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 /* Sectors dispatched in current dispatch round */
148 unsigned long nr_sectors;
152 * First index in the service_trees.
153 * IDLE is handled separately, so it has negative index
162 * Second index in the service_trees.
166 SYNC_NOIDLE_WORKLOAD = 1,
170 /* This is per cgroup per device grouping structure */
172 /* group service_tree member */
173 struct rb_node rb_node;
175 /* group service_tree key */
180 /* number of cfqq currently on this group */
183 /* Per group busy queus average. Useful for workload slice calc. */
184 unsigned int busy_queues_avg[2];
186 * rr lists of queues with requests, onle rr for each priority class.
187 * Counts are embedded in the cfq_rb_root
189 struct cfq_rb_root service_trees[2][3];
190 struct cfq_rb_root service_tree_idle;
192 unsigned long saved_workload_slice;
193 enum wl_type_t saved_workload;
194 enum wl_prio_t saved_serving_prio;
195 struct blkio_group blkg;
196 #ifdef CONFIG_CFQ_GROUP_IOSCHED
197 struct hlist_node cfqd_node;
203 * Per block device queue structure
206 struct request_queue *queue;
207 /* Root service tree for cfq_groups */
208 struct cfq_rb_root grp_service_tree;
209 struct cfq_group root_group;
210 /* Number of active cfq groups on group service tree */
214 * The priority currently being served
216 enum wl_prio_t serving_prio;
217 enum wl_type_t serving_type;
218 unsigned long workload_expires;
219 struct cfq_group *serving_group;
220 bool noidle_tree_requires_idle;
223 * Each priority tree is sorted by next_request position. These
224 * trees are used when determining if two or more queues are
225 * interleaving requests (see cfq_close_cooperator).
227 struct rb_root prio_trees[CFQ_PRIO_LISTS];
229 unsigned int busy_queues;
235 * queue-depth detection
241 * -1 => indeterminate, (cfq will behave as if NCQ is present, to allow better detection)
242 * 1 => NCQ is present (hw_tag_est_depth is the estimated max depth)
245 int hw_tag_est_depth;
246 unsigned int hw_tag_samples;
249 * idle window management
251 struct timer_list idle_slice_timer;
252 struct work_struct unplug_work;
254 struct cfq_queue *active_queue;
255 struct cfq_io_context *active_cic;
258 * async queue for each priority case
260 struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
261 struct cfq_queue *async_idle_cfqq;
263 sector_t last_position;
266 * tunables, see top of file
268 unsigned int cfq_quantum;
269 unsigned int cfq_fifo_expire[2];
270 unsigned int cfq_back_penalty;
271 unsigned int cfq_back_max;
272 unsigned int cfq_slice[2];
273 unsigned int cfq_slice_async_rq;
274 unsigned int cfq_slice_idle;
275 unsigned int cfq_latency;
277 struct list_head cic_list;
280 * Fallback dummy cfqq for extreme OOM conditions
282 struct cfq_queue oom_cfqq;
284 unsigned long last_end_sync_rq;
286 /* List of cfq groups being managed on this device*/
287 struct hlist_head cfqg_list;
290 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd);
292 static struct cfq_rb_root *service_tree_for(struct cfq_group *cfqg,
295 struct cfq_data *cfqd)
300 if (prio == IDLE_WORKLOAD)
301 return &cfqg->service_tree_idle;
303 return &cfqg->service_trees[prio][type];
306 enum cfqq_state_flags {
307 CFQ_CFQQ_FLAG_on_rr = 0, /* on round-robin busy list */
308 CFQ_CFQQ_FLAG_wait_request, /* waiting for a request */
309 CFQ_CFQQ_FLAG_must_dispatch, /* must be allowed a dispatch */
310 CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
311 CFQ_CFQQ_FLAG_fifo_expire, /* FIFO checked in this slice */
312 CFQ_CFQQ_FLAG_idle_window, /* slice idling enabled */
313 CFQ_CFQQ_FLAG_prio_changed, /* task priority has changed */
314 CFQ_CFQQ_FLAG_slice_new, /* no requests dispatched in slice */
315 CFQ_CFQQ_FLAG_sync, /* synchronous queue */
316 CFQ_CFQQ_FLAG_coop, /* cfqq is shared */
317 CFQ_CFQQ_FLAG_deep, /* sync cfqq experienced large depth */
318 CFQ_CFQQ_FLAG_wait_busy, /* Waiting for next request */
319 CFQ_CFQQ_FLAG_wait_busy_done, /* Got new request. Expire the queue */
322 #define CFQ_CFQQ_FNS(name) \
323 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \
325 (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name); \
327 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \
329 (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \
331 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \
333 return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \
337 CFQ_CFQQ_FNS(wait_request);
338 CFQ_CFQQ_FNS(must_dispatch);
339 CFQ_CFQQ_FNS(must_alloc_slice);
340 CFQ_CFQQ_FNS(fifo_expire);
341 CFQ_CFQQ_FNS(idle_window);
342 CFQ_CFQQ_FNS(prio_changed);
343 CFQ_CFQQ_FNS(slice_new);
347 CFQ_CFQQ_FNS(wait_busy);
348 CFQ_CFQQ_FNS(wait_busy_done);
351 #ifdef CONFIG_DEBUG_CFQ_IOSCHED
352 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
353 blk_add_trace_msg((cfqd)->queue, "cfq%d%c %s " fmt, (cfqq)->pid, \
354 cfq_cfqq_sync((cfqq)) ? 'S' : 'A', \
355 blkg_path(&(cfqq)->cfqg->blkg), ##args);
357 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...) \
358 blk_add_trace_msg((cfqd)->queue, "%s " fmt, \
359 blkg_path(&(cfqg)->blkg), ##args); \
362 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
363 blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args)
364 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...) do {} while (0);
366 #define cfq_log(cfqd, fmt, args...) \
367 blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
369 /* Traverses through cfq group service trees */
370 #define for_each_cfqg_st(cfqg, i, j, st) \
371 for (i = 0; i <= IDLE_WORKLOAD; i++) \
372 for (j = 0, st = i < IDLE_WORKLOAD ? &cfqg->service_trees[i][j]\
373 : &cfqg->service_tree_idle; \
374 (i < IDLE_WORKLOAD && j <= SYNC_WORKLOAD) || \
375 (i == IDLE_WORKLOAD && j == 0); \
376 j++, st = i < IDLE_WORKLOAD ? \
377 &cfqg->service_trees[i][j]: NULL) \
380 static inline enum wl_prio_t cfqq_prio(struct cfq_queue *cfqq)
382 if (cfq_class_idle(cfqq))
383 return IDLE_WORKLOAD;
384 if (cfq_class_rt(cfqq))
390 static enum wl_type_t cfqq_type(struct cfq_queue *cfqq)
392 if (!cfq_cfqq_sync(cfqq))
393 return ASYNC_WORKLOAD;
394 if (!cfq_cfqq_idle_window(cfqq))
395 return SYNC_NOIDLE_WORKLOAD;
396 return SYNC_WORKLOAD;
399 static inline int cfq_group_busy_queues_wl(enum wl_prio_t wl,
400 struct cfq_data *cfqd,
401 struct cfq_group *cfqg)
403 if (wl == IDLE_WORKLOAD)
404 return cfqg->service_tree_idle.count;
406 return cfqg->service_trees[wl][ASYNC_WORKLOAD].count
407 + cfqg->service_trees[wl][SYNC_NOIDLE_WORKLOAD].count
408 + cfqg->service_trees[wl][SYNC_WORKLOAD].count;
411 static inline int cfqg_busy_async_queues(struct cfq_data *cfqd,
412 struct cfq_group *cfqg)
414 return cfqg->service_trees[RT_WORKLOAD][ASYNC_WORKLOAD].count
415 + cfqg->service_trees[BE_WORKLOAD][ASYNC_WORKLOAD].count;
418 static void cfq_dispatch_insert(struct request_queue *, struct request *);
419 static struct cfq_queue *cfq_get_queue(struct cfq_data *, bool,
420 struct io_context *, gfp_t);
421 static struct cfq_io_context *cfq_cic_lookup(struct cfq_data *,
422 struct io_context *);
424 static inline int rq_in_driver(struct cfq_data *cfqd)
426 return cfqd->rq_in_driver[0] + cfqd->rq_in_driver[1];
429 static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_context *cic,
432 return cic->cfqq[is_sync];
435 static inline void cic_set_cfqq(struct cfq_io_context *cic,
436 struct cfq_queue *cfqq, bool is_sync)
438 cic->cfqq[is_sync] = cfqq;
442 * We regard a request as SYNC, if it's either a read or has the SYNC bit
443 * set (in which case it could also be direct WRITE).
445 static inline bool cfq_bio_sync(struct bio *bio)
447 return bio_data_dir(bio) == READ || bio_rw_flagged(bio, BIO_RW_SYNCIO);
451 * scheduler run of queue, if there are requests pending and no one in the
452 * driver that will restart queueing
454 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
456 if (cfqd->busy_queues) {
457 cfq_log(cfqd, "schedule dispatch");
458 kblockd_schedule_work(cfqd->queue, &cfqd->unplug_work);
462 static int cfq_queue_empty(struct request_queue *q)
464 struct cfq_data *cfqd = q->elevator->elevator_data;
466 return !cfqd->rq_queued;
470 * Scale schedule slice based on io priority. Use the sync time slice only
471 * if a queue is marked sync and has sync io queued. A sync queue with async
472 * io only, should not get full sync slice length.
474 static inline int cfq_prio_slice(struct cfq_data *cfqd, bool sync,
477 const int base_slice = cfqd->cfq_slice[sync];
479 WARN_ON(prio >= IOPRIO_BE_NR);
481 return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
485 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
487 return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
490 static inline u64 cfq_scale_slice(unsigned long delta, struct cfq_group *cfqg)
492 u64 d = delta << CFQ_SERVICE_SHIFT;
494 d = d * BLKIO_WEIGHT_DEFAULT;
495 do_div(d, cfqg->weight);
499 static inline u64 max_vdisktime(u64 min_vdisktime, u64 vdisktime)
501 s64 delta = (s64)(vdisktime - min_vdisktime);
503 min_vdisktime = vdisktime;
505 return min_vdisktime;
508 static inline u64 min_vdisktime(u64 min_vdisktime, u64 vdisktime)
510 s64 delta = (s64)(vdisktime - min_vdisktime);
512 min_vdisktime = vdisktime;
514 return min_vdisktime;
517 static void update_min_vdisktime(struct cfq_rb_root *st)
519 u64 vdisktime = st->min_vdisktime;
520 struct cfq_group *cfqg;
523 cfqg = rb_entry_cfqg(st->active);
524 vdisktime = cfqg->vdisktime;
528 cfqg = rb_entry_cfqg(st->left);
529 vdisktime = min_vdisktime(vdisktime, cfqg->vdisktime);
532 st->min_vdisktime = max_vdisktime(st->min_vdisktime, vdisktime);
536 * get averaged number of queues of RT/BE priority.
537 * average is updated, with a formula that gives more weight to higher numbers,
538 * to quickly follows sudden increases and decrease slowly
541 static inline unsigned cfq_group_get_avg_queues(struct cfq_data *cfqd,
542 struct cfq_group *cfqg, bool rt)
544 unsigned min_q, max_q;
545 unsigned mult = cfq_hist_divisor - 1;
546 unsigned round = cfq_hist_divisor / 2;
547 unsigned busy = cfq_group_busy_queues_wl(rt, cfqd, cfqg);
549 min_q = min(cfqg->busy_queues_avg[rt], busy);
550 max_q = max(cfqg->busy_queues_avg[rt], busy);
551 cfqg->busy_queues_avg[rt] = (mult * max_q + min_q + round) /
553 return cfqg->busy_queues_avg[rt];
556 static inline unsigned
557 cfq_group_slice(struct cfq_data *cfqd, struct cfq_group *cfqg)
559 struct cfq_rb_root *st = &cfqd->grp_service_tree;
561 return cfq_target_latency * cfqg->weight / st->total_weight;
565 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
567 unsigned slice = cfq_prio_to_slice(cfqd, cfqq);
568 if (cfqd->cfq_latency) {
570 * interested queues (we consider only the ones with the same
571 * priority class in the cfq group)
573 unsigned iq = cfq_group_get_avg_queues(cfqd, cfqq->cfqg,
575 unsigned sync_slice = cfqd->cfq_slice[1];
576 unsigned expect_latency = sync_slice * iq;
577 unsigned group_slice = cfq_group_slice(cfqd, cfqq->cfqg);
579 if (expect_latency > group_slice) {
580 unsigned base_low_slice = 2 * cfqd->cfq_slice_idle;
581 /* scale low_slice according to IO priority
582 * and sync vs async */
584 min(slice, base_low_slice * slice / sync_slice);
585 /* the adapted slice value is scaled to fit all iqs
586 * into the target latency */
587 slice = max(slice * group_slice / expect_latency,
591 cfqq->slice_start = jiffies;
592 cfqq->slice_end = jiffies + slice;
593 cfqq->allocated_slice = slice;
594 cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
598 * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
599 * isn't valid until the first request from the dispatch is activated
600 * and the slice time set.
602 static inline bool cfq_slice_used(struct cfq_queue *cfqq)
604 if (cfq_cfqq_slice_new(cfqq))
606 if (time_before(jiffies, cfqq->slice_end))
613 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
614 * We choose the request that is closest to the head right now. Distance
615 * behind the head is penalized and only allowed to a certain extent.
617 static struct request *
618 cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2, sector_t last)
620 sector_t s1, s2, d1 = 0, d2 = 0;
621 unsigned long back_max;
622 #define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */
623 #define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */
624 unsigned wrap = 0; /* bit mask: requests behind the disk head? */
626 if (rq1 == NULL || rq1 == rq2)
631 if (rq_is_sync(rq1) && !rq_is_sync(rq2))
633 else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
635 if (rq_is_meta(rq1) && !rq_is_meta(rq2))
637 else if (rq_is_meta(rq2) && !rq_is_meta(rq1))
640 s1 = blk_rq_pos(rq1);
641 s2 = blk_rq_pos(rq2);
644 * by definition, 1KiB is 2 sectors
646 back_max = cfqd->cfq_back_max * 2;
649 * Strict one way elevator _except_ in the case where we allow
650 * short backward seeks which are biased as twice the cost of a
651 * similar forward seek.
655 else if (s1 + back_max >= last)
656 d1 = (last - s1) * cfqd->cfq_back_penalty;
658 wrap |= CFQ_RQ1_WRAP;
662 else if (s2 + back_max >= last)
663 d2 = (last - s2) * cfqd->cfq_back_penalty;
665 wrap |= CFQ_RQ2_WRAP;
667 /* Found required data */
670 * By doing switch() on the bit mask "wrap" we avoid having to
671 * check two variables for all permutations: --> faster!
674 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
690 case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
693 * Since both rqs are wrapped,
694 * start with the one that's further behind head
695 * (--> only *one* back seek required),
696 * since back seek takes more time than forward.
706 * The below is leftmost cache rbtree addon
708 static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
710 /* Service tree is empty */
715 root->left = rb_first(&root->rb);
718 return rb_entry(root->left, struct cfq_queue, rb_node);
723 static struct cfq_group *cfq_rb_first_group(struct cfq_rb_root *root)
726 root->left = rb_first(&root->rb);
729 return rb_entry_cfqg(root->left);
734 static void rb_erase_init(struct rb_node *n, struct rb_root *root)
740 static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
744 rb_erase_init(n, &root->rb);
749 * would be nice to take fifo expire time into account as well
751 static struct request *
752 cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
753 struct request *last)
755 struct rb_node *rbnext = rb_next(&last->rb_node);
756 struct rb_node *rbprev = rb_prev(&last->rb_node);
757 struct request *next = NULL, *prev = NULL;
759 BUG_ON(RB_EMPTY_NODE(&last->rb_node));
762 prev = rb_entry_rq(rbprev);
765 next = rb_entry_rq(rbnext);
767 rbnext = rb_first(&cfqq->sort_list);
768 if (rbnext && rbnext != &last->rb_node)
769 next = rb_entry_rq(rbnext);
772 return cfq_choose_req(cfqd, next, prev, blk_rq_pos(last));
775 static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
776 struct cfq_queue *cfqq)
779 * just an approximation, should be ok.
781 return (cfqq->cfqg->nr_cfqq - 1) * (cfq_prio_slice(cfqd, 1, 0) -
782 cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
786 cfqg_key(struct cfq_rb_root *st, struct cfq_group *cfqg)
788 return cfqg->vdisktime - st->min_vdisktime;
792 __cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
794 struct rb_node **node = &st->rb.rb_node;
795 struct rb_node *parent = NULL;
796 struct cfq_group *__cfqg;
797 s64 key = cfqg_key(st, cfqg);
800 while (*node != NULL) {
802 __cfqg = rb_entry_cfqg(parent);
804 if (key < cfqg_key(st, __cfqg))
805 node = &parent->rb_left;
807 node = &parent->rb_right;
813 st->left = &cfqg->rb_node;
815 rb_link_node(&cfqg->rb_node, parent, node);
816 rb_insert_color(&cfqg->rb_node, &st->rb);
820 cfq_group_service_tree_add(struct cfq_data *cfqd, struct cfq_group *cfqg)
822 struct cfq_rb_root *st = &cfqd->grp_service_tree;
823 struct cfq_group *__cfqg;
831 * Currently put the group at the end. Later implement something
832 * so that groups get lesser vtime based on their weights, so that
833 * if group does not loose all if it was not continously backlogged.
835 n = rb_last(&st->rb);
837 __cfqg = rb_entry_cfqg(n);
838 cfqg->vdisktime = __cfqg->vdisktime + CFQ_IDLE_DELAY;
840 cfqg->vdisktime = st->min_vdisktime;
842 __cfq_group_service_tree_add(st, cfqg);
845 st->total_weight += cfqg->weight;
849 cfq_group_service_tree_del(struct cfq_data *cfqd, struct cfq_group *cfqg)
851 struct cfq_rb_root *st = &cfqd->grp_service_tree;
853 if (st->active == &cfqg->rb_node)
856 BUG_ON(cfqg->nr_cfqq < 1);
859 /* If there are other cfq queues under this group, don't delete it */
863 cfq_log_cfqg(cfqd, cfqg, "del_from_rr group");
866 st->total_weight -= cfqg->weight;
867 if (!RB_EMPTY_NODE(&cfqg->rb_node))
868 cfq_rb_erase(&cfqg->rb_node, st);
869 cfqg->saved_workload_slice = 0;
870 blkiocg_update_blkio_group_dequeue_stats(&cfqg->blkg, 1);
873 static inline unsigned int cfq_cfqq_slice_usage(struct cfq_queue *cfqq)
875 unsigned int slice_used;
878 * Queue got expired before even a single request completed or
879 * got expired immediately after first request completion.
881 if (!cfqq->slice_start || cfqq->slice_start == jiffies) {
883 * Also charge the seek time incurred to the group, otherwise
884 * if there are mutiple queues in the group, each can dispatch
885 * a single request on seeky media and cause lots of seek time
886 * and group will never know it.
888 slice_used = max_t(unsigned, (jiffies - cfqq->dispatch_start),
891 slice_used = jiffies - cfqq->slice_start;
892 if (slice_used > cfqq->allocated_slice)
893 slice_used = cfqq->allocated_slice;
896 cfq_log_cfqq(cfqq->cfqd, cfqq, "sl_used=%u sect=%lu", slice_used,
901 static void cfq_group_served(struct cfq_data *cfqd, struct cfq_group *cfqg,
902 struct cfq_queue *cfqq)
904 struct cfq_rb_root *st = &cfqd->grp_service_tree;
905 unsigned int used_sl, charge_sl;
906 int nr_sync = cfqg->nr_cfqq - cfqg_busy_async_queues(cfqd, cfqg)
907 - cfqg->service_tree_idle.count;
910 used_sl = charge_sl = cfq_cfqq_slice_usage(cfqq);
912 if (!cfq_cfqq_sync(cfqq) && !nr_sync)
913 charge_sl = cfqq->allocated_slice;
915 /* Can't update vdisktime while group is on service tree */
916 cfq_rb_erase(&cfqg->rb_node, st);
917 cfqg->vdisktime += cfq_scale_slice(charge_sl, cfqg);
918 __cfq_group_service_tree_add(st, cfqg);
920 /* This group is being expired. Save the context */
921 if (time_after(cfqd->workload_expires, jiffies)) {
922 cfqg->saved_workload_slice = cfqd->workload_expires
924 cfqg->saved_workload = cfqd->serving_type;
925 cfqg->saved_serving_prio = cfqd->serving_prio;
927 cfqg->saved_workload_slice = 0;
929 cfq_log_cfqg(cfqd, cfqg, "served: vt=%llu min_vt=%llu", cfqg->vdisktime,
931 blkiocg_update_blkio_group_stats(&cfqg->blkg, used_sl,
935 #ifdef CONFIG_CFQ_GROUP_IOSCHED
936 static inline struct cfq_group *cfqg_of_blkg(struct blkio_group *blkg)
939 return container_of(blkg, struct cfq_group, blkg);
944 cfq_update_blkio_group_weight(struct blkio_group *blkg, unsigned int weight)
946 cfqg_of_blkg(blkg)->weight = weight;
949 static struct cfq_group *
950 cfq_find_alloc_cfqg(struct cfq_data *cfqd, struct cgroup *cgroup, int create)
952 struct blkio_cgroup *blkcg = cgroup_to_blkio_cgroup(cgroup);
953 struct cfq_group *cfqg = NULL;
956 struct cfq_rb_root *st;
957 struct backing_dev_info *bdi = &cfqd->queue->backing_dev_info;
958 unsigned int major, minor;
960 /* Do we need to take this reference */
961 if (!css_tryget(&blkcg->css))
964 cfqg = cfqg_of_blkg(blkiocg_lookup_group(blkcg, key));
968 cfqg = kzalloc_node(sizeof(*cfqg), GFP_ATOMIC, cfqd->queue->node);
972 cfqg->weight = blkcg->weight;
973 for_each_cfqg_st(cfqg, i, j, st)
975 RB_CLEAR_NODE(&cfqg->rb_node);
978 * Take the initial reference that will be released on destroy
979 * This can be thought of a joint reference by cgroup and
980 * elevator which will be dropped by either elevator exit
981 * or cgroup deletion path depending on who is exiting first.
983 atomic_set(&cfqg->ref, 1);
985 /* Add group onto cgroup list */
986 sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
987 blkiocg_add_blkio_group(blkcg, &cfqg->blkg, (void *)cfqd,
988 MKDEV(major, minor));
990 /* Add group on cfqd list */
991 hlist_add_head(&cfqg->cfqd_node, &cfqd->cfqg_list);
994 css_put(&blkcg->css);
999 * Search for the cfq group current task belongs to. If create = 1, then also
1000 * create the cfq group if it does not exist. request_queue lock must be held.
1002 static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd, int create)
1004 struct cgroup *cgroup;
1005 struct cfq_group *cfqg = NULL;
1008 cgroup = task_cgroup(current, blkio_subsys_id);
1009 cfqg = cfq_find_alloc_cfqg(cfqd, cgroup, create);
1010 if (!cfqg && create)
1011 cfqg = &cfqd->root_group;
1016 static void cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg)
1018 /* Currently, all async queues are mapped to root group */
1019 if (!cfq_cfqq_sync(cfqq))
1020 cfqg = &cfqq->cfqd->root_group;
1023 /* cfqq reference on cfqg */
1024 atomic_inc(&cfqq->cfqg->ref);
1027 static void cfq_put_cfqg(struct cfq_group *cfqg)
1029 struct cfq_rb_root *st;
1032 BUG_ON(atomic_read(&cfqg->ref) <= 0);
1033 if (!atomic_dec_and_test(&cfqg->ref))
1035 for_each_cfqg_st(cfqg, i, j, st)
1036 BUG_ON(!RB_EMPTY_ROOT(&st->rb) || st->active != NULL);
1040 static void cfq_destroy_cfqg(struct cfq_data *cfqd, struct cfq_group *cfqg)
1042 /* Something wrong if we are trying to remove same group twice */
1043 BUG_ON(hlist_unhashed(&cfqg->cfqd_node));
1045 hlist_del_init(&cfqg->cfqd_node);
1048 * Put the reference taken at the time of creation so that when all
1049 * queues are gone, group can be destroyed.
1054 static void cfq_release_cfq_groups(struct cfq_data *cfqd)
1056 struct hlist_node *pos, *n;
1057 struct cfq_group *cfqg;
1059 hlist_for_each_entry_safe(cfqg, pos, n, &cfqd->cfqg_list, cfqd_node) {
1061 * If cgroup removal path got to blk_group first and removed
1062 * it from cgroup list, then it will take care of destroying
1065 if (!blkiocg_del_blkio_group(&cfqg->blkg))
1066 cfq_destroy_cfqg(cfqd, cfqg);
1071 * Blk cgroup controller notification saying that blkio_group object is being
1072 * delinked as associated cgroup object is going away. That also means that
1073 * no new IO will come in this group. So get rid of this group as soon as
1074 * any pending IO in the group is finished.
1076 * This function is called under rcu_read_lock(). key is the rcu protected
1077 * pointer. That means "key" is a valid cfq_data pointer as long as we are rcu
1080 * "key" was fetched from blkio_group under blkio_cgroup->lock. That means
1081 * it should not be NULL as even if elevator was exiting, cgroup deltion
1082 * path got to it first.
1084 void cfq_unlink_blkio_group(void *key, struct blkio_group *blkg)
1086 unsigned long flags;
1087 struct cfq_data *cfqd = key;
1089 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1090 cfq_destroy_cfqg(cfqd, cfqg_of_blkg(blkg));
1091 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1094 #else /* GROUP_IOSCHED */
1095 static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd, int create)
1097 return &cfqd->root_group;
1100 cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) {
1104 static void cfq_release_cfq_groups(struct cfq_data *cfqd) {}
1105 static inline void cfq_put_cfqg(struct cfq_group *cfqg) {}
1107 #endif /* GROUP_IOSCHED */
1110 * The cfqd->service_trees holds all pending cfq_queue's that have
1111 * requests waiting to be processed. It is sorted in the order that
1112 * we will service the queues.
1114 static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1117 struct rb_node **p, *parent;
1118 struct cfq_queue *__cfqq;
1119 unsigned long rb_key;
1120 struct cfq_rb_root *service_tree;
1124 service_tree = service_tree_for(cfqq->cfqg, cfqq_prio(cfqq),
1125 cfqq_type(cfqq), cfqd);
1126 if (cfq_class_idle(cfqq)) {
1127 rb_key = CFQ_IDLE_DELAY;
1128 parent = rb_last(&service_tree->rb);
1129 if (parent && parent != &cfqq->rb_node) {
1130 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1131 rb_key += __cfqq->rb_key;
1134 } else if (!add_front) {
1136 * Get our rb key offset. Subtract any residual slice
1137 * value carried from last service. A negative resid
1138 * count indicates slice overrun, and this should position
1139 * the next service time further away in the tree.
1141 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
1142 rb_key -= cfqq->slice_resid;
1143 cfqq->slice_resid = 0;
1146 __cfqq = cfq_rb_first(service_tree);
1147 rb_key += __cfqq ? __cfqq->rb_key : jiffies;
1150 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1153 * same position, nothing more to do
1155 if (rb_key == cfqq->rb_key &&
1156 cfqq->service_tree == service_tree)
1159 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1160 cfqq->service_tree = NULL;
1165 cfqq->service_tree = service_tree;
1166 p = &service_tree->rb.rb_node;
1171 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1174 * sort by key, that represents service time.
1176 if (time_before(rb_key, __cfqq->rb_key))
1179 n = &(*p)->rb_right;
1187 service_tree->left = &cfqq->rb_node;
1189 cfqq->rb_key = rb_key;
1190 rb_link_node(&cfqq->rb_node, parent, p);
1191 rb_insert_color(&cfqq->rb_node, &service_tree->rb);
1192 service_tree->count++;
1193 if (add_front || !new_cfqq)
1195 cfq_group_service_tree_add(cfqd, cfqq->cfqg);
1198 static struct cfq_queue *
1199 cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
1200 sector_t sector, struct rb_node **ret_parent,
1201 struct rb_node ***rb_link)
1203 struct rb_node **p, *parent;
1204 struct cfq_queue *cfqq = NULL;
1212 cfqq = rb_entry(parent, struct cfq_queue, p_node);
1215 * Sort strictly based on sector. Smallest to the left,
1216 * largest to the right.
1218 if (sector > blk_rq_pos(cfqq->next_rq))
1219 n = &(*p)->rb_right;
1220 else if (sector < blk_rq_pos(cfqq->next_rq))
1228 *ret_parent = parent;
1234 static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1236 struct rb_node **p, *parent;
1237 struct cfq_queue *__cfqq;
1240 rb_erase(&cfqq->p_node, cfqq->p_root);
1241 cfqq->p_root = NULL;
1244 if (cfq_class_idle(cfqq))
1249 cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
1250 __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
1251 blk_rq_pos(cfqq->next_rq), &parent, &p);
1253 rb_link_node(&cfqq->p_node, parent, p);
1254 rb_insert_color(&cfqq->p_node, cfqq->p_root);
1256 cfqq->p_root = NULL;
1260 * Update cfqq's position in the service tree.
1262 static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1265 * Resorting requires the cfqq to be on the RR list already.
1267 if (cfq_cfqq_on_rr(cfqq)) {
1268 cfq_service_tree_add(cfqd, cfqq, 0);
1269 cfq_prio_tree_add(cfqd, cfqq);
1274 * add to busy list of queues for service, trying to be fair in ordering
1275 * the pending list according to last request service
1277 static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1279 cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
1280 BUG_ON(cfq_cfqq_on_rr(cfqq));
1281 cfq_mark_cfqq_on_rr(cfqq);
1282 cfqd->busy_queues++;
1284 cfq_resort_rr_list(cfqd, cfqq);
1288 * Called when the cfqq no longer has requests pending, remove it from
1291 static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1293 cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
1294 BUG_ON(!cfq_cfqq_on_rr(cfqq));
1295 cfq_clear_cfqq_on_rr(cfqq);
1297 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1298 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1299 cfqq->service_tree = NULL;
1302 rb_erase(&cfqq->p_node, cfqq->p_root);
1303 cfqq->p_root = NULL;
1306 cfq_group_service_tree_del(cfqd, cfqq->cfqg);
1307 BUG_ON(!cfqd->busy_queues);
1308 cfqd->busy_queues--;
1312 * rb tree support functions
1314 static void cfq_del_rq_rb(struct request *rq)
1316 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1317 const int sync = rq_is_sync(rq);
1319 BUG_ON(!cfqq->queued[sync]);
1320 cfqq->queued[sync]--;
1322 elv_rb_del(&cfqq->sort_list, rq);
1324 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) {
1326 * Queue will be deleted from service tree when we actually
1327 * expire it later. Right now just remove it from prio tree
1331 rb_erase(&cfqq->p_node, cfqq->p_root);
1332 cfqq->p_root = NULL;
1337 static void cfq_add_rq_rb(struct request *rq)
1339 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1340 struct cfq_data *cfqd = cfqq->cfqd;
1341 struct request *__alias, *prev;
1343 cfqq->queued[rq_is_sync(rq)]++;
1346 * looks a little odd, but the first insert might return an alias.
1347 * if that happens, put the alias on the dispatch list
1349 while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL)
1350 cfq_dispatch_insert(cfqd->queue, __alias);
1352 if (!cfq_cfqq_on_rr(cfqq))
1353 cfq_add_cfqq_rr(cfqd, cfqq);
1356 * check if this request is a better next-serve candidate
1358 prev = cfqq->next_rq;
1359 cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq, cfqd->last_position);
1362 * adjust priority tree position, if ->next_rq changes
1364 if (prev != cfqq->next_rq)
1365 cfq_prio_tree_add(cfqd, cfqq);
1367 BUG_ON(!cfqq->next_rq);
1370 static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
1372 elv_rb_del(&cfqq->sort_list, rq);
1373 cfqq->queued[rq_is_sync(rq)]--;
1377 static struct request *
1378 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
1380 struct task_struct *tsk = current;
1381 struct cfq_io_context *cic;
1382 struct cfq_queue *cfqq;
1384 cic = cfq_cic_lookup(cfqd, tsk->io_context);
1388 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
1390 sector_t sector = bio->bi_sector + bio_sectors(bio);
1392 return elv_rb_find(&cfqq->sort_list, sector);
1398 static void cfq_activate_request(struct request_queue *q, struct request *rq)
1400 struct cfq_data *cfqd = q->elevator->elevator_data;
1402 cfqd->rq_in_driver[rq_is_sync(rq)]++;
1403 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
1404 rq_in_driver(cfqd));
1406 cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
1409 static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
1411 struct cfq_data *cfqd = q->elevator->elevator_data;
1412 const int sync = rq_is_sync(rq);
1414 WARN_ON(!cfqd->rq_in_driver[sync]);
1415 cfqd->rq_in_driver[sync]--;
1416 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
1417 rq_in_driver(cfqd));
1420 static void cfq_remove_request(struct request *rq)
1422 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1424 if (cfqq->next_rq == rq)
1425 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
1427 list_del_init(&rq->queuelist);
1430 cfqq->cfqd->rq_queued--;
1431 if (rq_is_meta(rq)) {
1432 WARN_ON(!cfqq->meta_pending);
1433 cfqq->meta_pending--;
1437 static int cfq_merge(struct request_queue *q, struct request **req,
1440 struct cfq_data *cfqd = q->elevator->elevator_data;
1441 struct request *__rq;
1443 __rq = cfq_find_rq_fmerge(cfqd, bio);
1444 if (__rq && elv_rq_merge_ok(__rq, bio)) {
1446 return ELEVATOR_FRONT_MERGE;
1449 return ELEVATOR_NO_MERGE;
1452 static void cfq_merged_request(struct request_queue *q, struct request *req,
1455 if (type == ELEVATOR_FRONT_MERGE) {
1456 struct cfq_queue *cfqq = RQ_CFQQ(req);
1458 cfq_reposition_rq_rb(cfqq, req);
1463 cfq_merged_requests(struct request_queue *q, struct request *rq,
1464 struct request *next)
1466 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1468 * reposition in fifo if next is older than rq
1470 if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
1471 time_before(rq_fifo_time(next), rq_fifo_time(rq))) {
1472 list_move(&rq->queuelist, &next->queuelist);
1473 rq_set_fifo_time(rq, rq_fifo_time(next));
1476 if (cfqq->next_rq == next)
1478 cfq_remove_request(next);
1481 static int cfq_allow_merge(struct request_queue *q, struct request *rq,
1484 struct cfq_data *cfqd = q->elevator->elevator_data;
1485 struct cfq_io_context *cic;
1486 struct cfq_queue *cfqq;
1488 /* Deny merge if bio and rq don't belong to same cfq group */
1489 if ((RQ_CFQQ(rq))->cfqg != cfq_get_cfqg(cfqd, 0))
1492 * Disallow merge of a sync bio into an async request.
1494 if (cfq_bio_sync(bio) && !rq_is_sync(rq))
1498 * Lookup the cfqq that this bio will be queued with. Allow
1499 * merge only if rq is queued there.
1501 cic = cfq_cic_lookup(cfqd, current->io_context);
1505 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
1506 return cfqq == RQ_CFQQ(rq);
1509 static void __cfq_set_active_queue(struct cfq_data *cfqd,
1510 struct cfq_queue *cfqq)
1513 cfq_log_cfqq(cfqd, cfqq, "set_active");
1514 cfqq->slice_start = 0;
1515 cfqq->dispatch_start = jiffies;
1516 cfqq->allocated_slice = 0;
1517 cfqq->slice_end = 0;
1518 cfqq->slice_dispatch = 0;
1519 cfqq->nr_sectors = 0;
1521 cfq_clear_cfqq_wait_request(cfqq);
1522 cfq_clear_cfqq_must_dispatch(cfqq);
1523 cfq_clear_cfqq_must_alloc_slice(cfqq);
1524 cfq_clear_cfqq_fifo_expire(cfqq);
1525 cfq_mark_cfqq_slice_new(cfqq);
1527 del_timer(&cfqd->idle_slice_timer);
1530 cfqd->active_queue = cfqq;
1534 * current cfqq expired its slice (or was too idle), select new one
1537 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1540 cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
1542 if (cfq_cfqq_wait_request(cfqq))
1543 del_timer(&cfqd->idle_slice_timer);
1545 cfq_clear_cfqq_wait_request(cfqq);
1546 cfq_clear_cfqq_wait_busy(cfqq);
1547 cfq_clear_cfqq_wait_busy_done(cfqq);
1550 * store what was left of this slice, if the queue idled/timed out
1552 if (timed_out && !cfq_cfqq_slice_new(cfqq)) {
1553 cfqq->slice_resid = cfqq->slice_end - jiffies;
1554 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
1557 cfq_group_served(cfqd, cfqq->cfqg, cfqq);
1559 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
1560 cfq_del_cfqq_rr(cfqd, cfqq);
1562 cfq_resort_rr_list(cfqd, cfqq);
1564 if (cfqq == cfqd->active_queue)
1565 cfqd->active_queue = NULL;
1567 if (&cfqq->cfqg->rb_node == cfqd->grp_service_tree.active)
1568 cfqd->grp_service_tree.active = NULL;
1570 if (cfqd->active_cic) {
1571 put_io_context(cfqd->active_cic->ioc);
1572 cfqd->active_cic = NULL;
1576 static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out)
1578 struct cfq_queue *cfqq = cfqd->active_queue;
1581 __cfq_slice_expired(cfqd, cfqq, timed_out);
1585 * Get next queue for service. Unless we have a queue preemption,
1586 * we'll simply select the first cfqq in the service tree.
1588 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
1590 struct cfq_rb_root *service_tree =
1591 service_tree_for(cfqd->serving_group, cfqd->serving_prio,
1592 cfqd->serving_type, cfqd);
1594 if (!cfqd->rq_queued)
1597 /* There is nothing to dispatch */
1600 if (RB_EMPTY_ROOT(&service_tree->rb))
1602 return cfq_rb_first(service_tree);
1605 static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd)
1607 struct cfq_group *cfqg;
1608 struct cfq_queue *cfqq;
1610 struct cfq_rb_root *st;
1612 if (!cfqd->rq_queued)
1615 cfqg = cfq_get_next_cfqg(cfqd);
1619 for_each_cfqg_st(cfqg, i, j, st)
1620 if ((cfqq = cfq_rb_first(st)) != NULL)
1626 * Get and set a new active queue for service.
1628 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
1629 struct cfq_queue *cfqq)
1632 cfqq = cfq_get_next_queue(cfqd);
1634 __cfq_set_active_queue(cfqd, cfqq);
1638 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
1641 if (blk_rq_pos(rq) >= cfqd->last_position)
1642 return blk_rq_pos(rq) - cfqd->last_position;
1644 return cfqd->last_position - blk_rq_pos(rq);
1647 #define CFQQ_SEEK_THR 8 * 1024
1648 #define CFQQ_SEEKY(cfqq) ((cfqq)->seek_mean > CFQQ_SEEK_THR)
1650 static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1653 sector_t sdist = cfqq->seek_mean;
1655 if (!sample_valid(cfqq->seek_samples))
1656 sdist = CFQQ_SEEK_THR;
1658 return cfq_dist_from_last(cfqd, rq) <= sdist;
1661 static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
1662 struct cfq_queue *cur_cfqq)
1664 struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
1665 struct rb_node *parent, *node;
1666 struct cfq_queue *__cfqq;
1667 sector_t sector = cfqd->last_position;
1669 if (RB_EMPTY_ROOT(root))
1673 * First, if we find a request starting at the end of the last
1674 * request, choose it.
1676 __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
1681 * If the exact sector wasn't found, the parent of the NULL leaf
1682 * will contain the closest sector.
1684 __cfqq = rb_entry(parent, struct cfq_queue, p_node);
1685 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1688 if (blk_rq_pos(__cfqq->next_rq) < sector)
1689 node = rb_next(&__cfqq->p_node);
1691 node = rb_prev(&__cfqq->p_node);
1695 __cfqq = rb_entry(node, struct cfq_queue, p_node);
1696 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1704 * cur_cfqq - passed in so that we don't decide that the current queue is
1705 * closely cooperating with itself.
1707 * So, basically we're assuming that that cur_cfqq has dispatched at least
1708 * one request, and that cfqd->last_position reflects a position on the disk
1709 * associated with the I/O issued by cur_cfqq. I'm not sure this is a valid
1712 static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
1713 struct cfq_queue *cur_cfqq)
1715 struct cfq_queue *cfqq;
1717 if (!cfq_cfqq_sync(cur_cfqq))
1719 if (CFQQ_SEEKY(cur_cfqq))
1723 * We should notice if some of the queues are cooperating, eg
1724 * working closely on the same area of the disk. In that case,
1725 * we can group them together and don't waste time idling.
1727 cfqq = cfqq_close(cfqd, cur_cfqq);
1731 /* If new queue belongs to different cfq_group, don't choose it */
1732 if (cur_cfqq->cfqg != cfqq->cfqg)
1736 * It only makes sense to merge sync queues.
1738 if (!cfq_cfqq_sync(cfqq))
1740 if (CFQQ_SEEKY(cfqq))
1744 * Do not merge queues of different priority classes
1746 if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq))
1753 * Determine whether we should enforce idle window for this queue.
1756 static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1758 enum wl_prio_t prio = cfqq_prio(cfqq);
1759 struct cfq_rb_root *service_tree = cfqq->service_tree;
1761 BUG_ON(!service_tree);
1762 BUG_ON(!service_tree->count);
1764 /* We never do for idle class queues. */
1765 if (prio == IDLE_WORKLOAD)
1768 /* We do for queues that were marked with idle window flag. */
1769 if (cfq_cfqq_idle_window(cfqq))
1773 * Otherwise, we do only if they are the last ones
1774 * in their service tree.
1776 return service_tree->count == 1;
1779 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
1781 struct cfq_queue *cfqq = cfqd->active_queue;
1782 struct cfq_io_context *cic;
1786 * SSD device without seek penalty, disable idling. But only do so
1787 * for devices that support queuing, otherwise we still have a problem
1788 * with sync vs async workloads.
1790 if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
1793 WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
1794 WARN_ON(cfq_cfqq_slice_new(cfqq));
1797 * idle is disabled, either manually or by past process history
1799 if (!cfqd->cfq_slice_idle || !cfq_should_idle(cfqd, cfqq))
1803 * still active requests from this queue, don't idle
1805 if (cfqq->dispatched)
1809 * task has exited, don't wait
1811 cic = cfqd->active_cic;
1812 if (!cic || !atomic_read(&cic->ioc->nr_tasks))
1816 * If our average think time is larger than the remaining time
1817 * slice, then don't idle. This avoids overrunning the allotted
1820 if (sample_valid(cic->ttime_samples) &&
1821 (cfqq->slice_end - jiffies < cic->ttime_mean))
1824 cfq_mark_cfqq_wait_request(cfqq);
1826 sl = cfqd->cfq_slice_idle;
1828 mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
1829 cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu", sl);
1833 * Move request from internal lists to the request queue dispatch list.
1835 static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
1837 struct cfq_data *cfqd = q->elevator->elevator_data;
1838 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1840 cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
1842 cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
1843 cfq_remove_request(rq);
1845 elv_dispatch_sort(q, rq);
1847 if (cfq_cfqq_sync(cfqq))
1848 cfqd->sync_flight++;
1849 cfqq->nr_sectors += blk_rq_sectors(rq);
1853 * return expired entry, or NULL to just start from scratch in rbtree
1855 static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
1857 struct request *rq = NULL;
1859 if (cfq_cfqq_fifo_expire(cfqq))
1862 cfq_mark_cfqq_fifo_expire(cfqq);
1864 if (list_empty(&cfqq->fifo))
1867 rq = rq_entry_fifo(cfqq->fifo.next);
1868 if (time_before(jiffies, rq_fifo_time(rq)))
1871 cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq);
1876 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1878 const int base_rq = cfqd->cfq_slice_async_rq;
1880 WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
1882 return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
1886 * Must be called with the queue_lock held.
1888 static int cfqq_process_refs(struct cfq_queue *cfqq)
1890 int process_refs, io_refs;
1892 io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE];
1893 process_refs = atomic_read(&cfqq->ref) - io_refs;
1894 BUG_ON(process_refs < 0);
1895 return process_refs;
1898 static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq)
1900 int process_refs, new_process_refs;
1901 struct cfq_queue *__cfqq;
1903 /* Avoid a circular list and skip interim queue merges */
1904 while ((__cfqq = new_cfqq->new_cfqq)) {
1910 process_refs = cfqq_process_refs(cfqq);
1912 * If the process for the cfqq has gone away, there is no
1913 * sense in merging the queues.
1915 if (process_refs == 0)
1919 * Merge in the direction of the lesser amount of work.
1921 new_process_refs = cfqq_process_refs(new_cfqq);
1922 if (new_process_refs >= process_refs) {
1923 cfqq->new_cfqq = new_cfqq;
1924 atomic_add(process_refs, &new_cfqq->ref);
1926 new_cfqq->new_cfqq = cfqq;
1927 atomic_add(new_process_refs, &cfqq->ref);
1931 static enum wl_type_t cfq_choose_wl(struct cfq_data *cfqd,
1932 struct cfq_group *cfqg, enum wl_prio_t prio,
1935 struct cfq_queue *queue;
1937 bool key_valid = false;
1938 unsigned long lowest_key = 0;
1939 enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD;
1943 * When priorities switched, we prefer starting
1944 * from SYNC_NOIDLE (first choice), or just SYNC
1947 if (service_tree_for(cfqg, prio, cur_best, cfqd)->count)
1949 cur_best = SYNC_WORKLOAD;
1950 if (service_tree_for(cfqg, prio, cur_best, cfqd)->count)
1953 return ASYNC_WORKLOAD;
1956 for (i = 0; i < 3; ++i) {
1957 /* otherwise, select the one with lowest rb_key */
1958 queue = cfq_rb_first(service_tree_for(cfqg, prio, i, cfqd));
1960 (!key_valid || time_before(queue->rb_key, lowest_key))) {
1961 lowest_key = queue->rb_key;
1970 static void choose_service_tree(struct cfq_data *cfqd, struct cfq_group *cfqg)
1972 enum wl_prio_t previous_prio = cfqd->serving_prio;
1976 struct cfq_rb_root *st;
1977 unsigned group_slice;
1980 cfqd->serving_prio = IDLE_WORKLOAD;
1981 cfqd->workload_expires = jiffies + 1;
1985 /* Choose next priority. RT > BE > IDLE */
1986 if (cfq_group_busy_queues_wl(RT_WORKLOAD, cfqd, cfqg))
1987 cfqd->serving_prio = RT_WORKLOAD;
1988 else if (cfq_group_busy_queues_wl(BE_WORKLOAD, cfqd, cfqg))
1989 cfqd->serving_prio = BE_WORKLOAD;
1991 cfqd->serving_prio = IDLE_WORKLOAD;
1992 cfqd->workload_expires = jiffies + 1;
1997 * For RT and BE, we have to choose also the type
1998 * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
2001 prio_changed = (cfqd->serving_prio != previous_prio);
2002 st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type,
2007 * If priority didn't change, check workload expiration,
2008 * and that we still have other queues ready
2010 if (!prio_changed && count &&
2011 !time_after(jiffies, cfqd->workload_expires))
2014 /* otherwise select new workload type */
2015 cfqd->serving_type =
2016 cfq_choose_wl(cfqd, cfqg, cfqd->serving_prio, prio_changed);
2017 st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type,
2022 * the workload slice is computed as a fraction of target latency
2023 * proportional to the number of queues in that workload, over
2024 * all the queues in the same priority class
2026 group_slice = cfq_group_slice(cfqd, cfqg);
2028 slice = group_slice * count /
2029 max_t(unsigned, cfqg->busy_queues_avg[cfqd->serving_prio],
2030 cfq_group_busy_queues_wl(cfqd->serving_prio, cfqd, cfqg));
2032 if (cfqd->serving_type == ASYNC_WORKLOAD) {
2036 * Async queues are currently system wide. Just taking
2037 * proportion of queues with-in same group will lead to higher
2038 * async ratio system wide as generally root group is going
2039 * to have higher weight. A more accurate thing would be to
2040 * calculate system wide asnc/sync ratio.
2042 tmp = cfq_target_latency * cfqg_busy_async_queues(cfqd, cfqg);
2043 tmp = tmp/cfqd->busy_queues;
2044 slice = min_t(unsigned, slice, tmp);
2046 /* async workload slice is scaled down according to
2047 * the sync/async slice ratio. */
2048 slice = slice * cfqd->cfq_slice[0] / cfqd->cfq_slice[1];
2050 /* sync workload slice is at least 2 * cfq_slice_idle */
2051 slice = max(slice, 2 * cfqd->cfq_slice_idle);
2053 slice = max_t(unsigned, slice, CFQ_MIN_TT);
2054 cfqd->workload_expires = jiffies + slice;
2055 cfqd->noidle_tree_requires_idle = false;
2058 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd)
2060 struct cfq_rb_root *st = &cfqd->grp_service_tree;
2061 struct cfq_group *cfqg;
2063 if (RB_EMPTY_ROOT(&st->rb))
2065 cfqg = cfq_rb_first_group(st);
2066 st->active = &cfqg->rb_node;
2067 update_min_vdisktime(st);
2071 static void cfq_choose_cfqg(struct cfq_data *cfqd)
2073 struct cfq_group *cfqg = cfq_get_next_cfqg(cfqd);
2075 cfqd->serving_group = cfqg;
2077 /* Restore the workload type data */
2078 if (cfqg->saved_workload_slice) {
2079 cfqd->workload_expires = jiffies + cfqg->saved_workload_slice;
2080 cfqd->serving_type = cfqg->saved_workload;
2081 cfqd->serving_prio = cfqg->saved_serving_prio;
2083 choose_service_tree(cfqd, cfqg);
2087 * Select a queue for service. If we have a current active queue,
2088 * check whether to continue servicing it, or retrieve and set a new one.
2090 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
2092 struct cfq_queue *cfqq, *new_cfqq = NULL;
2094 cfqq = cfqd->active_queue;
2098 if (!cfqd->rq_queued)
2101 * The active queue has run out of time, expire it and select new.
2103 if ((cfq_slice_used(cfqq) || cfq_cfqq_wait_busy_done(cfqq))
2104 && !cfq_cfqq_must_dispatch(cfqq))
2108 * The active queue has requests and isn't expired, allow it to
2111 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
2115 * If another queue has a request waiting within our mean seek
2116 * distance, let it run. The expire code will check for close
2117 * cooperators and put the close queue at the front of the service
2118 * tree. If possible, merge the expiring queue with the new cfqq.
2120 new_cfqq = cfq_close_cooperator(cfqd, cfqq);
2122 if (!cfqq->new_cfqq)
2123 cfq_setup_merge(cfqq, new_cfqq);
2128 * No requests pending. If the active queue still has requests in
2129 * flight or is idling for a new request, allow either of these
2130 * conditions to happen (or time out) before selecting a new queue.
2132 if (timer_pending(&cfqd->idle_slice_timer) ||
2133 (cfqq->dispatched && cfq_should_idle(cfqd, cfqq))) {
2139 cfq_slice_expired(cfqd, 0);
2142 * Current queue expired. Check if we have to switch to a new
2146 cfq_choose_cfqg(cfqd);
2148 cfqq = cfq_set_active_queue(cfqd, new_cfqq);
2153 static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
2157 while (cfqq->next_rq) {
2158 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
2162 BUG_ON(!list_empty(&cfqq->fifo));
2164 /* By default cfqq is not expired if it is empty. Do it explicitly */
2165 __cfq_slice_expired(cfqq->cfqd, cfqq, 0);
2170 * Drain our current requests. Used for barriers and when switching
2171 * io schedulers on-the-fly.
2173 static int cfq_forced_dispatch(struct cfq_data *cfqd)
2175 struct cfq_queue *cfqq;
2178 while ((cfqq = cfq_get_next_queue_forced(cfqd)) != NULL)
2179 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
2181 cfq_slice_expired(cfqd, 0);
2182 BUG_ON(cfqd->busy_queues);
2184 cfq_log(cfqd, "forced_dispatch=%d", dispatched);
2188 static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2190 unsigned int max_dispatch;
2193 * Drain async requests before we start sync IO
2195 if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_driver[BLK_RW_ASYNC])
2199 * If this is an async queue and we have sync IO in flight, let it wait
2201 if (cfqd->sync_flight && !cfq_cfqq_sync(cfqq))
2204 max_dispatch = cfqd->cfq_quantum;
2205 if (cfq_class_idle(cfqq))
2209 * Does this cfqq already have too much IO in flight?
2211 if (cfqq->dispatched >= max_dispatch) {
2213 * idle queue must always only have a single IO in flight
2215 if (cfq_class_idle(cfqq))
2219 * We have other queues, don't allow more IO from this one
2221 if (cfqd->busy_queues > 1)
2225 * Sole queue user, no limit
2231 * Async queues must wait a bit before being allowed dispatch.
2232 * We also ramp up the dispatch depth gradually for async IO,
2233 * based on the last sync IO we serviced
2235 if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
2236 unsigned long last_sync = jiffies - cfqd->last_end_sync_rq;
2239 depth = last_sync / cfqd->cfq_slice[1];
2240 if (!depth && !cfqq->dispatched)
2242 if (depth < max_dispatch)
2243 max_dispatch = depth;
2247 * If we're below the current max, allow a dispatch
2249 return cfqq->dispatched < max_dispatch;
2253 * Dispatch a request from cfqq, moving them to the request queue
2256 static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2260 BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
2262 if (!cfq_may_dispatch(cfqd, cfqq))
2266 * follow expired path, else get first next available
2268 rq = cfq_check_fifo(cfqq);
2273 * insert request into driver dispatch list
2275 cfq_dispatch_insert(cfqd->queue, rq);
2277 if (!cfqd->active_cic) {
2278 struct cfq_io_context *cic = RQ_CIC(rq);
2280 atomic_long_inc(&cic->ioc->refcount);
2281 cfqd->active_cic = cic;
2288 * Find the cfqq that we need to service and move a request from that to the
2291 static int cfq_dispatch_requests(struct request_queue *q, int force)
2293 struct cfq_data *cfqd = q->elevator->elevator_data;
2294 struct cfq_queue *cfqq;
2296 if (!cfqd->busy_queues)
2299 if (unlikely(force))
2300 return cfq_forced_dispatch(cfqd);
2302 cfqq = cfq_select_queue(cfqd);
2307 * Dispatch a request from this cfqq, if it is allowed
2309 if (!cfq_dispatch_request(cfqd, cfqq))
2312 cfqq->slice_dispatch++;
2313 cfq_clear_cfqq_must_dispatch(cfqq);
2316 * expire an async queue immediately if it has used up its slice. idle
2317 * queue always expire after 1 dispatch round.
2319 if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
2320 cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
2321 cfq_class_idle(cfqq))) {
2322 cfqq->slice_end = jiffies + 1;
2323 cfq_slice_expired(cfqd, 0);
2326 cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
2331 * task holds one reference to the queue, dropped when task exits. each rq
2332 * in-flight on this queue also holds a reference, dropped when rq is freed.
2334 * Each cfq queue took a reference on the parent group. Drop it now.
2335 * queue lock must be held here.
2337 static void cfq_put_queue(struct cfq_queue *cfqq)
2339 struct cfq_data *cfqd = cfqq->cfqd;
2340 struct cfq_group *cfqg;
2342 BUG_ON(atomic_read(&cfqq->ref) <= 0);
2344 if (!atomic_dec_and_test(&cfqq->ref))
2347 cfq_log_cfqq(cfqd, cfqq, "put_queue");
2348 BUG_ON(rb_first(&cfqq->sort_list));
2349 BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
2352 if (unlikely(cfqd->active_queue == cfqq)) {
2353 __cfq_slice_expired(cfqd, cfqq, 0);
2354 cfq_schedule_dispatch(cfqd);
2357 BUG_ON(cfq_cfqq_on_rr(cfqq));
2358 kmem_cache_free(cfq_pool, cfqq);
2363 * Must always be called with the rcu_read_lock() held
2366 __call_for_each_cic(struct io_context *ioc,
2367 void (*func)(struct io_context *, struct cfq_io_context *))
2369 struct cfq_io_context *cic;
2370 struct hlist_node *n;
2372 hlist_for_each_entry_rcu(cic, n, &ioc->cic_list, cic_list)
2377 * Call func for each cic attached to this ioc.
2380 call_for_each_cic(struct io_context *ioc,
2381 void (*func)(struct io_context *, struct cfq_io_context *))
2384 __call_for_each_cic(ioc, func);
2388 static void cfq_cic_free_rcu(struct rcu_head *head)
2390 struct cfq_io_context *cic;
2392 cic = container_of(head, struct cfq_io_context, rcu_head);
2394 kmem_cache_free(cfq_ioc_pool, cic);
2395 elv_ioc_count_dec(cfq_ioc_count);
2399 * CFQ scheduler is exiting, grab exit lock and check
2400 * the pending io context count. If it hits zero,
2401 * complete ioc_gone and set it back to NULL
2403 spin_lock(&ioc_gone_lock);
2404 if (ioc_gone && !elv_ioc_count_read(cfq_ioc_count)) {
2408 spin_unlock(&ioc_gone_lock);
2412 static void cfq_cic_free(struct cfq_io_context *cic)
2414 call_rcu(&cic->rcu_head, cfq_cic_free_rcu);
2417 static void cic_free_func(struct io_context *ioc, struct cfq_io_context *cic)
2419 unsigned long flags;
2421 BUG_ON(!cic->dead_key);
2423 spin_lock_irqsave(&ioc->lock, flags);
2424 radix_tree_delete(&ioc->radix_root, cic->dead_key);
2425 hlist_del_rcu(&cic->cic_list);
2426 spin_unlock_irqrestore(&ioc->lock, flags);
2432 * Must be called with rcu_read_lock() held or preemption otherwise disabled.
2433 * Only two callers of this - ->dtor() which is called with the rcu_read_lock(),
2434 * and ->trim() which is called with the task lock held
2436 static void cfq_free_io_context(struct io_context *ioc)
2439 * ioc->refcount is zero here, or we are called from elv_unregister(),
2440 * so no more cic's are allowed to be linked into this ioc. So it
2441 * should be ok to iterate over the known list, we will see all cic's
2442 * since no new ones are added.
2444 __call_for_each_cic(ioc, cic_free_func);
2447 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2449 struct cfq_queue *__cfqq, *next;
2451 if (unlikely(cfqq == cfqd->active_queue)) {
2452 __cfq_slice_expired(cfqd, cfqq, 0);
2453 cfq_schedule_dispatch(cfqd);
2457 * If this queue was scheduled to merge with another queue, be
2458 * sure to drop the reference taken on that queue (and others in
2459 * the merge chain). See cfq_setup_merge and cfq_merge_cfqqs.
2461 __cfqq = cfqq->new_cfqq;
2463 if (__cfqq == cfqq) {
2464 WARN(1, "cfqq->new_cfqq loop detected\n");
2467 next = __cfqq->new_cfqq;
2468 cfq_put_queue(__cfqq);
2472 cfq_put_queue(cfqq);
2475 static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
2476 struct cfq_io_context *cic)
2478 struct io_context *ioc = cic->ioc;
2480 list_del_init(&cic->queue_list);
2483 * Make sure key == NULL is seen for dead queues
2486 cic->dead_key = (unsigned long) cic->key;
2489 if (ioc->ioc_data == cic)
2490 rcu_assign_pointer(ioc->ioc_data, NULL);
2492 if (cic->cfqq[BLK_RW_ASYNC]) {
2493 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]);
2494 cic->cfqq[BLK_RW_ASYNC] = NULL;
2497 if (cic->cfqq[BLK_RW_SYNC]) {
2498 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]);
2499 cic->cfqq[BLK_RW_SYNC] = NULL;
2503 static void cfq_exit_single_io_context(struct io_context *ioc,
2504 struct cfq_io_context *cic)
2506 struct cfq_data *cfqd = cic->key;
2509 struct request_queue *q = cfqd->queue;
2510 unsigned long flags;
2512 spin_lock_irqsave(q->queue_lock, flags);
2515 * Ensure we get a fresh copy of the ->key to prevent
2516 * race between exiting task and queue
2518 smp_read_barrier_depends();
2520 __cfq_exit_single_io_context(cfqd, cic);
2522 spin_unlock_irqrestore(q->queue_lock, flags);
2527 * The process that ioc belongs to has exited, we need to clean up
2528 * and put the internal structures we have that belongs to that process.
2530 static void cfq_exit_io_context(struct io_context *ioc)
2532 call_for_each_cic(ioc, cfq_exit_single_io_context);
2535 static struct cfq_io_context *
2536 cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
2538 struct cfq_io_context *cic;
2540 cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO,
2543 cic->last_end_request = jiffies;
2544 INIT_LIST_HEAD(&cic->queue_list);
2545 INIT_HLIST_NODE(&cic->cic_list);
2546 cic->dtor = cfq_free_io_context;
2547 cic->exit = cfq_exit_io_context;
2548 elv_ioc_count_inc(cfq_ioc_count);
2554 static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc)
2556 struct task_struct *tsk = current;
2559 if (!cfq_cfqq_prio_changed(cfqq))
2562 ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio);
2563 switch (ioprio_class) {
2565 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
2566 case IOPRIO_CLASS_NONE:
2568 * no prio set, inherit CPU scheduling settings
2570 cfqq->ioprio = task_nice_ioprio(tsk);
2571 cfqq->ioprio_class = task_nice_ioclass(tsk);
2573 case IOPRIO_CLASS_RT:
2574 cfqq->ioprio = task_ioprio(ioc);
2575 cfqq->ioprio_class = IOPRIO_CLASS_RT;
2577 case IOPRIO_CLASS_BE:
2578 cfqq->ioprio = task_ioprio(ioc);
2579 cfqq->ioprio_class = IOPRIO_CLASS_BE;
2581 case IOPRIO_CLASS_IDLE:
2582 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
2584 cfq_clear_cfqq_idle_window(cfqq);
2589 * keep track of original prio settings in case we have to temporarily
2590 * elevate the priority of this queue
2592 cfqq->org_ioprio = cfqq->ioprio;
2593 cfqq->org_ioprio_class = cfqq->ioprio_class;
2594 cfq_clear_cfqq_prio_changed(cfqq);
2597 static void changed_ioprio(struct io_context *ioc, struct cfq_io_context *cic)
2599 struct cfq_data *cfqd = cic->key;
2600 struct cfq_queue *cfqq;
2601 unsigned long flags;
2603 if (unlikely(!cfqd))
2606 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2608 cfqq = cic->cfqq[BLK_RW_ASYNC];
2610 struct cfq_queue *new_cfqq;
2611 new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic->ioc,
2614 cic->cfqq[BLK_RW_ASYNC] = new_cfqq;
2615 cfq_put_queue(cfqq);
2619 cfqq = cic->cfqq[BLK_RW_SYNC];
2621 cfq_mark_cfqq_prio_changed(cfqq);
2623 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2626 static void cfq_ioc_set_ioprio(struct io_context *ioc)
2628 call_for_each_cic(ioc, changed_ioprio);
2629 ioc->ioprio_changed = 0;
2632 static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2633 pid_t pid, bool is_sync)
2635 RB_CLEAR_NODE(&cfqq->rb_node);
2636 RB_CLEAR_NODE(&cfqq->p_node);
2637 INIT_LIST_HEAD(&cfqq->fifo);
2639 atomic_set(&cfqq->ref, 0);
2642 cfq_mark_cfqq_prio_changed(cfqq);
2645 if (!cfq_class_idle(cfqq))
2646 cfq_mark_cfqq_idle_window(cfqq);
2647 cfq_mark_cfqq_sync(cfqq);
2652 #ifdef CONFIG_CFQ_GROUP_IOSCHED
2653 static void changed_cgroup(struct io_context *ioc, struct cfq_io_context *cic)
2655 struct cfq_queue *sync_cfqq = cic_to_cfqq(cic, 1);
2656 struct cfq_data *cfqd = cic->key;
2657 unsigned long flags;
2658 struct request_queue *q;
2660 if (unlikely(!cfqd))
2665 spin_lock_irqsave(q->queue_lock, flags);
2669 * Drop reference to sync queue. A new sync queue will be
2670 * assigned in new group upon arrival of a fresh request.
2672 cfq_log_cfqq(cfqd, sync_cfqq, "changed cgroup");
2673 cic_set_cfqq(cic, NULL, 1);
2674 cfq_put_queue(sync_cfqq);
2677 spin_unlock_irqrestore(q->queue_lock, flags);
2680 static void cfq_ioc_set_cgroup(struct io_context *ioc)
2682 call_for_each_cic(ioc, changed_cgroup);
2683 ioc->cgroup_changed = 0;
2685 #endif /* CONFIG_CFQ_GROUP_IOSCHED */
2687 static struct cfq_queue *
2688 cfq_find_alloc_queue(struct cfq_data *cfqd, bool is_sync,
2689 struct io_context *ioc, gfp_t gfp_mask)
2691 struct cfq_queue *cfqq, *new_cfqq = NULL;
2692 struct cfq_io_context *cic;
2693 struct cfq_group *cfqg;
2696 cfqg = cfq_get_cfqg(cfqd, 1);
2697 cic = cfq_cic_lookup(cfqd, ioc);
2698 /* cic always exists here */
2699 cfqq = cic_to_cfqq(cic, is_sync);
2702 * Always try a new alloc if we fell back to the OOM cfqq
2703 * originally, since it should just be a temporary situation.
2705 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
2710 } else if (gfp_mask & __GFP_WAIT) {
2711 spin_unlock_irq(cfqd->queue->queue_lock);
2712 new_cfqq = kmem_cache_alloc_node(cfq_pool,
2713 gfp_mask | __GFP_ZERO,
2715 spin_lock_irq(cfqd->queue->queue_lock);
2719 cfqq = kmem_cache_alloc_node(cfq_pool,
2720 gfp_mask | __GFP_ZERO,
2725 cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
2726 cfq_init_prio_data(cfqq, ioc);
2727 cfq_link_cfqq_cfqg(cfqq, cfqg);
2728 cfq_log_cfqq(cfqd, cfqq, "alloced");
2730 cfqq = &cfqd->oom_cfqq;
2734 kmem_cache_free(cfq_pool, new_cfqq);
2739 static struct cfq_queue **
2740 cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
2742 switch (ioprio_class) {
2743 case IOPRIO_CLASS_RT:
2744 return &cfqd->async_cfqq[0][ioprio];
2745 case IOPRIO_CLASS_BE:
2746 return &cfqd->async_cfqq[1][ioprio];
2747 case IOPRIO_CLASS_IDLE:
2748 return &cfqd->async_idle_cfqq;
2754 static struct cfq_queue *
2755 cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct io_context *ioc,
2758 const int ioprio = task_ioprio(ioc);
2759 const int ioprio_class = task_ioprio_class(ioc);
2760 struct cfq_queue **async_cfqq = NULL;
2761 struct cfq_queue *cfqq = NULL;
2764 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
2769 cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask);
2772 * pin the queue now that it's allocated, scheduler exit will prune it
2774 if (!is_sync && !(*async_cfqq)) {
2775 atomic_inc(&cfqq->ref);
2779 atomic_inc(&cfqq->ref);
2784 * We drop cfq io contexts lazily, so we may find a dead one.
2787 cfq_drop_dead_cic(struct cfq_data *cfqd, struct io_context *ioc,
2788 struct cfq_io_context *cic)
2790 unsigned long flags;
2792 WARN_ON(!list_empty(&cic->queue_list));
2794 spin_lock_irqsave(&ioc->lock, flags);
2796 BUG_ON(ioc->ioc_data == cic);
2798 radix_tree_delete(&ioc->radix_root, (unsigned long) cfqd);
2799 hlist_del_rcu(&cic->cic_list);
2800 spin_unlock_irqrestore(&ioc->lock, flags);
2805 static struct cfq_io_context *
2806 cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc)
2808 struct cfq_io_context *cic;
2809 unsigned long flags;
2818 * we maintain a last-hit cache, to avoid browsing over the tree
2820 cic = rcu_dereference(ioc->ioc_data);
2821 if (cic && cic->key == cfqd) {
2827 cic = radix_tree_lookup(&ioc->radix_root, (unsigned long) cfqd);
2831 /* ->key must be copied to avoid race with cfq_exit_queue() */
2834 cfq_drop_dead_cic(cfqd, ioc, cic);
2839 spin_lock_irqsave(&ioc->lock, flags);
2840 rcu_assign_pointer(ioc->ioc_data, cic);
2841 spin_unlock_irqrestore(&ioc->lock, flags);
2849 * Add cic into ioc, using cfqd as the search key. This enables us to lookup
2850 * the process specific cfq io context when entered from the block layer.
2851 * Also adds the cic to a per-cfqd list, used when this queue is removed.
2853 static int cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
2854 struct cfq_io_context *cic, gfp_t gfp_mask)
2856 unsigned long flags;
2859 ret = radix_tree_preload(gfp_mask);
2864 spin_lock_irqsave(&ioc->lock, flags);
2865 ret = radix_tree_insert(&ioc->radix_root,
2866 (unsigned long) cfqd, cic);
2868 hlist_add_head_rcu(&cic->cic_list, &ioc->cic_list);
2869 spin_unlock_irqrestore(&ioc->lock, flags);
2871 radix_tree_preload_end();
2874 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2875 list_add(&cic->queue_list, &cfqd->cic_list);
2876 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2881 printk(KERN_ERR "cfq: cic link failed!\n");
2887 * Setup general io context and cfq io context. There can be several cfq
2888 * io contexts per general io context, if this process is doing io to more
2889 * than one device managed by cfq.
2891 static struct cfq_io_context *
2892 cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
2894 struct io_context *ioc = NULL;
2895 struct cfq_io_context *cic;
2897 might_sleep_if(gfp_mask & __GFP_WAIT);
2899 ioc = get_io_context(gfp_mask, cfqd->queue->node);
2903 cic = cfq_cic_lookup(cfqd, ioc);
2907 cic = cfq_alloc_io_context(cfqd, gfp_mask);
2911 if (cfq_cic_link(cfqd, ioc, cic, gfp_mask))
2915 smp_read_barrier_depends();
2916 if (unlikely(ioc->ioprio_changed))
2917 cfq_ioc_set_ioprio(ioc);
2919 #ifdef CONFIG_CFQ_GROUP_IOSCHED
2920 if (unlikely(ioc->cgroup_changed))
2921 cfq_ioc_set_cgroup(ioc);
2927 put_io_context(ioc);
2932 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
2934 unsigned long elapsed = jiffies - cic->last_end_request;
2935 unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
2937 cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
2938 cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
2939 cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
2943 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2949 if (!cfqq->last_request_pos)
2951 else if (cfqq->last_request_pos < blk_rq_pos(rq))
2952 sdist = blk_rq_pos(rq) - cfqq->last_request_pos;
2954 sdist = cfqq->last_request_pos - blk_rq_pos(rq);
2957 * Don't allow the seek distance to get too large from the
2958 * odd fragment, pagein, etc
2960 if (cfqq->seek_samples <= 60) /* second&third seek */
2961 sdist = min(sdist, (cfqq->seek_mean * 4) + 2*1024*1024);
2963 sdist = min(sdist, (cfqq->seek_mean * 4) + 2*1024*64);
2965 cfqq->seek_samples = (7*cfqq->seek_samples + 256) / 8;
2966 cfqq->seek_total = (7*cfqq->seek_total + (u64)256*sdist) / 8;
2967 total = cfqq->seek_total + (cfqq->seek_samples/2);
2968 do_div(total, cfqq->seek_samples);
2969 cfqq->seek_mean = (sector_t)total;
2972 * If this cfqq is shared between multiple processes, check to
2973 * make sure that those processes are still issuing I/Os within
2974 * the mean seek distance. If not, it may be time to break the
2975 * queues apart again.
2977 if (cfq_cfqq_coop(cfqq)) {
2978 if (CFQQ_SEEKY(cfqq) && !cfqq->seeky_start)
2979 cfqq->seeky_start = jiffies;
2980 else if (!CFQQ_SEEKY(cfqq))
2981 cfqq->seeky_start = 0;
2986 * Disable idle window if the process thinks too long or seeks so much that
2990 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2991 struct cfq_io_context *cic)
2993 int old_idle, enable_idle;
2996 * Don't idle for async or idle io prio class
2998 if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
3001 enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
3003 if (cfqq->queued[0] + cfqq->queued[1] >= 4)
3004 cfq_mark_cfqq_deep(cfqq);
3006 if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle ||
3007 (!cfq_cfqq_deep(cfqq) && sample_valid(cfqq->seek_samples)
3008 && CFQQ_SEEKY(cfqq)))
3010 else if (sample_valid(cic->ttime_samples)) {
3011 if (cic->ttime_mean > cfqd->cfq_slice_idle)
3017 if (old_idle != enable_idle) {
3018 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
3020 cfq_mark_cfqq_idle_window(cfqq);
3022 cfq_clear_cfqq_idle_window(cfqq);
3027 * Check if new_cfqq should preempt the currently active queue. Return 0 for
3028 * no or if we aren't sure, a 1 will cause a preempt.
3031 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
3034 struct cfq_queue *cfqq;
3036 cfqq = cfqd->active_queue;
3040 if (cfq_class_idle(new_cfqq))
3043 if (cfq_class_idle(cfqq))
3047 * if the new request is sync, but the currently running queue is
3048 * not, let the sync request have priority.
3050 if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
3053 if (new_cfqq->cfqg != cfqq->cfqg)
3056 if (cfq_slice_used(cfqq))
3059 /* Allow preemption only if we are idling on sync-noidle tree */
3060 if (cfqd->serving_type == SYNC_NOIDLE_WORKLOAD &&
3061 cfqq_type(new_cfqq) == SYNC_NOIDLE_WORKLOAD &&
3062 new_cfqq->service_tree->count == 2 &&
3063 RB_EMPTY_ROOT(&cfqq->sort_list))
3067 * So both queues are sync. Let the new request get disk time if
3068 * it's a metadata request and the current queue is doing regular IO.
3070 if (rq_is_meta(rq) && !cfqq->meta_pending)
3074 * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
3076 if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
3079 if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
3083 * if this request is as-good as one we would expect from the
3084 * current cfqq, let it preempt
3086 if (cfq_rq_close(cfqd, cfqq, rq))
3093 * cfqq preempts the active queue. if we allowed preempt with no slice left,
3094 * let it have half of its nominal slice.
3096 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3098 cfq_log_cfqq(cfqd, cfqq, "preempt");
3099 cfq_slice_expired(cfqd, 1);
3102 * Put the new queue at the front of the of the current list,
3103 * so we know that it will be selected next.
3105 BUG_ON(!cfq_cfqq_on_rr(cfqq));
3107 cfq_service_tree_add(cfqd, cfqq, 1);
3109 cfqq->slice_end = 0;
3110 cfq_mark_cfqq_slice_new(cfqq);
3114 * Called when a new fs request (rq) is added (to cfqq). Check if there's
3115 * something we should do about it
3118 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3121 struct cfq_io_context *cic = RQ_CIC(rq);
3125 cfqq->meta_pending++;
3127 cfq_update_io_thinktime(cfqd, cic);
3128 cfq_update_io_seektime(cfqd, cfqq, rq);
3129 cfq_update_idle_window(cfqd, cfqq, cic);
3131 cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
3133 if (cfqq == cfqd->active_queue) {
3134 if (cfq_cfqq_wait_busy(cfqq)) {
3135 cfq_clear_cfqq_wait_busy(cfqq);
3136 cfq_mark_cfqq_wait_busy_done(cfqq);
3139 * Remember that we saw a request from this process, but
3140 * don't start queuing just yet. Otherwise we risk seeing lots
3141 * of tiny requests, because we disrupt the normal plugging
3142 * and merging. If the request is already larger than a single
3143 * page, let it rip immediately. For that case we assume that
3144 * merging is already done. Ditto for a busy system that
3145 * has other work pending, don't risk delaying until the
3146 * idle timer unplug to continue working.
3148 if (cfq_cfqq_wait_request(cfqq)) {
3149 if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
3150 cfqd->busy_queues > 1) {
3151 del_timer(&cfqd->idle_slice_timer);
3152 __blk_run_queue(cfqd->queue);
3154 cfq_mark_cfqq_must_dispatch(cfqq);
3156 } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
3158 * not the active queue - expire current slice if it is
3159 * idle and has expired it's mean thinktime or this new queue
3160 * has some old slice time left and is of higher priority or
3161 * this new queue is RT and the current one is BE
3163 cfq_preempt_queue(cfqd, cfqq);
3164 __blk_run_queue(cfqd->queue);
3168 static void cfq_insert_request(struct request_queue *q, struct request *rq)
3170 struct cfq_data *cfqd = q->elevator->elevator_data;
3171 struct cfq_queue *cfqq = RQ_CFQQ(rq);
3173 cfq_log_cfqq(cfqd, cfqq, "insert_request");
3174 cfq_init_prio_data(cfqq, RQ_CIC(rq)->ioc);
3176 rq_set_fifo_time(rq, jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)]);
3177 list_add_tail(&rq->queuelist, &cfqq->fifo);
3180 cfq_rq_enqueued(cfqd, cfqq, rq);
3184 * Update hw_tag based on peak queue depth over 50 samples under
3187 static void cfq_update_hw_tag(struct cfq_data *cfqd)
3189 struct cfq_queue *cfqq = cfqd->active_queue;
3191 if (rq_in_driver(cfqd) > cfqd->hw_tag_est_depth)
3192 cfqd->hw_tag_est_depth = rq_in_driver(cfqd);
3194 if (cfqd->hw_tag == 1)
3197 if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
3198 rq_in_driver(cfqd) <= CFQ_HW_QUEUE_MIN)
3202 * If active queue hasn't enough requests and can idle, cfq might not
3203 * dispatch sufficient requests to hardware. Don't zero hw_tag in this
3206 if (cfqq && cfq_cfqq_idle_window(cfqq) &&
3207 cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] <
3208 CFQ_HW_QUEUE_MIN && rq_in_driver(cfqd) < CFQ_HW_QUEUE_MIN)
3211 if (cfqd->hw_tag_samples++ < 50)
3214 if (cfqd->hw_tag_est_depth >= CFQ_HW_QUEUE_MIN)
3220 static void cfq_completed_request(struct request_queue *q, struct request *rq)
3222 struct cfq_queue *cfqq = RQ_CFQQ(rq);
3223 struct cfq_data *cfqd = cfqq->cfqd;
3224 const int sync = rq_is_sync(rq);
3228 cfq_log_cfqq(cfqd, cfqq, "complete rqnoidle %d", !!rq_noidle(rq));
3230 cfq_update_hw_tag(cfqd);
3232 WARN_ON(!cfqd->rq_in_driver[sync]);
3233 WARN_ON(!cfqq->dispatched);
3234 cfqd->rq_in_driver[sync]--;
3237 if (cfq_cfqq_sync(cfqq))
3238 cfqd->sync_flight--;
3241 RQ_CIC(rq)->last_end_request = now;
3242 cfqd->last_end_sync_rq = now;
3246 * If this is the active queue, check if it needs to be expired,
3247 * or if we want to idle in case it has no pending requests.
3249 if (cfqd->active_queue == cfqq) {
3250 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
3252 if (cfq_cfqq_slice_new(cfqq)) {
3253 cfq_set_prio_slice(cfqd, cfqq);
3254 cfq_clear_cfqq_slice_new(cfqq);
3258 * If this queue consumed its slice and this is last queue
3259 * in the group, wait for next request before we expire
3262 if (cfq_slice_used(cfqq) && cfqq->cfqg->nr_cfqq == 1) {
3263 cfqq->slice_end = jiffies + cfqd->cfq_slice_idle;
3264 cfq_mark_cfqq_wait_busy(cfqq);
3268 * Idling is not enabled on:
3270 * - idle-priority queues
3272 * - queues with still some requests queued
3273 * - when there is a close cooperator
3275 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
3276 cfq_slice_expired(cfqd, 1);
3277 else if (sync && cfqq_empty &&
3278 !cfq_close_cooperator(cfqd, cfqq)) {
3279 cfqd->noidle_tree_requires_idle |= !rq_noidle(rq);
3281 * Idling is enabled for SYNC_WORKLOAD.
3282 * SYNC_NOIDLE_WORKLOAD idles at the end of the tree
3283 * only if we processed at least one !rq_noidle request
3285 if (cfqd->serving_type == SYNC_WORKLOAD
3286 || cfqd->noidle_tree_requires_idle)
3287 cfq_arm_slice_timer(cfqd);
3291 if (!rq_in_driver(cfqd))
3292 cfq_schedule_dispatch(cfqd);
3296 * we temporarily boost lower priority queues if they are holding fs exclusive
3297 * resources. they are boosted to normal prio (CLASS_BE/4)
3299 static void cfq_prio_boost(struct cfq_queue *cfqq)
3301 if (has_fs_excl()) {
3303 * boost idle prio on transactions that would lock out other
3304 * users of the filesystem
3306 if (cfq_class_idle(cfqq))
3307 cfqq->ioprio_class = IOPRIO_CLASS_BE;
3308 if (cfqq->ioprio > IOPRIO_NORM)
3309 cfqq->ioprio = IOPRIO_NORM;
3312 * unboost the queue (if needed)
3314 cfqq->ioprio_class = cfqq->org_ioprio_class;
3315 cfqq->ioprio = cfqq->org_ioprio;
3319 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
3321 if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
3322 cfq_mark_cfqq_must_alloc_slice(cfqq);
3323 return ELV_MQUEUE_MUST;
3326 return ELV_MQUEUE_MAY;
3329 static int cfq_may_queue(struct request_queue *q, int rw)
3331 struct cfq_data *cfqd = q->elevator->elevator_data;
3332 struct task_struct *tsk = current;
3333 struct cfq_io_context *cic;
3334 struct cfq_queue *cfqq;
3337 * don't force setup of a queue from here, as a call to may_queue
3338 * does not necessarily imply that a request actually will be queued.
3339 * so just lookup a possibly existing queue, or return 'may queue'
3342 cic = cfq_cic_lookup(cfqd, tsk->io_context);
3344 return ELV_MQUEUE_MAY;
3346 cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
3348 cfq_init_prio_data(cfqq, cic->ioc);
3349 cfq_prio_boost(cfqq);
3351 return __cfq_may_queue(cfqq);
3354 return ELV_MQUEUE_MAY;
3358 * queue lock held here
3360 static void cfq_put_request(struct request *rq)
3362 struct cfq_queue *cfqq = RQ_CFQQ(rq);
3365 const int rw = rq_data_dir(rq);
3367 BUG_ON(!cfqq->allocated[rw]);
3368 cfqq->allocated[rw]--;
3370 put_io_context(RQ_CIC(rq)->ioc);
3372 rq->elevator_private = NULL;
3373 rq->elevator_private2 = NULL;
3375 cfq_put_queue(cfqq);
3379 static struct cfq_queue *
3380 cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_context *cic,
3381 struct cfq_queue *cfqq)
3383 cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq);
3384 cic_set_cfqq(cic, cfqq->new_cfqq, 1);
3385 cfq_mark_cfqq_coop(cfqq->new_cfqq);
3386 cfq_put_queue(cfqq);
3387 return cic_to_cfqq(cic, 1);
3390 static int should_split_cfqq(struct cfq_queue *cfqq)
3392 if (cfqq->seeky_start &&
3393 time_after(jiffies, cfqq->seeky_start + CFQQ_COOP_TOUT))
3399 * Returns NULL if a new cfqq should be allocated, or the old cfqq if this
3400 * was the last process referring to said cfqq.
3402 static struct cfq_queue *
3403 split_cfqq(struct cfq_io_context *cic, struct cfq_queue *cfqq)
3405 if (cfqq_process_refs(cfqq) == 1) {
3406 cfqq->seeky_start = 0;
3407 cfqq->pid = current->pid;
3408 cfq_clear_cfqq_coop(cfqq);
3412 cic_set_cfqq(cic, NULL, 1);
3413 cfq_put_queue(cfqq);
3417 * Allocate cfq data structures associated with this request.
3420 cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
3422 struct cfq_data *cfqd = q->elevator->elevator_data;
3423 struct cfq_io_context *cic;
3424 const int rw = rq_data_dir(rq);
3425 const bool is_sync = rq_is_sync(rq);
3426 struct cfq_queue *cfqq;
3427 unsigned long flags;
3429 might_sleep_if(gfp_mask & __GFP_WAIT);
3431 cic = cfq_get_io_context(cfqd, gfp_mask);
3433 spin_lock_irqsave(q->queue_lock, flags);
3439 cfqq = cic_to_cfqq(cic, is_sync);
3440 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
3441 cfqq = cfq_get_queue(cfqd, is_sync, cic->ioc, gfp_mask);
3442 cic_set_cfqq(cic, cfqq, is_sync);
3445 * If the queue was seeky for too long, break it apart.
3447 if (cfq_cfqq_coop(cfqq) && should_split_cfqq(cfqq)) {
3448 cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq");
3449 cfqq = split_cfqq(cic, cfqq);
3455 * Check to see if this queue is scheduled to merge with
3456 * another, closely cooperating queue. The merging of
3457 * queues happens here as it must be done in process context.
3458 * The reference on new_cfqq was taken in merge_cfqqs.
3461 cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq);
3464 cfqq->allocated[rw]++;
3465 atomic_inc(&cfqq->ref);
3467 spin_unlock_irqrestore(q->queue_lock, flags);
3469 rq->elevator_private = cic;
3470 rq->elevator_private2 = cfqq;
3475 put_io_context(cic->ioc);
3477 cfq_schedule_dispatch(cfqd);
3478 spin_unlock_irqrestore(q->queue_lock, flags);
3479 cfq_log(cfqd, "set_request fail");
3483 static void cfq_kick_queue(struct work_struct *work)
3485 struct cfq_data *cfqd =
3486 container_of(work, struct cfq_data, unplug_work);
3487 struct request_queue *q = cfqd->queue;
3489 spin_lock_irq(q->queue_lock);
3490 __blk_run_queue(cfqd->queue);
3491 spin_unlock_irq(q->queue_lock);
3495 * Timer running if the active_queue is currently idling inside its time slice
3497 static void cfq_idle_slice_timer(unsigned long data)
3499 struct cfq_data *cfqd = (struct cfq_data *) data;
3500 struct cfq_queue *cfqq;
3501 unsigned long flags;
3504 cfq_log(cfqd, "idle timer fired");
3506 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
3508 cfqq = cfqd->active_queue;
3513 * We saw a request before the queue expired, let it through
3515 if (cfq_cfqq_must_dispatch(cfqq))
3521 if (cfq_slice_used(cfqq))
3525 * only expire and reinvoke request handler, if there are
3526 * other queues with pending requests
3528 if (!cfqd->busy_queues)
3532 * not expired and it has a request pending, let it dispatch
3534 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3538 * Queue depth flag is reset only when the idle didn't succeed
3540 cfq_clear_cfqq_deep(cfqq);
3543 cfq_slice_expired(cfqd, timed_out);
3545 cfq_schedule_dispatch(cfqd);
3547 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
3550 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
3552 del_timer_sync(&cfqd->idle_slice_timer);
3553 cancel_work_sync(&cfqd->unplug_work);
3556 static void cfq_put_async_queues(struct cfq_data *cfqd)
3560 for (i = 0; i < IOPRIO_BE_NR; i++) {
3561 if (cfqd->async_cfqq[0][i])
3562 cfq_put_queue(cfqd->async_cfqq[0][i]);
3563 if (cfqd->async_cfqq[1][i])
3564 cfq_put_queue(cfqd->async_cfqq[1][i]);
3567 if (cfqd->async_idle_cfqq)
3568 cfq_put_queue(cfqd->async_idle_cfqq);
3571 static void cfq_exit_queue(struct elevator_queue *e)
3573 struct cfq_data *cfqd = e->elevator_data;
3574 struct request_queue *q = cfqd->queue;
3576 cfq_shutdown_timer_wq(cfqd);
3578 spin_lock_irq(q->queue_lock);
3580 if (cfqd->active_queue)
3581 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
3583 while (!list_empty(&cfqd->cic_list)) {
3584 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
3585 struct cfq_io_context,
3588 __cfq_exit_single_io_context(cfqd, cic);
3591 cfq_put_async_queues(cfqd);
3592 cfq_release_cfq_groups(cfqd);
3593 blkiocg_del_blkio_group(&cfqd->root_group.blkg);
3595 spin_unlock_irq(q->queue_lock);
3597 cfq_shutdown_timer_wq(cfqd);
3599 /* Wait for cfqg->blkg->key accessors to exit their grace periods. */
3604 static void *cfq_init_queue(struct request_queue *q)
3606 struct cfq_data *cfqd;
3608 struct cfq_group *cfqg;
3609 struct cfq_rb_root *st;
3611 cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
3615 /* Init root service tree */
3616 cfqd->grp_service_tree = CFQ_RB_ROOT;
3618 /* Init root group */
3619 cfqg = &cfqd->root_group;
3620 for_each_cfqg_st(cfqg, i, j, st)
3622 RB_CLEAR_NODE(&cfqg->rb_node);
3624 /* Give preference to root group over other groups */
3625 cfqg->weight = 2*BLKIO_WEIGHT_DEFAULT;
3627 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3629 * Take a reference to root group which we never drop. This is just
3630 * to make sure that cfq_put_cfqg() does not try to kfree root group
3632 atomic_set(&cfqg->ref, 1);
3633 blkiocg_add_blkio_group(&blkio_root_cgroup, &cfqg->blkg, (void *)cfqd,
3637 * Not strictly needed (since RB_ROOT just clears the node and we
3638 * zeroed cfqd on alloc), but better be safe in case someone decides
3639 * to add magic to the rb code
3641 for (i = 0; i < CFQ_PRIO_LISTS; i++)
3642 cfqd->prio_trees[i] = RB_ROOT;
3645 * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
3646 * Grab a permanent reference to it, so that the normal code flow
3647 * will not attempt to free it.
3649 cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
3650 atomic_inc(&cfqd->oom_cfqq.ref);
3651 cfq_link_cfqq_cfqg(&cfqd->oom_cfqq, &cfqd->root_group);
3653 INIT_LIST_HEAD(&cfqd->cic_list);
3657 init_timer(&cfqd->idle_slice_timer);
3658 cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
3659 cfqd->idle_slice_timer.data = (unsigned long) cfqd;
3661 INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
3663 cfqd->cfq_quantum = cfq_quantum;
3664 cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
3665 cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
3666 cfqd->cfq_back_max = cfq_back_max;
3667 cfqd->cfq_back_penalty = cfq_back_penalty;
3668 cfqd->cfq_slice[0] = cfq_slice_async;
3669 cfqd->cfq_slice[1] = cfq_slice_sync;
3670 cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
3671 cfqd->cfq_slice_idle = cfq_slice_idle;
3672 cfqd->cfq_latency = 1;
3674 cfqd->last_end_sync_rq = jiffies;
3678 static void cfq_slab_kill(void)
3681 * Caller already ensured that pending RCU callbacks are completed,
3682 * so we should have no busy allocations at this point.
3685 kmem_cache_destroy(cfq_pool);
3687 kmem_cache_destroy(cfq_ioc_pool);
3690 static int __init cfq_slab_setup(void)
3692 cfq_pool = KMEM_CACHE(cfq_queue, 0);
3696 cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0);
3707 * sysfs parts below -->
3710 cfq_var_show(unsigned int var, char *page)
3712 return sprintf(page, "%d\n", var);
3716 cfq_var_store(unsigned int *var, const char *page, size_t count)
3718 char *p = (char *) page;
3720 *var = simple_strtoul(p, &p, 10);
3724 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
3725 static ssize_t __FUNC(struct elevator_queue *e, char *page) \
3727 struct cfq_data *cfqd = e->elevator_data; \
3728 unsigned int __data = __VAR; \
3730 __data = jiffies_to_msecs(__data); \
3731 return cfq_var_show(__data, (page)); \
3733 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
3734 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
3735 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
3736 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
3737 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
3738 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
3739 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
3740 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
3741 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
3742 SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
3743 #undef SHOW_FUNCTION
3745 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
3746 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
3748 struct cfq_data *cfqd = e->elevator_data; \
3749 unsigned int __data; \
3750 int ret = cfq_var_store(&__data, (page), count); \
3751 if (__data < (MIN)) \
3753 else if (__data > (MAX)) \
3756 *(__PTR) = msecs_to_jiffies(__data); \
3758 *(__PTR) = __data; \
3761 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
3762 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
3764 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
3766 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
3767 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
3769 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
3770 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
3771 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
3772 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
3774 STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
3775 #undef STORE_FUNCTION
3777 #define CFQ_ATTR(name) \
3778 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
3780 static struct elv_fs_entry cfq_attrs[] = {
3782 CFQ_ATTR(fifo_expire_sync),
3783 CFQ_ATTR(fifo_expire_async),
3784 CFQ_ATTR(back_seek_max),
3785 CFQ_ATTR(back_seek_penalty),
3786 CFQ_ATTR(slice_sync),
3787 CFQ_ATTR(slice_async),
3788 CFQ_ATTR(slice_async_rq),
3789 CFQ_ATTR(slice_idle),
3790 CFQ_ATTR(low_latency),
3794 static struct elevator_type iosched_cfq = {
3796 .elevator_merge_fn = cfq_merge,
3797 .elevator_merged_fn = cfq_merged_request,
3798 .elevator_merge_req_fn = cfq_merged_requests,
3799 .elevator_allow_merge_fn = cfq_allow_merge,
3800 .elevator_dispatch_fn = cfq_dispatch_requests,
3801 .elevator_add_req_fn = cfq_insert_request,
3802 .elevator_activate_req_fn = cfq_activate_request,
3803 .elevator_deactivate_req_fn = cfq_deactivate_request,
3804 .elevator_queue_empty_fn = cfq_queue_empty,
3805 .elevator_completed_req_fn = cfq_completed_request,
3806 .elevator_former_req_fn = elv_rb_former_request,
3807 .elevator_latter_req_fn = elv_rb_latter_request,
3808 .elevator_set_req_fn = cfq_set_request,
3809 .elevator_put_req_fn = cfq_put_request,
3810 .elevator_may_queue_fn = cfq_may_queue,
3811 .elevator_init_fn = cfq_init_queue,
3812 .elevator_exit_fn = cfq_exit_queue,
3813 .trim = cfq_free_io_context,
3815 .elevator_attrs = cfq_attrs,
3816 .elevator_name = "cfq",
3817 .elevator_owner = THIS_MODULE,
3820 static int __init cfq_init(void)
3823 * could be 0 on HZ < 1000 setups
3825 if (!cfq_slice_async)
3826 cfq_slice_async = 1;
3827 if (!cfq_slice_idle)
3830 if (cfq_slab_setup())
3833 elv_register(&iosched_cfq);
3838 static void __exit cfq_exit(void)
3840 DECLARE_COMPLETION_ONSTACK(all_gone);
3841 elv_unregister(&iosched_cfq);
3842 ioc_gone = &all_gone;
3843 /* ioc_gone's update must be visible before reading ioc_count */
3847 * this also protects us from entering cfq_slab_kill() with
3848 * pending RCU callbacks
3850 if (elv_ioc_count_read(cfq_ioc_count))
3851 wait_for_completion(&all_gone);
3855 module_init(cfq_init);
3856 module_exit(cfq_exit);
3858 MODULE_AUTHOR("Jens Axboe");
3859 MODULE_LICENSE("GPL");
3860 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");