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 unsigned int allocated_slice;
121 /* time when first request from queue completed and slice started. */
122 unsigned long slice_start;
123 unsigned long slice_end;
125 unsigned int slice_dispatch;
127 /* pending metadata requests */
129 /* number of requests that are on the dispatch list or inside driver */
132 /* io prio of this group */
133 unsigned short ioprio, org_ioprio;
134 unsigned short ioprio_class, org_ioprio_class;
136 unsigned int seek_samples;
139 sector_t last_request_pos;
140 unsigned long seeky_start;
144 struct cfq_rb_root *service_tree;
145 struct cfq_queue *new_cfqq;
146 struct cfq_group *cfqg;
147 struct cfq_group *orig_cfqg;
148 /* Sectors dispatched in current dispatch round */
149 unsigned long nr_sectors;
153 * First index in the service_trees.
154 * IDLE is handled separately, so it has negative index
163 * Second index in the service_trees.
167 SYNC_NOIDLE_WORKLOAD = 1,
171 /* This is per cgroup per device grouping structure */
173 /* group service_tree member */
174 struct rb_node rb_node;
176 /* group service_tree key */
181 /* number of cfqq currently on this group */
184 /* Per group busy queus average. Useful for workload slice calc. */
185 unsigned int busy_queues_avg[2];
187 * rr lists of queues with requests, onle rr for each priority class.
188 * Counts are embedded in the cfq_rb_root
190 struct cfq_rb_root service_trees[2][3];
191 struct cfq_rb_root service_tree_idle;
193 unsigned long saved_workload_slice;
194 enum wl_type_t saved_workload;
195 enum wl_prio_t saved_serving_prio;
196 struct blkio_group blkg;
197 #ifdef CONFIG_CFQ_GROUP_IOSCHED
198 struct hlist_node cfqd_node;
204 * Per block device queue structure
207 struct request_queue *queue;
208 /* Root service tree for cfq_groups */
209 struct cfq_rb_root grp_service_tree;
210 struct cfq_group root_group;
211 /* Number of active cfq groups on group service tree */
215 * The priority currently being served
217 enum wl_prio_t serving_prio;
218 enum wl_type_t serving_type;
219 unsigned long workload_expires;
220 struct cfq_group *serving_group;
221 bool noidle_tree_requires_idle;
224 * Each priority tree is sorted by next_request position. These
225 * trees are used when determining if two or more queues are
226 * interleaving requests (see cfq_close_cooperator).
228 struct rb_root prio_trees[CFQ_PRIO_LISTS];
230 unsigned int busy_queues;
236 * queue-depth detection
242 * -1 => indeterminate, (cfq will behave as if NCQ is present, to allow better detection)
243 * 1 => NCQ is present (hw_tag_est_depth is the estimated max depth)
246 int hw_tag_est_depth;
247 unsigned int hw_tag_samples;
250 * idle window management
252 struct timer_list idle_slice_timer;
253 struct work_struct unplug_work;
255 struct cfq_queue *active_queue;
256 struct cfq_io_context *active_cic;
259 * async queue for each priority case
261 struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
262 struct cfq_queue *async_idle_cfqq;
264 sector_t last_position;
267 * tunables, see top of file
269 unsigned int cfq_quantum;
270 unsigned int cfq_fifo_expire[2];
271 unsigned int cfq_back_penalty;
272 unsigned int cfq_back_max;
273 unsigned int cfq_slice[2];
274 unsigned int cfq_slice_async_rq;
275 unsigned int cfq_slice_idle;
276 unsigned int cfq_latency;
277 unsigned int cfq_group_isolation;
279 struct list_head cic_list;
282 * Fallback dummy cfqq for extreme OOM conditions
284 struct cfq_queue oom_cfqq;
286 unsigned long last_end_sync_rq;
288 /* List of cfq groups being managed on this device*/
289 struct hlist_head cfqg_list;
292 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd);
294 static struct cfq_rb_root *service_tree_for(struct cfq_group *cfqg,
297 struct cfq_data *cfqd)
302 if (prio == IDLE_WORKLOAD)
303 return &cfqg->service_tree_idle;
305 return &cfqg->service_trees[prio][type];
308 enum cfqq_state_flags {
309 CFQ_CFQQ_FLAG_on_rr = 0, /* on round-robin busy list */
310 CFQ_CFQQ_FLAG_wait_request, /* waiting for a request */
311 CFQ_CFQQ_FLAG_must_dispatch, /* must be allowed a dispatch */
312 CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
313 CFQ_CFQQ_FLAG_fifo_expire, /* FIFO checked in this slice */
314 CFQ_CFQQ_FLAG_idle_window, /* slice idling enabled */
315 CFQ_CFQQ_FLAG_prio_changed, /* task priority has changed */
316 CFQ_CFQQ_FLAG_slice_new, /* no requests dispatched in slice */
317 CFQ_CFQQ_FLAG_sync, /* synchronous queue */
318 CFQ_CFQQ_FLAG_coop, /* cfqq is shared */
319 CFQ_CFQQ_FLAG_deep, /* sync cfqq experienced large depth */
320 CFQ_CFQQ_FLAG_wait_busy, /* Waiting for next request */
321 CFQ_CFQQ_FLAG_wait_busy_done, /* Got new request. Expire the queue */
324 #define CFQ_CFQQ_FNS(name) \
325 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \
327 (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name); \
329 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \
331 (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \
333 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \
335 return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \
339 CFQ_CFQQ_FNS(wait_request);
340 CFQ_CFQQ_FNS(must_dispatch);
341 CFQ_CFQQ_FNS(must_alloc_slice);
342 CFQ_CFQQ_FNS(fifo_expire);
343 CFQ_CFQQ_FNS(idle_window);
344 CFQ_CFQQ_FNS(prio_changed);
345 CFQ_CFQQ_FNS(slice_new);
349 CFQ_CFQQ_FNS(wait_busy);
350 CFQ_CFQQ_FNS(wait_busy_done);
353 #ifdef CONFIG_DEBUG_CFQ_IOSCHED
354 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
355 blk_add_trace_msg((cfqd)->queue, "cfq%d%c %s " fmt, (cfqq)->pid, \
356 cfq_cfqq_sync((cfqq)) ? 'S' : 'A', \
357 blkg_path(&(cfqq)->cfqg->blkg), ##args);
359 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...) \
360 blk_add_trace_msg((cfqd)->queue, "%s " fmt, \
361 blkg_path(&(cfqg)->blkg), ##args); \
364 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
365 blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args)
366 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...) do {} while (0);
368 #define cfq_log(cfqd, fmt, args...) \
369 blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
371 /* Traverses through cfq group service trees */
372 #define for_each_cfqg_st(cfqg, i, j, st) \
373 for (i = 0; i <= IDLE_WORKLOAD; i++) \
374 for (j = 0, st = i < IDLE_WORKLOAD ? &cfqg->service_trees[i][j]\
375 : &cfqg->service_tree_idle; \
376 (i < IDLE_WORKLOAD && j <= SYNC_WORKLOAD) || \
377 (i == IDLE_WORKLOAD && j == 0); \
378 j++, st = i < IDLE_WORKLOAD ? \
379 &cfqg->service_trees[i][j]: NULL) \
382 static inline enum wl_prio_t cfqq_prio(struct cfq_queue *cfqq)
384 if (cfq_class_idle(cfqq))
385 return IDLE_WORKLOAD;
386 if (cfq_class_rt(cfqq))
392 static enum wl_type_t cfqq_type(struct cfq_queue *cfqq)
394 if (!cfq_cfqq_sync(cfqq))
395 return ASYNC_WORKLOAD;
396 if (!cfq_cfqq_idle_window(cfqq))
397 return SYNC_NOIDLE_WORKLOAD;
398 return SYNC_WORKLOAD;
401 static inline int cfq_group_busy_queues_wl(enum wl_prio_t wl,
402 struct cfq_data *cfqd,
403 struct cfq_group *cfqg)
405 if (wl == IDLE_WORKLOAD)
406 return cfqg->service_tree_idle.count;
408 return cfqg->service_trees[wl][ASYNC_WORKLOAD].count
409 + cfqg->service_trees[wl][SYNC_NOIDLE_WORKLOAD].count
410 + cfqg->service_trees[wl][SYNC_WORKLOAD].count;
413 static inline int cfqg_busy_async_queues(struct cfq_data *cfqd,
414 struct cfq_group *cfqg)
416 return cfqg->service_trees[RT_WORKLOAD][ASYNC_WORKLOAD].count
417 + cfqg->service_trees[BE_WORKLOAD][ASYNC_WORKLOAD].count;
420 static void cfq_dispatch_insert(struct request_queue *, struct request *);
421 static struct cfq_queue *cfq_get_queue(struct cfq_data *, bool,
422 struct io_context *, gfp_t);
423 static struct cfq_io_context *cfq_cic_lookup(struct cfq_data *,
424 struct io_context *);
426 static inline int rq_in_driver(struct cfq_data *cfqd)
428 return cfqd->rq_in_driver[0] + cfqd->rq_in_driver[1];
431 static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_context *cic,
434 return cic->cfqq[is_sync];
437 static inline void cic_set_cfqq(struct cfq_io_context *cic,
438 struct cfq_queue *cfqq, bool is_sync)
440 cic->cfqq[is_sync] = cfqq;
444 * We regard a request as SYNC, if it's either a read or has the SYNC bit
445 * set (in which case it could also be direct WRITE).
447 static inline bool cfq_bio_sync(struct bio *bio)
449 return bio_data_dir(bio) == READ || bio_rw_flagged(bio, BIO_RW_SYNCIO);
453 * scheduler run of queue, if there are requests pending and no one in the
454 * driver that will restart queueing
456 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
458 if (cfqd->busy_queues) {
459 cfq_log(cfqd, "schedule dispatch");
460 kblockd_schedule_work(cfqd->queue, &cfqd->unplug_work);
464 static int cfq_queue_empty(struct request_queue *q)
466 struct cfq_data *cfqd = q->elevator->elevator_data;
468 return !cfqd->rq_queued;
472 * Scale schedule slice based on io priority. Use the sync time slice only
473 * if a queue is marked sync and has sync io queued. A sync queue with async
474 * io only, should not get full sync slice length.
476 static inline int cfq_prio_slice(struct cfq_data *cfqd, bool sync,
479 const int base_slice = cfqd->cfq_slice[sync];
481 WARN_ON(prio >= IOPRIO_BE_NR);
483 return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
487 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
489 return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
492 static inline u64 cfq_scale_slice(unsigned long delta, struct cfq_group *cfqg)
494 u64 d = delta << CFQ_SERVICE_SHIFT;
496 d = d * BLKIO_WEIGHT_DEFAULT;
497 do_div(d, cfqg->weight);
501 static inline u64 max_vdisktime(u64 min_vdisktime, u64 vdisktime)
503 s64 delta = (s64)(vdisktime - min_vdisktime);
505 min_vdisktime = vdisktime;
507 return min_vdisktime;
510 static inline u64 min_vdisktime(u64 min_vdisktime, u64 vdisktime)
512 s64 delta = (s64)(vdisktime - min_vdisktime);
514 min_vdisktime = vdisktime;
516 return min_vdisktime;
519 static void update_min_vdisktime(struct cfq_rb_root *st)
521 u64 vdisktime = st->min_vdisktime;
522 struct cfq_group *cfqg;
525 cfqg = rb_entry_cfqg(st->active);
526 vdisktime = cfqg->vdisktime;
530 cfqg = rb_entry_cfqg(st->left);
531 vdisktime = min_vdisktime(vdisktime, cfqg->vdisktime);
534 st->min_vdisktime = max_vdisktime(st->min_vdisktime, vdisktime);
538 * get averaged number of queues of RT/BE priority.
539 * average is updated, with a formula that gives more weight to higher numbers,
540 * to quickly follows sudden increases and decrease slowly
543 static inline unsigned cfq_group_get_avg_queues(struct cfq_data *cfqd,
544 struct cfq_group *cfqg, bool rt)
546 unsigned min_q, max_q;
547 unsigned mult = cfq_hist_divisor - 1;
548 unsigned round = cfq_hist_divisor / 2;
549 unsigned busy = cfq_group_busy_queues_wl(rt, cfqd, cfqg);
551 min_q = min(cfqg->busy_queues_avg[rt], busy);
552 max_q = max(cfqg->busy_queues_avg[rt], busy);
553 cfqg->busy_queues_avg[rt] = (mult * max_q + min_q + round) /
555 return cfqg->busy_queues_avg[rt];
558 static inline unsigned
559 cfq_group_slice(struct cfq_data *cfqd, struct cfq_group *cfqg)
561 struct cfq_rb_root *st = &cfqd->grp_service_tree;
563 return cfq_target_latency * cfqg->weight / st->total_weight;
567 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
569 unsigned slice = cfq_prio_to_slice(cfqd, cfqq);
570 if (cfqd->cfq_latency) {
572 * interested queues (we consider only the ones with the same
573 * priority class in the cfq group)
575 unsigned iq = cfq_group_get_avg_queues(cfqd, cfqq->cfqg,
577 unsigned sync_slice = cfqd->cfq_slice[1];
578 unsigned expect_latency = sync_slice * iq;
579 unsigned group_slice = cfq_group_slice(cfqd, cfqq->cfqg);
581 if (expect_latency > group_slice) {
582 unsigned base_low_slice = 2 * cfqd->cfq_slice_idle;
583 /* scale low_slice according to IO priority
584 * and sync vs async */
586 min(slice, base_low_slice * slice / sync_slice);
587 /* the adapted slice value is scaled to fit all iqs
588 * into the target latency */
589 slice = max(slice * group_slice / expect_latency,
593 cfqq->slice_start = jiffies;
594 cfqq->slice_end = jiffies + slice;
595 cfqq->allocated_slice = slice;
596 cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
600 * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
601 * isn't valid until the first request from the dispatch is activated
602 * and the slice time set.
604 static inline bool cfq_slice_used(struct cfq_queue *cfqq)
606 if (cfq_cfqq_slice_new(cfqq))
608 if (time_before(jiffies, cfqq->slice_end))
615 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
616 * We choose the request that is closest to the head right now. Distance
617 * behind the head is penalized and only allowed to a certain extent.
619 static struct request *
620 cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2, sector_t last)
622 sector_t s1, s2, d1 = 0, d2 = 0;
623 unsigned long back_max;
624 #define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */
625 #define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */
626 unsigned wrap = 0; /* bit mask: requests behind the disk head? */
628 if (rq1 == NULL || rq1 == rq2)
633 if (rq_is_sync(rq1) && !rq_is_sync(rq2))
635 else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
637 if (rq_is_meta(rq1) && !rq_is_meta(rq2))
639 else if (rq_is_meta(rq2) && !rq_is_meta(rq1))
642 s1 = blk_rq_pos(rq1);
643 s2 = blk_rq_pos(rq2);
646 * by definition, 1KiB is 2 sectors
648 back_max = cfqd->cfq_back_max * 2;
651 * Strict one way elevator _except_ in the case where we allow
652 * short backward seeks which are biased as twice the cost of a
653 * similar forward seek.
657 else if (s1 + back_max >= last)
658 d1 = (last - s1) * cfqd->cfq_back_penalty;
660 wrap |= CFQ_RQ1_WRAP;
664 else if (s2 + back_max >= last)
665 d2 = (last - s2) * cfqd->cfq_back_penalty;
667 wrap |= CFQ_RQ2_WRAP;
669 /* Found required data */
672 * By doing switch() on the bit mask "wrap" we avoid having to
673 * check two variables for all permutations: --> faster!
676 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
692 case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
695 * Since both rqs are wrapped,
696 * start with the one that's further behind head
697 * (--> only *one* back seek required),
698 * since back seek takes more time than forward.
708 * The below is leftmost cache rbtree addon
710 static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
712 /* Service tree is empty */
717 root->left = rb_first(&root->rb);
720 return rb_entry(root->left, struct cfq_queue, rb_node);
725 static struct cfq_group *cfq_rb_first_group(struct cfq_rb_root *root)
728 root->left = rb_first(&root->rb);
731 return rb_entry_cfqg(root->left);
736 static void rb_erase_init(struct rb_node *n, struct rb_root *root)
742 static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
746 rb_erase_init(n, &root->rb);
751 * would be nice to take fifo expire time into account as well
753 static struct request *
754 cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
755 struct request *last)
757 struct rb_node *rbnext = rb_next(&last->rb_node);
758 struct rb_node *rbprev = rb_prev(&last->rb_node);
759 struct request *next = NULL, *prev = NULL;
761 BUG_ON(RB_EMPTY_NODE(&last->rb_node));
764 prev = rb_entry_rq(rbprev);
767 next = rb_entry_rq(rbnext);
769 rbnext = rb_first(&cfqq->sort_list);
770 if (rbnext && rbnext != &last->rb_node)
771 next = rb_entry_rq(rbnext);
774 return cfq_choose_req(cfqd, next, prev, blk_rq_pos(last));
777 static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
778 struct cfq_queue *cfqq)
781 * just an approximation, should be ok.
783 return (cfqq->cfqg->nr_cfqq - 1) * (cfq_prio_slice(cfqd, 1, 0) -
784 cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
788 cfqg_key(struct cfq_rb_root *st, struct cfq_group *cfqg)
790 return cfqg->vdisktime - st->min_vdisktime;
794 __cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
796 struct rb_node **node = &st->rb.rb_node;
797 struct rb_node *parent = NULL;
798 struct cfq_group *__cfqg;
799 s64 key = cfqg_key(st, cfqg);
802 while (*node != NULL) {
804 __cfqg = rb_entry_cfqg(parent);
806 if (key < cfqg_key(st, __cfqg))
807 node = &parent->rb_left;
809 node = &parent->rb_right;
815 st->left = &cfqg->rb_node;
817 rb_link_node(&cfqg->rb_node, parent, node);
818 rb_insert_color(&cfqg->rb_node, &st->rb);
822 cfq_group_service_tree_add(struct cfq_data *cfqd, struct cfq_group *cfqg)
824 struct cfq_rb_root *st = &cfqd->grp_service_tree;
825 struct cfq_group *__cfqg;
833 * Currently put the group at the end. Later implement something
834 * so that groups get lesser vtime based on their weights, so that
835 * if group does not loose all if it was not continously backlogged.
837 n = rb_last(&st->rb);
839 __cfqg = rb_entry_cfqg(n);
840 cfqg->vdisktime = __cfqg->vdisktime + CFQ_IDLE_DELAY;
842 cfqg->vdisktime = st->min_vdisktime;
844 __cfq_group_service_tree_add(st, cfqg);
847 st->total_weight += cfqg->weight;
851 cfq_group_service_tree_del(struct cfq_data *cfqd, struct cfq_group *cfqg)
853 struct cfq_rb_root *st = &cfqd->grp_service_tree;
855 if (st->active == &cfqg->rb_node)
858 BUG_ON(cfqg->nr_cfqq < 1);
861 /* If there are other cfq queues under this group, don't delete it */
865 cfq_log_cfqg(cfqd, cfqg, "del_from_rr group");
868 st->total_weight -= cfqg->weight;
869 if (!RB_EMPTY_NODE(&cfqg->rb_node))
870 cfq_rb_erase(&cfqg->rb_node, st);
871 cfqg->saved_workload_slice = 0;
872 blkiocg_update_blkio_group_dequeue_stats(&cfqg->blkg, 1);
875 static inline unsigned int cfq_cfqq_slice_usage(struct cfq_queue *cfqq)
877 unsigned int slice_used;
880 * Queue got expired before even a single request completed or
881 * got expired immediately after first request completion.
883 if (!cfqq->slice_start || cfqq->slice_start == jiffies) {
885 * Also charge the seek time incurred to the group, otherwise
886 * if there are mutiple queues in the group, each can dispatch
887 * a single request on seeky media and cause lots of seek time
888 * and group will never know it.
890 slice_used = max_t(unsigned, (jiffies - cfqq->dispatch_start),
893 slice_used = jiffies - cfqq->slice_start;
894 if (slice_used > cfqq->allocated_slice)
895 slice_used = cfqq->allocated_slice;
898 cfq_log_cfqq(cfqq->cfqd, cfqq, "sl_used=%u sect=%lu", slice_used,
903 static void cfq_group_served(struct cfq_data *cfqd, struct cfq_group *cfqg,
904 struct cfq_queue *cfqq)
906 struct cfq_rb_root *st = &cfqd->grp_service_tree;
907 unsigned int used_sl, charge_sl;
908 int nr_sync = cfqg->nr_cfqq - cfqg_busy_async_queues(cfqd, cfqg)
909 - cfqg->service_tree_idle.count;
912 used_sl = charge_sl = cfq_cfqq_slice_usage(cfqq);
914 if (!cfq_cfqq_sync(cfqq) && !nr_sync)
915 charge_sl = cfqq->allocated_slice;
917 /* Can't update vdisktime while group is on service tree */
918 cfq_rb_erase(&cfqg->rb_node, st);
919 cfqg->vdisktime += cfq_scale_slice(charge_sl, cfqg);
920 __cfq_group_service_tree_add(st, cfqg);
922 /* This group is being expired. Save the context */
923 if (time_after(cfqd->workload_expires, jiffies)) {
924 cfqg->saved_workload_slice = cfqd->workload_expires
926 cfqg->saved_workload = cfqd->serving_type;
927 cfqg->saved_serving_prio = cfqd->serving_prio;
929 cfqg->saved_workload_slice = 0;
931 cfq_log_cfqg(cfqd, cfqg, "served: vt=%llu min_vt=%llu", cfqg->vdisktime,
933 blkiocg_update_blkio_group_stats(&cfqg->blkg, used_sl,
937 #ifdef CONFIG_CFQ_GROUP_IOSCHED
938 static inline struct cfq_group *cfqg_of_blkg(struct blkio_group *blkg)
941 return container_of(blkg, struct cfq_group, blkg);
946 cfq_update_blkio_group_weight(struct blkio_group *blkg, unsigned int weight)
948 cfqg_of_blkg(blkg)->weight = weight;
951 static struct cfq_group *
952 cfq_find_alloc_cfqg(struct cfq_data *cfqd, struct cgroup *cgroup, int create)
954 struct blkio_cgroup *blkcg = cgroup_to_blkio_cgroup(cgroup);
955 struct cfq_group *cfqg = NULL;
958 struct cfq_rb_root *st;
959 struct backing_dev_info *bdi = &cfqd->queue->backing_dev_info;
960 unsigned int major, minor;
962 /* Do we need to take this reference */
963 if (!blkiocg_css_tryget(blkcg))
966 cfqg = cfqg_of_blkg(blkiocg_lookup_group(blkcg, key));
970 cfqg = kzalloc_node(sizeof(*cfqg), GFP_ATOMIC, cfqd->queue->node);
974 cfqg->weight = blkcg->weight;
975 for_each_cfqg_st(cfqg, i, j, st)
977 RB_CLEAR_NODE(&cfqg->rb_node);
980 * Take the initial reference that will be released on destroy
981 * This can be thought of a joint reference by cgroup and
982 * elevator which will be dropped by either elevator exit
983 * or cgroup deletion path depending on who is exiting first.
985 atomic_set(&cfqg->ref, 1);
987 /* Add group onto cgroup list */
988 sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
989 blkiocg_add_blkio_group(blkcg, &cfqg->blkg, (void *)cfqd,
990 MKDEV(major, minor));
992 /* Add group on cfqd list */
993 hlist_add_head(&cfqg->cfqd_node, &cfqd->cfqg_list);
996 blkiocg_css_put(blkcg);
1001 * Search for the cfq group current task belongs to. If create = 1, then also
1002 * create the cfq group if it does not exist. request_queue lock must be held.
1004 static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd, int create)
1006 struct cgroup *cgroup;
1007 struct cfq_group *cfqg = NULL;
1010 cgroup = task_cgroup(current, blkio_subsys_id);
1011 cfqg = cfq_find_alloc_cfqg(cfqd, cgroup, create);
1012 if (!cfqg && create)
1013 cfqg = &cfqd->root_group;
1018 static void cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg)
1020 /* Currently, all async queues are mapped to root group */
1021 if (!cfq_cfqq_sync(cfqq))
1022 cfqg = &cfqq->cfqd->root_group;
1025 /* cfqq reference on cfqg */
1026 atomic_inc(&cfqq->cfqg->ref);
1029 static void cfq_put_cfqg(struct cfq_group *cfqg)
1031 struct cfq_rb_root *st;
1034 BUG_ON(atomic_read(&cfqg->ref) <= 0);
1035 if (!atomic_dec_and_test(&cfqg->ref))
1037 for_each_cfqg_st(cfqg, i, j, st)
1038 BUG_ON(!RB_EMPTY_ROOT(&st->rb) || st->active != NULL);
1042 static void cfq_destroy_cfqg(struct cfq_data *cfqd, struct cfq_group *cfqg)
1044 /* Something wrong if we are trying to remove same group twice */
1045 BUG_ON(hlist_unhashed(&cfqg->cfqd_node));
1047 hlist_del_init(&cfqg->cfqd_node);
1050 * Put the reference taken at the time of creation so that when all
1051 * queues are gone, group can be destroyed.
1056 static void cfq_release_cfq_groups(struct cfq_data *cfqd)
1058 struct hlist_node *pos, *n;
1059 struct cfq_group *cfqg;
1061 hlist_for_each_entry_safe(cfqg, pos, n, &cfqd->cfqg_list, cfqd_node) {
1063 * If cgroup removal path got to blk_group first and removed
1064 * it from cgroup list, then it will take care of destroying
1067 if (!blkiocg_del_blkio_group(&cfqg->blkg))
1068 cfq_destroy_cfqg(cfqd, cfqg);
1073 * Blk cgroup controller notification saying that blkio_group object is being
1074 * delinked as associated cgroup object is going away. That also means that
1075 * no new IO will come in this group. So get rid of this group as soon as
1076 * any pending IO in the group is finished.
1078 * This function is called under rcu_read_lock(). key is the rcu protected
1079 * pointer. That means "key" is a valid cfq_data pointer as long as we are rcu
1082 * "key" was fetched from blkio_group under blkio_cgroup->lock. That means
1083 * it should not be NULL as even if elevator was exiting, cgroup deltion
1084 * path got to it first.
1086 void cfq_unlink_blkio_group(void *key, struct blkio_group *blkg)
1088 unsigned long flags;
1089 struct cfq_data *cfqd = key;
1091 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1092 cfq_destroy_cfqg(cfqd, cfqg_of_blkg(blkg));
1093 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1096 #else /* GROUP_IOSCHED */
1097 static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd, int create)
1099 return &cfqd->root_group;
1102 cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) {
1106 static void cfq_release_cfq_groups(struct cfq_data *cfqd) {}
1107 static inline void cfq_put_cfqg(struct cfq_group *cfqg) {}
1109 #endif /* GROUP_IOSCHED */
1112 * The cfqd->service_trees holds all pending cfq_queue's that have
1113 * requests waiting to be processed. It is sorted in the order that
1114 * we will service the queues.
1116 static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1119 struct rb_node **p, *parent;
1120 struct cfq_queue *__cfqq;
1121 unsigned long rb_key;
1122 struct cfq_rb_root *service_tree;
1125 int group_changed = 0;
1127 #ifdef CONFIG_CFQ_GROUP_IOSCHED
1128 if (!cfqd->cfq_group_isolation
1129 && cfqq_type(cfqq) == SYNC_NOIDLE_WORKLOAD
1130 && cfqq->cfqg && cfqq->cfqg != &cfqd->root_group) {
1131 /* Move this cfq to root group */
1132 cfq_log_cfqq(cfqd, cfqq, "moving to root group");
1133 if (!RB_EMPTY_NODE(&cfqq->rb_node))
1134 cfq_group_service_tree_del(cfqd, cfqq->cfqg);
1135 cfqq->orig_cfqg = cfqq->cfqg;
1136 cfqq->cfqg = &cfqd->root_group;
1137 atomic_inc(&cfqd->root_group.ref);
1139 } else if (!cfqd->cfq_group_isolation
1140 && cfqq_type(cfqq) == SYNC_WORKLOAD && cfqq->orig_cfqg) {
1141 /* cfqq is sequential now needs to go to its original group */
1142 BUG_ON(cfqq->cfqg != &cfqd->root_group);
1143 if (!RB_EMPTY_NODE(&cfqq->rb_node))
1144 cfq_group_service_tree_del(cfqd, cfqq->cfqg);
1145 cfq_put_cfqg(cfqq->cfqg);
1146 cfqq->cfqg = cfqq->orig_cfqg;
1147 cfqq->orig_cfqg = NULL;
1149 cfq_log_cfqq(cfqd, cfqq, "moved to origin group");
1153 service_tree = service_tree_for(cfqq->cfqg, cfqq_prio(cfqq),
1154 cfqq_type(cfqq), cfqd);
1155 if (cfq_class_idle(cfqq)) {
1156 rb_key = CFQ_IDLE_DELAY;
1157 parent = rb_last(&service_tree->rb);
1158 if (parent && parent != &cfqq->rb_node) {
1159 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1160 rb_key += __cfqq->rb_key;
1163 } else if (!add_front) {
1165 * Get our rb key offset. Subtract any residual slice
1166 * value carried from last service. A negative resid
1167 * count indicates slice overrun, and this should position
1168 * the next service time further away in the tree.
1170 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
1171 rb_key -= cfqq->slice_resid;
1172 cfqq->slice_resid = 0;
1175 __cfqq = cfq_rb_first(service_tree);
1176 rb_key += __cfqq ? __cfqq->rb_key : jiffies;
1179 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1182 * same position, nothing more to do
1184 if (rb_key == cfqq->rb_key &&
1185 cfqq->service_tree == service_tree)
1188 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1189 cfqq->service_tree = NULL;
1194 cfqq->service_tree = service_tree;
1195 p = &service_tree->rb.rb_node;
1200 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1203 * sort by key, that represents service time.
1205 if (time_before(rb_key, __cfqq->rb_key))
1208 n = &(*p)->rb_right;
1216 service_tree->left = &cfqq->rb_node;
1218 cfqq->rb_key = rb_key;
1219 rb_link_node(&cfqq->rb_node, parent, p);
1220 rb_insert_color(&cfqq->rb_node, &service_tree->rb);
1221 service_tree->count++;
1222 if ((add_front || !new_cfqq) && !group_changed)
1224 cfq_group_service_tree_add(cfqd, cfqq->cfqg);
1227 static struct cfq_queue *
1228 cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
1229 sector_t sector, struct rb_node **ret_parent,
1230 struct rb_node ***rb_link)
1232 struct rb_node **p, *parent;
1233 struct cfq_queue *cfqq = NULL;
1241 cfqq = rb_entry(parent, struct cfq_queue, p_node);
1244 * Sort strictly based on sector. Smallest to the left,
1245 * largest to the right.
1247 if (sector > blk_rq_pos(cfqq->next_rq))
1248 n = &(*p)->rb_right;
1249 else if (sector < blk_rq_pos(cfqq->next_rq))
1257 *ret_parent = parent;
1263 static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1265 struct rb_node **p, *parent;
1266 struct cfq_queue *__cfqq;
1269 rb_erase(&cfqq->p_node, cfqq->p_root);
1270 cfqq->p_root = NULL;
1273 if (cfq_class_idle(cfqq))
1278 cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
1279 __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
1280 blk_rq_pos(cfqq->next_rq), &parent, &p);
1282 rb_link_node(&cfqq->p_node, parent, p);
1283 rb_insert_color(&cfqq->p_node, cfqq->p_root);
1285 cfqq->p_root = NULL;
1289 * Update cfqq's position in the service tree.
1291 static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1294 * Resorting requires the cfqq to be on the RR list already.
1296 if (cfq_cfqq_on_rr(cfqq)) {
1297 cfq_service_tree_add(cfqd, cfqq, 0);
1298 cfq_prio_tree_add(cfqd, cfqq);
1303 * add to busy list of queues for service, trying to be fair in ordering
1304 * the pending list according to last request service
1306 static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1308 cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
1309 BUG_ON(cfq_cfqq_on_rr(cfqq));
1310 cfq_mark_cfqq_on_rr(cfqq);
1311 cfqd->busy_queues++;
1313 cfq_resort_rr_list(cfqd, cfqq);
1317 * Called when the cfqq no longer has requests pending, remove it from
1320 static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1322 cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
1323 BUG_ON(!cfq_cfqq_on_rr(cfqq));
1324 cfq_clear_cfqq_on_rr(cfqq);
1326 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1327 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1328 cfqq->service_tree = NULL;
1331 rb_erase(&cfqq->p_node, cfqq->p_root);
1332 cfqq->p_root = NULL;
1335 cfq_group_service_tree_del(cfqd, cfqq->cfqg);
1336 BUG_ON(!cfqd->busy_queues);
1337 cfqd->busy_queues--;
1341 * rb tree support functions
1343 static void cfq_del_rq_rb(struct request *rq)
1345 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1346 const int sync = rq_is_sync(rq);
1348 BUG_ON(!cfqq->queued[sync]);
1349 cfqq->queued[sync]--;
1351 elv_rb_del(&cfqq->sort_list, rq);
1353 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) {
1355 * Queue will be deleted from service tree when we actually
1356 * expire it later. Right now just remove it from prio tree
1360 rb_erase(&cfqq->p_node, cfqq->p_root);
1361 cfqq->p_root = NULL;
1366 static void cfq_add_rq_rb(struct request *rq)
1368 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1369 struct cfq_data *cfqd = cfqq->cfqd;
1370 struct request *__alias, *prev;
1372 cfqq->queued[rq_is_sync(rq)]++;
1375 * looks a little odd, but the first insert might return an alias.
1376 * if that happens, put the alias on the dispatch list
1378 while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL)
1379 cfq_dispatch_insert(cfqd->queue, __alias);
1381 if (!cfq_cfqq_on_rr(cfqq))
1382 cfq_add_cfqq_rr(cfqd, cfqq);
1385 * check if this request is a better next-serve candidate
1387 prev = cfqq->next_rq;
1388 cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq, cfqd->last_position);
1391 * adjust priority tree position, if ->next_rq changes
1393 if (prev != cfqq->next_rq)
1394 cfq_prio_tree_add(cfqd, cfqq);
1396 BUG_ON(!cfqq->next_rq);
1399 static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
1401 elv_rb_del(&cfqq->sort_list, rq);
1402 cfqq->queued[rq_is_sync(rq)]--;
1406 static struct request *
1407 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
1409 struct task_struct *tsk = current;
1410 struct cfq_io_context *cic;
1411 struct cfq_queue *cfqq;
1413 cic = cfq_cic_lookup(cfqd, tsk->io_context);
1417 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
1419 sector_t sector = bio->bi_sector + bio_sectors(bio);
1421 return elv_rb_find(&cfqq->sort_list, sector);
1427 static void cfq_activate_request(struct request_queue *q, struct request *rq)
1429 struct cfq_data *cfqd = q->elevator->elevator_data;
1431 cfqd->rq_in_driver[rq_is_sync(rq)]++;
1432 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
1433 rq_in_driver(cfqd));
1435 cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
1438 static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
1440 struct cfq_data *cfqd = q->elevator->elevator_data;
1441 const int sync = rq_is_sync(rq);
1443 WARN_ON(!cfqd->rq_in_driver[sync]);
1444 cfqd->rq_in_driver[sync]--;
1445 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
1446 rq_in_driver(cfqd));
1449 static void cfq_remove_request(struct request *rq)
1451 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1453 if (cfqq->next_rq == rq)
1454 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
1456 list_del_init(&rq->queuelist);
1459 cfqq->cfqd->rq_queued--;
1460 if (rq_is_meta(rq)) {
1461 WARN_ON(!cfqq->meta_pending);
1462 cfqq->meta_pending--;
1466 static int cfq_merge(struct request_queue *q, struct request **req,
1469 struct cfq_data *cfqd = q->elevator->elevator_data;
1470 struct request *__rq;
1472 __rq = cfq_find_rq_fmerge(cfqd, bio);
1473 if (__rq && elv_rq_merge_ok(__rq, bio)) {
1475 return ELEVATOR_FRONT_MERGE;
1478 return ELEVATOR_NO_MERGE;
1481 static void cfq_merged_request(struct request_queue *q, struct request *req,
1484 if (type == ELEVATOR_FRONT_MERGE) {
1485 struct cfq_queue *cfqq = RQ_CFQQ(req);
1487 cfq_reposition_rq_rb(cfqq, req);
1492 cfq_merged_requests(struct request_queue *q, struct request *rq,
1493 struct request *next)
1495 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1497 * reposition in fifo if next is older than rq
1499 if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
1500 time_before(rq_fifo_time(next), rq_fifo_time(rq))) {
1501 list_move(&rq->queuelist, &next->queuelist);
1502 rq_set_fifo_time(rq, rq_fifo_time(next));
1505 if (cfqq->next_rq == next)
1507 cfq_remove_request(next);
1510 static int cfq_allow_merge(struct request_queue *q, struct request *rq,
1513 struct cfq_data *cfqd = q->elevator->elevator_data;
1514 struct cfq_io_context *cic;
1515 struct cfq_queue *cfqq;
1517 /* Deny merge if bio and rq don't belong to same cfq group */
1518 if ((RQ_CFQQ(rq))->cfqg != cfq_get_cfqg(cfqd, 0))
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 void __cfq_set_active_queue(struct cfq_data *cfqd,
1539 struct cfq_queue *cfqq)
1542 cfq_log_cfqq(cfqd, cfqq, "set_active");
1543 cfqq->slice_start = 0;
1544 cfqq->dispatch_start = jiffies;
1545 cfqq->allocated_slice = 0;
1546 cfqq->slice_end = 0;
1547 cfqq->slice_dispatch = 0;
1548 cfqq->nr_sectors = 0;
1550 cfq_clear_cfqq_wait_request(cfqq);
1551 cfq_clear_cfqq_must_dispatch(cfqq);
1552 cfq_clear_cfqq_must_alloc_slice(cfqq);
1553 cfq_clear_cfqq_fifo_expire(cfqq);
1554 cfq_mark_cfqq_slice_new(cfqq);
1556 del_timer(&cfqd->idle_slice_timer);
1559 cfqd->active_queue = cfqq;
1563 * current cfqq expired its slice (or was too idle), select new one
1566 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1569 cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
1571 if (cfq_cfqq_wait_request(cfqq))
1572 del_timer(&cfqd->idle_slice_timer);
1574 cfq_clear_cfqq_wait_request(cfqq);
1575 cfq_clear_cfqq_wait_busy(cfqq);
1576 cfq_clear_cfqq_wait_busy_done(cfqq);
1579 * store what was left of this slice, if the queue idled/timed out
1581 if (timed_out && !cfq_cfqq_slice_new(cfqq)) {
1582 cfqq->slice_resid = cfqq->slice_end - jiffies;
1583 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
1586 cfq_group_served(cfqd, cfqq->cfqg, cfqq);
1588 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
1589 cfq_del_cfqq_rr(cfqd, cfqq);
1591 cfq_resort_rr_list(cfqd, cfqq);
1593 if (cfqq == cfqd->active_queue)
1594 cfqd->active_queue = NULL;
1596 if (&cfqq->cfqg->rb_node == cfqd->grp_service_tree.active)
1597 cfqd->grp_service_tree.active = NULL;
1599 if (cfqd->active_cic) {
1600 put_io_context(cfqd->active_cic->ioc);
1601 cfqd->active_cic = NULL;
1605 static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out)
1607 struct cfq_queue *cfqq = cfqd->active_queue;
1610 __cfq_slice_expired(cfqd, cfqq, timed_out);
1614 * Get next queue for service. Unless we have a queue preemption,
1615 * we'll simply select the first cfqq in the service tree.
1617 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
1619 struct cfq_rb_root *service_tree =
1620 service_tree_for(cfqd->serving_group, cfqd->serving_prio,
1621 cfqd->serving_type, cfqd);
1623 if (!cfqd->rq_queued)
1626 /* There is nothing to dispatch */
1629 if (RB_EMPTY_ROOT(&service_tree->rb))
1631 return cfq_rb_first(service_tree);
1634 static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd)
1636 struct cfq_group *cfqg;
1637 struct cfq_queue *cfqq;
1639 struct cfq_rb_root *st;
1641 if (!cfqd->rq_queued)
1644 cfqg = cfq_get_next_cfqg(cfqd);
1648 for_each_cfqg_st(cfqg, i, j, st)
1649 if ((cfqq = cfq_rb_first(st)) != NULL)
1655 * Get and set a new active queue for service.
1657 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
1658 struct cfq_queue *cfqq)
1661 cfqq = cfq_get_next_queue(cfqd);
1663 __cfq_set_active_queue(cfqd, cfqq);
1667 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
1670 if (blk_rq_pos(rq) >= cfqd->last_position)
1671 return blk_rq_pos(rq) - cfqd->last_position;
1673 return cfqd->last_position - blk_rq_pos(rq);
1676 #define CFQQ_SEEK_THR 8 * 1024
1677 #define CFQQ_SEEKY(cfqq) ((cfqq)->seek_mean > CFQQ_SEEK_THR)
1679 static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1682 sector_t sdist = cfqq->seek_mean;
1684 if (!sample_valid(cfqq->seek_samples))
1685 sdist = CFQQ_SEEK_THR;
1687 return cfq_dist_from_last(cfqd, rq) <= sdist;
1690 static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
1691 struct cfq_queue *cur_cfqq)
1693 struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
1694 struct rb_node *parent, *node;
1695 struct cfq_queue *__cfqq;
1696 sector_t sector = cfqd->last_position;
1698 if (RB_EMPTY_ROOT(root))
1702 * First, if we find a request starting at the end of the last
1703 * request, choose it.
1705 __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
1710 * If the exact sector wasn't found, the parent of the NULL leaf
1711 * will contain the closest sector.
1713 __cfqq = rb_entry(parent, struct cfq_queue, p_node);
1714 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1717 if (blk_rq_pos(__cfqq->next_rq) < sector)
1718 node = rb_next(&__cfqq->p_node);
1720 node = rb_prev(&__cfqq->p_node);
1724 __cfqq = rb_entry(node, struct cfq_queue, p_node);
1725 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1733 * cur_cfqq - passed in so that we don't decide that the current queue is
1734 * closely cooperating with itself.
1736 * So, basically we're assuming that that cur_cfqq has dispatched at least
1737 * one request, and that cfqd->last_position reflects a position on the disk
1738 * associated with the I/O issued by cur_cfqq. I'm not sure this is a valid
1741 static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
1742 struct cfq_queue *cur_cfqq)
1744 struct cfq_queue *cfqq;
1746 if (!cfq_cfqq_sync(cur_cfqq))
1748 if (CFQQ_SEEKY(cur_cfqq))
1752 * We should notice if some of the queues are cooperating, eg
1753 * working closely on the same area of the disk. In that case,
1754 * we can group them together and don't waste time idling.
1756 cfqq = cfqq_close(cfqd, cur_cfqq);
1760 /* If new queue belongs to different cfq_group, don't choose it */
1761 if (cur_cfqq->cfqg != cfqq->cfqg)
1765 * It only makes sense to merge sync queues.
1767 if (!cfq_cfqq_sync(cfqq))
1769 if (CFQQ_SEEKY(cfqq))
1773 * Do not merge queues of different priority classes
1775 if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq))
1782 * Determine whether we should enforce idle window for this queue.
1785 static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1787 enum wl_prio_t prio = cfqq_prio(cfqq);
1788 struct cfq_rb_root *service_tree = cfqq->service_tree;
1790 BUG_ON(!service_tree);
1791 BUG_ON(!service_tree->count);
1793 /* We never do for idle class queues. */
1794 if (prio == IDLE_WORKLOAD)
1797 /* We do for queues that were marked with idle window flag. */
1798 if (cfq_cfqq_idle_window(cfqq) &&
1799 !(blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag))
1803 * Otherwise, we do only if they are the last ones
1804 * in their service tree.
1806 return service_tree->count == 1;
1809 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
1811 struct cfq_queue *cfqq = cfqd->active_queue;
1812 struct cfq_io_context *cic;
1816 * SSD device without seek penalty, disable idling. But only do so
1817 * for devices that support queuing, otherwise we still have a problem
1818 * with sync vs async workloads.
1820 if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
1823 WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
1824 WARN_ON(cfq_cfqq_slice_new(cfqq));
1827 * idle is disabled, either manually or by past process history
1829 if (!cfqd->cfq_slice_idle || !cfq_should_idle(cfqd, cfqq))
1833 * still active requests from this queue, don't idle
1835 if (cfqq->dispatched)
1839 * task has exited, don't wait
1841 cic = cfqd->active_cic;
1842 if (!cic || !atomic_read(&cic->ioc->nr_tasks))
1846 * If our average think time is larger than the remaining time
1847 * slice, then don't idle. This avoids overrunning the allotted
1850 if (sample_valid(cic->ttime_samples) &&
1851 (cfqq->slice_end - jiffies < cic->ttime_mean))
1854 cfq_mark_cfqq_wait_request(cfqq);
1856 sl = cfqd->cfq_slice_idle;
1858 mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
1859 cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu", sl);
1863 * Move request from internal lists to the request queue dispatch list.
1865 static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
1867 struct cfq_data *cfqd = q->elevator->elevator_data;
1868 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1870 cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
1872 cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
1873 cfq_remove_request(rq);
1875 elv_dispatch_sort(q, rq);
1877 if (cfq_cfqq_sync(cfqq))
1878 cfqd->sync_flight++;
1879 cfqq->nr_sectors += blk_rq_sectors(rq);
1883 * return expired entry, or NULL to just start from scratch in rbtree
1885 static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
1887 struct request *rq = NULL;
1889 if (cfq_cfqq_fifo_expire(cfqq))
1892 cfq_mark_cfqq_fifo_expire(cfqq);
1894 if (list_empty(&cfqq->fifo))
1897 rq = rq_entry_fifo(cfqq->fifo.next);
1898 if (time_before(jiffies, rq_fifo_time(rq)))
1901 cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq);
1906 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1908 const int base_rq = cfqd->cfq_slice_async_rq;
1910 WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
1912 return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
1916 * Must be called with the queue_lock held.
1918 static int cfqq_process_refs(struct cfq_queue *cfqq)
1920 int process_refs, io_refs;
1922 io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE];
1923 process_refs = atomic_read(&cfqq->ref) - io_refs;
1924 BUG_ON(process_refs < 0);
1925 return process_refs;
1928 static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq)
1930 int process_refs, new_process_refs;
1931 struct cfq_queue *__cfqq;
1933 /* Avoid a circular list and skip interim queue merges */
1934 while ((__cfqq = new_cfqq->new_cfqq)) {
1940 process_refs = cfqq_process_refs(cfqq);
1942 * If the process for the cfqq has gone away, there is no
1943 * sense in merging the queues.
1945 if (process_refs == 0)
1949 * Merge in the direction of the lesser amount of work.
1951 new_process_refs = cfqq_process_refs(new_cfqq);
1952 if (new_process_refs >= process_refs) {
1953 cfqq->new_cfqq = new_cfqq;
1954 atomic_add(process_refs, &new_cfqq->ref);
1956 new_cfqq->new_cfqq = cfqq;
1957 atomic_add(new_process_refs, &cfqq->ref);
1961 static enum wl_type_t cfq_choose_wl(struct cfq_data *cfqd,
1962 struct cfq_group *cfqg, enum wl_prio_t prio,
1965 struct cfq_queue *queue;
1967 bool key_valid = false;
1968 unsigned long lowest_key = 0;
1969 enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD;
1973 * When priorities switched, we prefer starting
1974 * from SYNC_NOIDLE (first choice), or just SYNC
1977 if (service_tree_for(cfqg, prio, cur_best, cfqd)->count)
1979 cur_best = SYNC_WORKLOAD;
1980 if (service_tree_for(cfqg, prio, cur_best, cfqd)->count)
1983 return ASYNC_WORKLOAD;
1986 for (i = 0; i < 3; ++i) {
1987 /* otherwise, select the one with lowest rb_key */
1988 queue = cfq_rb_first(service_tree_for(cfqg, prio, i, cfqd));
1990 (!key_valid || time_before(queue->rb_key, lowest_key))) {
1991 lowest_key = queue->rb_key;
2000 static void choose_service_tree(struct cfq_data *cfqd, struct cfq_group *cfqg)
2002 enum wl_prio_t previous_prio = cfqd->serving_prio;
2006 struct cfq_rb_root *st;
2007 unsigned group_slice;
2010 cfqd->serving_prio = IDLE_WORKLOAD;
2011 cfqd->workload_expires = jiffies + 1;
2015 /* Choose next priority. RT > BE > IDLE */
2016 if (cfq_group_busy_queues_wl(RT_WORKLOAD, cfqd, cfqg))
2017 cfqd->serving_prio = RT_WORKLOAD;
2018 else if (cfq_group_busy_queues_wl(BE_WORKLOAD, cfqd, cfqg))
2019 cfqd->serving_prio = BE_WORKLOAD;
2021 cfqd->serving_prio = IDLE_WORKLOAD;
2022 cfqd->workload_expires = jiffies + 1;
2027 * For RT and BE, we have to choose also the type
2028 * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
2031 prio_changed = (cfqd->serving_prio != previous_prio);
2032 st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type,
2037 * If priority didn't change, check workload expiration,
2038 * and that we still have other queues ready
2040 if (!prio_changed && count &&
2041 !time_after(jiffies, cfqd->workload_expires))
2044 /* otherwise select new workload type */
2045 cfqd->serving_type =
2046 cfq_choose_wl(cfqd, cfqg, cfqd->serving_prio, prio_changed);
2047 st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type,
2052 * the workload slice is computed as a fraction of target latency
2053 * proportional to the number of queues in that workload, over
2054 * all the queues in the same priority class
2056 group_slice = cfq_group_slice(cfqd, cfqg);
2058 slice = group_slice * count /
2059 max_t(unsigned, cfqg->busy_queues_avg[cfqd->serving_prio],
2060 cfq_group_busy_queues_wl(cfqd->serving_prio, cfqd, cfqg));
2062 if (cfqd->serving_type == ASYNC_WORKLOAD) {
2066 * Async queues are currently system wide. Just taking
2067 * proportion of queues with-in same group will lead to higher
2068 * async ratio system wide as generally root group is going
2069 * to have higher weight. A more accurate thing would be to
2070 * calculate system wide asnc/sync ratio.
2072 tmp = cfq_target_latency * cfqg_busy_async_queues(cfqd, cfqg);
2073 tmp = tmp/cfqd->busy_queues;
2074 slice = min_t(unsigned, slice, tmp);
2076 /* async workload slice is scaled down according to
2077 * the sync/async slice ratio. */
2078 slice = slice * cfqd->cfq_slice[0] / cfqd->cfq_slice[1];
2080 /* sync workload slice is at least 2 * cfq_slice_idle */
2081 slice = max(slice, 2 * cfqd->cfq_slice_idle);
2083 slice = max_t(unsigned, slice, CFQ_MIN_TT);
2084 cfqd->workload_expires = jiffies + slice;
2085 cfqd->noidle_tree_requires_idle = false;
2088 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd)
2090 struct cfq_rb_root *st = &cfqd->grp_service_tree;
2091 struct cfq_group *cfqg;
2093 if (RB_EMPTY_ROOT(&st->rb))
2095 cfqg = cfq_rb_first_group(st);
2096 st->active = &cfqg->rb_node;
2097 update_min_vdisktime(st);
2101 static void cfq_choose_cfqg(struct cfq_data *cfqd)
2103 struct cfq_group *cfqg = cfq_get_next_cfqg(cfqd);
2105 cfqd->serving_group = cfqg;
2107 /* Restore the workload type data */
2108 if (cfqg->saved_workload_slice) {
2109 cfqd->workload_expires = jiffies + cfqg->saved_workload_slice;
2110 cfqd->serving_type = cfqg->saved_workload;
2111 cfqd->serving_prio = cfqg->saved_serving_prio;
2113 choose_service_tree(cfqd, cfqg);
2117 * Select a queue for service. If we have a current active queue,
2118 * check whether to continue servicing it, or retrieve and set a new one.
2120 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
2122 struct cfq_queue *cfqq, *new_cfqq = NULL;
2124 cfqq = cfqd->active_queue;
2128 if (!cfqd->rq_queued)
2131 * The active queue has run out of time, expire it and select new.
2133 if ((cfq_slice_used(cfqq) || cfq_cfqq_wait_busy_done(cfqq))
2134 && !cfq_cfqq_must_dispatch(cfqq))
2138 * The active queue has requests and isn't expired, allow it to
2141 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
2145 * If another queue has a request waiting within our mean seek
2146 * distance, let it run. The expire code will check for close
2147 * cooperators and put the close queue at the front of the service
2148 * tree. If possible, merge the expiring queue with the new cfqq.
2150 new_cfqq = cfq_close_cooperator(cfqd, cfqq);
2152 if (!cfqq->new_cfqq)
2153 cfq_setup_merge(cfqq, new_cfqq);
2158 * No requests pending. If the active queue still has requests in
2159 * flight or is idling for a new request, allow either of these
2160 * conditions to happen (or time out) before selecting a new queue.
2162 if (timer_pending(&cfqd->idle_slice_timer) ||
2163 (cfqq->dispatched && cfq_should_idle(cfqd, cfqq))) {
2169 cfq_slice_expired(cfqd, 0);
2172 * Current queue expired. Check if we have to switch to a new
2176 cfq_choose_cfqg(cfqd);
2178 cfqq = cfq_set_active_queue(cfqd, new_cfqq);
2183 static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
2187 while (cfqq->next_rq) {
2188 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
2192 BUG_ON(!list_empty(&cfqq->fifo));
2194 /* By default cfqq is not expired if it is empty. Do it explicitly */
2195 __cfq_slice_expired(cfqq->cfqd, cfqq, 0);
2200 * Drain our current requests. Used for barriers and when switching
2201 * io schedulers on-the-fly.
2203 static int cfq_forced_dispatch(struct cfq_data *cfqd)
2205 struct cfq_queue *cfqq;
2208 while ((cfqq = cfq_get_next_queue_forced(cfqd)) != NULL)
2209 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
2211 cfq_slice_expired(cfqd, 0);
2212 BUG_ON(cfqd->busy_queues);
2214 cfq_log(cfqd, "forced_dispatch=%d", dispatched);
2218 static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2220 unsigned int max_dispatch;
2223 * Drain async requests before we start sync IO
2225 if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_driver[BLK_RW_ASYNC])
2229 * If this is an async queue and we have sync IO in flight, let it wait
2231 if (cfqd->sync_flight && !cfq_cfqq_sync(cfqq))
2234 max_dispatch = cfqd->cfq_quantum;
2235 if (cfq_class_idle(cfqq))
2239 * Does this cfqq already have too much IO in flight?
2241 if (cfqq->dispatched >= max_dispatch) {
2243 * idle queue must always only have a single IO in flight
2245 if (cfq_class_idle(cfqq))
2249 * We have other queues, don't allow more IO from this one
2251 if (cfqd->busy_queues > 1)
2255 * Sole queue user, no limit
2261 * Async queues must wait a bit before being allowed dispatch.
2262 * We also ramp up the dispatch depth gradually for async IO,
2263 * based on the last sync IO we serviced
2265 if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
2266 unsigned long last_sync = jiffies - cfqd->last_end_sync_rq;
2269 depth = last_sync / cfqd->cfq_slice[1];
2270 if (!depth && !cfqq->dispatched)
2272 if (depth < max_dispatch)
2273 max_dispatch = depth;
2277 * If we're below the current max, allow a dispatch
2279 return cfqq->dispatched < max_dispatch;
2283 * Dispatch a request from cfqq, moving them to the request queue
2286 static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2290 BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
2292 if (!cfq_may_dispatch(cfqd, cfqq))
2296 * follow expired path, else get first next available
2298 rq = cfq_check_fifo(cfqq);
2303 * insert request into driver dispatch list
2305 cfq_dispatch_insert(cfqd->queue, rq);
2307 if (!cfqd->active_cic) {
2308 struct cfq_io_context *cic = RQ_CIC(rq);
2310 atomic_long_inc(&cic->ioc->refcount);
2311 cfqd->active_cic = cic;
2318 * Find the cfqq that we need to service and move a request from that to the
2321 static int cfq_dispatch_requests(struct request_queue *q, int force)
2323 struct cfq_data *cfqd = q->elevator->elevator_data;
2324 struct cfq_queue *cfqq;
2326 if (!cfqd->busy_queues)
2329 if (unlikely(force))
2330 return cfq_forced_dispatch(cfqd);
2332 cfqq = cfq_select_queue(cfqd);
2337 * Dispatch a request from this cfqq, if it is allowed
2339 if (!cfq_dispatch_request(cfqd, cfqq))
2342 cfqq->slice_dispatch++;
2343 cfq_clear_cfqq_must_dispatch(cfqq);
2346 * expire an async queue immediately if it has used up its slice. idle
2347 * queue always expire after 1 dispatch round.
2349 if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
2350 cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
2351 cfq_class_idle(cfqq))) {
2352 cfqq->slice_end = jiffies + 1;
2353 cfq_slice_expired(cfqd, 0);
2356 cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
2361 * task holds one reference to the queue, dropped when task exits. each rq
2362 * in-flight on this queue also holds a reference, dropped when rq is freed.
2364 * Each cfq queue took a reference on the parent group. Drop it now.
2365 * queue lock must be held here.
2367 static void cfq_put_queue(struct cfq_queue *cfqq)
2369 struct cfq_data *cfqd = cfqq->cfqd;
2370 struct cfq_group *cfqg;
2372 BUG_ON(atomic_read(&cfqq->ref) <= 0);
2374 if (!atomic_dec_and_test(&cfqq->ref))
2377 cfq_log_cfqq(cfqd, cfqq, "put_queue");
2378 BUG_ON(rb_first(&cfqq->sort_list));
2379 BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
2382 if (unlikely(cfqd->active_queue == cfqq)) {
2383 __cfq_slice_expired(cfqd, cfqq, 0);
2384 cfq_schedule_dispatch(cfqd);
2387 BUG_ON(cfq_cfqq_on_rr(cfqq));
2388 kmem_cache_free(cfq_pool, cfqq);
2390 if (cfqq->orig_cfqg)
2391 cfq_put_cfqg(cfqq->orig_cfqg);
2395 * Must always be called with the rcu_read_lock() held
2398 __call_for_each_cic(struct io_context *ioc,
2399 void (*func)(struct io_context *, struct cfq_io_context *))
2401 struct cfq_io_context *cic;
2402 struct hlist_node *n;
2404 hlist_for_each_entry_rcu(cic, n, &ioc->cic_list, cic_list)
2409 * Call func for each cic attached to this ioc.
2412 call_for_each_cic(struct io_context *ioc,
2413 void (*func)(struct io_context *, struct cfq_io_context *))
2416 __call_for_each_cic(ioc, func);
2420 static void cfq_cic_free_rcu(struct rcu_head *head)
2422 struct cfq_io_context *cic;
2424 cic = container_of(head, struct cfq_io_context, rcu_head);
2426 kmem_cache_free(cfq_ioc_pool, cic);
2427 elv_ioc_count_dec(cfq_ioc_count);
2431 * CFQ scheduler is exiting, grab exit lock and check
2432 * the pending io context count. If it hits zero,
2433 * complete ioc_gone and set it back to NULL
2435 spin_lock(&ioc_gone_lock);
2436 if (ioc_gone && !elv_ioc_count_read(cfq_ioc_count)) {
2440 spin_unlock(&ioc_gone_lock);
2444 static void cfq_cic_free(struct cfq_io_context *cic)
2446 call_rcu(&cic->rcu_head, cfq_cic_free_rcu);
2449 static void cic_free_func(struct io_context *ioc, struct cfq_io_context *cic)
2451 unsigned long flags;
2453 BUG_ON(!cic->dead_key);
2455 spin_lock_irqsave(&ioc->lock, flags);
2456 radix_tree_delete(&ioc->radix_root, cic->dead_key);
2457 hlist_del_rcu(&cic->cic_list);
2458 spin_unlock_irqrestore(&ioc->lock, flags);
2464 * Must be called with rcu_read_lock() held or preemption otherwise disabled.
2465 * Only two callers of this - ->dtor() which is called with the rcu_read_lock(),
2466 * and ->trim() which is called with the task lock held
2468 static void cfq_free_io_context(struct io_context *ioc)
2471 * ioc->refcount is zero here, or we are called from elv_unregister(),
2472 * so no more cic's are allowed to be linked into this ioc. So it
2473 * should be ok to iterate over the known list, we will see all cic's
2474 * since no new ones are added.
2476 __call_for_each_cic(ioc, cic_free_func);
2479 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2481 struct cfq_queue *__cfqq, *next;
2483 if (unlikely(cfqq == cfqd->active_queue)) {
2484 __cfq_slice_expired(cfqd, cfqq, 0);
2485 cfq_schedule_dispatch(cfqd);
2489 * If this queue was scheduled to merge with another queue, be
2490 * sure to drop the reference taken on that queue (and others in
2491 * the merge chain). See cfq_setup_merge and cfq_merge_cfqqs.
2493 __cfqq = cfqq->new_cfqq;
2495 if (__cfqq == cfqq) {
2496 WARN(1, "cfqq->new_cfqq loop detected\n");
2499 next = __cfqq->new_cfqq;
2500 cfq_put_queue(__cfqq);
2504 cfq_put_queue(cfqq);
2507 static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
2508 struct cfq_io_context *cic)
2510 struct io_context *ioc = cic->ioc;
2512 list_del_init(&cic->queue_list);
2515 * Make sure key == NULL is seen for dead queues
2518 cic->dead_key = (unsigned long) cic->key;
2521 if (ioc->ioc_data == cic)
2522 rcu_assign_pointer(ioc->ioc_data, NULL);
2524 if (cic->cfqq[BLK_RW_ASYNC]) {
2525 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]);
2526 cic->cfqq[BLK_RW_ASYNC] = NULL;
2529 if (cic->cfqq[BLK_RW_SYNC]) {
2530 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]);
2531 cic->cfqq[BLK_RW_SYNC] = NULL;
2535 static void cfq_exit_single_io_context(struct io_context *ioc,
2536 struct cfq_io_context *cic)
2538 struct cfq_data *cfqd = cic->key;
2541 struct request_queue *q = cfqd->queue;
2542 unsigned long flags;
2544 spin_lock_irqsave(q->queue_lock, flags);
2547 * Ensure we get a fresh copy of the ->key to prevent
2548 * race between exiting task and queue
2550 smp_read_barrier_depends();
2552 __cfq_exit_single_io_context(cfqd, cic);
2554 spin_unlock_irqrestore(q->queue_lock, flags);
2559 * The process that ioc belongs to has exited, we need to clean up
2560 * and put the internal structures we have that belongs to that process.
2562 static void cfq_exit_io_context(struct io_context *ioc)
2564 call_for_each_cic(ioc, cfq_exit_single_io_context);
2567 static struct cfq_io_context *
2568 cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
2570 struct cfq_io_context *cic;
2572 cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO,
2575 cic->last_end_request = jiffies;
2576 INIT_LIST_HEAD(&cic->queue_list);
2577 INIT_HLIST_NODE(&cic->cic_list);
2578 cic->dtor = cfq_free_io_context;
2579 cic->exit = cfq_exit_io_context;
2580 elv_ioc_count_inc(cfq_ioc_count);
2586 static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc)
2588 struct task_struct *tsk = current;
2591 if (!cfq_cfqq_prio_changed(cfqq))
2594 ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio);
2595 switch (ioprio_class) {
2597 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
2598 case IOPRIO_CLASS_NONE:
2600 * no prio set, inherit CPU scheduling settings
2602 cfqq->ioprio = task_nice_ioprio(tsk);
2603 cfqq->ioprio_class = task_nice_ioclass(tsk);
2605 case IOPRIO_CLASS_RT:
2606 cfqq->ioprio = task_ioprio(ioc);
2607 cfqq->ioprio_class = IOPRIO_CLASS_RT;
2609 case IOPRIO_CLASS_BE:
2610 cfqq->ioprio = task_ioprio(ioc);
2611 cfqq->ioprio_class = IOPRIO_CLASS_BE;
2613 case IOPRIO_CLASS_IDLE:
2614 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
2616 cfq_clear_cfqq_idle_window(cfqq);
2621 * keep track of original prio settings in case we have to temporarily
2622 * elevate the priority of this queue
2624 cfqq->org_ioprio = cfqq->ioprio;
2625 cfqq->org_ioprio_class = cfqq->ioprio_class;
2626 cfq_clear_cfqq_prio_changed(cfqq);
2629 static void changed_ioprio(struct io_context *ioc, struct cfq_io_context *cic)
2631 struct cfq_data *cfqd = cic->key;
2632 struct cfq_queue *cfqq;
2633 unsigned long flags;
2635 if (unlikely(!cfqd))
2638 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2640 cfqq = cic->cfqq[BLK_RW_ASYNC];
2642 struct cfq_queue *new_cfqq;
2643 new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic->ioc,
2646 cic->cfqq[BLK_RW_ASYNC] = new_cfqq;
2647 cfq_put_queue(cfqq);
2651 cfqq = cic->cfqq[BLK_RW_SYNC];
2653 cfq_mark_cfqq_prio_changed(cfqq);
2655 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2658 static void cfq_ioc_set_ioprio(struct io_context *ioc)
2660 call_for_each_cic(ioc, changed_ioprio);
2661 ioc->ioprio_changed = 0;
2664 static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2665 pid_t pid, bool is_sync)
2667 RB_CLEAR_NODE(&cfqq->rb_node);
2668 RB_CLEAR_NODE(&cfqq->p_node);
2669 INIT_LIST_HEAD(&cfqq->fifo);
2671 atomic_set(&cfqq->ref, 0);
2674 cfq_mark_cfqq_prio_changed(cfqq);
2677 if (!cfq_class_idle(cfqq))
2678 cfq_mark_cfqq_idle_window(cfqq);
2679 cfq_mark_cfqq_sync(cfqq);
2684 #ifdef CONFIG_CFQ_GROUP_IOSCHED
2685 static void changed_cgroup(struct io_context *ioc, struct cfq_io_context *cic)
2687 struct cfq_queue *sync_cfqq = cic_to_cfqq(cic, 1);
2688 struct cfq_data *cfqd = cic->key;
2689 unsigned long flags;
2690 struct request_queue *q;
2692 if (unlikely(!cfqd))
2697 spin_lock_irqsave(q->queue_lock, flags);
2701 * Drop reference to sync queue. A new sync queue will be
2702 * assigned in new group upon arrival of a fresh request.
2704 cfq_log_cfqq(cfqd, sync_cfqq, "changed cgroup");
2705 cic_set_cfqq(cic, NULL, 1);
2706 cfq_put_queue(sync_cfqq);
2709 spin_unlock_irqrestore(q->queue_lock, flags);
2712 static void cfq_ioc_set_cgroup(struct io_context *ioc)
2714 call_for_each_cic(ioc, changed_cgroup);
2715 ioc->cgroup_changed = 0;
2717 #endif /* CONFIG_CFQ_GROUP_IOSCHED */
2719 static struct cfq_queue *
2720 cfq_find_alloc_queue(struct cfq_data *cfqd, bool is_sync,
2721 struct io_context *ioc, gfp_t gfp_mask)
2723 struct cfq_queue *cfqq, *new_cfqq = NULL;
2724 struct cfq_io_context *cic;
2725 struct cfq_group *cfqg;
2728 cfqg = cfq_get_cfqg(cfqd, 1);
2729 cic = cfq_cic_lookup(cfqd, ioc);
2730 /* cic always exists here */
2731 cfqq = cic_to_cfqq(cic, is_sync);
2734 * Always try a new alloc if we fell back to the OOM cfqq
2735 * originally, since it should just be a temporary situation.
2737 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
2742 } else if (gfp_mask & __GFP_WAIT) {
2743 spin_unlock_irq(cfqd->queue->queue_lock);
2744 new_cfqq = kmem_cache_alloc_node(cfq_pool,
2745 gfp_mask | __GFP_ZERO,
2747 spin_lock_irq(cfqd->queue->queue_lock);
2751 cfqq = kmem_cache_alloc_node(cfq_pool,
2752 gfp_mask | __GFP_ZERO,
2757 cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
2758 cfq_init_prio_data(cfqq, ioc);
2759 cfq_link_cfqq_cfqg(cfqq, cfqg);
2760 cfq_log_cfqq(cfqd, cfqq, "alloced");
2762 cfqq = &cfqd->oom_cfqq;
2766 kmem_cache_free(cfq_pool, new_cfqq);
2771 static struct cfq_queue **
2772 cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
2774 switch (ioprio_class) {
2775 case IOPRIO_CLASS_RT:
2776 return &cfqd->async_cfqq[0][ioprio];
2777 case IOPRIO_CLASS_BE:
2778 return &cfqd->async_cfqq[1][ioprio];
2779 case IOPRIO_CLASS_IDLE:
2780 return &cfqd->async_idle_cfqq;
2786 static struct cfq_queue *
2787 cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct io_context *ioc,
2790 const int ioprio = task_ioprio(ioc);
2791 const int ioprio_class = task_ioprio_class(ioc);
2792 struct cfq_queue **async_cfqq = NULL;
2793 struct cfq_queue *cfqq = NULL;
2796 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
2801 cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask);
2804 * pin the queue now that it's allocated, scheduler exit will prune it
2806 if (!is_sync && !(*async_cfqq)) {
2807 atomic_inc(&cfqq->ref);
2811 atomic_inc(&cfqq->ref);
2816 * We drop cfq io contexts lazily, so we may find a dead one.
2819 cfq_drop_dead_cic(struct cfq_data *cfqd, struct io_context *ioc,
2820 struct cfq_io_context *cic)
2822 unsigned long flags;
2824 WARN_ON(!list_empty(&cic->queue_list));
2826 spin_lock_irqsave(&ioc->lock, flags);
2828 BUG_ON(ioc->ioc_data == cic);
2830 radix_tree_delete(&ioc->radix_root, (unsigned long) cfqd);
2831 hlist_del_rcu(&cic->cic_list);
2832 spin_unlock_irqrestore(&ioc->lock, flags);
2837 static struct cfq_io_context *
2838 cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc)
2840 struct cfq_io_context *cic;
2841 unsigned long flags;
2850 * we maintain a last-hit cache, to avoid browsing over the tree
2852 cic = rcu_dereference(ioc->ioc_data);
2853 if (cic && cic->key == cfqd) {
2859 cic = radix_tree_lookup(&ioc->radix_root, (unsigned long) cfqd);
2863 /* ->key must be copied to avoid race with cfq_exit_queue() */
2866 cfq_drop_dead_cic(cfqd, ioc, cic);
2871 spin_lock_irqsave(&ioc->lock, flags);
2872 rcu_assign_pointer(ioc->ioc_data, cic);
2873 spin_unlock_irqrestore(&ioc->lock, flags);
2881 * Add cic into ioc, using cfqd as the search key. This enables us to lookup
2882 * the process specific cfq io context when entered from the block layer.
2883 * Also adds the cic to a per-cfqd list, used when this queue is removed.
2885 static int cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
2886 struct cfq_io_context *cic, gfp_t gfp_mask)
2888 unsigned long flags;
2891 ret = radix_tree_preload(gfp_mask);
2896 spin_lock_irqsave(&ioc->lock, flags);
2897 ret = radix_tree_insert(&ioc->radix_root,
2898 (unsigned long) cfqd, cic);
2900 hlist_add_head_rcu(&cic->cic_list, &ioc->cic_list);
2901 spin_unlock_irqrestore(&ioc->lock, flags);
2903 radix_tree_preload_end();
2906 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2907 list_add(&cic->queue_list, &cfqd->cic_list);
2908 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2913 printk(KERN_ERR "cfq: cic link failed!\n");
2919 * Setup general io context and cfq io context. There can be several cfq
2920 * io contexts per general io context, if this process is doing io to more
2921 * than one device managed by cfq.
2923 static struct cfq_io_context *
2924 cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
2926 struct io_context *ioc = NULL;
2927 struct cfq_io_context *cic;
2929 might_sleep_if(gfp_mask & __GFP_WAIT);
2931 ioc = get_io_context(gfp_mask, cfqd->queue->node);
2935 cic = cfq_cic_lookup(cfqd, ioc);
2939 cic = cfq_alloc_io_context(cfqd, gfp_mask);
2943 if (cfq_cic_link(cfqd, ioc, cic, gfp_mask))
2947 smp_read_barrier_depends();
2948 if (unlikely(ioc->ioprio_changed))
2949 cfq_ioc_set_ioprio(ioc);
2951 #ifdef CONFIG_CFQ_GROUP_IOSCHED
2952 if (unlikely(ioc->cgroup_changed))
2953 cfq_ioc_set_cgroup(ioc);
2959 put_io_context(ioc);
2964 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
2966 unsigned long elapsed = jiffies - cic->last_end_request;
2967 unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
2969 cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
2970 cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
2971 cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
2975 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2981 if (!cfqq->last_request_pos)
2983 else if (cfqq->last_request_pos < blk_rq_pos(rq))
2984 sdist = blk_rq_pos(rq) - cfqq->last_request_pos;
2986 sdist = cfqq->last_request_pos - blk_rq_pos(rq);
2989 * Don't allow the seek distance to get too large from the
2990 * odd fragment, pagein, etc
2992 if (cfqq->seek_samples <= 60) /* second&third seek */
2993 sdist = min(sdist, (cfqq->seek_mean * 4) + 2*1024*1024);
2995 sdist = min(sdist, (cfqq->seek_mean * 4) + 2*1024*64);
2997 cfqq->seek_samples = (7*cfqq->seek_samples + 256) / 8;
2998 cfqq->seek_total = (7*cfqq->seek_total + (u64)256*sdist) / 8;
2999 total = cfqq->seek_total + (cfqq->seek_samples/2);
3000 do_div(total, cfqq->seek_samples);
3001 cfqq->seek_mean = (sector_t)total;
3004 * If this cfqq is shared between multiple processes, check to
3005 * make sure that those processes are still issuing I/Os within
3006 * the mean seek distance. If not, it may be time to break the
3007 * queues apart again.
3009 if (cfq_cfqq_coop(cfqq)) {
3010 if (CFQQ_SEEKY(cfqq) && !cfqq->seeky_start)
3011 cfqq->seeky_start = jiffies;
3012 else if (!CFQQ_SEEKY(cfqq))
3013 cfqq->seeky_start = 0;
3018 * Disable idle window if the process thinks too long or seeks so much that
3022 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3023 struct cfq_io_context *cic)
3025 int old_idle, enable_idle;
3028 * Don't idle for async or idle io prio class
3030 if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
3033 enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
3035 if (cfqq->queued[0] + cfqq->queued[1] >= 4)
3036 cfq_mark_cfqq_deep(cfqq);
3038 if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle ||
3039 (!cfq_cfqq_deep(cfqq) && sample_valid(cfqq->seek_samples)
3040 && CFQQ_SEEKY(cfqq)))
3042 else if (sample_valid(cic->ttime_samples)) {
3043 if (cic->ttime_mean > cfqd->cfq_slice_idle)
3049 if (old_idle != enable_idle) {
3050 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
3052 cfq_mark_cfqq_idle_window(cfqq);
3054 cfq_clear_cfqq_idle_window(cfqq);
3059 * Check if new_cfqq should preempt the currently active queue. Return 0 for
3060 * no or if we aren't sure, a 1 will cause a preempt.
3063 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
3066 struct cfq_queue *cfqq;
3068 cfqq = cfqd->active_queue;
3072 if (cfq_class_idle(new_cfqq))
3075 if (cfq_class_idle(cfqq))
3079 * if the new request is sync, but the currently running queue is
3080 * not, let the sync request have priority.
3082 if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
3085 if (new_cfqq->cfqg != cfqq->cfqg)
3088 if (cfq_slice_used(cfqq))
3091 /* Allow preemption only if we are idling on sync-noidle tree */
3092 if (cfqd->serving_type == SYNC_NOIDLE_WORKLOAD &&
3093 cfqq_type(new_cfqq) == SYNC_NOIDLE_WORKLOAD &&
3094 new_cfqq->service_tree->count == 2 &&
3095 RB_EMPTY_ROOT(&cfqq->sort_list))
3099 * So both queues are sync. Let the new request get disk time if
3100 * it's a metadata request and the current queue is doing regular IO.
3102 if (rq_is_meta(rq) && !cfqq->meta_pending)
3106 * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
3108 if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
3111 if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
3115 * if this request is as-good as one we would expect from the
3116 * current cfqq, let it preempt
3118 if (cfq_rq_close(cfqd, cfqq, rq))
3125 * cfqq preempts the active queue. if we allowed preempt with no slice left,
3126 * let it have half of its nominal slice.
3128 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3130 cfq_log_cfqq(cfqd, cfqq, "preempt");
3131 cfq_slice_expired(cfqd, 1);
3134 * Put the new queue at the front of the of the current list,
3135 * so we know that it will be selected next.
3137 BUG_ON(!cfq_cfqq_on_rr(cfqq));
3139 cfq_service_tree_add(cfqd, cfqq, 1);
3141 cfqq->slice_end = 0;
3142 cfq_mark_cfqq_slice_new(cfqq);
3146 * Called when a new fs request (rq) is added (to cfqq). Check if there's
3147 * something we should do about it
3150 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3153 struct cfq_io_context *cic = RQ_CIC(rq);
3157 cfqq->meta_pending++;
3159 cfq_update_io_thinktime(cfqd, cic);
3160 cfq_update_io_seektime(cfqd, cfqq, rq);
3161 cfq_update_idle_window(cfqd, cfqq, cic);
3163 cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
3165 if (cfqq == cfqd->active_queue) {
3166 if (cfq_cfqq_wait_busy(cfqq)) {
3167 cfq_clear_cfqq_wait_busy(cfqq);
3168 cfq_mark_cfqq_wait_busy_done(cfqq);
3171 * Remember that we saw a request from this process, but
3172 * don't start queuing just yet. Otherwise we risk seeing lots
3173 * of tiny requests, because we disrupt the normal plugging
3174 * and merging. If the request is already larger than a single
3175 * page, let it rip immediately. For that case we assume that
3176 * merging is already done. Ditto for a busy system that
3177 * has other work pending, don't risk delaying until the
3178 * idle timer unplug to continue working.
3180 if (cfq_cfqq_wait_request(cfqq)) {
3181 if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
3182 cfqd->busy_queues > 1) {
3183 del_timer(&cfqd->idle_slice_timer);
3184 __blk_run_queue(cfqd->queue);
3186 cfq_mark_cfqq_must_dispatch(cfqq);
3188 } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
3190 * not the active queue - expire current slice if it is
3191 * idle and has expired it's mean thinktime or this new queue
3192 * has some old slice time left and is of higher priority or
3193 * this new queue is RT and the current one is BE
3195 cfq_preempt_queue(cfqd, cfqq);
3196 __blk_run_queue(cfqd->queue);
3200 static void cfq_insert_request(struct request_queue *q, struct request *rq)
3202 struct cfq_data *cfqd = q->elevator->elevator_data;
3203 struct cfq_queue *cfqq = RQ_CFQQ(rq);
3205 cfq_log_cfqq(cfqd, cfqq, "insert_request");
3206 cfq_init_prio_data(cfqq, RQ_CIC(rq)->ioc);
3208 rq_set_fifo_time(rq, jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)]);
3209 list_add_tail(&rq->queuelist, &cfqq->fifo);
3212 cfq_rq_enqueued(cfqd, cfqq, rq);
3216 * Update hw_tag based on peak queue depth over 50 samples under
3219 static void cfq_update_hw_tag(struct cfq_data *cfqd)
3221 struct cfq_queue *cfqq = cfqd->active_queue;
3223 if (rq_in_driver(cfqd) > cfqd->hw_tag_est_depth)
3224 cfqd->hw_tag_est_depth = rq_in_driver(cfqd);
3226 if (cfqd->hw_tag == 1)
3229 if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
3230 rq_in_driver(cfqd) <= CFQ_HW_QUEUE_MIN)
3234 * If active queue hasn't enough requests and can idle, cfq might not
3235 * dispatch sufficient requests to hardware. Don't zero hw_tag in this
3238 if (cfqq && cfq_cfqq_idle_window(cfqq) &&
3239 cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] <
3240 CFQ_HW_QUEUE_MIN && rq_in_driver(cfqd) < CFQ_HW_QUEUE_MIN)
3243 if (cfqd->hw_tag_samples++ < 50)
3246 if (cfqd->hw_tag_est_depth >= CFQ_HW_QUEUE_MIN)
3252 static void cfq_completed_request(struct request_queue *q, struct request *rq)
3254 struct cfq_queue *cfqq = RQ_CFQQ(rq);
3255 struct cfq_data *cfqd = cfqq->cfqd;
3256 const int sync = rq_is_sync(rq);
3260 cfq_log_cfqq(cfqd, cfqq, "complete rqnoidle %d", !!rq_noidle(rq));
3262 cfq_update_hw_tag(cfqd);
3264 WARN_ON(!cfqd->rq_in_driver[sync]);
3265 WARN_ON(!cfqq->dispatched);
3266 cfqd->rq_in_driver[sync]--;
3269 if (cfq_cfqq_sync(cfqq))
3270 cfqd->sync_flight--;
3273 RQ_CIC(rq)->last_end_request = now;
3274 cfqd->last_end_sync_rq = now;
3278 * If this is the active queue, check if it needs to be expired,
3279 * or if we want to idle in case it has no pending requests.
3281 if (cfqd->active_queue == cfqq) {
3282 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
3284 if (cfq_cfqq_slice_new(cfqq)) {
3285 cfq_set_prio_slice(cfqd, cfqq);
3286 cfq_clear_cfqq_slice_new(cfqq);
3290 * If this queue consumed its slice and this is last queue
3291 * in the group, wait for next request before we expire
3294 if (cfq_slice_used(cfqq) && cfqq->cfqg->nr_cfqq == 1) {
3295 cfqq->slice_end = jiffies + cfqd->cfq_slice_idle;
3296 cfq_mark_cfqq_wait_busy(cfqq);
3300 * Idling is not enabled on:
3302 * - idle-priority queues
3304 * - queues with still some requests queued
3305 * - when there is a close cooperator
3307 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
3308 cfq_slice_expired(cfqd, 1);
3309 else if (sync && cfqq_empty &&
3310 !cfq_close_cooperator(cfqd, cfqq)) {
3311 cfqd->noidle_tree_requires_idle |= !rq_noidle(rq);
3313 * Idling is enabled for SYNC_WORKLOAD.
3314 * SYNC_NOIDLE_WORKLOAD idles at the end of the tree
3315 * only if we processed at least one !rq_noidle request
3317 if (cfqd->serving_type == SYNC_WORKLOAD
3318 || cfqd->noidle_tree_requires_idle
3319 || cfqq->cfqg->nr_cfqq == 1)
3320 cfq_arm_slice_timer(cfqd);
3324 if (!rq_in_driver(cfqd))
3325 cfq_schedule_dispatch(cfqd);
3329 * we temporarily boost lower priority queues if they are holding fs exclusive
3330 * resources. they are boosted to normal prio (CLASS_BE/4)
3332 static void cfq_prio_boost(struct cfq_queue *cfqq)
3334 if (has_fs_excl()) {
3336 * boost idle prio on transactions that would lock out other
3337 * users of the filesystem
3339 if (cfq_class_idle(cfqq))
3340 cfqq->ioprio_class = IOPRIO_CLASS_BE;
3341 if (cfqq->ioprio > IOPRIO_NORM)
3342 cfqq->ioprio = IOPRIO_NORM;
3345 * unboost the queue (if needed)
3347 cfqq->ioprio_class = cfqq->org_ioprio_class;
3348 cfqq->ioprio = cfqq->org_ioprio;
3352 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
3354 if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
3355 cfq_mark_cfqq_must_alloc_slice(cfqq);
3356 return ELV_MQUEUE_MUST;
3359 return ELV_MQUEUE_MAY;
3362 static int cfq_may_queue(struct request_queue *q, int rw)
3364 struct cfq_data *cfqd = q->elevator->elevator_data;
3365 struct task_struct *tsk = current;
3366 struct cfq_io_context *cic;
3367 struct cfq_queue *cfqq;
3370 * don't force setup of a queue from here, as a call to may_queue
3371 * does not necessarily imply that a request actually will be queued.
3372 * so just lookup a possibly existing queue, or return 'may queue'
3375 cic = cfq_cic_lookup(cfqd, tsk->io_context);
3377 return ELV_MQUEUE_MAY;
3379 cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
3381 cfq_init_prio_data(cfqq, cic->ioc);
3382 cfq_prio_boost(cfqq);
3384 return __cfq_may_queue(cfqq);
3387 return ELV_MQUEUE_MAY;
3391 * queue lock held here
3393 static void cfq_put_request(struct request *rq)
3395 struct cfq_queue *cfqq = RQ_CFQQ(rq);
3398 const int rw = rq_data_dir(rq);
3400 BUG_ON(!cfqq->allocated[rw]);
3401 cfqq->allocated[rw]--;
3403 put_io_context(RQ_CIC(rq)->ioc);
3405 rq->elevator_private = NULL;
3406 rq->elevator_private2 = NULL;
3408 cfq_put_queue(cfqq);
3412 static struct cfq_queue *
3413 cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_context *cic,
3414 struct cfq_queue *cfqq)
3416 cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq);
3417 cic_set_cfqq(cic, cfqq->new_cfqq, 1);
3418 cfq_mark_cfqq_coop(cfqq->new_cfqq);
3419 cfq_put_queue(cfqq);
3420 return cic_to_cfqq(cic, 1);
3423 static int should_split_cfqq(struct cfq_queue *cfqq)
3425 if (cfqq->seeky_start &&
3426 time_after(jiffies, cfqq->seeky_start + CFQQ_COOP_TOUT))
3432 * Returns NULL if a new cfqq should be allocated, or the old cfqq if this
3433 * was the last process referring to said cfqq.
3435 static struct cfq_queue *
3436 split_cfqq(struct cfq_io_context *cic, struct cfq_queue *cfqq)
3438 if (cfqq_process_refs(cfqq) == 1) {
3439 cfqq->seeky_start = 0;
3440 cfqq->pid = current->pid;
3441 cfq_clear_cfqq_coop(cfqq);
3445 cic_set_cfqq(cic, NULL, 1);
3446 cfq_put_queue(cfqq);
3450 * Allocate cfq data structures associated with this request.
3453 cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
3455 struct cfq_data *cfqd = q->elevator->elevator_data;
3456 struct cfq_io_context *cic;
3457 const int rw = rq_data_dir(rq);
3458 const bool is_sync = rq_is_sync(rq);
3459 struct cfq_queue *cfqq;
3460 unsigned long flags;
3462 might_sleep_if(gfp_mask & __GFP_WAIT);
3464 cic = cfq_get_io_context(cfqd, gfp_mask);
3466 spin_lock_irqsave(q->queue_lock, flags);
3472 cfqq = cic_to_cfqq(cic, is_sync);
3473 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
3474 cfqq = cfq_get_queue(cfqd, is_sync, cic->ioc, gfp_mask);
3475 cic_set_cfqq(cic, cfqq, is_sync);
3478 * If the queue was seeky for too long, break it apart.
3480 if (cfq_cfqq_coop(cfqq) && should_split_cfqq(cfqq)) {
3481 cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq");
3482 cfqq = split_cfqq(cic, cfqq);
3488 * Check to see if this queue is scheduled to merge with
3489 * another, closely cooperating queue. The merging of
3490 * queues happens here as it must be done in process context.
3491 * The reference on new_cfqq was taken in merge_cfqqs.
3494 cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq);
3497 cfqq->allocated[rw]++;
3498 atomic_inc(&cfqq->ref);
3500 spin_unlock_irqrestore(q->queue_lock, flags);
3502 rq->elevator_private = cic;
3503 rq->elevator_private2 = cfqq;
3508 put_io_context(cic->ioc);
3510 cfq_schedule_dispatch(cfqd);
3511 spin_unlock_irqrestore(q->queue_lock, flags);
3512 cfq_log(cfqd, "set_request fail");
3516 static void cfq_kick_queue(struct work_struct *work)
3518 struct cfq_data *cfqd =
3519 container_of(work, struct cfq_data, unplug_work);
3520 struct request_queue *q = cfqd->queue;
3522 spin_lock_irq(q->queue_lock);
3523 __blk_run_queue(cfqd->queue);
3524 spin_unlock_irq(q->queue_lock);
3528 * Timer running if the active_queue is currently idling inside its time slice
3530 static void cfq_idle_slice_timer(unsigned long data)
3532 struct cfq_data *cfqd = (struct cfq_data *) data;
3533 struct cfq_queue *cfqq;
3534 unsigned long flags;
3537 cfq_log(cfqd, "idle timer fired");
3539 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
3541 cfqq = cfqd->active_queue;
3546 * We saw a request before the queue expired, let it through
3548 if (cfq_cfqq_must_dispatch(cfqq))
3554 if (cfq_slice_used(cfqq))
3558 * only expire and reinvoke request handler, if there are
3559 * other queues with pending requests
3561 if (!cfqd->busy_queues)
3565 * not expired and it has a request pending, let it dispatch
3567 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3571 * Queue depth flag is reset only when the idle didn't succeed
3573 cfq_clear_cfqq_deep(cfqq);
3576 cfq_slice_expired(cfqd, timed_out);
3578 cfq_schedule_dispatch(cfqd);
3580 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
3583 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
3585 del_timer_sync(&cfqd->idle_slice_timer);
3586 cancel_work_sync(&cfqd->unplug_work);
3589 static void cfq_put_async_queues(struct cfq_data *cfqd)
3593 for (i = 0; i < IOPRIO_BE_NR; i++) {
3594 if (cfqd->async_cfqq[0][i])
3595 cfq_put_queue(cfqd->async_cfqq[0][i]);
3596 if (cfqd->async_cfqq[1][i])
3597 cfq_put_queue(cfqd->async_cfqq[1][i]);
3600 if (cfqd->async_idle_cfqq)
3601 cfq_put_queue(cfqd->async_idle_cfqq);
3604 static void cfq_exit_queue(struct elevator_queue *e)
3606 struct cfq_data *cfqd = e->elevator_data;
3607 struct request_queue *q = cfqd->queue;
3609 cfq_shutdown_timer_wq(cfqd);
3611 spin_lock_irq(q->queue_lock);
3613 if (cfqd->active_queue)
3614 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
3616 while (!list_empty(&cfqd->cic_list)) {
3617 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
3618 struct cfq_io_context,
3621 __cfq_exit_single_io_context(cfqd, cic);
3624 cfq_put_async_queues(cfqd);
3625 cfq_release_cfq_groups(cfqd);
3626 blkiocg_del_blkio_group(&cfqd->root_group.blkg);
3628 spin_unlock_irq(q->queue_lock);
3630 cfq_shutdown_timer_wq(cfqd);
3632 /* Wait for cfqg->blkg->key accessors to exit their grace periods. */
3637 static void *cfq_init_queue(struct request_queue *q)
3639 struct cfq_data *cfqd;
3641 struct cfq_group *cfqg;
3642 struct cfq_rb_root *st;
3644 cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
3648 /* Init root service tree */
3649 cfqd->grp_service_tree = CFQ_RB_ROOT;
3651 /* Init root group */
3652 cfqg = &cfqd->root_group;
3653 for_each_cfqg_st(cfqg, i, j, st)
3655 RB_CLEAR_NODE(&cfqg->rb_node);
3657 /* Give preference to root group over other groups */
3658 cfqg->weight = 2*BLKIO_WEIGHT_DEFAULT;
3660 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3662 * Take a reference to root group which we never drop. This is just
3663 * to make sure that cfq_put_cfqg() does not try to kfree root group
3665 atomic_set(&cfqg->ref, 1);
3666 blkiocg_add_blkio_group(&blkio_root_cgroup, &cfqg->blkg, (void *)cfqd,
3670 * Not strictly needed (since RB_ROOT just clears the node and we
3671 * zeroed cfqd on alloc), but better be safe in case someone decides
3672 * to add magic to the rb code
3674 for (i = 0; i < CFQ_PRIO_LISTS; i++)
3675 cfqd->prio_trees[i] = RB_ROOT;
3678 * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
3679 * Grab a permanent reference to it, so that the normal code flow
3680 * will not attempt to free it.
3682 cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
3683 atomic_inc(&cfqd->oom_cfqq.ref);
3684 cfq_link_cfqq_cfqg(&cfqd->oom_cfqq, &cfqd->root_group);
3686 INIT_LIST_HEAD(&cfqd->cic_list);
3690 init_timer(&cfqd->idle_slice_timer);
3691 cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
3692 cfqd->idle_slice_timer.data = (unsigned long) cfqd;
3694 INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
3696 cfqd->cfq_quantum = cfq_quantum;
3697 cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
3698 cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
3699 cfqd->cfq_back_max = cfq_back_max;
3700 cfqd->cfq_back_penalty = cfq_back_penalty;
3701 cfqd->cfq_slice[0] = cfq_slice_async;
3702 cfqd->cfq_slice[1] = cfq_slice_sync;
3703 cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
3704 cfqd->cfq_slice_idle = cfq_slice_idle;
3705 cfqd->cfq_latency = 1;
3706 cfqd->cfq_group_isolation = 0;
3708 cfqd->last_end_sync_rq = jiffies;
3712 static void cfq_slab_kill(void)
3715 * Caller already ensured that pending RCU callbacks are completed,
3716 * so we should have no busy allocations at this point.
3719 kmem_cache_destroy(cfq_pool);
3721 kmem_cache_destroy(cfq_ioc_pool);
3724 static int __init cfq_slab_setup(void)
3726 cfq_pool = KMEM_CACHE(cfq_queue, 0);
3730 cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0);
3741 * sysfs parts below -->
3744 cfq_var_show(unsigned int var, char *page)
3746 return sprintf(page, "%d\n", var);
3750 cfq_var_store(unsigned int *var, const char *page, size_t count)
3752 char *p = (char *) page;
3754 *var = simple_strtoul(p, &p, 10);
3758 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
3759 static ssize_t __FUNC(struct elevator_queue *e, char *page) \
3761 struct cfq_data *cfqd = e->elevator_data; \
3762 unsigned int __data = __VAR; \
3764 __data = jiffies_to_msecs(__data); \
3765 return cfq_var_show(__data, (page)); \
3767 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
3768 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
3769 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
3770 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
3771 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
3772 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
3773 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
3774 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
3775 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
3776 SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
3777 SHOW_FUNCTION(cfq_group_isolation_show, cfqd->cfq_group_isolation, 0);
3778 #undef SHOW_FUNCTION
3780 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
3781 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
3783 struct cfq_data *cfqd = e->elevator_data; \
3784 unsigned int __data; \
3785 int ret = cfq_var_store(&__data, (page), count); \
3786 if (__data < (MIN)) \
3788 else if (__data > (MAX)) \
3791 *(__PTR) = msecs_to_jiffies(__data); \
3793 *(__PTR) = __data; \
3796 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
3797 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
3799 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
3801 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
3802 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
3804 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
3805 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
3806 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
3807 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
3809 STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
3810 STORE_FUNCTION(cfq_group_isolation_store, &cfqd->cfq_group_isolation, 0, 1, 0);
3811 #undef STORE_FUNCTION
3813 #define CFQ_ATTR(name) \
3814 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
3816 static struct elv_fs_entry cfq_attrs[] = {
3818 CFQ_ATTR(fifo_expire_sync),
3819 CFQ_ATTR(fifo_expire_async),
3820 CFQ_ATTR(back_seek_max),
3821 CFQ_ATTR(back_seek_penalty),
3822 CFQ_ATTR(slice_sync),
3823 CFQ_ATTR(slice_async),
3824 CFQ_ATTR(slice_async_rq),
3825 CFQ_ATTR(slice_idle),
3826 CFQ_ATTR(low_latency),
3827 CFQ_ATTR(group_isolation),
3831 static struct elevator_type iosched_cfq = {
3833 .elevator_merge_fn = cfq_merge,
3834 .elevator_merged_fn = cfq_merged_request,
3835 .elevator_merge_req_fn = cfq_merged_requests,
3836 .elevator_allow_merge_fn = cfq_allow_merge,
3837 .elevator_dispatch_fn = cfq_dispatch_requests,
3838 .elevator_add_req_fn = cfq_insert_request,
3839 .elevator_activate_req_fn = cfq_activate_request,
3840 .elevator_deactivate_req_fn = cfq_deactivate_request,
3841 .elevator_queue_empty_fn = cfq_queue_empty,
3842 .elevator_completed_req_fn = cfq_completed_request,
3843 .elevator_former_req_fn = elv_rb_former_request,
3844 .elevator_latter_req_fn = elv_rb_latter_request,
3845 .elevator_set_req_fn = cfq_set_request,
3846 .elevator_put_req_fn = cfq_put_request,
3847 .elevator_may_queue_fn = cfq_may_queue,
3848 .elevator_init_fn = cfq_init_queue,
3849 .elevator_exit_fn = cfq_exit_queue,
3850 .trim = cfq_free_io_context,
3852 .elevator_attrs = cfq_attrs,
3853 .elevator_name = "cfq",
3854 .elevator_owner = THIS_MODULE,
3857 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3858 static struct blkio_policy_type blkio_policy_cfq = {
3860 .blkio_unlink_group_fn = cfq_unlink_blkio_group,
3861 .blkio_update_group_weight_fn = cfq_update_blkio_group_weight,
3865 static struct blkio_policy_type blkio_policy_cfq;
3868 static int __init cfq_init(void)
3871 * could be 0 on HZ < 1000 setups
3873 if (!cfq_slice_async)
3874 cfq_slice_async = 1;
3875 if (!cfq_slice_idle)
3878 if (cfq_slab_setup())
3881 elv_register(&iosched_cfq);
3882 blkio_policy_register(&blkio_policy_cfq);
3887 static void __exit cfq_exit(void)
3889 DECLARE_COMPLETION_ONSTACK(all_gone);
3890 blkio_policy_unregister(&blkio_policy_cfq);
3891 elv_unregister(&iosched_cfq);
3892 ioc_gone = &all_gone;
3893 /* ioc_gone's update must be visible before reading ioc_count */
3897 * this also protects us from entering cfq_slab_kill() with
3898 * pending RCU callbacks
3900 if (elv_ioc_count_read(cfq_ioc_count))
3901 wait_for_completion(&all_gone);
3905 module_init(cfq_init);
3906 module_exit(cfq_exit);
3908 MODULE_AUTHOR("Jens Axboe");
3909 MODULE_LICENSE("GPL");
3910 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");