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 * Index in the service_trees.
138 * IDLE is handled separately, so it has negative index
147 * Per block device queue structure
150 struct request_queue *queue;
153 * rr lists of queues with requests, onle rr for each priority class.
154 * Counts are embedded in the cfq_rb_root
156 struct cfq_rb_root service_trees[2];
157 struct cfq_rb_root service_tree_idle;
159 * The priority currently being served
161 enum wl_prio_t serving_prio;
164 * Each priority tree is sorted by next_request position. These
165 * trees are used when determining if two or more queues are
166 * interleaving requests (see cfq_close_cooperator).
168 struct rb_root prio_trees[CFQ_PRIO_LISTS];
170 unsigned int busy_queues;
171 unsigned int busy_queues_avg[2];
177 * queue-depth detection
182 int rq_in_driver_peak;
185 * idle window management
187 struct timer_list idle_slice_timer;
188 struct work_struct unplug_work;
190 struct cfq_queue *active_queue;
191 struct cfq_io_context *active_cic;
194 * async queue for each priority case
196 struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
197 struct cfq_queue *async_idle_cfqq;
199 sector_t last_position;
202 * tunables, see top of file
204 unsigned int cfq_quantum;
205 unsigned int cfq_fifo_expire[2];
206 unsigned int cfq_back_penalty;
207 unsigned int cfq_back_max;
208 unsigned int cfq_slice[2];
209 unsigned int cfq_slice_async_rq;
210 unsigned int cfq_slice_idle;
211 unsigned int cfq_latency;
213 struct list_head cic_list;
216 * Fallback dummy cfqq for extreme OOM conditions
218 struct cfq_queue oom_cfqq;
220 unsigned long last_end_sync_rq;
223 static struct cfq_rb_root *service_tree_for(enum wl_prio_t prio,
224 struct cfq_data *cfqd)
226 if (prio == IDLE_WORKLOAD)
227 return &cfqd->service_tree_idle;
229 return &cfqd->service_trees[prio];
232 enum cfqq_state_flags {
233 CFQ_CFQQ_FLAG_on_rr = 0, /* on round-robin busy list */
234 CFQ_CFQQ_FLAG_wait_request, /* waiting for a request */
235 CFQ_CFQQ_FLAG_must_dispatch, /* must be allowed a dispatch */
236 CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
237 CFQ_CFQQ_FLAG_fifo_expire, /* FIFO checked in this slice */
238 CFQ_CFQQ_FLAG_idle_window, /* slice idling enabled */
239 CFQ_CFQQ_FLAG_prio_changed, /* task priority has changed */
240 CFQ_CFQQ_FLAG_slice_new, /* no requests dispatched in slice */
241 CFQ_CFQQ_FLAG_sync, /* synchronous queue */
242 CFQ_CFQQ_FLAG_coop, /* cfqq is shared */
245 #define CFQ_CFQQ_FNS(name) \
246 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \
248 (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name); \
250 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \
252 (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \
254 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \
256 return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \
260 CFQ_CFQQ_FNS(wait_request);
261 CFQ_CFQQ_FNS(must_dispatch);
262 CFQ_CFQQ_FNS(must_alloc_slice);
263 CFQ_CFQQ_FNS(fifo_expire);
264 CFQ_CFQQ_FNS(idle_window);
265 CFQ_CFQQ_FNS(prio_changed);
266 CFQ_CFQQ_FNS(slice_new);
271 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
272 blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args)
273 #define cfq_log(cfqd, fmt, args...) \
274 blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
276 static inline enum wl_prio_t cfqq_prio(struct cfq_queue *cfqq)
278 if (cfq_class_idle(cfqq))
279 return IDLE_WORKLOAD;
280 if (cfq_class_rt(cfqq))
285 static inline int cfq_busy_queues_wl(enum wl_prio_t wl, struct cfq_data *cfqd)
287 if (wl == IDLE_WORKLOAD)
288 return cfqd->service_tree_idle.count;
290 return cfqd->service_trees[wl].count;
293 static void cfq_dispatch_insert(struct request_queue *, struct request *);
294 static struct cfq_queue *cfq_get_queue(struct cfq_data *, bool,
295 struct io_context *, gfp_t);
296 static struct cfq_io_context *cfq_cic_lookup(struct cfq_data *,
297 struct io_context *);
299 static inline int rq_in_driver(struct cfq_data *cfqd)
301 return cfqd->rq_in_driver[0] + cfqd->rq_in_driver[1];
304 static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_context *cic,
307 return cic->cfqq[is_sync];
310 static inline void cic_set_cfqq(struct cfq_io_context *cic,
311 struct cfq_queue *cfqq, bool is_sync)
313 cic->cfqq[is_sync] = cfqq;
317 * We regard a request as SYNC, if it's either a read or has the SYNC bit
318 * set (in which case it could also be direct WRITE).
320 static inline bool cfq_bio_sync(struct bio *bio)
322 return bio_data_dir(bio) == READ || bio_rw_flagged(bio, BIO_RW_SYNCIO);
326 * scheduler run of queue, if there are requests pending and no one in the
327 * driver that will restart queueing
329 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
331 if (cfqd->busy_queues) {
332 cfq_log(cfqd, "schedule dispatch");
333 kblockd_schedule_work(cfqd->queue, &cfqd->unplug_work);
337 static int cfq_queue_empty(struct request_queue *q)
339 struct cfq_data *cfqd = q->elevator->elevator_data;
341 return !cfqd->busy_queues;
345 * Scale schedule slice based on io priority. Use the sync time slice only
346 * if a queue is marked sync and has sync io queued. A sync queue with async
347 * io only, should not get full sync slice length.
349 static inline int cfq_prio_slice(struct cfq_data *cfqd, bool sync,
352 const int base_slice = cfqd->cfq_slice[sync];
354 WARN_ON(prio >= IOPRIO_BE_NR);
356 return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
360 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
362 return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
366 * get averaged number of queues of RT/BE priority.
367 * average is updated, with a formula that gives more weight to higher numbers,
368 * to quickly follows sudden increases and decrease slowly
371 static inline unsigned
372 cfq_get_avg_queues(struct cfq_data *cfqd, bool rt) {
373 unsigned min_q, max_q;
374 unsigned mult = cfq_hist_divisor - 1;
375 unsigned round = cfq_hist_divisor / 2;
376 unsigned busy = cfq_busy_queues_wl(rt, cfqd);
378 min_q = min(cfqd->busy_queues_avg[rt], busy);
379 max_q = max(cfqd->busy_queues_avg[rt], busy);
380 cfqd->busy_queues_avg[rt] = (mult * max_q + min_q + round) /
382 return cfqd->busy_queues_avg[rt];
386 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
388 unsigned slice = cfq_prio_to_slice(cfqd, cfqq);
389 if (cfqd->cfq_latency) {
390 /* interested queues (we consider only the ones with the same
392 unsigned iq = cfq_get_avg_queues(cfqd, cfq_class_rt(cfqq));
393 unsigned sync_slice = cfqd->cfq_slice[1];
394 unsigned expect_latency = sync_slice * iq;
395 if (expect_latency > cfq_target_latency) {
396 unsigned base_low_slice = 2 * cfqd->cfq_slice_idle;
397 /* scale low_slice according to IO priority
398 * and sync vs async */
400 min(slice, base_low_slice * slice / sync_slice);
401 /* the adapted slice value is scaled to fit all iqs
402 * into the target latency */
403 slice = max(slice * cfq_target_latency / expect_latency,
407 cfqq->slice_end = jiffies + slice;
408 cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
412 * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
413 * isn't valid until the first request from the dispatch is activated
414 * and the slice time set.
416 static inline bool cfq_slice_used(struct cfq_queue *cfqq)
418 if (cfq_cfqq_slice_new(cfqq))
420 if (time_before(jiffies, cfqq->slice_end))
427 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
428 * We choose the request that is closest to the head right now. Distance
429 * behind the head is penalized and only allowed to a certain extent.
431 static struct request *
432 cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2)
434 sector_t last, s1, s2, d1 = 0, d2 = 0;
435 unsigned long back_max;
436 #define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */
437 #define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */
438 unsigned wrap = 0; /* bit mask: requests behind the disk head? */
440 if (rq1 == NULL || rq1 == rq2)
445 if (rq_is_sync(rq1) && !rq_is_sync(rq2))
447 else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
449 if (rq_is_meta(rq1) && !rq_is_meta(rq2))
451 else if (rq_is_meta(rq2) && !rq_is_meta(rq1))
454 s1 = blk_rq_pos(rq1);
455 s2 = blk_rq_pos(rq2);
457 last = cfqd->last_position;
460 * by definition, 1KiB is 2 sectors
462 back_max = cfqd->cfq_back_max * 2;
465 * Strict one way elevator _except_ in the case where we allow
466 * short backward seeks which are biased as twice the cost of a
467 * similar forward seek.
471 else if (s1 + back_max >= last)
472 d1 = (last - s1) * cfqd->cfq_back_penalty;
474 wrap |= CFQ_RQ1_WRAP;
478 else if (s2 + back_max >= last)
479 d2 = (last - s2) * cfqd->cfq_back_penalty;
481 wrap |= CFQ_RQ2_WRAP;
483 /* Found required data */
486 * By doing switch() on the bit mask "wrap" we avoid having to
487 * check two variables for all permutations: --> faster!
490 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
506 case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
509 * Since both rqs are wrapped,
510 * start with the one that's further behind head
511 * (--> only *one* back seek required),
512 * since back seek takes more time than forward.
522 * The below is leftmost cache rbtree addon
524 static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
527 root->left = rb_first(&root->rb);
530 return rb_entry(root->left, struct cfq_queue, rb_node);
535 static void rb_erase_init(struct rb_node *n, struct rb_root *root)
541 static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
545 rb_erase_init(n, &root->rb);
550 * would be nice to take fifo expire time into account as well
552 static struct request *
553 cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
554 struct request *last)
556 struct rb_node *rbnext = rb_next(&last->rb_node);
557 struct rb_node *rbprev = rb_prev(&last->rb_node);
558 struct request *next = NULL, *prev = NULL;
560 BUG_ON(RB_EMPTY_NODE(&last->rb_node));
563 prev = rb_entry_rq(rbprev);
566 next = rb_entry_rq(rbnext);
568 rbnext = rb_first(&cfqq->sort_list);
569 if (rbnext && rbnext != &last->rb_node)
570 next = rb_entry_rq(rbnext);
573 return cfq_choose_req(cfqd, next, prev);
576 static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
577 struct cfq_queue *cfqq)
580 * just an approximation, should be ok.
582 return (cfqd->busy_queues - 1) * (cfq_prio_slice(cfqd, 1, 0) -
583 cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
587 * The cfqd->service_trees holds all pending cfq_queue's that have
588 * requests waiting to be processed. It is sorted in the order that
589 * we will service the queues.
591 static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
594 struct rb_node **p, *parent;
595 struct cfq_queue *__cfqq;
596 unsigned long rb_key;
597 struct cfq_rb_root *service_tree;
600 service_tree = service_tree_for(cfqq_prio(cfqq), cfqd);
601 if (cfq_class_idle(cfqq)) {
602 rb_key = CFQ_IDLE_DELAY;
603 parent = rb_last(&service_tree->rb);
604 if (parent && parent != &cfqq->rb_node) {
605 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
606 rb_key += __cfqq->rb_key;
609 } else if (!add_front) {
611 * Get our rb key offset. Subtract any residual slice
612 * value carried from last service. A negative resid
613 * count indicates slice overrun, and this should position
614 * the next service time further away in the tree.
616 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
617 rb_key -= cfqq->slice_resid;
618 cfqq->slice_resid = 0;
621 __cfqq = cfq_rb_first(service_tree);
622 rb_key += __cfqq ? __cfqq->rb_key : jiffies;
625 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
627 * same position, nothing more to do
629 if (rb_key == cfqq->rb_key &&
630 cfqq->service_tree == service_tree)
633 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
634 cfqq->service_tree = NULL;
639 cfqq->service_tree = service_tree;
640 p = &service_tree->rb.rb_node;
645 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
648 * sort by key, that represents service time.
650 if (time_before(rb_key, __cfqq->rb_key))
661 service_tree->left = &cfqq->rb_node;
663 cfqq->rb_key = rb_key;
664 rb_link_node(&cfqq->rb_node, parent, p);
665 rb_insert_color(&cfqq->rb_node, &service_tree->rb);
666 service_tree->count++;
669 static struct cfq_queue *
670 cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
671 sector_t sector, struct rb_node **ret_parent,
672 struct rb_node ***rb_link)
674 struct rb_node **p, *parent;
675 struct cfq_queue *cfqq = NULL;
683 cfqq = rb_entry(parent, struct cfq_queue, p_node);
686 * Sort strictly based on sector. Smallest to the left,
687 * largest to the right.
689 if (sector > blk_rq_pos(cfqq->next_rq))
691 else if (sector < blk_rq_pos(cfqq->next_rq))
699 *ret_parent = parent;
705 static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
707 struct rb_node **p, *parent;
708 struct cfq_queue *__cfqq;
711 rb_erase(&cfqq->p_node, cfqq->p_root);
715 if (cfq_class_idle(cfqq))
720 cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
721 __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
722 blk_rq_pos(cfqq->next_rq), &parent, &p);
724 rb_link_node(&cfqq->p_node, parent, p);
725 rb_insert_color(&cfqq->p_node, cfqq->p_root);
731 * Update cfqq's position in the service tree.
733 static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
736 * Resorting requires the cfqq to be on the RR list already.
738 if (cfq_cfqq_on_rr(cfqq)) {
739 cfq_service_tree_add(cfqd, cfqq, 0);
740 cfq_prio_tree_add(cfqd, cfqq);
745 * add to busy list of queues for service, trying to be fair in ordering
746 * the pending list according to last request service
748 static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
750 cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
751 BUG_ON(cfq_cfqq_on_rr(cfqq));
752 cfq_mark_cfqq_on_rr(cfqq);
755 cfq_resort_rr_list(cfqd, cfqq);
759 * Called when the cfqq no longer has requests pending, remove it from
762 static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
764 cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
765 BUG_ON(!cfq_cfqq_on_rr(cfqq));
766 cfq_clear_cfqq_on_rr(cfqq);
768 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
769 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
770 cfqq->service_tree = NULL;
773 rb_erase(&cfqq->p_node, cfqq->p_root);
777 BUG_ON(!cfqd->busy_queues);
782 * rb tree support functions
784 static void cfq_del_rq_rb(struct request *rq)
786 struct cfq_queue *cfqq = RQ_CFQQ(rq);
787 struct cfq_data *cfqd = cfqq->cfqd;
788 const int sync = rq_is_sync(rq);
790 BUG_ON(!cfqq->queued[sync]);
791 cfqq->queued[sync]--;
793 elv_rb_del(&cfqq->sort_list, rq);
795 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
796 cfq_del_cfqq_rr(cfqd, cfqq);
799 static void cfq_add_rq_rb(struct request *rq)
801 struct cfq_queue *cfqq = RQ_CFQQ(rq);
802 struct cfq_data *cfqd = cfqq->cfqd;
803 struct request *__alias, *prev;
805 cfqq->queued[rq_is_sync(rq)]++;
808 * looks a little odd, but the first insert might return an alias.
809 * if that happens, put the alias on the dispatch list
811 while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL)
812 cfq_dispatch_insert(cfqd->queue, __alias);
814 if (!cfq_cfqq_on_rr(cfqq))
815 cfq_add_cfqq_rr(cfqd, cfqq);
818 * check if this request is a better next-serve candidate
820 prev = cfqq->next_rq;
821 cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq);
824 * adjust priority tree position, if ->next_rq changes
826 if (prev != cfqq->next_rq)
827 cfq_prio_tree_add(cfqd, cfqq);
829 BUG_ON(!cfqq->next_rq);
832 static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
834 elv_rb_del(&cfqq->sort_list, rq);
835 cfqq->queued[rq_is_sync(rq)]--;
839 static struct request *
840 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
842 struct task_struct *tsk = current;
843 struct cfq_io_context *cic;
844 struct cfq_queue *cfqq;
846 cic = cfq_cic_lookup(cfqd, tsk->io_context);
850 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
852 sector_t sector = bio->bi_sector + bio_sectors(bio);
854 return elv_rb_find(&cfqq->sort_list, sector);
860 static void cfq_activate_request(struct request_queue *q, struct request *rq)
862 struct cfq_data *cfqd = q->elevator->elevator_data;
864 cfqd->rq_in_driver[rq_is_sync(rq)]++;
865 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
868 cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
871 static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
873 struct cfq_data *cfqd = q->elevator->elevator_data;
874 const int sync = rq_is_sync(rq);
876 WARN_ON(!cfqd->rq_in_driver[sync]);
877 cfqd->rq_in_driver[sync]--;
878 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
882 static void cfq_remove_request(struct request *rq)
884 struct cfq_queue *cfqq = RQ_CFQQ(rq);
886 if (cfqq->next_rq == rq)
887 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
889 list_del_init(&rq->queuelist);
892 cfqq->cfqd->rq_queued--;
893 if (rq_is_meta(rq)) {
894 WARN_ON(!cfqq->meta_pending);
895 cfqq->meta_pending--;
899 static int cfq_merge(struct request_queue *q, struct request **req,
902 struct cfq_data *cfqd = q->elevator->elevator_data;
903 struct request *__rq;
905 __rq = cfq_find_rq_fmerge(cfqd, bio);
906 if (__rq && elv_rq_merge_ok(__rq, bio)) {
908 return ELEVATOR_FRONT_MERGE;
911 return ELEVATOR_NO_MERGE;
914 static void cfq_merged_request(struct request_queue *q, struct request *req,
917 if (type == ELEVATOR_FRONT_MERGE) {
918 struct cfq_queue *cfqq = RQ_CFQQ(req);
920 cfq_reposition_rq_rb(cfqq, req);
925 cfq_merged_requests(struct request_queue *q, struct request *rq,
926 struct request *next)
929 * reposition in fifo if next is older than rq
931 if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
932 time_before(rq_fifo_time(next), rq_fifo_time(rq))) {
933 list_move(&rq->queuelist, &next->queuelist);
934 rq_set_fifo_time(rq, rq_fifo_time(next));
937 cfq_remove_request(next);
940 static int cfq_allow_merge(struct request_queue *q, struct request *rq,
943 struct cfq_data *cfqd = q->elevator->elevator_data;
944 struct cfq_io_context *cic;
945 struct cfq_queue *cfqq;
948 * Disallow merge of a sync bio into an async request.
950 if (cfq_bio_sync(bio) && !rq_is_sync(rq))
954 * Lookup the cfqq that this bio will be queued with. Allow
955 * merge only if rq is queued there.
957 cic = cfq_cic_lookup(cfqd, current->io_context);
961 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
962 return cfqq == RQ_CFQQ(rq);
965 static void __cfq_set_active_queue(struct cfq_data *cfqd,
966 struct cfq_queue *cfqq)
969 cfq_log_cfqq(cfqd, cfqq, "set_active");
971 cfqq->slice_dispatch = 0;
973 cfq_clear_cfqq_wait_request(cfqq);
974 cfq_clear_cfqq_must_dispatch(cfqq);
975 cfq_clear_cfqq_must_alloc_slice(cfqq);
976 cfq_clear_cfqq_fifo_expire(cfqq);
977 cfq_mark_cfqq_slice_new(cfqq);
979 del_timer(&cfqd->idle_slice_timer);
982 cfqd->active_queue = cfqq;
986 * current cfqq expired its slice (or was too idle), select new one
989 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
992 cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
994 if (cfq_cfqq_wait_request(cfqq))
995 del_timer(&cfqd->idle_slice_timer);
997 cfq_clear_cfqq_wait_request(cfqq);
1000 * store what was left of this slice, if the queue idled/timed out
1002 if (timed_out && !cfq_cfqq_slice_new(cfqq)) {
1003 cfqq->slice_resid = cfqq->slice_end - jiffies;
1004 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
1007 cfq_resort_rr_list(cfqd, cfqq);
1009 if (cfqq == cfqd->active_queue)
1010 cfqd->active_queue = NULL;
1012 if (cfqd->active_cic) {
1013 put_io_context(cfqd->active_cic->ioc);
1014 cfqd->active_cic = NULL;
1018 static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out)
1020 struct cfq_queue *cfqq = cfqd->active_queue;
1023 __cfq_slice_expired(cfqd, cfqq, timed_out);
1027 * Get next queue for service. Unless we have a queue preemption,
1028 * we'll simply select the first cfqq in the service tree.
1030 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
1032 struct cfq_rb_root *service_tree =
1033 service_tree_for(cfqd->serving_prio, cfqd);
1035 if (RB_EMPTY_ROOT(&service_tree->rb))
1037 return cfq_rb_first(service_tree);
1041 * Get and set a new active queue for service.
1043 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
1044 struct cfq_queue *cfqq)
1047 cfqq = cfq_get_next_queue(cfqd);
1049 __cfq_set_active_queue(cfqd, cfqq);
1053 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
1056 if (blk_rq_pos(rq) >= cfqd->last_position)
1057 return blk_rq_pos(rq) - cfqd->last_position;
1059 return cfqd->last_position - blk_rq_pos(rq);
1062 #define CFQQ_SEEK_THR 8 * 1024
1063 #define CFQQ_SEEKY(cfqq) ((cfqq)->seek_mean > CFQQ_SEEK_THR)
1065 static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1068 sector_t sdist = cfqq->seek_mean;
1070 if (!sample_valid(cfqq->seek_samples))
1071 sdist = CFQQ_SEEK_THR;
1073 return cfq_dist_from_last(cfqd, rq) <= sdist;
1076 static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
1077 struct cfq_queue *cur_cfqq)
1079 struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
1080 struct rb_node *parent, *node;
1081 struct cfq_queue *__cfqq;
1082 sector_t sector = cfqd->last_position;
1084 if (RB_EMPTY_ROOT(root))
1088 * First, if we find a request starting at the end of the last
1089 * request, choose it.
1091 __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
1096 * If the exact sector wasn't found, the parent of the NULL leaf
1097 * will contain the closest sector.
1099 __cfqq = rb_entry(parent, struct cfq_queue, p_node);
1100 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1103 if (blk_rq_pos(__cfqq->next_rq) < sector)
1104 node = rb_next(&__cfqq->p_node);
1106 node = rb_prev(&__cfqq->p_node);
1110 __cfqq = rb_entry(node, struct cfq_queue, p_node);
1111 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1119 * cur_cfqq - passed in so that we don't decide that the current queue is
1120 * closely cooperating with itself.
1122 * So, basically we're assuming that that cur_cfqq has dispatched at least
1123 * one request, and that cfqd->last_position reflects a position on the disk
1124 * associated with the I/O issued by cur_cfqq. I'm not sure this is a valid
1127 static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
1128 struct cfq_queue *cur_cfqq)
1130 struct cfq_queue *cfqq;
1132 if (!cfq_cfqq_sync(cur_cfqq))
1134 if (CFQQ_SEEKY(cur_cfqq))
1138 * We should notice if some of the queues are cooperating, eg
1139 * working closely on the same area of the disk. In that case,
1140 * we can group them together and don't waste time idling.
1142 cfqq = cfqq_close(cfqd, cur_cfqq);
1147 * It only makes sense to merge sync queues.
1149 if (!cfq_cfqq_sync(cfqq))
1151 if (CFQQ_SEEKY(cfqq))
1155 * Do not merge queues of different priority classes
1157 if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq))
1163 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
1165 struct cfq_queue *cfqq = cfqd->active_queue;
1166 struct cfq_io_context *cic;
1170 * SSD device without seek penalty, disable idling. But only do so
1171 * for devices that support queuing, otherwise we still have a problem
1172 * with sync vs async workloads.
1174 if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
1177 WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
1178 WARN_ON(cfq_cfqq_slice_new(cfqq));
1181 * idle is disabled, either manually or by past process history
1183 if (!cfqd->cfq_slice_idle || !cfq_cfqq_idle_window(cfqq))
1187 * still requests with the driver, don't idle
1189 if (rq_in_driver(cfqd))
1193 * task has exited, don't wait
1195 cic = cfqd->active_cic;
1196 if (!cic || !atomic_read(&cic->ioc->nr_tasks))
1200 * If our average think time is larger than the remaining time
1201 * slice, then don't idle. This avoids overrunning the allotted
1204 if (sample_valid(cic->ttime_samples) &&
1205 (cfqq->slice_end - jiffies < cic->ttime_mean))
1208 cfq_mark_cfqq_wait_request(cfqq);
1211 * we don't want to idle for seeks, but we do want to allow
1212 * fair distribution of slice time for a process doing back-to-back
1213 * seeks. so allow a little bit of time for him to submit a new rq
1215 sl = cfqd->cfq_slice_idle;
1216 if (sample_valid(cfqq->seek_samples) && CFQQ_SEEKY(cfqq))
1217 sl = min(sl, msecs_to_jiffies(CFQ_MIN_TT));
1219 mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
1220 cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu", sl);
1224 * Move request from internal lists to the request queue dispatch list.
1226 static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
1228 struct cfq_data *cfqd = q->elevator->elevator_data;
1229 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1231 cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
1233 cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
1234 cfq_remove_request(rq);
1236 elv_dispatch_sort(q, rq);
1238 if (cfq_cfqq_sync(cfqq))
1239 cfqd->sync_flight++;
1243 * return expired entry, or NULL to just start from scratch in rbtree
1245 static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
1247 struct request *rq = NULL;
1249 if (cfq_cfqq_fifo_expire(cfqq))
1252 cfq_mark_cfqq_fifo_expire(cfqq);
1254 if (list_empty(&cfqq->fifo))
1257 rq = rq_entry_fifo(cfqq->fifo.next);
1258 if (time_before(jiffies, rq_fifo_time(rq)))
1261 cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq);
1266 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1268 const int base_rq = cfqd->cfq_slice_async_rq;
1270 WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
1272 return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
1276 * Must be called with the queue_lock held.
1278 static int cfqq_process_refs(struct cfq_queue *cfqq)
1280 int process_refs, io_refs;
1282 io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE];
1283 process_refs = atomic_read(&cfqq->ref) - io_refs;
1284 BUG_ON(process_refs < 0);
1285 return process_refs;
1288 static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq)
1290 int process_refs, new_process_refs;
1291 struct cfq_queue *__cfqq;
1293 /* Avoid a circular list and skip interim queue merges */
1294 while ((__cfqq = new_cfqq->new_cfqq)) {
1300 process_refs = cfqq_process_refs(cfqq);
1302 * If the process for the cfqq has gone away, there is no
1303 * sense in merging the queues.
1305 if (process_refs == 0)
1309 * Merge in the direction of the lesser amount of work.
1311 new_process_refs = cfqq_process_refs(new_cfqq);
1312 if (new_process_refs >= process_refs) {
1313 cfqq->new_cfqq = new_cfqq;
1314 atomic_add(process_refs, &new_cfqq->ref);
1316 new_cfqq->new_cfqq = cfqq;
1317 atomic_add(new_process_refs, &cfqq->ref);
1322 * Select a queue for service. If we have a current active queue,
1323 * check whether to continue servicing it, or retrieve and set a new one.
1325 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
1327 struct cfq_queue *cfqq, *new_cfqq = NULL;
1329 cfqq = cfqd->active_queue;
1334 * The active queue has run out of time, expire it and select new.
1336 if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq))
1340 * The active queue has requests and isn't expired, allow it to
1343 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
1347 * If another queue has a request waiting within our mean seek
1348 * distance, let it run. The expire code will check for close
1349 * cooperators and put the close queue at the front of the service
1350 * tree. If possible, merge the expiring queue with the new cfqq.
1352 new_cfqq = cfq_close_cooperator(cfqd, cfqq);
1354 if (!cfqq->new_cfqq)
1355 cfq_setup_merge(cfqq, new_cfqq);
1360 * No requests pending. If the active queue still has requests in
1361 * flight or is idling for a new request, allow either of these
1362 * conditions to happen (or time out) before selecting a new queue.
1364 if (timer_pending(&cfqd->idle_slice_timer) ||
1365 (cfqq->dispatched && cfq_cfqq_idle_window(cfqq))) {
1371 cfq_slice_expired(cfqd, 0);
1374 if (cfq_busy_queues_wl(RT_WORKLOAD, cfqd))
1375 cfqd->serving_prio = RT_WORKLOAD;
1376 else if (cfq_busy_queues_wl(BE_WORKLOAD, cfqd))
1377 cfqd->serving_prio = BE_WORKLOAD;
1379 cfqd->serving_prio = IDLE_WORKLOAD;
1381 cfqq = cfq_set_active_queue(cfqd, new_cfqq);
1386 static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
1390 while (cfqq->next_rq) {
1391 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
1395 BUG_ON(!list_empty(&cfqq->fifo));
1400 * Drain our current requests. Used for barriers and when switching
1401 * io schedulers on-the-fly.
1403 static int cfq_forced_dispatch(struct cfq_data *cfqd)
1405 struct cfq_queue *cfqq;
1408 for (i = 0; i < 2; ++i)
1409 while ((cfqq = cfq_rb_first(&cfqd->service_trees[i])) != NULL)
1410 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
1412 while ((cfqq = cfq_rb_first(&cfqd->service_tree_idle)) != NULL)
1413 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
1415 cfq_slice_expired(cfqd, 0);
1417 BUG_ON(cfqd->busy_queues);
1419 cfq_log(cfqd, "forced_dispatch=%d", dispatched);
1423 static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1425 unsigned int max_dispatch;
1428 * Drain async requests before we start sync IO
1430 if (cfq_cfqq_idle_window(cfqq) && cfqd->rq_in_driver[BLK_RW_ASYNC])
1434 * If this is an async queue and we have sync IO in flight, let it wait
1436 if (cfqd->sync_flight && !cfq_cfqq_sync(cfqq))
1439 max_dispatch = cfqd->cfq_quantum;
1440 if (cfq_class_idle(cfqq))
1444 * Does this cfqq already have too much IO in flight?
1446 if (cfqq->dispatched >= max_dispatch) {
1448 * idle queue must always only have a single IO in flight
1450 if (cfq_class_idle(cfqq))
1454 * We have other queues, don't allow more IO from this one
1456 if (cfqd->busy_queues > 1)
1460 * Sole queue user, allow bigger slice
1466 * Async queues must wait a bit before being allowed dispatch.
1467 * We also ramp up the dispatch depth gradually for async IO,
1468 * based on the last sync IO we serviced
1470 if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
1471 unsigned long last_sync = jiffies - cfqd->last_end_sync_rq;
1474 depth = last_sync / cfqd->cfq_slice[1];
1475 if (!depth && !cfqq->dispatched)
1477 if (depth < max_dispatch)
1478 max_dispatch = depth;
1482 * If we're below the current max, allow a dispatch
1484 return cfqq->dispatched < max_dispatch;
1488 * Dispatch a request from cfqq, moving them to the request queue
1491 static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1495 BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
1497 if (!cfq_may_dispatch(cfqd, cfqq))
1501 * follow expired path, else get first next available
1503 rq = cfq_check_fifo(cfqq);
1508 * insert request into driver dispatch list
1510 cfq_dispatch_insert(cfqd->queue, rq);
1512 if (!cfqd->active_cic) {
1513 struct cfq_io_context *cic = RQ_CIC(rq);
1515 atomic_long_inc(&cic->ioc->refcount);
1516 cfqd->active_cic = cic;
1523 * Find the cfqq that we need to service and move a request from that to the
1526 static int cfq_dispatch_requests(struct request_queue *q, int force)
1528 struct cfq_data *cfqd = q->elevator->elevator_data;
1529 struct cfq_queue *cfqq;
1531 if (!cfqd->busy_queues)
1534 if (unlikely(force))
1535 return cfq_forced_dispatch(cfqd);
1537 cfqq = cfq_select_queue(cfqd);
1542 * Dispatch a request from this cfqq, if it is allowed
1544 if (!cfq_dispatch_request(cfqd, cfqq))
1547 cfqq->slice_dispatch++;
1548 cfq_clear_cfqq_must_dispatch(cfqq);
1551 * expire an async queue immediately if it has used up its slice. idle
1552 * queue always expire after 1 dispatch round.
1554 if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
1555 cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
1556 cfq_class_idle(cfqq))) {
1557 cfqq->slice_end = jiffies + 1;
1558 cfq_slice_expired(cfqd, 0);
1561 cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
1566 * task holds one reference to the queue, dropped when task exits. each rq
1567 * in-flight on this queue also holds a reference, dropped when rq is freed.
1569 * queue lock must be held here.
1571 static void cfq_put_queue(struct cfq_queue *cfqq)
1573 struct cfq_data *cfqd = cfqq->cfqd;
1575 BUG_ON(atomic_read(&cfqq->ref) <= 0);
1577 if (!atomic_dec_and_test(&cfqq->ref))
1580 cfq_log_cfqq(cfqd, cfqq, "put_queue");
1581 BUG_ON(rb_first(&cfqq->sort_list));
1582 BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
1583 BUG_ON(cfq_cfqq_on_rr(cfqq));
1585 if (unlikely(cfqd->active_queue == cfqq)) {
1586 __cfq_slice_expired(cfqd, cfqq, 0);
1587 cfq_schedule_dispatch(cfqd);
1590 kmem_cache_free(cfq_pool, cfqq);
1594 * Must always be called with the rcu_read_lock() held
1597 __call_for_each_cic(struct io_context *ioc,
1598 void (*func)(struct io_context *, struct cfq_io_context *))
1600 struct cfq_io_context *cic;
1601 struct hlist_node *n;
1603 hlist_for_each_entry_rcu(cic, n, &ioc->cic_list, cic_list)
1608 * Call func for each cic attached to this ioc.
1611 call_for_each_cic(struct io_context *ioc,
1612 void (*func)(struct io_context *, struct cfq_io_context *))
1615 __call_for_each_cic(ioc, func);
1619 static void cfq_cic_free_rcu(struct rcu_head *head)
1621 struct cfq_io_context *cic;
1623 cic = container_of(head, struct cfq_io_context, rcu_head);
1625 kmem_cache_free(cfq_ioc_pool, cic);
1626 elv_ioc_count_dec(cfq_ioc_count);
1630 * CFQ scheduler is exiting, grab exit lock and check
1631 * the pending io context count. If it hits zero,
1632 * complete ioc_gone and set it back to NULL
1634 spin_lock(&ioc_gone_lock);
1635 if (ioc_gone && !elv_ioc_count_read(cfq_ioc_count)) {
1639 spin_unlock(&ioc_gone_lock);
1643 static void cfq_cic_free(struct cfq_io_context *cic)
1645 call_rcu(&cic->rcu_head, cfq_cic_free_rcu);
1648 static void cic_free_func(struct io_context *ioc, struct cfq_io_context *cic)
1650 unsigned long flags;
1652 BUG_ON(!cic->dead_key);
1654 spin_lock_irqsave(&ioc->lock, flags);
1655 radix_tree_delete(&ioc->radix_root, cic->dead_key);
1656 hlist_del_rcu(&cic->cic_list);
1657 spin_unlock_irqrestore(&ioc->lock, flags);
1663 * Must be called with rcu_read_lock() held or preemption otherwise disabled.
1664 * Only two callers of this - ->dtor() which is called with the rcu_read_lock(),
1665 * and ->trim() which is called with the task lock held
1667 static void cfq_free_io_context(struct io_context *ioc)
1670 * ioc->refcount is zero here, or we are called from elv_unregister(),
1671 * so no more cic's are allowed to be linked into this ioc. So it
1672 * should be ok to iterate over the known list, we will see all cic's
1673 * since no new ones are added.
1675 __call_for_each_cic(ioc, cic_free_func);
1678 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1680 struct cfq_queue *__cfqq, *next;
1682 if (unlikely(cfqq == cfqd->active_queue)) {
1683 __cfq_slice_expired(cfqd, cfqq, 0);
1684 cfq_schedule_dispatch(cfqd);
1688 * If this queue was scheduled to merge with another queue, be
1689 * sure to drop the reference taken on that queue (and others in
1690 * the merge chain). See cfq_setup_merge and cfq_merge_cfqqs.
1692 __cfqq = cfqq->new_cfqq;
1694 if (__cfqq == cfqq) {
1695 WARN(1, "cfqq->new_cfqq loop detected\n");
1698 next = __cfqq->new_cfqq;
1699 cfq_put_queue(__cfqq);
1703 cfq_put_queue(cfqq);
1706 static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
1707 struct cfq_io_context *cic)
1709 struct io_context *ioc = cic->ioc;
1711 list_del_init(&cic->queue_list);
1714 * Make sure key == NULL is seen for dead queues
1717 cic->dead_key = (unsigned long) cic->key;
1720 if (ioc->ioc_data == cic)
1721 rcu_assign_pointer(ioc->ioc_data, NULL);
1723 if (cic->cfqq[BLK_RW_ASYNC]) {
1724 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]);
1725 cic->cfqq[BLK_RW_ASYNC] = NULL;
1728 if (cic->cfqq[BLK_RW_SYNC]) {
1729 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]);
1730 cic->cfqq[BLK_RW_SYNC] = NULL;
1734 static void cfq_exit_single_io_context(struct io_context *ioc,
1735 struct cfq_io_context *cic)
1737 struct cfq_data *cfqd = cic->key;
1740 struct request_queue *q = cfqd->queue;
1741 unsigned long flags;
1743 spin_lock_irqsave(q->queue_lock, flags);
1746 * Ensure we get a fresh copy of the ->key to prevent
1747 * race between exiting task and queue
1749 smp_read_barrier_depends();
1751 __cfq_exit_single_io_context(cfqd, cic);
1753 spin_unlock_irqrestore(q->queue_lock, flags);
1758 * The process that ioc belongs to has exited, we need to clean up
1759 * and put the internal structures we have that belongs to that process.
1761 static void cfq_exit_io_context(struct io_context *ioc)
1763 call_for_each_cic(ioc, cfq_exit_single_io_context);
1766 static struct cfq_io_context *
1767 cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
1769 struct cfq_io_context *cic;
1771 cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO,
1774 cic->last_end_request = jiffies;
1775 INIT_LIST_HEAD(&cic->queue_list);
1776 INIT_HLIST_NODE(&cic->cic_list);
1777 cic->dtor = cfq_free_io_context;
1778 cic->exit = cfq_exit_io_context;
1779 elv_ioc_count_inc(cfq_ioc_count);
1785 static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc)
1787 struct task_struct *tsk = current;
1790 if (!cfq_cfqq_prio_changed(cfqq))
1793 ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio);
1794 switch (ioprio_class) {
1796 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
1797 case IOPRIO_CLASS_NONE:
1799 * no prio set, inherit CPU scheduling settings
1801 cfqq->ioprio = task_nice_ioprio(tsk);
1802 cfqq->ioprio_class = task_nice_ioclass(tsk);
1804 case IOPRIO_CLASS_RT:
1805 cfqq->ioprio = task_ioprio(ioc);
1806 cfqq->ioprio_class = IOPRIO_CLASS_RT;
1808 case IOPRIO_CLASS_BE:
1809 cfqq->ioprio = task_ioprio(ioc);
1810 cfqq->ioprio_class = IOPRIO_CLASS_BE;
1812 case IOPRIO_CLASS_IDLE:
1813 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
1815 cfq_clear_cfqq_idle_window(cfqq);
1820 * keep track of original prio settings in case we have to temporarily
1821 * elevate the priority of this queue
1823 cfqq->org_ioprio = cfqq->ioprio;
1824 cfqq->org_ioprio_class = cfqq->ioprio_class;
1825 cfq_clear_cfqq_prio_changed(cfqq);
1828 static void changed_ioprio(struct io_context *ioc, struct cfq_io_context *cic)
1830 struct cfq_data *cfqd = cic->key;
1831 struct cfq_queue *cfqq;
1832 unsigned long flags;
1834 if (unlikely(!cfqd))
1837 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1839 cfqq = cic->cfqq[BLK_RW_ASYNC];
1841 struct cfq_queue *new_cfqq;
1842 new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic->ioc,
1845 cic->cfqq[BLK_RW_ASYNC] = new_cfqq;
1846 cfq_put_queue(cfqq);
1850 cfqq = cic->cfqq[BLK_RW_SYNC];
1852 cfq_mark_cfqq_prio_changed(cfqq);
1854 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1857 static void cfq_ioc_set_ioprio(struct io_context *ioc)
1859 call_for_each_cic(ioc, changed_ioprio);
1860 ioc->ioprio_changed = 0;
1863 static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1864 pid_t pid, bool is_sync)
1866 RB_CLEAR_NODE(&cfqq->rb_node);
1867 RB_CLEAR_NODE(&cfqq->p_node);
1868 INIT_LIST_HEAD(&cfqq->fifo);
1870 atomic_set(&cfqq->ref, 0);
1873 cfq_mark_cfqq_prio_changed(cfqq);
1876 if (!cfq_class_idle(cfqq))
1877 cfq_mark_cfqq_idle_window(cfqq);
1878 cfq_mark_cfqq_sync(cfqq);
1883 static struct cfq_queue *
1884 cfq_find_alloc_queue(struct cfq_data *cfqd, bool is_sync,
1885 struct io_context *ioc, gfp_t gfp_mask)
1887 struct cfq_queue *cfqq, *new_cfqq = NULL;
1888 struct cfq_io_context *cic;
1891 cic = cfq_cic_lookup(cfqd, ioc);
1892 /* cic always exists here */
1893 cfqq = cic_to_cfqq(cic, is_sync);
1896 * Always try a new alloc if we fell back to the OOM cfqq
1897 * originally, since it should just be a temporary situation.
1899 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
1904 } else if (gfp_mask & __GFP_WAIT) {
1905 spin_unlock_irq(cfqd->queue->queue_lock);
1906 new_cfqq = kmem_cache_alloc_node(cfq_pool,
1907 gfp_mask | __GFP_ZERO,
1909 spin_lock_irq(cfqd->queue->queue_lock);
1913 cfqq = kmem_cache_alloc_node(cfq_pool,
1914 gfp_mask | __GFP_ZERO,
1919 cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
1920 cfq_init_prio_data(cfqq, ioc);
1921 cfq_log_cfqq(cfqd, cfqq, "alloced");
1923 cfqq = &cfqd->oom_cfqq;
1927 kmem_cache_free(cfq_pool, new_cfqq);
1932 static struct cfq_queue **
1933 cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
1935 switch (ioprio_class) {
1936 case IOPRIO_CLASS_RT:
1937 return &cfqd->async_cfqq[0][ioprio];
1938 case IOPRIO_CLASS_BE:
1939 return &cfqd->async_cfqq[1][ioprio];
1940 case IOPRIO_CLASS_IDLE:
1941 return &cfqd->async_idle_cfqq;
1947 static struct cfq_queue *
1948 cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct io_context *ioc,
1951 const int ioprio = task_ioprio(ioc);
1952 const int ioprio_class = task_ioprio_class(ioc);
1953 struct cfq_queue **async_cfqq = NULL;
1954 struct cfq_queue *cfqq = NULL;
1957 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
1962 cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask);
1965 * pin the queue now that it's allocated, scheduler exit will prune it
1967 if (!is_sync && !(*async_cfqq)) {
1968 atomic_inc(&cfqq->ref);
1972 atomic_inc(&cfqq->ref);
1977 * We drop cfq io contexts lazily, so we may find a dead one.
1980 cfq_drop_dead_cic(struct cfq_data *cfqd, struct io_context *ioc,
1981 struct cfq_io_context *cic)
1983 unsigned long flags;
1985 WARN_ON(!list_empty(&cic->queue_list));
1987 spin_lock_irqsave(&ioc->lock, flags);
1989 BUG_ON(ioc->ioc_data == cic);
1991 radix_tree_delete(&ioc->radix_root, (unsigned long) cfqd);
1992 hlist_del_rcu(&cic->cic_list);
1993 spin_unlock_irqrestore(&ioc->lock, flags);
1998 static struct cfq_io_context *
1999 cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc)
2001 struct cfq_io_context *cic;
2002 unsigned long flags;
2011 * we maintain a last-hit cache, to avoid browsing over the tree
2013 cic = rcu_dereference(ioc->ioc_data);
2014 if (cic && cic->key == cfqd) {
2020 cic = radix_tree_lookup(&ioc->radix_root, (unsigned long) cfqd);
2024 /* ->key must be copied to avoid race with cfq_exit_queue() */
2027 cfq_drop_dead_cic(cfqd, ioc, cic);
2032 spin_lock_irqsave(&ioc->lock, flags);
2033 rcu_assign_pointer(ioc->ioc_data, cic);
2034 spin_unlock_irqrestore(&ioc->lock, flags);
2042 * Add cic into ioc, using cfqd as the search key. This enables us to lookup
2043 * the process specific cfq io context when entered from the block layer.
2044 * Also adds the cic to a per-cfqd list, used when this queue is removed.
2046 static int cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
2047 struct cfq_io_context *cic, gfp_t gfp_mask)
2049 unsigned long flags;
2052 ret = radix_tree_preload(gfp_mask);
2057 spin_lock_irqsave(&ioc->lock, flags);
2058 ret = radix_tree_insert(&ioc->radix_root,
2059 (unsigned long) cfqd, cic);
2061 hlist_add_head_rcu(&cic->cic_list, &ioc->cic_list);
2062 spin_unlock_irqrestore(&ioc->lock, flags);
2064 radix_tree_preload_end();
2067 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2068 list_add(&cic->queue_list, &cfqd->cic_list);
2069 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2074 printk(KERN_ERR "cfq: cic link failed!\n");
2080 * Setup general io context and cfq io context. There can be several cfq
2081 * io contexts per general io context, if this process is doing io to more
2082 * than one device managed by cfq.
2084 static struct cfq_io_context *
2085 cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
2087 struct io_context *ioc = NULL;
2088 struct cfq_io_context *cic;
2090 might_sleep_if(gfp_mask & __GFP_WAIT);
2092 ioc = get_io_context(gfp_mask, cfqd->queue->node);
2096 cic = cfq_cic_lookup(cfqd, ioc);
2100 cic = cfq_alloc_io_context(cfqd, gfp_mask);
2104 if (cfq_cic_link(cfqd, ioc, cic, gfp_mask))
2108 smp_read_barrier_depends();
2109 if (unlikely(ioc->ioprio_changed))
2110 cfq_ioc_set_ioprio(ioc);
2116 put_io_context(ioc);
2121 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
2123 unsigned long elapsed = jiffies - cic->last_end_request;
2124 unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
2126 cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
2127 cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
2128 cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
2132 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2138 if (!cfqq->last_request_pos)
2140 else if (cfqq->last_request_pos < blk_rq_pos(rq))
2141 sdist = blk_rq_pos(rq) - cfqq->last_request_pos;
2143 sdist = cfqq->last_request_pos - blk_rq_pos(rq);
2146 * Don't allow the seek distance to get too large from the
2147 * odd fragment, pagein, etc
2149 if (cfqq->seek_samples <= 60) /* second&third seek */
2150 sdist = min(sdist, (cfqq->seek_mean * 4) + 2*1024*1024);
2152 sdist = min(sdist, (cfqq->seek_mean * 4) + 2*1024*64);
2154 cfqq->seek_samples = (7*cfqq->seek_samples + 256) / 8;
2155 cfqq->seek_total = (7*cfqq->seek_total + (u64)256*sdist) / 8;
2156 total = cfqq->seek_total + (cfqq->seek_samples/2);
2157 do_div(total, cfqq->seek_samples);
2158 cfqq->seek_mean = (sector_t)total;
2161 * If this cfqq is shared between multiple processes, check to
2162 * make sure that those processes are still issuing I/Os within
2163 * the mean seek distance. If not, it may be time to break the
2164 * queues apart again.
2166 if (cfq_cfqq_coop(cfqq)) {
2167 if (CFQQ_SEEKY(cfqq) && !cfqq->seeky_start)
2168 cfqq->seeky_start = jiffies;
2169 else if (!CFQQ_SEEKY(cfqq))
2170 cfqq->seeky_start = 0;
2175 * Disable idle window if the process thinks too long or seeks so much that
2179 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2180 struct cfq_io_context *cic)
2182 int old_idle, enable_idle;
2185 * Don't idle for async or idle io prio class
2187 if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
2190 enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
2192 if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle ||
2193 (!cfqd->cfq_latency && cfqd->hw_tag && CFQQ_SEEKY(cfqq)))
2195 else if (sample_valid(cic->ttime_samples)) {
2196 unsigned int slice_idle = cfqd->cfq_slice_idle;
2197 if (sample_valid(cfqq->seek_samples) && CFQQ_SEEKY(cfqq))
2198 slice_idle = msecs_to_jiffies(CFQ_MIN_TT);
2199 if (cic->ttime_mean > slice_idle)
2205 if (old_idle != enable_idle) {
2206 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
2208 cfq_mark_cfqq_idle_window(cfqq);
2210 cfq_clear_cfqq_idle_window(cfqq);
2215 * Check if new_cfqq should preempt the currently active queue. Return 0 for
2216 * no or if we aren't sure, a 1 will cause a preempt.
2219 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
2222 struct cfq_queue *cfqq;
2224 cfqq = cfqd->active_queue;
2228 if (cfq_slice_used(cfqq))
2231 if (cfq_class_idle(new_cfqq))
2234 if (cfq_class_idle(cfqq))
2238 * if the new request is sync, but the currently running queue is
2239 * not, let the sync request have priority.
2241 if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
2245 * So both queues are sync. Let the new request get disk time if
2246 * it's a metadata request and the current queue is doing regular IO.
2248 if (rq_is_meta(rq) && !cfqq->meta_pending)
2252 * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
2254 if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
2257 if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
2261 * if this request is as-good as one we would expect from the
2262 * current cfqq, let it preempt
2264 if (cfq_rq_close(cfqd, cfqq, rq))
2271 * cfqq preempts the active queue. if we allowed preempt with no slice left,
2272 * let it have half of its nominal slice.
2274 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2276 cfq_log_cfqq(cfqd, cfqq, "preempt");
2277 cfq_slice_expired(cfqd, 1);
2280 * Put the new queue at the front of the of the current list,
2281 * so we know that it will be selected next.
2283 BUG_ON(!cfq_cfqq_on_rr(cfqq));
2285 cfq_service_tree_add(cfqd, cfqq, 1);
2287 cfqq->slice_end = 0;
2288 cfq_mark_cfqq_slice_new(cfqq);
2292 * Called when a new fs request (rq) is added (to cfqq). Check if there's
2293 * something we should do about it
2296 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2299 struct cfq_io_context *cic = RQ_CIC(rq);
2303 cfqq->meta_pending++;
2305 cfq_update_io_thinktime(cfqd, cic);
2306 cfq_update_io_seektime(cfqd, cfqq, rq);
2307 cfq_update_idle_window(cfqd, cfqq, cic);
2309 cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
2311 if (cfqq == cfqd->active_queue) {
2313 * Remember that we saw a request from this process, but
2314 * don't start queuing just yet. Otherwise we risk seeing lots
2315 * of tiny requests, because we disrupt the normal plugging
2316 * and merging. If the request is already larger than a single
2317 * page, let it rip immediately. For that case we assume that
2318 * merging is already done. Ditto for a busy system that
2319 * has other work pending, don't risk delaying until the
2320 * idle timer unplug to continue working.
2322 if (cfq_cfqq_wait_request(cfqq)) {
2323 if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
2324 cfqd->busy_queues > 1) {
2325 del_timer(&cfqd->idle_slice_timer);
2326 __blk_run_queue(cfqd->queue);
2328 cfq_mark_cfqq_must_dispatch(cfqq);
2330 } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
2332 * not the active queue - expire current slice if it is
2333 * idle and has expired it's mean thinktime or this new queue
2334 * has some old slice time left and is of higher priority or
2335 * this new queue is RT and the current one is BE
2337 cfq_preempt_queue(cfqd, cfqq);
2338 __blk_run_queue(cfqd->queue);
2342 static void cfq_insert_request(struct request_queue *q, struct request *rq)
2344 struct cfq_data *cfqd = q->elevator->elevator_data;
2345 struct cfq_queue *cfqq = RQ_CFQQ(rq);
2347 cfq_log_cfqq(cfqd, cfqq, "insert_request");
2348 cfq_init_prio_data(cfqq, RQ_CIC(rq)->ioc);
2350 rq_set_fifo_time(rq, jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)]);
2351 list_add_tail(&rq->queuelist, &cfqq->fifo);
2354 cfq_rq_enqueued(cfqd, cfqq, rq);
2358 * Update hw_tag based on peak queue depth over 50 samples under
2361 static void cfq_update_hw_tag(struct cfq_data *cfqd)
2363 struct cfq_queue *cfqq = cfqd->active_queue;
2365 if (rq_in_driver(cfqd) > cfqd->rq_in_driver_peak)
2366 cfqd->rq_in_driver_peak = rq_in_driver(cfqd);
2368 if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
2369 rq_in_driver(cfqd) <= CFQ_HW_QUEUE_MIN)
2373 * If active queue hasn't enough requests and can idle, cfq might not
2374 * dispatch sufficient requests to hardware. Don't zero hw_tag in this
2377 if (cfqq && cfq_cfqq_idle_window(cfqq) &&
2378 cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] <
2379 CFQ_HW_QUEUE_MIN && rq_in_driver(cfqd) < CFQ_HW_QUEUE_MIN)
2382 if (cfqd->hw_tag_samples++ < 50)
2385 if (cfqd->rq_in_driver_peak >= CFQ_HW_QUEUE_MIN)
2390 cfqd->hw_tag_samples = 0;
2391 cfqd->rq_in_driver_peak = 0;
2394 static void cfq_completed_request(struct request_queue *q, struct request *rq)
2396 struct cfq_queue *cfqq = RQ_CFQQ(rq);
2397 struct cfq_data *cfqd = cfqq->cfqd;
2398 const int sync = rq_is_sync(rq);
2402 cfq_log_cfqq(cfqd, cfqq, "complete");
2404 cfq_update_hw_tag(cfqd);
2406 WARN_ON(!cfqd->rq_in_driver[sync]);
2407 WARN_ON(!cfqq->dispatched);
2408 cfqd->rq_in_driver[sync]--;
2411 if (cfq_cfqq_sync(cfqq))
2412 cfqd->sync_flight--;
2415 RQ_CIC(rq)->last_end_request = now;
2416 cfqd->last_end_sync_rq = now;
2420 * If this is the active queue, check if it needs to be expired,
2421 * or if we want to idle in case it has no pending requests.
2423 if (cfqd->active_queue == cfqq) {
2424 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
2426 if (cfq_cfqq_slice_new(cfqq)) {
2427 cfq_set_prio_slice(cfqd, cfqq);
2428 cfq_clear_cfqq_slice_new(cfqq);
2431 * If there are no requests waiting in this queue, and
2432 * there are other queues ready to issue requests, AND
2433 * those other queues are issuing requests within our
2434 * mean seek distance, give them a chance to run instead
2437 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
2438 cfq_slice_expired(cfqd, 1);
2439 else if (cfqq_empty && !cfq_close_cooperator(cfqd, cfqq) &&
2440 sync && !rq_noidle(rq))
2441 cfq_arm_slice_timer(cfqd);
2444 if (!rq_in_driver(cfqd))
2445 cfq_schedule_dispatch(cfqd);
2449 * we temporarily boost lower priority queues if they are holding fs exclusive
2450 * resources. they are boosted to normal prio (CLASS_BE/4)
2452 static void cfq_prio_boost(struct cfq_queue *cfqq)
2454 if (has_fs_excl()) {
2456 * boost idle prio on transactions that would lock out other
2457 * users of the filesystem
2459 if (cfq_class_idle(cfqq))
2460 cfqq->ioprio_class = IOPRIO_CLASS_BE;
2461 if (cfqq->ioprio > IOPRIO_NORM)
2462 cfqq->ioprio = IOPRIO_NORM;
2465 * check if we need to unboost the queue
2467 if (cfqq->ioprio_class != cfqq->org_ioprio_class)
2468 cfqq->ioprio_class = cfqq->org_ioprio_class;
2469 if (cfqq->ioprio != cfqq->org_ioprio)
2470 cfqq->ioprio = cfqq->org_ioprio;
2474 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
2476 if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
2477 cfq_mark_cfqq_must_alloc_slice(cfqq);
2478 return ELV_MQUEUE_MUST;
2481 return ELV_MQUEUE_MAY;
2484 static int cfq_may_queue(struct request_queue *q, int rw)
2486 struct cfq_data *cfqd = q->elevator->elevator_data;
2487 struct task_struct *tsk = current;
2488 struct cfq_io_context *cic;
2489 struct cfq_queue *cfqq;
2492 * don't force setup of a queue from here, as a call to may_queue
2493 * does not necessarily imply that a request actually will be queued.
2494 * so just lookup a possibly existing queue, or return 'may queue'
2497 cic = cfq_cic_lookup(cfqd, tsk->io_context);
2499 return ELV_MQUEUE_MAY;
2501 cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
2503 cfq_init_prio_data(cfqq, cic->ioc);
2504 cfq_prio_boost(cfqq);
2506 return __cfq_may_queue(cfqq);
2509 return ELV_MQUEUE_MAY;
2513 * queue lock held here
2515 static void cfq_put_request(struct request *rq)
2517 struct cfq_queue *cfqq = RQ_CFQQ(rq);
2520 const int rw = rq_data_dir(rq);
2522 BUG_ON(!cfqq->allocated[rw]);
2523 cfqq->allocated[rw]--;
2525 put_io_context(RQ_CIC(rq)->ioc);
2527 rq->elevator_private = NULL;
2528 rq->elevator_private2 = NULL;
2530 cfq_put_queue(cfqq);
2534 static struct cfq_queue *
2535 cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_context *cic,
2536 struct cfq_queue *cfqq)
2538 cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq);
2539 cic_set_cfqq(cic, cfqq->new_cfqq, 1);
2540 cfq_mark_cfqq_coop(cfqq->new_cfqq);
2541 cfq_put_queue(cfqq);
2542 return cic_to_cfqq(cic, 1);
2545 static int should_split_cfqq(struct cfq_queue *cfqq)
2547 if (cfqq->seeky_start &&
2548 time_after(jiffies, cfqq->seeky_start + CFQQ_COOP_TOUT))
2554 * Returns NULL if a new cfqq should be allocated, or the old cfqq if this
2555 * was the last process referring to said cfqq.
2557 static struct cfq_queue *
2558 split_cfqq(struct cfq_io_context *cic, struct cfq_queue *cfqq)
2560 if (cfqq_process_refs(cfqq) == 1) {
2561 cfqq->seeky_start = 0;
2562 cfqq->pid = current->pid;
2563 cfq_clear_cfqq_coop(cfqq);
2567 cic_set_cfqq(cic, NULL, 1);
2568 cfq_put_queue(cfqq);
2572 * Allocate cfq data structures associated with this request.
2575 cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
2577 struct cfq_data *cfqd = q->elevator->elevator_data;
2578 struct cfq_io_context *cic;
2579 const int rw = rq_data_dir(rq);
2580 const bool is_sync = rq_is_sync(rq);
2581 struct cfq_queue *cfqq;
2582 unsigned long flags;
2584 might_sleep_if(gfp_mask & __GFP_WAIT);
2586 cic = cfq_get_io_context(cfqd, gfp_mask);
2588 spin_lock_irqsave(q->queue_lock, flags);
2594 cfqq = cic_to_cfqq(cic, is_sync);
2595 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
2596 cfqq = cfq_get_queue(cfqd, is_sync, cic->ioc, gfp_mask);
2597 cic_set_cfqq(cic, cfqq, is_sync);
2600 * If the queue was seeky for too long, break it apart.
2602 if (cfq_cfqq_coop(cfqq) && should_split_cfqq(cfqq)) {
2603 cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq");
2604 cfqq = split_cfqq(cic, cfqq);
2610 * Check to see if this queue is scheduled to merge with
2611 * another, closely cooperating queue. The merging of
2612 * queues happens here as it must be done in process context.
2613 * The reference on new_cfqq was taken in merge_cfqqs.
2616 cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq);
2619 cfqq->allocated[rw]++;
2620 atomic_inc(&cfqq->ref);
2622 spin_unlock_irqrestore(q->queue_lock, flags);
2624 rq->elevator_private = cic;
2625 rq->elevator_private2 = cfqq;
2630 put_io_context(cic->ioc);
2632 cfq_schedule_dispatch(cfqd);
2633 spin_unlock_irqrestore(q->queue_lock, flags);
2634 cfq_log(cfqd, "set_request fail");
2638 static void cfq_kick_queue(struct work_struct *work)
2640 struct cfq_data *cfqd =
2641 container_of(work, struct cfq_data, unplug_work);
2642 struct request_queue *q = cfqd->queue;
2644 spin_lock_irq(q->queue_lock);
2645 __blk_run_queue(cfqd->queue);
2646 spin_unlock_irq(q->queue_lock);
2650 * Timer running if the active_queue is currently idling inside its time slice
2652 static void cfq_idle_slice_timer(unsigned long data)
2654 struct cfq_data *cfqd = (struct cfq_data *) data;
2655 struct cfq_queue *cfqq;
2656 unsigned long flags;
2659 cfq_log(cfqd, "idle timer fired");
2661 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2663 cfqq = cfqd->active_queue;
2668 * We saw a request before the queue expired, let it through
2670 if (cfq_cfqq_must_dispatch(cfqq))
2676 if (cfq_slice_used(cfqq))
2680 * only expire and reinvoke request handler, if there are
2681 * other queues with pending requests
2683 if (!cfqd->busy_queues)
2687 * not expired and it has a request pending, let it dispatch
2689 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
2693 cfq_slice_expired(cfqd, timed_out);
2695 cfq_schedule_dispatch(cfqd);
2697 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2700 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
2702 del_timer_sync(&cfqd->idle_slice_timer);
2703 cancel_work_sync(&cfqd->unplug_work);
2706 static void cfq_put_async_queues(struct cfq_data *cfqd)
2710 for (i = 0; i < IOPRIO_BE_NR; i++) {
2711 if (cfqd->async_cfqq[0][i])
2712 cfq_put_queue(cfqd->async_cfqq[0][i]);
2713 if (cfqd->async_cfqq[1][i])
2714 cfq_put_queue(cfqd->async_cfqq[1][i]);
2717 if (cfqd->async_idle_cfqq)
2718 cfq_put_queue(cfqd->async_idle_cfqq);
2721 static void cfq_exit_queue(struct elevator_queue *e)
2723 struct cfq_data *cfqd = e->elevator_data;
2724 struct request_queue *q = cfqd->queue;
2726 cfq_shutdown_timer_wq(cfqd);
2728 spin_lock_irq(q->queue_lock);
2730 if (cfqd->active_queue)
2731 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
2733 while (!list_empty(&cfqd->cic_list)) {
2734 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
2735 struct cfq_io_context,
2738 __cfq_exit_single_io_context(cfqd, cic);
2741 cfq_put_async_queues(cfqd);
2743 spin_unlock_irq(q->queue_lock);
2745 cfq_shutdown_timer_wq(cfqd);
2750 static void *cfq_init_queue(struct request_queue *q)
2752 struct cfq_data *cfqd;
2755 cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
2759 for (i = 0; i < 2; ++i)
2760 cfqd->service_trees[i] = CFQ_RB_ROOT;
2761 cfqd->service_tree_idle = CFQ_RB_ROOT;
2764 * Not strictly needed (since RB_ROOT just clears the node and we
2765 * zeroed cfqd on alloc), but better be safe in case someone decides
2766 * to add magic to the rb code
2768 for (i = 0; i < CFQ_PRIO_LISTS; i++)
2769 cfqd->prio_trees[i] = RB_ROOT;
2772 * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
2773 * Grab a permanent reference to it, so that the normal code flow
2774 * will not attempt to free it.
2776 cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
2777 atomic_inc(&cfqd->oom_cfqq.ref);
2779 INIT_LIST_HEAD(&cfqd->cic_list);
2783 init_timer(&cfqd->idle_slice_timer);
2784 cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
2785 cfqd->idle_slice_timer.data = (unsigned long) cfqd;
2787 INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
2789 cfqd->cfq_quantum = cfq_quantum;
2790 cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
2791 cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
2792 cfqd->cfq_back_max = cfq_back_max;
2793 cfqd->cfq_back_penalty = cfq_back_penalty;
2794 cfqd->cfq_slice[0] = cfq_slice_async;
2795 cfqd->cfq_slice[1] = cfq_slice_sync;
2796 cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
2797 cfqd->cfq_slice_idle = cfq_slice_idle;
2798 cfqd->cfq_latency = 1;
2800 cfqd->last_end_sync_rq = jiffies;
2804 static void cfq_slab_kill(void)
2807 * Caller already ensured that pending RCU callbacks are completed,
2808 * so we should have no busy allocations at this point.
2811 kmem_cache_destroy(cfq_pool);
2813 kmem_cache_destroy(cfq_ioc_pool);
2816 static int __init cfq_slab_setup(void)
2818 cfq_pool = KMEM_CACHE(cfq_queue, 0);
2822 cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0);
2833 * sysfs parts below -->
2836 cfq_var_show(unsigned int var, char *page)
2838 return sprintf(page, "%d\n", var);
2842 cfq_var_store(unsigned int *var, const char *page, size_t count)
2844 char *p = (char *) page;
2846 *var = simple_strtoul(p, &p, 10);
2850 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
2851 static ssize_t __FUNC(struct elevator_queue *e, char *page) \
2853 struct cfq_data *cfqd = e->elevator_data; \
2854 unsigned int __data = __VAR; \
2856 __data = jiffies_to_msecs(__data); \
2857 return cfq_var_show(__data, (page)); \
2859 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
2860 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
2861 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
2862 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
2863 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
2864 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
2865 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
2866 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
2867 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
2868 SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
2869 #undef SHOW_FUNCTION
2871 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
2872 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
2874 struct cfq_data *cfqd = e->elevator_data; \
2875 unsigned int __data; \
2876 int ret = cfq_var_store(&__data, (page), count); \
2877 if (__data < (MIN)) \
2879 else if (__data > (MAX)) \
2882 *(__PTR) = msecs_to_jiffies(__data); \
2884 *(__PTR) = __data; \
2887 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
2888 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
2890 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
2892 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
2893 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
2895 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
2896 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
2897 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
2898 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
2900 STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
2901 #undef STORE_FUNCTION
2903 #define CFQ_ATTR(name) \
2904 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
2906 static struct elv_fs_entry cfq_attrs[] = {
2908 CFQ_ATTR(fifo_expire_sync),
2909 CFQ_ATTR(fifo_expire_async),
2910 CFQ_ATTR(back_seek_max),
2911 CFQ_ATTR(back_seek_penalty),
2912 CFQ_ATTR(slice_sync),
2913 CFQ_ATTR(slice_async),
2914 CFQ_ATTR(slice_async_rq),
2915 CFQ_ATTR(slice_idle),
2916 CFQ_ATTR(low_latency),
2920 static struct elevator_type iosched_cfq = {
2922 .elevator_merge_fn = cfq_merge,
2923 .elevator_merged_fn = cfq_merged_request,
2924 .elevator_merge_req_fn = cfq_merged_requests,
2925 .elevator_allow_merge_fn = cfq_allow_merge,
2926 .elevator_dispatch_fn = cfq_dispatch_requests,
2927 .elevator_add_req_fn = cfq_insert_request,
2928 .elevator_activate_req_fn = cfq_activate_request,
2929 .elevator_deactivate_req_fn = cfq_deactivate_request,
2930 .elevator_queue_empty_fn = cfq_queue_empty,
2931 .elevator_completed_req_fn = cfq_completed_request,
2932 .elevator_former_req_fn = elv_rb_former_request,
2933 .elevator_latter_req_fn = elv_rb_latter_request,
2934 .elevator_set_req_fn = cfq_set_request,
2935 .elevator_put_req_fn = cfq_put_request,
2936 .elevator_may_queue_fn = cfq_may_queue,
2937 .elevator_init_fn = cfq_init_queue,
2938 .elevator_exit_fn = cfq_exit_queue,
2939 .trim = cfq_free_io_context,
2941 .elevator_attrs = cfq_attrs,
2942 .elevator_name = "cfq",
2943 .elevator_owner = THIS_MODULE,
2946 static int __init cfq_init(void)
2949 * could be 0 on HZ < 1000 setups
2951 if (!cfq_slice_async)
2952 cfq_slice_async = 1;
2953 if (!cfq_slice_idle)
2956 if (cfq_slab_setup())
2959 elv_register(&iosched_cfq);
2964 static void __exit cfq_exit(void)
2966 DECLARE_COMPLETION_ONSTACK(all_gone);
2967 elv_unregister(&iosched_cfq);
2968 ioc_gone = &all_gone;
2969 /* ioc_gone's update must be visible before reading ioc_count */
2973 * this also protects us from entering cfq_slab_kill() with
2974 * pending RCU callbacks
2976 if (elv_ioc_count_read(cfq_ioc_count))
2977 wait_for_completion(&all_gone);
2981 module_init(cfq_init);
2982 module_exit(cfq_exit);
2984 MODULE_AUTHOR("Jens Axboe");
2985 MODULE_LICENSE("GPL");
2986 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");