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/slab.h>
11 #include <linux/blkdev.h>
12 #include <linux/elevator.h>
13 #include <linux/jiffies.h>
14 #include <linux/rbtree.h>
15 #include <linux/ioprio.h>
16 #include <linux/blktrace_api.h>
17 #include "blk-cgroup.h"
22 /* max queue in one round of service */
23 static const int cfq_quantum = 8;
24 static const int cfq_fifo_expire[2] = { HZ / 4, HZ / 8 };
25 /* maximum backwards seek, in KiB */
26 static const int cfq_back_max = 16 * 1024;
27 /* penalty of a backwards seek */
28 static const int cfq_back_penalty = 2;
29 static const int cfq_slice_sync = HZ / 10;
30 static int cfq_slice_async = HZ / 25;
31 static const int cfq_slice_async_rq = 2;
32 static int cfq_slice_idle = HZ / 125;
33 static const int cfq_target_latency = HZ * 3/10; /* 300 ms */
34 static const int cfq_hist_divisor = 4;
37 * offset from end of service tree
39 #define CFQ_IDLE_DELAY (HZ / 5)
42 * below this threshold, we consider thinktime immediate
44 #define CFQ_MIN_TT (2)
46 #define CFQ_SLICE_SCALE (5)
47 #define CFQ_HW_QUEUE_MIN (5)
48 #define CFQ_SERVICE_SHIFT 12
50 #define CFQQ_SEEK_THR (sector_t)(8 * 100)
51 #define CFQQ_CLOSE_THR (sector_t)(8 * 1024)
52 #define CFQQ_SECT_THR_NONROT (sector_t)(2 * 32)
53 #define CFQQ_SEEKY(cfqq) (hweight32(cfqq->seek_history) > 32/8)
56 ((struct cfq_io_context *) (rq)->elevator_private)
57 #define RQ_CFQQ(rq) (struct cfq_queue *) ((rq)->elevator_private2)
59 static struct kmem_cache *cfq_pool;
60 static struct kmem_cache *cfq_ioc_pool;
62 static DEFINE_PER_CPU(unsigned long, cfq_ioc_count);
63 static struct completion *ioc_gone;
64 static DEFINE_SPINLOCK(ioc_gone_lock);
66 #define CFQ_PRIO_LISTS IOPRIO_BE_NR
67 #define cfq_class_idle(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
68 #define cfq_class_rt(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
70 #define sample_valid(samples) ((samples) > 80)
71 #define rb_entry_cfqg(node) rb_entry((node), struct cfq_group, rb_node)
74 * Most of our rbtree usage is for sorting with min extraction, so
75 * if we cache the leftmost node we don't have to walk down the tree
76 * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should
77 * move this into the elevator for the rq sorting as well.
83 unsigned total_weight;
85 struct rb_node *active;
87 #define CFQ_RB_ROOT (struct cfq_rb_root) { .rb = RB_ROOT, .left = NULL, \
88 .count = 0, .min_vdisktime = 0, }
91 * Per process-grouping structure
96 /* various state flags, see below */
99 struct cfq_data *cfqd;
100 /* service_tree member */
101 struct rb_node rb_node;
102 /* service_tree key */
103 unsigned long rb_key;
104 /* prio tree member */
105 struct rb_node p_node;
106 /* prio tree root we belong to, if any */
107 struct rb_root *p_root;
108 /* sorted list of pending requests */
109 struct rb_root sort_list;
110 /* if fifo isn't expired, next request to serve */
111 struct request *next_rq;
112 /* requests queued in sort_list */
114 /* currently allocated requests */
116 /* fifo list of requests in sort_list */
117 struct list_head fifo;
119 /* time when queue got scheduled in to dispatch first request. */
120 unsigned long dispatch_start;
121 unsigned int allocated_slice;
122 unsigned int slice_dispatch;
123 /* time when first request from queue completed and slice started. */
124 unsigned long slice_start;
125 unsigned long slice_end;
128 /* pending metadata requests */
130 /* number of requests that are on the dispatch list or inside driver */
133 /* io prio of this group */
134 unsigned short ioprio, org_ioprio;
135 unsigned short ioprio_class, org_ioprio_class;
140 sector_t last_request_pos;
142 struct cfq_rb_root *service_tree;
143 struct cfq_queue *new_cfqq;
144 struct cfq_group *cfqg;
145 struct cfq_group *orig_cfqg;
149 * First index in the service_trees.
150 * IDLE is handled separately, so it has negative index
159 * Second index in the service_trees.
163 SYNC_NOIDLE_WORKLOAD = 1,
167 /* This is per cgroup per device grouping structure */
169 /* group service_tree member */
170 struct rb_node rb_node;
172 /* group service_tree key */
177 /* number of cfqq currently on this group */
180 /* Per group busy queus average. Useful for workload slice calc. */
181 unsigned int busy_queues_avg[2];
183 * rr lists of queues with requests, onle rr for each priority class.
184 * Counts are embedded in the cfq_rb_root
186 struct cfq_rb_root service_trees[2][3];
187 struct cfq_rb_root service_tree_idle;
189 unsigned long saved_workload_slice;
190 enum wl_type_t saved_workload;
191 enum wl_prio_t saved_serving_prio;
192 struct blkio_group blkg;
193 #ifdef CONFIG_CFQ_GROUP_IOSCHED
194 struct hlist_node cfqd_node;
200 * Per block device queue structure
203 struct request_queue *queue;
204 /* Root service tree for cfq_groups */
205 struct cfq_rb_root grp_service_tree;
206 struct cfq_group root_group;
209 * The priority currently being served
211 enum wl_prio_t serving_prio;
212 enum wl_type_t serving_type;
213 unsigned long workload_expires;
214 struct cfq_group *serving_group;
215 bool noidle_tree_requires_idle;
218 * Each priority tree is sorted by next_request position. These
219 * trees are used when determining if two or more queues are
220 * interleaving requests (see cfq_close_cooperator).
222 struct rb_root prio_trees[CFQ_PRIO_LISTS];
224 unsigned int busy_queues;
230 * queue-depth detection
236 * -1 => indeterminate, (cfq will behave as if NCQ is present, to allow better detection)
237 * 1 => NCQ is present (hw_tag_est_depth is the estimated max depth)
240 int hw_tag_est_depth;
241 unsigned int hw_tag_samples;
244 * idle window management
246 struct timer_list idle_slice_timer;
247 struct work_struct unplug_work;
249 struct cfq_queue *active_queue;
250 struct cfq_io_context *active_cic;
253 * async queue for each priority case
255 struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
256 struct cfq_queue *async_idle_cfqq;
258 sector_t last_position;
261 * tunables, see top of file
263 unsigned int cfq_quantum;
264 unsigned int cfq_fifo_expire[2];
265 unsigned int cfq_back_penalty;
266 unsigned int cfq_back_max;
267 unsigned int cfq_slice[2];
268 unsigned int cfq_slice_async_rq;
269 unsigned int cfq_slice_idle;
270 unsigned int cfq_latency;
271 unsigned int cfq_group_isolation;
273 struct list_head cic_list;
276 * Fallback dummy cfqq for extreme OOM conditions
278 struct cfq_queue oom_cfqq;
280 unsigned long last_delayed_sync;
282 /* List of cfq groups being managed on this device*/
283 struct hlist_head cfqg_list;
287 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd);
289 static struct cfq_rb_root *service_tree_for(struct cfq_group *cfqg,
296 if (prio == IDLE_WORKLOAD)
297 return &cfqg->service_tree_idle;
299 return &cfqg->service_trees[prio][type];
302 enum cfqq_state_flags {
303 CFQ_CFQQ_FLAG_on_rr = 0, /* on round-robin busy list */
304 CFQ_CFQQ_FLAG_wait_request, /* waiting for a request */
305 CFQ_CFQQ_FLAG_must_dispatch, /* must be allowed a dispatch */
306 CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
307 CFQ_CFQQ_FLAG_fifo_expire, /* FIFO checked in this slice */
308 CFQ_CFQQ_FLAG_idle_window, /* slice idling enabled */
309 CFQ_CFQQ_FLAG_prio_changed, /* task priority has changed */
310 CFQ_CFQQ_FLAG_slice_new, /* no requests dispatched in slice */
311 CFQ_CFQQ_FLAG_sync, /* synchronous queue */
312 CFQ_CFQQ_FLAG_coop, /* cfqq is shared */
313 CFQ_CFQQ_FLAG_split_coop, /* shared cfqq will be splitted */
314 CFQ_CFQQ_FLAG_deep, /* sync cfqq experienced large depth */
315 CFQ_CFQQ_FLAG_wait_busy, /* Waiting for next request */
318 #define CFQ_CFQQ_FNS(name) \
319 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \
321 (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name); \
323 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \
325 (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \
327 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \
329 return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \
333 CFQ_CFQQ_FNS(wait_request);
334 CFQ_CFQQ_FNS(must_dispatch);
335 CFQ_CFQQ_FNS(must_alloc_slice);
336 CFQ_CFQQ_FNS(fifo_expire);
337 CFQ_CFQQ_FNS(idle_window);
338 CFQ_CFQQ_FNS(prio_changed);
339 CFQ_CFQQ_FNS(slice_new);
342 CFQ_CFQQ_FNS(split_coop);
344 CFQ_CFQQ_FNS(wait_busy);
347 #ifdef CONFIG_DEBUG_CFQ_IOSCHED
348 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
349 blk_add_trace_msg((cfqd)->queue, "cfq%d%c %s " fmt, (cfqq)->pid, \
350 cfq_cfqq_sync((cfqq)) ? 'S' : 'A', \
351 blkg_path(&(cfqq)->cfqg->blkg), ##args);
353 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...) \
354 blk_add_trace_msg((cfqd)->queue, "%s " fmt, \
355 blkg_path(&(cfqg)->blkg), ##args); \
358 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
359 blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args)
360 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...) do {} while (0);
362 #define cfq_log(cfqd, fmt, args...) \
363 blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
365 /* Traverses through cfq group service trees */
366 #define for_each_cfqg_st(cfqg, i, j, st) \
367 for (i = 0; i <= IDLE_WORKLOAD; i++) \
368 for (j = 0, st = i < IDLE_WORKLOAD ? &cfqg->service_trees[i][j]\
369 : &cfqg->service_tree_idle; \
370 (i < IDLE_WORKLOAD && j <= SYNC_WORKLOAD) || \
371 (i == IDLE_WORKLOAD && j == 0); \
372 j++, st = i < IDLE_WORKLOAD ? \
373 &cfqg->service_trees[i][j]: NULL) \
376 static inline enum wl_prio_t cfqq_prio(struct cfq_queue *cfqq)
378 if (cfq_class_idle(cfqq))
379 return IDLE_WORKLOAD;
380 if (cfq_class_rt(cfqq))
386 static enum wl_type_t cfqq_type(struct cfq_queue *cfqq)
388 if (!cfq_cfqq_sync(cfqq))
389 return ASYNC_WORKLOAD;
390 if (!cfq_cfqq_idle_window(cfqq))
391 return SYNC_NOIDLE_WORKLOAD;
392 return SYNC_WORKLOAD;
395 static inline int cfq_group_busy_queues_wl(enum wl_prio_t wl,
396 struct cfq_data *cfqd,
397 struct cfq_group *cfqg)
399 if (wl == IDLE_WORKLOAD)
400 return cfqg->service_tree_idle.count;
402 return cfqg->service_trees[wl][ASYNC_WORKLOAD].count
403 + cfqg->service_trees[wl][SYNC_NOIDLE_WORKLOAD].count
404 + cfqg->service_trees[wl][SYNC_WORKLOAD].count;
407 static inline int cfqg_busy_async_queues(struct cfq_data *cfqd,
408 struct cfq_group *cfqg)
410 return cfqg->service_trees[RT_WORKLOAD][ASYNC_WORKLOAD].count
411 + cfqg->service_trees[BE_WORKLOAD][ASYNC_WORKLOAD].count;
414 static void cfq_dispatch_insert(struct request_queue *, struct request *);
415 static struct cfq_queue *cfq_get_queue(struct cfq_data *, bool,
416 struct io_context *, gfp_t);
417 static struct cfq_io_context *cfq_cic_lookup(struct cfq_data *,
418 struct io_context *);
420 static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_context *cic,
423 return cic->cfqq[is_sync];
426 static inline void cic_set_cfqq(struct cfq_io_context *cic,
427 struct cfq_queue *cfqq, bool is_sync)
429 cic->cfqq[is_sync] = cfqq;
433 * We regard a request as SYNC, if it's either a read or has the SYNC bit
434 * set (in which case it could also be direct WRITE).
436 static inline bool cfq_bio_sync(struct bio *bio)
438 return bio_data_dir(bio) == READ || bio_rw_flagged(bio, BIO_RW_SYNCIO);
442 * scheduler run of queue, if there are requests pending and no one in the
443 * driver that will restart queueing
445 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
447 if (cfqd->busy_queues) {
448 cfq_log(cfqd, "schedule dispatch");
449 kblockd_schedule_work(cfqd->queue, &cfqd->unplug_work);
453 static int cfq_queue_empty(struct request_queue *q)
455 struct cfq_data *cfqd = q->elevator->elevator_data;
457 return !cfqd->rq_queued;
461 * Scale schedule slice based on io priority. Use the sync time slice only
462 * if a queue is marked sync and has sync io queued. A sync queue with async
463 * io only, should not get full sync slice length.
465 static inline int cfq_prio_slice(struct cfq_data *cfqd, bool sync,
468 const int base_slice = cfqd->cfq_slice[sync];
470 WARN_ON(prio >= IOPRIO_BE_NR);
472 return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
476 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
478 return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
481 static inline u64 cfq_scale_slice(unsigned long delta, struct cfq_group *cfqg)
483 u64 d = delta << CFQ_SERVICE_SHIFT;
485 d = d * BLKIO_WEIGHT_DEFAULT;
486 do_div(d, cfqg->weight);
490 static inline u64 max_vdisktime(u64 min_vdisktime, u64 vdisktime)
492 s64 delta = (s64)(vdisktime - min_vdisktime);
494 min_vdisktime = vdisktime;
496 return min_vdisktime;
499 static inline u64 min_vdisktime(u64 min_vdisktime, u64 vdisktime)
501 s64 delta = (s64)(vdisktime - min_vdisktime);
503 min_vdisktime = vdisktime;
505 return min_vdisktime;
508 static void update_min_vdisktime(struct cfq_rb_root *st)
510 u64 vdisktime = st->min_vdisktime;
511 struct cfq_group *cfqg;
514 cfqg = rb_entry_cfqg(st->active);
515 vdisktime = cfqg->vdisktime;
519 cfqg = rb_entry_cfqg(st->left);
520 vdisktime = min_vdisktime(vdisktime, cfqg->vdisktime);
523 st->min_vdisktime = max_vdisktime(st->min_vdisktime, vdisktime);
527 * get averaged number of queues of RT/BE priority.
528 * average is updated, with a formula that gives more weight to higher numbers,
529 * to quickly follows sudden increases and decrease slowly
532 static inline unsigned cfq_group_get_avg_queues(struct cfq_data *cfqd,
533 struct cfq_group *cfqg, bool rt)
535 unsigned min_q, max_q;
536 unsigned mult = cfq_hist_divisor - 1;
537 unsigned round = cfq_hist_divisor / 2;
538 unsigned busy = cfq_group_busy_queues_wl(rt, cfqd, cfqg);
540 min_q = min(cfqg->busy_queues_avg[rt], busy);
541 max_q = max(cfqg->busy_queues_avg[rt], busy);
542 cfqg->busy_queues_avg[rt] = (mult * max_q + min_q + round) /
544 return cfqg->busy_queues_avg[rt];
547 static inline unsigned
548 cfq_group_slice(struct cfq_data *cfqd, struct cfq_group *cfqg)
550 struct cfq_rb_root *st = &cfqd->grp_service_tree;
552 return cfq_target_latency * cfqg->weight / st->total_weight;
556 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
558 unsigned slice = cfq_prio_to_slice(cfqd, cfqq);
559 if (cfqd->cfq_latency) {
561 * interested queues (we consider only the ones with the same
562 * priority class in the cfq group)
564 unsigned iq = cfq_group_get_avg_queues(cfqd, cfqq->cfqg,
566 unsigned sync_slice = cfqd->cfq_slice[1];
567 unsigned expect_latency = sync_slice * iq;
568 unsigned group_slice = cfq_group_slice(cfqd, cfqq->cfqg);
570 if (expect_latency > group_slice) {
571 unsigned base_low_slice = 2 * cfqd->cfq_slice_idle;
572 /* scale low_slice according to IO priority
573 * and sync vs async */
575 min(slice, base_low_slice * slice / sync_slice);
576 /* the adapted slice value is scaled to fit all iqs
577 * into the target latency */
578 slice = max(slice * group_slice / expect_latency,
582 cfqq->slice_start = jiffies;
583 cfqq->slice_end = jiffies + slice;
584 cfqq->allocated_slice = slice;
585 cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
589 * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
590 * isn't valid until the first request from the dispatch is activated
591 * and the slice time set.
593 static inline bool cfq_slice_used(struct cfq_queue *cfqq)
595 if (cfq_cfqq_slice_new(cfqq))
597 if (time_before(jiffies, cfqq->slice_end))
604 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
605 * We choose the request that is closest to the head right now. Distance
606 * behind the head is penalized and only allowed to a certain extent.
608 static struct request *
609 cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2, sector_t last)
611 sector_t s1, s2, d1 = 0, d2 = 0;
612 unsigned long back_max;
613 #define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */
614 #define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */
615 unsigned wrap = 0; /* bit mask: requests behind the disk head? */
617 if (rq1 == NULL || rq1 == rq2)
622 if (rq_is_sync(rq1) && !rq_is_sync(rq2))
624 else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
626 if (rq_is_meta(rq1) && !rq_is_meta(rq2))
628 else if (rq_is_meta(rq2) && !rq_is_meta(rq1))
631 s1 = blk_rq_pos(rq1);
632 s2 = blk_rq_pos(rq2);
635 * by definition, 1KiB is 2 sectors
637 back_max = cfqd->cfq_back_max * 2;
640 * Strict one way elevator _except_ in the case where we allow
641 * short backward seeks which are biased as twice the cost of a
642 * similar forward seek.
646 else if (s1 + back_max >= last)
647 d1 = (last - s1) * cfqd->cfq_back_penalty;
649 wrap |= CFQ_RQ1_WRAP;
653 else if (s2 + back_max >= last)
654 d2 = (last - s2) * cfqd->cfq_back_penalty;
656 wrap |= CFQ_RQ2_WRAP;
658 /* Found required data */
661 * By doing switch() on the bit mask "wrap" we avoid having to
662 * check two variables for all permutations: --> faster!
665 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
681 case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
684 * Since both rqs are wrapped,
685 * start with the one that's further behind head
686 * (--> only *one* back seek required),
687 * since back seek takes more time than forward.
697 * The below is leftmost cache rbtree addon
699 static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
701 /* Service tree is empty */
706 root->left = rb_first(&root->rb);
709 return rb_entry(root->left, struct cfq_queue, rb_node);
714 static struct cfq_group *cfq_rb_first_group(struct cfq_rb_root *root)
717 root->left = rb_first(&root->rb);
720 return rb_entry_cfqg(root->left);
725 static void rb_erase_init(struct rb_node *n, struct rb_root *root)
731 static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
735 rb_erase_init(n, &root->rb);
740 * would be nice to take fifo expire time into account as well
742 static struct request *
743 cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
744 struct request *last)
746 struct rb_node *rbnext = rb_next(&last->rb_node);
747 struct rb_node *rbprev = rb_prev(&last->rb_node);
748 struct request *next = NULL, *prev = NULL;
750 BUG_ON(RB_EMPTY_NODE(&last->rb_node));
753 prev = rb_entry_rq(rbprev);
756 next = rb_entry_rq(rbnext);
758 rbnext = rb_first(&cfqq->sort_list);
759 if (rbnext && rbnext != &last->rb_node)
760 next = rb_entry_rq(rbnext);
763 return cfq_choose_req(cfqd, next, prev, blk_rq_pos(last));
766 static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
767 struct cfq_queue *cfqq)
770 * just an approximation, should be ok.
772 return (cfqq->cfqg->nr_cfqq - 1) * (cfq_prio_slice(cfqd, 1, 0) -
773 cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
777 cfqg_key(struct cfq_rb_root *st, struct cfq_group *cfqg)
779 return cfqg->vdisktime - st->min_vdisktime;
783 __cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
785 struct rb_node **node = &st->rb.rb_node;
786 struct rb_node *parent = NULL;
787 struct cfq_group *__cfqg;
788 s64 key = cfqg_key(st, cfqg);
791 while (*node != NULL) {
793 __cfqg = rb_entry_cfqg(parent);
795 if (key < cfqg_key(st, __cfqg))
796 node = &parent->rb_left;
798 node = &parent->rb_right;
804 st->left = &cfqg->rb_node;
806 rb_link_node(&cfqg->rb_node, parent, node);
807 rb_insert_color(&cfqg->rb_node, &st->rb);
811 cfq_group_service_tree_add(struct cfq_data *cfqd, struct cfq_group *cfqg)
813 struct cfq_rb_root *st = &cfqd->grp_service_tree;
814 struct cfq_group *__cfqg;
822 * Currently put the group at the end. Later implement something
823 * so that groups get lesser vtime based on their weights, so that
824 * if group does not loose all if it was not continously backlogged.
826 n = rb_last(&st->rb);
828 __cfqg = rb_entry_cfqg(n);
829 cfqg->vdisktime = __cfqg->vdisktime + CFQ_IDLE_DELAY;
831 cfqg->vdisktime = st->min_vdisktime;
833 __cfq_group_service_tree_add(st, cfqg);
835 st->total_weight += cfqg->weight;
839 cfq_group_service_tree_del(struct cfq_data *cfqd, struct cfq_group *cfqg)
841 struct cfq_rb_root *st = &cfqd->grp_service_tree;
843 if (st->active == &cfqg->rb_node)
846 BUG_ON(cfqg->nr_cfqq < 1);
849 /* If there are other cfq queues under this group, don't delete it */
853 cfq_log_cfqg(cfqd, cfqg, "del_from_rr group");
855 st->total_weight -= cfqg->weight;
856 if (!RB_EMPTY_NODE(&cfqg->rb_node))
857 cfq_rb_erase(&cfqg->rb_node, st);
858 cfqg->saved_workload_slice = 0;
859 blkiocg_update_dequeue_stats(&cfqg->blkg, 1);
862 static inline unsigned int cfq_cfqq_slice_usage(struct cfq_queue *cfqq)
864 unsigned int slice_used;
867 * Queue got expired before even a single request completed or
868 * got expired immediately after first request completion.
870 if (!cfqq->slice_start || cfqq->slice_start == jiffies) {
872 * Also charge the seek time incurred to the group, otherwise
873 * if there are mutiple queues in the group, each can dispatch
874 * a single request on seeky media and cause lots of seek time
875 * and group will never know it.
877 slice_used = max_t(unsigned, (jiffies - cfqq->dispatch_start),
880 slice_used = jiffies - cfqq->slice_start;
881 if (slice_used > cfqq->allocated_slice)
882 slice_used = cfqq->allocated_slice;
885 cfq_log_cfqq(cfqq->cfqd, cfqq, "sl_used=%u", slice_used);
889 static void cfq_group_served(struct cfq_data *cfqd, struct cfq_group *cfqg,
890 struct cfq_queue *cfqq, bool forced)
892 struct cfq_rb_root *st = &cfqd->grp_service_tree;
893 unsigned int used_sl, charge_sl;
894 int nr_sync = cfqg->nr_cfqq - cfqg_busy_async_queues(cfqd, cfqg)
895 - cfqg->service_tree_idle.count;
898 used_sl = charge_sl = cfq_cfqq_slice_usage(cfqq);
900 if (!cfq_cfqq_sync(cfqq) && !nr_sync)
901 charge_sl = cfqq->allocated_slice;
903 /* Can't update vdisktime while group is on service tree */
904 cfq_rb_erase(&cfqg->rb_node, st);
905 cfqg->vdisktime += cfq_scale_slice(charge_sl, cfqg);
906 __cfq_group_service_tree_add(st, cfqg);
908 /* This group is being expired. Save the context */
909 if (time_after(cfqd->workload_expires, jiffies)) {
910 cfqg->saved_workload_slice = cfqd->workload_expires
912 cfqg->saved_workload = cfqd->serving_type;
913 cfqg->saved_serving_prio = cfqd->serving_prio;
915 cfqg->saved_workload_slice = 0;
917 cfq_log_cfqg(cfqd, cfqg, "served: vt=%llu min_vt=%llu", cfqg->vdisktime,
919 blkiocg_update_timeslice_used(&cfqg->blkg, used_sl);
920 blkiocg_set_start_empty_time(&cfqg->blkg, forced);
923 #ifdef CONFIG_CFQ_GROUP_IOSCHED
924 static inline struct cfq_group *cfqg_of_blkg(struct blkio_group *blkg)
927 return container_of(blkg, struct cfq_group, blkg);
932 cfq_update_blkio_group_weight(struct blkio_group *blkg, unsigned int weight)
934 cfqg_of_blkg(blkg)->weight = weight;
937 static struct cfq_group *
938 cfq_find_alloc_cfqg(struct cfq_data *cfqd, struct cgroup *cgroup, int create)
940 struct blkio_cgroup *blkcg = cgroup_to_blkio_cgroup(cgroup);
941 struct cfq_group *cfqg = NULL;
944 struct cfq_rb_root *st;
945 struct backing_dev_info *bdi = &cfqd->queue->backing_dev_info;
946 unsigned int major, minor;
948 cfqg = cfqg_of_blkg(blkiocg_lookup_group(blkcg, key));
949 if (cfqg && !cfqg->blkg.dev && bdi->dev && dev_name(bdi->dev)) {
950 sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
951 cfqg->blkg.dev = MKDEV(major, minor);
957 cfqg = kzalloc_node(sizeof(*cfqg), GFP_ATOMIC, cfqd->queue->node);
961 for_each_cfqg_st(cfqg, i, j, st)
963 RB_CLEAR_NODE(&cfqg->rb_node);
964 blkio_group_init(&cfqg->blkg);
967 * Take the initial reference that will be released on destroy
968 * This can be thought of a joint reference by cgroup and
969 * elevator which will be dropped by either elevator exit
970 * or cgroup deletion path depending on who is exiting first.
972 atomic_set(&cfqg->ref, 1);
974 /* Add group onto cgroup list */
975 sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
976 blkiocg_add_blkio_group(blkcg, &cfqg->blkg, (void *)cfqd,
977 MKDEV(major, minor));
978 cfqg->weight = blkcg_get_weight(blkcg, cfqg->blkg.dev);
980 /* Add group on cfqd list */
981 hlist_add_head(&cfqg->cfqd_node, &cfqd->cfqg_list);
988 * Search for the cfq group current task belongs to. If create = 1, then also
989 * create the cfq group if it does not exist. request_queue lock must be held.
991 static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd, int create)
993 struct cgroup *cgroup;
994 struct cfq_group *cfqg = NULL;
997 cgroup = task_cgroup(current, blkio_subsys_id);
998 cfqg = cfq_find_alloc_cfqg(cfqd, cgroup, create);
1000 cfqg = &cfqd->root_group;
1005 static void cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg)
1007 /* Currently, all async queues are mapped to root group */
1008 if (!cfq_cfqq_sync(cfqq))
1009 cfqg = &cfqq->cfqd->root_group;
1012 /* cfqq reference on cfqg */
1013 atomic_inc(&cfqq->cfqg->ref);
1016 static void cfq_put_cfqg(struct cfq_group *cfqg)
1018 struct cfq_rb_root *st;
1021 BUG_ON(atomic_read(&cfqg->ref) <= 0);
1022 if (!atomic_dec_and_test(&cfqg->ref))
1024 for_each_cfqg_st(cfqg, i, j, st)
1025 BUG_ON(!RB_EMPTY_ROOT(&st->rb) || st->active != NULL);
1029 static void cfq_destroy_cfqg(struct cfq_data *cfqd, struct cfq_group *cfqg)
1031 /* Something wrong if we are trying to remove same group twice */
1032 BUG_ON(hlist_unhashed(&cfqg->cfqd_node));
1034 hlist_del_init(&cfqg->cfqd_node);
1037 * Put the reference taken at the time of creation so that when all
1038 * queues are gone, group can be destroyed.
1043 static void cfq_release_cfq_groups(struct cfq_data *cfqd)
1045 struct hlist_node *pos, *n;
1046 struct cfq_group *cfqg;
1048 hlist_for_each_entry_safe(cfqg, pos, n, &cfqd->cfqg_list, cfqd_node) {
1050 * If cgroup removal path got to blk_group first and removed
1051 * it from cgroup list, then it will take care of destroying
1054 if (!blkiocg_del_blkio_group(&cfqg->blkg))
1055 cfq_destroy_cfqg(cfqd, cfqg);
1060 * Blk cgroup controller notification saying that blkio_group object is being
1061 * delinked as associated cgroup object is going away. That also means that
1062 * no new IO will come in this group. So get rid of this group as soon as
1063 * any pending IO in the group is finished.
1065 * This function is called under rcu_read_lock(). key is the rcu protected
1066 * pointer. That means "key" is a valid cfq_data pointer as long as we are rcu
1069 * "key" was fetched from blkio_group under blkio_cgroup->lock. That means
1070 * it should not be NULL as even if elevator was exiting, cgroup deltion
1071 * path got to it first.
1073 void cfq_unlink_blkio_group(void *key, struct blkio_group *blkg)
1075 unsigned long flags;
1076 struct cfq_data *cfqd = key;
1078 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1079 cfq_destroy_cfqg(cfqd, cfqg_of_blkg(blkg));
1080 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1083 #else /* GROUP_IOSCHED */
1084 static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd, int create)
1086 return &cfqd->root_group;
1089 cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) {
1093 static void cfq_release_cfq_groups(struct cfq_data *cfqd) {}
1094 static inline void cfq_put_cfqg(struct cfq_group *cfqg) {}
1096 #endif /* GROUP_IOSCHED */
1099 * The cfqd->service_trees holds all pending cfq_queue's that have
1100 * requests waiting to be processed. It is sorted in the order that
1101 * we will service the queues.
1103 static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1106 struct rb_node **p, *parent;
1107 struct cfq_queue *__cfqq;
1108 unsigned long rb_key;
1109 struct cfq_rb_root *service_tree;
1112 int group_changed = 0;
1114 #ifdef CONFIG_CFQ_GROUP_IOSCHED
1115 if (!cfqd->cfq_group_isolation
1116 && cfqq_type(cfqq) == SYNC_NOIDLE_WORKLOAD
1117 && cfqq->cfqg && cfqq->cfqg != &cfqd->root_group) {
1118 /* Move this cfq to root group */
1119 cfq_log_cfqq(cfqd, cfqq, "moving to root group");
1120 if (!RB_EMPTY_NODE(&cfqq->rb_node))
1121 cfq_group_service_tree_del(cfqd, cfqq->cfqg);
1122 cfqq->orig_cfqg = cfqq->cfqg;
1123 cfqq->cfqg = &cfqd->root_group;
1124 atomic_inc(&cfqd->root_group.ref);
1126 } else if (!cfqd->cfq_group_isolation
1127 && cfqq_type(cfqq) == SYNC_WORKLOAD && cfqq->orig_cfqg) {
1128 /* cfqq is sequential now needs to go to its original group */
1129 BUG_ON(cfqq->cfqg != &cfqd->root_group);
1130 if (!RB_EMPTY_NODE(&cfqq->rb_node))
1131 cfq_group_service_tree_del(cfqd, cfqq->cfqg);
1132 cfq_put_cfqg(cfqq->cfqg);
1133 cfqq->cfqg = cfqq->orig_cfqg;
1134 cfqq->orig_cfqg = NULL;
1136 cfq_log_cfqq(cfqd, cfqq, "moved to origin group");
1140 service_tree = service_tree_for(cfqq->cfqg, cfqq_prio(cfqq),
1142 if (cfq_class_idle(cfqq)) {
1143 rb_key = CFQ_IDLE_DELAY;
1144 parent = rb_last(&service_tree->rb);
1145 if (parent && parent != &cfqq->rb_node) {
1146 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1147 rb_key += __cfqq->rb_key;
1150 } else if (!add_front) {
1152 * Get our rb key offset. Subtract any residual slice
1153 * value carried from last service. A negative resid
1154 * count indicates slice overrun, and this should position
1155 * the next service time further away in the tree.
1157 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
1158 rb_key -= cfqq->slice_resid;
1159 cfqq->slice_resid = 0;
1162 __cfqq = cfq_rb_first(service_tree);
1163 rb_key += __cfqq ? __cfqq->rb_key : jiffies;
1166 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1169 * same position, nothing more to do
1171 if (rb_key == cfqq->rb_key &&
1172 cfqq->service_tree == service_tree)
1175 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1176 cfqq->service_tree = NULL;
1181 cfqq->service_tree = service_tree;
1182 p = &service_tree->rb.rb_node;
1187 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1190 * sort by key, that represents service time.
1192 if (time_before(rb_key, __cfqq->rb_key))
1195 n = &(*p)->rb_right;
1203 service_tree->left = &cfqq->rb_node;
1205 cfqq->rb_key = rb_key;
1206 rb_link_node(&cfqq->rb_node, parent, p);
1207 rb_insert_color(&cfqq->rb_node, &service_tree->rb);
1208 service_tree->count++;
1209 if ((add_front || !new_cfqq) && !group_changed)
1211 cfq_group_service_tree_add(cfqd, cfqq->cfqg);
1214 static struct cfq_queue *
1215 cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
1216 sector_t sector, struct rb_node **ret_parent,
1217 struct rb_node ***rb_link)
1219 struct rb_node **p, *parent;
1220 struct cfq_queue *cfqq = NULL;
1228 cfqq = rb_entry(parent, struct cfq_queue, p_node);
1231 * Sort strictly based on sector. Smallest to the left,
1232 * largest to the right.
1234 if (sector > blk_rq_pos(cfqq->next_rq))
1235 n = &(*p)->rb_right;
1236 else if (sector < blk_rq_pos(cfqq->next_rq))
1244 *ret_parent = parent;
1250 static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1252 struct rb_node **p, *parent;
1253 struct cfq_queue *__cfqq;
1256 rb_erase(&cfqq->p_node, cfqq->p_root);
1257 cfqq->p_root = NULL;
1260 if (cfq_class_idle(cfqq))
1265 cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
1266 __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
1267 blk_rq_pos(cfqq->next_rq), &parent, &p);
1269 rb_link_node(&cfqq->p_node, parent, p);
1270 rb_insert_color(&cfqq->p_node, cfqq->p_root);
1272 cfqq->p_root = NULL;
1276 * Update cfqq's position in the service tree.
1278 static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1281 * Resorting requires the cfqq to be on the RR list already.
1283 if (cfq_cfqq_on_rr(cfqq)) {
1284 cfq_service_tree_add(cfqd, cfqq, 0);
1285 cfq_prio_tree_add(cfqd, cfqq);
1290 * add to busy list of queues for service, trying to be fair in ordering
1291 * the pending list according to last request service
1293 static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1295 cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
1296 BUG_ON(cfq_cfqq_on_rr(cfqq));
1297 cfq_mark_cfqq_on_rr(cfqq);
1298 cfqd->busy_queues++;
1300 cfq_resort_rr_list(cfqd, cfqq);
1304 * Called when the cfqq no longer has requests pending, remove it from
1307 static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1309 cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
1310 BUG_ON(!cfq_cfqq_on_rr(cfqq));
1311 cfq_clear_cfqq_on_rr(cfqq);
1313 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1314 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1315 cfqq->service_tree = NULL;
1318 rb_erase(&cfqq->p_node, cfqq->p_root);
1319 cfqq->p_root = NULL;
1322 cfq_group_service_tree_del(cfqd, cfqq->cfqg);
1323 BUG_ON(!cfqd->busy_queues);
1324 cfqd->busy_queues--;
1328 * rb tree support functions
1330 static void cfq_del_rq_rb(struct request *rq)
1332 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1333 const int sync = rq_is_sync(rq);
1335 BUG_ON(!cfqq->queued[sync]);
1336 cfqq->queued[sync]--;
1338 elv_rb_del(&cfqq->sort_list, rq);
1340 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) {
1342 * Queue will be deleted from service tree when we actually
1343 * expire it later. Right now just remove it from prio tree
1347 rb_erase(&cfqq->p_node, cfqq->p_root);
1348 cfqq->p_root = NULL;
1353 static void cfq_add_rq_rb(struct request *rq)
1355 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1356 struct cfq_data *cfqd = cfqq->cfqd;
1357 struct request *__alias, *prev;
1359 cfqq->queued[rq_is_sync(rq)]++;
1362 * looks a little odd, but the first insert might return an alias.
1363 * if that happens, put the alias on the dispatch list
1365 while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL)
1366 cfq_dispatch_insert(cfqd->queue, __alias);
1368 if (!cfq_cfqq_on_rr(cfqq))
1369 cfq_add_cfqq_rr(cfqd, cfqq);
1372 * check if this request is a better next-serve candidate
1374 prev = cfqq->next_rq;
1375 cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq, cfqd->last_position);
1378 * adjust priority tree position, if ->next_rq changes
1380 if (prev != cfqq->next_rq)
1381 cfq_prio_tree_add(cfqd, cfqq);
1383 BUG_ON(!cfqq->next_rq);
1386 static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
1388 elv_rb_del(&cfqq->sort_list, rq);
1389 cfqq->queued[rq_is_sync(rq)]--;
1390 blkiocg_update_io_remove_stats(&cfqq->cfqg->blkg, rq_data_dir(rq),
1393 blkiocg_update_io_add_stats(
1394 &cfqq->cfqg->blkg, &cfqq->cfqd->serving_group->blkg,
1395 rq_data_dir(rq), rq_is_sync(rq));
1398 static struct request *
1399 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
1401 struct task_struct *tsk = current;
1402 struct cfq_io_context *cic;
1403 struct cfq_queue *cfqq;
1405 cic = cfq_cic_lookup(cfqd, tsk->io_context);
1409 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
1411 sector_t sector = bio->bi_sector + bio_sectors(bio);
1413 return elv_rb_find(&cfqq->sort_list, sector);
1419 static void cfq_activate_request(struct request_queue *q, struct request *rq)
1421 struct cfq_data *cfqd = q->elevator->elevator_data;
1423 cfqd->rq_in_driver++;
1424 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
1425 cfqd->rq_in_driver);
1427 cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
1430 static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
1432 struct cfq_data *cfqd = q->elevator->elevator_data;
1434 WARN_ON(!cfqd->rq_in_driver);
1435 cfqd->rq_in_driver--;
1436 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
1437 cfqd->rq_in_driver);
1440 static void cfq_remove_request(struct request *rq)
1442 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1444 if (cfqq->next_rq == rq)
1445 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
1447 list_del_init(&rq->queuelist);
1450 cfqq->cfqd->rq_queued--;
1451 blkiocg_update_io_remove_stats(&cfqq->cfqg->blkg, rq_data_dir(rq),
1453 if (rq_is_meta(rq)) {
1454 WARN_ON(!cfqq->meta_pending);
1455 cfqq->meta_pending--;
1459 static int cfq_merge(struct request_queue *q, struct request **req,
1462 struct cfq_data *cfqd = q->elevator->elevator_data;
1463 struct request *__rq;
1465 __rq = cfq_find_rq_fmerge(cfqd, bio);
1466 if (__rq && elv_rq_merge_ok(__rq, bio)) {
1468 return ELEVATOR_FRONT_MERGE;
1471 return ELEVATOR_NO_MERGE;
1474 static void cfq_merged_request(struct request_queue *q, struct request *req,
1477 if (type == ELEVATOR_FRONT_MERGE) {
1478 struct cfq_queue *cfqq = RQ_CFQQ(req);
1480 cfq_reposition_rq_rb(cfqq, req);
1484 static void cfq_bio_merged(struct request_queue *q, struct request *req,
1487 struct cfq_queue *cfqq = RQ_CFQQ(req);
1488 blkiocg_update_io_merged_stats(&cfqq->cfqg->blkg, bio_data_dir(bio),
1493 cfq_merged_requests(struct request_queue *q, struct request *rq,
1494 struct request *next)
1496 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1498 * reposition in fifo if next is older than rq
1500 if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
1501 time_before(rq_fifo_time(next), rq_fifo_time(rq))) {
1502 list_move(&rq->queuelist, &next->queuelist);
1503 rq_set_fifo_time(rq, rq_fifo_time(next));
1506 if (cfqq->next_rq == next)
1508 cfq_remove_request(next);
1509 blkiocg_update_io_merged_stats(&cfqq->cfqg->blkg, rq_data_dir(next),
1513 static int cfq_allow_merge(struct request_queue *q, struct request *rq,
1516 struct cfq_data *cfqd = q->elevator->elevator_data;
1517 struct cfq_io_context *cic;
1518 struct cfq_queue *cfqq;
1521 * Disallow merge of a sync bio into an async request.
1523 if (cfq_bio_sync(bio) && !rq_is_sync(rq))
1527 * Lookup the cfqq that this bio will be queued with. Allow
1528 * merge only if rq is queued there.
1530 cic = cfq_cic_lookup(cfqd, current->io_context);
1534 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
1535 return cfqq == RQ_CFQQ(rq);
1538 static inline void cfq_del_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1540 del_timer(&cfqd->idle_slice_timer);
1541 blkiocg_update_idle_time_stats(&cfqq->cfqg->blkg);
1544 static void __cfq_set_active_queue(struct cfq_data *cfqd,
1545 struct cfq_queue *cfqq)
1548 cfq_log_cfqq(cfqd, cfqq, "set_active wl_prio:%d wl_type:%d",
1549 cfqd->serving_prio, cfqd->serving_type);
1550 blkiocg_update_avg_queue_size_stats(&cfqq->cfqg->blkg);
1551 cfqq->slice_start = 0;
1552 cfqq->dispatch_start = jiffies;
1553 cfqq->allocated_slice = 0;
1554 cfqq->slice_end = 0;
1555 cfqq->slice_dispatch = 0;
1557 cfq_clear_cfqq_wait_request(cfqq);
1558 cfq_clear_cfqq_must_dispatch(cfqq);
1559 cfq_clear_cfqq_must_alloc_slice(cfqq);
1560 cfq_clear_cfqq_fifo_expire(cfqq);
1561 cfq_mark_cfqq_slice_new(cfqq);
1563 cfq_del_timer(cfqd, cfqq);
1566 cfqd->active_queue = cfqq;
1570 * current cfqq expired its slice (or was too idle), select new one
1573 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1574 bool timed_out, bool forced)
1576 cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
1578 if (cfq_cfqq_wait_request(cfqq))
1579 cfq_del_timer(cfqd, cfqq);
1581 cfq_clear_cfqq_wait_request(cfqq);
1582 cfq_clear_cfqq_wait_busy(cfqq);
1585 * If this cfqq is shared between multiple processes, check to
1586 * make sure that those processes are still issuing I/Os within
1587 * the mean seek distance. If not, it may be time to break the
1588 * queues apart again.
1590 if (cfq_cfqq_coop(cfqq) && CFQQ_SEEKY(cfqq))
1591 cfq_mark_cfqq_split_coop(cfqq);
1594 * store what was left of this slice, if the queue idled/timed out
1596 if (timed_out && !cfq_cfqq_slice_new(cfqq)) {
1597 cfqq->slice_resid = cfqq->slice_end - jiffies;
1598 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
1601 cfq_group_served(cfqd, cfqq->cfqg, cfqq, forced);
1603 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
1604 cfq_del_cfqq_rr(cfqd, cfqq);
1606 cfq_resort_rr_list(cfqd, cfqq);
1608 if (cfqq == cfqd->active_queue)
1609 cfqd->active_queue = NULL;
1611 if (&cfqq->cfqg->rb_node == cfqd->grp_service_tree.active)
1612 cfqd->grp_service_tree.active = NULL;
1614 if (cfqd->active_cic) {
1615 put_io_context(cfqd->active_cic->ioc);
1616 cfqd->active_cic = NULL;
1620 static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out,
1623 struct cfq_queue *cfqq = cfqd->active_queue;
1626 __cfq_slice_expired(cfqd, cfqq, timed_out, forced);
1630 * Get next queue for service. Unless we have a queue preemption,
1631 * we'll simply select the first cfqq in the service tree.
1633 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
1635 struct cfq_rb_root *service_tree =
1636 service_tree_for(cfqd->serving_group, cfqd->serving_prio,
1637 cfqd->serving_type);
1639 if (!cfqd->rq_queued)
1642 /* There is nothing to dispatch */
1645 if (RB_EMPTY_ROOT(&service_tree->rb))
1647 return cfq_rb_first(service_tree);
1650 static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd)
1652 struct cfq_group *cfqg;
1653 struct cfq_queue *cfqq;
1655 struct cfq_rb_root *st;
1657 if (!cfqd->rq_queued)
1660 cfqg = cfq_get_next_cfqg(cfqd);
1664 for_each_cfqg_st(cfqg, i, j, st)
1665 if ((cfqq = cfq_rb_first(st)) != NULL)
1671 * Get and set a new active queue for service.
1673 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
1674 struct cfq_queue *cfqq)
1677 cfqq = cfq_get_next_queue(cfqd);
1679 __cfq_set_active_queue(cfqd, cfqq);
1683 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
1686 if (blk_rq_pos(rq) >= cfqd->last_position)
1687 return blk_rq_pos(rq) - cfqd->last_position;
1689 return cfqd->last_position - blk_rq_pos(rq);
1692 static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1695 return cfq_dist_from_last(cfqd, rq) <= CFQQ_CLOSE_THR;
1698 static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
1699 struct cfq_queue *cur_cfqq)
1701 struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
1702 struct rb_node *parent, *node;
1703 struct cfq_queue *__cfqq;
1704 sector_t sector = cfqd->last_position;
1706 if (RB_EMPTY_ROOT(root))
1710 * First, if we find a request starting at the end of the last
1711 * request, choose it.
1713 __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
1718 * If the exact sector wasn't found, the parent of the NULL leaf
1719 * will contain the closest sector.
1721 __cfqq = rb_entry(parent, struct cfq_queue, p_node);
1722 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1725 if (blk_rq_pos(__cfqq->next_rq) < sector)
1726 node = rb_next(&__cfqq->p_node);
1728 node = rb_prev(&__cfqq->p_node);
1732 __cfqq = rb_entry(node, struct cfq_queue, p_node);
1733 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1741 * cur_cfqq - passed in so that we don't decide that the current queue is
1742 * closely cooperating with itself.
1744 * So, basically we're assuming that that cur_cfqq has dispatched at least
1745 * one request, and that cfqd->last_position reflects a position on the disk
1746 * associated with the I/O issued by cur_cfqq. I'm not sure this is a valid
1749 static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
1750 struct cfq_queue *cur_cfqq)
1752 struct cfq_queue *cfqq;
1754 if (cfq_class_idle(cur_cfqq))
1756 if (!cfq_cfqq_sync(cur_cfqq))
1758 if (CFQQ_SEEKY(cur_cfqq))
1762 * Don't search priority tree if it's the only queue in the group.
1764 if (cur_cfqq->cfqg->nr_cfqq == 1)
1768 * We should notice if some of the queues are cooperating, eg
1769 * working closely on the same area of the disk. In that case,
1770 * we can group them together and don't waste time idling.
1772 cfqq = cfqq_close(cfqd, cur_cfqq);
1776 /* If new queue belongs to different cfq_group, don't choose it */
1777 if (cur_cfqq->cfqg != cfqq->cfqg)
1781 * It only makes sense to merge sync queues.
1783 if (!cfq_cfqq_sync(cfqq))
1785 if (CFQQ_SEEKY(cfqq))
1789 * Do not merge queues of different priority classes
1791 if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq))
1798 * Determine whether we should enforce idle window for this queue.
1801 static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1803 enum wl_prio_t prio = cfqq_prio(cfqq);
1804 struct cfq_rb_root *service_tree = cfqq->service_tree;
1806 BUG_ON(!service_tree);
1807 BUG_ON(!service_tree->count);
1809 /* We never do for idle class queues. */
1810 if (prio == IDLE_WORKLOAD)
1813 /* We do for queues that were marked with idle window flag. */
1814 if (cfq_cfqq_idle_window(cfqq) &&
1815 !(blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag))
1819 * Otherwise, we do only if they are the last ones
1820 * in their service tree.
1822 if (service_tree->count == 1 && cfq_cfqq_sync(cfqq))
1824 cfq_log_cfqq(cfqd, cfqq, "Not idling. st->count:%d",
1825 service_tree->count);
1829 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
1831 struct cfq_queue *cfqq = cfqd->active_queue;
1832 struct cfq_io_context *cic;
1836 * SSD device without seek penalty, disable idling. But only do so
1837 * for devices that support queuing, otherwise we still have a problem
1838 * with sync vs async workloads.
1840 if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
1843 WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
1844 WARN_ON(cfq_cfqq_slice_new(cfqq));
1847 * idle is disabled, either manually or by past process history
1849 if (!cfqd->cfq_slice_idle || !cfq_should_idle(cfqd, cfqq))
1853 * still active requests from this queue, don't idle
1855 if (cfqq->dispatched)
1859 * task has exited, don't wait
1861 cic = cfqd->active_cic;
1862 if (!cic || !atomic_read(&cic->ioc->nr_tasks))
1866 * If our average think time is larger than the remaining time
1867 * slice, then don't idle. This avoids overrunning the allotted
1870 if (sample_valid(cic->ttime_samples) &&
1871 (cfqq->slice_end - jiffies < cic->ttime_mean)) {
1872 cfq_log_cfqq(cfqd, cfqq, "Not idling. think_time:%d",
1877 cfq_mark_cfqq_wait_request(cfqq);
1879 sl = cfqd->cfq_slice_idle;
1881 mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
1882 blkiocg_update_set_idle_time_stats(&cfqq->cfqg->blkg);
1883 cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu", sl);
1887 * Move request from internal lists to the request queue dispatch list.
1889 static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
1891 struct cfq_data *cfqd = q->elevator->elevator_data;
1892 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1894 cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
1896 cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
1897 cfq_remove_request(rq);
1899 elv_dispatch_sort(q, rq);
1901 cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]++;
1902 blkiocg_update_dispatch_stats(&cfqq->cfqg->blkg, blk_rq_bytes(rq),
1903 rq_data_dir(rq), rq_is_sync(rq));
1907 * return expired entry, or NULL to just start from scratch in rbtree
1909 static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
1911 struct request *rq = NULL;
1913 if (cfq_cfqq_fifo_expire(cfqq))
1916 cfq_mark_cfqq_fifo_expire(cfqq);
1918 if (list_empty(&cfqq->fifo))
1921 rq = rq_entry_fifo(cfqq->fifo.next);
1922 if (time_before(jiffies, rq_fifo_time(rq)))
1925 cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq);
1930 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1932 const int base_rq = cfqd->cfq_slice_async_rq;
1934 WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
1936 return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
1940 * Must be called with the queue_lock held.
1942 static int cfqq_process_refs(struct cfq_queue *cfqq)
1944 int process_refs, io_refs;
1946 io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE];
1947 process_refs = atomic_read(&cfqq->ref) - io_refs;
1948 BUG_ON(process_refs < 0);
1949 return process_refs;
1952 static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq)
1954 int process_refs, new_process_refs;
1955 struct cfq_queue *__cfqq;
1957 /* Avoid a circular list and skip interim queue merges */
1958 while ((__cfqq = new_cfqq->new_cfqq)) {
1964 process_refs = cfqq_process_refs(cfqq);
1966 * If the process for the cfqq has gone away, there is no
1967 * sense in merging the queues.
1969 if (process_refs == 0)
1973 * Merge in the direction of the lesser amount of work.
1975 new_process_refs = cfqq_process_refs(new_cfqq);
1976 if (new_process_refs >= process_refs) {
1977 cfqq->new_cfqq = new_cfqq;
1978 atomic_add(process_refs, &new_cfqq->ref);
1980 new_cfqq->new_cfqq = cfqq;
1981 atomic_add(new_process_refs, &cfqq->ref);
1985 static enum wl_type_t cfq_choose_wl(struct cfq_data *cfqd,
1986 struct cfq_group *cfqg, enum wl_prio_t prio)
1988 struct cfq_queue *queue;
1990 bool key_valid = false;
1991 unsigned long lowest_key = 0;
1992 enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD;
1994 for (i = 0; i <= SYNC_WORKLOAD; ++i) {
1995 /* select the one with lowest rb_key */
1996 queue = cfq_rb_first(service_tree_for(cfqg, prio, i));
1998 (!key_valid || time_before(queue->rb_key, lowest_key))) {
1999 lowest_key = queue->rb_key;
2008 static void choose_service_tree(struct cfq_data *cfqd, struct cfq_group *cfqg)
2012 struct cfq_rb_root *st;
2013 unsigned group_slice;
2016 cfqd->serving_prio = IDLE_WORKLOAD;
2017 cfqd->workload_expires = jiffies + 1;
2021 /* Choose next priority. RT > BE > IDLE */
2022 if (cfq_group_busy_queues_wl(RT_WORKLOAD, cfqd, cfqg))
2023 cfqd->serving_prio = RT_WORKLOAD;
2024 else if (cfq_group_busy_queues_wl(BE_WORKLOAD, cfqd, cfqg))
2025 cfqd->serving_prio = BE_WORKLOAD;
2027 cfqd->serving_prio = IDLE_WORKLOAD;
2028 cfqd->workload_expires = jiffies + 1;
2033 * For RT and BE, we have to choose also the type
2034 * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
2037 st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
2041 * check workload expiration, and that we still have other queues ready
2043 if (count && !time_after(jiffies, cfqd->workload_expires))
2046 /* otherwise select new workload type */
2047 cfqd->serving_type =
2048 cfq_choose_wl(cfqd, cfqg, cfqd->serving_prio);
2049 st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
2053 * the workload slice is computed as a fraction of target latency
2054 * proportional to the number of queues in that workload, over
2055 * all the queues in the same priority class
2057 group_slice = cfq_group_slice(cfqd, cfqg);
2059 slice = group_slice * count /
2060 max_t(unsigned, cfqg->busy_queues_avg[cfqd->serving_prio],
2061 cfq_group_busy_queues_wl(cfqd->serving_prio, cfqd, cfqg));
2063 if (cfqd->serving_type == ASYNC_WORKLOAD) {
2067 * Async queues are currently system wide. Just taking
2068 * proportion of queues with-in same group will lead to higher
2069 * async ratio system wide as generally root group is going
2070 * to have higher weight. A more accurate thing would be to
2071 * calculate system wide asnc/sync ratio.
2073 tmp = cfq_target_latency * cfqg_busy_async_queues(cfqd, cfqg);
2074 tmp = tmp/cfqd->busy_queues;
2075 slice = min_t(unsigned, slice, tmp);
2077 /* async workload slice is scaled down according to
2078 * the sync/async slice ratio. */
2079 slice = slice * cfqd->cfq_slice[0] / cfqd->cfq_slice[1];
2081 /* sync workload slice is at least 2 * cfq_slice_idle */
2082 slice = max(slice, 2 * cfqd->cfq_slice_idle);
2084 slice = max_t(unsigned, slice, CFQ_MIN_TT);
2085 cfq_log(cfqd, "workload slice:%d", slice);
2086 cfqd->workload_expires = jiffies + slice;
2087 cfqd->noidle_tree_requires_idle = false;
2090 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd)
2092 struct cfq_rb_root *st = &cfqd->grp_service_tree;
2093 struct cfq_group *cfqg;
2095 if (RB_EMPTY_ROOT(&st->rb))
2097 cfqg = cfq_rb_first_group(st);
2098 st->active = &cfqg->rb_node;
2099 update_min_vdisktime(st);
2103 static void cfq_choose_cfqg(struct cfq_data *cfqd)
2105 struct cfq_group *cfqg = cfq_get_next_cfqg(cfqd);
2107 cfqd->serving_group = cfqg;
2109 /* Restore the workload type data */
2110 if (cfqg->saved_workload_slice) {
2111 cfqd->workload_expires = jiffies + cfqg->saved_workload_slice;
2112 cfqd->serving_type = cfqg->saved_workload;
2113 cfqd->serving_prio = cfqg->saved_serving_prio;
2115 cfqd->workload_expires = jiffies - 1;
2117 choose_service_tree(cfqd, cfqg);
2121 * Select a queue for service. If we have a current active queue,
2122 * check whether to continue servicing it, or retrieve and set a new one.
2124 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
2126 struct cfq_queue *cfqq, *new_cfqq = NULL;
2128 cfqq = cfqd->active_queue;
2132 if (!cfqd->rq_queued)
2136 * We were waiting for group to get backlogged. Expire the queue
2138 if (cfq_cfqq_wait_busy(cfqq) && !RB_EMPTY_ROOT(&cfqq->sort_list))
2142 * The active queue has run out of time, expire it and select new.
2144 if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq)) {
2146 * If slice had not expired at the completion of last request
2147 * we might not have turned on wait_busy flag. Don't expire
2148 * the queue yet. Allow the group to get backlogged.
2150 * The very fact that we have used the slice, that means we
2151 * have been idling all along on this queue and it should be
2152 * ok to wait for this request to complete.
2154 if (cfqq->cfqg->nr_cfqq == 1 && RB_EMPTY_ROOT(&cfqq->sort_list)
2155 && cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2163 * The active queue has requests and isn't expired, allow it to
2166 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
2170 * If another queue has a request waiting within our mean seek
2171 * distance, let it run. The expire code will check for close
2172 * cooperators and put the close queue at the front of the service
2173 * tree. If possible, merge the expiring queue with the new cfqq.
2175 new_cfqq = cfq_close_cooperator(cfqd, cfqq);
2177 if (!cfqq->new_cfqq)
2178 cfq_setup_merge(cfqq, new_cfqq);
2183 * No requests pending. If the active queue still has requests in
2184 * flight or is idling for a new request, allow either of these
2185 * conditions to happen (or time out) before selecting a new queue.
2187 if (timer_pending(&cfqd->idle_slice_timer) ||
2188 (cfqq->dispatched && cfq_should_idle(cfqd, cfqq))) {
2194 cfq_slice_expired(cfqd, 0, false);
2197 * Current queue expired. Check if we have to switch to a new
2201 cfq_choose_cfqg(cfqd);
2203 cfqq = cfq_set_active_queue(cfqd, new_cfqq);
2208 static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
2212 while (cfqq->next_rq) {
2213 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
2217 BUG_ON(!list_empty(&cfqq->fifo));
2219 /* By default cfqq is not expired if it is empty. Do it explicitly */
2220 __cfq_slice_expired(cfqq->cfqd, cfqq, 0, true);
2225 * Drain our current requests. Used for barriers and when switching
2226 * io schedulers on-the-fly.
2228 static int cfq_forced_dispatch(struct cfq_data *cfqd)
2230 struct cfq_queue *cfqq;
2233 /* Expire the timeslice of the current active queue first */
2234 cfq_slice_expired(cfqd, 0);
2235 while ((cfqq = cfq_get_next_queue_forced(cfqd)) != NULL) {
2236 __cfq_set_active_queue(cfqd, cfqq);
2237 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
2240 BUG_ON(cfqd->busy_queues);
2242 cfq_log(cfqd, "forced_dispatch=%d", dispatched);
2246 static inline bool cfq_slice_used_soon(struct cfq_data *cfqd,
2247 struct cfq_queue *cfqq)
2249 /* the queue hasn't finished any request, can't estimate */
2250 if (cfq_cfqq_slice_new(cfqq))
2252 if (time_after(jiffies + cfqd->cfq_slice_idle * cfqq->dispatched,
2259 static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2261 unsigned int max_dispatch;
2264 * Drain async requests before we start sync IO
2266 if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_flight[BLK_RW_ASYNC])
2270 * If this is an async queue and we have sync IO in flight, let it wait
2272 if (cfqd->rq_in_flight[BLK_RW_SYNC] && !cfq_cfqq_sync(cfqq))
2275 max_dispatch = max_t(unsigned int, cfqd->cfq_quantum / 2, 1);
2276 if (cfq_class_idle(cfqq))
2280 * Does this cfqq already have too much IO in flight?
2282 if (cfqq->dispatched >= max_dispatch) {
2284 * idle queue must always only have a single IO in flight
2286 if (cfq_class_idle(cfqq))
2290 * We have other queues, don't allow more IO from this one
2292 if (cfqd->busy_queues > 1 && cfq_slice_used_soon(cfqd, cfqq))
2296 * Sole queue user, no limit
2298 if (cfqd->busy_queues == 1)
2302 * Normally we start throttling cfqq when cfq_quantum/2
2303 * requests have been dispatched. But we can drive
2304 * deeper queue depths at the beginning of slice
2305 * subjected to upper limit of cfq_quantum.
2307 max_dispatch = cfqd->cfq_quantum;
2311 * Async queues must wait a bit before being allowed dispatch.
2312 * We also ramp up the dispatch depth gradually for async IO,
2313 * based on the last sync IO we serviced
2315 if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
2316 unsigned long last_sync = jiffies - cfqd->last_delayed_sync;
2319 depth = last_sync / cfqd->cfq_slice[1];
2320 if (!depth && !cfqq->dispatched)
2322 if (depth < max_dispatch)
2323 max_dispatch = depth;
2327 * If we're below the current max, allow a dispatch
2329 return cfqq->dispatched < max_dispatch;
2333 * Dispatch a request from cfqq, moving them to the request queue
2336 static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2340 BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
2342 if (!cfq_may_dispatch(cfqd, cfqq))
2346 * follow expired path, else get first next available
2348 rq = cfq_check_fifo(cfqq);
2353 * insert request into driver dispatch list
2355 cfq_dispatch_insert(cfqd->queue, rq);
2357 if (!cfqd->active_cic) {
2358 struct cfq_io_context *cic = RQ_CIC(rq);
2360 atomic_long_inc(&cic->ioc->refcount);
2361 cfqd->active_cic = cic;
2368 * Find the cfqq that we need to service and move a request from that to the
2371 static int cfq_dispatch_requests(struct request_queue *q, int force)
2373 struct cfq_data *cfqd = q->elevator->elevator_data;
2374 struct cfq_queue *cfqq;
2376 if (!cfqd->busy_queues)
2379 if (unlikely(force))
2380 return cfq_forced_dispatch(cfqd);
2382 cfqq = cfq_select_queue(cfqd);
2387 * Dispatch a request from this cfqq, if it is allowed
2389 if (!cfq_dispatch_request(cfqd, cfqq))
2392 cfqq->slice_dispatch++;
2393 cfq_clear_cfqq_must_dispatch(cfqq);
2396 * expire an async queue immediately if it has used up its slice. idle
2397 * queue always expire after 1 dispatch round.
2399 if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
2400 cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
2401 cfq_class_idle(cfqq))) {
2402 cfqq->slice_end = jiffies + 1;
2403 cfq_slice_expired(cfqd, 0, false);
2406 cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
2411 * task holds one reference to the queue, dropped when task exits. each rq
2412 * in-flight on this queue also holds a reference, dropped when rq is freed.
2414 * Each cfq queue took a reference on the parent group. Drop it now.
2415 * queue lock must be held here.
2417 static void cfq_put_queue(struct cfq_queue *cfqq)
2419 struct cfq_data *cfqd = cfqq->cfqd;
2420 struct cfq_group *cfqg, *orig_cfqg;
2422 BUG_ON(atomic_read(&cfqq->ref) <= 0);
2424 if (!atomic_dec_and_test(&cfqq->ref))
2427 cfq_log_cfqq(cfqd, cfqq, "put_queue");
2428 BUG_ON(rb_first(&cfqq->sort_list));
2429 BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
2431 orig_cfqg = cfqq->orig_cfqg;
2433 if (unlikely(cfqd->active_queue == cfqq)) {
2434 __cfq_slice_expired(cfqd, cfqq, 0, false);
2435 cfq_schedule_dispatch(cfqd);
2438 BUG_ON(cfq_cfqq_on_rr(cfqq));
2439 kmem_cache_free(cfq_pool, cfqq);
2442 cfq_put_cfqg(orig_cfqg);
2446 * Must always be called with the rcu_read_lock() held
2449 __call_for_each_cic(struct io_context *ioc,
2450 void (*func)(struct io_context *, struct cfq_io_context *))
2452 struct cfq_io_context *cic;
2453 struct hlist_node *n;
2455 hlist_for_each_entry_rcu(cic, n, &ioc->cic_list, cic_list)
2460 * Call func for each cic attached to this ioc.
2463 call_for_each_cic(struct io_context *ioc,
2464 void (*func)(struct io_context *, struct cfq_io_context *))
2467 __call_for_each_cic(ioc, func);
2471 static void cfq_cic_free_rcu(struct rcu_head *head)
2473 struct cfq_io_context *cic;
2475 cic = container_of(head, struct cfq_io_context, rcu_head);
2477 kmem_cache_free(cfq_ioc_pool, cic);
2478 elv_ioc_count_dec(cfq_ioc_count);
2482 * CFQ scheduler is exiting, grab exit lock and check
2483 * the pending io context count. If it hits zero,
2484 * complete ioc_gone and set it back to NULL
2486 spin_lock(&ioc_gone_lock);
2487 if (ioc_gone && !elv_ioc_count_read(cfq_ioc_count)) {
2491 spin_unlock(&ioc_gone_lock);
2495 static void cfq_cic_free(struct cfq_io_context *cic)
2497 call_rcu(&cic->rcu_head, cfq_cic_free_rcu);
2500 static void cic_free_func(struct io_context *ioc, struct cfq_io_context *cic)
2502 unsigned long flags;
2504 BUG_ON(!cic->dead_key);
2506 spin_lock_irqsave(&ioc->lock, flags);
2507 radix_tree_delete(&ioc->radix_root, cic->dead_key);
2508 hlist_del_rcu(&cic->cic_list);
2509 spin_unlock_irqrestore(&ioc->lock, flags);
2515 * Must be called with rcu_read_lock() held or preemption otherwise disabled.
2516 * Only two callers of this - ->dtor() which is called with the rcu_read_lock(),
2517 * and ->trim() which is called with the task lock held
2519 static void cfq_free_io_context(struct io_context *ioc)
2522 * ioc->refcount is zero here, or we are called from elv_unregister(),
2523 * so no more cic's are allowed to be linked into this ioc. So it
2524 * should be ok to iterate over the known list, we will see all cic's
2525 * since no new ones are added.
2527 __call_for_each_cic(ioc, cic_free_func);
2530 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2532 struct cfq_queue *__cfqq, *next;
2534 if (unlikely(cfqq == cfqd->active_queue)) {
2535 __cfq_slice_expired(cfqd, cfqq, 0, false);
2536 cfq_schedule_dispatch(cfqd);
2540 * If this queue was scheduled to merge with another queue, be
2541 * sure to drop the reference taken on that queue (and others in
2542 * the merge chain). See cfq_setup_merge and cfq_merge_cfqqs.
2544 __cfqq = cfqq->new_cfqq;
2546 if (__cfqq == cfqq) {
2547 WARN(1, "cfqq->new_cfqq loop detected\n");
2550 next = __cfqq->new_cfqq;
2551 cfq_put_queue(__cfqq);
2555 cfq_put_queue(cfqq);
2558 static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
2559 struct cfq_io_context *cic)
2561 struct io_context *ioc = cic->ioc;
2563 list_del_init(&cic->queue_list);
2566 * Make sure key == NULL is seen for dead queues
2569 cic->dead_key = (unsigned long) cic->key;
2572 if (ioc->ioc_data == cic)
2573 rcu_assign_pointer(ioc->ioc_data, NULL);
2575 if (cic->cfqq[BLK_RW_ASYNC]) {
2576 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]);
2577 cic->cfqq[BLK_RW_ASYNC] = NULL;
2580 if (cic->cfqq[BLK_RW_SYNC]) {
2581 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]);
2582 cic->cfqq[BLK_RW_SYNC] = NULL;
2586 static void cfq_exit_single_io_context(struct io_context *ioc,
2587 struct cfq_io_context *cic)
2589 struct cfq_data *cfqd = cic->key;
2592 struct request_queue *q = cfqd->queue;
2593 unsigned long flags;
2595 spin_lock_irqsave(q->queue_lock, flags);
2598 * Ensure we get a fresh copy of the ->key to prevent
2599 * race between exiting task and queue
2601 smp_read_barrier_depends();
2603 __cfq_exit_single_io_context(cfqd, cic);
2605 spin_unlock_irqrestore(q->queue_lock, flags);
2610 * The process that ioc belongs to has exited, we need to clean up
2611 * and put the internal structures we have that belongs to that process.
2613 static void cfq_exit_io_context(struct io_context *ioc)
2615 call_for_each_cic(ioc, cfq_exit_single_io_context);
2618 static struct cfq_io_context *
2619 cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
2621 struct cfq_io_context *cic;
2623 cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO,
2626 cic->last_end_request = jiffies;
2627 INIT_LIST_HEAD(&cic->queue_list);
2628 INIT_HLIST_NODE(&cic->cic_list);
2629 cic->dtor = cfq_free_io_context;
2630 cic->exit = cfq_exit_io_context;
2631 elv_ioc_count_inc(cfq_ioc_count);
2637 static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc)
2639 struct task_struct *tsk = current;
2642 if (!cfq_cfqq_prio_changed(cfqq))
2645 ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio);
2646 switch (ioprio_class) {
2648 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
2649 case IOPRIO_CLASS_NONE:
2651 * no prio set, inherit CPU scheduling settings
2653 cfqq->ioprio = task_nice_ioprio(tsk);
2654 cfqq->ioprio_class = task_nice_ioclass(tsk);
2656 case IOPRIO_CLASS_RT:
2657 cfqq->ioprio = task_ioprio(ioc);
2658 cfqq->ioprio_class = IOPRIO_CLASS_RT;
2660 case IOPRIO_CLASS_BE:
2661 cfqq->ioprio = task_ioprio(ioc);
2662 cfqq->ioprio_class = IOPRIO_CLASS_BE;
2664 case IOPRIO_CLASS_IDLE:
2665 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
2667 cfq_clear_cfqq_idle_window(cfqq);
2672 * keep track of original prio settings in case we have to temporarily
2673 * elevate the priority of this queue
2675 cfqq->org_ioprio = cfqq->ioprio;
2676 cfqq->org_ioprio_class = cfqq->ioprio_class;
2677 cfq_clear_cfqq_prio_changed(cfqq);
2680 static void changed_ioprio(struct io_context *ioc, struct cfq_io_context *cic)
2682 struct cfq_data *cfqd = cic->key;
2683 struct cfq_queue *cfqq;
2684 unsigned long flags;
2686 if (unlikely(!cfqd))
2689 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2691 cfqq = cic->cfqq[BLK_RW_ASYNC];
2693 struct cfq_queue *new_cfqq;
2694 new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic->ioc,
2697 cic->cfqq[BLK_RW_ASYNC] = new_cfqq;
2698 cfq_put_queue(cfqq);
2702 cfqq = cic->cfqq[BLK_RW_SYNC];
2704 cfq_mark_cfqq_prio_changed(cfqq);
2706 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2709 static void cfq_ioc_set_ioprio(struct io_context *ioc)
2711 call_for_each_cic(ioc, changed_ioprio);
2712 ioc->ioprio_changed = 0;
2715 static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2716 pid_t pid, bool is_sync)
2718 RB_CLEAR_NODE(&cfqq->rb_node);
2719 RB_CLEAR_NODE(&cfqq->p_node);
2720 INIT_LIST_HEAD(&cfqq->fifo);
2722 atomic_set(&cfqq->ref, 0);
2725 cfq_mark_cfqq_prio_changed(cfqq);
2728 if (!cfq_class_idle(cfqq))
2729 cfq_mark_cfqq_idle_window(cfqq);
2730 cfq_mark_cfqq_sync(cfqq);
2735 #ifdef CONFIG_CFQ_GROUP_IOSCHED
2736 static void changed_cgroup(struct io_context *ioc, struct cfq_io_context *cic)
2738 struct cfq_queue *sync_cfqq = cic_to_cfqq(cic, 1);
2739 struct cfq_data *cfqd = cic->key;
2740 unsigned long flags;
2741 struct request_queue *q;
2743 if (unlikely(!cfqd))
2748 spin_lock_irqsave(q->queue_lock, flags);
2752 * Drop reference to sync queue. A new sync queue will be
2753 * assigned in new group upon arrival of a fresh request.
2755 cfq_log_cfqq(cfqd, sync_cfqq, "changed cgroup");
2756 cic_set_cfqq(cic, NULL, 1);
2757 cfq_put_queue(sync_cfqq);
2760 spin_unlock_irqrestore(q->queue_lock, flags);
2763 static void cfq_ioc_set_cgroup(struct io_context *ioc)
2765 call_for_each_cic(ioc, changed_cgroup);
2766 ioc->cgroup_changed = 0;
2768 #endif /* CONFIG_CFQ_GROUP_IOSCHED */
2770 static struct cfq_queue *
2771 cfq_find_alloc_queue(struct cfq_data *cfqd, bool is_sync,
2772 struct io_context *ioc, gfp_t gfp_mask)
2774 struct cfq_queue *cfqq, *new_cfqq = NULL;
2775 struct cfq_io_context *cic;
2776 struct cfq_group *cfqg;
2779 cfqg = cfq_get_cfqg(cfqd, 1);
2780 cic = cfq_cic_lookup(cfqd, ioc);
2781 /* cic always exists here */
2782 cfqq = cic_to_cfqq(cic, is_sync);
2785 * Always try a new alloc if we fell back to the OOM cfqq
2786 * originally, since it should just be a temporary situation.
2788 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
2793 } else if (gfp_mask & __GFP_WAIT) {
2794 spin_unlock_irq(cfqd->queue->queue_lock);
2795 new_cfqq = kmem_cache_alloc_node(cfq_pool,
2796 gfp_mask | __GFP_ZERO,
2798 spin_lock_irq(cfqd->queue->queue_lock);
2802 cfqq = kmem_cache_alloc_node(cfq_pool,
2803 gfp_mask | __GFP_ZERO,
2808 cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
2809 cfq_init_prio_data(cfqq, ioc);
2810 cfq_link_cfqq_cfqg(cfqq, cfqg);
2811 cfq_log_cfqq(cfqd, cfqq, "alloced");
2813 cfqq = &cfqd->oom_cfqq;
2817 kmem_cache_free(cfq_pool, new_cfqq);
2822 static struct cfq_queue **
2823 cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
2825 switch (ioprio_class) {
2826 case IOPRIO_CLASS_RT:
2827 return &cfqd->async_cfqq[0][ioprio];
2828 case IOPRIO_CLASS_BE:
2829 return &cfqd->async_cfqq[1][ioprio];
2830 case IOPRIO_CLASS_IDLE:
2831 return &cfqd->async_idle_cfqq;
2837 static struct cfq_queue *
2838 cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct io_context *ioc,
2841 const int ioprio = task_ioprio(ioc);
2842 const int ioprio_class = task_ioprio_class(ioc);
2843 struct cfq_queue **async_cfqq = NULL;
2844 struct cfq_queue *cfqq = NULL;
2847 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
2852 cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask);
2855 * pin the queue now that it's allocated, scheduler exit will prune it
2857 if (!is_sync && !(*async_cfqq)) {
2858 atomic_inc(&cfqq->ref);
2862 atomic_inc(&cfqq->ref);
2867 * We drop cfq io contexts lazily, so we may find a dead one.
2870 cfq_drop_dead_cic(struct cfq_data *cfqd, struct io_context *ioc,
2871 struct cfq_io_context *cic)
2873 unsigned long flags;
2875 WARN_ON(!list_empty(&cic->queue_list));
2877 spin_lock_irqsave(&ioc->lock, flags);
2879 BUG_ON(ioc->ioc_data == cic);
2881 radix_tree_delete(&ioc->radix_root, (unsigned long) cfqd);
2882 hlist_del_rcu(&cic->cic_list);
2883 spin_unlock_irqrestore(&ioc->lock, flags);
2888 static struct cfq_io_context *
2889 cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc)
2891 struct cfq_io_context *cic;
2892 unsigned long flags;
2901 * we maintain a last-hit cache, to avoid browsing over the tree
2903 cic = rcu_dereference(ioc->ioc_data);
2904 if (cic && cic->key == cfqd) {
2910 cic = radix_tree_lookup(&ioc->radix_root, (unsigned long) cfqd);
2914 /* ->key must be copied to avoid race with cfq_exit_queue() */
2917 cfq_drop_dead_cic(cfqd, ioc, cic);
2922 spin_lock_irqsave(&ioc->lock, flags);
2923 rcu_assign_pointer(ioc->ioc_data, cic);
2924 spin_unlock_irqrestore(&ioc->lock, flags);
2932 * Add cic into ioc, using cfqd as the search key. This enables us to lookup
2933 * the process specific cfq io context when entered from the block layer.
2934 * Also adds the cic to a per-cfqd list, used when this queue is removed.
2936 static int cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
2937 struct cfq_io_context *cic, gfp_t gfp_mask)
2939 unsigned long flags;
2942 ret = radix_tree_preload(gfp_mask);
2947 spin_lock_irqsave(&ioc->lock, flags);
2948 ret = radix_tree_insert(&ioc->radix_root,
2949 (unsigned long) cfqd, cic);
2951 hlist_add_head_rcu(&cic->cic_list, &ioc->cic_list);
2952 spin_unlock_irqrestore(&ioc->lock, flags);
2954 radix_tree_preload_end();
2957 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2958 list_add(&cic->queue_list, &cfqd->cic_list);
2959 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2964 printk(KERN_ERR "cfq: cic link failed!\n");
2970 * Setup general io context and cfq io context. There can be several cfq
2971 * io contexts per general io context, if this process is doing io to more
2972 * than one device managed by cfq.
2974 static struct cfq_io_context *
2975 cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
2977 struct io_context *ioc = NULL;
2978 struct cfq_io_context *cic;
2980 might_sleep_if(gfp_mask & __GFP_WAIT);
2982 ioc = get_io_context(gfp_mask, cfqd->queue->node);
2986 cic = cfq_cic_lookup(cfqd, ioc);
2990 cic = cfq_alloc_io_context(cfqd, gfp_mask);
2994 if (cfq_cic_link(cfqd, ioc, cic, gfp_mask))
2998 smp_read_barrier_depends();
2999 if (unlikely(ioc->ioprio_changed))
3000 cfq_ioc_set_ioprio(ioc);
3002 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3003 if (unlikely(ioc->cgroup_changed))
3004 cfq_ioc_set_cgroup(ioc);
3010 put_io_context(ioc);
3015 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
3017 unsigned long elapsed = jiffies - cic->last_end_request;
3018 unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
3020 cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
3021 cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
3022 cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
3026 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3030 sector_t n_sec = blk_rq_sectors(rq);
3031 if (cfqq->last_request_pos) {
3032 if (cfqq->last_request_pos < blk_rq_pos(rq))
3033 sdist = blk_rq_pos(rq) - cfqq->last_request_pos;
3035 sdist = cfqq->last_request_pos - blk_rq_pos(rq);
3038 cfqq->seek_history <<= 1;
3039 if (blk_queue_nonrot(cfqd->queue))
3040 cfqq->seek_history |= (n_sec < CFQQ_SECT_THR_NONROT);
3042 cfqq->seek_history |= (sdist > CFQQ_SEEK_THR);
3046 * Disable idle window if the process thinks too long or seeks so much that
3050 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3051 struct cfq_io_context *cic)
3053 int old_idle, enable_idle;
3056 * Don't idle for async or idle io prio class
3058 if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
3061 enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
3063 if (cfqq->queued[0] + cfqq->queued[1] >= 4)
3064 cfq_mark_cfqq_deep(cfqq);
3066 if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle ||
3067 (!cfq_cfqq_deep(cfqq) && CFQQ_SEEKY(cfqq)))
3069 else if (sample_valid(cic->ttime_samples)) {
3070 if (cic->ttime_mean > cfqd->cfq_slice_idle)
3076 if (old_idle != enable_idle) {
3077 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
3079 cfq_mark_cfqq_idle_window(cfqq);
3081 cfq_clear_cfqq_idle_window(cfqq);
3086 * Check if new_cfqq should preempt the currently active queue. Return 0 for
3087 * no or if we aren't sure, a 1 will cause a preempt.
3090 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
3093 struct cfq_queue *cfqq;
3095 cfqq = cfqd->active_queue;
3099 if (cfq_class_idle(new_cfqq))
3102 if (cfq_class_idle(cfqq))
3106 * Don't allow a non-RT request to preempt an ongoing RT cfqq timeslice.
3108 if (cfq_class_rt(cfqq) && !cfq_class_rt(new_cfqq))
3112 * if the new request is sync, but the currently running queue is
3113 * not, let the sync request have priority.
3115 if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
3118 if (new_cfqq->cfqg != cfqq->cfqg)
3121 if (cfq_slice_used(cfqq))
3124 /* Allow preemption only if we are idling on sync-noidle tree */
3125 if (cfqd->serving_type == SYNC_NOIDLE_WORKLOAD &&
3126 cfqq_type(new_cfqq) == SYNC_NOIDLE_WORKLOAD &&
3127 new_cfqq->service_tree->count == 2 &&
3128 RB_EMPTY_ROOT(&cfqq->sort_list))
3132 * So both queues are sync. Let the new request get disk time if
3133 * it's a metadata request and the current queue is doing regular IO.
3135 if (rq_is_meta(rq) && !cfqq->meta_pending)
3139 * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
3141 if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
3144 if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
3148 * if this request is as-good as one we would expect from the
3149 * current cfqq, let it preempt
3151 if (cfq_rq_close(cfqd, cfqq, rq))
3158 * cfqq preempts the active queue. if we allowed preempt with no slice left,
3159 * let it have half of its nominal slice.
3161 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3163 cfq_log_cfqq(cfqd, cfqq, "preempt");
3164 cfq_slice_expired(cfqd, 1, false);
3167 * Put the new queue at the front of the of the current list,
3168 * so we know that it will be selected next.
3170 BUG_ON(!cfq_cfqq_on_rr(cfqq));
3172 cfq_service_tree_add(cfqd, cfqq, 1);
3174 cfqq->slice_end = 0;
3175 cfq_mark_cfqq_slice_new(cfqq);
3179 * Called when a new fs request (rq) is added (to cfqq). Check if there's
3180 * something we should do about it
3183 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3186 struct cfq_io_context *cic = RQ_CIC(rq);
3190 cfqq->meta_pending++;
3192 cfq_update_io_thinktime(cfqd, cic);
3193 cfq_update_io_seektime(cfqd, cfqq, rq);
3194 cfq_update_idle_window(cfqd, cfqq, cic);
3196 cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
3198 if (cfqq == cfqd->active_queue) {
3200 * Remember that we saw a request from this process, but
3201 * don't start queuing just yet. Otherwise we risk seeing lots
3202 * of tiny requests, because we disrupt the normal plugging
3203 * and merging. If the request is already larger than a single
3204 * page, let it rip immediately. For that case we assume that
3205 * merging is already done. Ditto for a busy system that
3206 * has other work pending, don't risk delaying until the
3207 * idle timer unplug to continue working.
3209 if (cfq_cfqq_wait_request(cfqq)) {
3210 if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
3211 cfqd->busy_queues > 1) {
3212 cfq_del_timer(cfqd, cfqq);
3213 cfq_clear_cfqq_wait_request(cfqq);
3214 __blk_run_queue(cfqd->queue);
3216 blkiocg_update_idle_time_stats(
3218 cfq_mark_cfqq_must_dispatch(cfqq);
3221 } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
3223 * not the active queue - expire current slice if it is
3224 * idle and has expired it's mean thinktime or this new queue
3225 * has some old slice time left and is of higher priority or
3226 * this new queue is RT and the current one is BE
3228 cfq_preempt_queue(cfqd, cfqq);
3229 __blk_run_queue(cfqd->queue);
3233 static void cfq_insert_request(struct request_queue *q, struct request *rq)
3235 struct cfq_data *cfqd = q->elevator->elevator_data;
3236 struct cfq_queue *cfqq = RQ_CFQQ(rq);
3238 cfq_log_cfqq(cfqd, cfqq, "insert_request");
3239 cfq_init_prio_data(cfqq, RQ_CIC(rq)->ioc);
3241 rq_set_fifo_time(rq, jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)]);
3242 list_add_tail(&rq->queuelist, &cfqq->fifo);
3245 blkiocg_update_io_add_stats(&cfqq->cfqg->blkg,
3246 &cfqd->serving_group->blkg, rq_data_dir(rq),
3248 cfq_rq_enqueued(cfqd, cfqq, rq);
3252 * Update hw_tag based on peak queue depth over 50 samples under
3255 static void cfq_update_hw_tag(struct cfq_data *cfqd)
3257 struct cfq_queue *cfqq = cfqd->active_queue;
3259 if (cfqd->rq_in_driver > cfqd->hw_tag_est_depth)
3260 cfqd->hw_tag_est_depth = cfqd->rq_in_driver;
3262 if (cfqd->hw_tag == 1)
3265 if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
3266 cfqd->rq_in_driver <= CFQ_HW_QUEUE_MIN)
3270 * If active queue hasn't enough requests and can idle, cfq might not
3271 * dispatch sufficient requests to hardware. Don't zero hw_tag in this
3274 if (cfqq && cfq_cfqq_idle_window(cfqq) &&
3275 cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] <
3276 CFQ_HW_QUEUE_MIN && cfqd->rq_in_driver < CFQ_HW_QUEUE_MIN)
3279 if (cfqd->hw_tag_samples++ < 50)
3282 if (cfqd->hw_tag_est_depth >= CFQ_HW_QUEUE_MIN)
3288 static bool cfq_should_wait_busy(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3290 struct cfq_io_context *cic = cfqd->active_cic;
3292 /* If there are other queues in the group, don't wait */
3293 if (cfqq->cfqg->nr_cfqq > 1)
3296 if (cfq_slice_used(cfqq))
3299 /* if slice left is less than think time, wait busy */
3300 if (cic && sample_valid(cic->ttime_samples)
3301 && (cfqq->slice_end - jiffies < cic->ttime_mean))
3305 * If think times is less than a jiffy than ttime_mean=0 and above
3306 * will not be true. It might happen that slice has not expired yet
3307 * but will expire soon (4-5 ns) during select_queue(). To cover the
3308 * case where think time is less than a jiffy, mark the queue wait
3309 * busy if only 1 jiffy is left in the slice.
3311 if (cfqq->slice_end - jiffies == 1)
3317 static void cfq_completed_request(struct request_queue *q, struct request *rq)
3319 struct cfq_queue *cfqq = RQ_CFQQ(rq);
3320 struct cfq_data *cfqd = cfqq->cfqd;
3321 const int sync = rq_is_sync(rq);
3325 cfq_log_cfqq(cfqd, cfqq, "complete rqnoidle %d", !!rq_noidle(rq));
3327 cfq_update_hw_tag(cfqd);
3329 WARN_ON(!cfqd->rq_in_driver);
3330 WARN_ON(!cfqq->dispatched);
3331 cfqd->rq_in_driver--;
3333 blkiocg_update_completion_stats(&cfqq->cfqg->blkg, rq_start_time_ns(rq),
3334 rq_io_start_time_ns(rq), rq_data_dir(rq),
3337 cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]--;
3340 RQ_CIC(rq)->last_end_request = now;
3341 if (!time_after(rq->start_time + cfqd->cfq_fifo_expire[1], now))
3342 cfqd->last_delayed_sync = now;
3346 * If this is the active queue, check if it needs to be expired,
3347 * or if we want to idle in case it has no pending requests.
3349 if (cfqd->active_queue == cfqq) {
3350 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
3352 if (cfq_cfqq_slice_new(cfqq)) {
3353 cfq_set_prio_slice(cfqd, cfqq);
3354 cfq_clear_cfqq_slice_new(cfqq);
3358 * Should we wait for next request to come in before we expire
3361 if (cfq_should_wait_busy(cfqd, cfqq)) {
3362 cfqq->slice_end = jiffies + cfqd->cfq_slice_idle;
3363 cfq_mark_cfqq_wait_busy(cfqq);
3364 cfq_log_cfqq(cfqd, cfqq, "will busy wait");
3368 * Idling is not enabled on:
3370 * - idle-priority queues
3372 * - queues with still some requests queued
3373 * - when there is a close cooperator
3375 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
3376 cfq_slice_expired(cfqd, 1, false);
3377 else if (sync && cfqq_empty &&
3378 !cfq_close_cooperator(cfqd, cfqq)) {
3379 cfqd->noidle_tree_requires_idle |= !rq_noidle(rq);
3381 * Idling is enabled for SYNC_WORKLOAD.
3382 * SYNC_NOIDLE_WORKLOAD idles at the end of the tree
3383 * only if we processed at least one !rq_noidle request
3385 if (cfqd->serving_type == SYNC_WORKLOAD
3386 || cfqd->noidle_tree_requires_idle
3387 || cfqq->cfqg->nr_cfqq == 1)
3388 cfq_arm_slice_timer(cfqd);
3392 if (!cfqd->rq_in_driver)
3393 cfq_schedule_dispatch(cfqd);
3397 * we temporarily boost lower priority queues if they are holding fs exclusive
3398 * resources. they are boosted to normal prio (CLASS_BE/4)
3400 static void cfq_prio_boost(struct cfq_queue *cfqq)
3402 if (has_fs_excl()) {
3404 * boost idle prio on transactions that would lock out other
3405 * users of the filesystem
3407 if (cfq_class_idle(cfqq))
3408 cfqq->ioprio_class = IOPRIO_CLASS_BE;
3409 if (cfqq->ioprio > IOPRIO_NORM)
3410 cfqq->ioprio = IOPRIO_NORM;
3413 * unboost the queue (if needed)
3415 cfqq->ioprio_class = cfqq->org_ioprio_class;
3416 cfqq->ioprio = cfqq->org_ioprio;
3420 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
3422 if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
3423 cfq_mark_cfqq_must_alloc_slice(cfqq);
3424 return ELV_MQUEUE_MUST;
3427 return ELV_MQUEUE_MAY;
3430 static int cfq_may_queue(struct request_queue *q, int rw)
3432 struct cfq_data *cfqd = q->elevator->elevator_data;
3433 struct task_struct *tsk = current;
3434 struct cfq_io_context *cic;
3435 struct cfq_queue *cfqq;
3438 * don't force setup of a queue from here, as a call to may_queue
3439 * does not necessarily imply that a request actually will be queued.
3440 * so just lookup a possibly existing queue, or return 'may queue'
3443 cic = cfq_cic_lookup(cfqd, tsk->io_context);
3445 return ELV_MQUEUE_MAY;
3447 cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
3449 cfq_init_prio_data(cfqq, cic->ioc);
3450 cfq_prio_boost(cfqq);
3452 return __cfq_may_queue(cfqq);
3455 return ELV_MQUEUE_MAY;
3459 * queue lock held here
3461 static void cfq_put_request(struct request *rq)
3463 struct cfq_queue *cfqq = RQ_CFQQ(rq);
3466 const int rw = rq_data_dir(rq);
3468 BUG_ON(!cfqq->allocated[rw]);
3469 cfqq->allocated[rw]--;
3471 put_io_context(RQ_CIC(rq)->ioc);
3473 rq->elevator_private = NULL;
3474 rq->elevator_private2 = NULL;
3476 cfq_put_queue(cfqq);
3480 static struct cfq_queue *
3481 cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_context *cic,
3482 struct cfq_queue *cfqq)
3484 cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq);
3485 cic_set_cfqq(cic, cfqq->new_cfqq, 1);
3486 cfq_mark_cfqq_coop(cfqq->new_cfqq);
3487 cfq_put_queue(cfqq);
3488 return cic_to_cfqq(cic, 1);
3492 * Returns NULL if a new cfqq should be allocated, or the old cfqq if this
3493 * was the last process referring to said cfqq.
3495 static struct cfq_queue *
3496 split_cfqq(struct cfq_io_context *cic, struct cfq_queue *cfqq)
3498 if (cfqq_process_refs(cfqq) == 1) {
3499 cfqq->pid = current->pid;
3500 cfq_clear_cfqq_coop(cfqq);
3501 cfq_clear_cfqq_split_coop(cfqq);
3505 cic_set_cfqq(cic, NULL, 1);
3506 cfq_put_queue(cfqq);
3510 * Allocate cfq data structures associated with this request.
3513 cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
3515 struct cfq_data *cfqd = q->elevator->elevator_data;
3516 struct cfq_io_context *cic;
3517 const int rw = rq_data_dir(rq);
3518 const bool is_sync = rq_is_sync(rq);
3519 struct cfq_queue *cfqq;
3520 unsigned long flags;
3522 might_sleep_if(gfp_mask & __GFP_WAIT);
3524 cic = cfq_get_io_context(cfqd, gfp_mask);
3526 spin_lock_irqsave(q->queue_lock, flags);
3532 cfqq = cic_to_cfqq(cic, is_sync);
3533 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
3534 cfqq = cfq_get_queue(cfqd, is_sync, cic->ioc, gfp_mask);
3535 cic_set_cfqq(cic, cfqq, is_sync);
3538 * If the queue was seeky for too long, break it apart.
3540 if (cfq_cfqq_coop(cfqq) && cfq_cfqq_split_coop(cfqq)) {
3541 cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq");
3542 cfqq = split_cfqq(cic, cfqq);
3548 * Check to see if this queue is scheduled to merge with
3549 * another, closely cooperating queue. The merging of
3550 * queues happens here as it must be done in process context.
3551 * The reference on new_cfqq was taken in merge_cfqqs.
3554 cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq);
3557 cfqq->allocated[rw]++;
3558 atomic_inc(&cfqq->ref);
3560 spin_unlock_irqrestore(q->queue_lock, flags);
3562 rq->elevator_private = cic;
3563 rq->elevator_private2 = cfqq;
3568 put_io_context(cic->ioc);
3570 cfq_schedule_dispatch(cfqd);
3571 spin_unlock_irqrestore(q->queue_lock, flags);
3572 cfq_log(cfqd, "set_request fail");
3576 static void cfq_kick_queue(struct work_struct *work)
3578 struct cfq_data *cfqd =
3579 container_of(work, struct cfq_data, unplug_work);
3580 struct request_queue *q = cfqd->queue;
3582 spin_lock_irq(q->queue_lock);
3583 __blk_run_queue(cfqd->queue);
3584 spin_unlock_irq(q->queue_lock);
3588 * Timer running if the active_queue is currently idling inside its time slice
3590 static void cfq_idle_slice_timer(unsigned long data)
3592 struct cfq_data *cfqd = (struct cfq_data *) data;
3593 struct cfq_queue *cfqq;
3594 unsigned long flags;
3597 cfq_log(cfqd, "idle timer fired");
3599 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
3601 cfqq = cfqd->active_queue;
3606 * We saw a request before the queue expired, let it through
3608 if (cfq_cfqq_must_dispatch(cfqq))
3614 if (cfq_slice_used(cfqq))
3618 * only expire and reinvoke request handler, if there are
3619 * other queues with pending requests
3621 if (!cfqd->busy_queues)
3625 * not expired and it has a request pending, let it dispatch
3627 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3631 * Queue depth flag is reset only when the idle didn't succeed
3633 cfq_clear_cfqq_deep(cfqq);
3636 cfq_slice_expired(cfqd, timed_out, false);
3638 cfq_schedule_dispatch(cfqd);
3640 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
3643 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
3645 del_timer_sync(&cfqd->idle_slice_timer);
3646 cancel_work_sync(&cfqd->unplug_work);
3649 static void cfq_put_async_queues(struct cfq_data *cfqd)
3653 for (i = 0; i < IOPRIO_BE_NR; i++) {
3654 if (cfqd->async_cfqq[0][i])
3655 cfq_put_queue(cfqd->async_cfqq[0][i]);
3656 if (cfqd->async_cfqq[1][i])
3657 cfq_put_queue(cfqd->async_cfqq[1][i]);
3660 if (cfqd->async_idle_cfqq)
3661 cfq_put_queue(cfqd->async_idle_cfqq);
3664 static void cfq_cfqd_free(struct rcu_head *head)
3666 kfree(container_of(head, struct cfq_data, rcu));
3669 static void cfq_exit_queue(struct elevator_queue *e)
3671 struct cfq_data *cfqd = e->elevator_data;
3672 struct request_queue *q = cfqd->queue;
3674 cfq_shutdown_timer_wq(cfqd);
3676 spin_lock_irq(q->queue_lock);
3678 if (cfqd->active_queue)
3679 __cfq_slice_expired(cfqd, cfqd->active_queue, 0, false);
3681 while (!list_empty(&cfqd->cic_list)) {
3682 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
3683 struct cfq_io_context,
3686 __cfq_exit_single_io_context(cfqd, cic);
3689 cfq_put_async_queues(cfqd);
3690 cfq_release_cfq_groups(cfqd);
3691 blkiocg_del_blkio_group(&cfqd->root_group.blkg);
3693 spin_unlock_irq(q->queue_lock);
3695 cfq_shutdown_timer_wq(cfqd);
3697 /* Wait for cfqg->blkg->key accessors to exit their grace periods. */
3698 call_rcu(&cfqd->rcu, cfq_cfqd_free);
3701 static void *cfq_init_queue(struct request_queue *q)
3703 struct cfq_data *cfqd;
3705 struct cfq_group *cfqg;
3706 struct cfq_rb_root *st;
3708 cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
3712 /* Init root service tree */
3713 cfqd->grp_service_tree = CFQ_RB_ROOT;
3715 /* Init root group */
3716 cfqg = &cfqd->root_group;
3717 for_each_cfqg_st(cfqg, i, j, st)
3719 RB_CLEAR_NODE(&cfqg->rb_node);
3721 /* Give preference to root group over other groups */
3722 cfqg->weight = 2*BLKIO_WEIGHT_DEFAULT;
3724 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3726 * Take a reference to root group which we never drop. This is just
3727 * to make sure that cfq_put_cfqg() does not try to kfree root group
3729 atomic_set(&cfqg->ref, 1);
3730 blkiocg_add_blkio_group(&blkio_root_cgroup, &cfqg->blkg, (void *)cfqd,
3734 * Not strictly needed (since RB_ROOT just clears the node and we
3735 * zeroed cfqd on alloc), but better be safe in case someone decides
3736 * to add magic to the rb code
3738 for (i = 0; i < CFQ_PRIO_LISTS; i++)
3739 cfqd->prio_trees[i] = RB_ROOT;
3742 * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
3743 * Grab a permanent reference to it, so that the normal code flow
3744 * will not attempt to free it.
3746 cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
3747 atomic_inc(&cfqd->oom_cfqq.ref);
3748 cfq_link_cfqq_cfqg(&cfqd->oom_cfqq, &cfqd->root_group);
3750 INIT_LIST_HEAD(&cfqd->cic_list);
3754 init_timer(&cfqd->idle_slice_timer);
3755 cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
3756 cfqd->idle_slice_timer.data = (unsigned long) cfqd;
3758 INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
3760 cfqd->cfq_quantum = cfq_quantum;
3761 cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
3762 cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
3763 cfqd->cfq_back_max = cfq_back_max;
3764 cfqd->cfq_back_penalty = cfq_back_penalty;
3765 cfqd->cfq_slice[0] = cfq_slice_async;
3766 cfqd->cfq_slice[1] = cfq_slice_sync;
3767 cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
3768 cfqd->cfq_slice_idle = cfq_slice_idle;
3769 cfqd->cfq_latency = 1;
3770 cfqd->cfq_group_isolation = 0;
3773 * we optimistically start assuming sync ops weren't delayed in last
3774 * second, in order to have larger depth for async operations.
3776 cfqd->last_delayed_sync = jiffies - HZ;
3777 INIT_RCU_HEAD(&cfqd->rcu);
3781 static void cfq_slab_kill(void)
3784 * Caller already ensured that pending RCU callbacks are completed,
3785 * so we should have no busy allocations at this point.
3788 kmem_cache_destroy(cfq_pool);
3790 kmem_cache_destroy(cfq_ioc_pool);
3793 static int __init cfq_slab_setup(void)
3795 cfq_pool = KMEM_CACHE(cfq_queue, 0);
3799 cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0);
3810 * sysfs parts below -->
3813 cfq_var_show(unsigned int var, char *page)
3815 return sprintf(page, "%d\n", var);
3819 cfq_var_store(unsigned int *var, const char *page, size_t count)
3821 char *p = (char *) page;
3823 *var = simple_strtoul(p, &p, 10);
3827 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
3828 static ssize_t __FUNC(struct elevator_queue *e, char *page) \
3830 struct cfq_data *cfqd = e->elevator_data; \
3831 unsigned int __data = __VAR; \
3833 __data = jiffies_to_msecs(__data); \
3834 return cfq_var_show(__data, (page)); \
3836 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
3837 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
3838 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
3839 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
3840 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
3841 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
3842 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
3843 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
3844 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
3845 SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
3846 SHOW_FUNCTION(cfq_group_isolation_show, cfqd->cfq_group_isolation, 0);
3847 #undef SHOW_FUNCTION
3849 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
3850 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
3852 struct cfq_data *cfqd = e->elevator_data; \
3853 unsigned int __data; \
3854 int ret = cfq_var_store(&__data, (page), count); \
3855 if (__data < (MIN)) \
3857 else if (__data > (MAX)) \
3860 *(__PTR) = msecs_to_jiffies(__data); \
3862 *(__PTR) = __data; \
3865 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
3866 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
3868 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
3870 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
3871 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
3873 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
3874 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
3875 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
3876 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
3878 STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
3879 STORE_FUNCTION(cfq_group_isolation_store, &cfqd->cfq_group_isolation, 0, 1, 0);
3880 #undef STORE_FUNCTION
3882 #define CFQ_ATTR(name) \
3883 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
3885 static struct elv_fs_entry cfq_attrs[] = {
3887 CFQ_ATTR(fifo_expire_sync),
3888 CFQ_ATTR(fifo_expire_async),
3889 CFQ_ATTR(back_seek_max),
3890 CFQ_ATTR(back_seek_penalty),
3891 CFQ_ATTR(slice_sync),
3892 CFQ_ATTR(slice_async),
3893 CFQ_ATTR(slice_async_rq),
3894 CFQ_ATTR(slice_idle),
3895 CFQ_ATTR(low_latency),
3896 CFQ_ATTR(group_isolation),
3900 static struct elevator_type iosched_cfq = {
3902 .elevator_merge_fn = cfq_merge,
3903 .elevator_merged_fn = cfq_merged_request,
3904 .elevator_merge_req_fn = cfq_merged_requests,
3905 .elevator_allow_merge_fn = cfq_allow_merge,
3906 .elevator_bio_merged_fn = cfq_bio_merged,
3907 .elevator_dispatch_fn = cfq_dispatch_requests,
3908 .elevator_add_req_fn = cfq_insert_request,
3909 .elevator_activate_req_fn = cfq_activate_request,
3910 .elevator_deactivate_req_fn = cfq_deactivate_request,
3911 .elevator_queue_empty_fn = cfq_queue_empty,
3912 .elevator_completed_req_fn = cfq_completed_request,
3913 .elevator_former_req_fn = elv_rb_former_request,
3914 .elevator_latter_req_fn = elv_rb_latter_request,
3915 .elevator_set_req_fn = cfq_set_request,
3916 .elevator_put_req_fn = cfq_put_request,
3917 .elevator_may_queue_fn = cfq_may_queue,
3918 .elevator_init_fn = cfq_init_queue,
3919 .elevator_exit_fn = cfq_exit_queue,
3920 .trim = cfq_free_io_context,
3922 .elevator_attrs = cfq_attrs,
3923 .elevator_name = "cfq",
3924 .elevator_owner = THIS_MODULE,
3927 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3928 static struct blkio_policy_type blkio_policy_cfq = {
3930 .blkio_unlink_group_fn = cfq_unlink_blkio_group,
3931 .blkio_update_group_weight_fn = cfq_update_blkio_group_weight,
3935 static struct blkio_policy_type blkio_policy_cfq;
3938 static int __init cfq_init(void)
3941 * could be 0 on HZ < 1000 setups
3943 if (!cfq_slice_async)
3944 cfq_slice_async = 1;
3945 if (!cfq_slice_idle)
3948 if (cfq_slab_setup())
3951 elv_register(&iosched_cfq);
3952 blkio_policy_register(&blkio_policy_cfq);
3957 static void __exit cfq_exit(void)
3959 DECLARE_COMPLETION_ONSTACK(all_gone);
3960 blkio_policy_unregister(&blkio_policy_cfq);
3961 elv_unregister(&iosched_cfq);
3962 ioc_gone = &all_gone;
3963 /* ioc_gone's update must be visible before reading ioc_count */
3967 * this also protects us from entering cfq_slab_kill() with
3968 * pending RCU callbacks
3970 if (elv_ioc_count_read(cfq_ioc_count))
3971 wait_for_completion(&all_gone);
3975 module_init(cfq_init);
3976 module_exit(cfq_exit);
3978 MODULE_AUTHOR("Jens Axboe");
3979 MODULE_LICENSE("GPL");
3980 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");