2 * CFQ, or complete fairness queueing, disk scheduler.
4 * Based on ideas from a previously unfinished io
5 * scheduler (round robin per-process disk scheduling) and Andrea Arcangeli.
7 * Copyright (C) 2003 Jens Axboe <axboe@kernel.dk>
9 #include <linux/module.h>
10 #include <linux/blkdev.h>
11 #include <linux/elevator.h>
12 #include <linux/rbtree.h>
13 #include <linux/ioprio.h>
14 #include <linux/blktrace_api.h>
19 /* max queue in one round of service */
20 static const int cfq_quantum = 4;
21 static const int cfq_fifo_expire[2] = { HZ / 4, HZ / 8 };
22 /* maximum backwards seek, in KiB */
23 static const int cfq_back_max = 16 * 1024;
24 /* penalty of a backwards seek */
25 static const int cfq_back_penalty = 2;
26 static const int cfq_slice_sync = HZ / 10;
27 static int cfq_slice_async = HZ / 25;
28 static const int cfq_slice_async_rq = 2;
29 static int cfq_slice_idle = HZ / 125;
32 * offset from end of service tree
34 #define CFQ_IDLE_DELAY (HZ / 5)
37 * below this threshold, we consider thinktime immediate
39 #define CFQ_MIN_TT (2)
41 #define CFQ_SLICE_SCALE (5)
42 #define CFQ_HW_QUEUE_MIN (5)
45 ((struct cfq_io_context *) (rq)->elevator_private)
46 #define RQ_CFQQ(rq) (struct cfq_queue *) ((rq)->elevator_private2)
48 static struct kmem_cache *cfq_pool;
49 static struct kmem_cache *cfq_ioc_pool;
51 static DEFINE_PER_CPU(unsigned long, ioc_count);
52 static struct completion *ioc_gone;
53 static DEFINE_SPINLOCK(ioc_gone_lock);
55 #define CFQ_PRIO_LISTS IOPRIO_BE_NR
56 #define cfq_class_idle(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
57 #define cfq_class_rt(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
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 block device queue structure
77 struct request_queue *queue;
80 * rr list of queues with requests and the count of them
82 struct cfq_rb_root service_tree;
83 unsigned int busy_queues;
85 * Used to track any pending rt requests so we can pre-empt current
86 * non-RT cfqq in service when this value is non-zero.
88 unsigned int busy_rt_queues;
94 * queue-depth detection
99 int rq_in_driver_peak;
102 * idle window management
104 struct timer_list idle_slice_timer;
105 struct work_struct unplug_work;
107 struct cfq_queue *active_queue;
108 struct cfq_io_context *active_cic;
111 * async queue for each priority case
113 struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
114 struct cfq_queue *async_idle_cfqq;
116 sector_t last_position;
117 unsigned long last_end_request;
120 * tunables, see top of file
122 unsigned int cfq_quantum;
123 unsigned int cfq_fifo_expire[2];
124 unsigned int cfq_back_penalty;
125 unsigned int cfq_back_max;
126 unsigned int cfq_slice[2];
127 unsigned int cfq_slice_async_rq;
128 unsigned int cfq_slice_idle;
130 struct list_head cic_list;
134 * Per process-grouping structure
137 /* reference count */
139 /* various state flags, see below */
141 /* parent cfq_data */
142 struct cfq_data *cfqd;
143 /* service_tree member */
144 struct rb_node rb_node;
145 /* service_tree key */
146 unsigned long rb_key;
147 /* sorted list of pending requests */
148 struct rb_root sort_list;
149 /* if fifo isn't expired, next request to serve */
150 struct request *next_rq;
151 /* requests queued in sort_list */
153 /* currently allocated requests */
155 /* fifo list of requests in sort_list */
156 struct list_head fifo;
158 unsigned long slice_end;
160 unsigned int slice_dispatch;
162 /* pending metadata requests */
164 /* number of requests that are on the dispatch list or inside driver */
167 /* io prio of this group */
168 unsigned short ioprio, org_ioprio;
169 unsigned short ioprio_class, org_ioprio_class;
174 enum cfqq_state_flags {
175 CFQ_CFQQ_FLAG_on_rr = 0, /* on round-robin busy list */
176 CFQ_CFQQ_FLAG_wait_request, /* waiting for a request */
177 CFQ_CFQQ_FLAG_must_dispatch, /* must be allowed a dispatch */
178 CFQ_CFQQ_FLAG_must_alloc, /* must be allowed rq alloc */
179 CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
180 CFQ_CFQQ_FLAG_fifo_expire, /* FIFO checked in this slice */
181 CFQ_CFQQ_FLAG_idle_window, /* slice idling enabled */
182 CFQ_CFQQ_FLAG_prio_changed, /* task priority has changed */
183 CFQ_CFQQ_FLAG_slice_new, /* no requests dispatched in slice */
184 CFQ_CFQQ_FLAG_sync, /* synchronous queue */
187 #define CFQ_CFQQ_FNS(name) \
188 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \
190 (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name); \
192 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \
194 (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \
196 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \
198 return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \
202 CFQ_CFQQ_FNS(wait_request);
203 CFQ_CFQQ_FNS(must_dispatch);
204 CFQ_CFQQ_FNS(must_alloc);
205 CFQ_CFQQ_FNS(must_alloc_slice);
206 CFQ_CFQQ_FNS(fifo_expire);
207 CFQ_CFQQ_FNS(idle_window);
208 CFQ_CFQQ_FNS(prio_changed);
209 CFQ_CFQQ_FNS(slice_new);
213 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
214 blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args)
215 #define cfq_log(cfqd, fmt, args...) \
216 blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
218 static void cfq_dispatch_insert(struct request_queue *, struct request *);
219 static struct cfq_queue *cfq_get_queue(struct cfq_data *, int,
220 struct io_context *, gfp_t);
221 static struct cfq_io_context *cfq_cic_lookup(struct cfq_data *,
222 struct io_context *);
224 static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_context *cic,
227 return cic->cfqq[!!is_sync];
230 static inline void cic_set_cfqq(struct cfq_io_context *cic,
231 struct cfq_queue *cfqq, int is_sync)
233 cic->cfqq[!!is_sync] = cfqq;
237 * We regard a request as SYNC, if it's either a read or has the SYNC bit
238 * set (in which case it could also be direct WRITE).
240 static inline int cfq_bio_sync(struct bio *bio)
242 if (bio_data_dir(bio) == READ || bio_sync(bio))
249 * scheduler run of queue, if there are requests pending and no one in the
250 * driver that will restart queueing
252 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
254 if (cfqd->busy_queues) {
255 cfq_log(cfqd, "schedule dispatch");
256 kblockd_schedule_work(cfqd->queue, &cfqd->unplug_work);
260 static int cfq_queue_empty(struct request_queue *q)
262 struct cfq_data *cfqd = q->elevator->elevator_data;
264 return !cfqd->busy_queues;
268 * Scale schedule slice based on io priority. Use the sync time slice only
269 * if a queue is marked sync and has sync io queued. A sync queue with async
270 * io only, should not get full sync slice length.
272 static inline int cfq_prio_slice(struct cfq_data *cfqd, int sync,
275 const int base_slice = cfqd->cfq_slice[sync];
277 WARN_ON(prio >= IOPRIO_BE_NR);
279 return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
283 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
285 return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
289 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
291 cfqq->slice_end = cfq_prio_to_slice(cfqd, cfqq) + jiffies;
292 cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
296 * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
297 * isn't valid until the first request from the dispatch is activated
298 * and the slice time set.
300 static inline int cfq_slice_used(struct cfq_queue *cfqq)
302 if (cfq_cfqq_slice_new(cfqq))
304 if (time_before(jiffies, cfqq->slice_end))
311 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
312 * We choose the request that is closest to the head right now. Distance
313 * behind the head is penalized and only allowed to a certain extent.
315 static struct request *
316 cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2)
318 sector_t last, s1, s2, d1 = 0, d2 = 0;
319 unsigned long back_max;
320 #define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */
321 #define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */
322 unsigned wrap = 0; /* bit mask: requests behind the disk head? */
324 if (rq1 == NULL || rq1 == rq2)
329 if (rq_is_sync(rq1) && !rq_is_sync(rq2))
331 else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
333 if (rq_is_meta(rq1) && !rq_is_meta(rq2))
335 else if (rq_is_meta(rq2) && !rq_is_meta(rq1))
341 last = cfqd->last_position;
344 * by definition, 1KiB is 2 sectors
346 back_max = cfqd->cfq_back_max * 2;
349 * Strict one way elevator _except_ in the case where we allow
350 * short backward seeks which are biased as twice the cost of a
351 * similar forward seek.
355 else if (s1 + back_max >= last)
356 d1 = (last - s1) * cfqd->cfq_back_penalty;
358 wrap |= CFQ_RQ1_WRAP;
362 else if (s2 + back_max >= last)
363 d2 = (last - s2) * cfqd->cfq_back_penalty;
365 wrap |= CFQ_RQ2_WRAP;
367 /* Found required data */
370 * By doing switch() on the bit mask "wrap" we avoid having to
371 * check two variables for all permutations: --> faster!
374 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
390 case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
393 * Since both rqs are wrapped,
394 * start with the one that's further behind head
395 * (--> only *one* back seek required),
396 * since back seek takes more time than forward.
406 * The below is leftmost cache rbtree addon
408 static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
411 root->left = rb_first(&root->rb);
414 return rb_entry(root->left, struct cfq_queue, rb_node);
419 static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
424 rb_erase(n, &root->rb);
429 * would be nice to take fifo expire time into account as well
431 static struct request *
432 cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
433 struct request *last)
435 struct rb_node *rbnext = rb_next(&last->rb_node);
436 struct rb_node *rbprev = rb_prev(&last->rb_node);
437 struct request *next = NULL, *prev = NULL;
439 BUG_ON(RB_EMPTY_NODE(&last->rb_node));
442 prev = rb_entry_rq(rbprev);
445 next = rb_entry_rq(rbnext);
447 rbnext = rb_first(&cfqq->sort_list);
448 if (rbnext && rbnext != &last->rb_node)
449 next = rb_entry_rq(rbnext);
452 return cfq_choose_req(cfqd, next, prev);
455 static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
456 struct cfq_queue *cfqq)
459 * just an approximation, should be ok.
461 return (cfqd->busy_queues - 1) * (cfq_prio_slice(cfqd, 1, 0) -
462 cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
466 * The cfqd->service_tree holds all pending cfq_queue's that have
467 * requests waiting to be processed. It is sorted in the order that
468 * we will service the queues.
470 static void cfq_service_tree_add(struct cfq_data *cfqd,
471 struct cfq_queue *cfqq, int add_front)
473 struct rb_node **p, *parent;
474 struct cfq_queue *__cfqq;
475 unsigned long rb_key;
478 if (cfq_class_idle(cfqq)) {
479 rb_key = CFQ_IDLE_DELAY;
480 parent = rb_last(&cfqd->service_tree.rb);
481 if (parent && parent != &cfqq->rb_node) {
482 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
483 rb_key += __cfqq->rb_key;
486 } else if (!add_front) {
487 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
488 rb_key += cfqq->slice_resid;
489 cfqq->slice_resid = 0;
493 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
495 * same position, nothing more to do
497 if (rb_key == cfqq->rb_key)
500 cfq_rb_erase(&cfqq->rb_node, &cfqd->service_tree);
505 p = &cfqd->service_tree.rb.rb_node;
510 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
513 * sort RT queues first, we always want to give
514 * preference to them. IDLE queues goes to the back.
515 * after that, sort on the next service time.
517 if (cfq_class_rt(cfqq) > cfq_class_rt(__cfqq))
519 else if (cfq_class_rt(cfqq) < cfq_class_rt(__cfqq))
521 else if (cfq_class_idle(cfqq) < cfq_class_idle(__cfqq))
523 else if (cfq_class_idle(cfqq) > cfq_class_idle(__cfqq))
525 else if (rb_key < __cfqq->rb_key)
530 if (n == &(*p)->rb_right)
537 cfqd->service_tree.left = &cfqq->rb_node;
539 cfqq->rb_key = rb_key;
540 rb_link_node(&cfqq->rb_node, parent, p);
541 rb_insert_color(&cfqq->rb_node, &cfqd->service_tree.rb);
545 * Update cfqq's position in the service tree.
547 static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
550 * Resorting requires the cfqq to be on the RR list already.
552 if (cfq_cfqq_on_rr(cfqq))
553 cfq_service_tree_add(cfqd, cfqq, 0);
557 * add to busy list of queues for service, trying to be fair in ordering
558 * the pending list according to last request service
560 static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
562 cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
563 BUG_ON(cfq_cfqq_on_rr(cfqq));
564 cfq_mark_cfqq_on_rr(cfqq);
566 if (cfq_class_rt(cfqq))
567 cfqd->busy_rt_queues++;
569 cfq_resort_rr_list(cfqd, cfqq);
573 * Called when the cfqq no longer has requests pending, remove it from
576 static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
578 cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
579 BUG_ON(!cfq_cfqq_on_rr(cfqq));
580 cfq_clear_cfqq_on_rr(cfqq);
582 if (!RB_EMPTY_NODE(&cfqq->rb_node))
583 cfq_rb_erase(&cfqq->rb_node, &cfqd->service_tree);
585 BUG_ON(!cfqd->busy_queues);
587 if (cfq_class_rt(cfqq))
588 cfqd->busy_rt_queues--;
592 * rb tree support functions
594 static void cfq_del_rq_rb(struct request *rq)
596 struct cfq_queue *cfqq = RQ_CFQQ(rq);
597 struct cfq_data *cfqd = cfqq->cfqd;
598 const int sync = rq_is_sync(rq);
600 BUG_ON(!cfqq->queued[sync]);
601 cfqq->queued[sync]--;
603 elv_rb_del(&cfqq->sort_list, rq);
605 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
606 cfq_del_cfqq_rr(cfqd, cfqq);
609 static void cfq_add_rq_rb(struct request *rq)
611 struct cfq_queue *cfqq = RQ_CFQQ(rq);
612 struct cfq_data *cfqd = cfqq->cfqd;
613 struct request *__alias;
615 cfqq->queued[rq_is_sync(rq)]++;
618 * looks a little odd, but the first insert might return an alias.
619 * if that happens, put the alias on the dispatch list
621 while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL)
622 cfq_dispatch_insert(cfqd->queue, __alias);
624 if (!cfq_cfqq_on_rr(cfqq))
625 cfq_add_cfqq_rr(cfqd, cfqq);
628 * check if this request is a better next-serve candidate
630 cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq);
631 BUG_ON(!cfqq->next_rq);
634 static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
636 elv_rb_del(&cfqq->sort_list, rq);
637 cfqq->queued[rq_is_sync(rq)]--;
641 static struct request *
642 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
644 struct task_struct *tsk = current;
645 struct cfq_io_context *cic;
646 struct cfq_queue *cfqq;
648 cic = cfq_cic_lookup(cfqd, tsk->io_context);
652 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
654 sector_t sector = bio->bi_sector + bio_sectors(bio);
656 return elv_rb_find(&cfqq->sort_list, sector);
662 static void cfq_activate_request(struct request_queue *q, struct request *rq)
664 struct cfq_data *cfqd = q->elevator->elevator_data;
666 cfqd->rq_in_driver++;
667 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
670 cfqd->last_position = rq->hard_sector + rq->hard_nr_sectors;
673 static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
675 struct cfq_data *cfqd = q->elevator->elevator_data;
677 WARN_ON(!cfqd->rq_in_driver);
678 cfqd->rq_in_driver--;
679 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
683 static void cfq_remove_request(struct request *rq)
685 struct cfq_queue *cfqq = RQ_CFQQ(rq);
687 if (cfqq->next_rq == rq)
688 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
690 list_del_init(&rq->queuelist);
693 cfqq->cfqd->rq_queued--;
694 if (rq_is_meta(rq)) {
695 WARN_ON(!cfqq->meta_pending);
696 cfqq->meta_pending--;
700 static int cfq_merge(struct request_queue *q, struct request **req,
703 struct cfq_data *cfqd = q->elevator->elevator_data;
704 struct request *__rq;
706 __rq = cfq_find_rq_fmerge(cfqd, bio);
707 if (__rq && elv_rq_merge_ok(__rq, bio)) {
709 return ELEVATOR_FRONT_MERGE;
712 return ELEVATOR_NO_MERGE;
715 static void cfq_merged_request(struct request_queue *q, struct request *req,
718 if (type == ELEVATOR_FRONT_MERGE) {
719 struct cfq_queue *cfqq = RQ_CFQQ(req);
721 cfq_reposition_rq_rb(cfqq, req);
726 cfq_merged_requests(struct request_queue *q, struct request *rq,
727 struct request *next)
730 * reposition in fifo if next is older than rq
732 if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
733 time_before(next->start_time, rq->start_time))
734 list_move(&rq->queuelist, &next->queuelist);
736 cfq_remove_request(next);
739 static int cfq_allow_merge(struct request_queue *q, struct request *rq,
742 struct cfq_data *cfqd = q->elevator->elevator_data;
743 struct cfq_io_context *cic;
744 struct cfq_queue *cfqq;
747 * Disallow merge of a sync bio into an async request.
749 if (cfq_bio_sync(bio) && !rq_is_sync(rq))
753 * Lookup the cfqq that this bio will be queued with. Allow
754 * merge only if rq is queued there.
756 cic = cfq_cic_lookup(cfqd, current->io_context);
760 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
761 if (cfqq == RQ_CFQQ(rq))
767 static void __cfq_set_active_queue(struct cfq_data *cfqd,
768 struct cfq_queue *cfqq)
771 cfq_log_cfqq(cfqd, cfqq, "set_active");
773 cfqq->slice_dispatch = 0;
775 cfq_clear_cfqq_wait_request(cfqq);
776 cfq_clear_cfqq_must_dispatch(cfqq);
777 cfq_clear_cfqq_must_alloc_slice(cfqq);
778 cfq_clear_cfqq_fifo_expire(cfqq);
779 cfq_mark_cfqq_slice_new(cfqq);
781 del_timer(&cfqd->idle_slice_timer);
784 cfqd->active_queue = cfqq;
788 * current cfqq expired its slice (or was too idle), select new one
791 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
794 cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
796 if (cfq_cfqq_wait_request(cfqq))
797 del_timer(&cfqd->idle_slice_timer);
799 cfq_clear_cfqq_wait_request(cfqq);
802 * store what was left of this slice, if the queue idled/timed out
804 if (timed_out && !cfq_cfqq_slice_new(cfqq)) {
805 cfqq->slice_resid = cfqq->slice_end - jiffies;
806 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
809 cfq_resort_rr_list(cfqd, cfqq);
811 if (cfqq == cfqd->active_queue)
812 cfqd->active_queue = NULL;
814 if (cfqd->active_cic) {
815 put_io_context(cfqd->active_cic->ioc);
816 cfqd->active_cic = NULL;
820 static inline void cfq_slice_expired(struct cfq_data *cfqd, int timed_out)
822 struct cfq_queue *cfqq = cfqd->active_queue;
825 __cfq_slice_expired(cfqd, cfqq, timed_out);
829 * Get next queue for service. Unless we have a queue preemption,
830 * we'll simply select the first cfqq in the service tree.
832 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
834 if (RB_EMPTY_ROOT(&cfqd->service_tree.rb))
837 return cfq_rb_first(&cfqd->service_tree);
841 * Get and set a new active queue for service.
843 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd)
845 struct cfq_queue *cfqq;
847 cfqq = cfq_get_next_queue(cfqd);
848 __cfq_set_active_queue(cfqd, cfqq);
852 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
855 if (rq->sector >= cfqd->last_position)
856 return rq->sector - cfqd->last_position;
858 return cfqd->last_position - rq->sector;
861 static inline int cfq_rq_close(struct cfq_data *cfqd, struct request *rq)
863 struct cfq_io_context *cic = cfqd->active_cic;
865 if (!sample_valid(cic->seek_samples))
868 return cfq_dist_from_last(cfqd, rq) <= cic->seek_mean;
871 static int cfq_close_cooperator(struct cfq_data *cfq_data,
872 struct cfq_queue *cfqq)
875 * We should notice if some of the queues are cooperating, eg
876 * working closely on the same area of the disk. In that case,
877 * we can group them together and don't waste time idling.
882 #define CIC_SEEKY(cic) ((cic)->seek_mean > (8 * 1024))
884 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
886 struct cfq_queue *cfqq = cfqd->active_queue;
887 struct cfq_io_context *cic;
891 * SSD device without seek penalty, disable idling. But only do so
892 * for devices that support queuing, otherwise we still have a problem
893 * with sync vs async workloads.
895 if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
898 WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
899 WARN_ON(cfq_cfqq_slice_new(cfqq));
902 * idle is disabled, either manually or by past process history
904 if (!cfqd->cfq_slice_idle || !cfq_cfqq_idle_window(cfqq))
908 * still requests with the driver, don't idle
910 if (cfqd->rq_in_driver)
914 * task has exited, don't wait
916 cic = cfqd->active_cic;
917 if (!cic || !atomic_read(&cic->ioc->nr_tasks))
921 * See if this prio level has a good candidate
923 if (cfq_close_cooperator(cfqd, cfqq) &&
924 (sample_valid(cic->ttime_samples) && cic->ttime_mean > 2))
927 cfq_mark_cfqq_wait_request(cfqq);
930 * we don't want to idle for seeks, but we do want to allow
931 * fair distribution of slice time for a process doing back-to-back
932 * seeks. so allow a little bit of time for him to submit a new rq
934 sl = cfqd->cfq_slice_idle;
935 if (sample_valid(cic->seek_samples) && CIC_SEEKY(cic))
936 sl = min(sl, msecs_to_jiffies(CFQ_MIN_TT));
938 mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
939 cfq_log(cfqd, "arm_idle: %lu", sl);
943 * Move request from internal lists to the request queue dispatch list.
945 static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
947 struct cfq_data *cfqd = q->elevator->elevator_data;
948 struct cfq_queue *cfqq = RQ_CFQQ(rq);
950 cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
952 cfq_remove_request(rq);
954 elv_dispatch_sort(q, rq);
956 if (cfq_cfqq_sync(cfqq))
961 * return expired entry, or NULL to just start from scratch in rbtree
963 static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
965 struct cfq_data *cfqd = cfqq->cfqd;
969 if (cfq_cfqq_fifo_expire(cfqq))
972 cfq_mark_cfqq_fifo_expire(cfqq);
974 if (list_empty(&cfqq->fifo))
977 fifo = cfq_cfqq_sync(cfqq);
978 rq = rq_entry_fifo(cfqq->fifo.next);
980 if (time_before(jiffies, rq->start_time + cfqd->cfq_fifo_expire[fifo]))
983 cfq_log_cfqq(cfqd, cfqq, "fifo=%p", rq);
988 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
990 const int base_rq = cfqd->cfq_slice_async_rq;
992 WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
994 return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
998 * Select a queue for service. If we have a current active queue,
999 * check whether to continue servicing it, or retrieve and set a new one.
1001 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
1003 struct cfq_queue *cfqq;
1005 cfqq = cfqd->active_queue;
1010 * The active queue has run out of time, expire it and select new.
1012 if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq))
1016 * If we have a RT cfqq waiting, then we pre-empt the current non-rt
1019 if (!cfq_class_rt(cfqq) && cfqd->busy_rt_queues) {
1021 * We simulate this as cfqq timed out so that it gets to bank
1022 * the remaining of its time slice.
1024 cfq_log_cfqq(cfqd, cfqq, "preempt");
1025 cfq_slice_expired(cfqd, 1);
1030 * The active queue has requests and isn't expired, allow it to
1033 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
1037 * No requests pending. If the active queue still has requests in
1038 * flight or is idling for a new request, allow either of these
1039 * conditions to happen (or time out) before selecting a new queue.
1041 if (timer_pending(&cfqd->idle_slice_timer) ||
1042 (cfqq->dispatched && cfq_cfqq_idle_window(cfqq))) {
1048 cfq_slice_expired(cfqd, 0);
1050 cfqq = cfq_set_active_queue(cfqd);
1055 static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
1059 while (cfqq->next_rq) {
1060 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
1064 BUG_ON(!list_empty(&cfqq->fifo));
1069 * Drain our current requests. Used for barriers and when switching
1070 * io schedulers on-the-fly.
1072 static int cfq_forced_dispatch(struct cfq_data *cfqd)
1074 struct cfq_queue *cfqq;
1077 while ((cfqq = cfq_rb_first(&cfqd->service_tree)) != NULL)
1078 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
1080 cfq_slice_expired(cfqd, 0);
1082 BUG_ON(cfqd->busy_queues);
1084 cfq_log(cfqd, "forced_dispatch=%d\n", dispatched);
1089 * Dispatch a request from cfqq, moving them to the request queue
1092 static void cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1096 BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
1099 * follow expired path, else get first next available
1101 rq = cfq_check_fifo(cfqq);
1106 * insert request into driver dispatch list
1108 cfq_dispatch_insert(cfqd->queue, rq);
1110 if (!cfqd->active_cic) {
1111 struct cfq_io_context *cic = RQ_CIC(rq);
1113 atomic_inc(&cic->ioc->refcount);
1114 cfqd->active_cic = cic;
1119 * Find the cfqq that we need to service and move a request from that to the
1122 static int cfq_dispatch_requests(struct request_queue *q, int force)
1124 struct cfq_data *cfqd = q->elevator->elevator_data;
1125 struct cfq_queue *cfqq;
1126 unsigned int max_dispatch;
1128 if (!cfqd->busy_queues)
1131 if (unlikely(force))
1132 return cfq_forced_dispatch(cfqd);
1134 cfqq = cfq_select_queue(cfqd);
1139 * If this is an async queue and we have sync IO in flight, let it wait
1141 if (cfqd->sync_flight && !cfq_cfqq_sync(cfqq))
1144 max_dispatch = cfqd->cfq_quantum;
1145 if (cfq_class_idle(cfqq))
1149 * Does this cfqq already have too much IO in flight?
1151 if (cfqq->dispatched >= max_dispatch) {
1153 * idle queue must always only have a single IO in flight
1155 if (cfq_class_idle(cfqq))
1159 * We have other queues, don't allow more IO from this one
1161 if (cfqd->busy_queues > 1)
1165 * we are the only queue, allow up to 4 times of 'quantum'
1167 if (cfqq->dispatched >= 4 * max_dispatch)
1172 * Dispatch a request from this cfqq
1174 cfq_dispatch_request(cfqd, cfqq);
1175 cfqq->slice_dispatch++;
1176 cfq_clear_cfqq_must_dispatch(cfqq);
1179 * expire an async queue immediately if it has used up its slice. idle
1180 * queue always expire after 1 dispatch round.
1182 if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
1183 cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
1184 cfq_class_idle(cfqq))) {
1185 cfqq->slice_end = jiffies + 1;
1186 cfq_slice_expired(cfqd, 0);
1189 cfq_log(cfqd, "dispatched a request");
1194 * task holds one reference to the queue, dropped when task exits. each rq
1195 * in-flight on this queue also holds a reference, dropped when rq is freed.
1197 * queue lock must be held here.
1199 static void cfq_put_queue(struct cfq_queue *cfqq)
1201 struct cfq_data *cfqd = cfqq->cfqd;
1203 BUG_ON(atomic_read(&cfqq->ref) <= 0);
1205 if (!atomic_dec_and_test(&cfqq->ref))
1208 cfq_log_cfqq(cfqd, cfqq, "put_queue");
1209 BUG_ON(rb_first(&cfqq->sort_list));
1210 BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
1211 BUG_ON(cfq_cfqq_on_rr(cfqq));
1213 if (unlikely(cfqd->active_queue == cfqq)) {
1214 __cfq_slice_expired(cfqd, cfqq, 0);
1215 cfq_schedule_dispatch(cfqd);
1218 kmem_cache_free(cfq_pool, cfqq);
1222 * Must always be called with the rcu_read_lock() held
1225 __call_for_each_cic(struct io_context *ioc,
1226 void (*func)(struct io_context *, struct cfq_io_context *))
1228 struct cfq_io_context *cic;
1229 struct hlist_node *n;
1231 hlist_for_each_entry_rcu(cic, n, &ioc->cic_list, cic_list)
1236 * Call func for each cic attached to this ioc.
1239 call_for_each_cic(struct io_context *ioc,
1240 void (*func)(struct io_context *, struct cfq_io_context *))
1243 __call_for_each_cic(ioc, func);
1247 static void cfq_cic_free_rcu(struct rcu_head *head)
1249 struct cfq_io_context *cic;
1251 cic = container_of(head, struct cfq_io_context, rcu_head);
1253 kmem_cache_free(cfq_ioc_pool, cic);
1254 elv_ioc_count_dec(ioc_count);
1258 * CFQ scheduler is exiting, grab exit lock and check
1259 * the pending io context count. If it hits zero,
1260 * complete ioc_gone and set it back to NULL
1262 spin_lock(&ioc_gone_lock);
1263 if (ioc_gone && !elv_ioc_count_read(ioc_count)) {
1267 spin_unlock(&ioc_gone_lock);
1271 static void cfq_cic_free(struct cfq_io_context *cic)
1273 call_rcu(&cic->rcu_head, cfq_cic_free_rcu);
1276 static void cic_free_func(struct io_context *ioc, struct cfq_io_context *cic)
1278 unsigned long flags;
1280 BUG_ON(!cic->dead_key);
1282 spin_lock_irqsave(&ioc->lock, flags);
1283 radix_tree_delete(&ioc->radix_root, cic->dead_key);
1284 hlist_del_rcu(&cic->cic_list);
1285 spin_unlock_irqrestore(&ioc->lock, flags);
1291 * Must be called with rcu_read_lock() held or preemption otherwise disabled.
1292 * Only two callers of this - ->dtor() which is called with the rcu_read_lock(),
1293 * and ->trim() which is called with the task lock held
1295 static void cfq_free_io_context(struct io_context *ioc)
1298 * ioc->refcount is zero here, or we are called from elv_unregister(),
1299 * so no more cic's are allowed to be linked into this ioc. So it
1300 * should be ok to iterate over the known list, we will see all cic's
1301 * since no new ones are added.
1303 __call_for_each_cic(ioc, cic_free_func);
1306 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1308 if (unlikely(cfqq == cfqd->active_queue)) {
1309 __cfq_slice_expired(cfqd, cfqq, 0);
1310 cfq_schedule_dispatch(cfqd);
1313 cfq_put_queue(cfqq);
1316 static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
1317 struct cfq_io_context *cic)
1319 struct io_context *ioc = cic->ioc;
1321 list_del_init(&cic->queue_list);
1324 * Make sure key == NULL is seen for dead queues
1327 cic->dead_key = (unsigned long) cic->key;
1330 if (ioc->ioc_data == cic)
1331 rcu_assign_pointer(ioc->ioc_data, NULL);
1333 if (cic->cfqq[BLK_RW_ASYNC]) {
1334 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]);
1335 cic->cfqq[BLK_RW_ASYNC] = NULL;
1338 if (cic->cfqq[BLK_RW_SYNC]) {
1339 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]);
1340 cic->cfqq[BLK_RW_SYNC] = NULL;
1344 static void cfq_exit_single_io_context(struct io_context *ioc,
1345 struct cfq_io_context *cic)
1347 struct cfq_data *cfqd = cic->key;
1350 struct request_queue *q = cfqd->queue;
1351 unsigned long flags;
1353 spin_lock_irqsave(q->queue_lock, flags);
1356 * Ensure we get a fresh copy of the ->key to prevent
1357 * race between exiting task and queue
1359 smp_read_barrier_depends();
1361 __cfq_exit_single_io_context(cfqd, cic);
1363 spin_unlock_irqrestore(q->queue_lock, flags);
1368 * The process that ioc belongs to has exited, we need to clean up
1369 * and put the internal structures we have that belongs to that process.
1371 static void cfq_exit_io_context(struct io_context *ioc)
1373 call_for_each_cic(ioc, cfq_exit_single_io_context);
1376 static struct cfq_io_context *
1377 cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
1379 struct cfq_io_context *cic;
1381 cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO,
1384 cic->last_end_request = jiffies;
1385 INIT_LIST_HEAD(&cic->queue_list);
1386 INIT_HLIST_NODE(&cic->cic_list);
1387 cic->dtor = cfq_free_io_context;
1388 cic->exit = cfq_exit_io_context;
1389 elv_ioc_count_inc(ioc_count);
1395 static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc)
1397 struct task_struct *tsk = current;
1400 if (!cfq_cfqq_prio_changed(cfqq))
1403 ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio);
1404 switch (ioprio_class) {
1406 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
1407 case IOPRIO_CLASS_NONE:
1409 * no prio set, inherit CPU scheduling settings
1411 cfqq->ioprio = task_nice_ioprio(tsk);
1412 cfqq->ioprio_class = task_nice_ioclass(tsk);
1414 case IOPRIO_CLASS_RT:
1415 cfqq->ioprio = task_ioprio(ioc);
1416 cfqq->ioprio_class = IOPRIO_CLASS_RT;
1418 case IOPRIO_CLASS_BE:
1419 cfqq->ioprio = task_ioprio(ioc);
1420 cfqq->ioprio_class = IOPRIO_CLASS_BE;
1422 case IOPRIO_CLASS_IDLE:
1423 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
1425 cfq_clear_cfqq_idle_window(cfqq);
1430 * keep track of original prio settings in case we have to temporarily
1431 * elevate the priority of this queue
1433 cfqq->org_ioprio = cfqq->ioprio;
1434 cfqq->org_ioprio_class = cfqq->ioprio_class;
1435 cfq_clear_cfqq_prio_changed(cfqq);
1438 static void changed_ioprio(struct io_context *ioc, struct cfq_io_context *cic)
1440 struct cfq_data *cfqd = cic->key;
1441 struct cfq_queue *cfqq;
1442 unsigned long flags;
1444 if (unlikely(!cfqd))
1447 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1449 cfqq = cic->cfqq[BLK_RW_ASYNC];
1451 struct cfq_queue *new_cfqq;
1452 new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic->ioc,
1455 cic->cfqq[BLK_RW_ASYNC] = new_cfqq;
1456 cfq_put_queue(cfqq);
1460 cfqq = cic->cfqq[BLK_RW_SYNC];
1462 cfq_mark_cfqq_prio_changed(cfqq);
1464 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1467 static void cfq_ioc_set_ioprio(struct io_context *ioc)
1469 call_for_each_cic(ioc, changed_ioprio);
1470 ioc->ioprio_changed = 0;
1473 static struct cfq_queue *
1474 cfq_find_alloc_queue(struct cfq_data *cfqd, int is_sync,
1475 struct io_context *ioc, gfp_t gfp_mask)
1477 struct cfq_queue *cfqq, *new_cfqq = NULL;
1478 struct cfq_io_context *cic;
1481 cic = cfq_cic_lookup(cfqd, ioc);
1482 /* cic always exists here */
1483 cfqq = cic_to_cfqq(cic, is_sync);
1489 } else if (gfp_mask & __GFP_WAIT) {
1491 * Inform the allocator of the fact that we will
1492 * just repeat this allocation if it fails, to allow
1493 * the allocator to do whatever it needs to attempt to
1496 spin_unlock_irq(cfqd->queue->queue_lock);
1497 new_cfqq = kmem_cache_alloc_node(cfq_pool,
1498 gfp_mask | __GFP_NOFAIL | __GFP_ZERO,
1500 spin_lock_irq(cfqd->queue->queue_lock);
1503 cfqq = kmem_cache_alloc_node(cfq_pool,
1504 gfp_mask | __GFP_ZERO,
1510 RB_CLEAR_NODE(&cfqq->rb_node);
1511 INIT_LIST_HEAD(&cfqq->fifo);
1513 atomic_set(&cfqq->ref, 0);
1516 cfq_mark_cfqq_prio_changed(cfqq);
1518 cfq_init_prio_data(cfqq, ioc);
1521 if (!cfq_class_idle(cfqq))
1522 cfq_mark_cfqq_idle_window(cfqq);
1523 cfq_mark_cfqq_sync(cfqq);
1525 cfqq->pid = current->pid;
1526 cfq_log_cfqq(cfqd, cfqq, "alloced");
1530 kmem_cache_free(cfq_pool, new_cfqq);
1533 WARN_ON((gfp_mask & __GFP_WAIT) && !cfqq);
1537 static struct cfq_queue **
1538 cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
1540 switch (ioprio_class) {
1541 case IOPRIO_CLASS_RT:
1542 return &cfqd->async_cfqq[0][ioprio];
1543 case IOPRIO_CLASS_BE:
1544 return &cfqd->async_cfqq[1][ioprio];
1545 case IOPRIO_CLASS_IDLE:
1546 return &cfqd->async_idle_cfqq;
1552 static struct cfq_queue *
1553 cfq_get_queue(struct cfq_data *cfqd, int is_sync, struct io_context *ioc,
1556 const int ioprio = task_ioprio(ioc);
1557 const int ioprio_class = task_ioprio_class(ioc);
1558 struct cfq_queue **async_cfqq = NULL;
1559 struct cfq_queue *cfqq = NULL;
1562 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
1567 cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask);
1573 * pin the queue now that it's allocated, scheduler exit will prune it
1575 if (!is_sync && !(*async_cfqq)) {
1576 atomic_inc(&cfqq->ref);
1580 atomic_inc(&cfqq->ref);
1585 * We drop cfq io contexts lazily, so we may find a dead one.
1588 cfq_drop_dead_cic(struct cfq_data *cfqd, struct io_context *ioc,
1589 struct cfq_io_context *cic)
1591 unsigned long flags;
1593 WARN_ON(!list_empty(&cic->queue_list));
1595 spin_lock_irqsave(&ioc->lock, flags);
1597 BUG_ON(ioc->ioc_data == cic);
1599 radix_tree_delete(&ioc->radix_root, (unsigned long) cfqd);
1600 hlist_del_rcu(&cic->cic_list);
1601 spin_unlock_irqrestore(&ioc->lock, flags);
1606 static struct cfq_io_context *
1607 cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc)
1609 struct cfq_io_context *cic;
1610 unsigned long flags;
1619 * we maintain a last-hit cache, to avoid browsing over the tree
1621 cic = rcu_dereference(ioc->ioc_data);
1622 if (cic && cic->key == cfqd) {
1628 cic = radix_tree_lookup(&ioc->radix_root, (unsigned long) cfqd);
1632 /* ->key must be copied to avoid race with cfq_exit_queue() */
1635 cfq_drop_dead_cic(cfqd, ioc, cic);
1640 spin_lock_irqsave(&ioc->lock, flags);
1641 rcu_assign_pointer(ioc->ioc_data, cic);
1642 spin_unlock_irqrestore(&ioc->lock, flags);
1650 * Add cic into ioc, using cfqd as the search key. This enables us to lookup
1651 * the process specific cfq io context when entered from the block layer.
1652 * Also adds the cic to a per-cfqd list, used when this queue is removed.
1654 static int cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
1655 struct cfq_io_context *cic, gfp_t gfp_mask)
1657 unsigned long flags;
1660 ret = radix_tree_preload(gfp_mask);
1665 spin_lock_irqsave(&ioc->lock, flags);
1666 ret = radix_tree_insert(&ioc->radix_root,
1667 (unsigned long) cfqd, cic);
1669 hlist_add_head_rcu(&cic->cic_list, &ioc->cic_list);
1670 spin_unlock_irqrestore(&ioc->lock, flags);
1672 radix_tree_preload_end();
1675 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1676 list_add(&cic->queue_list, &cfqd->cic_list);
1677 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1682 printk(KERN_ERR "cfq: cic link failed!\n");
1688 * Setup general io context and cfq io context. There can be several cfq
1689 * io contexts per general io context, if this process is doing io to more
1690 * than one device managed by cfq.
1692 static struct cfq_io_context *
1693 cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
1695 struct io_context *ioc = NULL;
1696 struct cfq_io_context *cic;
1698 might_sleep_if(gfp_mask & __GFP_WAIT);
1700 ioc = get_io_context(gfp_mask, cfqd->queue->node);
1704 cic = cfq_cic_lookup(cfqd, ioc);
1708 cic = cfq_alloc_io_context(cfqd, gfp_mask);
1712 if (cfq_cic_link(cfqd, ioc, cic, gfp_mask))
1716 smp_read_barrier_depends();
1717 if (unlikely(ioc->ioprio_changed))
1718 cfq_ioc_set_ioprio(ioc);
1724 put_io_context(ioc);
1729 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
1731 unsigned long elapsed = jiffies - cic->last_end_request;
1732 unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
1734 cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
1735 cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
1736 cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
1740 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_io_context *cic,
1746 if (cic->last_request_pos < rq->sector)
1747 sdist = rq->sector - cic->last_request_pos;
1749 sdist = cic->last_request_pos - rq->sector;
1752 * Don't allow the seek distance to get too large from the
1753 * odd fragment, pagein, etc
1755 if (cic->seek_samples <= 60) /* second&third seek */
1756 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*1024);
1758 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*64);
1760 cic->seek_samples = (7*cic->seek_samples + 256) / 8;
1761 cic->seek_total = (7*cic->seek_total + (u64)256*sdist) / 8;
1762 total = cic->seek_total + (cic->seek_samples/2);
1763 do_div(total, cic->seek_samples);
1764 cic->seek_mean = (sector_t)total;
1768 * Disable idle window if the process thinks too long or seeks so much that
1772 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1773 struct cfq_io_context *cic)
1775 int old_idle, enable_idle;
1778 * Don't idle for async or idle io prio class
1780 if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
1783 enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
1785 if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle ||
1786 (cfqd->hw_tag && CIC_SEEKY(cic)))
1788 else if (sample_valid(cic->ttime_samples)) {
1789 if (cic->ttime_mean > cfqd->cfq_slice_idle)
1795 if (old_idle != enable_idle) {
1796 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
1798 cfq_mark_cfqq_idle_window(cfqq);
1800 cfq_clear_cfqq_idle_window(cfqq);
1805 * Check if new_cfqq should preempt the currently active queue. Return 0 for
1806 * no or if we aren't sure, a 1 will cause a preempt.
1809 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
1812 struct cfq_queue *cfqq;
1814 cfqq = cfqd->active_queue;
1818 if (cfq_slice_used(cfqq))
1821 if (cfq_class_idle(new_cfqq))
1824 if (cfq_class_idle(cfqq))
1828 * if the new request is sync, but the currently running queue is
1829 * not, let the sync request have priority.
1831 if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
1835 * So both queues are sync. Let the new request get disk time if
1836 * it's a metadata request and the current queue is doing regular IO.
1838 if (rq_is_meta(rq) && !cfqq->meta_pending)
1842 * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
1844 if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
1847 if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
1851 * if this request is as-good as one we would expect from the
1852 * current cfqq, let it preempt
1854 if (cfq_rq_close(cfqd, rq))
1861 * cfqq preempts the active queue. if we allowed preempt with no slice left,
1862 * let it have half of its nominal slice.
1864 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1866 cfq_log_cfqq(cfqd, cfqq, "preempt");
1867 cfq_slice_expired(cfqd, 1);
1870 * Put the new queue at the front of the of the current list,
1871 * so we know that it will be selected next.
1873 BUG_ON(!cfq_cfqq_on_rr(cfqq));
1875 cfq_service_tree_add(cfqd, cfqq, 1);
1877 cfqq->slice_end = 0;
1878 cfq_mark_cfqq_slice_new(cfqq);
1882 * Called when a new fs request (rq) is added (to cfqq). Check if there's
1883 * something we should do about it
1886 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1889 struct cfq_io_context *cic = RQ_CIC(rq);
1893 cfqq->meta_pending++;
1895 cfq_update_io_thinktime(cfqd, cic);
1896 cfq_update_io_seektime(cfqd, cic, rq);
1897 cfq_update_idle_window(cfqd, cfqq, cic);
1899 cic->last_request_pos = rq->sector + rq->nr_sectors;
1901 if (cfqq == cfqd->active_queue) {
1903 * Remember that we saw a request from this process, but
1904 * don't start queuing just yet. Otherwise we risk seeing lots
1905 * of tiny requests, because we disrupt the normal plugging
1908 if (cfq_cfqq_wait_request(cfqq))
1909 cfq_mark_cfqq_must_dispatch(cfqq);
1910 } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
1912 * not the active queue - expire current slice if it is
1913 * idle and has expired it's mean thinktime or this new queue
1914 * has some old slice time left and is of higher priority or
1915 * this new queue is RT and the current one is BE
1917 cfq_preempt_queue(cfqd, cfqq);
1918 blk_start_queueing(cfqd->queue);
1922 static void cfq_insert_request(struct request_queue *q, struct request *rq)
1924 struct cfq_data *cfqd = q->elevator->elevator_data;
1925 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1927 cfq_log_cfqq(cfqd, cfqq, "insert_request");
1928 cfq_init_prio_data(cfqq, RQ_CIC(rq)->ioc);
1932 list_add_tail(&rq->queuelist, &cfqq->fifo);
1934 cfq_rq_enqueued(cfqd, cfqq, rq);
1938 * Update hw_tag based on peak queue depth over 50 samples under
1941 static void cfq_update_hw_tag(struct cfq_data *cfqd)
1943 if (cfqd->rq_in_driver > cfqd->rq_in_driver_peak)
1944 cfqd->rq_in_driver_peak = cfqd->rq_in_driver;
1946 if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
1947 cfqd->rq_in_driver <= CFQ_HW_QUEUE_MIN)
1950 if (cfqd->hw_tag_samples++ < 50)
1953 if (cfqd->rq_in_driver_peak >= CFQ_HW_QUEUE_MIN)
1958 cfqd->hw_tag_samples = 0;
1959 cfqd->rq_in_driver_peak = 0;
1962 static void cfq_completed_request(struct request_queue *q, struct request *rq)
1964 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1965 struct cfq_data *cfqd = cfqq->cfqd;
1966 const int sync = rq_is_sync(rq);
1970 cfq_log_cfqq(cfqd, cfqq, "complete");
1972 cfq_update_hw_tag(cfqd);
1974 WARN_ON(!cfqd->rq_in_driver);
1975 WARN_ON(!cfqq->dispatched);
1976 cfqd->rq_in_driver--;
1979 if (cfq_cfqq_sync(cfqq))
1980 cfqd->sync_flight--;
1982 if (!cfq_class_idle(cfqq))
1983 cfqd->last_end_request = now;
1986 RQ_CIC(rq)->last_end_request = now;
1989 * If this is the active queue, check if it needs to be expired,
1990 * or if we want to idle in case it has no pending requests.
1992 if (cfqd->active_queue == cfqq) {
1993 if (cfq_cfqq_slice_new(cfqq)) {
1994 cfq_set_prio_slice(cfqd, cfqq);
1995 cfq_clear_cfqq_slice_new(cfqq);
1997 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
1998 cfq_slice_expired(cfqd, 1);
1999 else if (sync && !rq_noidle(rq) &&
2000 RB_EMPTY_ROOT(&cfqq->sort_list)) {
2001 cfq_arm_slice_timer(cfqd);
2005 if (!cfqd->rq_in_driver)
2006 cfq_schedule_dispatch(cfqd);
2010 * we temporarily boost lower priority queues if they are holding fs exclusive
2011 * resources. they are boosted to normal prio (CLASS_BE/4)
2013 static void cfq_prio_boost(struct cfq_queue *cfqq)
2015 if (has_fs_excl()) {
2017 * boost idle prio on transactions that would lock out other
2018 * users of the filesystem
2020 if (cfq_class_idle(cfqq))
2021 cfqq->ioprio_class = IOPRIO_CLASS_BE;
2022 if (cfqq->ioprio > IOPRIO_NORM)
2023 cfqq->ioprio = IOPRIO_NORM;
2026 * check if we need to unboost the queue
2028 if (cfqq->ioprio_class != cfqq->org_ioprio_class)
2029 cfqq->ioprio_class = cfqq->org_ioprio_class;
2030 if (cfqq->ioprio != cfqq->org_ioprio)
2031 cfqq->ioprio = cfqq->org_ioprio;
2035 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
2037 if ((cfq_cfqq_wait_request(cfqq) || cfq_cfqq_must_alloc(cfqq)) &&
2038 !cfq_cfqq_must_alloc_slice(cfqq)) {
2039 cfq_mark_cfqq_must_alloc_slice(cfqq);
2040 return ELV_MQUEUE_MUST;
2043 return ELV_MQUEUE_MAY;
2046 static int cfq_may_queue(struct request_queue *q, int rw)
2048 struct cfq_data *cfqd = q->elevator->elevator_data;
2049 struct task_struct *tsk = current;
2050 struct cfq_io_context *cic;
2051 struct cfq_queue *cfqq;
2054 * don't force setup of a queue from here, as a call to may_queue
2055 * does not necessarily imply that a request actually will be queued.
2056 * so just lookup a possibly existing queue, or return 'may queue'
2059 cic = cfq_cic_lookup(cfqd, tsk->io_context);
2061 return ELV_MQUEUE_MAY;
2063 cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
2065 cfq_init_prio_data(cfqq, cic->ioc);
2066 cfq_prio_boost(cfqq);
2068 return __cfq_may_queue(cfqq);
2071 return ELV_MQUEUE_MAY;
2075 * queue lock held here
2077 static void cfq_put_request(struct request *rq)
2079 struct cfq_queue *cfqq = RQ_CFQQ(rq);
2082 const int rw = rq_data_dir(rq);
2084 BUG_ON(!cfqq->allocated[rw]);
2085 cfqq->allocated[rw]--;
2087 put_io_context(RQ_CIC(rq)->ioc);
2089 rq->elevator_private = NULL;
2090 rq->elevator_private2 = NULL;
2092 cfq_put_queue(cfqq);
2097 * Allocate cfq data structures associated with this request.
2100 cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
2102 struct cfq_data *cfqd = q->elevator->elevator_data;
2103 struct cfq_io_context *cic;
2104 const int rw = rq_data_dir(rq);
2105 const int is_sync = rq_is_sync(rq);
2106 struct cfq_queue *cfqq;
2107 unsigned long flags;
2109 might_sleep_if(gfp_mask & __GFP_WAIT);
2111 cic = cfq_get_io_context(cfqd, gfp_mask);
2113 spin_lock_irqsave(q->queue_lock, flags);
2118 cfqq = cic_to_cfqq(cic, is_sync);
2120 cfqq = cfq_get_queue(cfqd, is_sync, cic->ioc, gfp_mask);
2125 cic_set_cfqq(cic, cfqq, is_sync);
2128 cfqq->allocated[rw]++;
2129 cfq_clear_cfqq_must_alloc(cfqq);
2130 atomic_inc(&cfqq->ref);
2132 spin_unlock_irqrestore(q->queue_lock, flags);
2134 rq->elevator_private = cic;
2135 rq->elevator_private2 = cfqq;
2140 put_io_context(cic->ioc);
2142 cfq_schedule_dispatch(cfqd);
2143 spin_unlock_irqrestore(q->queue_lock, flags);
2144 cfq_log(cfqd, "set_request fail");
2148 static void cfq_kick_queue(struct work_struct *work)
2150 struct cfq_data *cfqd =
2151 container_of(work, struct cfq_data, unplug_work);
2152 struct request_queue *q = cfqd->queue;
2153 unsigned long flags;
2155 spin_lock_irqsave(q->queue_lock, flags);
2156 blk_start_queueing(q);
2157 spin_unlock_irqrestore(q->queue_lock, flags);
2161 * Timer running if the active_queue is currently idling inside its time slice
2163 static void cfq_idle_slice_timer(unsigned long data)
2165 struct cfq_data *cfqd = (struct cfq_data *) data;
2166 struct cfq_queue *cfqq;
2167 unsigned long flags;
2170 cfq_log(cfqd, "idle timer fired");
2172 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2174 cfqq = cfqd->active_queue;
2179 * We saw a request before the queue expired, let it through
2181 if (cfq_cfqq_must_dispatch(cfqq))
2187 if (cfq_slice_used(cfqq))
2191 * only expire and reinvoke request handler, if there are
2192 * other queues with pending requests
2194 if (!cfqd->busy_queues)
2198 * not expired and it has a request pending, let it dispatch
2200 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
2204 cfq_slice_expired(cfqd, timed_out);
2206 cfq_schedule_dispatch(cfqd);
2208 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2211 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
2213 del_timer_sync(&cfqd->idle_slice_timer);
2214 cancel_work_sync(&cfqd->unplug_work);
2217 static void cfq_put_async_queues(struct cfq_data *cfqd)
2221 for (i = 0; i < IOPRIO_BE_NR; i++) {
2222 if (cfqd->async_cfqq[0][i])
2223 cfq_put_queue(cfqd->async_cfqq[0][i]);
2224 if (cfqd->async_cfqq[1][i])
2225 cfq_put_queue(cfqd->async_cfqq[1][i]);
2228 if (cfqd->async_idle_cfqq)
2229 cfq_put_queue(cfqd->async_idle_cfqq);
2232 static void cfq_exit_queue(struct elevator_queue *e)
2234 struct cfq_data *cfqd = e->elevator_data;
2235 struct request_queue *q = cfqd->queue;
2237 cfq_shutdown_timer_wq(cfqd);
2239 spin_lock_irq(q->queue_lock);
2241 if (cfqd->active_queue)
2242 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
2244 while (!list_empty(&cfqd->cic_list)) {
2245 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
2246 struct cfq_io_context,
2249 __cfq_exit_single_io_context(cfqd, cic);
2252 cfq_put_async_queues(cfqd);
2254 spin_unlock_irq(q->queue_lock);
2256 cfq_shutdown_timer_wq(cfqd);
2261 static void *cfq_init_queue(struct request_queue *q)
2263 struct cfq_data *cfqd;
2265 cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
2269 cfqd->service_tree = CFQ_RB_ROOT;
2270 INIT_LIST_HEAD(&cfqd->cic_list);
2274 init_timer(&cfqd->idle_slice_timer);
2275 cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
2276 cfqd->idle_slice_timer.data = (unsigned long) cfqd;
2278 INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
2280 cfqd->last_end_request = jiffies;
2281 cfqd->cfq_quantum = cfq_quantum;
2282 cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
2283 cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
2284 cfqd->cfq_back_max = cfq_back_max;
2285 cfqd->cfq_back_penalty = cfq_back_penalty;
2286 cfqd->cfq_slice[0] = cfq_slice_async;
2287 cfqd->cfq_slice[1] = cfq_slice_sync;
2288 cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
2289 cfqd->cfq_slice_idle = cfq_slice_idle;
2295 static void cfq_slab_kill(void)
2298 * Caller already ensured that pending RCU callbacks are completed,
2299 * so we should have no busy allocations at this point.
2302 kmem_cache_destroy(cfq_pool);
2304 kmem_cache_destroy(cfq_ioc_pool);
2307 static int __init cfq_slab_setup(void)
2309 cfq_pool = KMEM_CACHE(cfq_queue, 0);
2313 cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0);
2324 * sysfs parts below -->
2327 cfq_var_show(unsigned int var, char *page)
2329 return sprintf(page, "%d\n", var);
2333 cfq_var_store(unsigned int *var, const char *page, size_t count)
2335 char *p = (char *) page;
2337 *var = simple_strtoul(p, &p, 10);
2341 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
2342 static ssize_t __FUNC(struct elevator_queue *e, char *page) \
2344 struct cfq_data *cfqd = e->elevator_data; \
2345 unsigned int __data = __VAR; \
2347 __data = jiffies_to_msecs(__data); \
2348 return cfq_var_show(__data, (page)); \
2350 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
2351 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
2352 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
2353 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
2354 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
2355 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
2356 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
2357 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
2358 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
2359 #undef SHOW_FUNCTION
2361 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
2362 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
2364 struct cfq_data *cfqd = e->elevator_data; \
2365 unsigned int __data; \
2366 int ret = cfq_var_store(&__data, (page), count); \
2367 if (__data < (MIN)) \
2369 else if (__data > (MAX)) \
2372 *(__PTR) = msecs_to_jiffies(__data); \
2374 *(__PTR) = __data; \
2377 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
2378 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
2380 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
2382 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
2383 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
2385 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
2386 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
2387 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
2388 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
2390 #undef STORE_FUNCTION
2392 #define CFQ_ATTR(name) \
2393 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
2395 static struct elv_fs_entry cfq_attrs[] = {
2397 CFQ_ATTR(fifo_expire_sync),
2398 CFQ_ATTR(fifo_expire_async),
2399 CFQ_ATTR(back_seek_max),
2400 CFQ_ATTR(back_seek_penalty),
2401 CFQ_ATTR(slice_sync),
2402 CFQ_ATTR(slice_async),
2403 CFQ_ATTR(slice_async_rq),
2404 CFQ_ATTR(slice_idle),
2408 static struct elevator_type iosched_cfq = {
2410 .elevator_merge_fn = cfq_merge,
2411 .elevator_merged_fn = cfq_merged_request,
2412 .elevator_merge_req_fn = cfq_merged_requests,
2413 .elevator_allow_merge_fn = cfq_allow_merge,
2414 .elevator_dispatch_fn = cfq_dispatch_requests,
2415 .elevator_add_req_fn = cfq_insert_request,
2416 .elevator_activate_req_fn = cfq_activate_request,
2417 .elevator_deactivate_req_fn = cfq_deactivate_request,
2418 .elevator_queue_empty_fn = cfq_queue_empty,
2419 .elevator_completed_req_fn = cfq_completed_request,
2420 .elevator_former_req_fn = elv_rb_former_request,
2421 .elevator_latter_req_fn = elv_rb_latter_request,
2422 .elevator_set_req_fn = cfq_set_request,
2423 .elevator_put_req_fn = cfq_put_request,
2424 .elevator_may_queue_fn = cfq_may_queue,
2425 .elevator_init_fn = cfq_init_queue,
2426 .elevator_exit_fn = cfq_exit_queue,
2427 .trim = cfq_free_io_context,
2429 .elevator_attrs = cfq_attrs,
2430 .elevator_name = "cfq",
2431 .elevator_owner = THIS_MODULE,
2434 static int __init cfq_init(void)
2437 * could be 0 on HZ < 1000 setups
2439 if (!cfq_slice_async)
2440 cfq_slice_async = 1;
2441 if (!cfq_slice_idle)
2444 if (cfq_slab_setup())
2447 elv_register(&iosched_cfq);
2452 static void __exit cfq_exit(void)
2454 DECLARE_COMPLETION_ONSTACK(all_gone);
2455 elv_unregister(&iosched_cfq);
2456 ioc_gone = &all_gone;
2457 /* ioc_gone's update must be visible before reading ioc_count */
2461 * this also protects us from entering cfq_slab_kill() with
2462 * pending RCU callbacks
2464 if (elv_ioc_count_read(ioc_count))
2465 wait_for_completion(&all_gone);
2469 module_init(cfq_init);
2470 module_exit(cfq_exit);
2472 MODULE_AUTHOR("Jens Axboe");
2473 MODULE_LICENSE("GPL");
2474 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");