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))
1164 * Determine whether we should enforce idle window for this queue.
1167 static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1169 enum wl_prio_t prio = cfqq_prio(cfqq);
1170 struct cfq_rb_root *service_tree;
1172 /* We never do for idle class queues. */
1173 if (prio == IDLE_WORKLOAD)
1176 /* We do for queues that were marked with idle window flag. */
1177 if (cfq_cfqq_idle_window(cfqq))
1181 * Otherwise, we do only if they are the last ones
1182 * in their service tree.
1184 service_tree = service_tree_for(prio, cfqd);
1185 if (service_tree->count == 0)
1188 return (service_tree->count == 1 && cfq_rb_first(service_tree) == cfqq);
1191 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
1193 struct cfq_queue *cfqq = cfqd->active_queue;
1194 struct cfq_io_context *cic;
1198 * SSD device without seek penalty, disable idling. But only do so
1199 * for devices that support queuing, otherwise we still have a problem
1200 * with sync vs async workloads.
1202 if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
1205 WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
1206 WARN_ON(cfq_cfqq_slice_new(cfqq));
1209 * idle is disabled, either manually or by past process history
1211 if (!cfqd->cfq_slice_idle || !cfq_should_idle(cfqd, cfqq))
1215 * still requests with the driver, don't idle
1217 if (rq_in_driver(cfqd))
1221 * task has exited, don't wait
1223 cic = cfqd->active_cic;
1224 if (!cic || !atomic_read(&cic->ioc->nr_tasks))
1228 * If our average think time is larger than the remaining time
1229 * slice, then don't idle. This avoids overrunning the allotted
1232 if (sample_valid(cic->ttime_samples) &&
1233 (cfqq->slice_end - jiffies < cic->ttime_mean))
1236 cfq_mark_cfqq_wait_request(cfqq);
1239 * we don't want to idle for seeks, but we do want to allow
1240 * fair distribution of slice time for a process doing back-to-back
1241 * seeks. so allow a little bit of time for him to submit a new rq
1243 sl = cfqd->cfq_slice_idle;
1244 if (sample_valid(cfqq->seek_samples) && CFQQ_SEEKY(cfqq))
1245 sl = min(sl, msecs_to_jiffies(CFQ_MIN_TT));
1247 mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
1248 cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu", sl);
1252 * Move request from internal lists to the request queue dispatch list.
1254 static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
1256 struct cfq_data *cfqd = q->elevator->elevator_data;
1257 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1259 cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
1261 cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
1262 cfq_remove_request(rq);
1264 elv_dispatch_sort(q, rq);
1266 if (cfq_cfqq_sync(cfqq))
1267 cfqd->sync_flight++;
1271 * return expired entry, or NULL to just start from scratch in rbtree
1273 static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
1275 struct request *rq = NULL;
1277 if (cfq_cfqq_fifo_expire(cfqq))
1280 cfq_mark_cfqq_fifo_expire(cfqq);
1282 if (list_empty(&cfqq->fifo))
1285 rq = rq_entry_fifo(cfqq->fifo.next);
1286 if (time_before(jiffies, rq_fifo_time(rq)))
1289 cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq);
1294 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1296 const int base_rq = cfqd->cfq_slice_async_rq;
1298 WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
1300 return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
1304 * Must be called with the queue_lock held.
1306 static int cfqq_process_refs(struct cfq_queue *cfqq)
1308 int process_refs, io_refs;
1310 io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE];
1311 process_refs = atomic_read(&cfqq->ref) - io_refs;
1312 BUG_ON(process_refs < 0);
1313 return process_refs;
1316 static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq)
1318 int process_refs, new_process_refs;
1319 struct cfq_queue *__cfqq;
1321 /* Avoid a circular list and skip interim queue merges */
1322 while ((__cfqq = new_cfqq->new_cfqq)) {
1328 process_refs = cfqq_process_refs(cfqq);
1330 * If the process for the cfqq has gone away, there is no
1331 * sense in merging the queues.
1333 if (process_refs == 0)
1337 * Merge in the direction of the lesser amount of work.
1339 new_process_refs = cfqq_process_refs(new_cfqq);
1340 if (new_process_refs >= process_refs) {
1341 cfqq->new_cfqq = new_cfqq;
1342 atomic_add(process_refs, &new_cfqq->ref);
1344 new_cfqq->new_cfqq = cfqq;
1345 atomic_add(new_process_refs, &cfqq->ref);
1350 * Select a queue for service. If we have a current active queue,
1351 * check whether to continue servicing it, or retrieve and set a new one.
1353 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
1355 struct cfq_queue *cfqq, *new_cfqq = NULL;
1357 cfqq = cfqd->active_queue;
1362 * The active queue has run out of time, expire it and select new.
1364 if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq))
1368 * The active queue has requests and isn't expired, allow it to
1371 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
1375 * If another queue has a request waiting within our mean seek
1376 * distance, let it run. The expire code will check for close
1377 * cooperators and put the close queue at the front of the service
1378 * tree. If possible, merge the expiring queue with the new cfqq.
1380 new_cfqq = cfq_close_cooperator(cfqd, cfqq);
1382 if (!cfqq->new_cfqq)
1383 cfq_setup_merge(cfqq, new_cfqq);
1388 * No requests pending. If the active queue still has requests in
1389 * flight or is idling for a new request, allow either of these
1390 * conditions to happen (or time out) before selecting a new queue.
1392 if (timer_pending(&cfqd->idle_slice_timer) ||
1393 (cfqq->dispatched && cfq_should_idle(cfqd, cfqq))) {
1399 cfq_slice_expired(cfqd, 0);
1402 if (cfq_busy_queues_wl(RT_WORKLOAD, cfqd))
1403 cfqd->serving_prio = RT_WORKLOAD;
1404 else if (cfq_busy_queues_wl(BE_WORKLOAD, cfqd))
1405 cfqd->serving_prio = BE_WORKLOAD;
1407 cfqd->serving_prio = IDLE_WORKLOAD;
1409 cfqq = cfq_set_active_queue(cfqd, new_cfqq);
1414 static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
1418 while (cfqq->next_rq) {
1419 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
1423 BUG_ON(!list_empty(&cfqq->fifo));
1428 * Drain our current requests. Used for barriers and when switching
1429 * io schedulers on-the-fly.
1431 static int cfq_forced_dispatch(struct cfq_data *cfqd)
1433 struct cfq_queue *cfqq;
1436 for (i = 0; i < 2; ++i)
1437 while ((cfqq = cfq_rb_first(&cfqd->service_trees[i])) != NULL)
1438 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
1440 while ((cfqq = cfq_rb_first(&cfqd->service_tree_idle)) != NULL)
1441 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
1443 cfq_slice_expired(cfqd, 0);
1445 BUG_ON(cfqd->busy_queues);
1447 cfq_log(cfqd, "forced_dispatch=%d", dispatched);
1451 static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1453 unsigned int max_dispatch;
1456 * Drain async requests before we start sync IO
1458 if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_driver[BLK_RW_ASYNC])
1462 * If this is an async queue and we have sync IO in flight, let it wait
1464 if (cfqd->sync_flight && !cfq_cfqq_sync(cfqq))
1467 max_dispatch = cfqd->cfq_quantum;
1468 if (cfq_class_idle(cfqq))
1472 * Does this cfqq already have too much IO in flight?
1474 if (cfqq->dispatched >= max_dispatch) {
1476 * idle queue must always only have a single IO in flight
1478 if (cfq_class_idle(cfqq))
1482 * We have other queues, don't allow more IO from this one
1484 if (cfqd->busy_queues > 1)
1488 * Sole queue user, allow bigger slice
1494 * Async queues must wait a bit before being allowed dispatch.
1495 * We also ramp up the dispatch depth gradually for async IO,
1496 * based on the last sync IO we serviced
1498 if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
1499 unsigned long last_sync = jiffies - cfqd->last_end_sync_rq;
1502 depth = last_sync / cfqd->cfq_slice[1];
1503 if (!depth && !cfqq->dispatched)
1505 if (depth < max_dispatch)
1506 max_dispatch = depth;
1510 * If we're below the current max, allow a dispatch
1512 return cfqq->dispatched < max_dispatch;
1516 * Dispatch a request from cfqq, moving them to the request queue
1519 static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1523 BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
1525 if (!cfq_may_dispatch(cfqd, cfqq))
1529 * follow expired path, else get first next available
1531 rq = cfq_check_fifo(cfqq);
1536 * insert request into driver dispatch list
1538 cfq_dispatch_insert(cfqd->queue, rq);
1540 if (!cfqd->active_cic) {
1541 struct cfq_io_context *cic = RQ_CIC(rq);
1543 atomic_long_inc(&cic->ioc->refcount);
1544 cfqd->active_cic = cic;
1551 * Find the cfqq that we need to service and move a request from that to the
1554 static int cfq_dispatch_requests(struct request_queue *q, int force)
1556 struct cfq_data *cfqd = q->elevator->elevator_data;
1557 struct cfq_queue *cfqq;
1559 if (!cfqd->busy_queues)
1562 if (unlikely(force))
1563 return cfq_forced_dispatch(cfqd);
1565 cfqq = cfq_select_queue(cfqd);
1570 * Dispatch a request from this cfqq, if it is allowed
1572 if (!cfq_dispatch_request(cfqd, cfqq))
1575 cfqq->slice_dispatch++;
1576 cfq_clear_cfqq_must_dispatch(cfqq);
1579 * expire an async queue immediately if it has used up its slice. idle
1580 * queue always expire after 1 dispatch round.
1582 if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
1583 cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
1584 cfq_class_idle(cfqq))) {
1585 cfqq->slice_end = jiffies + 1;
1586 cfq_slice_expired(cfqd, 0);
1589 cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
1594 * task holds one reference to the queue, dropped when task exits. each rq
1595 * in-flight on this queue also holds a reference, dropped when rq is freed.
1597 * queue lock must be held here.
1599 static void cfq_put_queue(struct cfq_queue *cfqq)
1601 struct cfq_data *cfqd = cfqq->cfqd;
1603 BUG_ON(atomic_read(&cfqq->ref) <= 0);
1605 if (!atomic_dec_and_test(&cfqq->ref))
1608 cfq_log_cfqq(cfqd, cfqq, "put_queue");
1609 BUG_ON(rb_first(&cfqq->sort_list));
1610 BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
1611 BUG_ON(cfq_cfqq_on_rr(cfqq));
1613 if (unlikely(cfqd->active_queue == cfqq)) {
1614 __cfq_slice_expired(cfqd, cfqq, 0);
1615 cfq_schedule_dispatch(cfqd);
1618 kmem_cache_free(cfq_pool, cfqq);
1622 * Must always be called with the rcu_read_lock() held
1625 __call_for_each_cic(struct io_context *ioc,
1626 void (*func)(struct io_context *, struct cfq_io_context *))
1628 struct cfq_io_context *cic;
1629 struct hlist_node *n;
1631 hlist_for_each_entry_rcu(cic, n, &ioc->cic_list, cic_list)
1636 * Call func for each cic attached to this ioc.
1639 call_for_each_cic(struct io_context *ioc,
1640 void (*func)(struct io_context *, struct cfq_io_context *))
1643 __call_for_each_cic(ioc, func);
1647 static void cfq_cic_free_rcu(struct rcu_head *head)
1649 struct cfq_io_context *cic;
1651 cic = container_of(head, struct cfq_io_context, rcu_head);
1653 kmem_cache_free(cfq_ioc_pool, cic);
1654 elv_ioc_count_dec(cfq_ioc_count);
1658 * CFQ scheduler is exiting, grab exit lock and check
1659 * the pending io context count. If it hits zero,
1660 * complete ioc_gone and set it back to NULL
1662 spin_lock(&ioc_gone_lock);
1663 if (ioc_gone && !elv_ioc_count_read(cfq_ioc_count)) {
1667 spin_unlock(&ioc_gone_lock);
1671 static void cfq_cic_free(struct cfq_io_context *cic)
1673 call_rcu(&cic->rcu_head, cfq_cic_free_rcu);
1676 static void cic_free_func(struct io_context *ioc, struct cfq_io_context *cic)
1678 unsigned long flags;
1680 BUG_ON(!cic->dead_key);
1682 spin_lock_irqsave(&ioc->lock, flags);
1683 radix_tree_delete(&ioc->radix_root, cic->dead_key);
1684 hlist_del_rcu(&cic->cic_list);
1685 spin_unlock_irqrestore(&ioc->lock, flags);
1691 * Must be called with rcu_read_lock() held or preemption otherwise disabled.
1692 * Only two callers of this - ->dtor() which is called with the rcu_read_lock(),
1693 * and ->trim() which is called with the task lock held
1695 static void cfq_free_io_context(struct io_context *ioc)
1698 * ioc->refcount is zero here, or we are called from elv_unregister(),
1699 * so no more cic's are allowed to be linked into this ioc. So it
1700 * should be ok to iterate over the known list, we will see all cic's
1701 * since no new ones are added.
1703 __call_for_each_cic(ioc, cic_free_func);
1706 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1708 struct cfq_queue *__cfqq, *next;
1710 if (unlikely(cfqq == cfqd->active_queue)) {
1711 __cfq_slice_expired(cfqd, cfqq, 0);
1712 cfq_schedule_dispatch(cfqd);
1716 * If this queue was scheduled to merge with another queue, be
1717 * sure to drop the reference taken on that queue (and others in
1718 * the merge chain). See cfq_setup_merge and cfq_merge_cfqqs.
1720 __cfqq = cfqq->new_cfqq;
1722 if (__cfqq == cfqq) {
1723 WARN(1, "cfqq->new_cfqq loop detected\n");
1726 next = __cfqq->new_cfqq;
1727 cfq_put_queue(__cfqq);
1731 cfq_put_queue(cfqq);
1734 static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
1735 struct cfq_io_context *cic)
1737 struct io_context *ioc = cic->ioc;
1739 list_del_init(&cic->queue_list);
1742 * Make sure key == NULL is seen for dead queues
1745 cic->dead_key = (unsigned long) cic->key;
1748 if (ioc->ioc_data == cic)
1749 rcu_assign_pointer(ioc->ioc_data, NULL);
1751 if (cic->cfqq[BLK_RW_ASYNC]) {
1752 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]);
1753 cic->cfqq[BLK_RW_ASYNC] = NULL;
1756 if (cic->cfqq[BLK_RW_SYNC]) {
1757 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]);
1758 cic->cfqq[BLK_RW_SYNC] = NULL;
1762 static void cfq_exit_single_io_context(struct io_context *ioc,
1763 struct cfq_io_context *cic)
1765 struct cfq_data *cfqd = cic->key;
1768 struct request_queue *q = cfqd->queue;
1769 unsigned long flags;
1771 spin_lock_irqsave(q->queue_lock, flags);
1774 * Ensure we get a fresh copy of the ->key to prevent
1775 * race between exiting task and queue
1777 smp_read_barrier_depends();
1779 __cfq_exit_single_io_context(cfqd, cic);
1781 spin_unlock_irqrestore(q->queue_lock, flags);
1786 * The process that ioc belongs to has exited, we need to clean up
1787 * and put the internal structures we have that belongs to that process.
1789 static void cfq_exit_io_context(struct io_context *ioc)
1791 call_for_each_cic(ioc, cfq_exit_single_io_context);
1794 static struct cfq_io_context *
1795 cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
1797 struct cfq_io_context *cic;
1799 cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO,
1802 cic->last_end_request = jiffies;
1803 INIT_LIST_HEAD(&cic->queue_list);
1804 INIT_HLIST_NODE(&cic->cic_list);
1805 cic->dtor = cfq_free_io_context;
1806 cic->exit = cfq_exit_io_context;
1807 elv_ioc_count_inc(cfq_ioc_count);
1813 static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc)
1815 struct task_struct *tsk = current;
1818 if (!cfq_cfqq_prio_changed(cfqq))
1821 ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio);
1822 switch (ioprio_class) {
1824 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
1825 case IOPRIO_CLASS_NONE:
1827 * no prio set, inherit CPU scheduling settings
1829 cfqq->ioprio = task_nice_ioprio(tsk);
1830 cfqq->ioprio_class = task_nice_ioclass(tsk);
1832 case IOPRIO_CLASS_RT:
1833 cfqq->ioprio = task_ioprio(ioc);
1834 cfqq->ioprio_class = IOPRIO_CLASS_RT;
1836 case IOPRIO_CLASS_BE:
1837 cfqq->ioprio = task_ioprio(ioc);
1838 cfqq->ioprio_class = IOPRIO_CLASS_BE;
1840 case IOPRIO_CLASS_IDLE:
1841 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
1843 cfq_clear_cfqq_idle_window(cfqq);
1848 * keep track of original prio settings in case we have to temporarily
1849 * elevate the priority of this queue
1851 cfqq->org_ioprio = cfqq->ioprio;
1852 cfqq->org_ioprio_class = cfqq->ioprio_class;
1853 cfq_clear_cfqq_prio_changed(cfqq);
1856 static void changed_ioprio(struct io_context *ioc, struct cfq_io_context *cic)
1858 struct cfq_data *cfqd = cic->key;
1859 struct cfq_queue *cfqq;
1860 unsigned long flags;
1862 if (unlikely(!cfqd))
1865 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1867 cfqq = cic->cfqq[BLK_RW_ASYNC];
1869 struct cfq_queue *new_cfqq;
1870 new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic->ioc,
1873 cic->cfqq[BLK_RW_ASYNC] = new_cfqq;
1874 cfq_put_queue(cfqq);
1878 cfqq = cic->cfqq[BLK_RW_SYNC];
1880 cfq_mark_cfqq_prio_changed(cfqq);
1882 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1885 static void cfq_ioc_set_ioprio(struct io_context *ioc)
1887 call_for_each_cic(ioc, changed_ioprio);
1888 ioc->ioprio_changed = 0;
1891 static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1892 pid_t pid, bool is_sync)
1894 RB_CLEAR_NODE(&cfqq->rb_node);
1895 RB_CLEAR_NODE(&cfqq->p_node);
1896 INIT_LIST_HEAD(&cfqq->fifo);
1898 atomic_set(&cfqq->ref, 0);
1901 cfq_mark_cfqq_prio_changed(cfqq);
1904 if (!cfq_class_idle(cfqq))
1905 cfq_mark_cfqq_idle_window(cfqq);
1906 cfq_mark_cfqq_sync(cfqq);
1911 static struct cfq_queue *
1912 cfq_find_alloc_queue(struct cfq_data *cfqd, bool is_sync,
1913 struct io_context *ioc, gfp_t gfp_mask)
1915 struct cfq_queue *cfqq, *new_cfqq = NULL;
1916 struct cfq_io_context *cic;
1919 cic = cfq_cic_lookup(cfqd, ioc);
1920 /* cic always exists here */
1921 cfqq = cic_to_cfqq(cic, is_sync);
1924 * Always try a new alloc if we fell back to the OOM cfqq
1925 * originally, since it should just be a temporary situation.
1927 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
1932 } else if (gfp_mask & __GFP_WAIT) {
1933 spin_unlock_irq(cfqd->queue->queue_lock);
1934 new_cfqq = kmem_cache_alloc_node(cfq_pool,
1935 gfp_mask | __GFP_ZERO,
1937 spin_lock_irq(cfqd->queue->queue_lock);
1941 cfqq = kmem_cache_alloc_node(cfq_pool,
1942 gfp_mask | __GFP_ZERO,
1947 cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
1948 cfq_init_prio_data(cfqq, ioc);
1949 cfq_log_cfqq(cfqd, cfqq, "alloced");
1951 cfqq = &cfqd->oom_cfqq;
1955 kmem_cache_free(cfq_pool, new_cfqq);
1960 static struct cfq_queue **
1961 cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
1963 switch (ioprio_class) {
1964 case IOPRIO_CLASS_RT:
1965 return &cfqd->async_cfqq[0][ioprio];
1966 case IOPRIO_CLASS_BE:
1967 return &cfqd->async_cfqq[1][ioprio];
1968 case IOPRIO_CLASS_IDLE:
1969 return &cfqd->async_idle_cfqq;
1975 static struct cfq_queue *
1976 cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct io_context *ioc,
1979 const int ioprio = task_ioprio(ioc);
1980 const int ioprio_class = task_ioprio_class(ioc);
1981 struct cfq_queue **async_cfqq = NULL;
1982 struct cfq_queue *cfqq = NULL;
1985 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
1990 cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask);
1993 * pin the queue now that it's allocated, scheduler exit will prune it
1995 if (!is_sync && !(*async_cfqq)) {
1996 atomic_inc(&cfqq->ref);
2000 atomic_inc(&cfqq->ref);
2005 * We drop cfq io contexts lazily, so we may find a dead one.
2008 cfq_drop_dead_cic(struct cfq_data *cfqd, struct io_context *ioc,
2009 struct cfq_io_context *cic)
2011 unsigned long flags;
2013 WARN_ON(!list_empty(&cic->queue_list));
2015 spin_lock_irqsave(&ioc->lock, flags);
2017 BUG_ON(ioc->ioc_data == cic);
2019 radix_tree_delete(&ioc->radix_root, (unsigned long) cfqd);
2020 hlist_del_rcu(&cic->cic_list);
2021 spin_unlock_irqrestore(&ioc->lock, flags);
2026 static struct cfq_io_context *
2027 cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc)
2029 struct cfq_io_context *cic;
2030 unsigned long flags;
2039 * we maintain a last-hit cache, to avoid browsing over the tree
2041 cic = rcu_dereference(ioc->ioc_data);
2042 if (cic && cic->key == cfqd) {
2048 cic = radix_tree_lookup(&ioc->radix_root, (unsigned long) cfqd);
2052 /* ->key must be copied to avoid race with cfq_exit_queue() */
2055 cfq_drop_dead_cic(cfqd, ioc, cic);
2060 spin_lock_irqsave(&ioc->lock, flags);
2061 rcu_assign_pointer(ioc->ioc_data, cic);
2062 spin_unlock_irqrestore(&ioc->lock, flags);
2070 * Add cic into ioc, using cfqd as the search key. This enables us to lookup
2071 * the process specific cfq io context when entered from the block layer.
2072 * Also adds the cic to a per-cfqd list, used when this queue is removed.
2074 static int cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
2075 struct cfq_io_context *cic, gfp_t gfp_mask)
2077 unsigned long flags;
2080 ret = radix_tree_preload(gfp_mask);
2085 spin_lock_irqsave(&ioc->lock, flags);
2086 ret = radix_tree_insert(&ioc->radix_root,
2087 (unsigned long) cfqd, cic);
2089 hlist_add_head_rcu(&cic->cic_list, &ioc->cic_list);
2090 spin_unlock_irqrestore(&ioc->lock, flags);
2092 radix_tree_preload_end();
2095 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2096 list_add(&cic->queue_list, &cfqd->cic_list);
2097 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2102 printk(KERN_ERR "cfq: cic link failed!\n");
2108 * Setup general io context and cfq io context. There can be several cfq
2109 * io contexts per general io context, if this process is doing io to more
2110 * than one device managed by cfq.
2112 static struct cfq_io_context *
2113 cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
2115 struct io_context *ioc = NULL;
2116 struct cfq_io_context *cic;
2118 might_sleep_if(gfp_mask & __GFP_WAIT);
2120 ioc = get_io_context(gfp_mask, cfqd->queue->node);
2124 cic = cfq_cic_lookup(cfqd, ioc);
2128 cic = cfq_alloc_io_context(cfqd, gfp_mask);
2132 if (cfq_cic_link(cfqd, ioc, cic, gfp_mask))
2136 smp_read_barrier_depends();
2137 if (unlikely(ioc->ioprio_changed))
2138 cfq_ioc_set_ioprio(ioc);
2144 put_io_context(ioc);
2149 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
2151 unsigned long elapsed = jiffies - cic->last_end_request;
2152 unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
2154 cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
2155 cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
2156 cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
2160 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2166 if (!cfqq->last_request_pos)
2168 else if (cfqq->last_request_pos < blk_rq_pos(rq))
2169 sdist = blk_rq_pos(rq) - cfqq->last_request_pos;
2171 sdist = cfqq->last_request_pos - blk_rq_pos(rq);
2174 * Don't allow the seek distance to get too large from the
2175 * odd fragment, pagein, etc
2177 if (cfqq->seek_samples <= 60) /* second&third seek */
2178 sdist = min(sdist, (cfqq->seek_mean * 4) + 2*1024*1024);
2180 sdist = min(sdist, (cfqq->seek_mean * 4) + 2*1024*64);
2182 cfqq->seek_samples = (7*cfqq->seek_samples + 256) / 8;
2183 cfqq->seek_total = (7*cfqq->seek_total + (u64)256*sdist) / 8;
2184 total = cfqq->seek_total + (cfqq->seek_samples/2);
2185 do_div(total, cfqq->seek_samples);
2186 cfqq->seek_mean = (sector_t)total;
2189 * If this cfqq is shared between multiple processes, check to
2190 * make sure that those processes are still issuing I/Os within
2191 * the mean seek distance. If not, it may be time to break the
2192 * queues apart again.
2194 if (cfq_cfqq_coop(cfqq)) {
2195 if (CFQQ_SEEKY(cfqq) && !cfqq->seeky_start)
2196 cfqq->seeky_start = jiffies;
2197 else if (!CFQQ_SEEKY(cfqq))
2198 cfqq->seeky_start = 0;
2203 * Disable idle window if the process thinks too long or seeks so much that
2207 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2208 struct cfq_io_context *cic)
2210 int old_idle, enable_idle;
2213 * Don't idle for async or idle io prio class
2215 if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
2218 enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
2220 if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle ||
2221 (!cfqd->cfq_latency && cfqd->hw_tag && CFQQ_SEEKY(cfqq)))
2223 else if (sample_valid(cic->ttime_samples)) {
2224 unsigned int slice_idle = cfqd->cfq_slice_idle;
2225 if (sample_valid(cfqq->seek_samples) && CFQQ_SEEKY(cfqq))
2226 slice_idle = msecs_to_jiffies(CFQ_MIN_TT);
2227 if (cic->ttime_mean > slice_idle)
2233 if (old_idle != enable_idle) {
2234 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
2236 cfq_mark_cfqq_idle_window(cfqq);
2238 cfq_clear_cfqq_idle_window(cfqq);
2243 * Check if new_cfqq should preempt the currently active queue. Return 0 for
2244 * no or if we aren't sure, a 1 will cause a preempt.
2247 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
2250 struct cfq_queue *cfqq;
2252 cfqq = cfqd->active_queue;
2256 if (cfq_slice_used(cfqq))
2259 if (cfq_class_idle(new_cfqq))
2262 if (cfq_class_idle(cfqq))
2266 * if the new request is sync, but the currently running queue is
2267 * not, let the sync request have priority.
2269 if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
2273 * So both queues are sync. Let the new request get disk time if
2274 * it's a metadata request and the current queue is doing regular IO.
2276 if (rq_is_meta(rq) && !cfqq->meta_pending)
2280 * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
2282 if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
2285 if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
2289 * if this request is as-good as one we would expect from the
2290 * current cfqq, let it preempt
2292 if (cfq_rq_close(cfqd, cfqq, rq))
2299 * cfqq preempts the active queue. if we allowed preempt with no slice left,
2300 * let it have half of its nominal slice.
2302 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2304 cfq_log_cfqq(cfqd, cfqq, "preempt");
2305 cfq_slice_expired(cfqd, 1);
2308 * Put the new queue at the front of the of the current list,
2309 * so we know that it will be selected next.
2311 BUG_ON(!cfq_cfqq_on_rr(cfqq));
2313 cfq_service_tree_add(cfqd, cfqq, 1);
2315 cfqq->slice_end = 0;
2316 cfq_mark_cfqq_slice_new(cfqq);
2320 * Called when a new fs request (rq) is added (to cfqq). Check if there's
2321 * something we should do about it
2324 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2327 struct cfq_io_context *cic = RQ_CIC(rq);
2331 cfqq->meta_pending++;
2333 cfq_update_io_thinktime(cfqd, cic);
2334 cfq_update_io_seektime(cfqd, cfqq, rq);
2335 cfq_update_idle_window(cfqd, cfqq, cic);
2337 cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
2339 if (cfqq == cfqd->active_queue) {
2341 * Remember that we saw a request from this process, but
2342 * don't start queuing just yet. Otherwise we risk seeing lots
2343 * of tiny requests, because we disrupt the normal plugging
2344 * and merging. If the request is already larger than a single
2345 * page, let it rip immediately. For that case we assume that
2346 * merging is already done. Ditto for a busy system that
2347 * has other work pending, don't risk delaying until the
2348 * idle timer unplug to continue working.
2350 if (cfq_cfqq_wait_request(cfqq)) {
2351 if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
2352 cfqd->busy_queues > 1) {
2353 del_timer(&cfqd->idle_slice_timer);
2354 __blk_run_queue(cfqd->queue);
2356 cfq_mark_cfqq_must_dispatch(cfqq);
2358 } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
2360 * not the active queue - expire current slice if it is
2361 * idle and has expired it's mean thinktime or this new queue
2362 * has some old slice time left and is of higher priority or
2363 * this new queue is RT and the current one is BE
2365 cfq_preempt_queue(cfqd, cfqq);
2366 __blk_run_queue(cfqd->queue);
2370 static void cfq_insert_request(struct request_queue *q, struct request *rq)
2372 struct cfq_data *cfqd = q->elevator->elevator_data;
2373 struct cfq_queue *cfqq = RQ_CFQQ(rq);
2375 cfq_log_cfqq(cfqd, cfqq, "insert_request");
2376 cfq_init_prio_data(cfqq, RQ_CIC(rq)->ioc);
2378 rq_set_fifo_time(rq, jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)]);
2379 list_add_tail(&rq->queuelist, &cfqq->fifo);
2382 cfq_rq_enqueued(cfqd, cfqq, rq);
2386 * Update hw_tag based on peak queue depth over 50 samples under
2389 static void cfq_update_hw_tag(struct cfq_data *cfqd)
2391 struct cfq_queue *cfqq = cfqd->active_queue;
2393 if (rq_in_driver(cfqd) > cfqd->rq_in_driver_peak)
2394 cfqd->rq_in_driver_peak = rq_in_driver(cfqd);
2396 if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
2397 rq_in_driver(cfqd) <= CFQ_HW_QUEUE_MIN)
2401 * If active queue hasn't enough requests and can idle, cfq might not
2402 * dispatch sufficient requests to hardware. Don't zero hw_tag in this
2405 if (cfqq && cfq_cfqq_idle_window(cfqq) &&
2406 cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] <
2407 CFQ_HW_QUEUE_MIN && rq_in_driver(cfqd) < CFQ_HW_QUEUE_MIN)
2410 if (cfqd->hw_tag_samples++ < 50)
2413 if (cfqd->rq_in_driver_peak >= CFQ_HW_QUEUE_MIN)
2418 cfqd->hw_tag_samples = 0;
2419 cfqd->rq_in_driver_peak = 0;
2422 static void cfq_completed_request(struct request_queue *q, struct request *rq)
2424 struct cfq_queue *cfqq = RQ_CFQQ(rq);
2425 struct cfq_data *cfqd = cfqq->cfqd;
2426 const int sync = rq_is_sync(rq);
2430 cfq_log_cfqq(cfqd, cfqq, "complete");
2432 cfq_update_hw_tag(cfqd);
2434 WARN_ON(!cfqd->rq_in_driver[sync]);
2435 WARN_ON(!cfqq->dispatched);
2436 cfqd->rq_in_driver[sync]--;
2439 if (cfq_cfqq_sync(cfqq))
2440 cfqd->sync_flight--;
2443 RQ_CIC(rq)->last_end_request = now;
2444 cfqd->last_end_sync_rq = now;
2448 * If this is the active queue, check if it needs to be expired,
2449 * or if we want to idle in case it has no pending requests.
2451 if (cfqd->active_queue == cfqq) {
2452 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
2454 if (cfq_cfqq_slice_new(cfqq)) {
2455 cfq_set_prio_slice(cfqd, cfqq);
2456 cfq_clear_cfqq_slice_new(cfqq);
2459 * If there are no requests waiting in this queue, and
2460 * there are other queues ready to issue requests, AND
2461 * those other queues are issuing requests within our
2462 * mean seek distance, give them a chance to run instead
2465 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
2466 cfq_slice_expired(cfqd, 1);
2467 else if (cfqq_empty && !cfq_close_cooperator(cfqd, cfqq) &&
2468 sync && !rq_noidle(rq))
2469 cfq_arm_slice_timer(cfqd);
2472 if (!rq_in_driver(cfqd))
2473 cfq_schedule_dispatch(cfqd);
2477 * we temporarily boost lower priority queues if they are holding fs exclusive
2478 * resources. they are boosted to normal prio (CLASS_BE/4)
2480 static void cfq_prio_boost(struct cfq_queue *cfqq)
2482 if (has_fs_excl()) {
2484 * boost idle prio on transactions that would lock out other
2485 * users of the filesystem
2487 if (cfq_class_idle(cfqq))
2488 cfqq->ioprio_class = IOPRIO_CLASS_BE;
2489 if (cfqq->ioprio > IOPRIO_NORM)
2490 cfqq->ioprio = IOPRIO_NORM;
2493 * check if we need to unboost the queue
2495 if (cfqq->ioprio_class != cfqq->org_ioprio_class)
2496 cfqq->ioprio_class = cfqq->org_ioprio_class;
2497 if (cfqq->ioprio != cfqq->org_ioprio)
2498 cfqq->ioprio = cfqq->org_ioprio;
2502 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
2504 if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
2505 cfq_mark_cfqq_must_alloc_slice(cfqq);
2506 return ELV_MQUEUE_MUST;
2509 return ELV_MQUEUE_MAY;
2512 static int cfq_may_queue(struct request_queue *q, int rw)
2514 struct cfq_data *cfqd = q->elevator->elevator_data;
2515 struct task_struct *tsk = current;
2516 struct cfq_io_context *cic;
2517 struct cfq_queue *cfqq;
2520 * don't force setup of a queue from here, as a call to may_queue
2521 * does not necessarily imply that a request actually will be queued.
2522 * so just lookup a possibly existing queue, or return 'may queue'
2525 cic = cfq_cic_lookup(cfqd, tsk->io_context);
2527 return ELV_MQUEUE_MAY;
2529 cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
2531 cfq_init_prio_data(cfqq, cic->ioc);
2532 cfq_prio_boost(cfqq);
2534 return __cfq_may_queue(cfqq);
2537 return ELV_MQUEUE_MAY;
2541 * queue lock held here
2543 static void cfq_put_request(struct request *rq)
2545 struct cfq_queue *cfqq = RQ_CFQQ(rq);
2548 const int rw = rq_data_dir(rq);
2550 BUG_ON(!cfqq->allocated[rw]);
2551 cfqq->allocated[rw]--;
2553 put_io_context(RQ_CIC(rq)->ioc);
2555 rq->elevator_private = NULL;
2556 rq->elevator_private2 = NULL;
2558 cfq_put_queue(cfqq);
2562 static struct cfq_queue *
2563 cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_context *cic,
2564 struct cfq_queue *cfqq)
2566 cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq);
2567 cic_set_cfqq(cic, cfqq->new_cfqq, 1);
2568 cfq_mark_cfqq_coop(cfqq->new_cfqq);
2569 cfq_put_queue(cfqq);
2570 return cic_to_cfqq(cic, 1);
2573 static int should_split_cfqq(struct cfq_queue *cfqq)
2575 if (cfqq->seeky_start &&
2576 time_after(jiffies, cfqq->seeky_start + CFQQ_COOP_TOUT))
2582 * Returns NULL if a new cfqq should be allocated, or the old cfqq if this
2583 * was the last process referring to said cfqq.
2585 static struct cfq_queue *
2586 split_cfqq(struct cfq_io_context *cic, struct cfq_queue *cfqq)
2588 if (cfqq_process_refs(cfqq) == 1) {
2589 cfqq->seeky_start = 0;
2590 cfqq->pid = current->pid;
2591 cfq_clear_cfqq_coop(cfqq);
2595 cic_set_cfqq(cic, NULL, 1);
2596 cfq_put_queue(cfqq);
2600 * Allocate cfq data structures associated with this request.
2603 cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
2605 struct cfq_data *cfqd = q->elevator->elevator_data;
2606 struct cfq_io_context *cic;
2607 const int rw = rq_data_dir(rq);
2608 const bool is_sync = rq_is_sync(rq);
2609 struct cfq_queue *cfqq;
2610 unsigned long flags;
2612 might_sleep_if(gfp_mask & __GFP_WAIT);
2614 cic = cfq_get_io_context(cfqd, gfp_mask);
2616 spin_lock_irqsave(q->queue_lock, flags);
2622 cfqq = cic_to_cfqq(cic, is_sync);
2623 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
2624 cfqq = cfq_get_queue(cfqd, is_sync, cic->ioc, gfp_mask);
2625 cic_set_cfqq(cic, cfqq, is_sync);
2628 * If the queue was seeky for too long, break it apart.
2630 if (cfq_cfqq_coop(cfqq) && should_split_cfqq(cfqq)) {
2631 cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq");
2632 cfqq = split_cfqq(cic, cfqq);
2638 * Check to see if this queue is scheduled to merge with
2639 * another, closely cooperating queue. The merging of
2640 * queues happens here as it must be done in process context.
2641 * The reference on new_cfqq was taken in merge_cfqqs.
2644 cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq);
2647 cfqq->allocated[rw]++;
2648 atomic_inc(&cfqq->ref);
2650 spin_unlock_irqrestore(q->queue_lock, flags);
2652 rq->elevator_private = cic;
2653 rq->elevator_private2 = cfqq;
2658 put_io_context(cic->ioc);
2660 cfq_schedule_dispatch(cfqd);
2661 spin_unlock_irqrestore(q->queue_lock, flags);
2662 cfq_log(cfqd, "set_request fail");
2666 static void cfq_kick_queue(struct work_struct *work)
2668 struct cfq_data *cfqd =
2669 container_of(work, struct cfq_data, unplug_work);
2670 struct request_queue *q = cfqd->queue;
2672 spin_lock_irq(q->queue_lock);
2673 __blk_run_queue(cfqd->queue);
2674 spin_unlock_irq(q->queue_lock);
2678 * Timer running if the active_queue is currently idling inside its time slice
2680 static void cfq_idle_slice_timer(unsigned long data)
2682 struct cfq_data *cfqd = (struct cfq_data *) data;
2683 struct cfq_queue *cfqq;
2684 unsigned long flags;
2687 cfq_log(cfqd, "idle timer fired");
2689 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2691 cfqq = cfqd->active_queue;
2696 * We saw a request before the queue expired, let it through
2698 if (cfq_cfqq_must_dispatch(cfqq))
2704 if (cfq_slice_used(cfqq))
2708 * only expire and reinvoke request handler, if there are
2709 * other queues with pending requests
2711 if (!cfqd->busy_queues)
2715 * not expired and it has a request pending, let it dispatch
2717 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
2721 cfq_slice_expired(cfqd, timed_out);
2723 cfq_schedule_dispatch(cfqd);
2725 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2728 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
2730 del_timer_sync(&cfqd->idle_slice_timer);
2731 cancel_work_sync(&cfqd->unplug_work);
2734 static void cfq_put_async_queues(struct cfq_data *cfqd)
2738 for (i = 0; i < IOPRIO_BE_NR; i++) {
2739 if (cfqd->async_cfqq[0][i])
2740 cfq_put_queue(cfqd->async_cfqq[0][i]);
2741 if (cfqd->async_cfqq[1][i])
2742 cfq_put_queue(cfqd->async_cfqq[1][i]);
2745 if (cfqd->async_idle_cfqq)
2746 cfq_put_queue(cfqd->async_idle_cfqq);
2749 static void cfq_exit_queue(struct elevator_queue *e)
2751 struct cfq_data *cfqd = e->elevator_data;
2752 struct request_queue *q = cfqd->queue;
2754 cfq_shutdown_timer_wq(cfqd);
2756 spin_lock_irq(q->queue_lock);
2758 if (cfqd->active_queue)
2759 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
2761 while (!list_empty(&cfqd->cic_list)) {
2762 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
2763 struct cfq_io_context,
2766 __cfq_exit_single_io_context(cfqd, cic);
2769 cfq_put_async_queues(cfqd);
2771 spin_unlock_irq(q->queue_lock);
2773 cfq_shutdown_timer_wq(cfqd);
2778 static void *cfq_init_queue(struct request_queue *q)
2780 struct cfq_data *cfqd;
2783 cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
2787 for (i = 0; i < 2; ++i)
2788 cfqd->service_trees[i] = CFQ_RB_ROOT;
2789 cfqd->service_tree_idle = CFQ_RB_ROOT;
2792 * Not strictly needed (since RB_ROOT just clears the node and we
2793 * zeroed cfqd on alloc), but better be safe in case someone decides
2794 * to add magic to the rb code
2796 for (i = 0; i < CFQ_PRIO_LISTS; i++)
2797 cfqd->prio_trees[i] = RB_ROOT;
2800 * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
2801 * Grab a permanent reference to it, so that the normal code flow
2802 * will not attempt to free it.
2804 cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
2805 atomic_inc(&cfqd->oom_cfqq.ref);
2807 INIT_LIST_HEAD(&cfqd->cic_list);
2811 init_timer(&cfqd->idle_slice_timer);
2812 cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
2813 cfqd->idle_slice_timer.data = (unsigned long) cfqd;
2815 INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
2817 cfqd->cfq_quantum = cfq_quantum;
2818 cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
2819 cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
2820 cfqd->cfq_back_max = cfq_back_max;
2821 cfqd->cfq_back_penalty = cfq_back_penalty;
2822 cfqd->cfq_slice[0] = cfq_slice_async;
2823 cfqd->cfq_slice[1] = cfq_slice_sync;
2824 cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
2825 cfqd->cfq_slice_idle = cfq_slice_idle;
2826 cfqd->cfq_latency = 1;
2828 cfqd->last_end_sync_rq = jiffies;
2832 static void cfq_slab_kill(void)
2835 * Caller already ensured that pending RCU callbacks are completed,
2836 * so we should have no busy allocations at this point.
2839 kmem_cache_destroy(cfq_pool);
2841 kmem_cache_destroy(cfq_ioc_pool);
2844 static int __init cfq_slab_setup(void)
2846 cfq_pool = KMEM_CACHE(cfq_queue, 0);
2850 cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0);
2861 * sysfs parts below -->
2864 cfq_var_show(unsigned int var, char *page)
2866 return sprintf(page, "%d\n", var);
2870 cfq_var_store(unsigned int *var, const char *page, size_t count)
2872 char *p = (char *) page;
2874 *var = simple_strtoul(p, &p, 10);
2878 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
2879 static ssize_t __FUNC(struct elevator_queue *e, char *page) \
2881 struct cfq_data *cfqd = e->elevator_data; \
2882 unsigned int __data = __VAR; \
2884 __data = jiffies_to_msecs(__data); \
2885 return cfq_var_show(__data, (page)); \
2887 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
2888 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
2889 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
2890 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
2891 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
2892 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
2893 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
2894 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
2895 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
2896 SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
2897 #undef SHOW_FUNCTION
2899 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
2900 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
2902 struct cfq_data *cfqd = e->elevator_data; \
2903 unsigned int __data; \
2904 int ret = cfq_var_store(&__data, (page), count); \
2905 if (__data < (MIN)) \
2907 else if (__data > (MAX)) \
2910 *(__PTR) = msecs_to_jiffies(__data); \
2912 *(__PTR) = __data; \
2915 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
2916 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
2918 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
2920 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
2921 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
2923 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
2924 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
2925 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
2926 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
2928 STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
2929 #undef STORE_FUNCTION
2931 #define CFQ_ATTR(name) \
2932 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
2934 static struct elv_fs_entry cfq_attrs[] = {
2936 CFQ_ATTR(fifo_expire_sync),
2937 CFQ_ATTR(fifo_expire_async),
2938 CFQ_ATTR(back_seek_max),
2939 CFQ_ATTR(back_seek_penalty),
2940 CFQ_ATTR(slice_sync),
2941 CFQ_ATTR(slice_async),
2942 CFQ_ATTR(slice_async_rq),
2943 CFQ_ATTR(slice_idle),
2944 CFQ_ATTR(low_latency),
2948 static struct elevator_type iosched_cfq = {
2950 .elevator_merge_fn = cfq_merge,
2951 .elevator_merged_fn = cfq_merged_request,
2952 .elevator_merge_req_fn = cfq_merged_requests,
2953 .elevator_allow_merge_fn = cfq_allow_merge,
2954 .elevator_dispatch_fn = cfq_dispatch_requests,
2955 .elevator_add_req_fn = cfq_insert_request,
2956 .elevator_activate_req_fn = cfq_activate_request,
2957 .elevator_deactivate_req_fn = cfq_deactivate_request,
2958 .elevator_queue_empty_fn = cfq_queue_empty,
2959 .elevator_completed_req_fn = cfq_completed_request,
2960 .elevator_former_req_fn = elv_rb_former_request,
2961 .elevator_latter_req_fn = elv_rb_latter_request,
2962 .elevator_set_req_fn = cfq_set_request,
2963 .elevator_put_req_fn = cfq_put_request,
2964 .elevator_may_queue_fn = cfq_may_queue,
2965 .elevator_init_fn = cfq_init_queue,
2966 .elevator_exit_fn = cfq_exit_queue,
2967 .trim = cfq_free_io_context,
2969 .elevator_attrs = cfq_attrs,
2970 .elevator_name = "cfq",
2971 .elevator_owner = THIS_MODULE,
2974 static int __init cfq_init(void)
2977 * could be 0 on HZ < 1000 setups
2979 if (!cfq_slice_async)
2980 cfq_slice_async = 1;
2981 if (!cfq_slice_idle)
2984 if (cfq_slab_setup())
2987 elv_register(&iosched_cfq);
2992 static void __exit cfq_exit(void)
2994 DECLARE_COMPLETION_ONSTACK(all_gone);
2995 elv_unregister(&iosched_cfq);
2996 ioc_gone = &all_gone;
2997 /* ioc_gone's update must be visible before reading ioc_count */
3001 * this also protects us from entering cfq_slab_kill() with
3002 * pending RCU callbacks
3004 if (elv_ioc_count_read(cfq_ioc_count))
3005 wait_for_completion(&all_gone);
3009 module_init(cfq_init);
3010 module_exit(cfq_exit);
3012 MODULE_AUTHOR("Jens Axboe");
3013 MODULE_LICENSE("GPL");
3014 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");