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_delayed_sync;
288 /* List of cfq groups being managed on this device*/
289 struct hlist_head cfqg_list;
293 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd);
295 static struct cfq_rb_root *service_tree_for(struct cfq_group *cfqg,
298 struct cfq_data *cfqd)
303 if (prio == IDLE_WORKLOAD)
304 return &cfqg->service_tree_idle;
306 return &cfqg->service_trees[prio][type];
309 enum cfqq_state_flags {
310 CFQ_CFQQ_FLAG_on_rr = 0, /* on round-robin busy list */
311 CFQ_CFQQ_FLAG_wait_request, /* waiting for a request */
312 CFQ_CFQQ_FLAG_must_dispatch, /* must be allowed a dispatch */
313 CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
314 CFQ_CFQQ_FLAG_fifo_expire, /* FIFO checked in this slice */
315 CFQ_CFQQ_FLAG_idle_window, /* slice idling enabled */
316 CFQ_CFQQ_FLAG_prio_changed, /* task priority has changed */
317 CFQ_CFQQ_FLAG_slice_new, /* no requests dispatched in slice */
318 CFQ_CFQQ_FLAG_sync, /* synchronous queue */
319 CFQ_CFQQ_FLAG_coop, /* cfqq is shared */
320 CFQ_CFQQ_FLAG_deep, /* sync cfqq experienced large depth */
321 CFQ_CFQQ_FLAG_wait_busy, /* Waiting for next request */
322 CFQ_CFQQ_FLAG_wait_busy_done, /* Got new request. Expire the queue */
325 #define CFQ_CFQQ_FNS(name) \
326 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \
328 (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name); \
330 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \
332 (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \
334 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \
336 return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \
340 CFQ_CFQQ_FNS(wait_request);
341 CFQ_CFQQ_FNS(must_dispatch);
342 CFQ_CFQQ_FNS(must_alloc_slice);
343 CFQ_CFQQ_FNS(fifo_expire);
344 CFQ_CFQQ_FNS(idle_window);
345 CFQ_CFQQ_FNS(prio_changed);
346 CFQ_CFQQ_FNS(slice_new);
350 CFQ_CFQQ_FNS(wait_busy);
351 CFQ_CFQQ_FNS(wait_busy_done);
354 #ifdef CONFIG_DEBUG_CFQ_IOSCHED
355 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
356 blk_add_trace_msg((cfqd)->queue, "cfq%d%c %s " fmt, (cfqq)->pid, \
357 cfq_cfqq_sync((cfqq)) ? 'S' : 'A', \
358 blkg_path(&(cfqq)->cfqg->blkg), ##args);
360 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...) \
361 blk_add_trace_msg((cfqd)->queue, "%s " fmt, \
362 blkg_path(&(cfqg)->blkg), ##args); \
365 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
366 blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args)
367 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...) do {} while (0);
369 #define cfq_log(cfqd, fmt, args...) \
370 blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
372 /* Traverses through cfq group service trees */
373 #define for_each_cfqg_st(cfqg, i, j, st) \
374 for (i = 0; i <= IDLE_WORKLOAD; i++) \
375 for (j = 0, st = i < IDLE_WORKLOAD ? &cfqg->service_trees[i][j]\
376 : &cfqg->service_tree_idle; \
377 (i < IDLE_WORKLOAD && j <= SYNC_WORKLOAD) || \
378 (i == IDLE_WORKLOAD && j == 0); \
379 j++, st = i < IDLE_WORKLOAD ? \
380 &cfqg->service_trees[i][j]: NULL) \
383 static inline enum wl_prio_t cfqq_prio(struct cfq_queue *cfqq)
385 if (cfq_class_idle(cfqq))
386 return IDLE_WORKLOAD;
387 if (cfq_class_rt(cfqq))
393 static enum wl_type_t cfqq_type(struct cfq_queue *cfqq)
395 if (!cfq_cfqq_sync(cfqq))
396 return ASYNC_WORKLOAD;
397 if (!cfq_cfqq_idle_window(cfqq))
398 return SYNC_NOIDLE_WORKLOAD;
399 return SYNC_WORKLOAD;
402 static inline int cfq_group_busy_queues_wl(enum wl_prio_t wl,
403 struct cfq_data *cfqd,
404 struct cfq_group *cfqg)
406 if (wl == IDLE_WORKLOAD)
407 return cfqg->service_tree_idle.count;
409 return cfqg->service_trees[wl][ASYNC_WORKLOAD].count
410 + cfqg->service_trees[wl][SYNC_NOIDLE_WORKLOAD].count
411 + cfqg->service_trees[wl][SYNC_WORKLOAD].count;
414 static inline int cfqg_busy_async_queues(struct cfq_data *cfqd,
415 struct cfq_group *cfqg)
417 return cfqg->service_trees[RT_WORKLOAD][ASYNC_WORKLOAD].count
418 + cfqg->service_trees[BE_WORKLOAD][ASYNC_WORKLOAD].count;
421 static void cfq_dispatch_insert(struct request_queue *, struct request *);
422 static struct cfq_queue *cfq_get_queue(struct cfq_data *, bool,
423 struct io_context *, gfp_t);
424 static struct cfq_io_context *cfq_cic_lookup(struct cfq_data *,
425 struct io_context *);
427 static inline int rq_in_driver(struct cfq_data *cfqd)
429 return cfqd->rq_in_driver[0] + cfqd->rq_in_driver[1];
432 static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_context *cic,
435 return cic->cfqq[is_sync];
438 static inline void cic_set_cfqq(struct cfq_io_context *cic,
439 struct cfq_queue *cfqq, bool is_sync)
441 cic->cfqq[is_sync] = cfqq;
445 * We regard a request as SYNC, if it's either a read or has the SYNC bit
446 * set (in which case it could also be direct WRITE).
448 static inline bool cfq_bio_sync(struct bio *bio)
450 return bio_data_dir(bio) == READ || bio_rw_flagged(bio, BIO_RW_SYNCIO);
454 * scheduler run of queue, if there are requests pending and no one in the
455 * driver that will restart queueing
457 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
459 if (cfqd->busy_queues) {
460 cfq_log(cfqd, "schedule dispatch");
461 kblockd_schedule_work(cfqd->queue, &cfqd->unplug_work);
465 static int cfq_queue_empty(struct request_queue *q)
467 struct cfq_data *cfqd = q->elevator->elevator_data;
469 return !cfqd->rq_queued;
473 * Scale schedule slice based on io priority. Use the sync time slice only
474 * if a queue is marked sync and has sync io queued. A sync queue with async
475 * io only, should not get full sync slice length.
477 static inline int cfq_prio_slice(struct cfq_data *cfqd, bool sync,
480 const int base_slice = cfqd->cfq_slice[sync];
482 WARN_ON(prio >= IOPRIO_BE_NR);
484 return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
488 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
490 return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
493 static inline u64 cfq_scale_slice(unsigned long delta, struct cfq_group *cfqg)
495 u64 d = delta << CFQ_SERVICE_SHIFT;
497 d = d * BLKIO_WEIGHT_DEFAULT;
498 do_div(d, cfqg->weight);
502 static inline u64 max_vdisktime(u64 min_vdisktime, u64 vdisktime)
504 s64 delta = (s64)(vdisktime - min_vdisktime);
506 min_vdisktime = vdisktime;
508 return min_vdisktime;
511 static inline u64 min_vdisktime(u64 min_vdisktime, u64 vdisktime)
513 s64 delta = (s64)(vdisktime - min_vdisktime);
515 min_vdisktime = vdisktime;
517 return min_vdisktime;
520 static void update_min_vdisktime(struct cfq_rb_root *st)
522 u64 vdisktime = st->min_vdisktime;
523 struct cfq_group *cfqg;
526 cfqg = rb_entry_cfqg(st->active);
527 vdisktime = cfqg->vdisktime;
531 cfqg = rb_entry_cfqg(st->left);
532 vdisktime = min_vdisktime(vdisktime, cfqg->vdisktime);
535 st->min_vdisktime = max_vdisktime(st->min_vdisktime, vdisktime);
539 * get averaged number of queues of RT/BE priority.
540 * average is updated, with a formula that gives more weight to higher numbers,
541 * to quickly follows sudden increases and decrease slowly
544 static inline unsigned cfq_group_get_avg_queues(struct cfq_data *cfqd,
545 struct cfq_group *cfqg, bool rt)
547 unsigned min_q, max_q;
548 unsigned mult = cfq_hist_divisor - 1;
549 unsigned round = cfq_hist_divisor / 2;
550 unsigned busy = cfq_group_busy_queues_wl(rt, cfqd, cfqg);
552 min_q = min(cfqg->busy_queues_avg[rt], busy);
553 max_q = max(cfqg->busy_queues_avg[rt], busy);
554 cfqg->busy_queues_avg[rt] = (mult * max_q + min_q + round) /
556 return cfqg->busy_queues_avg[rt];
559 static inline unsigned
560 cfq_group_slice(struct cfq_data *cfqd, struct cfq_group *cfqg)
562 struct cfq_rb_root *st = &cfqd->grp_service_tree;
564 return cfq_target_latency * cfqg->weight / st->total_weight;
568 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
570 unsigned slice = cfq_prio_to_slice(cfqd, cfqq);
571 if (cfqd->cfq_latency) {
573 * interested queues (we consider only the ones with the same
574 * priority class in the cfq group)
576 unsigned iq = cfq_group_get_avg_queues(cfqd, cfqq->cfqg,
578 unsigned sync_slice = cfqd->cfq_slice[1];
579 unsigned expect_latency = sync_slice * iq;
580 unsigned group_slice = cfq_group_slice(cfqd, cfqq->cfqg);
582 if (expect_latency > group_slice) {
583 unsigned base_low_slice = 2 * cfqd->cfq_slice_idle;
584 /* scale low_slice according to IO priority
585 * and sync vs async */
587 min(slice, base_low_slice * slice / sync_slice);
588 /* the adapted slice value is scaled to fit all iqs
589 * into the target latency */
590 slice = max(slice * group_slice / expect_latency,
594 cfqq->slice_start = jiffies;
595 cfqq->slice_end = jiffies + slice;
596 cfqq->allocated_slice = slice;
597 cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
601 * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
602 * isn't valid until the first request from the dispatch is activated
603 * and the slice time set.
605 static inline bool cfq_slice_used(struct cfq_queue *cfqq)
607 if (cfq_cfqq_slice_new(cfqq))
609 if (time_before(jiffies, cfqq->slice_end))
616 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
617 * We choose the request that is closest to the head right now. Distance
618 * behind the head is penalized and only allowed to a certain extent.
620 static struct request *
621 cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2, sector_t last)
623 sector_t s1, s2, d1 = 0, d2 = 0;
624 unsigned long back_max;
625 #define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */
626 #define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */
627 unsigned wrap = 0; /* bit mask: requests behind the disk head? */
629 if (rq1 == NULL || rq1 == rq2)
634 if (rq_is_sync(rq1) && !rq_is_sync(rq2))
636 else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
638 if (rq_is_meta(rq1) && !rq_is_meta(rq2))
640 else if (rq_is_meta(rq2) && !rq_is_meta(rq1))
643 s1 = blk_rq_pos(rq1);
644 s2 = blk_rq_pos(rq2);
647 * by definition, 1KiB is 2 sectors
649 back_max = cfqd->cfq_back_max * 2;
652 * Strict one way elevator _except_ in the case where we allow
653 * short backward seeks which are biased as twice the cost of a
654 * similar forward seek.
658 else if (s1 + back_max >= last)
659 d1 = (last - s1) * cfqd->cfq_back_penalty;
661 wrap |= CFQ_RQ1_WRAP;
665 else if (s2 + back_max >= last)
666 d2 = (last - s2) * cfqd->cfq_back_penalty;
668 wrap |= CFQ_RQ2_WRAP;
670 /* Found required data */
673 * By doing switch() on the bit mask "wrap" we avoid having to
674 * check two variables for all permutations: --> faster!
677 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
693 case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
696 * Since both rqs are wrapped,
697 * start with the one that's further behind head
698 * (--> only *one* back seek required),
699 * since back seek takes more time than forward.
709 * The below is leftmost cache rbtree addon
711 static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
713 /* Service tree is empty */
718 root->left = rb_first(&root->rb);
721 return rb_entry(root->left, struct cfq_queue, rb_node);
726 static struct cfq_group *cfq_rb_first_group(struct cfq_rb_root *root)
729 root->left = rb_first(&root->rb);
732 return rb_entry_cfqg(root->left);
737 static void rb_erase_init(struct rb_node *n, struct rb_root *root)
743 static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
747 rb_erase_init(n, &root->rb);
752 * would be nice to take fifo expire time into account as well
754 static struct request *
755 cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
756 struct request *last)
758 struct rb_node *rbnext = rb_next(&last->rb_node);
759 struct rb_node *rbprev = rb_prev(&last->rb_node);
760 struct request *next = NULL, *prev = NULL;
762 BUG_ON(RB_EMPTY_NODE(&last->rb_node));
765 prev = rb_entry_rq(rbprev);
768 next = rb_entry_rq(rbnext);
770 rbnext = rb_first(&cfqq->sort_list);
771 if (rbnext && rbnext != &last->rb_node)
772 next = rb_entry_rq(rbnext);
775 return cfq_choose_req(cfqd, next, prev, blk_rq_pos(last));
778 static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
779 struct cfq_queue *cfqq)
782 * just an approximation, should be ok.
784 return (cfqq->cfqg->nr_cfqq - 1) * (cfq_prio_slice(cfqd, 1, 0) -
785 cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
789 cfqg_key(struct cfq_rb_root *st, struct cfq_group *cfqg)
791 return cfqg->vdisktime - st->min_vdisktime;
795 __cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
797 struct rb_node **node = &st->rb.rb_node;
798 struct rb_node *parent = NULL;
799 struct cfq_group *__cfqg;
800 s64 key = cfqg_key(st, cfqg);
803 while (*node != NULL) {
805 __cfqg = rb_entry_cfqg(parent);
807 if (key < cfqg_key(st, __cfqg))
808 node = &parent->rb_left;
810 node = &parent->rb_right;
816 st->left = &cfqg->rb_node;
818 rb_link_node(&cfqg->rb_node, parent, node);
819 rb_insert_color(&cfqg->rb_node, &st->rb);
823 cfq_group_service_tree_add(struct cfq_data *cfqd, struct cfq_group *cfqg)
825 struct cfq_rb_root *st = &cfqd->grp_service_tree;
826 struct cfq_group *__cfqg;
834 * Currently put the group at the end. Later implement something
835 * so that groups get lesser vtime based on their weights, so that
836 * if group does not loose all if it was not continously backlogged.
838 n = rb_last(&st->rb);
840 __cfqg = rb_entry_cfqg(n);
841 cfqg->vdisktime = __cfqg->vdisktime + CFQ_IDLE_DELAY;
843 cfqg->vdisktime = st->min_vdisktime;
845 __cfq_group_service_tree_add(st, cfqg);
848 st->total_weight += cfqg->weight;
852 cfq_group_service_tree_del(struct cfq_data *cfqd, struct cfq_group *cfqg)
854 struct cfq_rb_root *st = &cfqd->grp_service_tree;
856 if (st->active == &cfqg->rb_node)
859 BUG_ON(cfqg->nr_cfqq < 1);
862 /* If there are other cfq queues under this group, don't delete it */
866 cfq_log_cfqg(cfqd, cfqg, "del_from_rr group");
869 st->total_weight -= cfqg->weight;
870 if (!RB_EMPTY_NODE(&cfqg->rb_node))
871 cfq_rb_erase(&cfqg->rb_node, st);
872 cfqg->saved_workload_slice = 0;
873 blkiocg_update_blkio_group_dequeue_stats(&cfqg->blkg, 1);
876 static inline unsigned int cfq_cfqq_slice_usage(struct cfq_queue *cfqq)
878 unsigned int slice_used;
881 * Queue got expired before even a single request completed or
882 * got expired immediately after first request completion.
884 if (!cfqq->slice_start || cfqq->slice_start == jiffies) {
886 * Also charge the seek time incurred to the group, otherwise
887 * if there are mutiple queues in the group, each can dispatch
888 * a single request on seeky media and cause lots of seek time
889 * and group will never know it.
891 slice_used = max_t(unsigned, (jiffies - cfqq->dispatch_start),
894 slice_used = jiffies - cfqq->slice_start;
895 if (slice_used > cfqq->allocated_slice)
896 slice_used = cfqq->allocated_slice;
899 cfq_log_cfqq(cfqq->cfqd, cfqq, "sl_used=%u sect=%lu", slice_used,
904 static void cfq_group_served(struct cfq_data *cfqd, struct cfq_group *cfqg,
905 struct cfq_queue *cfqq)
907 struct cfq_rb_root *st = &cfqd->grp_service_tree;
908 unsigned int used_sl, charge_sl;
909 int nr_sync = cfqg->nr_cfqq - cfqg_busy_async_queues(cfqd, cfqg)
910 - cfqg->service_tree_idle.count;
913 used_sl = charge_sl = cfq_cfqq_slice_usage(cfqq);
915 if (!cfq_cfqq_sync(cfqq) && !nr_sync)
916 charge_sl = cfqq->allocated_slice;
918 /* Can't update vdisktime while group is on service tree */
919 cfq_rb_erase(&cfqg->rb_node, st);
920 cfqg->vdisktime += cfq_scale_slice(charge_sl, cfqg);
921 __cfq_group_service_tree_add(st, cfqg);
923 /* This group is being expired. Save the context */
924 if (time_after(cfqd->workload_expires, jiffies)) {
925 cfqg->saved_workload_slice = cfqd->workload_expires
927 cfqg->saved_workload = cfqd->serving_type;
928 cfqg->saved_serving_prio = cfqd->serving_prio;
930 cfqg->saved_workload_slice = 0;
932 cfq_log_cfqg(cfqd, cfqg, "served: vt=%llu min_vt=%llu", cfqg->vdisktime,
934 blkiocg_update_blkio_group_stats(&cfqg->blkg, used_sl,
938 #ifdef CONFIG_CFQ_GROUP_IOSCHED
939 static inline struct cfq_group *cfqg_of_blkg(struct blkio_group *blkg)
942 return container_of(blkg, struct cfq_group, blkg);
947 cfq_update_blkio_group_weight(struct blkio_group *blkg, unsigned int weight)
949 cfqg_of_blkg(blkg)->weight = weight;
952 static struct cfq_group *
953 cfq_find_alloc_cfqg(struct cfq_data *cfqd, struct cgroup *cgroup, int create)
955 struct blkio_cgroup *blkcg = cgroup_to_blkio_cgroup(cgroup);
956 struct cfq_group *cfqg = NULL;
959 struct cfq_rb_root *st;
960 struct backing_dev_info *bdi = &cfqd->queue->backing_dev_info;
961 unsigned int major, minor;
963 /* Do we need to take this reference */
964 if (!blkiocg_css_tryget(blkcg))
967 cfqg = cfqg_of_blkg(blkiocg_lookup_group(blkcg, key));
971 cfqg = kzalloc_node(sizeof(*cfqg), GFP_ATOMIC, cfqd->queue->node);
975 cfqg->weight = blkcg->weight;
976 for_each_cfqg_st(cfqg, i, j, st)
978 RB_CLEAR_NODE(&cfqg->rb_node);
981 * Take the initial reference that will be released on destroy
982 * This can be thought of a joint reference by cgroup and
983 * elevator which will be dropped by either elevator exit
984 * or cgroup deletion path depending on who is exiting first.
986 atomic_set(&cfqg->ref, 1);
988 /* Add group onto cgroup list */
989 sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
990 blkiocg_add_blkio_group(blkcg, &cfqg->blkg, (void *)cfqd,
991 MKDEV(major, minor));
993 /* Add group on cfqd list */
994 hlist_add_head(&cfqg->cfqd_node, &cfqd->cfqg_list);
997 blkiocg_css_put(blkcg);
1002 * Search for the cfq group current task belongs to. If create = 1, then also
1003 * create the cfq group if it does not exist. request_queue lock must be held.
1005 static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd, int create)
1007 struct cgroup *cgroup;
1008 struct cfq_group *cfqg = NULL;
1011 cgroup = task_cgroup(current, blkio_subsys_id);
1012 cfqg = cfq_find_alloc_cfqg(cfqd, cgroup, create);
1013 if (!cfqg && create)
1014 cfqg = &cfqd->root_group;
1019 static void cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg)
1021 /* Currently, all async queues are mapped to root group */
1022 if (!cfq_cfqq_sync(cfqq))
1023 cfqg = &cfqq->cfqd->root_group;
1026 /* cfqq reference on cfqg */
1027 atomic_inc(&cfqq->cfqg->ref);
1030 static void cfq_put_cfqg(struct cfq_group *cfqg)
1032 struct cfq_rb_root *st;
1035 BUG_ON(atomic_read(&cfqg->ref) <= 0);
1036 if (!atomic_dec_and_test(&cfqg->ref))
1038 for_each_cfqg_st(cfqg, i, j, st)
1039 BUG_ON(!RB_EMPTY_ROOT(&st->rb) || st->active != NULL);
1043 static void cfq_destroy_cfqg(struct cfq_data *cfqd, struct cfq_group *cfqg)
1045 /* Something wrong if we are trying to remove same group twice */
1046 BUG_ON(hlist_unhashed(&cfqg->cfqd_node));
1048 hlist_del_init(&cfqg->cfqd_node);
1051 * Put the reference taken at the time of creation so that when all
1052 * queues are gone, group can be destroyed.
1057 static void cfq_release_cfq_groups(struct cfq_data *cfqd)
1059 struct hlist_node *pos, *n;
1060 struct cfq_group *cfqg;
1062 hlist_for_each_entry_safe(cfqg, pos, n, &cfqd->cfqg_list, cfqd_node) {
1064 * If cgroup removal path got to blk_group first and removed
1065 * it from cgroup list, then it will take care of destroying
1068 if (!blkiocg_del_blkio_group(&cfqg->blkg))
1069 cfq_destroy_cfqg(cfqd, cfqg);
1074 * Blk cgroup controller notification saying that blkio_group object is being
1075 * delinked as associated cgroup object is going away. That also means that
1076 * no new IO will come in this group. So get rid of this group as soon as
1077 * any pending IO in the group is finished.
1079 * This function is called under rcu_read_lock(). key is the rcu protected
1080 * pointer. That means "key" is a valid cfq_data pointer as long as we are rcu
1083 * "key" was fetched from blkio_group under blkio_cgroup->lock. That means
1084 * it should not be NULL as even if elevator was exiting, cgroup deltion
1085 * path got to it first.
1087 void cfq_unlink_blkio_group(void *key, struct blkio_group *blkg)
1089 unsigned long flags;
1090 struct cfq_data *cfqd = key;
1092 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1093 cfq_destroy_cfqg(cfqd, cfqg_of_blkg(blkg));
1094 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1097 #else /* GROUP_IOSCHED */
1098 static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd, int create)
1100 return &cfqd->root_group;
1103 cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) {
1107 static void cfq_release_cfq_groups(struct cfq_data *cfqd) {}
1108 static inline void cfq_put_cfqg(struct cfq_group *cfqg) {}
1110 #endif /* GROUP_IOSCHED */
1113 * The cfqd->service_trees holds all pending cfq_queue's that have
1114 * requests waiting to be processed. It is sorted in the order that
1115 * we will service the queues.
1117 static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1120 struct rb_node **p, *parent;
1121 struct cfq_queue *__cfqq;
1122 unsigned long rb_key;
1123 struct cfq_rb_root *service_tree;
1126 int group_changed = 0;
1128 #ifdef CONFIG_CFQ_GROUP_IOSCHED
1129 if (!cfqd->cfq_group_isolation
1130 && cfqq_type(cfqq) == SYNC_NOIDLE_WORKLOAD
1131 && cfqq->cfqg && cfqq->cfqg != &cfqd->root_group) {
1132 /* Move this cfq to root group */
1133 cfq_log_cfqq(cfqd, cfqq, "moving to root group");
1134 if (!RB_EMPTY_NODE(&cfqq->rb_node))
1135 cfq_group_service_tree_del(cfqd, cfqq->cfqg);
1136 cfqq->orig_cfqg = cfqq->cfqg;
1137 cfqq->cfqg = &cfqd->root_group;
1138 atomic_inc(&cfqd->root_group.ref);
1140 } else if (!cfqd->cfq_group_isolation
1141 && cfqq_type(cfqq) == SYNC_WORKLOAD && cfqq->orig_cfqg) {
1142 /* cfqq is sequential now needs to go to its original group */
1143 BUG_ON(cfqq->cfqg != &cfqd->root_group);
1144 if (!RB_EMPTY_NODE(&cfqq->rb_node))
1145 cfq_group_service_tree_del(cfqd, cfqq->cfqg);
1146 cfq_put_cfqg(cfqq->cfqg);
1147 cfqq->cfqg = cfqq->orig_cfqg;
1148 cfqq->orig_cfqg = NULL;
1150 cfq_log_cfqq(cfqd, cfqq, "moved to origin group");
1154 service_tree = service_tree_for(cfqq->cfqg, cfqq_prio(cfqq),
1155 cfqq_type(cfqq), cfqd);
1156 if (cfq_class_idle(cfqq)) {
1157 rb_key = CFQ_IDLE_DELAY;
1158 parent = rb_last(&service_tree->rb);
1159 if (parent && parent != &cfqq->rb_node) {
1160 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1161 rb_key += __cfqq->rb_key;
1164 } else if (!add_front) {
1166 * Get our rb key offset. Subtract any residual slice
1167 * value carried from last service. A negative resid
1168 * count indicates slice overrun, and this should position
1169 * the next service time further away in the tree.
1171 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
1172 rb_key -= cfqq->slice_resid;
1173 cfqq->slice_resid = 0;
1176 __cfqq = cfq_rb_first(service_tree);
1177 rb_key += __cfqq ? __cfqq->rb_key : jiffies;
1180 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1183 * same position, nothing more to do
1185 if (rb_key == cfqq->rb_key &&
1186 cfqq->service_tree == service_tree)
1189 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1190 cfqq->service_tree = NULL;
1195 cfqq->service_tree = service_tree;
1196 p = &service_tree->rb.rb_node;
1201 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1204 * sort by key, that represents service time.
1206 if (time_before(rb_key, __cfqq->rb_key))
1209 n = &(*p)->rb_right;
1217 service_tree->left = &cfqq->rb_node;
1219 cfqq->rb_key = rb_key;
1220 rb_link_node(&cfqq->rb_node, parent, p);
1221 rb_insert_color(&cfqq->rb_node, &service_tree->rb);
1222 service_tree->count++;
1223 if ((add_front || !new_cfqq) && !group_changed)
1225 cfq_group_service_tree_add(cfqd, cfqq->cfqg);
1228 static struct cfq_queue *
1229 cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
1230 sector_t sector, struct rb_node **ret_parent,
1231 struct rb_node ***rb_link)
1233 struct rb_node **p, *parent;
1234 struct cfq_queue *cfqq = NULL;
1242 cfqq = rb_entry(parent, struct cfq_queue, p_node);
1245 * Sort strictly based on sector. Smallest to the left,
1246 * largest to the right.
1248 if (sector > blk_rq_pos(cfqq->next_rq))
1249 n = &(*p)->rb_right;
1250 else if (sector < blk_rq_pos(cfqq->next_rq))
1258 *ret_parent = parent;
1264 static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1266 struct rb_node **p, *parent;
1267 struct cfq_queue *__cfqq;
1270 rb_erase(&cfqq->p_node, cfqq->p_root);
1271 cfqq->p_root = NULL;
1274 if (cfq_class_idle(cfqq))
1279 cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
1280 __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
1281 blk_rq_pos(cfqq->next_rq), &parent, &p);
1283 rb_link_node(&cfqq->p_node, parent, p);
1284 rb_insert_color(&cfqq->p_node, cfqq->p_root);
1286 cfqq->p_root = NULL;
1290 * Update cfqq's position in the service tree.
1292 static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1295 * Resorting requires the cfqq to be on the RR list already.
1297 if (cfq_cfqq_on_rr(cfqq)) {
1298 cfq_service_tree_add(cfqd, cfqq, 0);
1299 cfq_prio_tree_add(cfqd, cfqq);
1304 * add to busy list of queues for service, trying to be fair in ordering
1305 * the pending list according to last request service
1307 static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1309 cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
1310 BUG_ON(cfq_cfqq_on_rr(cfqq));
1311 cfq_mark_cfqq_on_rr(cfqq);
1312 cfqd->busy_queues++;
1314 cfq_resort_rr_list(cfqd, cfqq);
1318 * Called when the cfqq no longer has requests pending, remove it from
1321 static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1323 cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
1324 BUG_ON(!cfq_cfqq_on_rr(cfqq));
1325 cfq_clear_cfqq_on_rr(cfqq);
1327 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1328 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1329 cfqq->service_tree = NULL;
1332 rb_erase(&cfqq->p_node, cfqq->p_root);
1333 cfqq->p_root = NULL;
1336 cfq_group_service_tree_del(cfqd, cfqq->cfqg);
1337 BUG_ON(!cfqd->busy_queues);
1338 cfqd->busy_queues--;
1342 * rb tree support functions
1344 static void cfq_del_rq_rb(struct request *rq)
1346 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1347 const int sync = rq_is_sync(rq);
1349 BUG_ON(!cfqq->queued[sync]);
1350 cfqq->queued[sync]--;
1352 elv_rb_del(&cfqq->sort_list, rq);
1354 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) {
1356 * Queue will be deleted from service tree when we actually
1357 * expire it later. Right now just remove it from prio tree
1361 rb_erase(&cfqq->p_node, cfqq->p_root);
1362 cfqq->p_root = NULL;
1367 static void cfq_add_rq_rb(struct request *rq)
1369 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1370 struct cfq_data *cfqd = cfqq->cfqd;
1371 struct request *__alias, *prev;
1373 cfqq->queued[rq_is_sync(rq)]++;
1376 * looks a little odd, but the first insert might return an alias.
1377 * if that happens, put the alias on the dispatch list
1379 while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL)
1380 cfq_dispatch_insert(cfqd->queue, __alias);
1382 if (!cfq_cfqq_on_rr(cfqq))
1383 cfq_add_cfqq_rr(cfqd, cfqq);
1386 * check if this request is a better next-serve candidate
1388 prev = cfqq->next_rq;
1389 cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq, cfqd->last_position);
1392 * adjust priority tree position, if ->next_rq changes
1394 if (prev != cfqq->next_rq)
1395 cfq_prio_tree_add(cfqd, cfqq);
1397 BUG_ON(!cfqq->next_rq);
1400 static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
1402 elv_rb_del(&cfqq->sort_list, rq);
1403 cfqq->queued[rq_is_sync(rq)]--;
1407 static struct request *
1408 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
1410 struct task_struct *tsk = current;
1411 struct cfq_io_context *cic;
1412 struct cfq_queue *cfqq;
1414 cic = cfq_cic_lookup(cfqd, tsk->io_context);
1418 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
1420 sector_t sector = bio->bi_sector + bio_sectors(bio);
1422 return elv_rb_find(&cfqq->sort_list, sector);
1428 static void cfq_activate_request(struct request_queue *q, struct request *rq)
1430 struct cfq_data *cfqd = q->elevator->elevator_data;
1432 cfqd->rq_in_driver[rq_is_sync(rq)]++;
1433 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
1434 rq_in_driver(cfqd));
1436 cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
1439 static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
1441 struct cfq_data *cfqd = q->elevator->elevator_data;
1442 const int sync = rq_is_sync(rq);
1444 WARN_ON(!cfqd->rq_in_driver[sync]);
1445 cfqd->rq_in_driver[sync]--;
1446 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
1447 rq_in_driver(cfqd));
1450 static void cfq_remove_request(struct request *rq)
1452 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1454 if (cfqq->next_rq == rq)
1455 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
1457 list_del_init(&rq->queuelist);
1460 cfqq->cfqd->rq_queued--;
1461 if (rq_is_meta(rq)) {
1462 WARN_ON(!cfqq->meta_pending);
1463 cfqq->meta_pending--;
1467 static int cfq_merge(struct request_queue *q, struct request **req,
1470 struct cfq_data *cfqd = q->elevator->elevator_data;
1471 struct request *__rq;
1473 __rq = cfq_find_rq_fmerge(cfqd, bio);
1474 if (__rq && elv_rq_merge_ok(__rq, bio)) {
1476 return ELEVATOR_FRONT_MERGE;
1479 return ELEVATOR_NO_MERGE;
1482 static void cfq_merged_request(struct request_queue *q, struct request *req,
1485 if (type == ELEVATOR_FRONT_MERGE) {
1486 struct cfq_queue *cfqq = RQ_CFQQ(req);
1488 cfq_reposition_rq_rb(cfqq, req);
1493 cfq_merged_requests(struct request_queue *q, struct request *rq,
1494 struct request *next)
1496 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1498 * reposition in fifo if next is older than rq
1500 if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
1501 time_before(rq_fifo_time(next), rq_fifo_time(rq))) {
1502 list_move(&rq->queuelist, &next->queuelist);
1503 rq_set_fifo_time(rq, rq_fifo_time(next));
1506 if (cfqq->next_rq == next)
1508 cfq_remove_request(next);
1511 static int cfq_allow_merge(struct request_queue *q, struct request *rq,
1514 struct cfq_data *cfqd = q->elevator->elevator_data;
1515 struct cfq_io_context *cic;
1516 struct cfq_queue *cfqq;
1518 /* Deny merge if bio and rq don't belong to same cfq group */
1519 if ((RQ_CFQQ(rq))->cfqg != cfq_get_cfqg(cfqd, 0))
1522 * Disallow merge of a sync bio into an async request.
1524 if (cfq_bio_sync(bio) && !rq_is_sync(rq))
1528 * Lookup the cfqq that this bio will be queued with. Allow
1529 * merge only if rq is queued there.
1531 cic = cfq_cic_lookup(cfqd, current->io_context);
1535 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
1536 return cfqq == RQ_CFQQ(rq);
1539 static void __cfq_set_active_queue(struct cfq_data *cfqd,
1540 struct cfq_queue *cfqq)
1543 cfq_log_cfqq(cfqd, cfqq, "set_active");
1544 cfqq->slice_start = 0;
1545 cfqq->dispatch_start = jiffies;
1546 cfqq->allocated_slice = 0;
1547 cfqq->slice_end = 0;
1548 cfqq->slice_dispatch = 0;
1549 cfqq->nr_sectors = 0;
1551 cfq_clear_cfqq_wait_request(cfqq);
1552 cfq_clear_cfqq_must_dispatch(cfqq);
1553 cfq_clear_cfqq_must_alloc_slice(cfqq);
1554 cfq_clear_cfqq_fifo_expire(cfqq);
1555 cfq_mark_cfqq_slice_new(cfqq);
1557 del_timer(&cfqd->idle_slice_timer);
1560 cfqd->active_queue = cfqq;
1564 * current cfqq expired its slice (or was too idle), select new one
1567 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1570 cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
1572 if (cfq_cfqq_wait_request(cfqq))
1573 del_timer(&cfqd->idle_slice_timer);
1575 cfq_clear_cfqq_wait_request(cfqq);
1576 cfq_clear_cfqq_wait_busy(cfqq);
1577 cfq_clear_cfqq_wait_busy_done(cfqq);
1580 * store what was left of this slice, if the queue idled/timed out
1582 if (timed_out && !cfq_cfqq_slice_new(cfqq)) {
1583 cfqq->slice_resid = cfqq->slice_end - jiffies;
1584 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
1587 cfq_group_served(cfqd, cfqq->cfqg, cfqq);
1589 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
1590 cfq_del_cfqq_rr(cfqd, cfqq);
1592 cfq_resort_rr_list(cfqd, cfqq);
1594 if (cfqq == cfqd->active_queue)
1595 cfqd->active_queue = NULL;
1597 if (&cfqq->cfqg->rb_node == cfqd->grp_service_tree.active)
1598 cfqd->grp_service_tree.active = NULL;
1600 if (cfqd->active_cic) {
1601 put_io_context(cfqd->active_cic->ioc);
1602 cfqd->active_cic = NULL;
1606 static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out)
1608 struct cfq_queue *cfqq = cfqd->active_queue;
1611 __cfq_slice_expired(cfqd, cfqq, timed_out);
1615 * Get next queue for service. Unless we have a queue preemption,
1616 * we'll simply select the first cfqq in the service tree.
1618 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
1620 struct cfq_rb_root *service_tree =
1621 service_tree_for(cfqd->serving_group, cfqd->serving_prio,
1622 cfqd->serving_type, cfqd);
1624 if (!cfqd->rq_queued)
1627 /* There is nothing to dispatch */
1630 if (RB_EMPTY_ROOT(&service_tree->rb))
1632 return cfq_rb_first(service_tree);
1635 static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd)
1637 struct cfq_group *cfqg;
1638 struct cfq_queue *cfqq;
1640 struct cfq_rb_root *st;
1642 if (!cfqd->rq_queued)
1645 cfqg = cfq_get_next_cfqg(cfqd);
1649 for_each_cfqg_st(cfqg, i, j, st)
1650 if ((cfqq = cfq_rb_first(st)) != NULL)
1656 * Get and set a new active queue for service.
1658 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
1659 struct cfq_queue *cfqq)
1662 cfqq = cfq_get_next_queue(cfqd);
1664 __cfq_set_active_queue(cfqd, cfqq);
1668 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
1671 if (blk_rq_pos(rq) >= cfqd->last_position)
1672 return blk_rq_pos(rq) - cfqd->last_position;
1674 return cfqd->last_position - blk_rq_pos(rq);
1677 #define CFQQ_SEEK_THR 8 * 1024
1678 #define CFQQ_SEEKY(cfqq) ((cfqq)->seek_mean > CFQQ_SEEK_THR)
1680 static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1683 sector_t sdist = cfqq->seek_mean;
1685 if (!sample_valid(cfqq->seek_samples))
1686 sdist = CFQQ_SEEK_THR;
1688 return cfq_dist_from_last(cfqd, rq) <= sdist;
1691 static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
1692 struct cfq_queue *cur_cfqq)
1694 struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
1695 struct rb_node *parent, *node;
1696 struct cfq_queue *__cfqq;
1697 sector_t sector = cfqd->last_position;
1699 if (RB_EMPTY_ROOT(root))
1703 * First, if we find a request starting at the end of the last
1704 * request, choose it.
1706 __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
1711 * If the exact sector wasn't found, the parent of the NULL leaf
1712 * will contain the closest sector.
1714 __cfqq = rb_entry(parent, struct cfq_queue, p_node);
1715 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1718 if (blk_rq_pos(__cfqq->next_rq) < sector)
1719 node = rb_next(&__cfqq->p_node);
1721 node = rb_prev(&__cfqq->p_node);
1725 __cfqq = rb_entry(node, struct cfq_queue, p_node);
1726 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1734 * cur_cfqq - passed in so that we don't decide that the current queue is
1735 * closely cooperating with itself.
1737 * So, basically we're assuming that that cur_cfqq has dispatched at least
1738 * one request, and that cfqd->last_position reflects a position on the disk
1739 * associated with the I/O issued by cur_cfqq. I'm not sure this is a valid
1742 static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
1743 struct cfq_queue *cur_cfqq)
1745 struct cfq_queue *cfqq;
1747 if (!cfq_cfqq_sync(cur_cfqq))
1749 if (CFQQ_SEEKY(cur_cfqq))
1753 * Don't search priority tree if it's the only queue in the group.
1755 if (cur_cfqq->cfqg->nr_cfqq == 1)
1759 * We should notice if some of the queues are cooperating, eg
1760 * working closely on the same area of the disk. In that case,
1761 * we can group them together and don't waste time idling.
1763 cfqq = cfqq_close(cfqd, cur_cfqq);
1767 /* If new queue belongs to different cfq_group, don't choose it */
1768 if (cur_cfqq->cfqg != cfqq->cfqg)
1772 * It only makes sense to merge sync queues.
1774 if (!cfq_cfqq_sync(cfqq))
1776 if (CFQQ_SEEKY(cfqq))
1780 * Do not merge queues of different priority classes
1782 if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq))
1789 * Determine whether we should enforce idle window for this queue.
1792 static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1794 enum wl_prio_t prio = cfqq_prio(cfqq);
1795 struct cfq_rb_root *service_tree = cfqq->service_tree;
1797 BUG_ON(!service_tree);
1798 BUG_ON(!service_tree->count);
1800 /* We never do for idle class queues. */
1801 if (prio == IDLE_WORKLOAD)
1804 /* We do for queues that were marked with idle window flag. */
1805 if (cfq_cfqq_idle_window(cfqq) &&
1806 !(blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag))
1810 * Otherwise, we do only if they are the last ones
1811 * in their service tree.
1813 return service_tree->count == 1;
1816 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
1818 struct cfq_queue *cfqq = cfqd->active_queue;
1819 struct cfq_io_context *cic;
1823 * SSD device without seek penalty, disable idling. But only do so
1824 * for devices that support queuing, otherwise we still have a problem
1825 * with sync vs async workloads.
1827 if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
1830 WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
1831 WARN_ON(cfq_cfqq_slice_new(cfqq));
1834 * idle is disabled, either manually or by past process history
1836 if (!cfqd->cfq_slice_idle || !cfq_should_idle(cfqd, cfqq))
1840 * still active requests from this queue, don't idle
1842 if (cfqq->dispatched)
1846 * task has exited, don't wait
1848 cic = cfqd->active_cic;
1849 if (!cic || !atomic_read(&cic->ioc->nr_tasks))
1853 * If our average think time is larger than the remaining time
1854 * slice, then don't idle. This avoids overrunning the allotted
1857 if (sample_valid(cic->ttime_samples) &&
1858 (cfqq->slice_end - jiffies < cic->ttime_mean))
1861 cfq_mark_cfqq_wait_request(cfqq);
1863 sl = cfqd->cfq_slice_idle;
1865 mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
1866 cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu", sl);
1870 * Move request from internal lists to the request queue dispatch list.
1872 static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
1874 struct cfq_data *cfqd = q->elevator->elevator_data;
1875 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1877 cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
1879 cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
1880 cfq_remove_request(rq);
1882 elv_dispatch_sort(q, rq);
1884 if (cfq_cfqq_sync(cfqq))
1885 cfqd->sync_flight++;
1886 cfqq->nr_sectors += blk_rq_sectors(rq);
1890 * return expired entry, or NULL to just start from scratch in rbtree
1892 static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
1894 struct request *rq = NULL;
1896 if (cfq_cfqq_fifo_expire(cfqq))
1899 cfq_mark_cfqq_fifo_expire(cfqq);
1901 if (list_empty(&cfqq->fifo))
1904 rq = rq_entry_fifo(cfqq->fifo.next);
1905 if (time_before(jiffies, rq_fifo_time(rq)))
1908 cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq);
1913 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1915 const int base_rq = cfqd->cfq_slice_async_rq;
1917 WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
1919 return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
1923 * Must be called with the queue_lock held.
1925 static int cfqq_process_refs(struct cfq_queue *cfqq)
1927 int process_refs, io_refs;
1929 io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE];
1930 process_refs = atomic_read(&cfqq->ref) - io_refs;
1931 BUG_ON(process_refs < 0);
1932 return process_refs;
1935 static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq)
1937 int process_refs, new_process_refs;
1938 struct cfq_queue *__cfqq;
1940 /* Avoid a circular list and skip interim queue merges */
1941 while ((__cfqq = new_cfqq->new_cfqq)) {
1947 process_refs = cfqq_process_refs(cfqq);
1949 * If the process for the cfqq has gone away, there is no
1950 * sense in merging the queues.
1952 if (process_refs == 0)
1956 * Merge in the direction of the lesser amount of work.
1958 new_process_refs = cfqq_process_refs(new_cfqq);
1959 if (new_process_refs >= process_refs) {
1960 cfqq->new_cfqq = new_cfqq;
1961 atomic_add(process_refs, &new_cfqq->ref);
1963 new_cfqq->new_cfqq = cfqq;
1964 atomic_add(new_process_refs, &cfqq->ref);
1968 static enum wl_type_t cfq_choose_wl(struct cfq_data *cfqd,
1969 struct cfq_group *cfqg, enum wl_prio_t prio,
1972 struct cfq_queue *queue;
1974 bool key_valid = false;
1975 unsigned long lowest_key = 0;
1976 enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD;
1980 * When priorities switched, we prefer starting
1981 * from SYNC_NOIDLE (first choice), or just SYNC
1984 if (service_tree_for(cfqg, prio, cur_best, cfqd)->count)
1986 cur_best = SYNC_WORKLOAD;
1987 if (service_tree_for(cfqg, prio, cur_best, cfqd)->count)
1990 return ASYNC_WORKLOAD;
1993 for (i = 0; i < 3; ++i) {
1994 /* otherwise, select the one with lowest rb_key */
1995 queue = cfq_rb_first(service_tree_for(cfqg, prio, i, cfqd));
1997 (!key_valid || time_before(queue->rb_key, lowest_key))) {
1998 lowest_key = queue->rb_key;
2007 static void choose_service_tree(struct cfq_data *cfqd, struct cfq_group *cfqg)
2009 enum wl_prio_t previous_prio = cfqd->serving_prio;
2013 struct cfq_rb_root *st;
2014 unsigned group_slice;
2017 cfqd->serving_prio = IDLE_WORKLOAD;
2018 cfqd->workload_expires = jiffies + 1;
2022 /* Choose next priority. RT > BE > IDLE */
2023 if (cfq_group_busy_queues_wl(RT_WORKLOAD, cfqd, cfqg))
2024 cfqd->serving_prio = RT_WORKLOAD;
2025 else if (cfq_group_busy_queues_wl(BE_WORKLOAD, cfqd, cfqg))
2026 cfqd->serving_prio = BE_WORKLOAD;
2028 cfqd->serving_prio = IDLE_WORKLOAD;
2029 cfqd->workload_expires = jiffies + 1;
2034 * For RT and BE, we have to choose also the type
2035 * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
2038 prio_changed = (cfqd->serving_prio != previous_prio);
2039 st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type,
2044 * If priority didn't change, check workload expiration,
2045 * and that we still have other queues ready
2047 if (!prio_changed && count &&
2048 !time_after(jiffies, cfqd->workload_expires))
2051 /* otherwise select new workload type */
2052 cfqd->serving_type =
2053 cfq_choose_wl(cfqd, cfqg, cfqd->serving_prio, prio_changed);
2054 st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type,
2059 * the workload slice is computed as a fraction of target latency
2060 * proportional to the number of queues in that workload, over
2061 * all the queues in the same priority class
2063 group_slice = cfq_group_slice(cfqd, cfqg);
2065 slice = group_slice * count /
2066 max_t(unsigned, cfqg->busy_queues_avg[cfqd->serving_prio],
2067 cfq_group_busy_queues_wl(cfqd->serving_prio, cfqd, cfqg));
2069 if (cfqd->serving_type == ASYNC_WORKLOAD) {
2073 * Async queues are currently system wide. Just taking
2074 * proportion of queues with-in same group will lead to higher
2075 * async ratio system wide as generally root group is going
2076 * to have higher weight. A more accurate thing would be to
2077 * calculate system wide asnc/sync ratio.
2079 tmp = cfq_target_latency * cfqg_busy_async_queues(cfqd, cfqg);
2080 tmp = tmp/cfqd->busy_queues;
2081 slice = min_t(unsigned, slice, tmp);
2083 /* async workload slice is scaled down according to
2084 * the sync/async slice ratio. */
2085 slice = slice * cfqd->cfq_slice[0] / cfqd->cfq_slice[1];
2087 /* sync workload slice is at least 2 * cfq_slice_idle */
2088 slice = max(slice, 2 * cfqd->cfq_slice_idle);
2090 slice = max_t(unsigned, slice, CFQ_MIN_TT);
2091 cfqd->workload_expires = jiffies + slice;
2092 cfqd->noidle_tree_requires_idle = false;
2095 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd)
2097 struct cfq_rb_root *st = &cfqd->grp_service_tree;
2098 struct cfq_group *cfqg;
2100 if (RB_EMPTY_ROOT(&st->rb))
2102 cfqg = cfq_rb_first_group(st);
2103 st->active = &cfqg->rb_node;
2104 update_min_vdisktime(st);
2108 static void cfq_choose_cfqg(struct cfq_data *cfqd)
2110 struct cfq_group *cfqg = cfq_get_next_cfqg(cfqd);
2112 cfqd->serving_group = cfqg;
2114 /* Restore the workload type data */
2115 if (cfqg->saved_workload_slice) {
2116 cfqd->workload_expires = jiffies + cfqg->saved_workload_slice;
2117 cfqd->serving_type = cfqg->saved_workload;
2118 cfqd->serving_prio = cfqg->saved_serving_prio;
2120 choose_service_tree(cfqd, cfqg);
2124 * Select a queue for service. If we have a current active queue,
2125 * check whether to continue servicing it, or retrieve and set a new one.
2127 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
2129 struct cfq_queue *cfqq, *new_cfqq = NULL;
2131 cfqq = cfqd->active_queue;
2135 if (!cfqd->rq_queued)
2138 * The active queue has run out of time, expire it and select new.
2140 if ((cfq_slice_used(cfqq) || cfq_cfqq_wait_busy_done(cfqq))
2141 && !cfq_cfqq_must_dispatch(cfqq))
2145 * The active queue has requests and isn't expired, allow it to
2148 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
2152 * If another queue has a request waiting within our mean seek
2153 * distance, let it run. The expire code will check for close
2154 * cooperators and put the close queue at the front of the service
2155 * tree. If possible, merge the expiring queue with the new cfqq.
2157 new_cfqq = cfq_close_cooperator(cfqd, cfqq);
2159 if (!cfqq->new_cfqq)
2160 cfq_setup_merge(cfqq, new_cfqq);
2165 * No requests pending. If the active queue still has requests in
2166 * flight or is idling for a new request, allow either of these
2167 * conditions to happen (or time out) before selecting a new queue.
2169 if (timer_pending(&cfqd->idle_slice_timer) ||
2170 (cfqq->dispatched && cfq_should_idle(cfqd, cfqq))) {
2176 cfq_slice_expired(cfqd, 0);
2179 * Current queue expired. Check if we have to switch to a new
2183 cfq_choose_cfqg(cfqd);
2185 cfqq = cfq_set_active_queue(cfqd, new_cfqq);
2190 static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
2194 while (cfqq->next_rq) {
2195 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
2199 BUG_ON(!list_empty(&cfqq->fifo));
2201 /* By default cfqq is not expired if it is empty. Do it explicitly */
2202 __cfq_slice_expired(cfqq->cfqd, cfqq, 0);
2207 * Drain our current requests. Used for barriers and when switching
2208 * io schedulers on-the-fly.
2210 static int cfq_forced_dispatch(struct cfq_data *cfqd)
2212 struct cfq_queue *cfqq;
2215 while ((cfqq = cfq_get_next_queue_forced(cfqd)) != NULL)
2216 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
2218 cfq_slice_expired(cfqd, 0);
2219 BUG_ON(cfqd->busy_queues);
2221 cfq_log(cfqd, "forced_dispatch=%d", dispatched);
2225 static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2227 unsigned int max_dispatch;
2230 * Drain async requests before we start sync IO
2232 if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_driver[BLK_RW_ASYNC])
2236 * If this is an async queue and we have sync IO in flight, let it wait
2238 if (cfqd->sync_flight && !cfq_cfqq_sync(cfqq))
2241 max_dispatch = cfqd->cfq_quantum;
2242 if (cfq_class_idle(cfqq))
2246 * Does this cfqq already have too much IO in flight?
2248 if (cfqq->dispatched >= max_dispatch) {
2250 * idle queue must always only have a single IO in flight
2252 if (cfq_class_idle(cfqq))
2256 * We have other queues, don't allow more IO from this one
2258 if (cfqd->busy_queues > 1)
2262 * Sole queue user, no limit
2268 * Async queues must wait a bit before being allowed dispatch.
2269 * We also ramp up the dispatch depth gradually for async IO,
2270 * based on the last sync IO we serviced
2272 if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
2273 unsigned long last_sync = jiffies - cfqd->last_delayed_sync;
2276 depth = last_sync / cfqd->cfq_slice[1];
2277 if (!depth && !cfqq->dispatched)
2279 if (depth < max_dispatch)
2280 max_dispatch = depth;
2284 * If we're below the current max, allow a dispatch
2286 return cfqq->dispatched < max_dispatch;
2290 * Dispatch a request from cfqq, moving them to the request queue
2293 static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2297 BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
2299 if (!cfq_may_dispatch(cfqd, cfqq))
2303 * follow expired path, else get first next available
2305 rq = cfq_check_fifo(cfqq);
2310 * insert request into driver dispatch list
2312 cfq_dispatch_insert(cfqd->queue, rq);
2314 if (!cfqd->active_cic) {
2315 struct cfq_io_context *cic = RQ_CIC(rq);
2317 atomic_long_inc(&cic->ioc->refcount);
2318 cfqd->active_cic = cic;
2325 * Find the cfqq that we need to service and move a request from that to the
2328 static int cfq_dispatch_requests(struct request_queue *q, int force)
2330 struct cfq_data *cfqd = q->elevator->elevator_data;
2331 struct cfq_queue *cfqq;
2333 if (!cfqd->busy_queues)
2336 if (unlikely(force))
2337 return cfq_forced_dispatch(cfqd);
2339 cfqq = cfq_select_queue(cfqd);
2344 * Dispatch a request from this cfqq, if it is allowed
2346 if (!cfq_dispatch_request(cfqd, cfqq))
2349 cfqq->slice_dispatch++;
2350 cfq_clear_cfqq_must_dispatch(cfqq);
2353 * expire an async queue immediately if it has used up its slice. idle
2354 * queue always expire after 1 dispatch round.
2356 if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
2357 cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
2358 cfq_class_idle(cfqq))) {
2359 cfqq->slice_end = jiffies + 1;
2360 cfq_slice_expired(cfqd, 0);
2363 cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
2368 * task holds one reference to the queue, dropped when task exits. each rq
2369 * in-flight on this queue also holds a reference, dropped when rq is freed.
2371 * Each cfq queue took a reference on the parent group. Drop it now.
2372 * queue lock must be held here.
2374 static void cfq_put_queue(struct cfq_queue *cfqq)
2376 struct cfq_data *cfqd = cfqq->cfqd;
2377 struct cfq_group *cfqg, *orig_cfqg;
2379 BUG_ON(atomic_read(&cfqq->ref) <= 0);
2381 if (!atomic_dec_and_test(&cfqq->ref))
2384 cfq_log_cfqq(cfqd, cfqq, "put_queue");
2385 BUG_ON(rb_first(&cfqq->sort_list));
2386 BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
2388 orig_cfqg = cfqq->orig_cfqg;
2390 if (unlikely(cfqd->active_queue == cfqq)) {
2391 __cfq_slice_expired(cfqd, cfqq, 0);
2392 cfq_schedule_dispatch(cfqd);
2395 BUG_ON(cfq_cfqq_on_rr(cfqq));
2396 kmem_cache_free(cfq_pool, cfqq);
2399 cfq_put_cfqg(orig_cfqg);
2403 * Must always be called with the rcu_read_lock() held
2406 __call_for_each_cic(struct io_context *ioc,
2407 void (*func)(struct io_context *, struct cfq_io_context *))
2409 struct cfq_io_context *cic;
2410 struct hlist_node *n;
2412 hlist_for_each_entry_rcu(cic, n, &ioc->cic_list, cic_list)
2417 * Call func for each cic attached to this ioc.
2420 call_for_each_cic(struct io_context *ioc,
2421 void (*func)(struct io_context *, struct cfq_io_context *))
2424 __call_for_each_cic(ioc, func);
2428 static void cfq_cic_free_rcu(struct rcu_head *head)
2430 struct cfq_io_context *cic;
2432 cic = container_of(head, struct cfq_io_context, rcu_head);
2434 kmem_cache_free(cfq_ioc_pool, cic);
2435 elv_ioc_count_dec(cfq_ioc_count);
2439 * CFQ scheduler is exiting, grab exit lock and check
2440 * the pending io context count. If it hits zero,
2441 * complete ioc_gone and set it back to NULL
2443 spin_lock(&ioc_gone_lock);
2444 if (ioc_gone && !elv_ioc_count_read(cfq_ioc_count)) {
2448 spin_unlock(&ioc_gone_lock);
2452 static void cfq_cic_free(struct cfq_io_context *cic)
2454 call_rcu(&cic->rcu_head, cfq_cic_free_rcu);
2457 static void cic_free_func(struct io_context *ioc, struct cfq_io_context *cic)
2459 unsigned long flags;
2461 BUG_ON(!cic->dead_key);
2463 spin_lock_irqsave(&ioc->lock, flags);
2464 radix_tree_delete(&ioc->radix_root, cic->dead_key);
2465 hlist_del_rcu(&cic->cic_list);
2466 spin_unlock_irqrestore(&ioc->lock, flags);
2472 * Must be called with rcu_read_lock() held or preemption otherwise disabled.
2473 * Only two callers of this - ->dtor() which is called with the rcu_read_lock(),
2474 * and ->trim() which is called with the task lock held
2476 static void cfq_free_io_context(struct io_context *ioc)
2479 * ioc->refcount is zero here, or we are called from elv_unregister(),
2480 * so no more cic's are allowed to be linked into this ioc. So it
2481 * should be ok to iterate over the known list, we will see all cic's
2482 * since no new ones are added.
2484 __call_for_each_cic(ioc, cic_free_func);
2487 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2489 struct cfq_queue *__cfqq, *next;
2491 if (unlikely(cfqq == cfqd->active_queue)) {
2492 __cfq_slice_expired(cfqd, cfqq, 0);
2493 cfq_schedule_dispatch(cfqd);
2497 * If this queue was scheduled to merge with another queue, be
2498 * sure to drop the reference taken on that queue (and others in
2499 * the merge chain). See cfq_setup_merge and cfq_merge_cfqqs.
2501 __cfqq = cfqq->new_cfqq;
2503 if (__cfqq == cfqq) {
2504 WARN(1, "cfqq->new_cfqq loop detected\n");
2507 next = __cfqq->new_cfqq;
2508 cfq_put_queue(__cfqq);
2512 cfq_put_queue(cfqq);
2515 static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
2516 struct cfq_io_context *cic)
2518 struct io_context *ioc = cic->ioc;
2520 list_del_init(&cic->queue_list);
2523 * Make sure key == NULL is seen for dead queues
2526 cic->dead_key = (unsigned long) cic->key;
2529 if (ioc->ioc_data == cic)
2530 rcu_assign_pointer(ioc->ioc_data, NULL);
2532 if (cic->cfqq[BLK_RW_ASYNC]) {
2533 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]);
2534 cic->cfqq[BLK_RW_ASYNC] = NULL;
2537 if (cic->cfqq[BLK_RW_SYNC]) {
2538 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]);
2539 cic->cfqq[BLK_RW_SYNC] = NULL;
2543 static void cfq_exit_single_io_context(struct io_context *ioc,
2544 struct cfq_io_context *cic)
2546 struct cfq_data *cfqd = cic->key;
2549 struct request_queue *q = cfqd->queue;
2550 unsigned long flags;
2552 spin_lock_irqsave(q->queue_lock, flags);
2555 * Ensure we get a fresh copy of the ->key to prevent
2556 * race between exiting task and queue
2558 smp_read_barrier_depends();
2560 __cfq_exit_single_io_context(cfqd, cic);
2562 spin_unlock_irqrestore(q->queue_lock, flags);
2567 * The process that ioc belongs to has exited, we need to clean up
2568 * and put the internal structures we have that belongs to that process.
2570 static void cfq_exit_io_context(struct io_context *ioc)
2572 call_for_each_cic(ioc, cfq_exit_single_io_context);
2575 static struct cfq_io_context *
2576 cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
2578 struct cfq_io_context *cic;
2580 cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO,
2583 cic->last_end_request = jiffies;
2584 INIT_LIST_HEAD(&cic->queue_list);
2585 INIT_HLIST_NODE(&cic->cic_list);
2586 cic->dtor = cfq_free_io_context;
2587 cic->exit = cfq_exit_io_context;
2588 elv_ioc_count_inc(cfq_ioc_count);
2594 static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc)
2596 struct task_struct *tsk = current;
2599 if (!cfq_cfqq_prio_changed(cfqq))
2602 ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio);
2603 switch (ioprio_class) {
2605 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
2606 case IOPRIO_CLASS_NONE:
2608 * no prio set, inherit CPU scheduling settings
2610 cfqq->ioprio = task_nice_ioprio(tsk);
2611 cfqq->ioprio_class = task_nice_ioclass(tsk);
2613 case IOPRIO_CLASS_RT:
2614 cfqq->ioprio = task_ioprio(ioc);
2615 cfqq->ioprio_class = IOPRIO_CLASS_RT;
2617 case IOPRIO_CLASS_BE:
2618 cfqq->ioprio = task_ioprio(ioc);
2619 cfqq->ioprio_class = IOPRIO_CLASS_BE;
2621 case IOPRIO_CLASS_IDLE:
2622 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
2624 cfq_clear_cfqq_idle_window(cfqq);
2629 * keep track of original prio settings in case we have to temporarily
2630 * elevate the priority of this queue
2632 cfqq->org_ioprio = cfqq->ioprio;
2633 cfqq->org_ioprio_class = cfqq->ioprio_class;
2634 cfq_clear_cfqq_prio_changed(cfqq);
2637 static void changed_ioprio(struct io_context *ioc, struct cfq_io_context *cic)
2639 struct cfq_data *cfqd = cic->key;
2640 struct cfq_queue *cfqq;
2641 unsigned long flags;
2643 if (unlikely(!cfqd))
2646 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2648 cfqq = cic->cfqq[BLK_RW_ASYNC];
2650 struct cfq_queue *new_cfqq;
2651 new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic->ioc,
2654 cic->cfqq[BLK_RW_ASYNC] = new_cfqq;
2655 cfq_put_queue(cfqq);
2659 cfqq = cic->cfqq[BLK_RW_SYNC];
2661 cfq_mark_cfqq_prio_changed(cfqq);
2663 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2666 static void cfq_ioc_set_ioprio(struct io_context *ioc)
2668 call_for_each_cic(ioc, changed_ioprio);
2669 ioc->ioprio_changed = 0;
2672 static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2673 pid_t pid, bool is_sync)
2675 RB_CLEAR_NODE(&cfqq->rb_node);
2676 RB_CLEAR_NODE(&cfqq->p_node);
2677 INIT_LIST_HEAD(&cfqq->fifo);
2679 atomic_set(&cfqq->ref, 0);
2682 cfq_mark_cfqq_prio_changed(cfqq);
2685 if (!cfq_class_idle(cfqq))
2686 cfq_mark_cfqq_idle_window(cfqq);
2687 cfq_mark_cfqq_sync(cfqq);
2692 #ifdef CONFIG_CFQ_GROUP_IOSCHED
2693 static void changed_cgroup(struct io_context *ioc, struct cfq_io_context *cic)
2695 struct cfq_queue *sync_cfqq = cic_to_cfqq(cic, 1);
2696 struct cfq_data *cfqd = cic->key;
2697 unsigned long flags;
2698 struct request_queue *q;
2700 if (unlikely(!cfqd))
2705 spin_lock_irqsave(q->queue_lock, flags);
2709 * Drop reference to sync queue. A new sync queue will be
2710 * assigned in new group upon arrival of a fresh request.
2712 cfq_log_cfqq(cfqd, sync_cfqq, "changed cgroup");
2713 cic_set_cfqq(cic, NULL, 1);
2714 cfq_put_queue(sync_cfqq);
2717 spin_unlock_irqrestore(q->queue_lock, flags);
2720 static void cfq_ioc_set_cgroup(struct io_context *ioc)
2722 call_for_each_cic(ioc, changed_cgroup);
2723 ioc->cgroup_changed = 0;
2725 #endif /* CONFIG_CFQ_GROUP_IOSCHED */
2727 static struct cfq_queue *
2728 cfq_find_alloc_queue(struct cfq_data *cfqd, bool is_sync,
2729 struct io_context *ioc, gfp_t gfp_mask)
2731 struct cfq_queue *cfqq, *new_cfqq = NULL;
2732 struct cfq_io_context *cic;
2733 struct cfq_group *cfqg;
2736 cfqg = cfq_get_cfqg(cfqd, 1);
2737 cic = cfq_cic_lookup(cfqd, ioc);
2738 /* cic always exists here */
2739 cfqq = cic_to_cfqq(cic, is_sync);
2742 * Always try a new alloc if we fell back to the OOM cfqq
2743 * originally, since it should just be a temporary situation.
2745 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
2750 } else if (gfp_mask & __GFP_WAIT) {
2751 spin_unlock_irq(cfqd->queue->queue_lock);
2752 new_cfqq = kmem_cache_alloc_node(cfq_pool,
2753 gfp_mask | __GFP_ZERO,
2755 spin_lock_irq(cfqd->queue->queue_lock);
2759 cfqq = kmem_cache_alloc_node(cfq_pool,
2760 gfp_mask | __GFP_ZERO,
2765 cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
2766 cfq_init_prio_data(cfqq, ioc);
2767 cfq_link_cfqq_cfqg(cfqq, cfqg);
2768 cfq_log_cfqq(cfqd, cfqq, "alloced");
2770 cfqq = &cfqd->oom_cfqq;
2774 kmem_cache_free(cfq_pool, new_cfqq);
2779 static struct cfq_queue **
2780 cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
2782 switch (ioprio_class) {
2783 case IOPRIO_CLASS_RT:
2784 return &cfqd->async_cfqq[0][ioprio];
2785 case IOPRIO_CLASS_BE:
2786 return &cfqd->async_cfqq[1][ioprio];
2787 case IOPRIO_CLASS_IDLE:
2788 return &cfqd->async_idle_cfqq;
2794 static struct cfq_queue *
2795 cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct io_context *ioc,
2798 const int ioprio = task_ioprio(ioc);
2799 const int ioprio_class = task_ioprio_class(ioc);
2800 struct cfq_queue **async_cfqq = NULL;
2801 struct cfq_queue *cfqq = NULL;
2804 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
2809 cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask);
2812 * pin the queue now that it's allocated, scheduler exit will prune it
2814 if (!is_sync && !(*async_cfqq)) {
2815 atomic_inc(&cfqq->ref);
2819 atomic_inc(&cfqq->ref);
2824 * We drop cfq io contexts lazily, so we may find a dead one.
2827 cfq_drop_dead_cic(struct cfq_data *cfqd, struct io_context *ioc,
2828 struct cfq_io_context *cic)
2830 unsigned long flags;
2832 WARN_ON(!list_empty(&cic->queue_list));
2834 spin_lock_irqsave(&ioc->lock, flags);
2836 BUG_ON(ioc->ioc_data == cic);
2838 radix_tree_delete(&ioc->radix_root, (unsigned long) cfqd);
2839 hlist_del_rcu(&cic->cic_list);
2840 spin_unlock_irqrestore(&ioc->lock, flags);
2845 static struct cfq_io_context *
2846 cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc)
2848 struct cfq_io_context *cic;
2849 unsigned long flags;
2858 * we maintain a last-hit cache, to avoid browsing over the tree
2860 cic = rcu_dereference(ioc->ioc_data);
2861 if (cic && cic->key == cfqd) {
2867 cic = radix_tree_lookup(&ioc->radix_root, (unsigned long) cfqd);
2871 /* ->key must be copied to avoid race with cfq_exit_queue() */
2874 cfq_drop_dead_cic(cfqd, ioc, cic);
2879 spin_lock_irqsave(&ioc->lock, flags);
2880 rcu_assign_pointer(ioc->ioc_data, cic);
2881 spin_unlock_irqrestore(&ioc->lock, flags);
2889 * Add cic into ioc, using cfqd as the search key. This enables us to lookup
2890 * the process specific cfq io context when entered from the block layer.
2891 * Also adds the cic to a per-cfqd list, used when this queue is removed.
2893 static int cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
2894 struct cfq_io_context *cic, gfp_t gfp_mask)
2896 unsigned long flags;
2899 ret = radix_tree_preload(gfp_mask);
2904 spin_lock_irqsave(&ioc->lock, flags);
2905 ret = radix_tree_insert(&ioc->radix_root,
2906 (unsigned long) cfqd, cic);
2908 hlist_add_head_rcu(&cic->cic_list, &ioc->cic_list);
2909 spin_unlock_irqrestore(&ioc->lock, flags);
2911 radix_tree_preload_end();
2914 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2915 list_add(&cic->queue_list, &cfqd->cic_list);
2916 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2921 printk(KERN_ERR "cfq: cic link failed!\n");
2927 * Setup general io context and cfq io context. There can be several cfq
2928 * io contexts per general io context, if this process is doing io to more
2929 * than one device managed by cfq.
2931 static struct cfq_io_context *
2932 cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
2934 struct io_context *ioc = NULL;
2935 struct cfq_io_context *cic;
2937 might_sleep_if(gfp_mask & __GFP_WAIT);
2939 ioc = get_io_context(gfp_mask, cfqd->queue->node);
2943 cic = cfq_cic_lookup(cfqd, ioc);
2947 cic = cfq_alloc_io_context(cfqd, gfp_mask);
2951 if (cfq_cic_link(cfqd, ioc, cic, gfp_mask))
2955 smp_read_barrier_depends();
2956 if (unlikely(ioc->ioprio_changed))
2957 cfq_ioc_set_ioprio(ioc);
2959 #ifdef CONFIG_CFQ_GROUP_IOSCHED
2960 if (unlikely(ioc->cgroup_changed))
2961 cfq_ioc_set_cgroup(ioc);
2967 put_io_context(ioc);
2972 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
2974 unsigned long elapsed = jiffies - cic->last_end_request;
2975 unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
2977 cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
2978 cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
2979 cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
2983 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2989 if (!cfqq->last_request_pos)
2991 else if (cfqq->last_request_pos < blk_rq_pos(rq))
2992 sdist = blk_rq_pos(rq) - cfqq->last_request_pos;
2994 sdist = cfqq->last_request_pos - blk_rq_pos(rq);
2997 * Don't allow the seek distance to get too large from the
2998 * odd fragment, pagein, etc
3000 if (cfqq->seek_samples <= 60) /* second&third seek */
3001 sdist = min(sdist, (cfqq->seek_mean * 4) + 2*1024*1024);
3003 sdist = min(sdist, (cfqq->seek_mean * 4) + 2*1024*64);
3005 cfqq->seek_samples = (7*cfqq->seek_samples + 256) / 8;
3006 cfqq->seek_total = (7*cfqq->seek_total + (u64)256*sdist) / 8;
3007 total = cfqq->seek_total + (cfqq->seek_samples/2);
3008 do_div(total, cfqq->seek_samples);
3009 cfqq->seek_mean = (sector_t)total;
3012 * If this cfqq is shared between multiple processes, check to
3013 * make sure that those processes are still issuing I/Os within
3014 * the mean seek distance. If not, it may be time to break the
3015 * queues apart again.
3017 if (cfq_cfqq_coop(cfqq)) {
3018 if (CFQQ_SEEKY(cfqq) && !cfqq->seeky_start)
3019 cfqq->seeky_start = jiffies;
3020 else if (!CFQQ_SEEKY(cfqq))
3021 cfqq->seeky_start = 0;
3026 * Disable idle window if the process thinks too long or seeks so much that
3030 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3031 struct cfq_io_context *cic)
3033 int old_idle, enable_idle;
3036 * Don't idle for async or idle io prio class
3038 if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
3041 enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
3043 if (cfqq->queued[0] + cfqq->queued[1] >= 4)
3044 cfq_mark_cfqq_deep(cfqq);
3046 if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle ||
3047 (!cfq_cfqq_deep(cfqq) && sample_valid(cfqq->seek_samples)
3048 && CFQQ_SEEKY(cfqq)))
3050 else if (sample_valid(cic->ttime_samples)) {
3051 if (cic->ttime_mean > cfqd->cfq_slice_idle)
3057 if (old_idle != enable_idle) {
3058 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
3060 cfq_mark_cfqq_idle_window(cfqq);
3062 cfq_clear_cfqq_idle_window(cfqq);
3067 * Check if new_cfqq should preempt the currently active queue. Return 0 for
3068 * no or if we aren't sure, a 1 will cause a preempt.
3071 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
3074 struct cfq_queue *cfqq;
3076 cfqq = cfqd->active_queue;
3080 if (cfq_class_idle(new_cfqq))
3083 if (cfq_class_idle(cfqq))
3087 * if the new request is sync, but the currently running queue is
3088 * not, let the sync request have priority.
3090 if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
3093 if (new_cfqq->cfqg != cfqq->cfqg)
3096 if (cfq_slice_used(cfqq))
3099 /* Allow preemption only if we are idling on sync-noidle tree */
3100 if (cfqd->serving_type == SYNC_NOIDLE_WORKLOAD &&
3101 cfqq_type(new_cfqq) == SYNC_NOIDLE_WORKLOAD &&
3102 new_cfqq->service_tree->count == 2 &&
3103 RB_EMPTY_ROOT(&cfqq->sort_list))
3107 * So both queues are sync. Let the new request get disk time if
3108 * it's a metadata request and the current queue is doing regular IO.
3110 if (rq_is_meta(rq) && !cfqq->meta_pending)
3114 * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
3116 if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
3119 if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
3123 * if this request is as-good as one we would expect from the
3124 * current cfqq, let it preempt
3126 if (cfq_rq_close(cfqd, cfqq, rq))
3133 * cfqq preempts the active queue. if we allowed preempt with no slice left,
3134 * let it have half of its nominal slice.
3136 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3138 cfq_log_cfqq(cfqd, cfqq, "preempt");
3139 cfq_slice_expired(cfqd, 1);
3142 * Put the new queue at the front of the of the current list,
3143 * so we know that it will be selected next.
3145 BUG_ON(!cfq_cfqq_on_rr(cfqq));
3147 cfq_service_tree_add(cfqd, cfqq, 1);
3149 cfqq->slice_end = 0;
3150 cfq_mark_cfqq_slice_new(cfqq);
3154 * Called when a new fs request (rq) is added (to cfqq). Check if there's
3155 * something we should do about it
3158 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3161 struct cfq_io_context *cic = RQ_CIC(rq);
3165 cfqq->meta_pending++;
3167 cfq_update_io_thinktime(cfqd, cic);
3168 cfq_update_io_seektime(cfqd, cfqq, rq);
3169 cfq_update_idle_window(cfqd, cfqq, cic);
3171 cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
3173 if (cfqq == cfqd->active_queue) {
3174 if (cfq_cfqq_wait_busy(cfqq)) {
3175 cfq_clear_cfqq_wait_busy(cfqq);
3176 cfq_mark_cfqq_wait_busy_done(cfqq);
3179 * Remember that we saw a request from this process, but
3180 * don't start queuing just yet. Otherwise we risk seeing lots
3181 * of tiny requests, because we disrupt the normal plugging
3182 * and merging. If the request is already larger than a single
3183 * page, let it rip immediately. For that case we assume that
3184 * merging is already done. Ditto for a busy system that
3185 * has other work pending, don't risk delaying until the
3186 * idle timer unplug to continue working.
3188 if (cfq_cfqq_wait_request(cfqq)) {
3189 if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
3190 cfqd->busy_queues > 1) {
3191 del_timer(&cfqd->idle_slice_timer);
3192 __blk_run_queue(cfqd->queue);
3194 cfq_mark_cfqq_must_dispatch(cfqq);
3196 } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
3198 * not the active queue - expire current slice if it is
3199 * idle and has expired it's mean thinktime or this new queue
3200 * has some old slice time left and is of higher priority or
3201 * this new queue is RT and the current one is BE
3203 cfq_preempt_queue(cfqd, cfqq);
3204 __blk_run_queue(cfqd->queue);
3208 static void cfq_insert_request(struct request_queue *q, struct request *rq)
3210 struct cfq_data *cfqd = q->elevator->elevator_data;
3211 struct cfq_queue *cfqq = RQ_CFQQ(rq);
3213 cfq_log_cfqq(cfqd, cfqq, "insert_request");
3214 cfq_init_prio_data(cfqq, RQ_CIC(rq)->ioc);
3216 rq_set_fifo_time(rq, jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)]);
3217 list_add_tail(&rq->queuelist, &cfqq->fifo);
3220 cfq_rq_enqueued(cfqd, cfqq, rq);
3224 * Update hw_tag based on peak queue depth over 50 samples under
3227 static void cfq_update_hw_tag(struct cfq_data *cfqd)
3229 struct cfq_queue *cfqq = cfqd->active_queue;
3231 if (rq_in_driver(cfqd) > cfqd->hw_tag_est_depth)
3232 cfqd->hw_tag_est_depth = rq_in_driver(cfqd);
3234 if (cfqd->hw_tag == 1)
3237 if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
3238 rq_in_driver(cfqd) <= CFQ_HW_QUEUE_MIN)
3242 * If active queue hasn't enough requests and can idle, cfq might not
3243 * dispatch sufficient requests to hardware. Don't zero hw_tag in this
3246 if (cfqq && cfq_cfqq_idle_window(cfqq) &&
3247 cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] <
3248 CFQ_HW_QUEUE_MIN && rq_in_driver(cfqd) < CFQ_HW_QUEUE_MIN)
3251 if (cfqd->hw_tag_samples++ < 50)
3254 if (cfqd->hw_tag_est_depth >= CFQ_HW_QUEUE_MIN)
3260 static void cfq_completed_request(struct request_queue *q, struct request *rq)
3262 struct cfq_queue *cfqq = RQ_CFQQ(rq);
3263 struct cfq_data *cfqd = cfqq->cfqd;
3264 const int sync = rq_is_sync(rq);
3268 cfq_log_cfqq(cfqd, cfqq, "complete rqnoidle %d", !!rq_noidle(rq));
3270 cfq_update_hw_tag(cfqd);
3272 WARN_ON(!cfqd->rq_in_driver[sync]);
3273 WARN_ON(!cfqq->dispatched);
3274 cfqd->rq_in_driver[sync]--;
3277 if (cfq_cfqq_sync(cfqq))
3278 cfqd->sync_flight--;
3281 RQ_CIC(rq)->last_end_request = now;
3282 if (!time_after(rq->start_time + cfqd->cfq_fifo_expire[1], now))
3283 cfqd->last_delayed_sync = now;
3287 * If this is the active queue, check if it needs to be expired,
3288 * or if we want to idle in case it has no pending requests.
3290 if (cfqd->active_queue == cfqq) {
3291 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
3293 if (cfq_cfqq_slice_new(cfqq)) {
3294 cfq_set_prio_slice(cfqd, cfqq);
3295 cfq_clear_cfqq_slice_new(cfqq);
3299 * If this queue consumed its slice and this is last queue
3300 * in the group, wait for next request before we expire
3303 if (cfq_slice_used(cfqq) && cfqq->cfqg->nr_cfqq == 1) {
3304 cfqq->slice_end = jiffies + cfqd->cfq_slice_idle;
3305 cfq_mark_cfqq_wait_busy(cfqq);
3309 * Idling is not enabled on:
3311 * - idle-priority queues
3313 * - queues with still some requests queued
3314 * - when there is a close cooperator
3316 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
3317 cfq_slice_expired(cfqd, 1);
3318 else if (sync && cfqq_empty &&
3319 !cfq_close_cooperator(cfqd, cfqq)) {
3320 cfqd->noidle_tree_requires_idle |= !rq_noidle(rq);
3322 * Idling is enabled for SYNC_WORKLOAD.
3323 * SYNC_NOIDLE_WORKLOAD idles at the end of the tree
3324 * only if we processed at least one !rq_noidle request
3326 if (cfqd->serving_type == SYNC_WORKLOAD
3327 || cfqd->noidle_tree_requires_idle
3328 || cfqq->cfqg->nr_cfqq == 1)
3329 cfq_arm_slice_timer(cfqd);
3333 if (!rq_in_driver(cfqd))
3334 cfq_schedule_dispatch(cfqd);
3338 * we temporarily boost lower priority queues if they are holding fs exclusive
3339 * resources. they are boosted to normal prio (CLASS_BE/4)
3341 static void cfq_prio_boost(struct cfq_queue *cfqq)
3343 if (has_fs_excl()) {
3345 * boost idle prio on transactions that would lock out other
3346 * users of the filesystem
3348 if (cfq_class_idle(cfqq))
3349 cfqq->ioprio_class = IOPRIO_CLASS_BE;
3350 if (cfqq->ioprio > IOPRIO_NORM)
3351 cfqq->ioprio = IOPRIO_NORM;
3354 * unboost the queue (if needed)
3356 cfqq->ioprio_class = cfqq->org_ioprio_class;
3357 cfqq->ioprio = cfqq->org_ioprio;
3361 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
3363 if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
3364 cfq_mark_cfqq_must_alloc_slice(cfqq);
3365 return ELV_MQUEUE_MUST;
3368 return ELV_MQUEUE_MAY;
3371 static int cfq_may_queue(struct request_queue *q, int rw)
3373 struct cfq_data *cfqd = q->elevator->elevator_data;
3374 struct task_struct *tsk = current;
3375 struct cfq_io_context *cic;
3376 struct cfq_queue *cfqq;
3379 * don't force setup of a queue from here, as a call to may_queue
3380 * does not necessarily imply that a request actually will be queued.
3381 * so just lookup a possibly existing queue, or return 'may queue'
3384 cic = cfq_cic_lookup(cfqd, tsk->io_context);
3386 return ELV_MQUEUE_MAY;
3388 cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
3390 cfq_init_prio_data(cfqq, cic->ioc);
3391 cfq_prio_boost(cfqq);
3393 return __cfq_may_queue(cfqq);
3396 return ELV_MQUEUE_MAY;
3400 * queue lock held here
3402 static void cfq_put_request(struct request *rq)
3404 struct cfq_queue *cfqq = RQ_CFQQ(rq);
3407 const int rw = rq_data_dir(rq);
3409 BUG_ON(!cfqq->allocated[rw]);
3410 cfqq->allocated[rw]--;
3412 put_io_context(RQ_CIC(rq)->ioc);
3414 rq->elevator_private = NULL;
3415 rq->elevator_private2 = NULL;
3417 cfq_put_queue(cfqq);
3421 static struct cfq_queue *
3422 cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_context *cic,
3423 struct cfq_queue *cfqq)
3425 cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq);
3426 cic_set_cfqq(cic, cfqq->new_cfqq, 1);
3427 cfq_mark_cfqq_coop(cfqq->new_cfqq);
3428 cfq_put_queue(cfqq);
3429 return cic_to_cfqq(cic, 1);
3432 static int should_split_cfqq(struct cfq_queue *cfqq)
3434 if (cfqq->seeky_start &&
3435 time_after(jiffies, cfqq->seeky_start + CFQQ_COOP_TOUT))
3441 * Returns NULL if a new cfqq should be allocated, or the old cfqq if this
3442 * was the last process referring to said cfqq.
3444 static struct cfq_queue *
3445 split_cfqq(struct cfq_io_context *cic, struct cfq_queue *cfqq)
3447 if (cfqq_process_refs(cfqq) == 1) {
3448 cfqq->seeky_start = 0;
3449 cfqq->pid = current->pid;
3450 cfq_clear_cfqq_coop(cfqq);
3454 cic_set_cfqq(cic, NULL, 1);
3455 cfq_put_queue(cfqq);
3459 * Allocate cfq data structures associated with this request.
3462 cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
3464 struct cfq_data *cfqd = q->elevator->elevator_data;
3465 struct cfq_io_context *cic;
3466 const int rw = rq_data_dir(rq);
3467 const bool is_sync = rq_is_sync(rq);
3468 struct cfq_queue *cfqq;
3469 unsigned long flags;
3471 might_sleep_if(gfp_mask & __GFP_WAIT);
3473 cic = cfq_get_io_context(cfqd, gfp_mask);
3475 spin_lock_irqsave(q->queue_lock, flags);
3481 cfqq = cic_to_cfqq(cic, is_sync);
3482 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
3483 cfqq = cfq_get_queue(cfqd, is_sync, cic->ioc, gfp_mask);
3484 cic_set_cfqq(cic, cfqq, is_sync);
3487 * If the queue was seeky for too long, break it apart.
3489 if (cfq_cfqq_coop(cfqq) && should_split_cfqq(cfqq)) {
3490 cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq");
3491 cfqq = split_cfqq(cic, cfqq);
3497 * Check to see if this queue is scheduled to merge with
3498 * another, closely cooperating queue. The merging of
3499 * queues happens here as it must be done in process context.
3500 * The reference on new_cfqq was taken in merge_cfqqs.
3503 cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq);
3506 cfqq->allocated[rw]++;
3507 atomic_inc(&cfqq->ref);
3509 spin_unlock_irqrestore(q->queue_lock, flags);
3511 rq->elevator_private = cic;
3512 rq->elevator_private2 = cfqq;
3517 put_io_context(cic->ioc);
3519 cfq_schedule_dispatch(cfqd);
3520 spin_unlock_irqrestore(q->queue_lock, flags);
3521 cfq_log(cfqd, "set_request fail");
3525 static void cfq_kick_queue(struct work_struct *work)
3527 struct cfq_data *cfqd =
3528 container_of(work, struct cfq_data, unplug_work);
3529 struct request_queue *q = cfqd->queue;
3531 spin_lock_irq(q->queue_lock);
3532 __blk_run_queue(cfqd->queue);
3533 spin_unlock_irq(q->queue_lock);
3537 * Timer running if the active_queue is currently idling inside its time slice
3539 static void cfq_idle_slice_timer(unsigned long data)
3541 struct cfq_data *cfqd = (struct cfq_data *) data;
3542 struct cfq_queue *cfqq;
3543 unsigned long flags;
3546 cfq_log(cfqd, "idle timer fired");
3548 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
3550 cfqq = cfqd->active_queue;
3555 * We saw a request before the queue expired, let it through
3557 if (cfq_cfqq_must_dispatch(cfqq))
3563 if (cfq_slice_used(cfqq))
3567 * only expire and reinvoke request handler, if there are
3568 * other queues with pending requests
3570 if (!cfqd->busy_queues)
3574 * not expired and it has a request pending, let it dispatch
3576 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3580 * Queue depth flag is reset only when the idle didn't succeed
3582 cfq_clear_cfqq_deep(cfqq);
3585 cfq_slice_expired(cfqd, timed_out);
3587 cfq_schedule_dispatch(cfqd);
3589 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
3592 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
3594 del_timer_sync(&cfqd->idle_slice_timer);
3595 cancel_work_sync(&cfqd->unplug_work);
3598 static void cfq_put_async_queues(struct cfq_data *cfqd)
3602 for (i = 0; i < IOPRIO_BE_NR; i++) {
3603 if (cfqd->async_cfqq[0][i])
3604 cfq_put_queue(cfqd->async_cfqq[0][i]);
3605 if (cfqd->async_cfqq[1][i])
3606 cfq_put_queue(cfqd->async_cfqq[1][i]);
3609 if (cfqd->async_idle_cfqq)
3610 cfq_put_queue(cfqd->async_idle_cfqq);
3613 static void cfq_cfqd_free(struct rcu_head *head)
3615 kfree(container_of(head, struct cfq_data, rcu));
3618 static void cfq_exit_queue(struct elevator_queue *e)
3620 struct cfq_data *cfqd = e->elevator_data;
3621 struct request_queue *q = cfqd->queue;
3623 cfq_shutdown_timer_wq(cfqd);
3625 spin_lock_irq(q->queue_lock);
3627 if (cfqd->active_queue)
3628 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
3630 while (!list_empty(&cfqd->cic_list)) {
3631 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
3632 struct cfq_io_context,
3635 __cfq_exit_single_io_context(cfqd, cic);
3638 cfq_put_async_queues(cfqd);
3639 cfq_release_cfq_groups(cfqd);
3640 blkiocg_del_blkio_group(&cfqd->root_group.blkg);
3642 spin_unlock_irq(q->queue_lock);
3644 cfq_shutdown_timer_wq(cfqd);
3646 /* Wait for cfqg->blkg->key accessors to exit their grace periods. */
3647 call_rcu(&cfqd->rcu, cfq_cfqd_free);
3650 static void *cfq_init_queue(struct request_queue *q)
3652 struct cfq_data *cfqd;
3654 struct cfq_group *cfqg;
3655 struct cfq_rb_root *st;
3657 cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
3661 /* Init root service tree */
3662 cfqd->grp_service_tree = CFQ_RB_ROOT;
3664 /* Init root group */
3665 cfqg = &cfqd->root_group;
3666 for_each_cfqg_st(cfqg, i, j, st)
3668 RB_CLEAR_NODE(&cfqg->rb_node);
3670 /* Give preference to root group over other groups */
3671 cfqg->weight = 2*BLKIO_WEIGHT_DEFAULT;
3673 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3675 * Take a reference to root group which we never drop. This is just
3676 * to make sure that cfq_put_cfqg() does not try to kfree root group
3678 atomic_set(&cfqg->ref, 1);
3679 blkiocg_add_blkio_group(&blkio_root_cgroup, &cfqg->blkg, (void *)cfqd,
3683 * Not strictly needed (since RB_ROOT just clears the node and we
3684 * zeroed cfqd on alloc), but better be safe in case someone decides
3685 * to add magic to the rb code
3687 for (i = 0; i < CFQ_PRIO_LISTS; i++)
3688 cfqd->prio_trees[i] = RB_ROOT;
3691 * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
3692 * Grab a permanent reference to it, so that the normal code flow
3693 * will not attempt to free it.
3695 cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
3696 atomic_inc(&cfqd->oom_cfqq.ref);
3697 cfq_link_cfqq_cfqg(&cfqd->oom_cfqq, &cfqd->root_group);
3699 INIT_LIST_HEAD(&cfqd->cic_list);
3703 init_timer(&cfqd->idle_slice_timer);
3704 cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
3705 cfqd->idle_slice_timer.data = (unsigned long) cfqd;
3707 INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
3709 cfqd->cfq_quantum = cfq_quantum;
3710 cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
3711 cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
3712 cfqd->cfq_back_max = cfq_back_max;
3713 cfqd->cfq_back_penalty = cfq_back_penalty;
3714 cfqd->cfq_slice[0] = cfq_slice_async;
3715 cfqd->cfq_slice[1] = cfq_slice_sync;
3716 cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
3717 cfqd->cfq_slice_idle = cfq_slice_idle;
3718 cfqd->cfq_latency = 1;
3719 cfqd->cfq_group_isolation = 0;
3721 cfqd->last_delayed_sync = jiffies - HZ;
3722 INIT_RCU_HEAD(&cfqd->rcu);
3726 static void cfq_slab_kill(void)
3729 * Caller already ensured that pending RCU callbacks are completed,
3730 * so we should have no busy allocations at this point.
3733 kmem_cache_destroy(cfq_pool);
3735 kmem_cache_destroy(cfq_ioc_pool);
3738 static int __init cfq_slab_setup(void)
3740 cfq_pool = KMEM_CACHE(cfq_queue, 0);
3744 cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0);
3755 * sysfs parts below -->
3758 cfq_var_show(unsigned int var, char *page)
3760 return sprintf(page, "%d\n", var);
3764 cfq_var_store(unsigned int *var, const char *page, size_t count)
3766 char *p = (char *) page;
3768 *var = simple_strtoul(p, &p, 10);
3772 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
3773 static ssize_t __FUNC(struct elevator_queue *e, char *page) \
3775 struct cfq_data *cfqd = e->elevator_data; \
3776 unsigned int __data = __VAR; \
3778 __data = jiffies_to_msecs(__data); \
3779 return cfq_var_show(__data, (page)); \
3781 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
3782 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
3783 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
3784 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
3785 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
3786 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
3787 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
3788 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
3789 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
3790 SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
3791 SHOW_FUNCTION(cfq_group_isolation_show, cfqd->cfq_group_isolation, 0);
3792 #undef SHOW_FUNCTION
3794 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
3795 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
3797 struct cfq_data *cfqd = e->elevator_data; \
3798 unsigned int __data; \
3799 int ret = cfq_var_store(&__data, (page), count); \
3800 if (__data < (MIN)) \
3802 else if (__data > (MAX)) \
3805 *(__PTR) = msecs_to_jiffies(__data); \
3807 *(__PTR) = __data; \
3810 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
3811 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
3813 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
3815 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
3816 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
3818 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
3819 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
3820 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
3821 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
3823 STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
3824 STORE_FUNCTION(cfq_group_isolation_store, &cfqd->cfq_group_isolation, 0, 1, 0);
3825 #undef STORE_FUNCTION
3827 #define CFQ_ATTR(name) \
3828 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
3830 static struct elv_fs_entry cfq_attrs[] = {
3832 CFQ_ATTR(fifo_expire_sync),
3833 CFQ_ATTR(fifo_expire_async),
3834 CFQ_ATTR(back_seek_max),
3835 CFQ_ATTR(back_seek_penalty),
3836 CFQ_ATTR(slice_sync),
3837 CFQ_ATTR(slice_async),
3838 CFQ_ATTR(slice_async_rq),
3839 CFQ_ATTR(slice_idle),
3840 CFQ_ATTR(low_latency),
3841 CFQ_ATTR(group_isolation),
3845 static struct elevator_type iosched_cfq = {
3847 .elevator_merge_fn = cfq_merge,
3848 .elevator_merged_fn = cfq_merged_request,
3849 .elevator_merge_req_fn = cfq_merged_requests,
3850 .elevator_allow_merge_fn = cfq_allow_merge,
3851 .elevator_dispatch_fn = cfq_dispatch_requests,
3852 .elevator_add_req_fn = cfq_insert_request,
3853 .elevator_activate_req_fn = cfq_activate_request,
3854 .elevator_deactivate_req_fn = cfq_deactivate_request,
3855 .elevator_queue_empty_fn = cfq_queue_empty,
3856 .elevator_completed_req_fn = cfq_completed_request,
3857 .elevator_former_req_fn = elv_rb_former_request,
3858 .elevator_latter_req_fn = elv_rb_latter_request,
3859 .elevator_set_req_fn = cfq_set_request,
3860 .elevator_put_req_fn = cfq_put_request,
3861 .elevator_may_queue_fn = cfq_may_queue,
3862 .elevator_init_fn = cfq_init_queue,
3863 .elevator_exit_fn = cfq_exit_queue,
3864 .trim = cfq_free_io_context,
3866 .elevator_attrs = cfq_attrs,
3867 .elevator_name = "cfq",
3868 .elevator_owner = THIS_MODULE,
3871 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3872 static struct blkio_policy_type blkio_policy_cfq = {
3874 .blkio_unlink_group_fn = cfq_unlink_blkio_group,
3875 .blkio_update_group_weight_fn = cfq_update_blkio_group_weight,
3879 static struct blkio_policy_type blkio_policy_cfq;
3882 static int __init cfq_init(void)
3885 * could be 0 on HZ < 1000 setups
3887 if (!cfq_slice_async)
3888 cfq_slice_async = 1;
3889 if (!cfq_slice_idle)
3892 if (cfq_slab_setup())
3895 elv_register(&iosched_cfq);
3896 blkio_policy_register(&blkio_policy_cfq);
3901 static void __exit cfq_exit(void)
3903 DECLARE_COMPLETION_ONSTACK(all_gone);
3904 blkio_policy_unregister(&blkio_policy_cfq);
3905 elv_unregister(&iosched_cfq);
3906 ioc_gone = &all_gone;
3907 /* ioc_gone's update must be visible before reading ioc_count */
3911 * this also protects us from entering cfq_slab_kill() with
3912 * pending RCU callbacks
3914 if (elv_ioc_count_read(cfq_ioc_count))
3915 wait_for_completion(&all_gone);
3919 module_init(cfq_init);
3920 module_exit(cfq_exit);
3922 MODULE_AUTHOR("Jens Axboe");
3923 MODULE_LICENSE("GPL");
3924 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");