2 * CFQ, or complete fairness queueing, disk scheduler.
4 * Based on ideas from a previously unfinished io
5 * scheduler (round robin per-process disk scheduling) and Andrea Arcangeli.
7 * Copyright (C) 2003 Jens Axboe <axboe@kernel.dk>
9 #include <linux/module.h>
10 #include <linux/blkdev.h>
11 #include <linux/elevator.h>
12 #include <linux/jiffies.h>
13 #include <linux/rbtree.h>
14 #include <linux/ioprio.h>
15 #include <linux/blktrace_api.h>
16 #include "blk-cgroup.h"
21 /* max queue in one round of service */
22 static const int cfq_quantum = 4;
23 static const int cfq_fifo_expire[2] = { HZ / 4, HZ / 8 };
24 /* maximum backwards seek, in KiB */
25 static const int cfq_back_max = 16 * 1024;
26 /* penalty of a backwards seek */
27 static const int cfq_back_penalty = 2;
28 static const int cfq_slice_sync = HZ / 10;
29 static int cfq_slice_async = HZ / 25;
30 static const int cfq_slice_async_rq = 2;
31 static int cfq_slice_idle = HZ / 125;
32 static const int cfq_target_latency = HZ * 3/10; /* 300 ms */
33 static const int cfq_hist_divisor = 4;
36 * offset from end of service tree
38 #define CFQ_IDLE_DELAY (HZ / 5)
41 * below this threshold, we consider thinktime immediate
43 #define CFQ_MIN_TT (2)
46 * Allow merged cfqqs to perform this amount of seeky I/O before
47 * deciding to break the queues up again.
49 #define CFQQ_COOP_TOUT (HZ)
51 #define CFQ_SLICE_SCALE (5)
52 #define CFQ_HW_QUEUE_MIN (5)
53 #define CFQ_SERVICE_SHIFT 12
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.
84 struct rb_node *active;
85 unsigned total_weight;
87 #define CFQ_RB_ROOT (struct cfq_rb_root) { RB_ROOT, NULL, 0, 0, }
90 * Per process-grouping structure
95 /* various state flags, see below */
98 struct cfq_data *cfqd;
99 /* service_tree member */
100 struct rb_node rb_node;
101 /* service_tree key */
102 unsigned long rb_key;
103 /* prio tree member */
104 struct rb_node p_node;
105 /* prio tree root we belong to, if any */
106 struct rb_root *p_root;
107 /* sorted list of pending requests */
108 struct rb_root sort_list;
109 /* if fifo isn't expired, next request to serve */
110 struct request *next_rq;
111 /* requests queued in sort_list */
113 /* currently allocated requests */
115 /* fifo list of requests in sort_list */
116 struct list_head fifo;
118 /* time when queue got scheduled in to dispatch first request. */
119 unsigned long dispatch_start;
120 /* time when first request from queue completed and slice started. */
121 unsigned long slice_start;
122 unsigned long slice_end;
124 unsigned int slice_dispatch;
126 /* pending metadata requests */
128 /* number of requests that are on the dispatch list or inside driver */
131 /* io prio of this group */
132 unsigned short ioprio, org_ioprio;
133 unsigned short ioprio_class, org_ioprio_class;
135 unsigned int seek_samples;
138 sector_t last_request_pos;
139 unsigned long seeky_start;
143 struct cfq_rb_root *service_tree;
144 struct cfq_queue *new_cfqq;
145 struct cfq_group *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;
195 * Per block device queue structure
198 struct request_queue *queue;
199 /* Root service tree for cfq_groups */
200 struct cfq_rb_root grp_service_tree;
201 struct cfq_group root_group;
202 /* Number of active cfq groups on group service tree */
206 * The priority currently being served
208 enum wl_prio_t serving_prio;
209 enum wl_type_t serving_type;
210 unsigned long workload_expires;
211 struct cfq_group *serving_group;
212 bool noidle_tree_requires_idle;
215 * Each priority tree is sorted by next_request position. These
216 * trees are used when determining if two or more queues are
217 * interleaving requests (see cfq_close_cooperator).
219 struct rb_root prio_trees[CFQ_PRIO_LISTS];
221 unsigned int busy_queues;
227 * queue-depth detection
233 * -1 => indeterminate, (cfq will behave as if NCQ is present, to allow better detection)
234 * 1 => NCQ is present (hw_tag_est_depth is the estimated max depth)
237 int hw_tag_est_depth;
238 unsigned int hw_tag_samples;
241 * idle window management
243 struct timer_list idle_slice_timer;
244 struct work_struct unplug_work;
246 struct cfq_queue *active_queue;
247 struct cfq_io_context *active_cic;
250 * async queue for each priority case
252 struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
253 struct cfq_queue *async_idle_cfqq;
255 sector_t last_position;
258 * tunables, see top of file
260 unsigned int cfq_quantum;
261 unsigned int cfq_fifo_expire[2];
262 unsigned int cfq_back_penalty;
263 unsigned int cfq_back_max;
264 unsigned int cfq_slice[2];
265 unsigned int cfq_slice_async_rq;
266 unsigned int cfq_slice_idle;
267 unsigned int cfq_latency;
269 struct list_head cic_list;
272 * Fallback dummy cfqq for extreme OOM conditions
274 struct cfq_queue oom_cfqq;
276 unsigned long last_end_sync_rq;
279 static struct cfq_rb_root *service_tree_for(struct cfq_group *cfqg,
282 struct cfq_data *cfqd)
287 if (prio == IDLE_WORKLOAD)
288 return &cfqg->service_tree_idle;
290 return &cfqg->service_trees[prio][type];
293 enum cfqq_state_flags {
294 CFQ_CFQQ_FLAG_on_rr = 0, /* on round-robin busy list */
295 CFQ_CFQQ_FLAG_wait_request, /* waiting for a request */
296 CFQ_CFQQ_FLAG_must_dispatch, /* must be allowed a dispatch */
297 CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
298 CFQ_CFQQ_FLAG_fifo_expire, /* FIFO checked in this slice */
299 CFQ_CFQQ_FLAG_idle_window, /* slice idling enabled */
300 CFQ_CFQQ_FLAG_prio_changed, /* task priority has changed */
301 CFQ_CFQQ_FLAG_slice_new, /* no requests dispatched in slice */
302 CFQ_CFQQ_FLAG_sync, /* synchronous queue */
303 CFQ_CFQQ_FLAG_coop, /* cfqq is shared */
304 CFQ_CFQQ_FLAG_deep, /* sync cfqq experienced large depth */
307 #define CFQ_CFQQ_FNS(name) \
308 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \
310 (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name); \
312 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \
314 (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \
316 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \
318 return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \
322 CFQ_CFQQ_FNS(wait_request);
323 CFQ_CFQQ_FNS(must_dispatch);
324 CFQ_CFQQ_FNS(must_alloc_slice);
325 CFQ_CFQQ_FNS(fifo_expire);
326 CFQ_CFQQ_FNS(idle_window);
327 CFQ_CFQQ_FNS(prio_changed);
328 CFQ_CFQQ_FNS(slice_new);
334 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
335 blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args)
336 #define cfq_log(cfqd, fmt, args...) \
337 blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
339 /* Traverses through cfq group service trees */
340 #define for_each_cfqg_st(cfqg, i, j, st) \
341 for (i = 0; i <= IDLE_WORKLOAD; i++) \
342 for (j = 0, st = i < IDLE_WORKLOAD ? &cfqg->service_trees[i][j]\
343 : &cfqg->service_tree_idle; \
344 (i < IDLE_WORKLOAD && j <= SYNC_WORKLOAD) || \
345 (i == IDLE_WORKLOAD && j == 0); \
346 j++, st = i < IDLE_WORKLOAD ? \
347 &cfqg->service_trees[i][j]: NULL) \
350 static inline enum wl_prio_t cfqq_prio(struct cfq_queue *cfqq)
352 if (cfq_class_idle(cfqq))
353 return IDLE_WORKLOAD;
354 if (cfq_class_rt(cfqq))
360 static enum wl_type_t cfqq_type(struct cfq_queue *cfqq)
362 if (!cfq_cfqq_sync(cfqq))
363 return ASYNC_WORKLOAD;
364 if (!cfq_cfqq_idle_window(cfqq))
365 return SYNC_NOIDLE_WORKLOAD;
366 return SYNC_WORKLOAD;
369 static inline int cfq_group_busy_queues_wl(enum wl_prio_t wl,
370 struct cfq_data *cfqd,
371 struct cfq_group *cfqg)
373 if (wl == IDLE_WORKLOAD)
374 return cfqg->service_tree_idle.count;
376 return cfqg->service_trees[wl][ASYNC_WORKLOAD].count
377 + cfqg->service_trees[wl][SYNC_NOIDLE_WORKLOAD].count
378 + cfqg->service_trees[wl][SYNC_WORKLOAD].count;
381 static void cfq_dispatch_insert(struct request_queue *, struct request *);
382 static struct cfq_queue *cfq_get_queue(struct cfq_data *, bool,
383 struct io_context *, gfp_t);
384 static struct cfq_io_context *cfq_cic_lookup(struct cfq_data *,
385 struct io_context *);
387 static inline int rq_in_driver(struct cfq_data *cfqd)
389 return cfqd->rq_in_driver[0] + cfqd->rq_in_driver[1];
392 static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_context *cic,
395 return cic->cfqq[is_sync];
398 static inline void cic_set_cfqq(struct cfq_io_context *cic,
399 struct cfq_queue *cfqq, bool is_sync)
401 cic->cfqq[is_sync] = cfqq;
405 * We regard a request as SYNC, if it's either a read or has the SYNC bit
406 * set (in which case it could also be direct WRITE).
408 static inline bool cfq_bio_sync(struct bio *bio)
410 return bio_data_dir(bio) == READ || bio_rw_flagged(bio, BIO_RW_SYNCIO);
414 * scheduler run of queue, if there are requests pending and no one in the
415 * driver that will restart queueing
417 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
419 if (cfqd->busy_queues) {
420 cfq_log(cfqd, "schedule dispatch");
421 kblockd_schedule_work(cfqd->queue, &cfqd->unplug_work);
425 static int cfq_queue_empty(struct request_queue *q)
427 struct cfq_data *cfqd = q->elevator->elevator_data;
429 return !cfqd->rq_queued;
433 * Scale schedule slice based on io priority. Use the sync time slice only
434 * if a queue is marked sync and has sync io queued. A sync queue with async
435 * io only, should not get full sync slice length.
437 static inline int cfq_prio_slice(struct cfq_data *cfqd, bool sync,
440 const int base_slice = cfqd->cfq_slice[sync];
442 WARN_ON(prio >= IOPRIO_BE_NR);
444 return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
448 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
450 return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
453 static inline u64 cfq_scale_slice(unsigned long delta, struct cfq_group *cfqg)
455 u64 d = delta << CFQ_SERVICE_SHIFT;
457 d = d * BLKIO_WEIGHT_DEFAULT;
458 do_div(d, cfqg->weight);
462 static inline u64 max_vdisktime(u64 min_vdisktime, u64 vdisktime)
464 s64 delta = (s64)(vdisktime - min_vdisktime);
466 min_vdisktime = vdisktime;
468 return min_vdisktime;
471 static inline u64 min_vdisktime(u64 min_vdisktime, u64 vdisktime)
473 s64 delta = (s64)(vdisktime - min_vdisktime);
475 min_vdisktime = vdisktime;
477 return min_vdisktime;
480 static void update_min_vdisktime(struct cfq_rb_root *st)
482 u64 vdisktime = st->min_vdisktime;
483 struct cfq_group *cfqg;
486 cfqg = rb_entry_cfqg(st->active);
487 vdisktime = cfqg->vdisktime;
491 cfqg = rb_entry_cfqg(st->left);
492 vdisktime = min_vdisktime(vdisktime, cfqg->vdisktime);
495 st->min_vdisktime = max_vdisktime(st->min_vdisktime, vdisktime);
499 * get averaged number of queues of RT/BE priority.
500 * average is updated, with a formula that gives more weight to higher numbers,
501 * to quickly follows sudden increases and decrease slowly
504 static inline unsigned cfq_group_get_avg_queues(struct cfq_data *cfqd,
505 struct cfq_group *cfqg, bool rt)
507 unsigned min_q, max_q;
508 unsigned mult = cfq_hist_divisor - 1;
509 unsigned round = cfq_hist_divisor / 2;
510 unsigned busy = cfq_group_busy_queues_wl(rt, cfqd, cfqg);
512 min_q = min(cfqg->busy_queues_avg[rt], busy);
513 max_q = max(cfqg->busy_queues_avg[rt], busy);
514 cfqg->busy_queues_avg[rt] = (mult * max_q + min_q + round) /
516 return cfqg->busy_queues_avg[rt];
519 static inline unsigned
520 cfq_group_slice(struct cfq_data *cfqd, struct cfq_group *cfqg)
522 struct cfq_rb_root *st = &cfqd->grp_service_tree;
524 return cfq_target_latency * cfqg->weight / st->total_weight;
528 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
530 unsigned slice = cfq_prio_to_slice(cfqd, cfqq);
531 if (cfqd->cfq_latency) {
533 * interested queues (we consider only the ones with the same
534 * priority class in the cfq group)
536 unsigned iq = cfq_group_get_avg_queues(cfqd, cfqq->cfqg,
538 unsigned sync_slice = cfqd->cfq_slice[1];
539 unsigned expect_latency = sync_slice * iq;
540 unsigned group_slice = cfq_group_slice(cfqd, cfqq->cfqg);
542 if (expect_latency > group_slice) {
543 unsigned base_low_slice = 2 * cfqd->cfq_slice_idle;
544 /* scale low_slice according to IO priority
545 * and sync vs async */
547 min(slice, base_low_slice * slice / sync_slice);
548 /* the adapted slice value is scaled to fit all iqs
549 * into the target latency */
550 slice = max(slice * group_slice / expect_latency,
554 cfqq->slice_start = jiffies;
555 cfqq->slice_end = jiffies + slice;
556 cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
560 * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
561 * isn't valid until the first request from the dispatch is activated
562 * and the slice time set.
564 static inline bool cfq_slice_used(struct cfq_queue *cfqq)
566 if (cfq_cfqq_slice_new(cfqq))
568 if (time_before(jiffies, cfqq->slice_end))
575 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
576 * We choose the request that is closest to the head right now. Distance
577 * behind the head is penalized and only allowed to a certain extent.
579 static struct request *
580 cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2, sector_t last)
582 sector_t s1, s2, d1 = 0, d2 = 0;
583 unsigned long back_max;
584 #define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */
585 #define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */
586 unsigned wrap = 0; /* bit mask: requests behind the disk head? */
588 if (rq1 == NULL || rq1 == rq2)
593 if (rq_is_sync(rq1) && !rq_is_sync(rq2))
595 else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
597 if (rq_is_meta(rq1) && !rq_is_meta(rq2))
599 else if (rq_is_meta(rq2) && !rq_is_meta(rq1))
602 s1 = blk_rq_pos(rq1);
603 s2 = blk_rq_pos(rq2);
606 * by definition, 1KiB is 2 sectors
608 back_max = cfqd->cfq_back_max * 2;
611 * Strict one way elevator _except_ in the case where we allow
612 * short backward seeks which are biased as twice the cost of a
613 * similar forward seek.
617 else if (s1 + back_max >= last)
618 d1 = (last - s1) * cfqd->cfq_back_penalty;
620 wrap |= CFQ_RQ1_WRAP;
624 else if (s2 + back_max >= last)
625 d2 = (last - s2) * cfqd->cfq_back_penalty;
627 wrap |= CFQ_RQ2_WRAP;
629 /* Found required data */
632 * By doing switch() on the bit mask "wrap" we avoid having to
633 * check two variables for all permutations: --> faster!
636 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
652 case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
655 * Since both rqs are wrapped,
656 * start with the one that's further behind head
657 * (--> only *one* back seek required),
658 * since back seek takes more time than forward.
668 * The below is leftmost cache rbtree addon
670 static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
672 /* Service tree is empty */
677 root->left = rb_first(&root->rb);
680 return rb_entry(root->left, struct cfq_queue, rb_node);
685 static struct cfq_group *cfq_rb_first_group(struct cfq_rb_root *root)
688 root->left = rb_first(&root->rb);
691 return rb_entry_cfqg(root->left);
696 static void rb_erase_init(struct rb_node *n, struct rb_root *root)
702 static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
706 rb_erase_init(n, &root->rb);
711 * would be nice to take fifo expire time into account as well
713 static struct request *
714 cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
715 struct request *last)
717 struct rb_node *rbnext = rb_next(&last->rb_node);
718 struct rb_node *rbprev = rb_prev(&last->rb_node);
719 struct request *next = NULL, *prev = NULL;
721 BUG_ON(RB_EMPTY_NODE(&last->rb_node));
724 prev = rb_entry_rq(rbprev);
727 next = rb_entry_rq(rbnext);
729 rbnext = rb_first(&cfqq->sort_list);
730 if (rbnext && rbnext != &last->rb_node)
731 next = rb_entry_rq(rbnext);
734 return cfq_choose_req(cfqd, next, prev, blk_rq_pos(last));
737 static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
738 struct cfq_queue *cfqq)
741 * just an approximation, should be ok.
743 return (cfqq->cfqg->nr_cfqq - 1) * (cfq_prio_slice(cfqd, 1, 0) -
744 cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
748 cfqg_key(struct cfq_rb_root *st, struct cfq_group *cfqg)
750 return cfqg->vdisktime - st->min_vdisktime;
754 __cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
756 struct rb_node **node = &st->rb.rb_node;
757 struct rb_node *parent = NULL;
758 struct cfq_group *__cfqg;
759 s64 key = cfqg_key(st, cfqg);
762 while (*node != NULL) {
764 __cfqg = rb_entry_cfqg(parent);
766 if (key < cfqg_key(st, __cfqg))
767 node = &parent->rb_left;
769 node = &parent->rb_right;
775 st->left = &cfqg->rb_node;
777 rb_link_node(&cfqg->rb_node, parent, node);
778 rb_insert_color(&cfqg->rb_node, &st->rb);
782 cfq_group_service_tree_add(struct cfq_data *cfqd, struct cfq_group *cfqg)
784 struct cfq_rb_root *st = &cfqd->grp_service_tree;
785 struct cfq_group *__cfqg;
793 * Currently put the group at the end. Later implement something
794 * so that groups get lesser vtime based on their weights, so that
795 * if group does not loose all if it was not continously backlogged.
797 n = rb_last(&st->rb);
799 __cfqg = rb_entry_cfqg(n);
800 cfqg->vdisktime = __cfqg->vdisktime + CFQ_IDLE_DELAY;
802 cfqg->vdisktime = st->min_vdisktime;
804 __cfq_group_service_tree_add(st, cfqg);
807 st->total_weight += cfqg->weight;
811 cfq_group_service_tree_del(struct cfq_data *cfqd, struct cfq_group *cfqg)
813 struct cfq_rb_root *st = &cfqd->grp_service_tree;
815 if (st->active == &cfqg->rb_node)
818 BUG_ON(cfqg->nr_cfqq < 1);
821 /* If there are other cfq queues under this group, don't delete it */
827 st->total_weight -= cfqg->weight;
828 if (!RB_EMPTY_NODE(&cfqg->rb_node))
829 cfq_rb_erase(&cfqg->rb_node, st);
830 cfqg->saved_workload_slice = 0;
833 static inline unsigned int cfq_cfqq_slice_usage(struct cfq_queue *cfqq)
835 unsigned int slice_used, allocated_slice;
838 * Queue got expired before even a single request completed or
839 * got expired immediately after first request completion.
841 if (!cfqq->slice_start || cfqq->slice_start == jiffies) {
843 * Also charge the seek time incurred to the group, otherwise
844 * if there are mutiple queues in the group, each can dispatch
845 * a single request on seeky media and cause lots of seek time
846 * and group will never know it.
848 slice_used = max_t(unsigned, (jiffies - cfqq->dispatch_start),
851 slice_used = jiffies - cfqq->slice_start;
852 allocated_slice = cfqq->slice_end - cfqq->slice_start;
853 if (slice_used > allocated_slice)
854 slice_used = allocated_slice;
857 cfq_log_cfqq(cfqq->cfqd, cfqq, "sl_used=%u", slice_used);
861 static void cfq_group_served(struct cfq_data *cfqd, struct cfq_group *cfqg,
862 struct cfq_queue *cfqq)
864 struct cfq_rb_root *st = &cfqd->grp_service_tree;
865 unsigned int used_sl;
867 used_sl = cfq_cfqq_slice_usage(cfqq);
869 /* Can't update vdisktime while group is on service tree */
870 cfq_rb_erase(&cfqg->rb_node, st);
871 cfqg->vdisktime += cfq_scale_slice(used_sl, cfqg);
872 __cfq_group_service_tree_add(st, cfqg);
874 /* This group is being expired. Save the context */
875 if (time_after(cfqd->workload_expires, jiffies)) {
876 cfqg->saved_workload_slice = cfqd->workload_expires
878 cfqg->saved_workload = cfqd->serving_type;
879 cfqg->saved_serving_prio = cfqd->serving_prio;
881 cfqg->saved_workload_slice = 0;
885 * The cfqd->service_trees holds all pending cfq_queue's that have
886 * requests waiting to be processed. It is sorted in the order that
887 * we will service the queues.
889 static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
892 struct rb_node **p, *parent;
893 struct cfq_queue *__cfqq;
894 unsigned long rb_key;
895 struct cfq_rb_root *service_tree;
899 service_tree = service_tree_for(cfqq->cfqg, cfqq_prio(cfqq),
900 cfqq_type(cfqq), cfqd);
901 if (cfq_class_idle(cfqq)) {
902 rb_key = CFQ_IDLE_DELAY;
903 parent = rb_last(&service_tree->rb);
904 if (parent && parent != &cfqq->rb_node) {
905 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
906 rb_key += __cfqq->rb_key;
909 } else if (!add_front) {
911 * Get our rb key offset. Subtract any residual slice
912 * value carried from last service. A negative resid
913 * count indicates slice overrun, and this should position
914 * the next service time further away in the tree.
916 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
917 rb_key -= cfqq->slice_resid;
918 cfqq->slice_resid = 0;
921 __cfqq = cfq_rb_first(service_tree);
922 rb_key += __cfqq ? __cfqq->rb_key : jiffies;
925 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
928 * same position, nothing more to do
930 if (rb_key == cfqq->rb_key &&
931 cfqq->service_tree == service_tree)
934 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
935 cfqq->service_tree = NULL;
940 cfqq->service_tree = service_tree;
941 p = &service_tree->rb.rb_node;
946 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
949 * sort by key, that represents service time.
951 if (time_before(rb_key, __cfqq->rb_key))
962 service_tree->left = &cfqq->rb_node;
964 cfqq->rb_key = rb_key;
965 rb_link_node(&cfqq->rb_node, parent, p);
966 rb_insert_color(&cfqq->rb_node, &service_tree->rb);
967 service_tree->count++;
968 if (add_front || !new_cfqq)
970 cfq_group_service_tree_add(cfqd, cfqq->cfqg);
973 static struct cfq_queue *
974 cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
975 sector_t sector, struct rb_node **ret_parent,
976 struct rb_node ***rb_link)
978 struct rb_node **p, *parent;
979 struct cfq_queue *cfqq = NULL;
987 cfqq = rb_entry(parent, struct cfq_queue, p_node);
990 * Sort strictly based on sector. Smallest to the left,
991 * largest to the right.
993 if (sector > blk_rq_pos(cfqq->next_rq))
995 else if (sector < blk_rq_pos(cfqq->next_rq))
1003 *ret_parent = parent;
1009 static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1011 struct rb_node **p, *parent;
1012 struct cfq_queue *__cfqq;
1015 rb_erase(&cfqq->p_node, cfqq->p_root);
1016 cfqq->p_root = NULL;
1019 if (cfq_class_idle(cfqq))
1024 cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
1025 __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
1026 blk_rq_pos(cfqq->next_rq), &parent, &p);
1028 rb_link_node(&cfqq->p_node, parent, p);
1029 rb_insert_color(&cfqq->p_node, cfqq->p_root);
1031 cfqq->p_root = NULL;
1035 * Update cfqq's position in the service tree.
1037 static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1040 * Resorting requires the cfqq to be on the RR list already.
1042 if (cfq_cfqq_on_rr(cfqq)) {
1043 cfq_service_tree_add(cfqd, cfqq, 0);
1044 cfq_prio_tree_add(cfqd, cfqq);
1049 * add to busy list of queues for service, trying to be fair in ordering
1050 * the pending list according to last request service
1052 static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1054 cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
1055 BUG_ON(cfq_cfqq_on_rr(cfqq));
1056 cfq_mark_cfqq_on_rr(cfqq);
1057 cfqd->busy_queues++;
1059 cfq_resort_rr_list(cfqd, cfqq);
1063 * Called when the cfqq no longer has requests pending, remove it from
1066 static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1068 cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
1069 BUG_ON(!cfq_cfqq_on_rr(cfqq));
1070 cfq_clear_cfqq_on_rr(cfqq);
1072 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1073 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1074 cfqq->service_tree = NULL;
1077 rb_erase(&cfqq->p_node, cfqq->p_root);
1078 cfqq->p_root = NULL;
1081 cfq_group_service_tree_del(cfqd, cfqq->cfqg);
1082 BUG_ON(!cfqd->busy_queues);
1083 cfqd->busy_queues--;
1087 * rb tree support functions
1089 static void cfq_del_rq_rb(struct request *rq)
1091 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1092 const int sync = rq_is_sync(rq);
1094 BUG_ON(!cfqq->queued[sync]);
1095 cfqq->queued[sync]--;
1097 elv_rb_del(&cfqq->sort_list, rq);
1099 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) {
1101 * Queue will be deleted from service tree when we actually
1102 * expire it later. Right now just remove it from prio tree
1106 rb_erase(&cfqq->p_node, cfqq->p_root);
1107 cfqq->p_root = NULL;
1112 static void cfq_add_rq_rb(struct request *rq)
1114 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1115 struct cfq_data *cfqd = cfqq->cfqd;
1116 struct request *__alias, *prev;
1118 cfqq->queued[rq_is_sync(rq)]++;
1121 * looks a little odd, but the first insert might return an alias.
1122 * if that happens, put the alias on the dispatch list
1124 while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL)
1125 cfq_dispatch_insert(cfqd->queue, __alias);
1127 if (!cfq_cfqq_on_rr(cfqq))
1128 cfq_add_cfqq_rr(cfqd, cfqq);
1131 * check if this request is a better next-serve candidate
1133 prev = cfqq->next_rq;
1134 cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq, cfqd->last_position);
1137 * adjust priority tree position, if ->next_rq changes
1139 if (prev != cfqq->next_rq)
1140 cfq_prio_tree_add(cfqd, cfqq);
1142 BUG_ON(!cfqq->next_rq);
1145 static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
1147 elv_rb_del(&cfqq->sort_list, rq);
1148 cfqq->queued[rq_is_sync(rq)]--;
1152 static struct request *
1153 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
1155 struct task_struct *tsk = current;
1156 struct cfq_io_context *cic;
1157 struct cfq_queue *cfqq;
1159 cic = cfq_cic_lookup(cfqd, tsk->io_context);
1163 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
1165 sector_t sector = bio->bi_sector + bio_sectors(bio);
1167 return elv_rb_find(&cfqq->sort_list, sector);
1173 static void cfq_activate_request(struct request_queue *q, struct request *rq)
1175 struct cfq_data *cfqd = q->elevator->elevator_data;
1177 cfqd->rq_in_driver[rq_is_sync(rq)]++;
1178 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
1179 rq_in_driver(cfqd));
1181 cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
1184 static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
1186 struct cfq_data *cfqd = q->elevator->elevator_data;
1187 const int sync = rq_is_sync(rq);
1189 WARN_ON(!cfqd->rq_in_driver[sync]);
1190 cfqd->rq_in_driver[sync]--;
1191 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
1192 rq_in_driver(cfqd));
1195 static void cfq_remove_request(struct request *rq)
1197 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1199 if (cfqq->next_rq == rq)
1200 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
1202 list_del_init(&rq->queuelist);
1205 cfqq->cfqd->rq_queued--;
1206 if (rq_is_meta(rq)) {
1207 WARN_ON(!cfqq->meta_pending);
1208 cfqq->meta_pending--;
1212 static int cfq_merge(struct request_queue *q, struct request **req,
1215 struct cfq_data *cfqd = q->elevator->elevator_data;
1216 struct request *__rq;
1218 __rq = cfq_find_rq_fmerge(cfqd, bio);
1219 if (__rq && elv_rq_merge_ok(__rq, bio)) {
1221 return ELEVATOR_FRONT_MERGE;
1224 return ELEVATOR_NO_MERGE;
1227 static void cfq_merged_request(struct request_queue *q, struct request *req,
1230 if (type == ELEVATOR_FRONT_MERGE) {
1231 struct cfq_queue *cfqq = RQ_CFQQ(req);
1233 cfq_reposition_rq_rb(cfqq, req);
1238 cfq_merged_requests(struct request_queue *q, struct request *rq,
1239 struct request *next)
1241 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1243 * reposition in fifo if next is older than rq
1245 if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
1246 time_before(rq_fifo_time(next), rq_fifo_time(rq))) {
1247 list_move(&rq->queuelist, &next->queuelist);
1248 rq_set_fifo_time(rq, rq_fifo_time(next));
1251 if (cfqq->next_rq == next)
1253 cfq_remove_request(next);
1256 static int cfq_allow_merge(struct request_queue *q, struct request *rq,
1259 struct cfq_data *cfqd = q->elevator->elevator_data;
1260 struct cfq_io_context *cic;
1261 struct cfq_queue *cfqq;
1264 * Disallow merge of a sync bio into an async request.
1266 if (cfq_bio_sync(bio) && !rq_is_sync(rq))
1270 * Lookup the cfqq that this bio will be queued with. Allow
1271 * merge only if rq is queued there.
1273 cic = cfq_cic_lookup(cfqd, current->io_context);
1277 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
1278 return cfqq == RQ_CFQQ(rq);
1281 static void __cfq_set_active_queue(struct cfq_data *cfqd,
1282 struct cfq_queue *cfqq)
1285 cfq_log_cfqq(cfqd, cfqq, "set_active");
1286 cfqq->slice_start = 0;
1287 cfqq->dispatch_start = jiffies;
1288 cfqq->slice_end = 0;
1289 cfqq->slice_dispatch = 0;
1291 cfq_clear_cfqq_wait_request(cfqq);
1292 cfq_clear_cfqq_must_dispatch(cfqq);
1293 cfq_clear_cfqq_must_alloc_slice(cfqq);
1294 cfq_clear_cfqq_fifo_expire(cfqq);
1295 cfq_mark_cfqq_slice_new(cfqq);
1297 del_timer(&cfqd->idle_slice_timer);
1300 cfqd->active_queue = cfqq;
1304 * current cfqq expired its slice (or was too idle), select new one
1307 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1310 cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
1312 if (cfq_cfqq_wait_request(cfqq))
1313 del_timer(&cfqd->idle_slice_timer);
1315 cfq_clear_cfqq_wait_request(cfqq);
1318 * store what was left of this slice, if the queue idled/timed out
1320 if (timed_out && !cfq_cfqq_slice_new(cfqq)) {
1321 cfqq->slice_resid = cfqq->slice_end - jiffies;
1322 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
1325 cfq_group_served(cfqd, cfqq->cfqg, cfqq);
1327 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
1328 cfq_del_cfqq_rr(cfqd, cfqq);
1330 cfq_resort_rr_list(cfqd, cfqq);
1332 if (cfqq == cfqd->active_queue)
1333 cfqd->active_queue = NULL;
1335 if (&cfqq->cfqg->rb_node == cfqd->grp_service_tree.active)
1336 cfqd->grp_service_tree.active = NULL;
1338 if (cfqd->active_cic) {
1339 put_io_context(cfqd->active_cic->ioc);
1340 cfqd->active_cic = NULL;
1344 static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out)
1346 struct cfq_queue *cfqq = cfqd->active_queue;
1349 __cfq_slice_expired(cfqd, cfqq, timed_out);
1353 * Get next queue for service. Unless we have a queue preemption,
1354 * we'll simply select the first cfqq in the service tree.
1356 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
1358 struct cfq_rb_root *service_tree =
1359 service_tree_for(cfqd->serving_group, cfqd->serving_prio,
1360 cfqd->serving_type, cfqd);
1362 if (!cfqd->rq_queued)
1365 /* There is nothing to dispatch */
1368 if (RB_EMPTY_ROOT(&service_tree->rb))
1370 return cfq_rb_first(service_tree);
1373 static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd)
1375 struct cfq_group *cfqg = &cfqd->root_group;
1376 struct cfq_queue *cfqq;
1378 struct cfq_rb_root *st;
1380 if (!cfqd->rq_queued)
1383 for_each_cfqg_st(cfqg, i, j, st)
1384 if ((cfqq = cfq_rb_first(st)) != NULL)
1390 * Get and set a new active queue for service.
1392 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
1393 struct cfq_queue *cfqq)
1396 cfqq = cfq_get_next_queue(cfqd);
1398 __cfq_set_active_queue(cfqd, cfqq);
1402 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
1405 if (blk_rq_pos(rq) >= cfqd->last_position)
1406 return blk_rq_pos(rq) - cfqd->last_position;
1408 return cfqd->last_position - blk_rq_pos(rq);
1411 #define CFQQ_SEEK_THR 8 * 1024
1412 #define CFQQ_SEEKY(cfqq) ((cfqq)->seek_mean > CFQQ_SEEK_THR)
1414 static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1417 sector_t sdist = cfqq->seek_mean;
1419 if (!sample_valid(cfqq->seek_samples))
1420 sdist = CFQQ_SEEK_THR;
1422 return cfq_dist_from_last(cfqd, rq) <= sdist;
1425 static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
1426 struct cfq_queue *cur_cfqq)
1428 struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
1429 struct rb_node *parent, *node;
1430 struct cfq_queue *__cfqq;
1431 sector_t sector = cfqd->last_position;
1433 if (RB_EMPTY_ROOT(root))
1437 * First, if we find a request starting at the end of the last
1438 * request, choose it.
1440 __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
1445 * If the exact sector wasn't found, the parent of the NULL leaf
1446 * will contain the closest sector.
1448 __cfqq = rb_entry(parent, struct cfq_queue, p_node);
1449 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1452 if (blk_rq_pos(__cfqq->next_rq) < sector)
1453 node = rb_next(&__cfqq->p_node);
1455 node = rb_prev(&__cfqq->p_node);
1459 __cfqq = rb_entry(node, struct cfq_queue, p_node);
1460 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1468 * cur_cfqq - passed in so that we don't decide that the current queue is
1469 * closely cooperating with itself.
1471 * So, basically we're assuming that that cur_cfqq has dispatched at least
1472 * one request, and that cfqd->last_position reflects a position on the disk
1473 * associated with the I/O issued by cur_cfqq. I'm not sure this is a valid
1476 static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
1477 struct cfq_queue *cur_cfqq)
1479 struct cfq_queue *cfqq;
1481 if (!cfq_cfqq_sync(cur_cfqq))
1483 if (CFQQ_SEEKY(cur_cfqq))
1487 * We should notice if some of the queues are cooperating, eg
1488 * working closely on the same area of the disk. In that case,
1489 * we can group them together and don't waste time idling.
1491 cfqq = cfqq_close(cfqd, cur_cfqq);
1496 * It only makes sense to merge sync queues.
1498 if (!cfq_cfqq_sync(cfqq))
1500 if (CFQQ_SEEKY(cfqq))
1504 * Do not merge queues of different priority classes
1506 if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq))
1513 * Determine whether we should enforce idle window for this queue.
1516 static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1518 enum wl_prio_t prio = cfqq_prio(cfqq);
1519 struct cfq_rb_root *service_tree = cfqq->service_tree;
1521 BUG_ON(!service_tree);
1522 BUG_ON(!service_tree->count);
1524 /* We never do for idle class queues. */
1525 if (prio == IDLE_WORKLOAD)
1528 /* We do for queues that were marked with idle window flag. */
1529 if (cfq_cfqq_idle_window(cfqq))
1533 * Otherwise, we do only if they are the last ones
1534 * in their service tree.
1536 return service_tree->count == 1;
1539 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
1541 struct cfq_queue *cfqq = cfqd->active_queue;
1542 struct cfq_io_context *cic;
1546 * SSD device without seek penalty, disable idling. But only do so
1547 * for devices that support queuing, otherwise we still have a problem
1548 * with sync vs async workloads.
1550 if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
1553 WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
1554 WARN_ON(cfq_cfqq_slice_new(cfqq));
1557 * idle is disabled, either manually or by past process history
1559 if (!cfqd->cfq_slice_idle || !cfq_should_idle(cfqd, cfqq))
1563 * still active requests from this queue, don't idle
1565 if (cfqq->dispatched)
1569 * task has exited, don't wait
1571 cic = cfqd->active_cic;
1572 if (!cic || !atomic_read(&cic->ioc->nr_tasks))
1576 * If our average think time is larger than the remaining time
1577 * slice, then don't idle. This avoids overrunning the allotted
1580 if (sample_valid(cic->ttime_samples) &&
1581 (cfqq->slice_end - jiffies < cic->ttime_mean))
1584 cfq_mark_cfqq_wait_request(cfqq);
1586 sl = cfqd->cfq_slice_idle;
1588 mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
1589 cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu", sl);
1593 * Move request from internal lists to the request queue dispatch list.
1595 static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
1597 struct cfq_data *cfqd = q->elevator->elevator_data;
1598 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1600 cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
1602 cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
1603 cfq_remove_request(rq);
1605 elv_dispatch_sort(q, rq);
1607 if (cfq_cfqq_sync(cfqq))
1608 cfqd->sync_flight++;
1612 * return expired entry, or NULL to just start from scratch in rbtree
1614 static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
1616 struct request *rq = NULL;
1618 if (cfq_cfqq_fifo_expire(cfqq))
1621 cfq_mark_cfqq_fifo_expire(cfqq);
1623 if (list_empty(&cfqq->fifo))
1626 rq = rq_entry_fifo(cfqq->fifo.next);
1627 if (time_before(jiffies, rq_fifo_time(rq)))
1630 cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq);
1635 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1637 const int base_rq = cfqd->cfq_slice_async_rq;
1639 WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
1641 return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
1645 * Must be called with the queue_lock held.
1647 static int cfqq_process_refs(struct cfq_queue *cfqq)
1649 int process_refs, io_refs;
1651 io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE];
1652 process_refs = atomic_read(&cfqq->ref) - io_refs;
1653 BUG_ON(process_refs < 0);
1654 return process_refs;
1657 static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq)
1659 int process_refs, new_process_refs;
1660 struct cfq_queue *__cfqq;
1662 /* Avoid a circular list and skip interim queue merges */
1663 while ((__cfqq = new_cfqq->new_cfqq)) {
1669 process_refs = cfqq_process_refs(cfqq);
1671 * If the process for the cfqq has gone away, there is no
1672 * sense in merging the queues.
1674 if (process_refs == 0)
1678 * Merge in the direction of the lesser amount of work.
1680 new_process_refs = cfqq_process_refs(new_cfqq);
1681 if (new_process_refs >= process_refs) {
1682 cfqq->new_cfqq = new_cfqq;
1683 atomic_add(process_refs, &new_cfqq->ref);
1685 new_cfqq->new_cfqq = cfqq;
1686 atomic_add(new_process_refs, &cfqq->ref);
1690 static enum wl_type_t cfq_choose_wl(struct cfq_data *cfqd,
1691 struct cfq_group *cfqg, enum wl_prio_t prio,
1694 struct cfq_queue *queue;
1696 bool key_valid = false;
1697 unsigned long lowest_key = 0;
1698 enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD;
1702 * When priorities switched, we prefer starting
1703 * from SYNC_NOIDLE (first choice), or just SYNC
1706 if (service_tree_for(cfqg, prio, cur_best, cfqd)->count)
1708 cur_best = SYNC_WORKLOAD;
1709 if (service_tree_for(cfqg, prio, cur_best, cfqd)->count)
1712 return ASYNC_WORKLOAD;
1715 for (i = 0; i < 3; ++i) {
1716 /* otherwise, select the one with lowest rb_key */
1717 queue = cfq_rb_first(service_tree_for(cfqg, prio, i, cfqd));
1719 (!key_valid || time_before(queue->rb_key, lowest_key))) {
1720 lowest_key = queue->rb_key;
1729 static void choose_service_tree(struct cfq_data *cfqd, struct cfq_group *cfqg)
1731 enum wl_prio_t previous_prio = cfqd->serving_prio;
1735 struct cfq_rb_root *st;
1736 unsigned group_slice;
1739 cfqd->serving_prio = IDLE_WORKLOAD;
1740 cfqd->workload_expires = jiffies + 1;
1744 /* Choose next priority. RT > BE > IDLE */
1745 if (cfq_group_busy_queues_wl(RT_WORKLOAD, cfqd, cfqg))
1746 cfqd->serving_prio = RT_WORKLOAD;
1747 else if (cfq_group_busy_queues_wl(BE_WORKLOAD, cfqd, cfqg))
1748 cfqd->serving_prio = BE_WORKLOAD;
1750 cfqd->serving_prio = IDLE_WORKLOAD;
1751 cfqd->workload_expires = jiffies + 1;
1756 * For RT and BE, we have to choose also the type
1757 * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
1760 prio_changed = (cfqd->serving_prio != previous_prio);
1761 st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type,
1766 * If priority didn't change, check workload expiration,
1767 * and that we still have other queues ready
1769 if (!prio_changed && count &&
1770 !time_after(jiffies, cfqd->workload_expires))
1773 /* otherwise select new workload type */
1774 cfqd->serving_type =
1775 cfq_choose_wl(cfqd, cfqg, cfqd->serving_prio, prio_changed);
1776 st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type,
1781 * the workload slice is computed as a fraction of target latency
1782 * proportional to the number of queues in that workload, over
1783 * all the queues in the same priority class
1785 group_slice = cfq_group_slice(cfqd, cfqg);
1787 slice = group_slice * count /
1788 max_t(unsigned, cfqg->busy_queues_avg[cfqd->serving_prio],
1789 cfq_group_busy_queues_wl(cfqd->serving_prio, cfqd, cfqg));
1791 if (cfqd->serving_type == ASYNC_WORKLOAD)
1792 /* async workload slice is scaled down according to
1793 * the sync/async slice ratio. */
1794 slice = slice * cfqd->cfq_slice[0] / cfqd->cfq_slice[1];
1796 /* sync workload slice is at least 2 * cfq_slice_idle */
1797 slice = max(slice, 2 * cfqd->cfq_slice_idle);
1799 slice = max_t(unsigned, slice, CFQ_MIN_TT);
1800 cfqd->workload_expires = jiffies + slice;
1801 cfqd->noidle_tree_requires_idle = false;
1804 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd)
1806 struct cfq_rb_root *st = &cfqd->grp_service_tree;
1807 struct cfq_group *cfqg;
1809 if (RB_EMPTY_ROOT(&st->rb))
1811 cfqg = cfq_rb_first_group(st);
1812 st->active = &cfqg->rb_node;
1813 update_min_vdisktime(st);
1817 static void cfq_choose_cfqg(struct cfq_data *cfqd)
1819 struct cfq_group *cfqg = cfq_get_next_cfqg(cfqd);
1821 cfqd->serving_group = cfqg;
1823 /* Restore the workload type data */
1824 if (cfqg->saved_workload_slice) {
1825 cfqd->workload_expires = jiffies + cfqg->saved_workload_slice;
1826 cfqd->serving_type = cfqg->saved_workload;
1827 cfqd->serving_prio = cfqg->saved_serving_prio;
1829 choose_service_tree(cfqd, cfqg);
1833 * Select a queue for service. If we have a current active queue,
1834 * check whether to continue servicing it, or retrieve and set a new one.
1836 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
1838 struct cfq_queue *cfqq, *new_cfqq = NULL;
1840 cfqq = cfqd->active_queue;
1844 if (!cfqd->rq_queued)
1847 * The active queue has run out of time, expire it and select new.
1849 if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq))
1853 * The active queue has requests and isn't expired, allow it to
1856 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
1860 * If another queue has a request waiting within our mean seek
1861 * distance, let it run. The expire code will check for close
1862 * cooperators and put the close queue at the front of the service
1863 * tree. If possible, merge the expiring queue with the new cfqq.
1865 new_cfqq = cfq_close_cooperator(cfqd, cfqq);
1867 if (!cfqq->new_cfqq)
1868 cfq_setup_merge(cfqq, new_cfqq);
1873 * No requests pending. If the active queue still has requests in
1874 * flight or is idling for a new request, allow either of these
1875 * conditions to happen (or time out) before selecting a new queue.
1877 if (timer_pending(&cfqd->idle_slice_timer) ||
1878 (cfqq->dispatched && cfq_should_idle(cfqd, cfqq))) {
1884 cfq_slice_expired(cfqd, 0);
1887 * Current queue expired. Check if we have to switch to a new
1891 cfq_choose_cfqg(cfqd);
1893 cfqq = cfq_set_active_queue(cfqd, new_cfqq);
1898 static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
1902 while (cfqq->next_rq) {
1903 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
1907 BUG_ON(!list_empty(&cfqq->fifo));
1909 /* By default cfqq is not expired if it is empty. Do it explicitly */
1910 __cfq_slice_expired(cfqq->cfqd, cfqq, 0);
1915 * Drain our current requests. Used for barriers and when switching
1916 * io schedulers on-the-fly.
1918 static int cfq_forced_dispatch(struct cfq_data *cfqd)
1920 struct cfq_queue *cfqq;
1923 while ((cfqq = cfq_get_next_queue_forced(cfqd)) != NULL)
1924 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
1926 cfq_slice_expired(cfqd, 0);
1927 BUG_ON(cfqd->busy_queues);
1929 cfq_log(cfqd, "forced_dispatch=%d", dispatched);
1933 static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1935 unsigned int max_dispatch;
1938 * Drain async requests before we start sync IO
1940 if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_driver[BLK_RW_ASYNC])
1944 * If this is an async queue and we have sync IO in flight, let it wait
1946 if (cfqd->sync_flight && !cfq_cfqq_sync(cfqq))
1949 max_dispatch = cfqd->cfq_quantum;
1950 if (cfq_class_idle(cfqq))
1954 * Does this cfqq already have too much IO in flight?
1956 if (cfqq->dispatched >= max_dispatch) {
1958 * idle queue must always only have a single IO in flight
1960 if (cfq_class_idle(cfqq))
1964 * We have other queues, don't allow more IO from this one
1966 if (cfqd->busy_queues > 1)
1970 * Sole queue user, no limit
1976 * Async queues must wait a bit before being allowed dispatch.
1977 * We also ramp up the dispatch depth gradually for async IO,
1978 * based on the last sync IO we serviced
1980 if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
1981 unsigned long last_sync = jiffies - cfqd->last_end_sync_rq;
1984 depth = last_sync / cfqd->cfq_slice[1];
1985 if (!depth && !cfqq->dispatched)
1987 if (depth < max_dispatch)
1988 max_dispatch = depth;
1992 * If we're below the current max, allow a dispatch
1994 return cfqq->dispatched < max_dispatch;
1998 * Dispatch a request from cfqq, moving them to the request queue
2001 static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2005 BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
2007 if (!cfq_may_dispatch(cfqd, cfqq))
2011 * follow expired path, else get first next available
2013 rq = cfq_check_fifo(cfqq);
2018 * insert request into driver dispatch list
2020 cfq_dispatch_insert(cfqd->queue, rq);
2022 if (!cfqd->active_cic) {
2023 struct cfq_io_context *cic = RQ_CIC(rq);
2025 atomic_long_inc(&cic->ioc->refcount);
2026 cfqd->active_cic = cic;
2033 * Find the cfqq that we need to service and move a request from that to the
2036 static int cfq_dispatch_requests(struct request_queue *q, int force)
2038 struct cfq_data *cfqd = q->elevator->elevator_data;
2039 struct cfq_queue *cfqq;
2041 if (!cfqd->busy_queues)
2044 if (unlikely(force))
2045 return cfq_forced_dispatch(cfqd);
2047 cfqq = cfq_select_queue(cfqd);
2052 * Dispatch a request from this cfqq, if it is allowed
2054 if (!cfq_dispatch_request(cfqd, cfqq))
2057 cfqq->slice_dispatch++;
2058 cfq_clear_cfqq_must_dispatch(cfqq);
2061 * expire an async queue immediately if it has used up its slice. idle
2062 * queue always expire after 1 dispatch round.
2064 if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
2065 cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
2066 cfq_class_idle(cfqq))) {
2067 cfqq->slice_end = jiffies + 1;
2068 cfq_slice_expired(cfqd, 0);
2071 cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
2076 * task holds one reference to the queue, dropped when task exits. each rq
2077 * in-flight on this queue also holds a reference, dropped when rq is freed.
2079 * queue lock must be held here.
2081 static void cfq_put_queue(struct cfq_queue *cfqq)
2083 struct cfq_data *cfqd = cfqq->cfqd;
2085 BUG_ON(atomic_read(&cfqq->ref) <= 0);
2087 if (!atomic_dec_and_test(&cfqq->ref))
2090 cfq_log_cfqq(cfqd, cfqq, "put_queue");
2091 BUG_ON(rb_first(&cfqq->sort_list));
2092 BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
2094 if (unlikely(cfqd->active_queue == cfqq)) {
2095 __cfq_slice_expired(cfqd, cfqq, 0);
2096 cfq_schedule_dispatch(cfqd);
2099 BUG_ON(cfq_cfqq_on_rr(cfqq));
2100 kmem_cache_free(cfq_pool, cfqq);
2104 * Must always be called with the rcu_read_lock() held
2107 __call_for_each_cic(struct io_context *ioc,
2108 void (*func)(struct io_context *, struct cfq_io_context *))
2110 struct cfq_io_context *cic;
2111 struct hlist_node *n;
2113 hlist_for_each_entry_rcu(cic, n, &ioc->cic_list, cic_list)
2118 * Call func for each cic attached to this ioc.
2121 call_for_each_cic(struct io_context *ioc,
2122 void (*func)(struct io_context *, struct cfq_io_context *))
2125 __call_for_each_cic(ioc, func);
2129 static void cfq_cic_free_rcu(struct rcu_head *head)
2131 struct cfq_io_context *cic;
2133 cic = container_of(head, struct cfq_io_context, rcu_head);
2135 kmem_cache_free(cfq_ioc_pool, cic);
2136 elv_ioc_count_dec(cfq_ioc_count);
2140 * CFQ scheduler is exiting, grab exit lock and check
2141 * the pending io context count. If it hits zero,
2142 * complete ioc_gone and set it back to NULL
2144 spin_lock(&ioc_gone_lock);
2145 if (ioc_gone && !elv_ioc_count_read(cfq_ioc_count)) {
2149 spin_unlock(&ioc_gone_lock);
2153 static void cfq_cic_free(struct cfq_io_context *cic)
2155 call_rcu(&cic->rcu_head, cfq_cic_free_rcu);
2158 static void cic_free_func(struct io_context *ioc, struct cfq_io_context *cic)
2160 unsigned long flags;
2162 BUG_ON(!cic->dead_key);
2164 spin_lock_irqsave(&ioc->lock, flags);
2165 radix_tree_delete(&ioc->radix_root, cic->dead_key);
2166 hlist_del_rcu(&cic->cic_list);
2167 spin_unlock_irqrestore(&ioc->lock, flags);
2173 * Must be called with rcu_read_lock() held or preemption otherwise disabled.
2174 * Only two callers of this - ->dtor() which is called with the rcu_read_lock(),
2175 * and ->trim() which is called with the task lock held
2177 static void cfq_free_io_context(struct io_context *ioc)
2180 * ioc->refcount is zero here, or we are called from elv_unregister(),
2181 * so no more cic's are allowed to be linked into this ioc. So it
2182 * should be ok to iterate over the known list, we will see all cic's
2183 * since no new ones are added.
2185 __call_for_each_cic(ioc, cic_free_func);
2188 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2190 struct cfq_queue *__cfqq, *next;
2192 if (unlikely(cfqq == cfqd->active_queue)) {
2193 __cfq_slice_expired(cfqd, cfqq, 0);
2194 cfq_schedule_dispatch(cfqd);
2198 * If this queue was scheduled to merge with another queue, be
2199 * sure to drop the reference taken on that queue (and others in
2200 * the merge chain). See cfq_setup_merge and cfq_merge_cfqqs.
2202 __cfqq = cfqq->new_cfqq;
2204 if (__cfqq == cfqq) {
2205 WARN(1, "cfqq->new_cfqq loop detected\n");
2208 next = __cfqq->new_cfqq;
2209 cfq_put_queue(__cfqq);
2213 cfq_put_queue(cfqq);
2216 static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
2217 struct cfq_io_context *cic)
2219 struct io_context *ioc = cic->ioc;
2221 list_del_init(&cic->queue_list);
2224 * Make sure key == NULL is seen for dead queues
2227 cic->dead_key = (unsigned long) cic->key;
2230 if (ioc->ioc_data == cic)
2231 rcu_assign_pointer(ioc->ioc_data, NULL);
2233 if (cic->cfqq[BLK_RW_ASYNC]) {
2234 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]);
2235 cic->cfqq[BLK_RW_ASYNC] = NULL;
2238 if (cic->cfqq[BLK_RW_SYNC]) {
2239 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]);
2240 cic->cfqq[BLK_RW_SYNC] = NULL;
2244 static void cfq_exit_single_io_context(struct io_context *ioc,
2245 struct cfq_io_context *cic)
2247 struct cfq_data *cfqd = cic->key;
2250 struct request_queue *q = cfqd->queue;
2251 unsigned long flags;
2253 spin_lock_irqsave(q->queue_lock, flags);
2256 * Ensure we get a fresh copy of the ->key to prevent
2257 * race between exiting task and queue
2259 smp_read_barrier_depends();
2261 __cfq_exit_single_io_context(cfqd, cic);
2263 spin_unlock_irqrestore(q->queue_lock, flags);
2268 * The process that ioc belongs to has exited, we need to clean up
2269 * and put the internal structures we have that belongs to that process.
2271 static void cfq_exit_io_context(struct io_context *ioc)
2273 call_for_each_cic(ioc, cfq_exit_single_io_context);
2276 static struct cfq_io_context *
2277 cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
2279 struct cfq_io_context *cic;
2281 cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO,
2284 cic->last_end_request = jiffies;
2285 INIT_LIST_HEAD(&cic->queue_list);
2286 INIT_HLIST_NODE(&cic->cic_list);
2287 cic->dtor = cfq_free_io_context;
2288 cic->exit = cfq_exit_io_context;
2289 elv_ioc_count_inc(cfq_ioc_count);
2295 static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc)
2297 struct task_struct *tsk = current;
2300 if (!cfq_cfqq_prio_changed(cfqq))
2303 ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio);
2304 switch (ioprio_class) {
2306 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
2307 case IOPRIO_CLASS_NONE:
2309 * no prio set, inherit CPU scheduling settings
2311 cfqq->ioprio = task_nice_ioprio(tsk);
2312 cfqq->ioprio_class = task_nice_ioclass(tsk);
2314 case IOPRIO_CLASS_RT:
2315 cfqq->ioprio = task_ioprio(ioc);
2316 cfqq->ioprio_class = IOPRIO_CLASS_RT;
2318 case IOPRIO_CLASS_BE:
2319 cfqq->ioprio = task_ioprio(ioc);
2320 cfqq->ioprio_class = IOPRIO_CLASS_BE;
2322 case IOPRIO_CLASS_IDLE:
2323 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
2325 cfq_clear_cfqq_idle_window(cfqq);
2330 * keep track of original prio settings in case we have to temporarily
2331 * elevate the priority of this queue
2333 cfqq->org_ioprio = cfqq->ioprio;
2334 cfqq->org_ioprio_class = cfqq->ioprio_class;
2335 cfq_clear_cfqq_prio_changed(cfqq);
2338 static void changed_ioprio(struct io_context *ioc, struct cfq_io_context *cic)
2340 struct cfq_data *cfqd = cic->key;
2341 struct cfq_queue *cfqq;
2342 unsigned long flags;
2344 if (unlikely(!cfqd))
2347 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2349 cfqq = cic->cfqq[BLK_RW_ASYNC];
2351 struct cfq_queue *new_cfqq;
2352 new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic->ioc,
2355 cic->cfqq[BLK_RW_ASYNC] = new_cfqq;
2356 cfq_put_queue(cfqq);
2360 cfqq = cic->cfqq[BLK_RW_SYNC];
2362 cfq_mark_cfqq_prio_changed(cfqq);
2364 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2367 static void cfq_ioc_set_ioprio(struct io_context *ioc)
2369 call_for_each_cic(ioc, changed_ioprio);
2370 ioc->ioprio_changed = 0;
2373 static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2374 pid_t pid, bool is_sync)
2376 RB_CLEAR_NODE(&cfqq->rb_node);
2377 RB_CLEAR_NODE(&cfqq->p_node);
2378 INIT_LIST_HEAD(&cfqq->fifo);
2380 atomic_set(&cfqq->ref, 0);
2383 cfq_mark_cfqq_prio_changed(cfqq);
2386 if (!cfq_class_idle(cfqq))
2387 cfq_mark_cfqq_idle_window(cfqq);
2388 cfq_mark_cfqq_sync(cfqq);
2393 static void cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg)
2398 static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd, int create)
2400 return &cfqd->root_group;
2403 static struct cfq_queue *
2404 cfq_find_alloc_queue(struct cfq_data *cfqd, bool is_sync,
2405 struct io_context *ioc, gfp_t gfp_mask)
2407 struct cfq_queue *cfqq, *new_cfqq = NULL;
2408 struct cfq_io_context *cic;
2409 struct cfq_group *cfqg;
2412 cfqg = cfq_get_cfqg(cfqd, 1);
2413 cic = cfq_cic_lookup(cfqd, ioc);
2414 /* cic always exists here */
2415 cfqq = cic_to_cfqq(cic, is_sync);
2418 * Always try a new alloc if we fell back to the OOM cfqq
2419 * originally, since it should just be a temporary situation.
2421 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
2426 } else if (gfp_mask & __GFP_WAIT) {
2427 spin_unlock_irq(cfqd->queue->queue_lock);
2428 new_cfqq = kmem_cache_alloc_node(cfq_pool,
2429 gfp_mask | __GFP_ZERO,
2431 spin_lock_irq(cfqd->queue->queue_lock);
2435 cfqq = kmem_cache_alloc_node(cfq_pool,
2436 gfp_mask | __GFP_ZERO,
2441 cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
2442 cfq_init_prio_data(cfqq, ioc);
2443 cfq_link_cfqq_cfqg(cfqq, cfqg);
2444 cfq_log_cfqq(cfqd, cfqq, "alloced");
2446 cfqq = &cfqd->oom_cfqq;
2450 kmem_cache_free(cfq_pool, new_cfqq);
2455 static struct cfq_queue **
2456 cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
2458 switch (ioprio_class) {
2459 case IOPRIO_CLASS_RT:
2460 return &cfqd->async_cfqq[0][ioprio];
2461 case IOPRIO_CLASS_BE:
2462 return &cfqd->async_cfqq[1][ioprio];
2463 case IOPRIO_CLASS_IDLE:
2464 return &cfqd->async_idle_cfqq;
2470 static struct cfq_queue *
2471 cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct io_context *ioc,
2474 const int ioprio = task_ioprio(ioc);
2475 const int ioprio_class = task_ioprio_class(ioc);
2476 struct cfq_queue **async_cfqq = NULL;
2477 struct cfq_queue *cfqq = NULL;
2480 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
2485 cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask);
2488 * pin the queue now that it's allocated, scheduler exit will prune it
2490 if (!is_sync && !(*async_cfqq)) {
2491 atomic_inc(&cfqq->ref);
2495 atomic_inc(&cfqq->ref);
2500 * We drop cfq io contexts lazily, so we may find a dead one.
2503 cfq_drop_dead_cic(struct cfq_data *cfqd, struct io_context *ioc,
2504 struct cfq_io_context *cic)
2506 unsigned long flags;
2508 WARN_ON(!list_empty(&cic->queue_list));
2510 spin_lock_irqsave(&ioc->lock, flags);
2512 BUG_ON(ioc->ioc_data == cic);
2514 radix_tree_delete(&ioc->radix_root, (unsigned long) cfqd);
2515 hlist_del_rcu(&cic->cic_list);
2516 spin_unlock_irqrestore(&ioc->lock, flags);
2521 static struct cfq_io_context *
2522 cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc)
2524 struct cfq_io_context *cic;
2525 unsigned long flags;
2534 * we maintain a last-hit cache, to avoid browsing over the tree
2536 cic = rcu_dereference(ioc->ioc_data);
2537 if (cic && cic->key == cfqd) {
2543 cic = radix_tree_lookup(&ioc->radix_root, (unsigned long) cfqd);
2547 /* ->key must be copied to avoid race with cfq_exit_queue() */
2550 cfq_drop_dead_cic(cfqd, ioc, cic);
2555 spin_lock_irqsave(&ioc->lock, flags);
2556 rcu_assign_pointer(ioc->ioc_data, cic);
2557 spin_unlock_irqrestore(&ioc->lock, flags);
2565 * Add cic into ioc, using cfqd as the search key. This enables us to lookup
2566 * the process specific cfq io context when entered from the block layer.
2567 * Also adds the cic to a per-cfqd list, used when this queue is removed.
2569 static int cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
2570 struct cfq_io_context *cic, gfp_t gfp_mask)
2572 unsigned long flags;
2575 ret = radix_tree_preload(gfp_mask);
2580 spin_lock_irqsave(&ioc->lock, flags);
2581 ret = radix_tree_insert(&ioc->radix_root,
2582 (unsigned long) cfqd, cic);
2584 hlist_add_head_rcu(&cic->cic_list, &ioc->cic_list);
2585 spin_unlock_irqrestore(&ioc->lock, flags);
2587 radix_tree_preload_end();
2590 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2591 list_add(&cic->queue_list, &cfqd->cic_list);
2592 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2597 printk(KERN_ERR "cfq: cic link failed!\n");
2603 * Setup general io context and cfq io context. There can be several cfq
2604 * io contexts per general io context, if this process is doing io to more
2605 * than one device managed by cfq.
2607 static struct cfq_io_context *
2608 cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
2610 struct io_context *ioc = NULL;
2611 struct cfq_io_context *cic;
2613 might_sleep_if(gfp_mask & __GFP_WAIT);
2615 ioc = get_io_context(gfp_mask, cfqd->queue->node);
2619 cic = cfq_cic_lookup(cfqd, ioc);
2623 cic = cfq_alloc_io_context(cfqd, gfp_mask);
2627 if (cfq_cic_link(cfqd, ioc, cic, gfp_mask))
2631 smp_read_barrier_depends();
2632 if (unlikely(ioc->ioprio_changed))
2633 cfq_ioc_set_ioprio(ioc);
2639 put_io_context(ioc);
2644 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
2646 unsigned long elapsed = jiffies - cic->last_end_request;
2647 unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
2649 cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
2650 cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
2651 cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
2655 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2661 if (!cfqq->last_request_pos)
2663 else if (cfqq->last_request_pos < blk_rq_pos(rq))
2664 sdist = blk_rq_pos(rq) - cfqq->last_request_pos;
2666 sdist = cfqq->last_request_pos - blk_rq_pos(rq);
2669 * Don't allow the seek distance to get too large from the
2670 * odd fragment, pagein, etc
2672 if (cfqq->seek_samples <= 60) /* second&third seek */
2673 sdist = min(sdist, (cfqq->seek_mean * 4) + 2*1024*1024);
2675 sdist = min(sdist, (cfqq->seek_mean * 4) + 2*1024*64);
2677 cfqq->seek_samples = (7*cfqq->seek_samples + 256) / 8;
2678 cfqq->seek_total = (7*cfqq->seek_total + (u64)256*sdist) / 8;
2679 total = cfqq->seek_total + (cfqq->seek_samples/2);
2680 do_div(total, cfqq->seek_samples);
2681 cfqq->seek_mean = (sector_t)total;
2684 * If this cfqq is shared between multiple processes, check to
2685 * make sure that those processes are still issuing I/Os within
2686 * the mean seek distance. If not, it may be time to break the
2687 * queues apart again.
2689 if (cfq_cfqq_coop(cfqq)) {
2690 if (CFQQ_SEEKY(cfqq) && !cfqq->seeky_start)
2691 cfqq->seeky_start = jiffies;
2692 else if (!CFQQ_SEEKY(cfqq))
2693 cfqq->seeky_start = 0;
2698 * Disable idle window if the process thinks too long or seeks so much that
2702 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2703 struct cfq_io_context *cic)
2705 int old_idle, enable_idle;
2708 * Don't idle for async or idle io prio class
2710 if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
2713 enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
2715 if (cfqq->queued[0] + cfqq->queued[1] >= 4)
2716 cfq_mark_cfqq_deep(cfqq);
2718 if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle ||
2719 (!cfq_cfqq_deep(cfqq) && sample_valid(cfqq->seek_samples)
2720 && CFQQ_SEEKY(cfqq)))
2722 else if (sample_valid(cic->ttime_samples)) {
2723 if (cic->ttime_mean > cfqd->cfq_slice_idle)
2729 if (old_idle != enable_idle) {
2730 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
2732 cfq_mark_cfqq_idle_window(cfqq);
2734 cfq_clear_cfqq_idle_window(cfqq);
2739 * Check if new_cfqq should preempt the currently active queue. Return 0 for
2740 * no or if we aren't sure, a 1 will cause a preempt.
2743 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
2746 struct cfq_queue *cfqq;
2748 cfqq = cfqd->active_queue;
2752 if (cfq_slice_used(cfqq))
2755 if (cfq_class_idle(new_cfqq))
2758 if (cfq_class_idle(cfqq))
2761 /* Allow preemption only if we are idling on sync-noidle tree */
2762 if (cfqd->serving_type == SYNC_NOIDLE_WORKLOAD &&
2763 cfqq_type(new_cfqq) == SYNC_NOIDLE_WORKLOAD &&
2764 new_cfqq->service_tree->count == 2 &&
2765 RB_EMPTY_ROOT(&cfqq->sort_list))
2769 * if the new request is sync, but the currently running queue is
2770 * not, let the sync request have priority.
2772 if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
2776 * So both queues are sync. Let the new request get disk time if
2777 * it's a metadata request and the current queue is doing regular IO.
2779 if (rq_is_meta(rq) && !cfqq->meta_pending)
2783 * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
2785 if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
2788 if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
2792 * if this request is as-good as one we would expect from the
2793 * current cfqq, let it preempt
2795 if (cfq_rq_close(cfqd, cfqq, rq))
2802 * cfqq preempts the active queue. if we allowed preempt with no slice left,
2803 * let it have half of its nominal slice.
2805 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2807 cfq_log_cfqq(cfqd, cfqq, "preempt");
2808 cfq_slice_expired(cfqd, 1);
2811 * Put the new queue at the front of the of the current list,
2812 * so we know that it will be selected next.
2814 BUG_ON(!cfq_cfqq_on_rr(cfqq));
2816 cfq_service_tree_add(cfqd, cfqq, 1);
2818 cfqq->slice_end = 0;
2819 cfq_mark_cfqq_slice_new(cfqq);
2823 * Called when a new fs request (rq) is added (to cfqq). Check if there's
2824 * something we should do about it
2827 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2830 struct cfq_io_context *cic = RQ_CIC(rq);
2834 cfqq->meta_pending++;
2836 cfq_update_io_thinktime(cfqd, cic);
2837 cfq_update_io_seektime(cfqd, cfqq, rq);
2838 cfq_update_idle_window(cfqd, cfqq, cic);
2840 cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
2842 if (cfqq == cfqd->active_queue) {
2844 * Remember that we saw a request from this process, but
2845 * don't start queuing just yet. Otherwise we risk seeing lots
2846 * of tiny requests, because we disrupt the normal plugging
2847 * and merging. If the request is already larger than a single
2848 * page, let it rip immediately. For that case we assume that
2849 * merging is already done. Ditto for a busy system that
2850 * has other work pending, don't risk delaying until the
2851 * idle timer unplug to continue working.
2853 if (cfq_cfqq_wait_request(cfqq)) {
2854 if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
2855 cfqd->busy_queues > 1) {
2856 del_timer(&cfqd->idle_slice_timer);
2857 __blk_run_queue(cfqd->queue);
2859 cfq_mark_cfqq_must_dispatch(cfqq);
2861 } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
2863 * not the active queue - expire current slice if it is
2864 * idle and has expired it's mean thinktime or this new queue
2865 * has some old slice time left and is of higher priority or
2866 * this new queue is RT and the current one is BE
2868 cfq_preempt_queue(cfqd, cfqq);
2869 __blk_run_queue(cfqd->queue);
2873 static void cfq_insert_request(struct request_queue *q, struct request *rq)
2875 struct cfq_data *cfqd = q->elevator->elevator_data;
2876 struct cfq_queue *cfqq = RQ_CFQQ(rq);
2878 cfq_log_cfqq(cfqd, cfqq, "insert_request");
2879 cfq_init_prio_data(cfqq, RQ_CIC(rq)->ioc);
2881 rq_set_fifo_time(rq, jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)]);
2882 list_add_tail(&rq->queuelist, &cfqq->fifo);
2885 cfq_rq_enqueued(cfqd, cfqq, rq);
2889 * Update hw_tag based on peak queue depth over 50 samples under
2892 static void cfq_update_hw_tag(struct cfq_data *cfqd)
2894 struct cfq_queue *cfqq = cfqd->active_queue;
2896 if (rq_in_driver(cfqd) > cfqd->hw_tag_est_depth)
2897 cfqd->hw_tag_est_depth = rq_in_driver(cfqd);
2899 if (cfqd->hw_tag == 1)
2902 if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
2903 rq_in_driver(cfqd) <= CFQ_HW_QUEUE_MIN)
2907 * If active queue hasn't enough requests and can idle, cfq might not
2908 * dispatch sufficient requests to hardware. Don't zero hw_tag in this
2911 if (cfqq && cfq_cfqq_idle_window(cfqq) &&
2912 cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] <
2913 CFQ_HW_QUEUE_MIN && rq_in_driver(cfqd) < CFQ_HW_QUEUE_MIN)
2916 if (cfqd->hw_tag_samples++ < 50)
2919 if (cfqd->hw_tag_est_depth >= CFQ_HW_QUEUE_MIN)
2925 static void cfq_completed_request(struct request_queue *q, struct request *rq)
2927 struct cfq_queue *cfqq = RQ_CFQQ(rq);
2928 struct cfq_data *cfqd = cfqq->cfqd;
2929 const int sync = rq_is_sync(rq);
2933 cfq_log_cfqq(cfqd, cfqq, "complete");
2935 cfq_update_hw_tag(cfqd);
2937 WARN_ON(!cfqd->rq_in_driver[sync]);
2938 WARN_ON(!cfqq->dispatched);
2939 cfqd->rq_in_driver[sync]--;
2942 if (cfq_cfqq_sync(cfqq))
2943 cfqd->sync_flight--;
2946 RQ_CIC(rq)->last_end_request = now;
2947 cfqd->last_end_sync_rq = now;
2951 * If this is the active queue, check if it needs to be expired,
2952 * or if we want to idle in case it has no pending requests.
2954 if (cfqd->active_queue == cfqq) {
2955 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
2957 if (cfq_cfqq_slice_new(cfqq)) {
2958 cfq_set_prio_slice(cfqd, cfqq);
2959 cfq_clear_cfqq_slice_new(cfqq);
2962 * Idling is not enabled on:
2964 * - idle-priority queues
2966 * - queues with still some requests queued
2967 * - when there is a close cooperator
2969 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
2970 cfq_slice_expired(cfqd, 1);
2971 else if (sync && cfqq_empty &&
2972 !cfq_close_cooperator(cfqd, cfqq)) {
2973 cfqd->noidle_tree_requires_idle |= !rq_noidle(rq);
2975 * Idling is enabled for SYNC_WORKLOAD.
2976 * SYNC_NOIDLE_WORKLOAD idles at the end of the tree
2977 * only if we processed at least one !rq_noidle request
2979 if (cfqd->serving_type == SYNC_WORKLOAD
2980 || cfqd->noidle_tree_requires_idle)
2981 cfq_arm_slice_timer(cfqd);
2985 if (!rq_in_driver(cfqd))
2986 cfq_schedule_dispatch(cfqd);
2990 * we temporarily boost lower priority queues if they are holding fs exclusive
2991 * resources. they are boosted to normal prio (CLASS_BE/4)
2993 static void cfq_prio_boost(struct cfq_queue *cfqq)
2995 if (has_fs_excl()) {
2997 * boost idle prio on transactions that would lock out other
2998 * users of the filesystem
3000 if (cfq_class_idle(cfqq))
3001 cfqq->ioprio_class = IOPRIO_CLASS_BE;
3002 if (cfqq->ioprio > IOPRIO_NORM)
3003 cfqq->ioprio = IOPRIO_NORM;
3006 * unboost the queue (if needed)
3008 cfqq->ioprio_class = cfqq->org_ioprio_class;
3009 cfqq->ioprio = cfqq->org_ioprio;
3013 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
3015 if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
3016 cfq_mark_cfqq_must_alloc_slice(cfqq);
3017 return ELV_MQUEUE_MUST;
3020 return ELV_MQUEUE_MAY;
3023 static int cfq_may_queue(struct request_queue *q, int rw)
3025 struct cfq_data *cfqd = q->elevator->elevator_data;
3026 struct task_struct *tsk = current;
3027 struct cfq_io_context *cic;
3028 struct cfq_queue *cfqq;
3031 * don't force setup of a queue from here, as a call to may_queue
3032 * does not necessarily imply that a request actually will be queued.
3033 * so just lookup a possibly existing queue, or return 'may queue'
3036 cic = cfq_cic_lookup(cfqd, tsk->io_context);
3038 return ELV_MQUEUE_MAY;
3040 cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
3042 cfq_init_prio_data(cfqq, cic->ioc);
3043 cfq_prio_boost(cfqq);
3045 return __cfq_may_queue(cfqq);
3048 return ELV_MQUEUE_MAY;
3052 * queue lock held here
3054 static void cfq_put_request(struct request *rq)
3056 struct cfq_queue *cfqq = RQ_CFQQ(rq);
3059 const int rw = rq_data_dir(rq);
3061 BUG_ON(!cfqq->allocated[rw]);
3062 cfqq->allocated[rw]--;
3064 put_io_context(RQ_CIC(rq)->ioc);
3066 rq->elevator_private = NULL;
3067 rq->elevator_private2 = NULL;
3069 cfq_put_queue(cfqq);
3073 static struct cfq_queue *
3074 cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_context *cic,
3075 struct cfq_queue *cfqq)
3077 cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq);
3078 cic_set_cfqq(cic, cfqq->new_cfqq, 1);
3079 cfq_mark_cfqq_coop(cfqq->new_cfqq);
3080 cfq_put_queue(cfqq);
3081 return cic_to_cfqq(cic, 1);
3084 static int should_split_cfqq(struct cfq_queue *cfqq)
3086 if (cfqq->seeky_start &&
3087 time_after(jiffies, cfqq->seeky_start + CFQQ_COOP_TOUT))
3093 * Returns NULL if a new cfqq should be allocated, or the old cfqq if this
3094 * was the last process referring to said cfqq.
3096 static struct cfq_queue *
3097 split_cfqq(struct cfq_io_context *cic, struct cfq_queue *cfqq)
3099 if (cfqq_process_refs(cfqq) == 1) {
3100 cfqq->seeky_start = 0;
3101 cfqq->pid = current->pid;
3102 cfq_clear_cfqq_coop(cfqq);
3106 cic_set_cfqq(cic, NULL, 1);
3107 cfq_put_queue(cfqq);
3111 * Allocate cfq data structures associated with this request.
3114 cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
3116 struct cfq_data *cfqd = q->elevator->elevator_data;
3117 struct cfq_io_context *cic;
3118 const int rw = rq_data_dir(rq);
3119 const bool is_sync = rq_is_sync(rq);
3120 struct cfq_queue *cfqq;
3121 unsigned long flags;
3123 might_sleep_if(gfp_mask & __GFP_WAIT);
3125 cic = cfq_get_io_context(cfqd, gfp_mask);
3127 spin_lock_irqsave(q->queue_lock, flags);
3133 cfqq = cic_to_cfqq(cic, is_sync);
3134 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
3135 cfqq = cfq_get_queue(cfqd, is_sync, cic->ioc, gfp_mask);
3136 cic_set_cfqq(cic, cfqq, is_sync);
3139 * If the queue was seeky for too long, break it apart.
3141 if (cfq_cfqq_coop(cfqq) && should_split_cfqq(cfqq)) {
3142 cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq");
3143 cfqq = split_cfqq(cic, cfqq);
3149 * Check to see if this queue is scheduled to merge with
3150 * another, closely cooperating queue. The merging of
3151 * queues happens here as it must be done in process context.
3152 * The reference on new_cfqq was taken in merge_cfqqs.
3155 cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq);
3158 cfqq->allocated[rw]++;
3159 atomic_inc(&cfqq->ref);
3161 spin_unlock_irqrestore(q->queue_lock, flags);
3163 rq->elevator_private = cic;
3164 rq->elevator_private2 = cfqq;
3169 put_io_context(cic->ioc);
3171 cfq_schedule_dispatch(cfqd);
3172 spin_unlock_irqrestore(q->queue_lock, flags);
3173 cfq_log(cfqd, "set_request fail");
3177 static void cfq_kick_queue(struct work_struct *work)
3179 struct cfq_data *cfqd =
3180 container_of(work, struct cfq_data, unplug_work);
3181 struct request_queue *q = cfqd->queue;
3183 spin_lock_irq(q->queue_lock);
3184 __blk_run_queue(cfqd->queue);
3185 spin_unlock_irq(q->queue_lock);
3189 * Timer running if the active_queue is currently idling inside its time slice
3191 static void cfq_idle_slice_timer(unsigned long data)
3193 struct cfq_data *cfqd = (struct cfq_data *) data;
3194 struct cfq_queue *cfqq;
3195 unsigned long flags;
3198 cfq_log(cfqd, "idle timer fired");
3200 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
3202 cfqq = cfqd->active_queue;
3207 * We saw a request before the queue expired, let it through
3209 if (cfq_cfqq_must_dispatch(cfqq))
3215 if (cfq_slice_used(cfqq))
3219 * only expire and reinvoke request handler, if there are
3220 * other queues with pending requests
3222 if (!cfqd->busy_queues)
3226 * not expired and it has a request pending, let it dispatch
3228 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3232 * Queue depth flag is reset only when the idle didn't succeed
3234 cfq_clear_cfqq_deep(cfqq);
3237 cfq_slice_expired(cfqd, timed_out);
3239 cfq_schedule_dispatch(cfqd);
3241 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
3244 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
3246 del_timer_sync(&cfqd->idle_slice_timer);
3247 cancel_work_sync(&cfqd->unplug_work);
3250 static void cfq_put_async_queues(struct cfq_data *cfqd)
3254 for (i = 0; i < IOPRIO_BE_NR; i++) {
3255 if (cfqd->async_cfqq[0][i])
3256 cfq_put_queue(cfqd->async_cfqq[0][i]);
3257 if (cfqd->async_cfqq[1][i])
3258 cfq_put_queue(cfqd->async_cfqq[1][i]);
3261 if (cfqd->async_idle_cfqq)
3262 cfq_put_queue(cfqd->async_idle_cfqq);
3265 static void cfq_exit_queue(struct elevator_queue *e)
3267 struct cfq_data *cfqd = e->elevator_data;
3268 struct request_queue *q = cfqd->queue;
3270 cfq_shutdown_timer_wq(cfqd);
3272 spin_lock_irq(q->queue_lock);
3274 if (cfqd->active_queue)
3275 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
3277 while (!list_empty(&cfqd->cic_list)) {
3278 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
3279 struct cfq_io_context,
3282 __cfq_exit_single_io_context(cfqd, cic);
3285 cfq_put_async_queues(cfqd);
3287 spin_unlock_irq(q->queue_lock);
3289 cfq_shutdown_timer_wq(cfqd);
3294 static void *cfq_init_queue(struct request_queue *q)
3296 struct cfq_data *cfqd;
3298 struct cfq_group *cfqg;
3299 struct cfq_rb_root *st;
3301 cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
3305 /* Init root service tree */
3306 cfqd->grp_service_tree = CFQ_RB_ROOT;
3308 /* Init root group */
3309 cfqg = &cfqd->root_group;
3310 for_each_cfqg_st(cfqg, i, j, st)
3312 RB_CLEAR_NODE(&cfqg->rb_node);
3314 /* Give preference to root group over other groups */
3315 cfqg->weight = 2*BLKIO_WEIGHT_DEFAULT;
3318 * Not strictly needed (since RB_ROOT just clears the node and we
3319 * zeroed cfqd on alloc), but better be safe in case someone decides
3320 * to add magic to the rb code
3322 for (i = 0; i < CFQ_PRIO_LISTS; i++)
3323 cfqd->prio_trees[i] = RB_ROOT;
3326 * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
3327 * Grab a permanent reference to it, so that the normal code flow
3328 * will not attempt to free it.
3330 cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
3331 atomic_inc(&cfqd->oom_cfqq.ref);
3332 cfq_link_cfqq_cfqg(&cfqd->oom_cfqq, &cfqd->root_group);
3334 INIT_LIST_HEAD(&cfqd->cic_list);
3338 init_timer(&cfqd->idle_slice_timer);
3339 cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
3340 cfqd->idle_slice_timer.data = (unsigned long) cfqd;
3342 INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
3344 cfqd->cfq_quantum = cfq_quantum;
3345 cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
3346 cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
3347 cfqd->cfq_back_max = cfq_back_max;
3348 cfqd->cfq_back_penalty = cfq_back_penalty;
3349 cfqd->cfq_slice[0] = cfq_slice_async;
3350 cfqd->cfq_slice[1] = cfq_slice_sync;
3351 cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
3352 cfqd->cfq_slice_idle = cfq_slice_idle;
3353 cfqd->cfq_latency = 1;
3355 cfqd->last_end_sync_rq = jiffies;
3359 static void cfq_slab_kill(void)
3362 * Caller already ensured that pending RCU callbacks are completed,
3363 * so we should have no busy allocations at this point.
3366 kmem_cache_destroy(cfq_pool);
3368 kmem_cache_destroy(cfq_ioc_pool);
3371 static int __init cfq_slab_setup(void)
3373 cfq_pool = KMEM_CACHE(cfq_queue, 0);
3377 cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0);
3388 * sysfs parts below -->
3391 cfq_var_show(unsigned int var, char *page)
3393 return sprintf(page, "%d\n", var);
3397 cfq_var_store(unsigned int *var, const char *page, size_t count)
3399 char *p = (char *) page;
3401 *var = simple_strtoul(p, &p, 10);
3405 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
3406 static ssize_t __FUNC(struct elevator_queue *e, char *page) \
3408 struct cfq_data *cfqd = e->elevator_data; \
3409 unsigned int __data = __VAR; \
3411 __data = jiffies_to_msecs(__data); \
3412 return cfq_var_show(__data, (page)); \
3414 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
3415 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
3416 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
3417 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
3418 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
3419 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
3420 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
3421 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
3422 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
3423 SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
3424 #undef SHOW_FUNCTION
3426 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
3427 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
3429 struct cfq_data *cfqd = e->elevator_data; \
3430 unsigned int __data; \
3431 int ret = cfq_var_store(&__data, (page), count); \
3432 if (__data < (MIN)) \
3434 else if (__data > (MAX)) \
3437 *(__PTR) = msecs_to_jiffies(__data); \
3439 *(__PTR) = __data; \
3442 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
3443 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
3445 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
3447 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
3448 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
3450 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
3451 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
3452 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
3453 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
3455 STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
3456 #undef STORE_FUNCTION
3458 #define CFQ_ATTR(name) \
3459 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
3461 static struct elv_fs_entry cfq_attrs[] = {
3463 CFQ_ATTR(fifo_expire_sync),
3464 CFQ_ATTR(fifo_expire_async),
3465 CFQ_ATTR(back_seek_max),
3466 CFQ_ATTR(back_seek_penalty),
3467 CFQ_ATTR(slice_sync),
3468 CFQ_ATTR(slice_async),
3469 CFQ_ATTR(slice_async_rq),
3470 CFQ_ATTR(slice_idle),
3471 CFQ_ATTR(low_latency),
3475 static struct elevator_type iosched_cfq = {
3477 .elevator_merge_fn = cfq_merge,
3478 .elevator_merged_fn = cfq_merged_request,
3479 .elevator_merge_req_fn = cfq_merged_requests,
3480 .elevator_allow_merge_fn = cfq_allow_merge,
3481 .elevator_dispatch_fn = cfq_dispatch_requests,
3482 .elevator_add_req_fn = cfq_insert_request,
3483 .elevator_activate_req_fn = cfq_activate_request,
3484 .elevator_deactivate_req_fn = cfq_deactivate_request,
3485 .elevator_queue_empty_fn = cfq_queue_empty,
3486 .elevator_completed_req_fn = cfq_completed_request,
3487 .elevator_former_req_fn = elv_rb_former_request,
3488 .elevator_latter_req_fn = elv_rb_latter_request,
3489 .elevator_set_req_fn = cfq_set_request,
3490 .elevator_put_req_fn = cfq_put_request,
3491 .elevator_may_queue_fn = cfq_may_queue,
3492 .elevator_init_fn = cfq_init_queue,
3493 .elevator_exit_fn = cfq_exit_queue,
3494 .trim = cfq_free_io_context,
3496 .elevator_attrs = cfq_attrs,
3497 .elevator_name = "cfq",
3498 .elevator_owner = THIS_MODULE,
3501 static int __init cfq_init(void)
3504 * could be 0 on HZ < 1000 setups
3506 if (!cfq_slice_async)
3507 cfq_slice_async = 1;
3508 if (!cfq_slice_idle)
3511 if (cfq_slab_setup())
3514 elv_register(&iosched_cfq);
3519 static void __exit cfq_exit(void)
3521 DECLARE_COMPLETION_ONSTACK(all_gone);
3522 elv_unregister(&iosched_cfq);
3523 ioc_gone = &all_gone;
3524 /* ioc_gone's update must be visible before reading ioc_count */
3528 * this also protects us from entering cfq_slab_kill() with
3529 * pending RCU callbacks
3531 if (elv_ioc_count_read(cfq_ioc_count))
3532 wait_for_completion(&all_gone);
3536 module_init(cfq_init);
3537 module_exit(cfq_exit);
3539 MODULE_AUTHOR("Jens Axboe");
3540 MODULE_LICENSE("GPL");
3541 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");