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, cfq_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)
59 #define sample_valid(samples) ((samples) > 80)
62 * Most of our rbtree usage is for sorting with min extraction, so
63 * if we cache the leftmost node we don't have to walk down the tree
64 * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should
65 * move this into the elevator for the rq sorting as well.
71 #define CFQ_RB_ROOT (struct cfq_rb_root) { RB_ROOT, NULL, }
74 * Per process-grouping structure
79 /* various state flags, see below */
82 struct cfq_data *cfqd;
83 /* service_tree member */
84 struct rb_node rb_node;
85 /* service_tree key */
87 /* prio tree member */
88 struct rb_node p_node;
89 /* prio tree root we belong to, if any */
90 struct rb_root *p_root;
91 /* sorted list of pending requests */
92 struct rb_root sort_list;
93 /* if fifo isn't expired, next request to serve */
94 struct request *next_rq;
95 /* requests queued in sort_list */
97 /* currently allocated requests */
99 /* fifo list of requests in sort_list */
100 struct list_head fifo;
102 unsigned long slice_end;
104 unsigned int slice_dispatch;
106 /* pending metadata requests */
108 /* number of requests that are on the dispatch list or inside driver */
111 /* io prio of this group */
112 unsigned short ioprio, org_ioprio;
113 unsigned short ioprio_class, org_ioprio_class;
115 unsigned int seek_samples;
118 sector_t last_request_pos;
124 * Per block device queue structure
127 struct request_queue *queue;
130 * rr list of queues with requests and the count of them
132 struct cfq_rb_root service_tree;
135 * Each priority tree is sorted by next_request position. These
136 * trees are used when determining if two or more queues are
137 * interleaving requests (see cfq_close_cooperator).
139 struct rb_root prio_trees[CFQ_PRIO_LISTS];
141 unsigned int busy_queues;
147 * queue-depth detection
152 int rq_in_driver_peak;
155 * idle window management
157 struct timer_list idle_slice_timer;
158 struct work_struct unplug_work;
160 struct cfq_queue *active_queue;
161 struct cfq_io_context *active_cic;
164 * async queue for each priority case
166 struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
167 struct cfq_queue *async_idle_cfqq;
169 sector_t last_position;
172 * tunables, see top of file
174 unsigned int cfq_quantum;
175 unsigned int cfq_fifo_expire[2];
176 unsigned int cfq_back_penalty;
177 unsigned int cfq_back_max;
178 unsigned int cfq_slice[2];
179 unsigned int cfq_slice_async_rq;
180 unsigned int cfq_slice_idle;
181 unsigned int cfq_latency;
183 struct list_head cic_list;
186 * Fallback dummy cfqq for extreme OOM conditions
188 struct cfq_queue oom_cfqq;
190 unsigned long last_end_sync_rq;
193 enum cfqq_state_flags {
194 CFQ_CFQQ_FLAG_on_rr = 0, /* on round-robin busy list */
195 CFQ_CFQQ_FLAG_wait_request, /* waiting for a request */
196 CFQ_CFQQ_FLAG_must_dispatch, /* must be allowed a dispatch */
197 CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
198 CFQ_CFQQ_FLAG_fifo_expire, /* FIFO checked in this slice */
199 CFQ_CFQQ_FLAG_idle_window, /* slice idling enabled */
200 CFQ_CFQQ_FLAG_prio_changed, /* task priority has changed */
201 CFQ_CFQQ_FLAG_slice_new, /* no requests dispatched in slice */
202 CFQ_CFQQ_FLAG_sync, /* synchronous queue */
203 CFQ_CFQQ_FLAG_coop, /* has done a coop jump of the queue */
206 #define CFQ_CFQQ_FNS(name) \
207 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \
209 (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name); \
211 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \
213 (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \
215 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \
217 return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \
221 CFQ_CFQQ_FNS(wait_request);
222 CFQ_CFQQ_FNS(must_dispatch);
223 CFQ_CFQQ_FNS(must_alloc_slice);
224 CFQ_CFQQ_FNS(fifo_expire);
225 CFQ_CFQQ_FNS(idle_window);
226 CFQ_CFQQ_FNS(prio_changed);
227 CFQ_CFQQ_FNS(slice_new);
232 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
233 blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args)
234 #define cfq_log(cfqd, fmt, args...) \
235 blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
237 static void cfq_dispatch_insert(struct request_queue *, struct request *);
238 static struct cfq_queue *cfq_get_queue(struct cfq_data *, bool,
239 struct io_context *, gfp_t);
240 static struct cfq_io_context *cfq_cic_lookup(struct cfq_data *,
241 struct io_context *);
243 static inline int rq_in_driver(struct cfq_data *cfqd)
245 return cfqd->rq_in_driver[0] + cfqd->rq_in_driver[1];
248 static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_context *cic,
251 return cic->cfqq[is_sync];
254 static inline void cic_set_cfqq(struct cfq_io_context *cic,
255 struct cfq_queue *cfqq, bool is_sync)
257 cic->cfqq[is_sync] = cfqq;
261 * We regard a request as SYNC, if it's either a read or has the SYNC bit
262 * set (in which case it could also be direct WRITE).
264 static inline bool cfq_bio_sync(struct bio *bio)
266 return bio_data_dir(bio) == READ || bio_rw_flagged(bio, BIO_RW_SYNCIO);
270 * scheduler run of queue, if there are requests pending and no one in the
271 * driver that will restart queueing
273 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
275 if (cfqd->busy_queues) {
276 cfq_log(cfqd, "schedule dispatch");
277 kblockd_schedule_work(cfqd->queue, &cfqd->unplug_work);
281 static int cfq_queue_empty(struct request_queue *q)
283 struct cfq_data *cfqd = q->elevator->elevator_data;
285 return !cfqd->busy_queues;
289 * Scale schedule slice based on io priority. Use the sync time slice only
290 * if a queue is marked sync and has sync io queued. A sync queue with async
291 * io only, should not get full sync slice length.
293 static inline int cfq_prio_slice(struct cfq_data *cfqd, bool sync,
296 const int base_slice = cfqd->cfq_slice[sync];
298 WARN_ON(prio >= IOPRIO_BE_NR);
300 return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
304 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
306 return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
310 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
312 cfqq->slice_end = cfq_prio_to_slice(cfqd, cfqq) + jiffies;
313 cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
317 * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
318 * isn't valid until the first request from the dispatch is activated
319 * and the slice time set.
321 static inline bool cfq_slice_used(struct cfq_queue *cfqq)
323 if (cfq_cfqq_slice_new(cfqq))
325 if (time_before(jiffies, cfqq->slice_end))
332 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
333 * We choose the request that is closest to the head right now. Distance
334 * behind the head is penalized and only allowed to a certain extent.
336 static struct request *
337 cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2)
339 sector_t last, s1, s2, d1 = 0, d2 = 0;
340 unsigned long back_max;
341 #define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */
342 #define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */
343 unsigned wrap = 0; /* bit mask: requests behind the disk head? */
345 if (rq1 == NULL || rq1 == rq2)
350 if (rq_is_sync(rq1) && !rq_is_sync(rq2))
352 else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
354 if (rq_is_meta(rq1) && !rq_is_meta(rq2))
356 else if (rq_is_meta(rq2) && !rq_is_meta(rq1))
359 s1 = blk_rq_pos(rq1);
360 s2 = blk_rq_pos(rq2);
362 last = cfqd->last_position;
365 * by definition, 1KiB is 2 sectors
367 back_max = cfqd->cfq_back_max * 2;
370 * Strict one way elevator _except_ in the case where we allow
371 * short backward seeks which are biased as twice the cost of a
372 * similar forward seek.
376 else if (s1 + back_max >= last)
377 d1 = (last - s1) * cfqd->cfq_back_penalty;
379 wrap |= CFQ_RQ1_WRAP;
383 else if (s2 + back_max >= last)
384 d2 = (last - s2) * cfqd->cfq_back_penalty;
386 wrap |= CFQ_RQ2_WRAP;
388 /* Found required data */
391 * By doing switch() on the bit mask "wrap" we avoid having to
392 * check two variables for all permutations: --> faster!
395 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
411 case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
414 * Since both rqs are wrapped,
415 * start with the one that's further behind head
416 * (--> only *one* back seek required),
417 * since back seek takes more time than forward.
427 * The below is leftmost cache rbtree addon
429 static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
432 root->left = rb_first(&root->rb);
435 return rb_entry(root->left, struct cfq_queue, rb_node);
440 static void rb_erase_init(struct rb_node *n, struct rb_root *root)
446 static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
450 rb_erase_init(n, &root->rb);
454 * would be nice to take fifo expire time into account as well
456 static struct request *
457 cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
458 struct request *last)
460 struct rb_node *rbnext = rb_next(&last->rb_node);
461 struct rb_node *rbprev = rb_prev(&last->rb_node);
462 struct request *next = NULL, *prev = NULL;
464 BUG_ON(RB_EMPTY_NODE(&last->rb_node));
467 prev = rb_entry_rq(rbprev);
470 next = rb_entry_rq(rbnext);
472 rbnext = rb_first(&cfqq->sort_list);
473 if (rbnext && rbnext != &last->rb_node)
474 next = rb_entry_rq(rbnext);
477 return cfq_choose_req(cfqd, next, prev);
480 static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
481 struct cfq_queue *cfqq)
484 * just an approximation, should be ok.
486 return (cfqd->busy_queues - 1) * (cfq_prio_slice(cfqd, 1, 0) -
487 cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
491 * The cfqd->service_tree holds all pending cfq_queue's that have
492 * requests waiting to be processed. It is sorted in the order that
493 * we will service the queues.
495 static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
498 struct rb_node **p, *parent;
499 struct cfq_queue *__cfqq;
500 unsigned long rb_key;
503 if (cfq_class_idle(cfqq)) {
504 rb_key = CFQ_IDLE_DELAY;
505 parent = rb_last(&cfqd->service_tree.rb);
506 if (parent && parent != &cfqq->rb_node) {
507 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
508 rb_key += __cfqq->rb_key;
511 } else if (!add_front) {
513 * Get our rb key offset. Subtract any residual slice
514 * value carried from last service. A negative resid
515 * count indicates slice overrun, and this should position
516 * the next service time further away in the tree.
518 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
519 rb_key -= cfqq->slice_resid;
520 cfqq->slice_resid = 0;
523 __cfqq = cfq_rb_first(&cfqd->service_tree);
524 rb_key += __cfqq ? __cfqq->rb_key : jiffies;
527 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
529 * same position, nothing more to do
531 if (rb_key == cfqq->rb_key)
534 cfq_rb_erase(&cfqq->rb_node, &cfqd->service_tree);
539 p = &cfqd->service_tree.rb.rb_node;
544 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
547 * sort RT queues first, we always want to give
548 * preference to them. IDLE queues goes to the back.
549 * after that, sort on the next service time.
551 if (cfq_class_rt(cfqq) > cfq_class_rt(__cfqq))
553 else if (cfq_class_rt(cfqq) < cfq_class_rt(__cfqq))
555 else if (cfq_class_idle(cfqq) < cfq_class_idle(__cfqq))
557 else if (cfq_class_idle(cfqq) > cfq_class_idle(__cfqq))
559 else if (time_before(rb_key, __cfqq->rb_key))
564 if (n == &(*p)->rb_right)
571 cfqd->service_tree.left = &cfqq->rb_node;
573 cfqq->rb_key = rb_key;
574 rb_link_node(&cfqq->rb_node, parent, p);
575 rb_insert_color(&cfqq->rb_node, &cfqd->service_tree.rb);
578 static struct cfq_queue *
579 cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
580 sector_t sector, struct rb_node **ret_parent,
581 struct rb_node ***rb_link)
583 struct rb_node **p, *parent;
584 struct cfq_queue *cfqq = NULL;
592 cfqq = rb_entry(parent, struct cfq_queue, p_node);
595 * Sort strictly based on sector. Smallest to the left,
596 * largest to the right.
598 if (sector > blk_rq_pos(cfqq->next_rq))
600 else if (sector < blk_rq_pos(cfqq->next_rq))
608 *ret_parent = parent;
614 static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
616 struct rb_node **p, *parent;
617 struct cfq_queue *__cfqq;
620 rb_erase(&cfqq->p_node, cfqq->p_root);
624 if (cfq_class_idle(cfqq))
629 cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
630 __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
631 blk_rq_pos(cfqq->next_rq), &parent, &p);
633 rb_link_node(&cfqq->p_node, parent, p);
634 rb_insert_color(&cfqq->p_node, cfqq->p_root);
640 * Update cfqq's position in the service tree.
642 static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
645 * Resorting requires the cfqq to be on the RR list already.
647 if (cfq_cfqq_on_rr(cfqq)) {
648 cfq_service_tree_add(cfqd, cfqq, 0);
649 cfq_prio_tree_add(cfqd, cfqq);
654 * add to busy list of queues for service, trying to be fair in ordering
655 * the pending list according to last request service
657 static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
659 cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
660 BUG_ON(cfq_cfqq_on_rr(cfqq));
661 cfq_mark_cfqq_on_rr(cfqq);
664 cfq_resort_rr_list(cfqd, cfqq);
668 * Called when the cfqq no longer has requests pending, remove it from
671 static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
673 cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
674 BUG_ON(!cfq_cfqq_on_rr(cfqq));
675 cfq_clear_cfqq_on_rr(cfqq);
677 if (!RB_EMPTY_NODE(&cfqq->rb_node))
678 cfq_rb_erase(&cfqq->rb_node, &cfqd->service_tree);
680 rb_erase(&cfqq->p_node, cfqq->p_root);
684 BUG_ON(!cfqd->busy_queues);
689 * rb tree support functions
691 static void cfq_del_rq_rb(struct request *rq)
693 struct cfq_queue *cfqq = RQ_CFQQ(rq);
694 struct cfq_data *cfqd = cfqq->cfqd;
695 const int sync = rq_is_sync(rq);
697 BUG_ON(!cfqq->queued[sync]);
698 cfqq->queued[sync]--;
700 elv_rb_del(&cfqq->sort_list, rq);
702 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
703 cfq_del_cfqq_rr(cfqd, cfqq);
706 static void cfq_add_rq_rb(struct request *rq)
708 struct cfq_queue *cfqq = RQ_CFQQ(rq);
709 struct cfq_data *cfqd = cfqq->cfqd;
710 struct request *__alias, *prev;
712 cfqq->queued[rq_is_sync(rq)]++;
715 * looks a little odd, but the first insert might return an alias.
716 * if that happens, put the alias on the dispatch list
718 while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL)
719 cfq_dispatch_insert(cfqd->queue, __alias);
721 if (!cfq_cfqq_on_rr(cfqq))
722 cfq_add_cfqq_rr(cfqd, cfqq);
725 * check if this request is a better next-serve candidate
727 prev = cfqq->next_rq;
728 cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq);
731 * adjust priority tree position, if ->next_rq changes
733 if (prev != cfqq->next_rq)
734 cfq_prio_tree_add(cfqd, cfqq);
736 BUG_ON(!cfqq->next_rq);
739 static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
741 elv_rb_del(&cfqq->sort_list, rq);
742 cfqq->queued[rq_is_sync(rq)]--;
746 static struct request *
747 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
749 struct task_struct *tsk = current;
750 struct cfq_io_context *cic;
751 struct cfq_queue *cfqq;
753 cic = cfq_cic_lookup(cfqd, tsk->io_context);
757 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
759 sector_t sector = bio->bi_sector + bio_sectors(bio);
761 return elv_rb_find(&cfqq->sort_list, sector);
767 static void cfq_activate_request(struct request_queue *q, struct request *rq)
769 struct cfq_data *cfqd = q->elevator->elevator_data;
771 cfqd->rq_in_driver[rq_is_sync(rq)]++;
772 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
775 cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
778 static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
780 struct cfq_data *cfqd = q->elevator->elevator_data;
781 const int sync = rq_is_sync(rq);
783 WARN_ON(!cfqd->rq_in_driver[sync]);
784 cfqd->rq_in_driver[sync]--;
785 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
789 static void cfq_remove_request(struct request *rq)
791 struct cfq_queue *cfqq = RQ_CFQQ(rq);
793 if (cfqq->next_rq == rq)
794 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
796 list_del_init(&rq->queuelist);
799 cfqq->cfqd->rq_queued--;
800 if (rq_is_meta(rq)) {
801 WARN_ON(!cfqq->meta_pending);
802 cfqq->meta_pending--;
806 static int cfq_merge(struct request_queue *q, struct request **req,
809 struct cfq_data *cfqd = q->elevator->elevator_data;
810 struct request *__rq;
812 __rq = cfq_find_rq_fmerge(cfqd, bio);
813 if (__rq && elv_rq_merge_ok(__rq, bio)) {
815 return ELEVATOR_FRONT_MERGE;
818 return ELEVATOR_NO_MERGE;
821 static void cfq_merged_request(struct request_queue *q, struct request *req,
824 if (type == ELEVATOR_FRONT_MERGE) {
825 struct cfq_queue *cfqq = RQ_CFQQ(req);
827 cfq_reposition_rq_rb(cfqq, req);
832 cfq_merged_requests(struct request_queue *q, struct request *rq,
833 struct request *next)
836 * reposition in fifo if next is older than rq
838 if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
839 time_before(rq_fifo_time(next), rq_fifo_time(rq))) {
840 list_move(&rq->queuelist, &next->queuelist);
841 rq_set_fifo_time(rq, rq_fifo_time(next));
844 cfq_remove_request(next);
847 static int cfq_allow_merge(struct request_queue *q, struct request *rq,
850 struct cfq_data *cfqd = q->elevator->elevator_data;
851 struct cfq_io_context *cic;
852 struct cfq_queue *cfqq;
855 * Disallow merge of a sync bio into an async request.
857 if (cfq_bio_sync(bio) && !rq_is_sync(rq))
861 * Lookup the cfqq that this bio will be queued with. Allow
862 * merge only if rq is queued there.
864 cic = cfq_cic_lookup(cfqd, current->io_context);
868 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
869 return cfqq == RQ_CFQQ(rq);
872 static void __cfq_set_active_queue(struct cfq_data *cfqd,
873 struct cfq_queue *cfqq)
876 cfq_log_cfqq(cfqd, cfqq, "set_active");
878 cfqq->slice_dispatch = 0;
880 cfq_clear_cfqq_wait_request(cfqq);
881 cfq_clear_cfqq_must_dispatch(cfqq);
882 cfq_clear_cfqq_must_alloc_slice(cfqq);
883 cfq_clear_cfqq_fifo_expire(cfqq);
884 cfq_mark_cfqq_slice_new(cfqq);
886 del_timer(&cfqd->idle_slice_timer);
889 cfqd->active_queue = cfqq;
893 * current cfqq expired its slice (or was too idle), select new one
896 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
899 cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
901 if (cfq_cfqq_wait_request(cfqq))
902 del_timer(&cfqd->idle_slice_timer);
904 cfq_clear_cfqq_wait_request(cfqq);
907 * store what was left of this slice, if the queue idled/timed out
909 if (timed_out && !cfq_cfqq_slice_new(cfqq)) {
910 cfqq->slice_resid = cfqq->slice_end - jiffies;
911 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
914 cfq_resort_rr_list(cfqd, cfqq);
916 if (cfqq == cfqd->active_queue)
917 cfqd->active_queue = NULL;
919 if (cfqd->active_cic) {
920 put_io_context(cfqd->active_cic->ioc);
921 cfqd->active_cic = NULL;
925 static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out)
927 struct cfq_queue *cfqq = cfqd->active_queue;
930 __cfq_slice_expired(cfqd, cfqq, timed_out);
934 * Get next queue for service. Unless we have a queue preemption,
935 * we'll simply select the first cfqq in the service tree.
937 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
939 if (RB_EMPTY_ROOT(&cfqd->service_tree.rb))
942 return cfq_rb_first(&cfqd->service_tree);
946 * Get and set a new active queue for service.
948 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
949 struct cfq_queue *cfqq)
952 cfqq = cfq_get_next_queue(cfqd);
954 cfq_clear_cfqq_coop(cfqq);
957 __cfq_set_active_queue(cfqd, cfqq);
961 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
964 if (blk_rq_pos(rq) >= cfqd->last_position)
965 return blk_rq_pos(rq) - cfqd->last_position;
967 return cfqd->last_position - blk_rq_pos(rq);
970 #define CFQQ_SEEK_THR 8 * 1024
971 #define CFQQ_SEEKY(cfqq) ((cfqq)->seek_mean > CFQQ_SEEK_THR)
973 static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq,
976 sector_t sdist = cfqq->seek_mean;
978 if (!sample_valid(cfqq->seek_samples))
979 sdist = CFQQ_SEEK_THR;
981 return cfq_dist_from_last(cfqd, rq) <= sdist;
984 static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
985 struct cfq_queue *cur_cfqq)
987 struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
988 struct rb_node *parent, *node;
989 struct cfq_queue *__cfqq;
990 sector_t sector = cfqd->last_position;
992 if (RB_EMPTY_ROOT(root))
996 * First, if we find a request starting at the end of the last
997 * request, choose it.
999 __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
1004 * If the exact sector wasn't found, the parent of the NULL leaf
1005 * will contain the closest sector.
1007 __cfqq = rb_entry(parent, struct cfq_queue, p_node);
1008 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1011 if (blk_rq_pos(__cfqq->next_rq) < sector)
1012 node = rb_next(&__cfqq->p_node);
1014 node = rb_prev(&__cfqq->p_node);
1018 __cfqq = rb_entry(node, struct cfq_queue, p_node);
1019 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1027 * cur_cfqq - passed in so that we don't decide that the current queue is
1028 * closely cooperating with itself.
1030 * So, basically we're assuming that that cur_cfqq has dispatched at least
1031 * one request, and that cfqd->last_position reflects a position on the disk
1032 * associated with the I/O issued by cur_cfqq. I'm not sure this is a valid
1035 static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
1036 struct cfq_queue *cur_cfqq,
1039 struct cfq_queue *cfqq;
1042 * We should notice if some of the queues are cooperating, eg
1043 * working closely on the same area of the disk. In that case,
1044 * we can group them together and don't waste time idling.
1046 cfqq = cfqq_close(cfqd, cur_cfqq);
1050 if (cfq_cfqq_coop(cfqq))
1054 cfq_mark_cfqq_coop(cfqq);
1058 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
1060 struct cfq_queue *cfqq = cfqd->active_queue;
1061 struct cfq_io_context *cic;
1065 * SSD device without seek penalty, disable idling. But only do so
1066 * for devices that support queuing, otherwise we still have a problem
1067 * with sync vs async workloads.
1069 if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
1072 WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
1073 WARN_ON(cfq_cfqq_slice_new(cfqq));
1076 * idle is disabled, either manually or by past process history
1078 if (!cfqd->cfq_slice_idle || !cfq_cfqq_idle_window(cfqq))
1082 * still requests with the driver, don't idle
1084 if (rq_in_driver(cfqd))
1088 * task has exited, don't wait
1090 cic = cfqd->active_cic;
1091 if (!cic || !atomic_read(&cic->ioc->nr_tasks))
1095 * If our average think time is larger than the remaining time
1096 * slice, then don't idle. This avoids overrunning the allotted
1099 if (sample_valid(cic->ttime_samples) &&
1100 (cfqq->slice_end - jiffies < cic->ttime_mean))
1103 cfq_mark_cfqq_wait_request(cfqq);
1106 * we don't want to idle for seeks, but we do want to allow
1107 * fair distribution of slice time for a process doing back-to-back
1108 * seeks. so allow a little bit of time for him to submit a new rq
1110 sl = cfqd->cfq_slice_idle;
1111 if (sample_valid(cfqq->seek_samples) && CFQQ_SEEKY(cfqq))
1112 sl = min(sl, msecs_to_jiffies(CFQ_MIN_TT));
1114 mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
1115 cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu", sl);
1119 * Move request from internal lists to the request queue dispatch list.
1121 static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
1123 struct cfq_data *cfqd = q->elevator->elevator_data;
1124 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1126 cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
1128 cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
1129 cfq_remove_request(rq);
1131 elv_dispatch_sort(q, rq);
1133 if (cfq_cfqq_sync(cfqq))
1134 cfqd->sync_flight++;
1138 * return expired entry, or NULL to just start from scratch in rbtree
1140 static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
1142 struct request *rq = NULL;
1144 if (cfq_cfqq_fifo_expire(cfqq))
1147 cfq_mark_cfqq_fifo_expire(cfqq);
1149 if (list_empty(&cfqq->fifo))
1152 rq = rq_entry_fifo(cfqq->fifo.next);
1153 if (time_before(jiffies, rq_fifo_time(rq)))
1156 cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq);
1161 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1163 const int base_rq = cfqd->cfq_slice_async_rq;
1165 WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
1167 return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
1171 * Select a queue for service. If we have a current active queue,
1172 * check whether to continue servicing it, or retrieve and set a new one.
1174 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
1176 struct cfq_queue *cfqq, *new_cfqq = NULL;
1178 cfqq = cfqd->active_queue;
1183 * The active queue has run out of time, expire it and select new.
1185 if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq))
1189 * The active queue has requests and isn't expired, allow it to
1192 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
1196 * If another queue has a request waiting within our mean seek
1197 * distance, let it run. The expire code will check for close
1198 * cooperators and put the close queue at the front of the service
1201 new_cfqq = cfq_close_cooperator(cfqd, cfqq, 0);
1206 * No requests pending. If the active queue still has requests in
1207 * flight or is idling for a new request, allow either of these
1208 * conditions to happen (or time out) before selecting a new queue.
1210 if (timer_pending(&cfqd->idle_slice_timer) ||
1211 (cfqq->dispatched && cfq_cfqq_idle_window(cfqq))) {
1217 cfq_slice_expired(cfqd, 0);
1219 cfqq = cfq_set_active_queue(cfqd, new_cfqq);
1224 static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
1228 while (cfqq->next_rq) {
1229 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
1233 BUG_ON(!list_empty(&cfqq->fifo));
1238 * Drain our current requests. Used for barriers and when switching
1239 * io schedulers on-the-fly.
1241 static int cfq_forced_dispatch(struct cfq_data *cfqd)
1243 struct cfq_queue *cfqq;
1246 while ((cfqq = cfq_rb_first(&cfqd->service_tree)) != NULL)
1247 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
1249 cfq_slice_expired(cfqd, 0);
1251 BUG_ON(cfqd->busy_queues);
1253 cfq_log(cfqd, "forced_dispatch=%d", dispatched);
1257 static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1259 unsigned int max_dispatch;
1262 * Drain async requests before we start sync IO
1264 if (cfq_cfqq_idle_window(cfqq) && cfqd->rq_in_driver[BLK_RW_ASYNC])
1268 * If this is an async queue and we have sync IO in flight, let it wait
1270 if (cfqd->sync_flight && !cfq_cfqq_sync(cfqq))
1273 max_dispatch = cfqd->cfq_quantum;
1274 if (cfq_class_idle(cfqq))
1278 * Does this cfqq already have too much IO in flight?
1280 if (cfqq->dispatched >= max_dispatch) {
1282 * idle queue must always only have a single IO in flight
1284 if (cfq_class_idle(cfqq))
1288 * We have other queues, don't allow more IO from this one
1290 if (cfqd->busy_queues > 1)
1294 * Sole queue user, allow bigger slice
1300 * Async queues must wait a bit before being allowed dispatch.
1301 * We also ramp up the dispatch depth gradually for async IO,
1302 * based on the last sync IO we serviced
1304 if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
1305 unsigned long last_sync = jiffies - cfqd->last_end_sync_rq;
1308 depth = last_sync / cfqd->cfq_slice[1];
1309 if (!depth && !cfqq->dispatched)
1311 if (depth < max_dispatch)
1312 max_dispatch = depth;
1316 * If we're below the current max, allow a dispatch
1318 return cfqq->dispatched < max_dispatch;
1322 * Dispatch a request from cfqq, moving them to the request queue
1325 static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1329 BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
1331 if (!cfq_may_dispatch(cfqd, cfqq))
1335 * follow expired path, else get first next available
1337 rq = cfq_check_fifo(cfqq);
1342 * insert request into driver dispatch list
1344 cfq_dispatch_insert(cfqd->queue, rq);
1346 if (!cfqd->active_cic) {
1347 struct cfq_io_context *cic = RQ_CIC(rq);
1349 atomic_long_inc(&cic->ioc->refcount);
1350 cfqd->active_cic = cic;
1357 * Find the cfqq that we need to service and move a request from that to the
1360 static int cfq_dispatch_requests(struct request_queue *q, int force)
1362 struct cfq_data *cfqd = q->elevator->elevator_data;
1363 struct cfq_queue *cfqq;
1365 if (!cfqd->busy_queues)
1368 if (unlikely(force))
1369 return cfq_forced_dispatch(cfqd);
1371 cfqq = cfq_select_queue(cfqd);
1376 * Dispatch a request from this cfqq, if it is allowed
1378 if (!cfq_dispatch_request(cfqd, cfqq))
1381 cfqq->slice_dispatch++;
1382 cfq_clear_cfqq_must_dispatch(cfqq);
1385 * expire an async queue immediately if it has used up its slice. idle
1386 * queue always expire after 1 dispatch round.
1388 if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
1389 cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
1390 cfq_class_idle(cfqq))) {
1391 cfqq->slice_end = jiffies + 1;
1392 cfq_slice_expired(cfqd, 0);
1395 cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
1400 * task holds one reference to the queue, dropped when task exits. each rq
1401 * in-flight on this queue also holds a reference, dropped when rq is freed.
1403 * queue lock must be held here.
1405 static void cfq_put_queue(struct cfq_queue *cfqq)
1407 struct cfq_data *cfqd = cfqq->cfqd;
1409 BUG_ON(atomic_read(&cfqq->ref) <= 0);
1411 if (!atomic_dec_and_test(&cfqq->ref))
1414 cfq_log_cfqq(cfqd, cfqq, "put_queue");
1415 BUG_ON(rb_first(&cfqq->sort_list));
1416 BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
1417 BUG_ON(cfq_cfqq_on_rr(cfqq));
1419 if (unlikely(cfqd->active_queue == cfqq)) {
1420 __cfq_slice_expired(cfqd, cfqq, 0);
1421 cfq_schedule_dispatch(cfqd);
1424 kmem_cache_free(cfq_pool, cfqq);
1428 * Must always be called with the rcu_read_lock() held
1431 __call_for_each_cic(struct io_context *ioc,
1432 void (*func)(struct io_context *, struct cfq_io_context *))
1434 struct cfq_io_context *cic;
1435 struct hlist_node *n;
1437 hlist_for_each_entry_rcu(cic, n, &ioc->cic_list, cic_list)
1442 * Call func for each cic attached to this ioc.
1445 call_for_each_cic(struct io_context *ioc,
1446 void (*func)(struct io_context *, struct cfq_io_context *))
1449 __call_for_each_cic(ioc, func);
1453 static void cfq_cic_free_rcu(struct rcu_head *head)
1455 struct cfq_io_context *cic;
1457 cic = container_of(head, struct cfq_io_context, rcu_head);
1459 kmem_cache_free(cfq_ioc_pool, cic);
1460 elv_ioc_count_dec(cfq_ioc_count);
1464 * CFQ scheduler is exiting, grab exit lock and check
1465 * the pending io context count. If it hits zero,
1466 * complete ioc_gone and set it back to NULL
1468 spin_lock(&ioc_gone_lock);
1469 if (ioc_gone && !elv_ioc_count_read(cfq_ioc_count)) {
1473 spin_unlock(&ioc_gone_lock);
1477 static void cfq_cic_free(struct cfq_io_context *cic)
1479 call_rcu(&cic->rcu_head, cfq_cic_free_rcu);
1482 static void cic_free_func(struct io_context *ioc, struct cfq_io_context *cic)
1484 unsigned long flags;
1486 BUG_ON(!cic->dead_key);
1488 spin_lock_irqsave(&ioc->lock, flags);
1489 radix_tree_delete(&ioc->radix_root, cic->dead_key);
1490 hlist_del_rcu(&cic->cic_list);
1491 spin_unlock_irqrestore(&ioc->lock, flags);
1497 * Must be called with rcu_read_lock() held or preemption otherwise disabled.
1498 * Only two callers of this - ->dtor() which is called with the rcu_read_lock(),
1499 * and ->trim() which is called with the task lock held
1501 static void cfq_free_io_context(struct io_context *ioc)
1504 * ioc->refcount is zero here, or we are called from elv_unregister(),
1505 * so no more cic's are allowed to be linked into this ioc. So it
1506 * should be ok to iterate over the known list, we will see all cic's
1507 * since no new ones are added.
1509 __call_for_each_cic(ioc, cic_free_func);
1512 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1514 if (unlikely(cfqq == cfqd->active_queue)) {
1515 __cfq_slice_expired(cfqd, cfqq, 0);
1516 cfq_schedule_dispatch(cfqd);
1519 cfq_put_queue(cfqq);
1522 static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
1523 struct cfq_io_context *cic)
1525 struct io_context *ioc = cic->ioc;
1527 list_del_init(&cic->queue_list);
1530 * Make sure key == NULL is seen for dead queues
1533 cic->dead_key = (unsigned long) cic->key;
1536 if (ioc->ioc_data == cic)
1537 rcu_assign_pointer(ioc->ioc_data, NULL);
1539 if (cic->cfqq[BLK_RW_ASYNC]) {
1540 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]);
1541 cic->cfqq[BLK_RW_ASYNC] = NULL;
1544 if (cic->cfqq[BLK_RW_SYNC]) {
1545 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]);
1546 cic->cfqq[BLK_RW_SYNC] = NULL;
1550 static void cfq_exit_single_io_context(struct io_context *ioc,
1551 struct cfq_io_context *cic)
1553 struct cfq_data *cfqd = cic->key;
1556 struct request_queue *q = cfqd->queue;
1557 unsigned long flags;
1559 spin_lock_irqsave(q->queue_lock, flags);
1562 * Ensure we get a fresh copy of the ->key to prevent
1563 * race between exiting task and queue
1565 smp_read_barrier_depends();
1567 __cfq_exit_single_io_context(cfqd, cic);
1569 spin_unlock_irqrestore(q->queue_lock, flags);
1574 * The process that ioc belongs to has exited, we need to clean up
1575 * and put the internal structures we have that belongs to that process.
1577 static void cfq_exit_io_context(struct io_context *ioc)
1579 call_for_each_cic(ioc, cfq_exit_single_io_context);
1582 static struct cfq_io_context *
1583 cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
1585 struct cfq_io_context *cic;
1587 cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO,
1590 cic->last_end_request = jiffies;
1591 INIT_LIST_HEAD(&cic->queue_list);
1592 INIT_HLIST_NODE(&cic->cic_list);
1593 cic->dtor = cfq_free_io_context;
1594 cic->exit = cfq_exit_io_context;
1595 elv_ioc_count_inc(cfq_ioc_count);
1601 static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc)
1603 struct task_struct *tsk = current;
1606 if (!cfq_cfqq_prio_changed(cfqq))
1609 ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio);
1610 switch (ioprio_class) {
1612 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
1613 case IOPRIO_CLASS_NONE:
1615 * no prio set, inherit CPU scheduling settings
1617 cfqq->ioprio = task_nice_ioprio(tsk);
1618 cfqq->ioprio_class = task_nice_ioclass(tsk);
1620 case IOPRIO_CLASS_RT:
1621 cfqq->ioprio = task_ioprio(ioc);
1622 cfqq->ioprio_class = IOPRIO_CLASS_RT;
1624 case IOPRIO_CLASS_BE:
1625 cfqq->ioprio = task_ioprio(ioc);
1626 cfqq->ioprio_class = IOPRIO_CLASS_BE;
1628 case IOPRIO_CLASS_IDLE:
1629 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
1631 cfq_clear_cfqq_idle_window(cfqq);
1636 * keep track of original prio settings in case we have to temporarily
1637 * elevate the priority of this queue
1639 cfqq->org_ioprio = cfqq->ioprio;
1640 cfqq->org_ioprio_class = cfqq->ioprio_class;
1641 cfq_clear_cfqq_prio_changed(cfqq);
1644 static void changed_ioprio(struct io_context *ioc, struct cfq_io_context *cic)
1646 struct cfq_data *cfqd = cic->key;
1647 struct cfq_queue *cfqq;
1648 unsigned long flags;
1650 if (unlikely(!cfqd))
1653 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1655 cfqq = cic->cfqq[BLK_RW_ASYNC];
1657 struct cfq_queue *new_cfqq;
1658 new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic->ioc,
1661 cic->cfqq[BLK_RW_ASYNC] = new_cfqq;
1662 cfq_put_queue(cfqq);
1666 cfqq = cic->cfqq[BLK_RW_SYNC];
1668 cfq_mark_cfqq_prio_changed(cfqq);
1670 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1673 static void cfq_ioc_set_ioprio(struct io_context *ioc)
1675 call_for_each_cic(ioc, changed_ioprio);
1676 ioc->ioprio_changed = 0;
1679 static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1680 pid_t pid, bool is_sync)
1682 RB_CLEAR_NODE(&cfqq->rb_node);
1683 RB_CLEAR_NODE(&cfqq->p_node);
1684 INIT_LIST_HEAD(&cfqq->fifo);
1686 atomic_set(&cfqq->ref, 0);
1689 cfq_mark_cfqq_prio_changed(cfqq);
1692 if (!cfq_class_idle(cfqq))
1693 cfq_mark_cfqq_idle_window(cfqq);
1694 cfq_mark_cfqq_sync(cfqq);
1699 static struct cfq_queue *
1700 cfq_find_alloc_queue(struct cfq_data *cfqd, bool is_sync,
1701 struct io_context *ioc, gfp_t gfp_mask)
1703 struct cfq_queue *cfqq, *new_cfqq = NULL;
1704 struct cfq_io_context *cic;
1707 cic = cfq_cic_lookup(cfqd, ioc);
1708 /* cic always exists here */
1709 cfqq = cic_to_cfqq(cic, is_sync);
1712 * Always try a new alloc if we fell back to the OOM cfqq
1713 * originally, since it should just be a temporary situation.
1715 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
1720 } else if (gfp_mask & __GFP_WAIT) {
1721 spin_unlock_irq(cfqd->queue->queue_lock);
1722 new_cfqq = kmem_cache_alloc_node(cfq_pool,
1723 gfp_mask | __GFP_ZERO,
1725 spin_lock_irq(cfqd->queue->queue_lock);
1729 cfqq = kmem_cache_alloc_node(cfq_pool,
1730 gfp_mask | __GFP_ZERO,
1735 cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
1736 cfq_init_prio_data(cfqq, ioc);
1737 cfq_log_cfqq(cfqd, cfqq, "alloced");
1739 cfqq = &cfqd->oom_cfqq;
1743 kmem_cache_free(cfq_pool, new_cfqq);
1748 static struct cfq_queue **
1749 cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
1751 switch (ioprio_class) {
1752 case IOPRIO_CLASS_RT:
1753 return &cfqd->async_cfqq[0][ioprio];
1754 case IOPRIO_CLASS_BE:
1755 return &cfqd->async_cfqq[1][ioprio];
1756 case IOPRIO_CLASS_IDLE:
1757 return &cfqd->async_idle_cfqq;
1763 static struct cfq_queue *
1764 cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct io_context *ioc,
1767 const int ioprio = task_ioprio(ioc);
1768 const int ioprio_class = task_ioprio_class(ioc);
1769 struct cfq_queue **async_cfqq = NULL;
1770 struct cfq_queue *cfqq = NULL;
1773 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
1778 cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask);
1781 * pin the queue now that it's allocated, scheduler exit will prune it
1783 if (!is_sync && !(*async_cfqq)) {
1784 atomic_inc(&cfqq->ref);
1788 atomic_inc(&cfqq->ref);
1793 * We drop cfq io contexts lazily, so we may find a dead one.
1796 cfq_drop_dead_cic(struct cfq_data *cfqd, struct io_context *ioc,
1797 struct cfq_io_context *cic)
1799 unsigned long flags;
1801 WARN_ON(!list_empty(&cic->queue_list));
1803 spin_lock_irqsave(&ioc->lock, flags);
1805 BUG_ON(ioc->ioc_data == cic);
1807 radix_tree_delete(&ioc->radix_root, (unsigned long) cfqd);
1808 hlist_del_rcu(&cic->cic_list);
1809 spin_unlock_irqrestore(&ioc->lock, flags);
1814 static struct cfq_io_context *
1815 cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc)
1817 struct cfq_io_context *cic;
1818 unsigned long flags;
1827 * we maintain a last-hit cache, to avoid browsing over the tree
1829 cic = rcu_dereference(ioc->ioc_data);
1830 if (cic && cic->key == cfqd) {
1836 cic = radix_tree_lookup(&ioc->radix_root, (unsigned long) cfqd);
1840 /* ->key must be copied to avoid race with cfq_exit_queue() */
1843 cfq_drop_dead_cic(cfqd, ioc, cic);
1848 spin_lock_irqsave(&ioc->lock, flags);
1849 rcu_assign_pointer(ioc->ioc_data, cic);
1850 spin_unlock_irqrestore(&ioc->lock, flags);
1858 * Add cic into ioc, using cfqd as the search key. This enables us to lookup
1859 * the process specific cfq io context when entered from the block layer.
1860 * Also adds the cic to a per-cfqd list, used when this queue is removed.
1862 static int cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
1863 struct cfq_io_context *cic, gfp_t gfp_mask)
1865 unsigned long flags;
1868 ret = radix_tree_preload(gfp_mask);
1873 spin_lock_irqsave(&ioc->lock, flags);
1874 ret = radix_tree_insert(&ioc->radix_root,
1875 (unsigned long) cfqd, cic);
1877 hlist_add_head_rcu(&cic->cic_list, &ioc->cic_list);
1878 spin_unlock_irqrestore(&ioc->lock, flags);
1880 radix_tree_preload_end();
1883 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1884 list_add(&cic->queue_list, &cfqd->cic_list);
1885 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1890 printk(KERN_ERR "cfq: cic link failed!\n");
1896 * Setup general io context and cfq io context. There can be several cfq
1897 * io contexts per general io context, if this process is doing io to more
1898 * than one device managed by cfq.
1900 static struct cfq_io_context *
1901 cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
1903 struct io_context *ioc = NULL;
1904 struct cfq_io_context *cic;
1906 might_sleep_if(gfp_mask & __GFP_WAIT);
1908 ioc = get_io_context(gfp_mask, cfqd->queue->node);
1912 cic = cfq_cic_lookup(cfqd, ioc);
1916 cic = cfq_alloc_io_context(cfqd, gfp_mask);
1920 if (cfq_cic_link(cfqd, ioc, cic, gfp_mask))
1924 smp_read_barrier_depends();
1925 if (unlikely(ioc->ioprio_changed))
1926 cfq_ioc_set_ioprio(ioc);
1932 put_io_context(ioc);
1937 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
1939 unsigned long elapsed = jiffies - cic->last_end_request;
1940 unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
1942 cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
1943 cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
1944 cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
1948 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1954 if (!cfqq->last_request_pos)
1956 else if (cfqq->last_request_pos < blk_rq_pos(rq))
1957 sdist = blk_rq_pos(rq) - cfqq->last_request_pos;
1959 sdist = cfqq->last_request_pos - blk_rq_pos(rq);
1962 * Don't allow the seek distance to get too large from the
1963 * odd fragment, pagein, etc
1965 if (cfqq->seek_samples <= 60) /* second&third seek */
1966 sdist = min(sdist, (cfqq->seek_mean * 4) + 2*1024*1024);
1968 sdist = min(sdist, (cfqq->seek_mean * 4) + 2*1024*64);
1970 cfqq->seek_samples = (7*cfqq->seek_samples + 256) / 8;
1971 cfqq->seek_total = (7*cfqq->seek_total + (u64)256*sdist) / 8;
1972 total = cfqq->seek_total + (cfqq->seek_samples/2);
1973 do_div(total, cfqq->seek_samples);
1974 cfqq->seek_mean = (sector_t)total;
1978 * Disable idle window if the process thinks too long or seeks so much that
1982 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1983 struct cfq_io_context *cic)
1985 int old_idle, enable_idle;
1988 * Don't idle for async or idle io prio class
1990 if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
1993 enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
1995 if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle ||
1996 (!cfqd->cfq_latency && cfqd->hw_tag && CFQQ_SEEKY(cfqq)))
1998 else if (sample_valid(cic->ttime_samples)) {
1999 unsigned int slice_idle = cfqd->cfq_slice_idle;
2000 if (sample_valid(cfqq->seek_samples) && CFQQ_SEEKY(cfqq))
2001 slice_idle = msecs_to_jiffies(CFQ_MIN_TT);
2002 if (cic->ttime_mean > slice_idle)
2008 if (old_idle != enable_idle) {
2009 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
2011 cfq_mark_cfqq_idle_window(cfqq);
2013 cfq_clear_cfqq_idle_window(cfqq);
2018 * Check if new_cfqq should preempt the currently active queue. Return 0 for
2019 * no or if we aren't sure, a 1 will cause a preempt.
2022 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
2025 struct cfq_queue *cfqq;
2027 cfqq = cfqd->active_queue;
2031 if (cfq_slice_used(cfqq))
2034 if (cfq_class_idle(new_cfqq))
2037 if (cfq_class_idle(cfqq))
2041 * if the new request is sync, but the currently running queue is
2042 * not, let the sync request have priority.
2044 if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
2048 * So both queues are sync. Let the new request get disk time if
2049 * it's a metadata request and the current queue is doing regular IO.
2051 if (rq_is_meta(rq) && !cfqq->meta_pending)
2055 * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
2057 if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
2060 if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
2064 * if this request is as-good as one we would expect from the
2065 * current cfqq, let it preempt
2067 if (cfq_rq_close(cfqd, cfqq, rq))
2074 * cfqq preempts the active queue. if we allowed preempt with no slice left,
2075 * let it have half of its nominal slice.
2077 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2079 cfq_log_cfqq(cfqd, cfqq, "preempt");
2080 cfq_slice_expired(cfqd, 1);
2083 * Put the new queue at the front of the of the current list,
2084 * so we know that it will be selected next.
2086 BUG_ON(!cfq_cfqq_on_rr(cfqq));
2088 cfq_service_tree_add(cfqd, cfqq, 1);
2090 cfqq->slice_end = 0;
2091 cfq_mark_cfqq_slice_new(cfqq);
2095 * Called when a new fs request (rq) is added (to cfqq). Check if there's
2096 * something we should do about it
2099 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2102 struct cfq_io_context *cic = RQ_CIC(rq);
2106 cfqq->meta_pending++;
2108 cfq_update_io_thinktime(cfqd, cic);
2109 cfq_update_io_seektime(cfqd, cfqq, rq);
2110 cfq_update_idle_window(cfqd, cfqq, cic);
2112 cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
2114 if (cfqq == cfqd->active_queue) {
2116 * Remember that we saw a request from this process, but
2117 * don't start queuing just yet. Otherwise we risk seeing lots
2118 * of tiny requests, because we disrupt the normal plugging
2119 * and merging. If the request is already larger than a single
2120 * page, let it rip immediately. For that case we assume that
2121 * merging is already done. Ditto for a busy system that
2122 * has other work pending, don't risk delaying until the
2123 * idle timer unplug to continue working.
2125 if (cfq_cfqq_wait_request(cfqq)) {
2126 if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
2127 cfqd->busy_queues > 1) {
2128 del_timer(&cfqd->idle_slice_timer);
2129 __blk_run_queue(cfqd->queue);
2131 cfq_mark_cfqq_must_dispatch(cfqq);
2133 } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
2135 * not the active queue - expire current slice if it is
2136 * idle and has expired it's mean thinktime or this new queue
2137 * has some old slice time left and is of higher priority or
2138 * this new queue is RT and the current one is BE
2140 cfq_preempt_queue(cfqd, cfqq);
2141 __blk_run_queue(cfqd->queue);
2145 static void cfq_insert_request(struct request_queue *q, struct request *rq)
2147 struct cfq_data *cfqd = q->elevator->elevator_data;
2148 struct cfq_queue *cfqq = RQ_CFQQ(rq);
2150 cfq_log_cfqq(cfqd, cfqq, "insert_request");
2151 cfq_init_prio_data(cfqq, RQ_CIC(rq)->ioc);
2155 rq_set_fifo_time(rq, jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)]);
2156 list_add_tail(&rq->queuelist, &cfqq->fifo);
2158 cfq_rq_enqueued(cfqd, cfqq, rq);
2162 * Update hw_tag based on peak queue depth over 50 samples under
2165 static void cfq_update_hw_tag(struct cfq_data *cfqd)
2167 if (rq_in_driver(cfqd) > cfqd->rq_in_driver_peak)
2168 cfqd->rq_in_driver_peak = rq_in_driver(cfqd);
2170 if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
2171 rq_in_driver(cfqd) <= CFQ_HW_QUEUE_MIN)
2174 if (cfqd->hw_tag_samples++ < 50)
2177 if (cfqd->rq_in_driver_peak >= CFQ_HW_QUEUE_MIN)
2182 cfqd->hw_tag_samples = 0;
2183 cfqd->rq_in_driver_peak = 0;
2186 static void cfq_completed_request(struct request_queue *q, struct request *rq)
2188 struct cfq_queue *cfqq = RQ_CFQQ(rq);
2189 struct cfq_data *cfqd = cfqq->cfqd;
2190 const int sync = rq_is_sync(rq);
2194 cfq_log_cfqq(cfqd, cfqq, "complete");
2196 cfq_update_hw_tag(cfqd);
2198 WARN_ON(!cfqd->rq_in_driver[sync]);
2199 WARN_ON(!cfqq->dispatched);
2200 cfqd->rq_in_driver[sync]--;
2203 if (cfq_cfqq_sync(cfqq))
2204 cfqd->sync_flight--;
2207 RQ_CIC(rq)->last_end_request = now;
2208 cfqd->last_end_sync_rq = now;
2212 * If this is the active queue, check if it needs to be expired,
2213 * or if we want to idle in case it has no pending requests.
2215 if (cfqd->active_queue == cfqq) {
2216 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
2218 if (cfq_cfqq_slice_new(cfqq)) {
2219 cfq_set_prio_slice(cfqd, cfqq);
2220 cfq_clear_cfqq_slice_new(cfqq);
2223 * If there are no requests waiting in this queue, and
2224 * there are other queues ready to issue requests, AND
2225 * those other queues are issuing requests within our
2226 * mean seek distance, give them a chance to run instead
2229 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
2230 cfq_slice_expired(cfqd, 1);
2231 else if (cfqq_empty && !cfq_close_cooperator(cfqd, cfqq, 1) &&
2232 sync && !rq_noidle(rq))
2233 cfq_arm_slice_timer(cfqd);
2236 if (!rq_in_driver(cfqd))
2237 cfq_schedule_dispatch(cfqd);
2241 * we temporarily boost lower priority queues if they are holding fs exclusive
2242 * resources. they are boosted to normal prio (CLASS_BE/4)
2244 static void cfq_prio_boost(struct cfq_queue *cfqq)
2246 if (has_fs_excl()) {
2248 * boost idle prio on transactions that would lock out other
2249 * users of the filesystem
2251 if (cfq_class_idle(cfqq))
2252 cfqq->ioprio_class = IOPRIO_CLASS_BE;
2253 if (cfqq->ioprio > IOPRIO_NORM)
2254 cfqq->ioprio = IOPRIO_NORM;
2257 * check if we need to unboost the queue
2259 if (cfqq->ioprio_class != cfqq->org_ioprio_class)
2260 cfqq->ioprio_class = cfqq->org_ioprio_class;
2261 if (cfqq->ioprio != cfqq->org_ioprio)
2262 cfqq->ioprio = cfqq->org_ioprio;
2266 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
2268 if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
2269 cfq_mark_cfqq_must_alloc_slice(cfqq);
2270 return ELV_MQUEUE_MUST;
2273 return ELV_MQUEUE_MAY;
2276 static int cfq_may_queue(struct request_queue *q, int rw)
2278 struct cfq_data *cfqd = q->elevator->elevator_data;
2279 struct task_struct *tsk = current;
2280 struct cfq_io_context *cic;
2281 struct cfq_queue *cfqq;
2284 * don't force setup of a queue from here, as a call to may_queue
2285 * does not necessarily imply that a request actually will be queued.
2286 * so just lookup a possibly existing queue, or return 'may queue'
2289 cic = cfq_cic_lookup(cfqd, tsk->io_context);
2291 return ELV_MQUEUE_MAY;
2293 cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
2295 cfq_init_prio_data(cfqq, cic->ioc);
2296 cfq_prio_boost(cfqq);
2298 return __cfq_may_queue(cfqq);
2301 return ELV_MQUEUE_MAY;
2305 * queue lock held here
2307 static void cfq_put_request(struct request *rq)
2309 struct cfq_queue *cfqq = RQ_CFQQ(rq);
2312 const int rw = rq_data_dir(rq);
2314 BUG_ON(!cfqq->allocated[rw]);
2315 cfqq->allocated[rw]--;
2317 put_io_context(RQ_CIC(rq)->ioc);
2319 rq->elevator_private = NULL;
2320 rq->elevator_private2 = NULL;
2322 cfq_put_queue(cfqq);
2327 * Allocate cfq data structures associated with this request.
2330 cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
2332 struct cfq_data *cfqd = q->elevator->elevator_data;
2333 struct cfq_io_context *cic;
2334 const int rw = rq_data_dir(rq);
2335 const bool is_sync = rq_is_sync(rq);
2336 struct cfq_queue *cfqq;
2337 unsigned long flags;
2339 might_sleep_if(gfp_mask & __GFP_WAIT);
2341 cic = cfq_get_io_context(cfqd, gfp_mask);
2343 spin_lock_irqsave(q->queue_lock, flags);
2348 cfqq = cic_to_cfqq(cic, is_sync);
2349 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
2350 cfqq = cfq_get_queue(cfqd, is_sync, cic->ioc, gfp_mask);
2351 cic_set_cfqq(cic, cfqq, is_sync);
2354 cfqq->allocated[rw]++;
2355 atomic_inc(&cfqq->ref);
2357 spin_unlock_irqrestore(q->queue_lock, flags);
2359 rq->elevator_private = cic;
2360 rq->elevator_private2 = cfqq;
2365 put_io_context(cic->ioc);
2367 cfq_schedule_dispatch(cfqd);
2368 spin_unlock_irqrestore(q->queue_lock, flags);
2369 cfq_log(cfqd, "set_request fail");
2373 static void cfq_kick_queue(struct work_struct *work)
2375 struct cfq_data *cfqd =
2376 container_of(work, struct cfq_data, unplug_work);
2377 struct request_queue *q = cfqd->queue;
2379 spin_lock_irq(q->queue_lock);
2380 __blk_run_queue(cfqd->queue);
2381 spin_unlock_irq(q->queue_lock);
2385 * Timer running if the active_queue is currently idling inside its time slice
2387 static void cfq_idle_slice_timer(unsigned long data)
2389 struct cfq_data *cfqd = (struct cfq_data *) data;
2390 struct cfq_queue *cfqq;
2391 unsigned long flags;
2394 cfq_log(cfqd, "idle timer fired");
2396 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2398 cfqq = cfqd->active_queue;
2403 * We saw a request before the queue expired, let it through
2405 if (cfq_cfqq_must_dispatch(cfqq))
2411 if (cfq_slice_used(cfqq))
2415 * only expire and reinvoke request handler, if there are
2416 * other queues with pending requests
2418 if (!cfqd->busy_queues)
2422 * not expired and it has a request pending, let it dispatch
2424 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
2428 cfq_slice_expired(cfqd, timed_out);
2430 cfq_schedule_dispatch(cfqd);
2432 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2435 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
2437 del_timer_sync(&cfqd->idle_slice_timer);
2438 cancel_work_sync(&cfqd->unplug_work);
2441 static void cfq_put_async_queues(struct cfq_data *cfqd)
2445 for (i = 0; i < IOPRIO_BE_NR; i++) {
2446 if (cfqd->async_cfqq[0][i])
2447 cfq_put_queue(cfqd->async_cfqq[0][i]);
2448 if (cfqd->async_cfqq[1][i])
2449 cfq_put_queue(cfqd->async_cfqq[1][i]);
2452 if (cfqd->async_idle_cfqq)
2453 cfq_put_queue(cfqd->async_idle_cfqq);
2456 static void cfq_exit_queue(struct elevator_queue *e)
2458 struct cfq_data *cfqd = e->elevator_data;
2459 struct request_queue *q = cfqd->queue;
2461 cfq_shutdown_timer_wq(cfqd);
2463 spin_lock_irq(q->queue_lock);
2465 if (cfqd->active_queue)
2466 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
2468 while (!list_empty(&cfqd->cic_list)) {
2469 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
2470 struct cfq_io_context,
2473 __cfq_exit_single_io_context(cfqd, cic);
2476 cfq_put_async_queues(cfqd);
2478 spin_unlock_irq(q->queue_lock);
2480 cfq_shutdown_timer_wq(cfqd);
2485 static void *cfq_init_queue(struct request_queue *q)
2487 struct cfq_data *cfqd;
2490 cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
2494 cfqd->service_tree = CFQ_RB_ROOT;
2497 * Not strictly needed (since RB_ROOT just clears the node and we
2498 * zeroed cfqd on alloc), but better be safe in case someone decides
2499 * to add magic to the rb code
2501 for (i = 0; i < CFQ_PRIO_LISTS; i++)
2502 cfqd->prio_trees[i] = RB_ROOT;
2505 * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
2506 * Grab a permanent reference to it, so that the normal code flow
2507 * will not attempt to free it.
2509 cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
2510 atomic_inc(&cfqd->oom_cfqq.ref);
2512 INIT_LIST_HEAD(&cfqd->cic_list);
2516 init_timer(&cfqd->idle_slice_timer);
2517 cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
2518 cfqd->idle_slice_timer.data = (unsigned long) cfqd;
2520 INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
2522 cfqd->cfq_quantum = cfq_quantum;
2523 cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
2524 cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
2525 cfqd->cfq_back_max = cfq_back_max;
2526 cfqd->cfq_back_penalty = cfq_back_penalty;
2527 cfqd->cfq_slice[0] = cfq_slice_async;
2528 cfqd->cfq_slice[1] = cfq_slice_sync;
2529 cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
2530 cfqd->cfq_slice_idle = cfq_slice_idle;
2531 cfqd->cfq_latency = 1;
2533 cfqd->last_end_sync_rq = jiffies;
2537 static void cfq_slab_kill(void)
2540 * Caller already ensured that pending RCU callbacks are completed,
2541 * so we should have no busy allocations at this point.
2544 kmem_cache_destroy(cfq_pool);
2546 kmem_cache_destroy(cfq_ioc_pool);
2549 static int __init cfq_slab_setup(void)
2551 cfq_pool = KMEM_CACHE(cfq_queue, 0);
2555 cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0);
2566 * sysfs parts below -->
2569 cfq_var_show(unsigned int var, char *page)
2571 return sprintf(page, "%d\n", var);
2575 cfq_var_store(unsigned int *var, const char *page, size_t count)
2577 char *p = (char *) page;
2579 *var = simple_strtoul(p, &p, 10);
2583 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
2584 static ssize_t __FUNC(struct elevator_queue *e, char *page) \
2586 struct cfq_data *cfqd = e->elevator_data; \
2587 unsigned int __data = __VAR; \
2589 __data = jiffies_to_msecs(__data); \
2590 return cfq_var_show(__data, (page)); \
2592 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
2593 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
2594 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
2595 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
2596 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
2597 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
2598 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
2599 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
2600 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
2601 SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
2602 #undef SHOW_FUNCTION
2604 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
2605 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
2607 struct cfq_data *cfqd = e->elevator_data; \
2608 unsigned int __data; \
2609 int ret = cfq_var_store(&__data, (page), count); \
2610 if (__data < (MIN)) \
2612 else if (__data > (MAX)) \
2615 *(__PTR) = msecs_to_jiffies(__data); \
2617 *(__PTR) = __data; \
2620 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
2621 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
2623 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
2625 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
2626 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
2628 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
2629 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
2630 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
2631 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
2633 STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
2634 #undef STORE_FUNCTION
2636 #define CFQ_ATTR(name) \
2637 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
2639 static struct elv_fs_entry cfq_attrs[] = {
2641 CFQ_ATTR(fifo_expire_sync),
2642 CFQ_ATTR(fifo_expire_async),
2643 CFQ_ATTR(back_seek_max),
2644 CFQ_ATTR(back_seek_penalty),
2645 CFQ_ATTR(slice_sync),
2646 CFQ_ATTR(slice_async),
2647 CFQ_ATTR(slice_async_rq),
2648 CFQ_ATTR(slice_idle),
2649 CFQ_ATTR(low_latency),
2653 static struct elevator_type iosched_cfq = {
2655 .elevator_merge_fn = cfq_merge,
2656 .elevator_merged_fn = cfq_merged_request,
2657 .elevator_merge_req_fn = cfq_merged_requests,
2658 .elevator_allow_merge_fn = cfq_allow_merge,
2659 .elevator_dispatch_fn = cfq_dispatch_requests,
2660 .elevator_add_req_fn = cfq_insert_request,
2661 .elevator_activate_req_fn = cfq_activate_request,
2662 .elevator_deactivate_req_fn = cfq_deactivate_request,
2663 .elevator_queue_empty_fn = cfq_queue_empty,
2664 .elevator_completed_req_fn = cfq_completed_request,
2665 .elevator_former_req_fn = elv_rb_former_request,
2666 .elevator_latter_req_fn = elv_rb_latter_request,
2667 .elevator_set_req_fn = cfq_set_request,
2668 .elevator_put_req_fn = cfq_put_request,
2669 .elevator_may_queue_fn = cfq_may_queue,
2670 .elevator_init_fn = cfq_init_queue,
2671 .elevator_exit_fn = cfq_exit_queue,
2672 .trim = cfq_free_io_context,
2674 .elevator_attrs = cfq_attrs,
2675 .elevator_name = "cfq",
2676 .elevator_owner = THIS_MODULE,
2679 static int __init cfq_init(void)
2682 * could be 0 on HZ < 1000 setups
2684 if (!cfq_slice_async)
2685 cfq_slice_async = 1;
2686 if (!cfq_slice_idle)
2689 if (cfq_slab_setup())
2692 elv_register(&iosched_cfq);
2697 static void __exit cfq_exit(void)
2699 DECLARE_COMPLETION_ONSTACK(all_gone);
2700 elv_unregister(&iosched_cfq);
2701 ioc_gone = &all_gone;
2702 /* ioc_gone's update must be visible before reading ioc_count */
2706 * this also protects us from entering cfq_slab_kill() with
2707 * pending RCU callbacks
2709 if (elv_ioc_count_read(cfq_ioc_count))
2710 wait_for_completion(&all_gone);
2714 module_init(cfq_init);
2715 module_exit(cfq_exit);
2717 MODULE_AUTHOR("Jens Axboe");
2718 MODULE_LICENSE("GPL");
2719 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");