_io_context *cic;
739 struct cfq_queue *cfqq;
742 * Disallow merge of a sync bio into an async request.
744 if (cfq_bio_sync(bio) && !rq_is_sync(rq))
748 * Lookup the cfqq that this bio will be queued with. Allow
749 * merge only if rq is queued there.
751 cic = cfq_cic_lookup(cfqd, current->io_context);
755 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
756 if (cfqq == RQ_CFQQ(rq))
762 static void __cfq_set_active_queue(struct cfq_data *cfqd,
763 struct cfq_queue *cfqq)
766 cfq_log_cfqq(cfqd, cfqq, "set_active");
768 cfq_clear_cfqq_must_alloc_slice(cfqq);
769 cfq_clear_cfqq_fifo_expire(cfqq);
770 cfq_mark_cfqq_slice_new(cfqq);
771 cfq_clear_cfqq_queue_new(cfqq);
774 cfqd->active_queue = cfqq;
778 * current cfqq expired its slice (or was too idle), select new one
781 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
784 cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
786 if (cfq_cfqq_wait_request(cfqq))
787 del_timer(&cfqd->idle_slice_timer);
789 cfq_clear_cfqq_must_dispatch(cfqq);
790 cfq_clear_cfqq_wait_request(cfqq);
793 * store what was left of this slice, if the queue idled/timed out
795 if (timed_out && !cfq_cfqq_slice_new(cfqq)) {
796 cfqq->slice_resid = cfqq->slice_end - jiffies;
797 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
800 cfq_resort_rr_list(cfqd, cfqq);
802 if (cfqq == cfqd->active_queue)
803 cfqd->active_queue = NULL;
805 if (cfqd->active_cic) {
806 put_io_context(cfqd->active_cic->ioc);
807 cfqd->active_cic = NULL;
811 static inline void cfq_slice_expired(struct cfq_data *cfqd, int timed_out)
813 struct cfq_queue *cfqq = cfqd->active_queue;
816 __cfq_slice_expired(cfqd, cfqq, timed_out);
820 * Get next queue for service. Unless we have a queue preemption,
821 * we'll simply select the first cfqq in the service tree.
823 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
825 if (RB_EMPTY_ROOT(&cfqd->service_tree.rb))
828 return cfq_rb_first(&cfqd->service_tree);
832 * Get and set a new active queue for service.
834 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd)
836 struct cfq_queue *cfqq;
838 cfqq = cfq_get_next_queue(cfqd);
839 __cfq_set_active_queue(cfqd, cfqq);
843 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
846 if (rq->sector >= cfqd->last_position)
847 return rq->sector - cfqd->last_position;
849 return cfqd->last_position - rq->sector;
852 static inline int cfq_rq_close(struct cfq_data *cfqd, struct request *rq)
854 struct cfq_io_context *cic = cfqd->active_cic;
856 if (!sample_valid(cic->seek_samples))
859 return cfq_dist_from_last(cfqd, rq) <= cic->seek_mean;
862 static int cfq_close_cooperator(struct cfq_data *cfq_data,
863 struct cfq_queue *cfqq)
866 * We should notice if some of the queues are cooperating, eg
867 * working closely on the same area of the disk. In that case,
868 * we can group them together and don't waste time idling.
873 #define CIC_SEEKY(cic) ((cic)->seek_mean > (8 * 1024))
875 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
877 struct cfq_queue *cfqq = cfqd->active_queue;
878 struct cfq_io_context *cic;
881 WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
882 WARN_ON(cfq_cfqq_slice_new(cfqq));
885 * idle is disabled, either manually or by past process history
887 if (!cfqd->cfq_slice_idle || !cfq_cfqq_idle_window(cfqq))
891 * still requests with the driver, don't idle
893 if (cfqd->rq_in_driver)
897 * task has exited, don't wait
899 cic = cfqd->active_cic;
900 if (!cic || !atomic_read(&cic->ioc->nr_tasks))
904 * See if this prio level has a good candidate
906 if (cfq_close_cooperator(cfqd, cfqq) &&
907 (sample_valid(cic->ttime_samples) && cic->ttime_mean > 2))
910 cfq_mark_cfqq_must_dispatch(cfqq);
911 cfq_mark_cfqq_wait_request(cfqq);
914 * we don't want to idle for seeks, but we do want to allow
915 * fair distribution of slice time for a process doing back-to-back
916 * seeks. so allow a little bit of time for him to submit a new rq
918 sl = cfqd->cfq_slice_idle;
919 if (sample_valid(cic->seek_samples) && CIC_SEEKY(cic))
920 sl = min(sl, msecs_to_jiffies(CFQ_MIN_TT));
922 mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
923 cfq_log(cfqd, "arm_idle: %lu", sl);
927 * Move request from internal lists to the request queue dispatch list.
929 static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
931 struct cfq_data *cfqd = q->elevator->elevator_data;
932 struct cfq_queue *cfqq = RQ_CFQQ(rq);
934 cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
936 cfq_remove_request(rq);
938 elv_dispatch_sort(q, rq);
940 if (cfq_cfqq_sync(cfqq))
945 * return expired entry, or NULL to just start from scratch in rbtree
947 static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
949 struct cfq_data *cfqd = cfqq->cfqd;
953 if (cfq_cfqq_fifo_expire(cfqq))
956 cfq_mark_cfqq_fifo_expire(cfqq);
958 if (list_empty(&cfqq->fifo))
961 fifo = cfq_cfqq_sync(cfqq);
962 rq = rq_entry_fifo(cfqq->fifo.next);
964 if (time_before(jiffies, rq->start_time + cfqd->cfq_fifo_expire[fifo]))
967 cfq_log_cfqq(cfqd, cfqq, "fifo=%p", rq);
972 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
974 const int base_rq = cfqd->cfq_slice_async_rq;
976 WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
978 return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
982 * Select a queue for service. If we have a current active queue,
983 * check whether to continue servicing it, or retrieve and set a new one.
985 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
987 struct cfq_queue *cfqq;
989 cfqq = cfqd->active_queue;
994 * The active queue has run out of time, expire it and select new.
996 if (cfq_slice_used(cfqq))
1000 * The active queue has requests and isn't expired, allow it to
1003 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
1007 * No requests pending. If the active queue still has requests in
1008 * flight or is idling for a new request, allow either of these
1009 * conditions to happen (or time out) before selecting a new queue.
1011 if (timer_pending(&cfqd->idle_slice_timer) ||
1012 (cfqq->dispatched && cfq_cfqq_idle_window(cfqq))) {
1018 cfq_slice_expired(cfqd, 0);
1020 cfqq = cfq_set_active_queue(cfqd);
1026 * Dispatch some requests from cfqq, moving them to the request queue
1030 __cfq_dispatch_requests(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1035 BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
1041 * follow expired path, else get first next available
1043 rq = cfq_check_fifo(cfqq);
1048 * finally, insert request into driver dispatch list
1050 cfq_dispatch_insert(cfqd->queue, rq);
1054 if (!cfqd->active_cic) {
1055 atomic_inc(&RQ_CIC(rq)->ioc->refcount);
1056 cfqd->active_cic = RQ_CIC(rq);
1059 if (RB_EMPTY_ROOT(&cfqq->sort_list))
1062 } while (dispatched < max_dispatch);
1065 * expire an async queue immediately if it has used up its slice. idle
1066 * queue always expire after 1 dispatch round.
1068 if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
1069 dispatched >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
1070 cfq_class_idle(cfqq))) {
1071 cfqq->slice_end = jiffies + 1;
1072 cfq_slice_expired(cfqd, 0);
1078 static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
1082 while (cfqq->next_rq) {
1083 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
1087 BUG_ON(!list_empty(&cfqq->fifo));
1092 * Drain our current requests. Used for barriers and when switching
1093 * io schedulers on-the-fly.
1095 static int cfq_forced_dispatch(struct cfq_data *cfqd)
1097 struct cfq_queue *cfqq;
1100 while ((cfqq = cfq_rb_first(&cfqd->service_tree)) != NULL)
1101 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
1103 cfq_slice_expired(cfqd, 0);
1105 BUG_ON(cfqd->busy_queues);
1107 cfq_log(cfqd, "forced_dispatch=%d\n", dispatched);
1111 static int cfq_dispatch_requests(struct request_queue *q, int force)
1113 struct cfq_data *cfqd = q->elevator->elevator_data;
1114 struct cfq_queue *cfqq;
1117 if (!cfqd->busy_queues)
1120 if (unlikely(force))
1121 return cfq_forced_dispatch(cfqd);
1124 while ((cfqq = cfq_select_queue(cfqd)) != NULL) {
1127 max_dispatch = cfqd->cfq_quantum;
1128 if (cfq_class_idle(cfqq))
1131 if (cfqq->dispatched >= max_dispatch) {
1132 if (cfqd->busy_queues > 1)
1134 if (cfqq->dispatched >= 4 * max_dispatch)
1138 if (cfqd->sync_flight && !cfq_cfqq_sync(cfqq))
1141 cfq_clear_cfqq_must_dispatch(cfqq);
1142 cfq_clear_cfqq_wait_request(cfqq);
1143 del_timer(&cfqd->idle_slice_timer);
1145 dispatched += __cfq_dispatch_requests(cfqd, cfqq, max_dispatch);
1148 cfq_log(cfqd, "dispatched=%d", dispatched);
1153 * task holds one reference to the queue, dropped when task exits. each rq
1154 * in-flight on this queue also holds a reference, dropped when rq is freed.
1156 * queue lock must be held here.
1158 static void cfq_put_queue(struct cfq_queue *cfqq)
1160 struct cfq_data *cfqd = cfqq->cfqd;
1162 BUG_ON(atomic_read(&cfqq->ref) <= 0);
1164 if (!atomic_dec_and_test(&cfqq->ref))
1167 cfq_log_cfqq(cfqd, cfqq, "put_queue");
1168 BUG_ON(rb_first(&cfqq->sort_list));
1169 BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
1170 BUG_ON(cfq_cfqq_on_rr(cfqq));
1172 if (unlikely(cfqd->active_queue == cfqq)) {
1173 __cfq_slice_expired(cfqd, cfqq, 0);
1174 cfq_schedule_dispatch(cfqd);
1177 kmem_cache_free(cfq_pool, cfqq);
1181 * Must always be called with the rcu_read_lock() held
1184 __call_for_each_cic(struct io_context *ioc,
1185 void (*func)(struct io_context *, struct cfq_io_context *))
1187 struct cfq_io_context *cic;
1188 struct hlist_node *n;
1190 hlist_for_each_entry_rcu(cic, n, &ioc->cic_list, cic_list)
1195 * Call func for each cic attached to this ioc.
1198 call_for_each_cic(struct io_context *ioc,
1199 void (*func)(struct io_context *, struct cfq_io_context *))
1202 __call_for_each_cic(ioc, func);
1206 static void cfq_cic_free_rcu(struct rcu_head *head)
1208 struct cfq_io_context *cic;
1210 cic = container_of(head, struct cfq_io_context, rcu_head);
1212 kmem_cache_free(cfq_ioc_pool, cic);
1213 elv_ioc_count_dec(ioc_count);
1217 * CFQ scheduler is exiting, grab exit lock and check
1218 * the pending io context count. If it hits zero,
1219 * complete ioc_gone and set it back to NULL
1221 spin_lock(&ioc_gone_lock);
1222 if (ioc_gone && !elv_ioc_count_read(ioc_count)) {
1226 spin_unlock(&ioc_gone_lock);
1230 static void cfq_cic_free(struct cfq_io_context *cic)
1232 call_rcu(&cic->rcu_head, cfq_cic_free_rcu);
1235 static void cic_free_func(struct io_context *ioc, struct cfq_io_context *cic)
1237 unsigned long flags;
1239 BUG_ON(!cic->dead_key);
1241 spin_lock_irqsave(&ioc->lock, flags);
1242 radix_tree_delete(&ioc->radix_root, cic->dead_key);
1243 hlist_del_rcu(&cic->cic_list);
1244 spin_unlock_irqrestore(&ioc->lock, flags);
1250 * Must be called with rcu_read_lock() held or preemption otherwise disabled.
1251 * Only two callers of this - ->dtor() which is called with the rcu_read_lock(),
1252 * and ->trim() which is called with the task lock held
1254 static void cfq_free_io_context(struct io_context *ioc)
1257 * ioc->refcount is zero here, or we are called from elv_unregister(),
1258 * so no more cic's are allowed to be linked into this ioc. So it
1259 * should be ok to iterate over the known list, we will see all cic's
1260 * since no new ones are added.
1262 __call_for_each_cic(ioc, cic_free_func);
1265 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1267 if (unlikely(cfqq == cfqd->active_queue)) {
1268 __cfq_slice_expired(cfqd, cfqq, 0);
1269 cfq_schedule_dispatch(cfqd);
1272 cfq_put_queue(cfqq);
1275 static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
1276 struct cfq_io_context *cic)
1278 struct io_context *ioc = cic->ioc;
1280 list_del_init(&cic->queue_list);
1283 * Make sure key == NULL is seen for dead queues
1286 cic->dead_key = (unsigned long) cic->key;
1289 if (ioc->ioc_data == cic)
1290 rcu_assign_pointer(ioc->ioc_data, NULL);
1292 if (cic->cfqq[ASYNC]) {
1293 cfq_exit_cfqq(cfqd, cic->cfqq[ASYNC]);
1294 cic->cfqq[ASYNC] = NULL;
1297 if (cic->cfqq[SYNC]) {
1298 cfq_exit_cfqq(cfqd, cic->cfqq[SYNC]);
1299 cic->cfqq[SYNC] = NULL;
1303 static void cfq_exit_single_io_context(struct io_context *ioc,
1304 struct cfq_io_context *cic)
1306 struct cfq_data *cfqd = cic->key;
1309 struct request_queue *q = cfqd->queue;
1310 unsigned long flags;
1312 spin_lock_irqsave(q->queue_lock, flags);
1313 __cfq_exit_single_io_context(cfqd, cic);
1314 spin_unlock_irqrestore(q->queue_lock, flags);
1319 * The process that ioc belongs to has exited, we need to clean up
1320 * and put the internal structures we have that belongs to that process.
1322 static void cfq_exit_io_context(struct io_context *ioc)
1324 call_for_each_cic(ioc, cfq_exit_single_io_context);
1327 static struct cfq_io_context *
1328 cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
1330 struct cfq_io_context *cic;
1332 cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO,
1335 cic->last_end_request = jiffies;
1336 INIT_LIST_HEAD(&cic->queue_list);
1337 INIT_HLIST_NODE(&cic->cic_list);
1338 cic->dtor = cfq_free_io_context;
1339 cic->exit = cfq_exit_io_context;
1340 elv_ioc_count_inc(ioc_count);
1346 static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc)
1348 struct task_struct *tsk = current;
1351 if (!cfq_cfqq_prio_changed(cfqq))
1354 ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio);
1355 switch (ioprio_class) {
1357 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
1358 case IOPRIO_CLASS_NONE:
1360 * no prio set, inherit CPU scheduling settings
1362 cfqq->ioprio = task_nice_ioprio(tsk);
1363 cfqq->ioprio_class = task_nice_ioclass(tsk);
1365 case IOPRIO_CLASS_RT:
1366 cfqq->ioprio = task_ioprio(ioc);
1367 cfqq->ioprio_class = IOPRIO_CLASS_RT;
1369 case IOPRIO_CLASS_BE:
1370 cfqq->ioprio = task_ioprio(ioc);
1371 cfqq->ioprio_class = IOPRIO_CLASS_BE;
1373 case IOPRIO_CLASS_IDLE:
1374 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
1376 cfq_clear_cfqq_idle_window(cfqq);
1381 * keep track of original prio settings in case we have to temporarily
1382 * elevate the priority of this queue
1384 cfqq->org_ioprio = cfqq->ioprio;
1385 cfqq->org_ioprio_class = cfqq->ioprio_class;
1386 cfq_clear_cfqq_prio_changed(cfqq);
1389 static void changed_ioprio(struct io_context *ioc, struct cfq_io_context *cic)
1391 struct cfq_data *cfqd = cic->key;
1392 struct cfq_queue *cfqq;
1393 unsigned long flags;
1395 if (unlikely(!cfqd))
1398 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1400 cfqq = cic->cfqq[ASYNC];
1402 struct cfq_queue *new_cfqq;
1403 new_cfqq = cfq_get_queue(cfqd, ASYNC, cic->ioc, GFP_ATOMIC);
1405 cic->cfqq[ASYNC] = new_cfqq;
1406 cfq_put_queue(cfqq);
1410 cfqq = cic->cfqq[SYNC];
1412 cfq_mark_cfqq_prio_changed(cfqq);
1414 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1417 static void cfq_ioc_set_ioprio(struct io_context *ioc)
1419 call_for_each_cic(ioc, changed_ioprio);
1420 ioc->ioprio_changed = 0;
1423 static struct cfq_queue *
1424 cfq_find_alloc_queue(struct cfq_data *cfqd, int is_sync,
1425 struct io_context *ioc, gfp_t gfp_mask)
1427 struct cfq_queue *cfqq, *new_cfqq = NULL;
1428 struct cfq_io_context *cic;
1431 cic = cfq_cic_lookup(cfqd, ioc);
1432 /* cic always exists here */
1433 cfqq = cic_to_cfqq(cic, is_sync);
1439 } else if (gfp_mask & __GFP_WAIT) {
1441 * Inform the allocator of the fact that we will
1442 * just repeat this allocation if it fails, to allow
1443 * the allocator to do whatever it needs to attempt to
1446 spin_unlock_irq(cfqd->queue->queue_lock);
1447 new_cfqq = kmem_cache_alloc_node(cfq_pool,
1448 gfp_mask | __GFP_NOFAIL | __GFP_ZERO,
1450 spin_lock_irq(cfqd->queue->queue_lock);
1453 cfqq = kmem_cache_alloc_node(cfq_pool,
1454 gfp_mask | __GFP_ZERO,
1460 RB_CLEAR_NODE(&cfqq->rb_node);
1461 INIT_LIST_HEAD(&cfqq->fifo);
1463 atomic_set(&cfqq->ref, 0);
1466 cfq_mark_cfqq_prio_changed(cfqq);
1467 cfq_mark_cfqq_queue_new(cfqq);
1469 cfq_init_prio_data(cfqq, ioc);
1472 if (!cfq_class_idle(cfqq))
1473 cfq_mark_cfqq_idle_window(cfqq);
1474 cfq_mark_cfqq_sync(cfqq);
1476 cfqq->pid = current->pid;
1477 cfq_log_cfqq(cfqd, cfqq, "alloced");
1481 kmem_cache_free(cfq_pool, new_cfqq);
1484 WARN_ON((gfp_mask & __GFP_WAIT) && !cfqq);
1488 static struct cfq_queue **
1489 cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
1491 switch (ioprio_class) {
1492 case IOPRIO_CLASS_RT:
1493 return &cfqd->async_cfqq[0][ioprio];
1494 case IOPRIO_CLASS_BE:
1495 return &cfqd->async_cfqq[1][ioprio];
1496 case IOPRIO_CLASS_IDLE:
1497 return &cfqd->async_idle_cfqq;
1503 static struct cfq_queue *
1504 cfq_get_queue(struct cfq_data *cfqd, int is_sync, struct io_context *ioc,
1507 const int ioprio = task_ioprio(ioc);
1508 const int ioprio_class = task_ioprio_class(ioc);
1509 struct cfq_queue **async_cfqq = NULL;
1510 struct cfq_queue *cfqq = NULL;
1513 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
1518 cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask);
1524 * pin the queue now that it's allocated, scheduler exit will prune it
1526 if (!is_sync && !(*async_cfqq)) {
1527 atomic_inc(&cfqq->ref);
1531 atomic_inc(&cfqq->ref);
1536 * We drop cfq io contexts lazily, so we may find a dead one.
1539 cfq_drop_dead_cic(struct cfq_data *cfqd, struct io_context *ioc,
1540 struct cfq_io_context *cic)
1542 unsigned long flags;
1544 WARN_ON(!list_empty(&cic->queue_list));
1546 spin_lock_irqsave(&ioc->lock, flags);
1548 BUG_ON(ioc->ioc_data == cic);
1550 radix_tree_delete(&ioc->radix_root, (unsigned long) cfqd);
1551 hlist_del_rcu(&cic->cic_list);
1552 spin_unlock_irqrestore(&ioc->lock, flags);
1557 static struct cfq_io_context *
1558 cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc)
1560 struct cfq_io_context *cic;
1561 unsigned long flags;
1570 * we maintain a last-hit cache, to avoid browsing over the tree
1572 cic = rcu_dereference(ioc->ioc_data);
1573 if (cic && cic->key == cfqd) {
1579 cic = radix_tree_lookup(&ioc->radix_root, (unsigned long) cfqd);
1583 /* ->key must be copied to avoid race with cfq_exit_queue() */
1586 cfq_drop_dead_cic(cfqd, ioc, cic);
1591 spin_lock_irqsave(&ioc->lock, flags);
1592 rcu_assign_pointer(ioc->ioc_data, cic);
1593 spin_unlock_irqrestore(&ioc->lock, flags);
1601 * Add cic into ioc, using cfqd as the search key. This enables us to lookup
1602 * the process specific cfq io context when entered from the block layer.
1603 * Also adds the cic to a per-cfqd list, used when this queue is removed.
1605 static int cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
1606 struct cfq_io_context *cic, gfp_t gfp_mask)
1608 unsigned long flags;
1611 ret = radix_tree_preload(gfp_mask);
1616 spin_lock_irqsave(&ioc->lock, flags);
1617 ret = radix_tree_insert(&ioc->radix_root,
1618 (unsigned long) cfqd, cic);
1620 hlist_add_head_rcu(&cic->cic_list, &ioc->cic_list);
1621 spin_unlock_irqrestore(&ioc->lock, flags);
1623 radix_tree_preload_end();
1626 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1627 list_add(&cic->queue_list, &cfqd->cic_list);
1628 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1633 printk(KERN_ERR "cfq: cic link failed!\n");
1639 * Setup general io context and cfq io context. There can be several cfq
1640 * io contexts per general io context, if this process is doing io to more
1641 * than one device managed by cfq.
1643 static struct cfq_io_context *
1644 cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
1646 struct io_context *ioc = NULL;
1647 struct cfq_io_context *cic;
1649 might_sleep_if(gfp_mask & __GFP_WAIT);
1651 ioc = get_io_context(gfp_mask, cfqd->queue->node);
1655 cic = cfq_cic_lookup(cfqd, ioc);
1659 cic = cfq_alloc_io_context(cfqd, gfp_mask);
1663 if (cfq_cic_link(cfqd, ioc, cic, gfp_mask))
1667 smp_read_barrier_depends();
1668 if (unlikely(ioc->ioprio_changed))
1669 cfq_ioc_set_ioprio(ioc);
1675 put_io_context(ioc);
1680 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
1682 unsigned long elapsed = jiffies - cic->last_end_request;
1683 unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
1685 cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
1686 cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
1687 cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
1691 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_io_context *cic,
1697 if (cic->last_request_pos < rq->sector)
1698 sdist = rq->sector - cic->last_request_pos;
1700 sdist = cic->last_request_pos - rq->sector;
1703 * Don't allow the seek distance to get too large from the
1704 * odd fragment, pagein, etc
1706 if (cic->seek_samples <= 60) /* second&third seek */
1707 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*1024);
1709 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*64);
1711 cic->seek_samples = (7*cic->seek_samples + 256) / 8;
1712 cic->seek_total = (7*cic->seek_total + (u64)256*sdist) / 8;
1713 total = cic->seek_total + (cic->seek_samples/2);
1714 do_div(total, cic->seek_samples);
1715 cic->seek_mean = (sector_t)total;
1719 * Disable idle window if the process thinks too long or seeks so much that
1723 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1724 struct cfq_io_context *cic)
1726 int old_idle, enable_idle;
1729 * Don't idle for async or idle io prio class
1731 if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
1734 enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
1736 if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle ||
1737 (cfqd->hw_tag && CIC_SEEKY(cic)))
1739 else if (sample_valid(cic->ttime_samples)) {
1740 if (cic->ttime_mean > cfqd->cfq_slice_idle)
1746 if (old_idle != enable_idle) {
1747 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
1749 cfq_mark_cfqq_idle_window(cfqq);
1751 cfq_clear_cfqq_idle_window(cfqq);
1756 * Check if new_cfqq should preempt the currently active queue. Return 0 for
1757 * no or if we aren't sure, a 1 will cause a preempt.
1760 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
1763 struct cfq_queue *cfqq;
1765 cfqq = cfqd->active_queue;
1769 if (cfq_slice_used(cfqq))
1772 if (cfq_class_idle(new_cfqq))
1775 if (cfq_class_idle(cfqq))
1779 * if the new request is sync, but the currently running queue is
1780 * not, let the sync request have priority.
1782 if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
1786 * So both queues are sync. Let the new request get disk time if
1787 * it's a metadata request and the current queue is doing regular IO.
1789 if (rq_is_meta(rq) && !cfqq->meta_pending)
1792 if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
1796 * if this request is as-good as one we would expect from the
1797 * current cfqq, let it preempt
1799 if (cfq_rq_close(cfqd, rq))
1806 * cfqq preempts the active queue. if we allowed preempt with no slice left,
1807 * let it have half of its nominal slice.
1809 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1811 cfq_log_cfqq(cfqd, cfqq, "preempt");
1812 cfq_slice_expired(cfqd, 1);
1815 * Put the new queue at the front of the of the current list,
1816 * so we know that it will be selected next.
1818 BUG_ON(!cfq_cfqq_on_rr(cfqq));
1820 cfq_service_tree_add(cfqd, cfqq, 1);
1822 cfqq->slice_end = 0;
1823 cfq_mark_cfqq_slice_new(cfqq);
1827 * Called when a new fs request (rq) is added (to cfqq). Check if there's
1828 * something we should do about it
1831 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1834 struct cfq_io_context *cic = RQ_CIC(rq);
1838 cfqq->meta_pending++;
1840 cfq_update_io_thinktime(cfqd, cic);
1841 cfq_update_io_seektime(cfqd, cic, rq);
1842 cfq_update_idle_window(cfqd, cfqq, cic);
1844 cic->last_request_pos = rq->sector + rq->nr_sectors;
1846 if (cfqq == cfqd->active_queue) {
1848 * if we are waiting for a request for this queue, let it rip
1849 * immediately and flag that we must not expire this queue
1852 if (cfq_cfqq_wait_request(cfqq)) {
1853 cfq_mark_cfqq_must_dispatch(cfqq);
1854 del_timer(&cfqd->idle_slice_timer);
1855 blk_start_queueing(cfqd->queue);
1857 } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
1859 * not the active queue - expire current slice if it is
1860 * idle and has expired it's mean thinktime or this new queue
1861 * has some old slice time left and is of higher priority
1863 cfq_preempt_queue(cfqd, cfqq);
1864 cfq_mark_cfqq_must_dispatch(cfqq);
1865 blk_start_queueing(cfqd->queue);
1869 static void cfq_insert_request(struct request_queue *q, struct request *rq)
1871 struct cfq_data *cfqd = q->elevator->elevator_data;
1872 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1874 cfq_log_cfqq(cfqd, cfqq, "insert_request");
1875 cfq_init_prio_data(cfqq, RQ_CIC(rq)->ioc);
1879 list_add_tail(&rq->queuelist, &cfqq->fifo);
1881 cfq_rq_enqueued(cfqd, cfqq, rq);
1885 * Update hw_tag based on peak queue depth over 50 samples under
1888 static void cfq_update_hw_tag(struct cfq_data *cfqd)
1890 if (cfqd->rq_in_driver > cfqd->rq_in_driver_peak)
1891 cfqd->rq_in_driver_peak = cfqd->rq_in_driver;
1893 if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
1894 cfqd->rq_in_driver <= CFQ_HW_QUEUE_MIN)
1897 if (cfqd->hw_tag_samples++ < 50)
1900 if (cfqd->rq_in_driver_peak >= CFQ_HW_QUEUE_MIN)
1905 cfqd->hw_tag_samples = 0;
1906 cfqd->rq_in_driver_peak = 0;
1909 static void cfq_completed_request(struct request_queue *q, struct request *rq)
1911 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1912 struct cfq_data *cfqd = cfqq->cfqd;
1913 const int sync = rq_is_sync(rq);
1917 cfq_log_cfqq(cfqd, cfqq, "complete");
1919 cfq_update_hw_tag(cfqd);
1921 WARN_ON(!cfqd->rq_in_driver);
1922 WARN_ON(!cfqq->dispatched);
1923 cfqd->rq_in_driver--;
1926 if (cfq_cfqq_sync(cfqq))
1927 cfqd->sync_flight--;
1929 if (!cfq_class_idle(cfqq))
1930 cfqd->last_end_request = now;
1933 RQ_CIC(rq)->last_end_request = now;
1936 * If this is the active queue, check if it needs to be expired,
1937 * or if we want to idle in case it has no pending requests.
1939 if (cfqd->active_queue == cfqq) {
1940 if (cfq_cfqq_slice_new(cfqq)) {
1941 cfq_set_prio_slice(cfqd, cfqq);
1942 cfq_clear_cfqq_slice_new(cfqq);
1944 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
1945 cfq_slice_expired(cfqd, 1);
1946 else if (sync && RB_EMPTY_ROOT(&cfqq->sort_list))
1947 cfq_arm_slice_timer(cfqd);
1950 if (!cfqd->rq_in_driver)
1951 cfq_schedule_dispatch(cfqd);
1955 * we temporarily boost lower priority queues if they are holding fs exclusive
1956 * resources. they are boosted to normal prio (CLASS_BE/4)
1958 static void cfq_prio_boost(struct cfq_queue *cfqq)
1960 if (has_fs_excl()) {
1962 * boost idle prio on transactions that would lock out other
1963 * users of the filesystem
1965 if (cfq_class_idle(cfqq))
1966 cfqq->ioprio_class = IOPRIO_CLASS_BE;
1967 if (cfqq->ioprio > IOPRIO_NORM)
1968 cfqq->ioprio = IOPRIO_NORM;
1971 * check if we need to unboost the queue
1973 if (cfqq->ioprio_class != cfqq->org_ioprio_class)
1974 cfqq->ioprio_class = cfqq->org_ioprio_class;
1975 if (cfqq->ioprio != cfqq->org_ioprio)
1976 cfqq->ioprio = cfqq->org_ioprio;
1980 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
1982 if ((cfq_cfqq_wait_request(cfqq) || cfq_cfqq_must_alloc(cfqq)) &&
1983 !cfq_cfqq_must_alloc_slice(cfqq)) {
1984 cfq_mark_cfqq_must_alloc_slice(cfqq);
1985 return ELV_MQUEUE_MUST;
1988 return ELV_MQUEUE_MAY;
1991 static int cfq_may_queue(struct request_queue *q, int rw)
1993 struct cfq_data *cfqd = q->elevator->elevator_data;
1994 struct task_struct *tsk = current;
1995 struct cfq_io_context *cic;
1996 struct cfq_queue *cfqq;
1999 * don't force setup of a queue from here, as a call to may_queue
2000 * does not necessarily imply that a request actually will be queued.
2001 * so just lookup a possibly existing queue, or return 'may queue'
2004 cic = cfq_cic_lookup(cfqd, tsk->io_context);
2006 return ELV_MQUEUE_MAY;
2008 cfqq = cic_to_cfqq(cic, rw & REQ_RW_SYNC);
2010 cfq_init_prio_data(cfqq, cic->ioc);
2011 cfq_prio_boost(cfqq);
2013 return __cfq_may_queue(cfqq);
2016 return ELV_MQUEUE_MAY;
2020 * queue lock held here
2022 static void cfq_put_request(struct request *rq)
2024 struct cfq_queue *cfqq = RQ_CFQQ(rq);
2027 const int rw = rq_data_dir(rq);
2029 BUG_ON(!cfqq->allocated[rw]);
2030 cfqq->allocated[rw]--;
2032 put_io_context(RQ_CIC(rq)->ioc);
2034 rq->elevator_private = NULL;
2035 rq->elevator_private2 = NULL;
2037 cfq_put_queue(cfqq);
2042 * Allocate cfq data structures associated with this request.
2045 cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
2047 struct cfq_data *cfqd = q->elevator->elevator_data;
2048 struct cfq_io_context *cic;
2049 const int rw = rq_data_dir(rq);
2050 const int is_sync = rq_is_sync(rq);
2051 struct cfq_queue *cfqq;
2052 unsigned long flags;
2054 might_sleep_if(gfp_mask & __GFP_WAIT);
2056 cic = cfq_get_io_context(cfqd, gfp_mask);
2058 spin_lock_irqsave(q->queue_lock, flags);
2063 cfqq = cic_to_cfqq(cic, is_sync);
2065 cfqq = cfq_get_queue(cfqd, is_sync, cic->ioc, gfp_mask);
2070 cic_set_cfqq(cic, cfqq, is_sync);
2073 cfqq->allocated[rw]++;
2074 cfq_clear_cfqq_must_alloc(cfqq);
2075 atomic_inc(&cfqq->ref);
2077 spin_unlock_irqrestore(q->queue_lock, flags);
2079 rq->elevator_private = cic;
2080 rq->elevator_private2 = cfqq;
2085 put_io_context(cic->ioc);
2087 cfq_schedule_dispatch(cfqd);
2088 spin_unlock_irqrestore(q->queue_lock, flags);
2089 cfq_log(cfqd, "set_request fail");
2093 static void cfq_kick_queue(struct work_struct *work)
2095 struct cfq_data *cfqd =
2096 container_of(work, struct cfq_data, unplug_work);
2097 struct request_queue *q = cfqd->queue;
2098 unsigned long flags;
2100 spin_lock_irqsave(q->queue_lock, flags);
2101 blk_start_queueing(q);
2102 spin_unlock_irqrestore(q->queue_lock, flags);
2106 * Timer running if the active_queue is currently idling inside its time slice
2108 static void cfq_idle_slice_timer(unsigned long data)
2110 struct cfq_data *cfqd = (struct cfq_data *) data;
2111 struct cfq_queue *cfqq;
2112 unsigned long flags;
2115 cfq_log(cfqd, "idle timer fired");
2117 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2119 cfqq = cfqd->active_queue;
2126 if (cfq_slice_used(cfqq))
2130 * only expire and reinvoke request handler, if there are
2131 * other queues with pending requests
2133 if (!cfqd->busy_queues)
2137 * not expired and it has a request pending, let it dispatch
2139 if (!RB_EMPTY_ROOT(&cfqq->sort_list)) {
2140 cfq_mark_cfqq_must_dispatch(cfqq);
2145 cfq_slice_expired(cfqd, timed_out);
2147 cfq_schedule_dispatch(cfqd);
2149 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2152 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
2154 del_timer_sync(&cfqd->idle_slice_timer);
2155 kblockd_flush_work(&cfqd->unplug_work);
2158 static void cfq_put_async_queues(struct cfq_data *cfqd)
2162 for (i = 0; i < IOPRIO_BE_NR; i++) {
2163 if (cfqd->async_cfqq[0][i])
2164 cfq_put_queue(cfqd->async_cfqq[0][i]);
2165 if (cfqd->async_cfqq[1][i])
2166 cfq_put_queue(cfqd->async_cfqq[1][i]);
2169 if (cfqd->async_idle_cfqq)
2170 cfq_put_queue(cfqd->async_idle_cfqq);
2173 static void cfq_exit_queue(elevator_t *e)
2175 struct cfq_data *cfqd = e->elevator_data;
2176 struct request_queue *q = cfqd->queue;
2178 cfq_shutdown_timer_wq(cfqd);
2180 spin_lock_irq(q->queue_lock);
2182 if (cfqd->active_queue)
2183 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
2185 while (!list_empty(&cfqd->cic_list)) {
2186 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
2187 struct cfq_io_context,
2190 __cfq_exit_single_io_context(cfqd, cic);
2193 cfq_put_async_queues(cfqd);
2195 spin_unlock_irq(q->queue_lock);
2197 cfq_shutdown_timer_wq(cfqd);
2202 static void *cfq_init_queue(struct request_queue *q)
2204 struct cfq_data *cfqd;
2206 cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
2210 cfqd->service_tree = CFQ_RB_ROOT;
2211 INIT_LIST_HEAD(&cfqd->cic_list);
2215 init_timer(&cfqd->idle_slice_timer);
2216 cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
2217 cfqd->idle_slice_timer.data = (unsigned long) cfqd;
2219 INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
2221 cfqd->last_end_request = jiffies;
2222 cfqd->cfq_quantum = cfq_quantum;
2223 cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
2224 cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
2225 cfqd->cfq_back_max = cfq_back_max;
2226 cfqd->cfq_back_penalty = cfq_back_penalty;
2227 cfqd->cfq_slice[0] = cfq_slice_async;
2228 cfqd->cfq_slice[1] = cfq_slice_sync;
2229 cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
2230 cfqd->cfq_slice_idle = cfq_slice_idle;
2236 static void cfq_slab_kill(void)
2239 * Caller already ensured that pending RCU callbacks are completed,
2240 * so we should have no busy allocations at this point.
2243 kmem_cache_destroy(cfq_pool);
2245 kmem_cache_destroy(cfq_ioc_pool);
2248 static int __init cfq_slab_setup(void)
2250 cfq_pool = KMEM_CACHE(cfq_queue, 0);
2254 cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0);
2265 * sysfs parts below -->
2268 cfq_var_show(unsigned int var, char *page)
2270 return sprintf(page, "%d\n", var);
2274 cfq_var_store(unsigned int *var, const char *page, size_t count)
2276 char *p = (char *) page;
2278 *var = simple_strtoul(p, &p, 10);
2282 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
2283 static ssize_t __FUNC(elevator_t *e, char *page) \
2285 struct cfq_data *cfqd = e->elevator_data; \
2286 unsigned int __data = __VAR; \
2288 __data = jiffies_to_msecs(__data); \
2289 return cfq_var_show(__data, (page)); \
2291 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
2292 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
2293 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
2294 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
2295 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
2296 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
2297 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
2298 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
2299 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
2300 #undef SHOW_FUNCTION
2302 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
2303 static ssize_t __FUNC(elevator_t *e, const char *page, size_t count) \
2305 struct cfq_data *cfqd = e->elevator_data; \
2306 unsigned int __data; \
2307 int ret = cfq_var_store(&__data, (page), count); \
2308 if (__data < (MIN)) \
2310 else if (__data > (MAX)) \
2313 *(__PTR) = msecs_to_jiffies(__data); \
2315 *(__PTR) = __data; \
2318 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
2319 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
2321 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
2323 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
2324 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
2326 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
2327 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
2328 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
2329 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
2331 #undef STORE_FUNCTION
2333 #define CFQ_ATTR(name) \
2334 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
2336 static struct elv_fs_entry cfq_attrs[] = {
2338 CFQ_ATTR(fifo_expire_sync),
2339 CFQ_ATTR(fifo_expire_async),
2340 CFQ_ATTR(back_seek_max),
2341 CFQ_ATTR(back_seek_penalty),
2342 CFQ_ATTR(slice_sync),
2343 CFQ_ATTR(slice_async),
2344 CFQ_ATTR(slice_async_rq),
2345 CFQ_ATTR(slice_idle),
2349 static struct elevator_type iosched_cfq = {
2351 .elevator_merge_fn = cfq_merge,
2352 .elevator_merged_fn = cfq_merged_request,
2353 .elevator_merge_req_fn = cfq_merged_requests,
2354 .elevator_allow_merge_fn = cfq_allow_merge,
2355 .elevator_dispatch_fn = cfq_dispatch_requests,
2356 .elevator_add_req_fn = cfq_insert_request,
2357 .elevator_activate_req_fn = cfq_activate_request,
2358 .elevator_deactivate_req_fn = cfq_deactivate_request,
2359 .elevator_queue_empty_fn = cfq_queue_empty,
2360 .elevator_completed_req_fn = cfq_completed_request,
2361 .elevator_former_req_fn = elv_rb_former_request,
2362 .elevator_latter_req_fn = elv_rb_latter_request,
2363 .elevator_set_req_fn = cfq_set_request,
2364 .elevator_put_req_fn = cfq_put_request,
2365 .elevator_may_queue_fn = cfq_may_queue,
2366 .elevator_init_fn = cfq_init_queue,
2367 .elevator_exit_fn = cfq_exit_queue,
2368 .trim = cfq_free_io_context,
2370 .elevator_attrs = cfq_attrs,
2371 .elevator_name = "cfq",
2372 .elevator_owner = THIS_MODULE,
2375 static int __init cfq_init(void)
2378 * could be 0 on HZ < 1000 setups
2380 if (!cfq_slice_async)
2381 cfq_slice_async = 1;
2382 if (!cfq_slice_idle)
2385 if (cfq_slab_setup())
2388 elv_register(&iosched_cfq);
2393 static void __exit cfq_exit(void)
2395 DECLARE_COMPLETION_ONSTACK(all_gone);
2396 elv_unregister(&iosched_cfq);
2397 ioc_gone = &all_gone;
2398 /* ioc_gone's update must be visible before reading ioc_count */
2402 * this also protects us from entering cfq_slab_kill() with
2403 * pending RCU callbacks
2405 if (elv_ioc_count_read(ioc_count))
2406 wait_for_completion(&all_gone);
2410 module_init(cfq_init);
2411 module_exit(cfq_exit);
2413 MODULE_AUTHOR("Jens Axboe");
2414 MODULE_LICENSE("GPL");
2415 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");