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;
119 * Per block device queue structure
122 struct request_queue *queue;
125 * rr list of queues with requests and the count of them
127 struct cfq_rb_root service_tree;
130 * Each priority tree is sorted by next_request position. These
131 * trees are used when determining if two or more queues are
132 * interleaving requests (see cfq_close_cooperator).
134 struct rb_root prio_trees[CFQ_PRIO_LISTS];
136 unsigned int busy_queues;
142 * queue-depth detection
147 int rq_in_driver_peak;
150 * idle window management
152 struct timer_list idle_slice_timer;
153 struct work_struct unplug_work;
155 struct cfq_queue *active_queue;
156 struct cfq_io_context *active_cic;
159 * async queue for each priority case
161 struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
162 struct cfq_queue *async_idle_cfqq;
164 sector_t last_position;
167 * tunables, see top of file
169 unsigned int cfq_quantum;
170 unsigned int cfq_fifo_expire[2];
171 unsigned int cfq_back_penalty;
172 unsigned int cfq_back_max;
173 unsigned int cfq_slice[2];
174 unsigned int cfq_slice_async_rq;
175 unsigned int cfq_slice_idle;
176 unsigned int cfq_latency;
178 struct list_head cic_list;
181 * Fallback dummy cfqq for extreme OOM conditions
183 struct cfq_queue oom_cfqq;
185 unsigned long last_end_sync_rq;
188 enum cfqq_state_flags {
189 CFQ_CFQQ_FLAG_on_rr = 0, /* on round-robin busy list */
190 CFQ_CFQQ_FLAG_wait_request, /* waiting for a request */
191 CFQ_CFQQ_FLAG_must_dispatch, /* must be allowed a dispatch */
192 CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
193 CFQ_CFQQ_FLAG_fifo_expire, /* FIFO checked in this slice */
194 CFQ_CFQQ_FLAG_idle_window, /* slice idling enabled */
195 CFQ_CFQQ_FLAG_prio_changed, /* task priority has changed */
196 CFQ_CFQQ_FLAG_slice_new, /* no requests dispatched in slice */
197 CFQ_CFQQ_FLAG_sync, /* synchronous queue */
198 CFQ_CFQQ_FLAG_coop, /* has done a coop jump of the queue */
201 #define CFQ_CFQQ_FNS(name) \
202 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \
204 (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name); \
206 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \
208 (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \
210 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \
212 return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \
216 CFQ_CFQQ_FNS(wait_request);
217 CFQ_CFQQ_FNS(must_dispatch);
218 CFQ_CFQQ_FNS(must_alloc_slice);
219 CFQ_CFQQ_FNS(fifo_expire);
220 CFQ_CFQQ_FNS(idle_window);
221 CFQ_CFQQ_FNS(prio_changed);
222 CFQ_CFQQ_FNS(slice_new);
227 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
228 blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args)
229 #define cfq_log(cfqd, fmt, args...) \
230 blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
232 static void cfq_dispatch_insert(struct request_queue *, struct request *);
233 static struct cfq_queue *cfq_get_queue(struct cfq_data *, int,
234 struct io_context *, gfp_t);
235 static struct cfq_io_context *cfq_cic_lookup(struct cfq_data *,
236 struct io_context *);
238 static inline int rq_in_driver(struct cfq_data *cfqd)
240 return cfqd->rq_in_driver[0] + cfqd->rq_in_driver[1];
243 static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_context *cic,
246 return cic->cfqq[!!is_sync];
249 static inline void cic_set_cfqq(struct cfq_io_context *cic,
250 struct cfq_queue *cfqq, int is_sync)
252 cic->cfqq[!!is_sync] = cfqq;
256 * We regard a request as SYNC, if it's either a read or has the SYNC bit
257 * set (in which case it could also be direct WRITE).
259 static inline int cfq_bio_sync(struct bio *bio)
261 if (bio_data_dir(bio) == READ || bio_rw_flagged(bio, BIO_RW_SYNCIO))
268 * scheduler run of queue, if there are requests pending and no one in the
269 * driver that will restart queueing
271 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
273 if (cfqd->busy_queues) {
274 cfq_log(cfqd, "schedule dispatch");
275 kblockd_schedule_work(cfqd->queue, &cfqd->unplug_work);
279 static int cfq_queue_empty(struct request_queue *q)
281 struct cfq_data *cfqd = q->elevator->elevator_data;
283 return !cfqd->busy_queues;
287 * Scale schedule slice based on io priority. Use the sync time slice only
288 * if a queue is marked sync and has sync io queued. A sync queue with async
289 * io only, should not get full sync slice length.
291 static inline int cfq_prio_slice(struct cfq_data *cfqd, int sync,
294 const int base_slice = cfqd->cfq_slice[sync];
296 WARN_ON(prio >= IOPRIO_BE_NR);
298 return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
302 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
304 return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
308 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
310 cfqq->slice_end = cfq_prio_to_slice(cfqd, cfqq) + jiffies;
311 cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
315 * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
316 * isn't valid until the first request from the dispatch is activated
317 * and the slice time set.
319 static inline int cfq_slice_used(struct cfq_queue *cfqq)
321 if (cfq_cfqq_slice_new(cfqq))
323 if (time_before(jiffies, cfqq->slice_end))
330 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
331 * We choose the request that is closest to the head right now. Distance
332 * behind the head is penalized and only allowed to a certain extent.
334 static struct request *
335 cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2)
337 sector_t last, s1, s2, d1 = 0, d2 = 0;
338 unsigned long back_max;
339 #define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */
340 #define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */
341 unsigned wrap = 0; /* bit mask: requests behind the disk head? */
343 if (rq1 == NULL || rq1 == rq2)
348 if (rq_is_sync(rq1) && !rq_is_sync(rq2))
350 else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
352 if (rq_is_meta(rq1) && !rq_is_meta(rq2))
354 else if (rq_is_meta(rq2) && !rq_is_meta(rq1))
357 s1 = blk_rq_pos(rq1);
358 s2 = blk_rq_pos(rq2);
360 last = cfqd->last_position;
363 * by definition, 1KiB is 2 sectors
365 back_max = cfqd->cfq_back_max * 2;
368 * Strict one way elevator _except_ in the case where we allow
369 * short backward seeks which are biased as twice the cost of a
370 * similar forward seek.
374 else if (s1 + back_max >= last)
375 d1 = (last - s1) * cfqd->cfq_back_penalty;
377 wrap |= CFQ_RQ1_WRAP;
381 else if (s2 + back_max >= last)
382 d2 = (last - s2) * cfqd->cfq_back_penalty;
384 wrap |= CFQ_RQ2_WRAP;
386 /* Found required data */
389 * By doing switch() on the bit mask "wrap" we avoid having to
390 * check two variables for all permutations: --> faster!
393 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
409 case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
412 * Since both rqs are wrapped,
413 * start with the one that's further behind head
414 * (--> only *one* back seek required),
415 * since back seek takes more time than forward.
425 * The below is leftmost cache rbtree addon
427 static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
430 root->left = rb_first(&root->rb);
433 return rb_entry(root->left, struct cfq_queue, rb_node);
438 static void rb_erase_init(struct rb_node *n, struct rb_root *root)
444 static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
448 rb_erase_init(n, &root->rb);
452 * would be nice to take fifo expire time into account as well
454 static struct request *
455 cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
456 struct request *last)
458 struct rb_node *rbnext = rb_next(&last->rb_node);
459 struct rb_node *rbprev = rb_prev(&last->rb_node);
460 struct request *next = NULL, *prev = NULL;
462 BUG_ON(RB_EMPTY_NODE(&last->rb_node));
465 prev = rb_entry_rq(rbprev);
468 next = rb_entry_rq(rbnext);
470 rbnext = rb_first(&cfqq->sort_list);
471 if (rbnext && rbnext != &last->rb_node)
472 next = rb_entry_rq(rbnext);
475 return cfq_choose_req(cfqd, next, prev);
478 static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
479 struct cfq_queue *cfqq)
482 * just an approximation, should be ok.
484 return (cfqd->busy_queues - 1) * (cfq_prio_slice(cfqd, 1, 0) -
485 cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
489 * The cfqd->service_tree holds all pending cfq_queue's that have
490 * requests waiting to be processed. It is sorted in the order that
491 * we will service the queues.
493 static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
496 struct rb_node **p, *parent;
497 struct cfq_queue *__cfqq;
498 unsigned long rb_key;
501 if (cfq_class_idle(cfqq)) {
502 rb_key = CFQ_IDLE_DELAY;
503 parent = rb_last(&cfqd->service_tree.rb);
504 if (parent && parent != &cfqq->rb_node) {
505 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
506 rb_key += __cfqq->rb_key;
509 } else if (!add_front) {
511 * Get our rb key offset. Subtract any residual slice
512 * value carried from last service. A negative resid
513 * count indicates slice overrun, and this should position
514 * the next service time further away in the tree.
516 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
517 rb_key -= cfqq->slice_resid;
518 cfqq->slice_resid = 0;
521 __cfqq = cfq_rb_first(&cfqd->service_tree);
522 rb_key += __cfqq ? __cfqq->rb_key : jiffies;
525 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
527 * same position, nothing more to do
529 if (rb_key == cfqq->rb_key)
532 cfq_rb_erase(&cfqq->rb_node, &cfqd->service_tree);
537 p = &cfqd->service_tree.rb.rb_node;
542 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
545 * sort RT queues first, we always want to give
546 * preference to them. IDLE queues goes to the back.
547 * after that, sort on the next service time.
549 if (cfq_class_rt(cfqq) > cfq_class_rt(__cfqq))
551 else if (cfq_class_rt(cfqq) < cfq_class_rt(__cfqq))
553 else if (cfq_class_idle(cfqq) < cfq_class_idle(__cfqq))
555 else if (cfq_class_idle(cfqq) > cfq_class_idle(__cfqq))
557 else if (time_before(rb_key, __cfqq->rb_key))
562 if (n == &(*p)->rb_right)
569 cfqd->service_tree.left = &cfqq->rb_node;
571 cfqq->rb_key = rb_key;
572 rb_link_node(&cfqq->rb_node, parent, p);
573 rb_insert_color(&cfqq->rb_node, &cfqd->service_tree.rb);
576 static struct cfq_queue *
577 cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
578 sector_t sector, struct rb_node **ret_parent,
579 struct rb_node ***rb_link)
581 struct rb_node **p, *parent;
582 struct cfq_queue *cfqq = NULL;
590 cfqq = rb_entry(parent, struct cfq_queue, p_node);
593 * Sort strictly based on sector. Smallest to the left,
594 * largest to the right.
596 if (sector > blk_rq_pos(cfqq->next_rq))
598 else if (sector < blk_rq_pos(cfqq->next_rq))
606 *ret_parent = parent;
612 static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
614 struct rb_node **p, *parent;
615 struct cfq_queue *__cfqq;
618 rb_erase(&cfqq->p_node, cfqq->p_root);
622 if (cfq_class_idle(cfqq))
627 cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
628 __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
629 blk_rq_pos(cfqq->next_rq), &parent, &p);
631 rb_link_node(&cfqq->p_node, parent, p);
632 rb_insert_color(&cfqq->p_node, cfqq->p_root);
638 * Update cfqq's position in the service tree.
640 static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
643 * Resorting requires the cfqq to be on the RR list already.
645 if (cfq_cfqq_on_rr(cfqq)) {
646 cfq_service_tree_add(cfqd, cfqq, 0);
647 cfq_prio_tree_add(cfqd, cfqq);
652 * add to busy list of queues for service, trying to be fair in ordering
653 * the pending list according to last request service
655 static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
657 cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
658 BUG_ON(cfq_cfqq_on_rr(cfqq));
659 cfq_mark_cfqq_on_rr(cfqq);
662 cfq_resort_rr_list(cfqd, cfqq);
666 * Called when the cfqq no longer has requests pending, remove it from
669 static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
671 cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
672 BUG_ON(!cfq_cfqq_on_rr(cfqq));
673 cfq_clear_cfqq_on_rr(cfqq);
675 if (!RB_EMPTY_NODE(&cfqq->rb_node))
676 cfq_rb_erase(&cfqq->rb_node, &cfqd->service_tree);
678 rb_erase(&cfqq->p_node, cfqq->p_root);
682 BUG_ON(!cfqd->busy_queues);
687 * rb tree support functions
689 static void cfq_del_rq_rb(struct request *rq)
691 struct cfq_queue *cfqq = RQ_CFQQ(rq);
692 struct cfq_data *cfqd = cfqq->cfqd;
693 const int sync = rq_is_sync(rq);
695 BUG_ON(!cfqq->queued[sync]);
696 cfqq->queued[sync]--;
698 elv_rb_del(&cfqq->sort_list, rq);
700 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
701 cfq_del_cfqq_rr(cfqd, cfqq);
704 static void cfq_add_rq_rb(struct request *rq)
706 struct cfq_queue *cfqq = RQ_CFQQ(rq);
707 struct cfq_data *cfqd = cfqq->cfqd;
708 struct request *__alias, *prev;
710 cfqq->queued[rq_is_sync(rq)]++;
713 * looks a little odd, but the first insert might return an alias.
714 * if that happens, put the alias on the dispatch list
716 while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL)
717 cfq_dispatch_insert(cfqd->queue, __alias);
719 if (!cfq_cfqq_on_rr(cfqq))
720 cfq_add_cfqq_rr(cfqd, cfqq);
723 * check if this request is a better next-serve candidate
725 prev = cfqq->next_rq;
726 cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq);
729 * adjust priority tree position, if ->next_rq changes
731 if (prev != cfqq->next_rq)
732 cfq_prio_tree_add(cfqd, cfqq);
734 BUG_ON(!cfqq->next_rq);
737 static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
739 elv_rb_del(&cfqq->sort_list, rq);
740 cfqq->queued[rq_is_sync(rq)]--;
744 static struct request *
745 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
747 struct task_struct *tsk = current;
748 struct cfq_io_context *cic;
749 struct cfq_queue *cfqq;
751 cic = cfq_cic_lookup(cfqd, tsk->io_context);
755 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
757 sector_t sector = bio->bi_sector + bio_sectors(bio);
759 return elv_rb_find(&cfqq->sort_list, sector);
765 static void cfq_activate_request(struct request_queue *q, struct request *rq)
767 struct cfq_data *cfqd = q->elevator->elevator_data;
769 cfqd->rq_in_driver[rq_is_sync(rq)]++;
770 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
773 cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
776 static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
778 struct cfq_data *cfqd = q->elevator->elevator_data;
779 const int sync = rq_is_sync(rq);
781 WARN_ON(!cfqd->rq_in_driver[sync]);
782 cfqd->rq_in_driver[sync]--;
783 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
787 static void cfq_remove_request(struct request *rq)
789 struct cfq_queue *cfqq = RQ_CFQQ(rq);
791 if (cfqq->next_rq == rq)
792 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
794 list_del_init(&rq->queuelist);
797 cfqq->cfqd->rq_queued--;
798 if (rq_is_meta(rq)) {
799 WARN_ON(!cfqq->meta_pending);
800 cfqq->meta_pending--;
804 static int cfq_merge(struct request_queue *q, struct request **req,
807 struct cfq_data *cfqd = q->elevator->elevator_data;
808 struct request *__rq;
810 __rq = cfq_find_rq_fmerge(cfqd, bio);
811 if (__rq && elv_rq_merge_ok(__rq, bio)) {
813 return ELEVATOR_FRONT_MERGE;
816 return ELEVATOR_NO_MERGE;
819 static void cfq_merged_request(struct request_queue *q, struct request *req,
822 if (type == ELEVATOR_FRONT_MERGE) {
823 struct cfq_queue *cfqq = RQ_CFQQ(req);
825 cfq_reposition_rq_rb(cfqq, req);
830 cfq_merged_requests(struct request_queue *q, struct request *rq,
831 struct request *next)
834 * reposition in fifo if next is older than rq
836 if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
837 time_before(rq_fifo_time(next), rq_fifo_time(rq))) {
838 list_move(&rq->queuelist, &next->queuelist);
839 rq_set_fifo_time(rq, rq_fifo_time(next));
842 cfq_remove_request(next);
845 static int cfq_allow_merge(struct request_queue *q, struct request *rq,
848 struct cfq_data *cfqd = q->elevator->elevator_data;
849 struct cfq_io_context *cic;
850 struct cfq_queue *cfqq;
853 * Disallow merge of a sync bio into an async request.
855 if (cfq_bio_sync(bio) && !rq_is_sync(rq))
859 * Lookup the cfqq that this bio will be queued with. Allow
860 * merge only if rq is queued there.
862 cic = cfq_cic_lookup(cfqd, current->io_context);
866 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
867 if (cfqq == RQ_CFQQ(rq))
873 static void __cfq_set_active_queue(struct cfq_data *cfqd,
874 struct cfq_queue *cfqq)
877 cfq_log_cfqq(cfqd, cfqq, "set_active");
879 cfqq->slice_dispatch = 0;
881 cfq_clear_cfqq_wait_request(cfqq);
882 cfq_clear_cfqq_must_dispatch(cfqq);
883 cfq_clear_cfqq_must_alloc_slice(cfqq);
884 cfq_clear_cfqq_fifo_expire(cfqq);
885 cfq_mark_cfqq_slice_new(cfqq);
887 del_timer(&cfqd->idle_slice_timer);
890 cfqd->active_queue = cfqq;
894 * current cfqq expired its slice (or was too idle), select new one
897 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
900 cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
902 if (cfq_cfqq_wait_request(cfqq))
903 del_timer(&cfqd->idle_slice_timer);
905 cfq_clear_cfqq_wait_request(cfqq);
908 * store what was left of this slice, if the queue idled/timed out
910 if (timed_out && !cfq_cfqq_slice_new(cfqq)) {
911 cfqq->slice_resid = cfqq->slice_end - jiffies;
912 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
915 cfq_resort_rr_list(cfqd, cfqq);
917 if (cfqq == cfqd->active_queue)
918 cfqd->active_queue = NULL;
920 if (cfqd->active_cic) {
921 put_io_context(cfqd->active_cic->ioc);
922 cfqd->active_cic = NULL;
926 static inline void cfq_slice_expired(struct cfq_data *cfqd, int timed_out)
928 struct cfq_queue *cfqq = cfqd->active_queue;
931 __cfq_slice_expired(cfqd, cfqq, timed_out);
935 * Get next queue for service. Unless we have a queue preemption,
936 * we'll simply select the first cfqq in the service tree.
938 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
940 if (RB_EMPTY_ROOT(&cfqd->service_tree.rb))
943 return cfq_rb_first(&cfqd->service_tree);
947 * Get and set a new active queue for service.
949 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
950 struct cfq_queue *cfqq)
953 cfqq = cfq_get_next_queue(cfqd);
955 cfq_clear_cfqq_coop(cfqq);
958 __cfq_set_active_queue(cfqd, cfqq);
962 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
965 if (blk_rq_pos(rq) >= cfqd->last_position)
966 return blk_rq_pos(rq) - cfqd->last_position;
968 return cfqd->last_position - blk_rq_pos(rq);
971 #define CIC_SEEK_THR 8 * 1024
972 #define CIC_SEEKY(cic) ((cic)->seek_mean > CIC_SEEK_THR)
974 static inline int cfq_rq_close(struct cfq_data *cfqd, struct request *rq)
976 struct cfq_io_context *cic = cfqd->active_cic;
977 sector_t sdist = cic->seek_mean;
979 if (!sample_valid(cic->seek_samples))
980 sdist = CIC_SEEK_THR;
982 return cfq_dist_from_last(cfqd, rq) <= sdist;
985 static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
986 struct cfq_queue *cur_cfqq)
988 struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
989 struct rb_node *parent, *node;
990 struct cfq_queue *__cfqq;
991 sector_t sector = cfqd->last_position;
993 if (RB_EMPTY_ROOT(root))
997 * First, if we find a request starting at the end of the last
998 * request, choose it.
1000 __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
1005 * If the exact sector wasn't found, the parent of the NULL leaf
1006 * will contain the closest sector.
1008 __cfqq = rb_entry(parent, struct cfq_queue, p_node);
1009 if (cfq_rq_close(cfqd, __cfqq->next_rq))
1012 if (blk_rq_pos(__cfqq->next_rq) < sector)
1013 node = rb_next(&__cfqq->p_node);
1015 node = rb_prev(&__cfqq->p_node);
1019 __cfqq = rb_entry(node, struct cfq_queue, p_node);
1020 if (cfq_rq_close(cfqd, __cfqq->next_rq))
1028 * cur_cfqq - passed in so that we don't decide that the current queue is
1029 * closely cooperating with itself.
1031 * So, basically we're assuming that that cur_cfqq has dispatched at least
1032 * one request, and that cfqd->last_position reflects a position on the disk
1033 * associated with the I/O issued by cur_cfqq. I'm not sure this is a valid
1036 static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
1037 struct cfq_queue *cur_cfqq,
1040 struct cfq_queue *cfqq;
1043 * A valid cfq_io_context is necessary to compare requests against
1044 * the seek_mean of the current cfqq.
1046 if (!cfqd->active_cic)
1050 * We should notice if some of the queues are cooperating, eg
1051 * working closely on the same area of the disk. In that case,
1052 * we can group them together and don't waste time idling.
1054 cfqq = cfqq_close(cfqd, cur_cfqq);
1058 if (cfq_cfqq_coop(cfqq))
1062 cfq_mark_cfqq_coop(cfqq);
1066 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
1068 struct cfq_queue *cfqq = cfqd->active_queue;
1069 struct cfq_io_context *cic;
1073 * SSD device without seek penalty, disable idling. But only do so
1074 * for devices that support queuing, otherwise we still have a problem
1075 * with sync vs async workloads.
1077 if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
1080 WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
1081 WARN_ON(cfq_cfqq_slice_new(cfqq));
1084 * idle is disabled, either manually or by past process history
1086 if (!cfqd->cfq_slice_idle || !cfq_cfqq_idle_window(cfqq))
1090 * still requests with the driver, don't idle
1092 if (rq_in_driver(cfqd))
1096 * task has exited, don't wait
1098 cic = cfqd->active_cic;
1099 if (!cic || !atomic_read(&cic->ioc->nr_tasks))
1102 cfq_mark_cfqq_wait_request(cfqq);
1105 * we don't want to idle for seeks, but we do want to allow
1106 * fair distribution of slice time for a process doing back-to-back
1107 * seeks. so allow a little bit of time for him to submit a new rq
1109 sl = cfqd->cfq_slice_idle;
1110 if (sample_valid(cic->seek_samples) && CIC_SEEKY(cic))
1111 sl = min(sl, msecs_to_jiffies(CFQ_MIN_TT));
1113 mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
1114 cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu", sl);
1118 * Move request from internal lists to the request queue dispatch list.
1120 static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
1122 struct cfq_data *cfqd = q->elevator->elevator_data;
1123 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1125 cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
1127 cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
1128 cfq_remove_request(rq);
1130 elv_dispatch_sort(q, rq);
1132 if (cfq_cfqq_sync(cfqq))
1133 cfqd->sync_flight++;
1137 * return expired entry, or NULL to just start from scratch in rbtree
1139 static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
1141 struct request *rq = NULL;
1143 if (cfq_cfqq_fifo_expire(cfqq))
1146 cfq_mark_cfqq_fifo_expire(cfqq);
1148 if (list_empty(&cfqq->fifo))
1151 rq = rq_entry_fifo(cfqq->fifo.next);
1152 if (time_before(jiffies, rq_fifo_time(rq)))
1155 cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq);
1160 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1162 const int base_rq = cfqd->cfq_slice_async_rq;
1164 WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
1166 return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
1170 * Select a queue for service. If we have a current active queue,
1171 * check whether to continue servicing it, or retrieve and set a new one.
1173 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
1175 struct cfq_queue *cfqq, *new_cfqq = NULL;
1177 cfqq = cfqd->active_queue;
1182 * The active queue has run out of time, expire it and select new.
1184 if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq))
1188 * The active queue has requests and isn't expired, allow it to
1191 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
1195 * If another queue has a request waiting within our mean seek
1196 * distance, let it run. The expire code will check for close
1197 * cooperators and put the close queue at the front of the service
1200 new_cfqq = cfq_close_cooperator(cfqd, cfqq, 0);
1205 * No requests pending. If the active queue still has requests in
1206 * flight or is idling for a new request, allow either of these
1207 * conditions to happen (or time out) before selecting a new queue.
1209 if (timer_pending(&cfqd->idle_slice_timer) ||
1210 (cfqq->dispatched && cfq_cfqq_idle_window(cfqq))) {
1216 cfq_slice_expired(cfqd, 0);
1218 cfqq = cfq_set_active_queue(cfqd, new_cfqq);
1223 static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
1227 while (cfqq->next_rq) {
1228 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
1232 BUG_ON(!list_empty(&cfqq->fifo));
1237 * Drain our current requests. Used for barriers and when switching
1238 * io schedulers on-the-fly.
1240 static int cfq_forced_dispatch(struct cfq_data *cfqd)
1242 struct cfq_queue *cfqq;
1245 while ((cfqq = cfq_rb_first(&cfqd->service_tree)) != NULL)
1246 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
1248 cfq_slice_expired(cfqd, 0);
1250 BUG_ON(cfqd->busy_queues);
1252 cfq_log(cfqd, "forced_dispatch=%d", dispatched);
1256 static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1258 unsigned int max_dispatch;
1261 * Drain async requests before we start sync IO
1263 if (cfq_cfqq_idle_window(cfqq) && cfqd->rq_in_driver[BLK_RW_ASYNC])
1267 * If this is an async queue and we have sync IO in flight, let it wait
1269 if (cfqd->sync_flight && !cfq_cfqq_sync(cfqq))
1272 max_dispatch = cfqd->cfq_quantum;
1273 if (cfq_class_idle(cfqq))
1277 * Does this cfqq already have too much IO in flight?
1279 if (cfqq->dispatched >= max_dispatch) {
1281 * idle queue must always only have a single IO in flight
1283 if (cfq_class_idle(cfqq))
1287 * We have other queues, don't allow more IO from this one
1289 if (cfqd->busy_queues > 1)
1293 * Sole queue user, allow bigger slice
1299 * Async queues must wait a bit before being allowed dispatch.
1300 * We also ramp up the dispatch depth gradually for async IO,
1301 * based on the last sync IO we serviced
1303 if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
1304 unsigned long last_sync = jiffies - cfqd->last_end_sync_rq;
1307 depth = last_sync / cfqd->cfq_slice[1];
1308 if (!depth && !cfqq->dispatched)
1310 if (depth < max_dispatch)
1311 max_dispatch = depth;
1315 * If we're below the current max, allow a dispatch
1317 return cfqq->dispatched < max_dispatch;
1321 * Dispatch a request from cfqq, moving them to the request queue
1324 static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1328 BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
1330 if (!cfq_may_dispatch(cfqd, cfqq))
1334 * follow expired path, else get first next available
1336 rq = cfq_check_fifo(cfqq);
1341 * insert request into driver dispatch list
1343 cfq_dispatch_insert(cfqd->queue, rq);
1345 if (!cfqd->active_cic) {
1346 struct cfq_io_context *cic = RQ_CIC(rq);
1348 atomic_long_inc(&cic->ioc->refcount);
1349 cfqd->active_cic = cic;
1356 * Find the cfqq that we need to service and move a request from that to the
1359 static int cfq_dispatch_requests(struct request_queue *q, int force)
1361 struct cfq_data *cfqd = q->elevator->elevator_data;
1362 struct cfq_queue *cfqq;
1364 if (!cfqd->busy_queues)
1367 if (unlikely(force))
1368 return cfq_forced_dispatch(cfqd);
1370 cfqq = cfq_select_queue(cfqd);
1375 * Dispatch a request from this cfqq, if it is allowed
1377 if (!cfq_dispatch_request(cfqd, cfqq))
1380 cfqq->slice_dispatch++;
1381 cfq_clear_cfqq_must_dispatch(cfqq);
1384 * expire an async queue immediately if it has used up its slice. idle
1385 * queue always expire after 1 dispatch round.
1387 if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
1388 cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
1389 cfq_class_idle(cfqq))) {
1390 cfqq->slice_end = jiffies + 1;
1391 cfq_slice_expired(cfqd, 0);
1394 cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
1399 * task holds one reference to the queue, dropped when task exits. each rq
1400 * in-flight on this queue also holds a reference, dropped when rq is freed.
1402 * queue lock must be held here.
1404 static void cfq_put_queue(struct cfq_queue *cfqq)
1406 struct cfq_data *cfqd = cfqq->cfqd;
1408 BUG_ON(atomic_read(&cfqq->ref) <= 0);
1410 if (!atomic_dec_and_test(&cfqq->ref))
1413 cfq_log_cfqq(cfqd, cfqq, "put_queue");
1414 BUG_ON(rb_first(&cfqq->sort_list));
1415 BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
1416 BUG_ON(cfq_cfqq_on_rr(cfqq));
1418 if (unlikely(cfqd->active_queue == cfqq)) {
1419 __cfq_slice_expired(cfqd, cfqq, 0);
1420 cfq_schedule_dispatch(cfqd);
1423 kmem_cache_free(cfq_pool, cfqq);
1427 * Must always be called with the rcu_read_lock() held
1430 __call_for_each_cic(struct io_context *ioc,
1431 void (*func)(struct io_context *, struct cfq_io_context *))
1433 struct cfq_io_context *cic;
1434 struct hlist_node *n;
1436 hlist_for_each_entry_rcu(cic, n, &ioc->cic_list, cic_list)
1441 * Call func for each cic attached to this ioc.
1444 call_for_each_cic(struct io_context *ioc,
1445 void (*func)(struct io_context *, struct cfq_io_context *))
1448 __call_for_each_cic(ioc, func);
1452 static void cfq_cic_free_rcu(struct rcu_head *head)
1454 struct cfq_io_context *cic;
1456 cic = container_of(head, struct cfq_io_context, rcu_head);
1458 kmem_cache_free(cfq_ioc_pool, cic);
1459 elv_ioc_count_dec(cfq_ioc_count);
1463 * CFQ scheduler is exiting, grab exit lock and check
1464 * the pending io context count. If it hits zero,
1465 * complete ioc_gone and set it back to NULL
1467 spin_lock(&ioc_gone_lock);
1468 if (ioc_gone && !elv_ioc_count_read(cfq_ioc_count)) {
1472 spin_unlock(&ioc_gone_lock);
1476 static void cfq_cic_free(struct cfq_io_context *cic)
1478 call_rcu(&cic->rcu_head, cfq_cic_free_rcu);
1481 static void cic_free_func(struct io_context *ioc, struct cfq_io_context *cic)
1483 unsigned long flags;
1485 BUG_ON(!cic->dead_key);
1487 spin_lock_irqsave(&ioc->lock, flags);
1488 radix_tree_delete(&ioc->radix_root, cic->dead_key);
1489 hlist_del_rcu(&cic->cic_list);
1490 spin_unlock_irqrestore(&ioc->lock, flags);
1496 * Must be called with rcu_read_lock() held or preemption otherwise disabled.
1497 * Only two callers of this - ->dtor() which is called with the rcu_read_lock(),
1498 * and ->trim() which is called with the task lock held
1500 static void cfq_free_io_context(struct io_context *ioc)
1503 * ioc->refcount is zero here, or we are called from elv_unregister(),
1504 * so no more cic's are allowed to be linked into this ioc. So it
1505 * should be ok to iterate over the known list, we will see all cic's
1506 * since no new ones are added.
1508 __call_for_each_cic(ioc, cic_free_func);
1511 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1513 if (unlikely(cfqq == cfqd->active_queue)) {
1514 __cfq_slice_expired(cfqd, cfqq, 0);
1515 cfq_schedule_dispatch(cfqd);
1518 cfq_put_queue(cfqq);
1521 static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
1522 struct cfq_io_context *cic)
1524 struct io_context *ioc = cic->ioc;
1526 list_del_init(&cic->queue_list);
1529 * Make sure key == NULL is seen for dead queues
1532 cic->dead_key = (unsigned long) cic->key;
1535 if (ioc->ioc_data == cic)
1536 rcu_assign_pointer(ioc->ioc_data, NULL);
1538 if (cic->cfqq[BLK_RW_ASYNC]) {
1539 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]);
1540 cic->cfqq[BLK_RW_ASYNC] = NULL;
1543 if (cic->cfqq[BLK_RW_SYNC]) {
1544 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]);
1545 cic->cfqq[BLK_RW_SYNC] = NULL;
1549 static void cfq_exit_single_io_context(struct io_context *ioc,
1550 struct cfq_io_context *cic)
1552 struct cfq_data *cfqd = cic->key;
1555 struct request_queue *q = cfqd->queue;
1556 unsigned long flags;
1558 spin_lock_irqsave(q->queue_lock, flags);
1561 * Ensure we get a fresh copy of the ->key to prevent
1562 * race between exiting task and queue
1564 smp_read_barrier_depends();
1566 __cfq_exit_single_io_context(cfqd, cic);
1568 spin_unlock_irqrestore(q->queue_lock, flags);
1573 * The process that ioc belongs to has exited, we need to clean up
1574 * and put the internal structures we have that belongs to that process.
1576 static void cfq_exit_io_context(struct io_context *ioc)
1578 call_for_each_cic(ioc, cfq_exit_single_io_context);
1581 static struct cfq_io_context *
1582 cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
1584 struct cfq_io_context *cic;
1586 cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO,
1589 cic->last_end_request = jiffies;
1590 INIT_LIST_HEAD(&cic->queue_list);
1591 INIT_HLIST_NODE(&cic->cic_list);
1592 cic->dtor = cfq_free_io_context;
1593 cic->exit = cfq_exit_io_context;
1594 elv_ioc_count_inc(cfq_ioc_count);
1600 static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc)
1602 struct task_struct *tsk = current;
1605 if (!cfq_cfqq_prio_changed(cfqq))
1608 ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio);
1609 switch (ioprio_class) {
1611 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
1612 case IOPRIO_CLASS_NONE:
1614 * no prio set, inherit CPU scheduling settings
1616 cfqq->ioprio = task_nice_ioprio(tsk);
1617 cfqq->ioprio_class = task_nice_ioclass(tsk);
1619 case IOPRIO_CLASS_RT:
1620 cfqq->ioprio = task_ioprio(ioc);
1621 cfqq->ioprio_class = IOPRIO_CLASS_RT;
1623 case IOPRIO_CLASS_BE:
1624 cfqq->ioprio = task_ioprio(ioc);
1625 cfqq->ioprio_class = IOPRIO_CLASS_BE;
1627 case IOPRIO_CLASS_IDLE:
1628 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
1630 cfq_clear_cfqq_idle_window(cfqq);
1635 * keep track of original prio settings in case we have to temporarily
1636 * elevate the priority of this queue
1638 cfqq->org_ioprio = cfqq->ioprio;
1639 cfqq->org_ioprio_class = cfqq->ioprio_class;
1640 cfq_clear_cfqq_prio_changed(cfqq);
1643 static void changed_ioprio(struct io_context *ioc, struct cfq_io_context *cic)
1645 struct cfq_data *cfqd = cic->key;
1646 struct cfq_queue *cfqq;
1647 unsigned long flags;
1649 if (unlikely(!cfqd))
1652 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1654 cfqq = cic->cfqq[BLK_RW_ASYNC];
1656 struct cfq_queue *new_cfqq;
1657 new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic->ioc,
1660 cic->cfqq[BLK_RW_ASYNC] = new_cfqq;
1661 cfq_put_queue(cfqq);
1665 cfqq = cic->cfqq[BLK_RW_SYNC];
1667 cfq_mark_cfqq_prio_changed(cfqq);
1669 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1672 static void cfq_ioc_set_ioprio(struct io_context *ioc)
1674 call_for_each_cic(ioc, changed_ioprio);
1675 ioc->ioprio_changed = 0;
1678 static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1679 pid_t pid, int is_sync)
1681 RB_CLEAR_NODE(&cfqq->rb_node);
1682 RB_CLEAR_NODE(&cfqq->p_node);
1683 INIT_LIST_HEAD(&cfqq->fifo);
1685 atomic_set(&cfqq->ref, 0);
1688 cfq_mark_cfqq_prio_changed(cfqq);
1691 if (!cfq_class_idle(cfqq))
1692 cfq_mark_cfqq_idle_window(cfqq);
1693 cfq_mark_cfqq_sync(cfqq);
1698 static struct cfq_queue *
1699 cfq_find_alloc_queue(struct cfq_data *cfqd, int is_sync,
1700 struct io_context *ioc, gfp_t gfp_mask)
1702 struct cfq_queue *cfqq, *new_cfqq = NULL;
1703 struct cfq_io_context *cic;
1706 cic = cfq_cic_lookup(cfqd, ioc);
1707 /* cic always exists here */
1708 cfqq = cic_to_cfqq(cic, is_sync);
1711 * Always try a new alloc if we fell back to the OOM cfqq
1712 * originally, since it should just be a temporary situation.
1714 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
1719 } else if (gfp_mask & __GFP_WAIT) {
1720 spin_unlock_irq(cfqd->queue->queue_lock);
1721 new_cfqq = kmem_cache_alloc_node(cfq_pool,
1722 gfp_mask | __GFP_ZERO,
1724 spin_lock_irq(cfqd->queue->queue_lock);
1728 cfqq = kmem_cache_alloc_node(cfq_pool,
1729 gfp_mask | __GFP_ZERO,
1734 cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
1735 cfq_init_prio_data(cfqq, ioc);
1736 cfq_log_cfqq(cfqd, cfqq, "alloced");
1738 cfqq = &cfqd->oom_cfqq;
1742 kmem_cache_free(cfq_pool, new_cfqq);
1747 static struct cfq_queue **
1748 cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
1750 switch (ioprio_class) {
1751 case IOPRIO_CLASS_RT:
1752 return &cfqd->async_cfqq[0][ioprio];
1753 case IOPRIO_CLASS_BE:
1754 return &cfqd->async_cfqq[1][ioprio];
1755 case IOPRIO_CLASS_IDLE:
1756 return &cfqd->async_idle_cfqq;
1762 static struct cfq_queue *
1763 cfq_get_queue(struct cfq_data *cfqd, int is_sync, struct io_context *ioc,
1766 const int ioprio = task_ioprio(ioc);
1767 const int ioprio_class = task_ioprio_class(ioc);
1768 struct cfq_queue **async_cfqq = NULL;
1769 struct cfq_queue *cfqq = NULL;
1772 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
1777 cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask);
1780 * pin the queue now that it's allocated, scheduler exit will prune it
1782 if (!is_sync && !(*async_cfqq)) {
1783 atomic_inc(&cfqq->ref);
1787 atomic_inc(&cfqq->ref);
1792 * We drop cfq io contexts lazily, so we may find a dead one.
1795 cfq_drop_dead_cic(struct cfq_data *cfqd, struct io_context *ioc,
1796 struct cfq_io_context *cic)
1798 unsigned long flags;
1800 WARN_ON(!list_empty(&cic->queue_list));
1802 spin_lock_irqsave(&ioc->lock, flags);
1804 BUG_ON(ioc->ioc_data == cic);
1806 radix_tree_delete(&ioc->radix_root, (unsigned long) cfqd);
1807 hlist_del_rcu(&cic->cic_list);
1808 spin_unlock_irqrestore(&ioc->lock, flags);
1813 static struct cfq_io_context *
1814 cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc)
1816 struct cfq_io_context *cic;
1817 unsigned long flags;
1826 * we maintain a last-hit cache, to avoid browsing over the tree
1828 cic = rcu_dereference(ioc->ioc_data);
1829 if (cic && cic->key == cfqd) {
1835 cic = radix_tree_lookup(&ioc->radix_root, (unsigned long) cfqd);
1839 /* ->key must be copied to avoid race with cfq_exit_queue() */
1842 cfq_drop_dead_cic(cfqd, ioc, cic);
1847 spin_lock_irqsave(&ioc->lock, flags);
1848 rcu_assign_pointer(ioc->ioc_data, cic);
1849 spin_unlock_irqrestore(&ioc->lock, flags);
1857 * Add cic into ioc, using cfqd as the search key. This enables us to lookup
1858 * the process specific cfq io context when entered from the block layer.
1859 * Also adds the cic to a per-cfqd list, used when this queue is removed.
1861 static int cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
1862 struct cfq_io_context *cic, gfp_t gfp_mask)
1864 unsigned long flags;
1867 ret = radix_tree_preload(gfp_mask);
1872 spin_lock_irqsave(&ioc->lock, flags);
1873 ret = radix_tree_insert(&ioc->radix_root,
1874 (unsigned long) cfqd, cic);
1876 hlist_add_head_rcu(&cic->cic_list, &ioc->cic_list);
1877 spin_unlock_irqrestore(&ioc->lock, flags);
1879 radix_tree_preload_end();
1882 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1883 list_add(&cic->queue_list, &cfqd->cic_list);
1884 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1889 printk(KERN_ERR "cfq: cic link failed!\n");
1895 * Setup general io context and cfq io context. There can be several cfq
1896 * io contexts per general io context, if this process is doing io to more
1897 * than one device managed by cfq.
1899 static struct cfq_io_context *
1900 cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
1902 struct io_context *ioc = NULL;
1903 struct cfq_io_context *cic;
1905 might_sleep_if(gfp_mask & __GFP_WAIT);
1907 ioc = get_io_context(gfp_mask, cfqd->queue->node);
1911 cic = cfq_cic_lookup(cfqd, ioc);
1915 cic = cfq_alloc_io_context(cfqd, gfp_mask);
1919 if (cfq_cic_link(cfqd, ioc, cic, gfp_mask))
1923 smp_read_barrier_depends();
1924 if (unlikely(ioc->ioprio_changed))
1925 cfq_ioc_set_ioprio(ioc);
1931 put_io_context(ioc);
1936 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
1938 unsigned long elapsed = jiffies - cic->last_end_request;
1939 unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
1941 cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
1942 cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
1943 cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
1947 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_io_context *cic,
1953 if (!cic->last_request_pos)
1955 else if (cic->last_request_pos < blk_rq_pos(rq))
1956 sdist = blk_rq_pos(rq) - cic->last_request_pos;
1958 sdist = cic->last_request_pos - blk_rq_pos(rq);
1961 * Don't allow the seek distance to get too large from the
1962 * odd fragment, pagein, etc
1964 if (cic->seek_samples <= 60) /* second&third seek */
1965 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*1024);
1967 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*64);
1969 cic->seek_samples = (7*cic->seek_samples + 256) / 8;
1970 cic->seek_total = (7*cic->seek_total + (u64)256*sdist) / 8;
1971 total = cic->seek_total + (cic->seek_samples/2);
1972 do_div(total, cic->seek_samples);
1973 cic->seek_mean = (sector_t)total;
1977 * Disable idle window if the process thinks too long or seeks so much that
1981 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1982 struct cfq_io_context *cic)
1984 int old_idle, enable_idle;
1987 * Don't idle for async or idle io prio class
1989 if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
1992 enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
1994 if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle ||
1995 (!cfqd->cfq_latency && cfqd->hw_tag && CIC_SEEKY(cic)))
1997 else if (sample_valid(cic->ttime_samples)) {
1998 if (cic->ttime_mean > cfqd->cfq_slice_idle)
2004 if (old_idle != enable_idle) {
2005 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
2007 cfq_mark_cfqq_idle_window(cfqq);
2009 cfq_clear_cfqq_idle_window(cfqq);
2014 * Check if new_cfqq should preempt the currently active queue. Return 0 for
2015 * no or if we aren't sure, a 1 will cause a preempt.
2018 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
2021 struct cfq_queue *cfqq;
2023 cfqq = cfqd->active_queue;
2027 if (cfq_slice_used(cfqq))
2030 if (cfq_class_idle(new_cfqq))
2033 if (cfq_class_idle(cfqq))
2037 * if the new request is sync, but the currently running queue is
2038 * not, let the sync request have priority.
2040 if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
2044 * So both queues are sync. Let the new request get disk time if
2045 * it's a metadata request and the current queue is doing regular IO.
2047 if (rq_is_meta(rq) && !cfqq->meta_pending)
2051 * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
2053 if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
2056 if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
2060 * if this request is as-good as one we would expect from the
2061 * current cfqq, let it preempt
2063 if (cfq_rq_close(cfqd, rq))
2070 * cfqq preempts the active queue. if we allowed preempt with no slice left,
2071 * let it have half of its nominal slice.
2073 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2075 cfq_log_cfqq(cfqd, cfqq, "preempt");
2076 cfq_slice_expired(cfqd, 1);
2079 * Put the new queue at the front of the of the current list,
2080 * so we know that it will be selected next.
2082 BUG_ON(!cfq_cfqq_on_rr(cfqq));
2084 cfq_service_tree_add(cfqd, cfqq, 1);
2086 cfqq->slice_end = 0;
2087 cfq_mark_cfqq_slice_new(cfqq);
2091 * Called when a new fs request (rq) is added (to cfqq). Check if there's
2092 * something we should do about it
2095 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2098 struct cfq_io_context *cic = RQ_CIC(rq);
2102 cfqq->meta_pending++;
2104 cfq_update_io_thinktime(cfqd, cic);
2105 cfq_update_io_seektime(cfqd, cic, rq);
2106 cfq_update_idle_window(cfqd, cfqq, cic);
2108 cic->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
2110 if (cfqq == cfqd->active_queue) {
2112 * Remember that we saw a request from this process, but
2113 * don't start queuing just yet. Otherwise we risk seeing lots
2114 * of tiny requests, because we disrupt the normal plugging
2115 * and merging. If the request is already larger than a single
2116 * page, let it rip immediately. For that case we assume that
2117 * merging is already done. Ditto for a busy system that
2118 * has other work pending, don't risk delaying until the
2119 * idle timer unplug to continue working.
2121 if (cfq_cfqq_wait_request(cfqq)) {
2122 if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
2123 cfqd->busy_queues > 1) {
2124 del_timer(&cfqd->idle_slice_timer);
2125 __blk_run_queue(cfqd->queue);
2127 cfq_mark_cfqq_must_dispatch(cfqq);
2129 } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
2131 * not the active queue - expire current slice if it is
2132 * idle and has expired it's mean thinktime or this new queue
2133 * has some old slice time left and is of higher priority or
2134 * this new queue is RT and the current one is BE
2136 cfq_preempt_queue(cfqd, cfqq);
2137 __blk_run_queue(cfqd->queue);
2141 static void cfq_insert_request(struct request_queue *q, struct request *rq)
2143 struct cfq_data *cfqd = q->elevator->elevator_data;
2144 struct cfq_queue *cfqq = RQ_CFQQ(rq);
2146 cfq_log_cfqq(cfqd, cfqq, "insert_request");
2147 cfq_init_prio_data(cfqq, RQ_CIC(rq)->ioc);
2151 rq_set_fifo_time(rq, jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)]);
2152 list_add_tail(&rq->queuelist, &cfqq->fifo);
2154 cfq_rq_enqueued(cfqd, cfqq, rq);
2158 * Update hw_tag based on peak queue depth over 50 samples under
2161 static void cfq_update_hw_tag(struct cfq_data *cfqd)
2163 if (rq_in_driver(cfqd) > cfqd->rq_in_driver_peak)
2164 cfqd->rq_in_driver_peak = rq_in_driver(cfqd);
2166 if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
2167 rq_in_driver(cfqd) <= CFQ_HW_QUEUE_MIN)
2170 if (cfqd->hw_tag_samples++ < 50)
2173 if (cfqd->rq_in_driver_peak >= CFQ_HW_QUEUE_MIN)
2178 cfqd->hw_tag_samples = 0;
2179 cfqd->rq_in_driver_peak = 0;
2182 static void cfq_completed_request(struct request_queue *q, struct request *rq)
2184 struct cfq_queue *cfqq = RQ_CFQQ(rq);
2185 struct cfq_data *cfqd = cfqq->cfqd;
2186 const int sync = rq_is_sync(rq);
2190 cfq_log_cfqq(cfqd, cfqq, "complete");
2192 cfq_update_hw_tag(cfqd);
2194 WARN_ON(!cfqd->rq_in_driver[sync]);
2195 WARN_ON(!cfqq->dispatched);
2196 cfqd->rq_in_driver[sync]--;
2199 if (cfq_cfqq_sync(cfqq))
2200 cfqd->sync_flight--;
2203 RQ_CIC(rq)->last_end_request = now;
2204 cfqd->last_end_sync_rq = now;
2208 * If this is the active queue, check if it needs to be expired,
2209 * or if we want to idle in case it has no pending requests.
2211 if (cfqd->active_queue == cfqq) {
2212 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
2214 if (cfq_cfqq_slice_new(cfqq)) {
2215 cfq_set_prio_slice(cfqd, cfqq);
2216 cfq_clear_cfqq_slice_new(cfqq);
2219 * If there are no requests waiting in this queue, and
2220 * there are other queues ready to issue requests, AND
2221 * those other queues are issuing requests within our
2222 * mean seek distance, give them a chance to run instead
2225 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
2226 cfq_slice_expired(cfqd, 1);
2227 else if (cfqq_empty && !cfq_close_cooperator(cfqd, cfqq, 1) &&
2228 sync && !rq_noidle(rq))
2229 cfq_arm_slice_timer(cfqd);
2232 if (!rq_in_driver(cfqd))
2233 cfq_schedule_dispatch(cfqd);
2237 * we temporarily boost lower priority queues if they are holding fs exclusive
2238 * resources. they are boosted to normal prio (CLASS_BE/4)
2240 static void cfq_prio_boost(struct cfq_queue *cfqq)
2242 if (has_fs_excl()) {
2244 * boost idle prio on transactions that would lock out other
2245 * users of the filesystem
2247 if (cfq_class_idle(cfqq))
2248 cfqq->ioprio_class = IOPRIO_CLASS_BE;
2249 if (cfqq->ioprio > IOPRIO_NORM)
2250 cfqq->ioprio = IOPRIO_NORM;
2253 * check if we need to unboost the queue
2255 if (cfqq->ioprio_class != cfqq->org_ioprio_class)
2256 cfqq->ioprio_class = cfqq->org_ioprio_class;
2257 if (cfqq->ioprio != cfqq->org_ioprio)
2258 cfqq->ioprio = cfqq->org_ioprio;
2262 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
2264 if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
2265 cfq_mark_cfqq_must_alloc_slice(cfqq);
2266 return ELV_MQUEUE_MUST;
2269 return ELV_MQUEUE_MAY;
2272 static int cfq_may_queue(struct request_queue *q, int rw)
2274 struct cfq_data *cfqd = q->elevator->elevator_data;
2275 struct task_struct *tsk = current;
2276 struct cfq_io_context *cic;
2277 struct cfq_queue *cfqq;
2280 * don't force setup of a queue from here, as a call to may_queue
2281 * does not necessarily imply that a request actually will be queued.
2282 * so just lookup a possibly existing queue, or return 'may queue'
2285 cic = cfq_cic_lookup(cfqd, tsk->io_context);
2287 return ELV_MQUEUE_MAY;
2289 cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
2291 cfq_init_prio_data(cfqq, cic->ioc);
2292 cfq_prio_boost(cfqq);
2294 return __cfq_may_queue(cfqq);
2297 return ELV_MQUEUE_MAY;
2301 * queue lock held here
2303 static void cfq_put_request(struct request *rq)
2305 struct cfq_queue *cfqq = RQ_CFQQ(rq);
2308 const int rw = rq_data_dir(rq);
2310 BUG_ON(!cfqq->allocated[rw]);
2311 cfqq->allocated[rw]--;
2313 put_io_context(RQ_CIC(rq)->ioc);
2315 rq->elevator_private = NULL;
2316 rq->elevator_private2 = NULL;
2318 cfq_put_queue(cfqq);
2323 * Allocate cfq data structures associated with this request.
2326 cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
2328 struct cfq_data *cfqd = q->elevator->elevator_data;
2329 struct cfq_io_context *cic;
2330 const int rw = rq_data_dir(rq);
2331 const int is_sync = rq_is_sync(rq);
2332 struct cfq_queue *cfqq;
2333 unsigned long flags;
2335 might_sleep_if(gfp_mask & __GFP_WAIT);
2337 cic = cfq_get_io_context(cfqd, gfp_mask);
2339 spin_lock_irqsave(q->queue_lock, flags);
2344 cfqq = cic_to_cfqq(cic, is_sync);
2345 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
2346 cfqq = cfq_get_queue(cfqd, is_sync, cic->ioc, gfp_mask);
2347 cic_set_cfqq(cic, cfqq, is_sync);
2350 cfqq->allocated[rw]++;
2351 atomic_inc(&cfqq->ref);
2353 spin_unlock_irqrestore(q->queue_lock, flags);
2355 rq->elevator_private = cic;
2356 rq->elevator_private2 = cfqq;
2361 put_io_context(cic->ioc);
2363 cfq_schedule_dispatch(cfqd);
2364 spin_unlock_irqrestore(q->queue_lock, flags);
2365 cfq_log(cfqd, "set_request fail");
2369 static void cfq_kick_queue(struct work_struct *work)
2371 struct cfq_data *cfqd =
2372 container_of(work, struct cfq_data, unplug_work);
2373 struct request_queue *q = cfqd->queue;
2375 spin_lock_irq(q->queue_lock);
2376 __blk_run_queue(cfqd->queue);
2377 spin_unlock_irq(q->queue_lock);
2381 * Timer running if the active_queue is currently idling inside its time slice
2383 static void cfq_idle_slice_timer(unsigned long data)
2385 struct cfq_data *cfqd = (struct cfq_data *) data;
2386 struct cfq_queue *cfqq;
2387 unsigned long flags;
2390 cfq_log(cfqd, "idle timer fired");
2392 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2394 cfqq = cfqd->active_queue;
2399 * We saw a request before the queue expired, let it through
2401 if (cfq_cfqq_must_dispatch(cfqq))
2407 if (cfq_slice_used(cfqq))
2411 * only expire and reinvoke request handler, if there are
2412 * other queues with pending requests
2414 if (!cfqd->busy_queues)
2418 * not expired and it has a request pending, let it dispatch
2420 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
2424 cfq_slice_expired(cfqd, timed_out);
2426 cfq_schedule_dispatch(cfqd);
2428 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2431 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
2433 del_timer_sync(&cfqd->idle_slice_timer);
2434 cancel_work_sync(&cfqd->unplug_work);
2437 static void cfq_put_async_queues(struct cfq_data *cfqd)
2441 for (i = 0; i < IOPRIO_BE_NR; i++) {
2442 if (cfqd->async_cfqq[0][i])
2443 cfq_put_queue(cfqd->async_cfqq[0][i]);
2444 if (cfqd->async_cfqq[1][i])
2445 cfq_put_queue(cfqd->async_cfqq[1][i]);
2448 if (cfqd->async_idle_cfqq)
2449 cfq_put_queue(cfqd->async_idle_cfqq);
2452 static void cfq_exit_queue(struct elevator_queue *e)
2454 struct cfq_data *cfqd = e->elevator_data;
2455 struct request_queue *q = cfqd->queue;
2457 cfq_shutdown_timer_wq(cfqd);
2459 spin_lock_irq(q->queue_lock);
2461 if (cfqd->active_queue)
2462 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
2464 while (!list_empty(&cfqd->cic_list)) {
2465 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
2466 struct cfq_io_context,
2469 __cfq_exit_single_io_context(cfqd, cic);
2472 cfq_put_async_queues(cfqd);
2474 spin_unlock_irq(q->queue_lock);
2476 cfq_shutdown_timer_wq(cfqd);
2481 static void *cfq_init_queue(struct request_queue *q)
2483 struct cfq_data *cfqd;
2486 cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
2490 cfqd->service_tree = CFQ_RB_ROOT;
2493 * Not strictly needed (since RB_ROOT just clears the node and we
2494 * zeroed cfqd on alloc), but better be safe in case someone decides
2495 * to add magic to the rb code
2497 for (i = 0; i < CFQ_PRIO_LISTS; i++)
2498 cfqd->prio_trees[i] = RB_ROOT;
2501 * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
2502 * Grab a permanent reference to it, so that the normal code flow
2503 * will not attempt to free it.
2505 cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
2506 atomic_inc(&cfqd->oom_cfqq.ref);
2508 INIT_LIST_HEAD(&cfqd->cic_list);
2512 init_timer(&cfqd->idle_slice_timer);
2513 cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
2514 cfqd->idle_slice_timer.data = (unsigned long) cfqd;
2516 INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
2518 cfqd->cfq_quantum = cfq_quantum;
2519 cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
2520 cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
2521 cfqd->cfq_back_max = cfq_back_max;
2522 cfqd->cfq_back_penalty = cfq_back_penalty;
2523 cfqd->cfq_slice[0] = cfq_slice_async;
2524 cfqd->cfq_slice[1] = cfq_slice_sync;
2525 cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
2526 cfqd->cfq_slice_idle = cfq_slice_idle;
2527 cfqd->cfq_latency = 1;
2529 cfqd->last_end_sync_rq = jiffies;
2533 static void cfq_slab_kill(void)
2536 * Caller already ensured that pending RCU callbacks are completed,
2537 * so we should have no busy allocations at this point.
2540 kmem_cache_destroy(cfq_pool);
2542 kmem_cache_destroy(cfq_ioc_pool);
2545 static int __init cfq_slab_setup(void)
2547 cfq_pool = KMEM_CACHE(cfq_queue, 0);
2551 cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0);
2562 * sysfs parts below -->
2565 cfq_var_show(unsigned int var, char *page)
2567 return sprintf(page, "%d\n", var);
2571 cfq_var_store(unsigned int *var, const char *page, size_t count)
2573 char *p = (char *) page;
2575 *var = simple_strtoul(p, &p, 10);
2579 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
2580 static ssize_t __FUNC(struct elevator_queue *e, char *page) \
2582 struct cfq_data *cfqd = e->elevator_data; \
2583 unsigned int __data = __VAR; \
2585 __data = jiffies_to_msecs(__data); \
2586 return cfq_var_show(__data, (page)); \
2588 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
2589 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
2590 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
2591 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
2592 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
2593 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
2594 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
2595 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
2596 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
2597 SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
2598 #undef SHOW_FUNCTION
2600 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
2601 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
2603 struct cfq_data *cfqd = e->elevator_data; \
2604 unsigned int __data; \
2605 int ret = cfq_var_store(&__data, (page), count); \
2606 if (__data < (MIN)) \
2608 else if (__data > (MAX)) \
2611 *(__PTR) = msecs_to_jiffies(__data); \
2613 *(__PTR) = __data; \
2616 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
2617 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
2619 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
2621 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
2622 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
2624 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
2625 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
2626 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
2627 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
2629 STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
2630 #undef STORE_FUNCTION
2632 #define CFQ_ATTR(name) \
2633 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
2635 static struct elv_fs_entry cfq_attrs[] = {
2637 CFQ_ATTR(fifo_expire_sync),
2638 CFQ_ATTR(fifo_expire_async),
2639 CFQ_ATTR(back_seek_max),
2640 CFQ_ATTR(back_seek_penalty),
2641 CFQ_ATTR(slice_sync),
2642 CFQ_ATTR(slice_async),
2643 CFQ_ATTR(slice_async_rq),
2644 CFQ_ATTR(slice_idle),
2645 CFQ_ATTR(low_latency),
2649 static struct elevator_type iosched_cfq = {
2651 .elevator_merge_fn = cfq_merge,
2652 .elevator_merged_fn = cfq_merged_request,
2653 .elevator_merge_req_fn = cfq_merged_requests,
2654 .elevator_allow_merge_fn = cfq_allow_merge,
2655 .elevator_dispatch_fn = cfq_dispatch_requests,
2656 .elevator_add_req_fn = cfq_insert_request,
2657 .elevator_activate_req_fn = cfq_activate_request,
2658 .elevator_deactivate_req_fn = cfq_deactivate_request,
2659 .elevator_queue_empty_fn = cfq_queue_empty,
2660 .elevator_completed_req_fn = cfq_completed_request,
2661 .elevator_former_req_fn = elv_rb_former_request,
2662 .elevator_latter_req_fn = elv_rb_latter_request,
2663 .elevator_set_req_fn = cfq_set_request,
2664 .elevator_put_req_fn = cfq_put_request,
2665 .elevator_may_queue_fn = cfq_may_queue,
2666 .elevator_init_fn = cfq_init_queue,
2667 .elevator_exit_fn = cfq_exit_queue,
2668 .trim = cfq_free_io_context,
2670 .elevator_attrs = cfq_attrs,
2671 .elevator_name = "cfq",
2672 .elevator_owner = THIS_MODULE,
2675 static int __init cfq_init(void)
2678 * could be 0 on HZ < 1000 setups
2680 if (!cfq_slice_async)
2681 cfq_slice_async = 1;
2682 if (!cfq_slice_idle)
2685 if (cfq_slab_setup())
2688 elv_register(&iosched_cfq);
2693 static void __exit cfq_exit(void)
2695 DECLARE_COMPLETION_ONSTACK(all_gone);
2696 elv_unregister(&iosched_cfq);
2697 ioc_gone = &all_gone;
2698 /* ioc_gone's update must be visible before reading ioc_count */
2702 * this also protects us from entering cfq_slab_kill() with
2703 * pending RCU callbacks
2705 if (elv_ioc_count_read(cfq_ioc_count))
2706 wait_for_completion(&all_gone);
2710 module_init(cfq_init);
2711 module_exit(cfq_exit);
2713 MODULE_AUTHOR("Jens Axboe");
2714 MODULE_LICENSE("GPL");
2715 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");