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;
30 static const int cfq_target_latency = HZ * 3/10; /* 300 ms */
31 static const int cfq_hist_divisor = 4;
34 * offset from end of service tree
36 #define CFQ_IDLE_DELAY (HZ / 5)
39 * below this threshold, we consider thinktime immediate
41 #define CFQ_MIN_TT (2)
44 * Allow merged cfqqs to perform this amount of seeky I/O before
45 * deciding to break the queues up again.
47 #define CFQQ_COOP_TOUT (HZ)
49 #define CFQ_SLICE_SCALE (5)
50 #define CFQ_HW_QUEUE_MIN (5)
53 ((struct cfq_io_context *) (rq)->elevator_private)
54 #define RQ_CFQQ(rq) (struct cfq_queue *) ((rq)->elevator_private2)
56 static struct kmem_cache *cfq_pool;
57 static struct kmem_cache *cfq_ioc_pool;
59 static DEFINE_PER_CPU(unsigned long, cfq_ioc_count);
60 static struct completion *ioc_gone;
61 static DEFINE_SPINLOCK(ioc_gone_lock);
63 #define CFQ_PRIO_LISTS IOPRIO_BE_NR
64 #define cfq_class_idle(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
65 #define cfq_class_rt(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
67 #define sample_valid(samples) ((samples) > 80)
70 * Most of our rbtree usage is for sorting with min extraction, so
71 * if we cache the leftmost node we don't have to walk down the tree
72 * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should
73 * move this into the elevator for the rq sorting as well.
80 #define CFQ_RB_ROOT (struct cfq_rb_root) { RB_ROOT, NULL, 0, }
83 * Per process-grouping structure
88 /* various state flags, see below */
91 struct cfq_data *cfqd;
92 /* service_tree member */
93 struct rb_node rb_node;
94 /* service_tree key */
96 /* prio tree member */
97 struct rb_node p_node;
98 /* prio tree root we belong to, if any */
99 struct rb_root *p_root;
100 /* sorted list of pending requests */
101 struct rb_root sort_list;
102 /* if fifo isn't expired, next request to serve */
103 struct request *next_rq;
104 /* requests queued in sort_list */
106 /* currently allocated requests */
108 /* fifo list of requests in sort_list */
109 struct list_head fifo;
111 unsigned long slice_end;
113 unsigned int slice_dispatch;
115 /* pending metadata requests */
117 /* number of requests that are on the dispatch list or inside driver */
120 /* io prio of this group */
121 unsigned short ioprio, org_ioprio;
122 unsigned short ioprio_class, org_ioprio_class;
124 unsigned int seek_samples;
127 sector_t last_request_pos;
128 unsigned long seeky_start;
132 struct cfq_rb_root *service_tree;
133 struct cfq_queue *new_cfqq;
137 * Per block device queue structure
140 struct request_queue *queue;
143 * rr list of queues with requests and the count of them
145 struct cfq_rb_root service_tree;
148 * Each priority tree is sorted by next_request position. These
149 * trees are used when determining if two or more queues are
150 * interleaving requests (see cfq_close_cooperator).
152 struct rb_root prio_trees[CFQ_PRIO_LISTS];
154 unsigned int busy_queues;
155 unsigned int busy_rt_queues;
156 unsigned int busy_queues_avg[2];
162 * queue-depth detection
167 int rq_in_driver_peak;
170 * idle window management
172 struct timer_list idle_slice_timer;
173 struct work_struct unplug_work;
175 struct cfq_queue *active_queue;
176 struct cfq_io_context *active_cic;
179 * async queue for each priority case
181 struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
182 struct cfq_queue *async_idle_cfqq;
184 sector_t last_position;
187 * tunables, see top of file
189 unsigned int cfq_quantum;
190 unsigned int cfq_fifo_expire[2];
191 unsigned int cfq_back_penalty;
192 unsigned int cfq_back_max;
193 unsigned int cfq_slice[2];
194 unsigned int cfq_slice_async_rq;
195 unsigned int cfq_slice_idle;
196 unsigned int cfq_latency;
198 struct list_head cic_list;
201 * Fallback dummy cfqq for extreme OOM conditions
203 struct cfq_queue oom_cfqq;
205 unsigned long last_end_sync_rq;
208 enum cfqq_state_flags {
209 CFQ_CFQQ_FLAG_on_rr = 0, /* on round-robin busy list */
210 CFQ_CFQQ_FLAG_wait_request, /* waiting for a request */
211 CFQ_CFQQ_FLAG_must_dispatch, /* must be allowed a dispatch */
212 CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
213 CFQ_CFQQ_FLAG_fifo_expire, /* FIFO checked in this slice */
214 CFQ_CFQQ_FLAG_idle_window, /* slice idling enabled */
215 CFQ_CFQQ_FLAG_prio_changed, /* task priority has changed */
216 CFQ_CFQQ_FLAG_slice_new, /* no requests dispatched in slice */
217 CFQ_CFQQ_FLAG_sync, /* synchronous queue */
218 CFQ_CFQQ_FLAG_coop, /* cfqq is shared */
221 #define CFQ_CFQQ_FNS(name) \
222 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \
224 (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name); \
226 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \
228 (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \
230 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \
232 return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \
236 CFQ_CFQQ_FNS(wait_request);
237 CFQ_CFQQ_FNS(must_dispatch);
238 CFQ_CFQQ_FNS(must_alloc_slice);
239 CFQ_CFQQ_FNS(fifo_expire);
240 CFQ_CFQQ_FNS(idle_window);
241 CFQ_CFQQ_FNS(prio_changed);
242 CFQ_CFQQ_FNS(slice_new);
247 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
248 blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args)
249 #define cfq_log(cfqd, fmt, args...) \
250 blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
252 static void cfq_dispatch_insert(struct request_queue *, struct request *);
253 static struct cfq_queue *cfq_get_queue(struct cfq_data *, bool,
254 struct io_context *, gfp_t);
255 static struct cfq_io_context *cfq_cic_lookup(struct cfq_data *,
256 struct io_context *);
258 static inline int rq_in_driver(struct cfq_data *cfqd)
260 return cfqd->rq_in_driver[0] + cfqd->rq_in_driver[1];
263 static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_context *cic,
266 return cic->cfqq[is_sync];
269 static inline void cic_set_cfqq(struct cfq_io_context *cic,
270 struct cfq_queue *cfqq, bool is_sync)
272 cic->cfqq[is_sync] = cfqq;
276 * We regard a request as SYNC, if it's either a read or has the SYNC bit
277 * set (in which case it could also be direct WRITE).
279 static inline bool cfq_bio_sync(struct bio *bio)
281 return bio_data_dir(bio) == READ || bio_rw_flagged(bio, BIO_RW_SYNCIO);
285 * scheduler run of queue, if there are requests pending and no one in the
286 * driver that will restart queueing
288 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
290 if (cfqd->busy_queues) {
291 cfq_log(cfqd, "schedule dispatch");
292 kblockd_schedule_work(cfqd->queue, &cfqd->unplug_work);
296 static int cfq_queue_empty(struct request_queue *q)
298 struct cfq_data *cfqd = q->elevator->elevator_data;
300 return !cfqd->busy_queues;
304 * Scale schedule slice based on io priority. Use the sync time slice only
305 * if a queue is marked sync and has sync io queued. A sync queue with async
306 * io only, should not get full sync slice length.
308 static inline int cfq_prio_slice(struct cfq_data *cfqd, bool sync,
311 const int base_slice = cfqd->cfq_slice[sync];
313 WARN_ON(prio >= IOPRIO_BE_NR);
315 return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
319 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
321 return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
325 * get averaged number of queues of RT/BE priority.
326 * average is updated, with a formula that gives more weight to higher numbers,
327 * to quickly follows sudden increases and decrease slowly
330 static inline unsigned
331 cfq_get_avg_queues(struct cfq_data *cfqd, bool rt) {
332 unsigned min_q, max_q;
333 unsigned mult = cfq_hist_divisor - 1;
334 unsigned round = cfq_hist_divisor / 2;
335 unsigned busy = cfqd->busy_rt_queues;
338 busy = cfqd->busy_queues - cfqd->busy_rt_queues;
340 min_q = min(cfqd->busy_queues_avg[rt], busy);
341 max_q = max(cfqd->busy_queues_avg[rt], busy);
342 cfqd->busy_queues_avg[rt] = (mult * max_q + min_q + round) /
344 return cfqd->busy_queues_avg[rt];
348 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
350 unsigned slice = cfq_prio_to_slice(cfqd, cfqq);
351 if (cfqd->cfq_latency) {
352 /* interested queues (we consider only the ones with the same
354 unsigned iq = cfq_get_avg_queues(cfqd, cfq_class_rt(cfqq));
355 unsigned sync_slice = cfqd->cfq_slice[1];
356 unsigned expect_latency = sync_slice * iq;
357 if (expect_latency > cfq_target_latency) {
358 unsigned base_low_slice = 2 * cfqd->cfq_slice_idle;
359 /* scale low_slice according to IO priority
360 * and sync vs async */
362 min(slice, base_low_slice * slice / sync_slice);
363 /* the adapted slice value is scaled to fit all iqs
364 * into the target latency */
365 slice = max(slice * cfq_target_latency / expect_latency,
369 cfqq->slice_end = jiffies + slice;
370 cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
374 * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
375 * isn't valid until the first request from the dispatch is activated
376 * and the slice time set.
378 static inline bool cfq_slice_used(struct cfq_queue *cfqq)
380 if (cfq_cfqq_slice_new(cfqq))
382 if (time_before(jiffies, cfqq->slice_end))
389 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
390 * We choose the request that is closest to the head right now. Distance
391 * behind the head is penalized and only allowed to a certain extent.
393 static struct request *
394 cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2)
396 sector_t last, s1, s2, d1 = 0, d2 = 0;
397 unsigned long back_max;
398 #define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */
399 #define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */
400 unsigned wrap = 0; /* bit mask: requests behind the disk head? */
402 if (rq1 == NULL || rq1 == rq2)
407 if (rq_is_sync(rq1) && !rq_is_sync(rq2))
409 else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
411 if (rq_is_meta(rq1) && !rq_is_meta(rq2))
413 else if (rq_is_meta(rq2) && !rq_is_meta(rq1))
416 s1 = blk_rq_pos(rq1);
417 s2 = blk_rq_pos(rq2);
419 last = cfqd->last_position;
422 * by definition, 1KiB is 2 sectors
424 back_max = cfqd->cfq_back_max * 2;
427 * Strict one way elevator _except_ in the case where we allow
428 * short backward seeks which are biased as twice the cost of a
429 * similar forward seek.
433 else if (s1 + back_max >= last)
434 d1 = (last - s1) * cfqd->cfq_back_penalty;
436 wrap |= CFQ_RQ1_WRAP;
440 else if (s2 + back_max >= last)
441 d2 = (last - s2) * cfqd->cfq_back_penalty;
443 wrap |= CFQ_RQ2_WRAP;
445 /* Found required data */
448 * By doing switch() on the bit mask "wrap" we avoid having to
449 * check two variables for all permutations: --> faster!
452 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
468 case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
471 * Since both rqs are wrapped,
472 * start with the one that's further behind head
473 * (--> only *one* back seek required),
474 * since back seek takes more time than forward.
484 * The below is leftmost cache rbtree addon
486 static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
489 root->left = rb_first(&root->rb);
492 return rb_entry(root->left, struct cfq_queue, rb_node);
497 static void rb_erase_init(struct rb_node *n, struct rb_root *root)
503 static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
507 rb_erase_init(n, &root->rb);
512 * would be nice to take fifo expire time into account as well
514 static struct request *
515 cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
516 struct request *last)
518 struct rb_node *rbnext = rb_next(&last->rb_node);
519 struct rb_node *rbprev = rb_prev(&last->rb_node);
520 struct request *next = NULL, *prev = NULL;
522 BUG_ON(RB_EMPTY_NODE(&last->rb_node));
525 prev = rb_entry_rq(rbprev);
528 next = rb_entry_rq(rbnext);
530 rbnext = rb_first(&cfqq->sort_list);
531 if (rbnext && rbnext != &last->rb_node)
532 next = rb_entry_rq(rbnext);
535 return cfq_choose_req(cfqd, next, prev);
538 static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
539 struct cfq_queue *cfqq)
542 * just an approximation, should be ok.
544 return (cfqd->busy_queues - 1) * (cfq_prio_slice(cfqd, 1, 0) -
545 cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
549 * The cfqd->service_tree holds all pending cfq_queue's that have
550 * requests waiting to be processed. It is sorted in the order that
551 * we will service the queues.
553 static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
556 struct rb_node **p, *parent;
557 struct cfq_queue *__cfqq;
558 unsigned long rb_key;
559 struct cfq_rb_root *service_tree = &cfqd->service_tree;
562 if (cfq_class_idle(cfqq)) {
563 rb_key = CFQ_IDLE_DELAY;
564 parent = rb_last(&service_tree->rb);
565 if (parent && parent != &cfqq->rb_node) {
566 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
567 rb_key += __cfqq->rb_key;
570 } else if (!add_front) {
572 * Get our rb key offset. Subtract any residual slice
573 * value carried from last service. A negative resid
574 * count indicates slice overrun, and this should position
575 * the next service time further away in the tree.
577 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
578 rb_key -= cfqq->slice_resid;
579 cfqq->slice_resid = 0;
582 __cfqq = cfq_rb_first(service_tree);
583 rb_key += __cfqq ? __cfqq->rb_key : jiffies;
586 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
588 * same position, nothing more to do
590 if (rb_key == cfqq->rb_key)
593 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
594 cfqq->service_tree = NULL;
599 cfqq->service_tree = service_tree;
600 p = &service_tree->rb.rb_node;
605 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
608 * sort RT queues first, we always want to give
609 * preference to them. IDLE queues goes to the back.
610 * after that, sort on the next service time.
612 if (cfq_class_rt(cfqq) > cfq_class_rt(__cfqq))
614 else if (cfq_class_rt(cfqq) < cfq_class_rt(__cfqq))
616 else if (cfq_class_idle(cfqq) < cfq_class_idle(__cfqq))
618 else if (cfq_class_idle(cfqq) > cfq_class_idle(__cfqq))
620 else if (time_before(rb_key, __cfqq->rb_key))
625 if (n == &(*p)->rb_right)
632 service_tree->left = &cfqq->rb_node;
634 cfqq->rb_key = rb_key;
635 rb_link_node(&cfqq->rb_node, parent, p);
636 rb_insert_color(&cfqq->rb_node, &service_tree->rb);
637 service_tree->count++;
640 static struct cfq_queue *
641 cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
642 sector_t sector, struct rb_node **ret_parent,
643 struct rb_node ***rb_link)
645 struct rb_node **p, *parent;
646 struct cfq_queue *cfqq = NULL;
654 cfqq = rb_entry(parent, struct cfq_queue, p_node);
657 * Sort strictly based on sector. Smallest to the left,
658 * largest to the right.
660 if (sector > blk_rq_pos(cfqq->next_rq))
662 else if (sector < blk_rq_pos(cfqq->next_rq))
670 *ret_parent = parent;
676 static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
678 struct rb_node **p, *parent;
679 struct cfq_queue *__cfqq;
682 rb_erase(&cfqq->p_node, cfqq->p_root);
686 if (cfq_class_idle(cfqq))
691 cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
692 __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
693 blk_rq_pos(cfqq->next_rq), &parent, &p);
695 rb_link_node(&cfqq->p_node, parent, p);
696 rb_insert_color(&cfqq->p_node, cfqq->p_root);
702 * Update cfqq's position in the service tree.
704 static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
707 * Resorting requires the cfqq to be on the RR list already.
709 if (cfq_cfqq_on_rr(cfqq)) {
710 cfq_service_tree_add(cfqd, cfqq, 0);
711 cfq_prio_tree_add(cfqd, cfqq);
716 * add to busy list of queues for service, trying to be fair in ordering
717 * the pending list according to last request service
719 static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
721 cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
722 BUG_ON(cfq_cfqq_on_rr(cfqq));
723 cfq_mark_cfqq_on_rr(cfqq);
725 if (cfq_class_rt(cfqq))
726 cfqd->busy_rt_queues++;
727 cfq_resort_rr_list(cfqd, cfqq);
731 * Called when the cfqq no longer has requests pending, remove it from
734 static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
736 cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
737 BUG_ON(!cfq_cfqq_on_rr(cfqq));
738 cfq_clear_cfqq_on_rr(cfqq);
740 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
741 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
742 cfqq->service_tree = NULL;
745 rb_erase(&cfqq->p_node, cfqq->p_root);
749 BUG_ON(!cfqd->busy_queues);
751 if (cfq_class_rt(cfqq))
752 cfqd->busy_rt_queues--;
756 * rb tree support functions
758 static void cfq_del_rq_rb(struct request *rq)
760 struct cfq_queue *cfqq = RQ_CFQQ(rq);
761 struct cfq_data *cfqd = cfqq->cfqd;
762 const int sync = rq_is_sync(rq);
764 BUG_ON(!cfqq->queued[sync]);
765 cfqq->queued[sync]--;
767 elv_rb_del(&cfqq->sort_list, rq);
769 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
770 cfq_del_cfqq_rr(cfqd, cfqq);
773 static void cfq_add_rq_rb(struct request *rq)
775 struct cfq_queue *cfqq = RQ_CFQQ(rq);
776 struct cfq_data *cfqd = cfqq->cfqd;
777 struct request *__alias, *prev;
779 cfqq->queued[rq_is_sync(rq)]++;
782 * looks a little odd, but the first insert might return an alias.
783 * if that happens, put the alias on the dispatch list
785 while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL)
786 cfq_dispatch_insert(cfqd->queue, __alias);
788 if (!cfq_cfqq_on_rr(cfqq))
789 cfq_add_cfqq_rr(cfqd, cfqq);
792 * check if this request is a better next-serve candidate
794 prev = cfqq->next_rq;
795 cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq);
798 * adjust priority tree position, if ->next_rq changes
800 if (prev != cfqq->next_rq)
801 cfq_prio_tree_add(cfqd, cfqq);
803 BUG_ON(!cfqq->next_rq);
806 static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
808 elv_rb_del(&cfqq->sort_list, rq);
809 cfqq->queued[rq_is_sync(rq)]--;
813 static struct request *
814 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
816 struct task_struct *tsk = current;
817 struct cfq_io_context *cic;
818 struct cfq_queue *cfqq;
820 cic = cfq_cic_lookup(cfqd, tsk->io_context);
824 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
826 sector_t sector = bio->bi_sector + bio_sectors(bio);
828 return elv_rb_find(&cfqq->sort_list, sector);
834 static void cfq_activate_request(struct request_queue *q, struct request *rq)
836 struct cfq_data *cfqd = q->elevator->elevator_data;
838 cfqd->rq_in_driver[rq_is_sync(rq)]++;
839 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
842 cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
845 static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
847 struct cfq_data *cfqd = q->elevator->elevator_data;
848 const int sync = rq_is_sync(rq);
850 WARN_ON(!cfqd->rq_in_driver[sync]);
851 cfqd->rq_in_driver[sync]--;
852 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
856 static void cfq_remove_request(struct request *rq)
858 struct cfq_queue *cfqq = RQ_CFQQ(rq);
860 if (cfqq->next_rq == rq)
861 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
863 list_del_init(&rq->queuelist);
866 cfqq->cfqd->rq_queued--;
867 if (rq_is_meta(rq)) {
868 WARN_ON(!cfqq->meta_pending);
869 cfqq->meta_pending--;
873 static int cfq_merge(struct request_queue *q, struct request **req,
876 struct cfq_data *cfqd = q->elevator->elevator_data;
877 struct request *__rq;
879 __rq = cfq_find_rq_fmerge(cfqd, bio);
880 if (__rq && elv_rq_merge_ok(__rq, bio)) {
882 return ELEVATOR_FRONT_MERGE;
885 return ELEVATOR_NO_MERGE;
888 static void cfq_merged_request(struct request_queue *q, struct request *req,
891 if (type == ELEVATOR_FRONT_MERGE) {
892 struct cfq_queue *cfqq = RQ_CFQQ(req);
894 cfq_reposition_rq_rb(cfqq, req);
899 cfq_merged_requests(struct request_queue *q, struct request *rq,
900 struct request *next)
903 * reposition in fifo if next is older than rq
905 if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
906 time_before(rq_fifo_time(next), rq_fifo_time(rq))) {
907 list_move(&rq->queuelist, &next->queuelist);
908 rq_set_fifo_time(rq, rq_fifo_time(next));
911 cfq_remove_request(next);
914 static int cfq_allow_merge(struct request_queue *q, struct request *rq,
917 struct cfq_data *cfqd = q->elevator->elevator_data;
918 struct cfq_io_context *cic;
919 struct cfq_queue *cfqq;
922 * Disallow merge of a sync bio into an async request.
924 if (cfq_bio_sync(bio) && !rq_is_sync(rq))
928 * Lookup the cfqq that this bio will be queued with. Allow
929 * merge only if rq is queued there.
931 cic = cfq_cic_lookup(cfqd, current->io_context);
935 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
936 return cfqq == RQ_CFQQ(rq);
939 static void __cfq_set_active_queue(struct cfq_data *cfqd,
940 struct cfq_queue *cfqq)
943 cfq_log_cfqq(cfqd, cfqq, "set_active");
945 cfqq->slice_dispatch = 0;
947 cfq_clear_cfqq_wait_request(cfqq);
948 cfq_clear_cfqq_must_dispatch(cfqq);
949 cfq_clear_cfqq_must_alloc_slice(cfqq);
950 cfq_clear_cfqq_fifo_expire(cfqq);
951 cfq_mark_cfqq_slice_new(cfqq);
953 del_timer(&cfqd->idle_slice_timer);
956 cfqd->active_queue = cfqq;
960 * current cfqq expired its slice (or was too idle), select new one
963 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
966 cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
968 if (cfq_cfqq_wait_request(cfqq))
969 del_timer(&cfqd->idle_slice_timer);
971 cfq_clear_cfqq_wait_request(cfqq);
974 * store what was left of this slice, if the queue idled/timed out
976 if (timed_out && !cfq_cfqq_slice_new(cfqq)) {
977 cfqq->slice_resid = cfqq->slice_end - jiffies;
978 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
981 cfq_resort_rr_list(cfqd, cfqq);
983 if (cfqq == cfqd->active_queue)
984 cfqd->active_queue = NULL;
986 if (cfqd->active_cic) {
987 put_io_context(cfqd->active_cic->ioc);
988 cfqd->active_cic = NULL;
992 static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out)
994 struct cfq_queue *cfqq = cfqd->active_queue;
997 __cfq_slice_expired(cfqd, cfqq, timed_out);
1001 * Get next queue for service. Unless we have a queue preemption,
1002 * we'll simply select the first cfqq in the service tree.
1004 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
1006 if (RB_EMPTY_ROOT(&cfqd->service_tree.rb))
1009 return cfq_rb_first(&cfqd->service_tree);
1013 * Get and set a new active queue for service.
1015 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
1016 struct cfq_queue *cfqq)
1019 cfqq = cfq_get_next_queue(cfqd);
1021 __cfq_set_active_queue(cfqd, cfqq);
1025 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
1028 if (blk_rq_pos(rq) >= cfqd->last_position)
1029 return blk_rq_pos(rq) - cfqd->last_position;
1031 return cfqd->last_position - blk_rq_pos(rq);
1034 #define CFQQ_SEEK_THR 8 * 1024
1035 #define CFQQ_SEEKY(cfqq) ((cfqq)->seek_mean > CFQQ_SEEK_THR)
1037 static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1040 sector_t sdist = cfqq->seek_mean;
1042 if (!sample_valid(cfqq->seek_samples))
1043 sdist = CFQQ_SEEK_THR;
1045 return cfq_dist_from_last(cfqd, rq) <= sdist;
1048 static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
1049 struct cfq_queue *cur_cfqq)
1051 struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
1052 struct rb_node *parent, *node;
1053 struct cfq_queue *__cfqq;
1054 sector_t sector = cfqd->last_position;
1056 if (RB_EMPTY_ROOT(root))
1060 * First, if we find a request starting at the end of the last
1061 * request, choose it.
1063 __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
1068 * If the exact sector wasn't found, the parent of the NULL leaf
1069 * will contain the closest sector.
1071 __cfqq = rb_entry(parent, struct cfq_queue, p_node);
1072 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1075 if (blk_rq_pos(__cfqq->next_rq) < sector)
1076 node = rb_next(&__cfqq->p_node);
1078 node = rb_prev(&__cfqq->p_node);
1082 __cfqq = rb_entry(node, struct cfq_queue, p_node);
1083 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1091 * cur_cfqq - passed in so that we don't decide that the current queue is
1092 * closely cooperating with itself.
1094 * So, basically we're assuming that that cur_cfqq has dispatched at least
1095 * one request, and that cfqd->last_position reflects a position on the disk
1096 * associated with the I/O issued by cur_cfqq. I'm not sure this is a valid
1099 static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
1100 struct cfq_queue *cur_cfqq)
1102 struct cfq_queue *cfqq;
1104 if (!cfq_cfqq_sync(cur_cfqq))
1106 if (CFQQ_SEEKY(cur_cfqq))
1110 * We should notice if some of the queues are cooperating, eg
1111 * working closely on the same area of the disk. In that case,
1112 * we can group them together and don't waste time idling.
1114 cfqq = cfqq_close(cfqd, cur_cfqq);
1119 * It only makes sense to merge sync queues.
1121 if (!cfq_cfqq_sync(cfqq))
1123 if (CFQQ_SEEKY(cfqq))
1129 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
1131 struct cfq_queue *cfqq = cfqd->active_queue;
1132 struct cfq_io_context *cic;
1136 * SSD device without seek penalty, disable idling. But only do so
1137 * for devices that support queuing, otherwise we still have a problem
1138 * with sync vs async workloads.
1140 if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
1143 WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
1144 WARN_ON(cfq_cfqq_slice_new(cfqq));
1147 * idle is disabled, either manually or by past process history
1149 if (!cfqd->cfq_slice_idle || !cfq_cfqq_idle_window(cfqq))
1153 * still requests with the driver, don't idle
1155 if (rq_in_driver(cfqd))
1159 * task has exited, don't wait
1161 cic = cfqd->active_cic;
1162 if (!cic || !atomic_read(&cic->ioc->nr_tasks))
1166 * If our average think time is larger than the remaining time
1167 * slice, then don't idle. This avoids overrunning the allotted
1170 if (sample_valid(cic->ttime_samples) &&
1171 (cfqq->slice_end - jiffies < cic->ttime_mean))
1174 cfq_mark_cfqq_wait_request(cfqq);
1177 * we don't want to idle for seeks, but we do want to allow
1178 * fair distribution of slice time for a process doing back-to-back
1179 * seeks. so allow a little bit of time for him to submit a new rq
1181 sl = cfqd->cfq_slice_idle;
1182 if (sample_valid(cfqq->seek_samples) && CFQQ_SEEKY(cfqq))
1183 sl = min(sl, msecs_to_jiffies(CFQ_MIN_TT));
1185 mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
1186 cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu", sl);
1190 * Move request from internal lists to the request queue dispatch list.
1192 static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
1194 struct cfq_data *cfqd = q->elevator->elevator_data;
1195 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1197 cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
1199 cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
1200 cfq_remove_request(rq);
1202 elv_dispatch_sort(q, rq);
1204 if (cfq_cfqq_sync(cfqq))
1205 cfqd->sync_flight++;
1209 * return expired entry, or NULL to just start from scratch in rbtree
1211 static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
1213 struct request *rq = NULL;
1215 if (cfq_cfqq_fifo_expire(cfqq))
1218 cfq_mark_cfqq_fifo_expire(cfqq);
1220 if (list_empty(&cfqq->fifo))
1223 rq = rq_entry_fifo(cfqq->fifo.next);
1224 if (time_before(jiffies, rq_fifo_time(rq)))
1227 cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq);
1232 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1234 const int base_rq = cfqd->cfq_slice_async_rq;
1236 WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
1238 return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
1242 * Must be called with the queue_lock held.
1244 static int cfqq_process_refs(struct cfq_queue *cfqq)
1246 int process_refs, io_refs;
1248 io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE];
1249 process_refs = atomic_read(&cfqq->ref) - io_refs;
1250 BUG_ON(process_refs < 0);
1251 return process_refs;
1254 static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq)
1256 int process_refs, new_process_refs;
1257 struct cfq_queue *__cfqq;
1259 /* Avoid a circular list and skip interim queue merges */
1260 while ((__cfqq = new_cfqq->new_cfqq)) {
1266 process_refs = cfqq_process_refs(cfqq);
1268 * If the process for the cfqq has gone away, there is no
1269 * sense in merging the queues.
1271 if (process_refs == 0)
1275 * Merge in the direction of the lesser amount of work.
1277 new_process_refs = cfqq_process_refs(new_cfqq);
1278 if (new_process_refs >= process_refs) {
1279 cfqq->new_cfqq = new_cfqq;
1280 atomic_add(process_refs, &new_cfqq->ref);
1282 new_cfqq->new_cfqq = cfqq;
1283 atomic_add(new_process_refs, &cfqq->ref);
1288 * Select a queue for service. If we have a current active queue,
1289 * check whether to continue servicing it, or retrieve and set a new one.
1291 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
1293 struct cfq_queue *cfqq, *new_cfqq = NULL;
1295 cfqq = cfqd->active_queue;
1300 * The active queue has run out of time, expire it and select new.
1302 if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq))
1306 * The active queue has requests and isn't expired, allow it to
1309 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
1313 * If another queue has a request waiting within our mean seek
1314 * distance, let it run. The expire code will check for close
1315 * cooperators and put the close queue at the front of the service
1316 * tree. If possible, merge the expiring queue with the new cfqq.
1318 new_cfqq = cfq_close_cooperator(cfqd, cfqq);
1320 if (!cfqq->new_cfqq)
1321 cfq_setup_merge(cfqq, new_cfqq);
1326 * No requests pending. If the active queue still has requests in
1327 * flight or is idling for a new request, allow either of these
1328 * conditions to happen (or time out) before selecting a new queue.
1330 if (timer_pending(&cfqd->idle_slice_timer) ||
1331 (cfqq->dispatched && cfq_cfqq_idle_window(cfqq))) {
1337 cfq_slice_expired(cfqd, 0);
1339 cfqq = cfq_set_active_queue(cfqd, new_cfqq);
1344 static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
1348 while (cfqq->next_rq) {
1349 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
1353 BUG_ON(!list_empty(&cfqq->fifo));
1358 * Drain our current requests. Used for barriers and when switching
1359 * io schedulers on-the-fly.
1361 static int cfq_forced_dispatch(struct cfq_data *cfqd)
1363 struct cfq_queue *cfqq;
1366 while ((cfqq = cfq_rb_first(&cfqd->service_tree)) != NULL)
1367 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
1369 cfq_slice_expired(cfqd, 0);
1371 BUG_ON(cfqd->busy_queues);
1373 cfq_log(cfqd, "forced_dispatch=%d", dispatched);
1377 static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1379 unsigned int max_dispatch;
1382 * Drain async requests before we start sync IO
1384 if (cfq_cfqq_idle_window(cfqq) && cfqd->rq_in_driver[BLK_RW_ASYNC])
1388 * If this is an async queue and we have sync IO in flight, let it wait
1390 if (cfqd->sync_flight && !cfq_cfqq_sync(cfqq))
1393 max_dispatch = cfqd->cfq_quantum;
1394 if (cfq_class_idle(cfqq))
1398 * Does this cfqq already have too much IO in flight?
1400 if (cfqq->dispatched >= max_dispatch) {
1402 * idle queue must always only have a single IO in flight
1404 if (cfq_class_idle(cfqq))
1408 * We have other queues, don't allow more IO from this one
1410 if (cfqd->busy_queues > 1)
1414 * Sole queue user, allow bigger slice
1420 * Async queues must wait a bit before being allowed dispatch.
1421 * We also ramp up the dispatch depth gradually for async IO,
1422 * based on the last sync IO we serviced
1424 if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
1425 unsigned long last_sync = jiffies - cfqd->last_end_sync_rq;
1428 depth = last_sync / cfqd->cfq_slice[1];
1429 if (!depth && !cfqq->dispatched)
1431 if (depth < max_dispatch)
1432 max_dispatch = depth;
1436 * If we're below the current max, allow a dispatch
1438 return cfqq->dispatched < max_dispatch;
1442 * Dispatch a request from cfqq, moving them to the request queue
1445 static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1449 BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
1451 if (!cfq_may_dispatch(cfqd, cfqq))
1455 * follow expired path, else get first next available
1457 rq = cfq_check_fifo(cfqq);
1462 * insert request into driver dispatch list
1464 cfq_dispatch_insert(cfqd->queue, rq);
1466 if (!cfqd->active_cic) {
1467 struct cfq_io_context *cic = RQ_CIC(rq);
1469 atomic_long_inc(&cic->ioc->refcount);
1470 cfqd->active_cic = cic;
1477 * Find the cfqq that we need to service and move a request from that to the
1480 static int cfq_dispatch_requests(struct request_queue *q, int force)
1482 struct cfq_data *cfqd = q->elevator->elevator_data;
1483 struct cfq_queue *cfqq;
1485 if (!cfqd->busy_queues)
1488 if (unlikely(force))
1489 return cfq_forced_dispatch(cfqd);
1491 cfqq = cfq_select_queue(cfqd);
1496 * Dispatch a request from this cfqq, if it is allowed
1498 if (!cfq_dispatch_request(cfqd, cfqq))
1501 cfqq->slice_dispatch++;
1502 cfq_clear_cfqq_must_dispatch(cfqq);
1505 * expire an async queue immediately if it has used up its slice. idle
1506 * queue always expire after 1 dispatch round.
1508 if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
1509 cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
1510 cfq_class_idle(cfqq))) {
1511 cfqq->slice_end = jiffies + 1;
1512 cfq_slice_expired(cfqd, 0);
1515 cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
1520 * task holds one reference to the queue, dropped when task exits. each rq
1521 * in-flight on this queue also holds a reference, dropped when rq is freed.
1523 * queue lock must be held here.
1525 static void cfq_put_queue(struct cfq_queue *cfqq)
1527 struct cfq_data *cfqd = cfqq->cfqd;
1529 BUG_ON(atomic_read(&cfqq->ref) <= 0);
1531 if (!atomic_dec_and_test(&cfqq->ref))
1534 cfq_log_cfqq(cfqd, cfqq, "put_queue");
1535 BUG_ON(rb_first(&cfqq->sort_list));
1536 BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
1537 BUG_ON(cfq_cfqq_on_rr(cfqq));
1539 if (unlikely(cfqd->active_queue == cfqq)) {
1540 __cfq_slice_expired(cfqd, cfqq, 0);
1541 cfq_schedule_dispatch(cfqd);
1544 kmem_cache_free(cfq_pool, cfqq);
1548 * Must always be called with the rcu_read_lock() held
1551 __call_for_each_cic(struct io_context *ioc,
1552 void (*func)(struct io_context *, struct cfq_io_context *))
1554 struct cfq_io_context *cic;
1555 struct hlist_node *n;
1557 hlist_for_each_entry_rcu(cic, n, &ioc->cic_list, cic_list)
1562 * Call func for each cic attached to this ioc.
1565 call_for_each_cic(struct io_context *ioc,
1566 void (*func)(struct io_context *, struct cfq_io_context *))
1569 __call_for_each_cic(ioc, func);
1573 static void cfq_cic_free_rcu(struct rcu_head *head)
1575 struct cfq_io_context *cic;
1577 cic = container_of(head, struct cfq_io_context, rcu_head);
1579 kmem_cache_free(cfq_ioc_pool, cic);
1580 elv_ioc_count_dec(cfq_ioc_count);
1584 * CFQ scheduler is exiting, grab exit lock and check
1585 * the pending io context count. If it hits zero,
1586 * complete ioc_gone and set it back to NULL
1588 spin_lock(&ioc_gone_lock);
1589 if (ioc_gone && !elv_ioc_count_read(cfq_ioc_count)) {
1593 spin_unlock(&ioc_gone_lock);
1597 static void cfq_cic_free(struct cfq_io_context *cic)
1599 call_rcu(&cic->rcu_head, cfq_cic_free_rcu);
1602 static void cic_free_func(struct io_context *ioc, struct cfq_io_context *cic)
1604 unsigned long flags;
1606 BUG_ON(!cic->dead_key);
1608 spin_lock_irqsave(&ioc->lock, flags);
1609 radix_tree_delete(&ioc->radix_root, cic->dead_key);
1610 hlist_del_rcu(&cic->cic_list);
1611 spin_unlock_irqrestore(&ioc->lock, flags);
1617 * Must be called with rcu_read_lock() held or preemption otherwise disabled.
1618 * Only two callers of this - ->dtor() which is called with the rcu_read_lock(),
1619 * and ->trim() which is called with the task lock held
1621 static void cfq_free_io_context(struct io_context *ioc)
1624 * ioc->refcount is zero here, or we are called from elv_unregister(),
1625 * so no more cic's are allowed to be linked into this ioc. So it
1626 * should be ok to iterate over the known list, we will see all cic's
1627 * since no new ones are added.
1629 __call_for_each_cic(ioc, cic_free_func);
1632 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1634 struct cfq_queue *__cfqq, *next;
1636 if (unlikely(cfqq == cfqd->active_queue)) {
1637 __cfq_slice_expired(cfqd, cfqq, 0);
1638 cfq_schedule_dispatch(cfqd);
1642 * If this queue was scheduled to merge with another queue, be
1643 * sure to drop the reference taken on that queue (and others in
1644 * the merge chain). See cfq_setup_merge and cfq_merge_cfqqs.
1646 __cfqq = cfqq->new_cfqq;
1648 if (__cfqq == cfqq) {
1649 WARN(1, "cfqq->new_cfqq loop detected\n");
1652 next = __cfqq->new_cfqq;
1653 cfq_put_queue(__cfqq);
1657 cfq_put_queue(cfqq);
1660 static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
1661 struct cfq_io_context *cic)
1663 struct io_context *ioc = cic->ioc;
1665 list_del_init(&cic->queue_list);
1668 * Make sure key == NULL is seen for dead queues
1671 cic->dead_key = (unsigned long) cic->key;
1674 if (ioc->ioc_data == cic)
1675 rcu_assign_pointer(ioc->ioc_data, NULL);
1677 if (cic->cfqq[BLK_RW_ASYNC]) {
1678 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]);
1679 cic->cfqq[BLK_RW_ASYNC] = NULL;
1682 if (cic->cfqq[BLK_RW_SYNC]) {
1683 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]);
1684 cic->cfqq[BLK_RW_SYNC] = NULL;
1688 static void cfq_exit_single_io_context(struct io_context *ioc,
1689 struct cfq_io_context *cic)
1691 struct cfq_data *cfqd = cic->key;
1694 struct request_queue *q = cfqd->queue;
1695 unsigned long flags;
1697 spin_lock_irqsave(q->queue_lock, flags);
1700 * Ensure we get a fresh copy of the ->key to prevent
1701 * race between exiting task and queue
1703 smp_read_barrier_depends();
1705 __cfq_exit_single_io_context(cfqd, cic);
1707 spin_unlock_irqrestore(q->queue_lock, flags);
1712 * The process that ioc belongs to has exited, we need to clean up
1713 * and put the internal structures we have that belongs to that process.
1715 static void cfq_exit_io_context(struct io_context *ioc)
1717 call_for_each_cic(ioc, cfq_exit_single_io_context);
1720 static struct cfq_io_context *
1721 cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
1723 struct cfq_io_context *cic;
1725 cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO,
1728 cic->last_end_request = jiffies;
1729 INIT_LIST_HEAD(&cic->queue_list);
1730 INIT_HLIST_NODE(&cic->cic_list);
1731 cic->dtor = cfq_free_io_context;
1732 cic->exit = cfq_exit_io_context;
1733 elv_ioc_count_inc(cfq_ioc_count);
1739 static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc)
1741 struct task_struct *tsk = current;
1744 if (!cfq_cfqq_prio_changed(cfqq))
1747 ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio);
1748 switch (ioprio_class) {
1750 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
1751 case IOPRIO_CLASS_NONE:
1753 * no prio set, inherit CPU scheduling settings
1755 cfqq->ioprio = task_nice_ioprio(tsk);
1756 cfqq->ioprio_class = task_nice_ioclass(tsk);
1758 case IOPRIO_CLASS_RT:
1759 cfqq->ioprio = task_ioprio(ioc);
1760 cfqq->ioprio_class = IOPRIO_CLASS_RT;
1762 case IOPRIO_CLASS_BE:
1763 cfqq->ioprio = task_ioprio(ioc);
1764 cfqq->ioprio_class = IOPRIO_CLASS_BE;
1766 case IOPRIO_CLASS_IDLE:
1767 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
1769 cfq_clear_cfqq_idle_window(cfqq);
1774 * keep track of original prio settings in case we have to temporarily
1775 * elevate the priority of this queue
1777 cfqq->org_ioprio = cfqq->ioprio;
1778 cfqq->org_ioprio_class = cfqq->ioprio_class;
1779 cfq_clear_cfqq_prio_changed(cfqq);
1782 static void changed_ioprio(struct io_context *ioc, struct cfq_io_context *cic)
1784 struct cfq_data *cfqd = cic->key;
1785 struct cfq_queue *cfqq;
1786 unsigned long flags;
1788 if (unlikely(!cfqd))
1791 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1793 cfqq = cic->cfqq[BLK_RW_ASYNC];
1795 struct cfq_queue *new_cfqq;
1796 new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic->ioc,
1799 cic->cfqq[BLK_RW_ASYNC] = new_cfqq;
1800 cfq_put_queue(cfqq);
1804 cfqq = cic->cfqq[BLK_RW_SYNC];
1806 cfq_mark_cfqq_prio_changed(cfqq);
1808 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1811 static void cfq_ioc_set_ioprio(struct io_context *ioc)
1813 call_for_each_cic(ioc, changed_ioprio);
1814 ioc->ioprio_changed = 0;
1817 static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1818 pid_t pid, bool is_sync)
1820 RB_CLEAR_NODE(&cfqq->rb_node);
1821 RB_CLEAR_NODE(&cfqq->p_node);
1822 INIT_LIST_HEAD(&cfqq->fifo);
1824 atomic_set(&cfqq->ref, 0);
1827 cfq_mark_cfqq_prio_changed(cfqq);
1830 if (!cfq_class_idle(cfqq))
1831 cfq_mark_cfqq_idle_window(cfqq);
1832 cfq_mark_cfqq_sync(cfqq);
1837 static struct cfq_queue *
1838 cfq_find_alloc_queue(struct cfq_data *cfqd, bool is_sync,
1839 struct io_context *ioc, gfp_t gfp_mask)
1841 struct cfq_queue *cfqq, *new_cfqq = NULL;
1842 struct cfq_io_context *cic;
1845 cic = cfq_cic_lookup(cfqd, ioc);
1846 /* cic always exists here */
1847 cfqq = cic_to_cfqq(cic, is_sync);
1850 * Always try a new alloc if we fell back to the OOM cfqq
1851 * originally, since it should just be a temporary situation.
1853 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
1858 } else if (gfp_mask & __GFP_WAIT) {
1859 spin_unlock_irq(cfqd->queue->queue_lock);
1860 new_cfqq = kmem_cache_alloc_node(cfq_pool,
1861 gfp_mask | __GFP_ZERO,
1863 spin_lock_irq(cfqd->queue->queue_lock);
1867 cfqq = kmem_cache_alloc_node(cfq_pool,
1868 gfp_mask | __GFP_ZERO,
1873 cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
1874 cfq_init_prio_data(cfqq, ioc);
1875 cfq_log_cfqq(cfqd, cfqq, "alloced");
1877 cfqq = &cfqd->oom_cfqq;
1881 kmem_cache_free(cfq_pool, new_cfqq);
1886 static struct cfq_queue **
1887 cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
1889 switch (ioprio_class) {
1890 case IOPRIO_CLASS_RT:
1891 return &cfqd->async_cfqq[0][ioprio];
1892 case IOPRIO_CLASS_BE:
1893 return &cfqd->async_cfqq[1][ioprio];
1894 case IOPRIO_CLASS_IDLE:
1895 return &cfqd->async_idle_cfqq;
1901 static struct cfq_queue *
1902 cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct io_context *ioc,
1905 const int ioprio = task_ioprio(ioc);
1906 const int ioprio_class = task_ioprio_class(ioc);
1907 struct cfq_queue **async_cfqq = NULL;
1908 struct cfq_queue *cfqq = NULL;
1911 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
1916 cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask);
1919 * pin the queue now that it's allocated, scheduler exit will prune it
1921 if (!is_sync && !(*async_cfqq)) {
1922 atomic_inc(&cfqq->ref);
1926 atomic_inc(&cfqq->ref);
1931 * We drop cfq io contexts lazily, so we may find a dead one.
1934 cfq_drop_dead_cic(struct cfq_data *cfqd, struct io_context *ioc,
1935 struct cfq_io_context *cic)
1937 unsigned long flags;
1939 WARN_ON(!list_empty(&cic->queue_list));
1941 spin_lock_irqsave(&ioc->lock, flags);
1943 BUG_ON(ioc->ioc_data == cic);
1945 radix_tree_delete(&ioc->radix_root, (unsigned long) cfqd);
1946 hlist_del_rcu(&cic->cic_list);
1947 spin_unlock_irqrestore(&ioc->lock, flags);
1952 static struct cfq_io_context *
1953 cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc)
1955 struct cfq_io_context *cic;
1956 unsigned long flags;
1965 * we maintain a last-hit cache, to avoid browsing over the tree
1967 cic = rcu_dereference(ioc->ioc_data);
1968 if (cic && cic->key == cfqd) {
1974 cic = radix_tree_lookup(&ioc->radix_root, (unsigned long) cfqd);
1978 /* ->key must be copied to avoid race with cfq_exit_queue() */
1981 cfq_drop_dead_cic(cfqd, ioc, cic);
1986 spin_lock_irqsave(&ioc->lock, flags);
1987 rcu_assign_pointer(ioc->ioc_data, cic);
1988 spin_unlock_irqrestore(&ioc->lock, flags);
1996 * Add cic into ioc, using cfqd as the search key. This enables us to lookup
1997 * the process specific cfq io context when entered from the block layer.
1998 * Also adds the cic to a per-cfqd list, used when this queue is removed.
2000 static int cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
2001 struct cfq_io_context *cic, gfp_t gfp_mask)
2003 unsigned long flags;
2006 ret = radix_tree_preload(gfp_mask);
2011 spin_lock_irqsave(&ioc->lock, flags);
2012 ret = radix_tree_insert(&ioc->radix_root,
2013 (unsigned long) cfqd, cic);
2015 hlist_add_head_rcu(&cic->cic_list, &ioc->cic_list);
2016 spin_unlock_irqrestore(&ioc->lock, flags);
2018 radix_tree_preload_end();
2021 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2022 list_add(&cic->queue_list, &cfqd->cic_list);
2023 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2028 printk(KERN_ERR "cfq: cic link failed!\n");
2034 * Setup general io context and cfq io context. There can be several cfq
2035 * io contexts per general io context, if this process is doing io to more
2036 * than one device managed by cfq.
2038 static struct cfq_io_context *
2039 cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
2041 struct io_context *ioc = NULL;
2042 struct cfq_io_context *cic;
2044 might_sleep_if(gfp_mask & __GFP_WAIT);
2046 ioc = get_io_context(gfp_mask, cfqd->queue->node);
2050 cic = cfq_cic_lookup(cfqd, ioc);
2054 cic = cfq_alloc_io_context(cfqd, gfp_mask);
2058 if (cfq_cic_link(cfqd, ioc, cic, gfp_mask))
2062 smp_read_barrier_depends();
2063 if (unlikely(ioc->ioprio_changed))
2064 cfq_ioc_set_ioprio(ioc);
2070 put_io_context(ioc);
2075 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
2077 unsigned long elapsed = jiffies - cic->last_end_request;
2078 unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
2080 cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
2081 cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
2082 cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
2086 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2092 if (!cfqq->last_request_pos)
2094 else if (cfqq->last_request_pos < blk_rq_pos(rq))
2095 sdist = blk_rq_pos(rq) - cfqq->last_request_pos;
2097 sdist = cfqq->last_request_pos - blk_rq_pos(rq);
2100 * Don't allow the seek distance to get too large from the
2101 * odd fragment, pagein, etc
2103 if (cfqq->seek_samples <= 60) /* second&third seek */
2104 sdist = min(sdist, (cfqq->seek_mean * 4) + 2*1024*1024);
2106 sdist = min(sdist, (cfqq->seek_mean * 4) + 2*1024*64);
2108 cfqq->seek_samples = (7*cfqq->seek_samples + 256) / 8;
2109 cfqq->seek_total = (7*cfqq->seek_total + (u64)256*sdist) / 8;
2110 total = cfqq->seek_total + (cfqq->seek_samples/2);
2111 do_div(total, cfqq->seek_samples);
2112 cfqq->seek_mean = (sector_t)total;
2115 * If this cfqq is shared between multiple processes, check to
2116 * make sure that those processes are still issuing I/Os within
2117 * the mean seek distance. If not, it may be time to break the
2118 * queues apart again.
2120 if (cfq_cfqq_coop(cfqq)) {
2121 if (CFQQ_SEEKY(cfqq) && !cfqq->seeky_start)
2122 cfqq->seeky_start = jiffies;
2123 else if (!CFQQ_SEEKY(cfqq))
2124 cfqq->seeky_start = 0;
2129 * Disable idle window if the process thinks too long or seeks so much that
2133 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2134 struct cfq_io_context *cic)
2136 int old_idle, enable_idle;
2139 * Don't idle for async or idle io prio class
2141 if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
2144 enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
2146 if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle ||
2147 (!cfqd->cfq_latency && cfqd->hw_tag && CFQQ_SEEKY(cfqq)))
2149 else if (sample_valid(cic->ttime_samples)) {
2150 unsigned int slice_idle = cfqd->cfq_slice_idle;
2151 if (sample_valid(cfqq->seek_samples) && CFQQ_SEEKY(cfqq))
2152 slice_idle = msecs_to_jiffies(CFQ_MIN_TT);
2153 if (cic->ttime_mean > slice_idle)
2159 if (old_idle != enable_idle) {
2160 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
2162 cfq_mark_cfqq_idle_window(cfqq);
2164 cfq_clear_cfqq_idle_window(cfqq);
2169 * Check if new_cfqq should preempt the currently active queue. Return 0 for
2170 * no or if we aren't sure, a 1 will cause a preempt.
2173 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
2176 struct cfq_queue *cfqq;
2178 cfqq = cfqd->active_queue;
2182 if (cfq_slice_used(cfqq))
2185 if (cfq_class_idle(new_cfqq))
2188 if (cfq_class_idle(cfqq))
2192 * if the new request is sync, but the currently running queue is
2193 * not, let the sync request have priority.
2195 if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
2199 * So both queues are sync. Let the new request get disk time if
2200 * it's a metadata request and the current queue is doing regular IO.
2202 if (rq_is_meta(rq) && !cfqq->meta_pending)
2206 * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
2208 if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
2211 if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
2215 * if this request is as-good as one we would expect from the
2216 * current cfqq, let it preempt
2218 if (cfq_rq_close(cfqd, cfqq, rq))
2225 * cfqq preempts the active queue. if we allowed preempt with no slice left,
2226 * let it have half of its nominal slice.
2228 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2230 cfq_log_cfqq(cfqd, cfqq, "preempt");
2231 cfq_slice_expired(cfqd, 1);
2234 * Put the new queue at the front of the of the current list,
2235 * so we know that it will be selected next.
2237 BUG_ON(!cfq_cfqq_on_rr(cfqq));
2239 cfq_service_tree_add(cfqd, cfqq, 1);
2241 cfqq->slice_end = 0;
2242 cfq_mark_cfqq_slice_new(cfqq);
2246 * Called when a new fs request (rq) is added (to cfqq). Check if there's
2247 * something we should do about it
2250 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2253 struct cfq_io_context *cic = RQ_CIC(rq);
2257 cfqq->meta_pending++;
2259 cfq_update_io_thinktime(cfqd, cic);
2260 cfq_update_io_seektime(cfqd, cfqq, rq);
2261 cfq_update_idle_window(cfqd, cfqq, cic);
2263 cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
2265 if (cfqq == cfqd->active_queue) {
2267 * Remember that we saw a request from this process, but
2268 * don't start queuing just yet. Otherwise we risk seeing lots
2269 * of tiny requests, because we disrupt the normal plugging
2270 * and merging. If the request is already larger than a single
2271 * page, let it rip immediately. For that case we assume that
2272 * merging is already done. Ditto for a busy system that
2273 * has other work pending, don't risk delaying until the
2274 * idle timer unplug to continue working.
2276 if (cfq_cfqq_wait_request(cfqq)) {
2277 if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
2278 cfqd->busy_queues > 1) {
2279 del_timer(&cfqd->idle_slice_timer);
2280 __blk_run_queue(cfqd->queue);
2282 cfq_mark_cfqq_must_dispatch(cfqq);
2284 } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
2286 * not the active queue - expire current slice if it is
2287 * idle and has expired it's mean thinktime or this new queue
2288 * has some old slice time left and is of higher priority or
2289 * this new queue is RT and the current one is BE
2291 cfq_preempt_queue(cfqd, cfqq);
2292 __blk_run_queue(cfqd->queue);
2296 static void cfq_insert_request(struct request_queue *q, struct request *rq)
2298 struct cfq_data *cfqd = q->elevator->elevator_data;
2299 struct cfq_queue *cfqq = RQ_CFQQ(rq);
2301 cfq_log_cfqq(cfqd, cfqq, "insert_request");
2302 cfq_init_prio_data(cfqq, RQ_CIC(rq)->ioc);
2304 rq_set_fifo_time(rq, jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)]);
2305 list_add_tail(&rq->queuelist, &cfqq->fifo);
2308 cfq_rq_enqueued(cfqd, cfqq, rq);
2312 * Update hw_tag based on peak queue depth over 50 samples under
2315 static void cfq_update_hw_tag(struct cfq_data *cfqd)
2317 struct cfq_queue *cfqq = cfqd->active_queue;
2319 if (rq_in_driver(cfqd) > cfqd->rq_in_driver_peak)
2320 cfqd->rq_in_driver_peak = rq_in_driver(cfqd);
2322 if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
2323 rq_in_driver(cfqd) <= CFQ_HW_QUEUE_MIN)
2327 * If active queue hasn't enough requests and can idle, cfq might not
2328 * dispatch sufficient requests to hardware. Don't zero hw_tag in this
2331 if (cfqq && cfq_cfqq_idle_window(cfqq) &&
2332 cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] <
2333 CFQ_HW_QUEUE_MIN && rq_in_driver(cfqd) < CFQ_HW_QUEUE_MIN)
2336 if (cfqd->hw_tag_samples++ < 50)
2339 if (cfqd->rq_in_driver_peak >= CFQ_HW_QUEUE_MIN)
2344 cfqd->hw_tag_samples = 0;
2345 cfqd->rq_in_driver_peak = 0;
2348 static void cfq_completed_request(struct request_queue *q, struct request *rq)
2350 struct cfq_queue *cfqq = RQ_CFQQ(rq);
2351 struct cfq_data *cfqd = cfqq->cfqd;
2352 const int sync = rq_is_sync(rq);
2356 cfq_log_cfqq(cfqd, cfqq, "complete");
2358 cfq_update_hw_tag(cfqd);
2360 WARN_ON(!cfqd->rq_in_driver[sync]);
2361 WARN_ON(!cfqq->dispatched);
2362 cfqd->rq_in_driver[sync]--;
2365 if (cfq_cfqq_sync(cfqq))
2366 cfqd->sync_flight--;
2369 RQ_CIC(rq)->last_end_request = now;
2370 cfqd->last_end_sync_rq = now;
2374 * If this is the active queue, check if it needs to be expired,
2375 * or if we want to idle in case it has no pending requests.
2377 if (cfqd->active_queue == cfqq) {
2378 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
2380 if (cfq_cfqq_slice_new(cfqq)) {
2381 cfq_set_prio_slice(cfqd, cfqq);
2382 cfq_clear_cfqq_slice_new(cfqq);
2385 * If there are no requests waiting in this queue, and
2386 * there are other queues ready to issue requests, AND
2387 * those other queues are issuing requests within our
2388 * mean seek distance, give them a chance to run instead
2391 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
2392 cfq_slice_expired(cfqd, 1);
2393 else if (cfqq_empty && !cfq_close_cooperator(cfqd, cfqq) &&
2394 sync && !rq_noidle(rq))
2395 cfq_arm_slice_timer(cfqd);
2398 if (!rq_in_driver(cfqd))
2399 cfq_schedule_dispatch(cfqd);
2403 * we temporarily boost lower priority queues if they are holding fs exclusive
2404 * resources. they are boosted to normal prio (CLASS_BE/4)
2406 static void cfq_prio_boost(struct cfq_queue *cfqq)
2408 if (has_fs_excl()) {
2410 * boost idle prio on transactions that would lock out other
2411 * users of the filesystem
2413 if (cfq_class_idle(cfqq))
2414 cfqq->ioprio_class = IOPRIO_CLASS_BE;
2415 if (cfqq->ioprio > IOPRIO_NORM)
2416 cfqq->ioprio = IOPRIO_NORM;
2419 * check if we need to unboost the queue
2421 if (cfqq->ioprio_class != cfqq->org_ioprio_class)
2422 cfqq->ioprio_class = cfqq->org_ioprio_class;
2423 if (cfqq->ioprio != cfqq->org_ioprio)
2424 cfqq->ioprio = cfqq->org_ioprio;
2428 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
2430 if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
2431 cfq_mark_cfqq_must_alloc_slice(cfqq);
2432 return ELV_MQUEUE_MUST;
2435 return ELV_MQUEUE_MAY;
2438 static int cfq_may_queue(struct request_queue *q, int rw)
2440 struct cfq_data *cfqd = q->elevator->elevator_data;
2441 struct task_struct *tsk = current;
2442 struct cfq_io_context *cic;
2443 struct cfq_queue *cfqq;
2446 * don't force setup of a queue from here, as a call to may_queue
2447 * does not necessarily imply that a request actually will be queued.
2448 * so just lookup a possibly existing queue, or return 'may queue'
2451 cic = cfq_cic_lookup(cfqd, tsk->io_context);
2453 return ELV_MQUEUE_MAY;
2455 cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
2457 cfq_init_prio_data(cfqq, cic->ioc);
2458 cfq_prio_boost(cfqq);
2460 return __cfq_may_queue(cfqq);
2463 return ELV_MQUEUE_MAY;
2467 * queue lock held here
2469 static void cfq_put_request(struct request *rq)
2471 struct cfq_queue *cfqq = RQ_CFQQ(rq);
2474 const int rw = rq_data_dir(rq);
2476 BUG_ON(!cfqq->allocated[rw]);
2477 cfqq->allocated[rw]--;
2479 put_io_context(RQ_CIC(rq)->ioc);
2481 rq->elevator_private = NULL;
2482 rq->elevator_private2 = NULL;
2484 cfq_put_queue(cfqq);
2488 static struct cfq_queue *
2489 cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_context *cic,
2490 struct cfq_queue *cfqq)
2492 cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq);
2493 cic_set_cfqq(cic, cfqq->new_cfqq, 1);
2494 cfq_mark_cfqq_coop(cfqq->new_cfqq);
2495 cfq_put_queue(cfqq);
2496 return cic_to_cfqq(cic, 1);
2499 static int should_split_cfqq(struct cfq_queue *cfqq)
2501 if (cfqq->seeky_start &&
2502 time_after(jiffies, cfqq->seeky_start + CFQQ_COOP_TOUT))
2508 * Returns NULL if a new cfqq should be allocated, or the old cfqq if this
2509 * was the last process referring to said cfqq.
2511 static struct cfq_queue *
2512 split_cfqq(struct cfq_io_context *cic, struct cfq_queue *cfqq)
2514 if (cfqq_process_refs(cfqq) == 1) {
2515 cfqq->seeky_start = 0;
2516 cfqq->pid = current->pid;
2517 cfq_clear_cfqq_coop(cfqq);
2521 cic_set_cfqq(cic, NULL, 1);
2522 cfq_put_queue(cfqq);
2526 * Allocate cfq data structures associated with this request.
2529 cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
2531 struct cfq_data *cfqd = q->elevator->elevator_data;
2532 struct cfq_io_context *cic;
2533 const int rw = rq_data_dir(rq);
2534 const bool is_sync = rq_is_sync(rq);
2535 struct cfq_queue *cfqq;
2536 unsigned long flags;
2538 might_sleep_if(gfp_mask & __GFP_WAIT);
2540 cic = cfq_get_io_context(cfqd, gfp_mask);
2542 spin_lock_irqsave(q->queue_lock, flags);
2548 cfqq = cic_to_cfqq(cic, is_sync);
2549 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
2550 cfqq = cfq_get_queue(cfqd, is_sync, cic->ioc, gfp_mask);
2551 cic_set_cfqq(cic, cfqq, is_sync);
2554 * If the queue was seeky for too long, break it apart.
2556 if (cfq_cfqq_coop(cfqq) && should_split_cfqq(cfqq)) {
2557 cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq");
2558 cfqq = split_cfqq(cic, cfqq);
2564 * Check to see if this queue is scheduled to merge with
2565 * another, closely cooperating queue. The merging of
2566 * queues happens here as it must be done in process context.
2567 * The reference on new_cfqq was taken in merge_cfqqs.
2570 cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq);
2573 cfqq->allocated[rw]++;
2574 atomic_inc(&cfqq->ref);
2576 spin_unlock_irqrestore(q->queue_lock, flags);
2578 rq->elevator_private = cic;
2579 rq->elevator_private2 = cfqq;
2584 put_io_context(cic->ioc);
2586 cfq_schedule_dispatch(cfqd);
2587 spin_unlock_irqrestore(q->queue_lock, flags);
2588 cfq_log(cfqd, "set_request fail");
2592 static void cfq_kick_queue(struct work_struct *work)
2594 struct cfq_data *cfqd =
2595 container_of(work, struct cfq_data, unplug_work);
2596 struct request_queue *q = cfqd->queue;
2598 spin_lock_irq(q->queue_lock);
2599 __blk_run_queue(cfqd->queue);
2600 spin_unlock_irq(q->queue_lock);
2604 * Timer running if the active_queue is currently idling inside its time slice
2606 static void cfq_idle_slice_timer(unsigned long data)
2608 struct cfq_data *cfqd = (struct cfq_data *) data;
2609 struct cfq_queue *cfqq;
2610 unsigned long flags;
2613 cfq_log(cfqd, "idle timer fired");
2615 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2617 cfqq = cfqd->active_queue;
2622 * We saw a request before the queue expired, let it through
2624 if (cfq_cfqq_must_dispatch(cfqq))
2630 if (cfq_slice_used(cfqq))
2634 * only expire and reinvoke request handler, if there are
2635 * other queues with pending requests
2637 if (!cfqd->busy_queues)
2641 * not expired and it has a request pending, let it dispatch
2643 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
2647 cfq_slice_expired(cfqd, timed_out);
2649 cfq_schedule_dispatch(cfqd);
2651 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2654 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
2656 del_timer_sync(&cfqd->idle_slice_timer);
2657 cancel_work_sync(&cfqd->unplug_work);
2660 static void cfq_put_async_queues(struct cfq_data *cfqd)
2664 for (i = 0; i < IOPRIO_BE_NR; i++) {
2665 if (cfqd->async_cfqq[0][i])
2666 cfq_put_queue(cfqd->async_cfqq[0][i]);
2667 if (cfqd->async_cfqq[1][i])
2668 cfq_put_queue(cfqd->async_cfqq[1][i]);
2671 if (cfqd->async_idle_cfqq)
2672 cfq_put_queue(cfqd->async_idle_cfqq);
2675 static void cfq_exit_queue(struct elevator_queue *e)
2677 struct cfq_data *cfqd = e->elevator_data;
2678 struct request_queue *q = cfqd->queue;
2680 cfq_shutdown_timer_wq(cfqd);
2682 spin_lock_irq(q->queue_lock);
2684 if (cfqd->active_queue)
2685 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
2687 while (!list_empty(&cfqd->cic_list)) {
2688 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
2689 struct cfq_io_context,
2692 __cfq_exit_single_io_context(cfqd, cic);
2695 cfq_put_async_queues(cfqd);
2697 spin_unlock_irq(q->queue_lock);
2699 cfq_shutdown_timer_wq(cfqd);
2704 static void *cfq_init_queue(struct request_queue *q)
2706 struct cfq_data *cfqd;
2709 cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
2713 cfqd->service_tree = CFQ_RB_ROOT;
2716 * Not strictly needed (since RB_ROOT just clears the node and we
2717 * zeroed cfqd on alloc), but better be safe in case someone decides
2718 * to add magic to the rb code
2720 for (i = 0; i < CFQ_PRIO_LISTS; i++)
2721 cfqd->prio_trees[i] = RB_ROOT;
2724 * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
2725 * Grab a permanent reference to it, so that the normal code flow
2726 * will not attempt to free it.
2728 cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
2729 atomic_inc(&cfqd->oom_cfqq.ref);
2731 INIT_LIST_HEAD(&cfqd->cic_list);
2735 init_timer(&cfqd->idle_slice_timer);
2736 cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
2737 cfqd->idle_slice_timer.data = (unsigned long) cfqd;
2739 INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
2741 cfqd->cfq_quantum = cfq_quantum;
2742 cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
2743 cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
2744 cfqd->cfq_back_max = cfq_back_max;
2745 cfqd->cfq_back_penalty = cfq_back_penalty;
2746 cfqd->cfq_slice[0] = cfq_slice_async;
2747 cfqd->cfq_slice[1] = cfq_slice_sync;
2748 cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
2749 cfqd->cfq_slice_idle = cfq_slice_idle;
2750 cfqd->cfq_latency = 1;
2752 cfqd->last_end_sync_rq = jiffies;
2756 static void cfq_slab_kill(void)
2759 * Caller already ensured that pending RCU callbacks are completed,
2760 * so we should have no busy allocations at this point.
2763 kmem_cache_destroy(cfq_pool);
2765 kmem_cache_destroy(cfq_ioc_pool);
2768 static int __init cfq_slab_setup(void)
2770 cfq_pool = KMEM_CACHE(cfq_queue, 0);
2774 cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0);
2785 * sysfs parts below -->
2788 cfq_var_show(unsigned int var, char *page)
2790 return sprintf(page, "%d\n", var);
2794 cfq_var_store(unsigned int *var, const char *page, size_t count)
2796 char *p = (char *) page;
2798 *var = simple_strtoul(p, &p, 10);
2802 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
2803 static ssize_t __FUNC(struct elevator_queue *e, char *page) \
2805 struct cfq_data *cfqd = e->elevator_data; \
2806 unsigned int __data = __VAR; \
2808 __data = jiffies_to_msecs(__data); \
2809 return cfq_var_show(__data, (page)); \
2811 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
2812 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
2813 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
2814 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
2815 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
2816 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
2817 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
2818 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
2819 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
2820 SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
2821 #undef SHOW_FUNCTION
2823 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
2824 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
2826 struct cfq_data *cfqd = e->elevator_data; \
2827 unsigned int __data; \
2828 int ret = cfq_var_store(&__data, (page), count); \
2829 if (__data < (MIN)) \
2831 else if (__data > (MAX)) \
2834 *(__PTR) = msecs_to_jiffies(__data); \
2836 *(__PTR) = __data; \
2839 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
2840 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
2842 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
2844 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
2845 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
2847 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
2848 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
2849 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
2850 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
2852 STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
2853 #undef STORE_FUNCTION
2855 #define CFQ_ATTR(name) \
2856 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
2858 static struct elv_fs_entry cfq_attrs[] = {
2860 CFQ_ATTR(fifo_expire_sync),
2861 CFQ_ATTR(fifo_expire_async),
2862 CFQ_ATTR(back_seek_max),
2863 CFQ_ATTR(back_seek_penalty),
2864 CFQ_ATTR(slice_sync),
2865 CFQ_ATTR(slice_async),
2866 CFQ_ATTR(slice_async_rq),
2867 CFQ_ATTR(slice_idle),
2868 CFQ_ATTR(low_latency),
2872 static struct elevator_type iosched_cfq = {
2874 .elevator_merge_fn = cfq_merge,
2875 .elevator_merged_fn = cfq_merged_request,
2876 .elevator_merge_req_fn = cfq_merged_requests,
2877 .elevator_allow_merge_fn = cfq_allow_merge,
2878 .elevator_dispatch_fn = cfq_dispatch_requests,
2879 .elevator_add_req_fn = cfq_insert_request,
2880 .elevator_activate_req_fn = cfq_activate_request,
2881 .elevator_deactivate_req_fn = cfq_deactivate_request,
2882 .elevator_queue_empty_fn = cfq_queue_empty,
2883 .elevator_completed_req_fn = cfq_completed_request,
2884 .elevator_former_req_fn = elv_rb_former_request,
2885 .elevator_latter_req_fn = elv_rb_latter_request,
2886 .elevator_set_req_fn = cfq_set_request,
2887 .elevator_put_req_fn = cfq_put_request,
2888 .elevator_may_queue_fn = cfq_may_queue,
2889 .elevator_init_fn = cfq_init_queue,
2890 .elevator_exit_fn = cfq_exit_queue,
2891 .trim = cfq_free_io_context,
2893 .elevator_attrs = cfq_attrs,
2894 .elevator_name = "cfq",
2895 .elevator_owner = THIS_MODULE,
2898 static int __init cfq_init(void)
2901 * could be 0 on HZ < 1000 setups
2903 if (!cfq_slice_async)
2904 cfq_slice_async = 1;
2905 if (!cfq_slice_idle)
2908 if (cfq_slab_setup())
2911 elv_register(&iosched_cfq);
2916 static void __exit cfq_exit(void)
2918 DECLARE_COMPLETION_ONSTACK(all_gone);
2919 elv_unregister(&iosched_cfq);
2920 ioc_gone = &all_gone;
2921 /* ioc_gone's update must be visible before reading ioc_count */
2925 * this also protects us from entering cfq_slab_kill() with
2926 * pending RCU callbacks
2928 if (elv_ioc_count_read(cfq_ioc_count))
2929 wait_for_completion(&all_gone);
2933 module_init(cfq_init);
2934 module_exit(cfq_exit);
2936 MODULE_AUTHOR("Jens Axboe");
2937 MODULE_LICENSE("GPL");
2938 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");