2 * CFQ, or complete fairness queueing, disk scheduler.
4 * Based on ideas from a previously unfinished io
5 * scheduler (round robin per-process disk scheduling) and Andrea Arcangeli.
7 * Copyright (C) 2003 Jens Axboe <axboe@kernel.dk>
9 #include <linux/module.h>
10 #include <linux/slab.h>
11 #include <linux/blkdev.h>
12 #include <linux/elevator.h>
13 #include <linux/jiffies.h>
14 #include <linux/rbtree.h>
15 #include <linux/ioprio.h>
16 #include <linux/blktrace_api.h>
17 #include "blk-cgroup.h"
22 /* max queue in one round of service */
23 static const int cfq_quantum = 8;
24 static const int cfq_fifo_expire[2] = { HZ / 4, HZ / 8 };
25 /* maximum backwards seek, in KiB */
26 static const int cfq_back_max = 16 * 1024;
27 /* penalty of a backwards seek */
28 static const int cfq_back_penalty = 2;
29 static const int cfq_slice_sync = HZ / 10;
30 static int cfq_slice_async = HZ / 25;
31 static const int cfq_slice_async_rq = 2;
32 static int cfq_slice_idle = HZ / 125;
33 static const int cfq_target_latency = HZ * 3/10; /* 300 ms */
34 static const int cfq_hist_divisor = 4;
37 * offset from end of service tree
39 #define CFQ_IDLE_DELAY (HZ / 5)
42 * below this threshold, we consider thinktime immediate
44 #define CFQ_MIN_TT (2)
46 #define CFQ_SLICE_SCALE (5)
47 #define CFQ_HW_QUEUE_MIN (5)
48 #define CFQ_SERVICE_SHIFT 12
50 #define CFQQ_SEEK_THR (sector_t)(8 * 100)
51 #define CFQQ_CLOSE_THR (sector_t)(8 * 1024)
52 #define CFQQ_SECT_THR_NONROT (sector_t)(2 * 32)
53 #define CFQQ_SEEKY(cfqq) (hweight32(cfqq->seek_history) > 32/8)
56 ((struct cfq_io_context *) (rq)->elevator_private)
57 #define RQ_CFQQ(rq) (struct cfq_queue *) ((rq)->elevator_private2)
58 #define RQ_CFQG(rq) (struct cfq_group *) ((rq)->elevator_private3)
60 static struct kmem_cache *cfq_pool;
61 static struct kmem_cache *cfq_ioc_pool;
63 static DEFINE_PER_CPU(unsigned long, cfq_ioc_count);
64 static struct completion *ioc_gone;
65 static DEFINE_SPINLOCK(ioc_gone_lock);
67 #define CFQ_PRIO_LISTS IOPRIO_BE_NR
68 #define cfq_class_idle(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
69 #define cfq_class_rt(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
71 #define sample_valid(samples) ((samples) > 80)
72 #define rb_entry_cfqg(node) rb_entry((node), struct cfq_group, rb_node)
75 * Most of our rbtree usage is for sorting with min extraction, so
76 * if we cache the leftmost node we don't have to walk down the tree
77 * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should
78 * move this into the elevator for the rq sorting as well.
84 unsigned total_weight;
86 struct rb_node *active;
88 #define CFQ_RB_ROOT (struct cfq_rb_root) { .rb = RB_ROOT, .left = NULL, \
89 .count = 0, .min_vdisktime = 0, }
92 * Per process-grouping structure
97 /* various state flags, see below */
100 struct cfq_data *cfqd;
101 /* service_tree member */
102 struct rb_node rb_node;
103 /* service_tree key */
104 unsigned long rb_key;
105 /* prio tree member */
106 struct rb_node p_node;
107 /* prio tree root we belong to, if any */
108 struct rb_root *p_root;
109 /* sorted list of pending requests */
110 struct rb_root sort_list;
111 /* if fifo isn't expired, next request to serve */
112 struct request *next_rq;
113 /* requests queued in sort_list */
115 /* currently allocated requests */
117 /* fifo list of requests in sort_list */
118 struct list_head fifo;
120 /* time when queue got scheduled in to dispatch first request. */
121 unsigned long dispatch_start;
122 unsigned int allocated_slice;
123 unsigned int slice_dispatch;
124 /* time when first request from queue completed and slice started. */
125 unsigned long slice_start;
126 unsigned long slice_end;
129 /* pending metadata requests */
131 /* number of requests that are on the dispatch list or inside driver */
134 /* io prio of this group */
135 unsigned short ioprio, org_ioprio;
136 unsigned short ioprio_class, org_ioprio_class;
141 sector_t last_request_pos;
143 struct cfq_rb_root *service_tree;
144 struct cfq_queue *new_cfqq;
145 struct cfq_group *cfqg;
146 struct cfq_group *orig_cfqg;
150 * First index in the service_trees.
151 * IDLE is handled separately, so it has negative index
160 * Second index in the service_trees.
164 SYNC_NOIDLE_WORKLOAD = 1,
168 /* This is per cgroup per device grouping structure */
170 /* group service_tree member */
171 struct rb_node rb_node;
173 /* group service_tree key */
178 /* number of cfqq currently on this group */
181 /* Per group busy queus average. Useful for workload slice calc. */
182 unsigned int busy_queues_avg[2];
184 * rr lists of queues with requests, onle rr for each priority class.
185 * Counts are embedded in the cfq_rb_root
187 struct cfq_rb_root service_trees[2][3];
188 struct cfq_rb_root service_tree_idle;
190 unsigned long saved_workload_slice;
191 enum wl_type_t saved_workload;
192 enum wl_prio_t saved_serving_prio;
193 struct blkio_group blkg;
194 #ifdef CONFIG_CFQ_GROUP_IOSCHED
195 struct hlist_node cfqd_node;
201 * Per block device queue structure
204 struct request_queue *queue;
205 /* Root service tree for cfq_groups */
206 struct cfq_rb_root grp_service_tree;
207 struct cfq_group root_group;
210 * The priority currently being served
212 enum wl_prio_t serving_prio;
213 enum wl_type_t serving_type;
214 unsigned long workload_expires;
215 struct cfq_group *serving_group;
216 bool noidle_tree_requires_idle;
219 * Each priority tree is sorted by next_request position. These
220 * trees are used when determining if two or more queues are
221 * interleaving requests (see cfq_close_cooperator).
223 struct rb_root prio_trees[CFQ_PRIO_LISTS];
225 unsigned int busy_queues;
231 * queue-depth detection
237 * -1 => indeterminate, (cfq will behave as if NCQ is present, to allow better detection)
238 * 1 => NCQ is present (hw_tag_est_depth is the estimated max depth)
241 int hw_tag_est_depth;
242 unsigned int hw_tag_samples;
245 * idle window management
247 struct timer_list idle_slice_timer;
248 struct work_struct unplug_work;
250 struct cfq_queue *active_queue;
251 struct cfq_io_context *active_cic;
254 * async queue for each priority case
256 struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
257 struct cfq_queue *async_idle_cfqq;
259 sector_t last_position;
262 * tunables, see top of file
264 unsigned int cfq_quantum;
265 unsigned int cfq_fifo_expire[2];
266 unsigned int cfq_back_penalty;
267 unsigned int cfq_back_max;
268 unsigned int cfq_slice[2];
269 unsigned int cfq_slice_async_rq;
270 unsigned int cfq_slice_idle;
271 unsigned int cfq_latency;
272 unsigned int cfq_group_isolation;
274 struct list_head cic_list;
277 * Fallback dummy cfqq for extreme OOM conditions
279 struct cfq_queue oom_cfqq;
281 unsigned long last_delayed_sync;
283 /* List of cfq groups being managed on this device*/
284 struct hlist_head cfqg_list;
288 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd);
290 static struct cfq_rb_root *service_tree_for(struct cfq_group *cfqg,
297 if (prio == IDLE_WORKLOAD)
298 return &cfqg->service_tree_idle;
300 return &cfqg->service_trees[prio][type];
303 enum cfqq_state_flags {
304 CFQ_CFQQ_FLAG_on_rr = 0, /* on round-robin busy list */
305 CFQ_CFQQ_FLAG_wait_request, /* waiting for a request */
306 CFQ_CFQQ_FLAG_must_dispatch, /* must be allowed a dispatch */
307 CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
308 CFQ_CFQQ_FLAG_fifo_expire, /* FIFO checked in this slice */
309 CFQ_CFQQ_FLAG_idle_window, /* slice idling enabled */
310 CFQ_CFQQ_FLAG_prio_changed, /* task priority has changed */
311 CFQ_CFQQ_FLAG_slice_new, /* no requests dispatched in slice */
312 CFQ_CFQQ_FLAG_sync, /* synchronous queue */
313 CFQ_CFQQ_FLAG_coop, /* cfqq is shared */
314 CFQ_CFQQ_FLAG_split_coop, /* shared cfqq will be splitted */
315 CFQ_CFQQ_FLAG_deep, /* sync cfqq experienced large depth */
316 CFQ_CFQQ_FLAG_wait_busy, /* Waiting for next request */
319 #define CFQ_CFQQ_FNS(name) \
320 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \
322 (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name); \
324 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \
326 (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \
328 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \
330 return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \
334 CFQ_CFQQ_FNS(wait_request);
335 CFQ_CFQQ_FNS(must_dispatch);
336 CFQ_CFQQ_FNS(must_alloc_slice);
337 CFQ_CFQQ_FNS(fifo_expire);
338 CFQ_CFQQ_FNS(idle_window);
339 CFQ_CFQQ_FNS(prio_changed);
340 CFQ_CFQQ_FNS(slice_new);
343 CFQ_CFQQ_FNS(split_coop);
345 CFQ_CFQQ_FNS(wait_busy);
348 #ifdef CONFIG_DEBUG_CFQ_IOSCHED
349 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
350 blk_add_trace_msg((cfqd)->queue, "cfq%d%c %s " fmt, (cfqq)->pid, \
351 cfq_cfqq_sync((cfqq)) ? 'S' : 'A', \
352 blkg_path(&(cfqq)->cfqg->blkg), ##args);
354 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...) \
355 blk_add_trace_msg((cfqd)->queue, "%s " fmt, \
356 blkg_path(&(cfqg)->blkg), ##args); \
359 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
360 blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args)
361 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...) do {} while (0);
363 #define cfq_log(cfqd, fmt, args...) \
364 blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
366 /* Traverses through cfq group service trees */
367 #define for_each_cfqg_st(cfqg, i, j, st) \
368 for (i = 0; i <= IDLE_WORKLOAD; i++) \
369 for (j = 0, st = i < IDLE_WORKLOAD ? &cfqg->service_trees[i][j]\
370 : &cfqg->service_tree_idle; \
371 (i < IDLE_WORKLOAD && j <= SYNC_WORKLOAD) || \
372 (i == IDLE_WORKLOAD && j == 0); \
373 j++, st = i < IDLE_WORKLOAD ? \
374 &cfqg->service_trees[i][j]: NULL) \
377 static inline enum wl_prio_t cfqq_prio(struct cfq_queue *cfqq)
379 if (cfq_class_idle(cfqq))
380 return IDLE_WORKLOAD;
381 if (cfq_class_rt(cfqq))
387 static enum wl_type_t cfqq_type(struct cfq_queue *cfqq)
389 if (!cfq_cfqq_sync(cfqq))
390 return ASYNC_WORKLOAD;
391 if (!cfq_cfqq_idle_window(cfqq))
392 return SYNC_NOIDLE_WORKLOAD;
393 return SYNC_WORKLOAD;
396 static inline int cfq_group_busy_queues_wl(enum wl_prio_t wl,
397 struct cfq_data *cfqd,
398 struct cfq_group *cfqg)
400 if (wl == IDLE_WORKLOAD)
401 return cfqg->service_tree_idle.count;
403 return cfqg->service_trees[wl][ASYNC_WORKLOAD].count
404 + cfqg->service_trees[wl][SYNC_NOIDLE_WORKLOAD].count
405 + cfqg->service_trees[wl][SYNC_WORKLOAD].count;
408 static inline int cfqg_busy_async_queues(struct cfq_data *cfqd,
409 struct cfq_group *cfqg)
411 return cfqg->service_trees[RT_WORKLOAD][ASYNC_WORKLOAD].count
412 + cfqg->service_trees[BE_WORKLOAD][ASYNC_WORKLOAD].count;
415 static void cfq_dispatch_insert(struct request_queue *, struct request *);
416 static struct cfq_queue *cfq_get_queue(struct cfq_data *, bool,
417 struct io_context *, gfp_t);
418 static struct cfq_io_context *cfq_cic_lookup(struct cfq_data *,
419 struct io_context *);
421 static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_context *cic,
424 return cic->cfqq[is_sync];
427 static inline void cic_set_cfqq(struct cfq_io_context *cic,
428 struct cfq_queue *cfqq, bool is_sync)
430 cic->cfqq[is_sync] = cfqq;
434 * We regard a request as SYNC, if it's either a read or has the SYNC bit
435 * set (in which case it could also be direct WRITE).
437 static inline bool cfq_bio_sync(struct bio *bio)
439 return bio_data_dir(bio) == READ || bio_rw_flagged(bio, BIO_RW_SYNCIO);
443 * scheduler run of queue, if there are requests pending and no one in the
444 * driver that will restart queueing
446 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
448 if (cfqd->busy_queues) {
449 cfq_log(cfqd, "schedule dispatch");
450 kblockd_schedule_work(cfqd->queue, &cfqd->unplug_work);
454 static int cfq_queue_empty(struct request_queue *q)
456 struct cfq_data *cfqd = q->elevator->elevator_data;
458 return !cfqd->rq_queued;
462 * Scale schedule slice based on io priority. Use the sync time slice only
463 * if a queue is marked sync and has sync io queued. A sync queue with async
464 * io only, should not get full sync slice length.
466 static inline int cfq_prio_slice(struct cfq_data *cfqd, bool sync,
469 const int base_slice = cfqd->cfq_slice[sync];
471 WARN_ON(prio >= IOPRIO_BE_NR);
473 return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
477 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
479 return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
482 static inline u64 cfq_scale_slice(unsigned long delta, struct cfq_group *cfqg)
484 u64 d = delta << CFQ_SERVICE_SHIFT;
486 d = d * BLKIO_WEIGHT_DEFAULT;
487 do_div(d, cfqg->weight);
491 static inline u64 max_vdisktime(u64 min_vdisktime, u64 vdisktime)
493 s64 delta = (s64)(vdisktime - min_vdisktime);
495 min_vdisktime = vdisktime;
497 return min_vdisktime;
500 static inline u64 min_vdisktime(u64 min_vdisktime, u64 vdisktime)
502 s64 delta = (s64)(vdisktime - min_vdisktime);
504 min_vdisktime = vdisktime;
506 return min_vdisktime;
509 static void update_min_vdisktime(struct cfq_rb_root *st)
511 u64 vdisktime = st->min_vdisktime;
512 struct cfq_group *cfqg;
515 cfqg = rb_entry_cfqg(st->active);
516 vdisktime = cfqg->vdisktime;
520 cfqg = rb_entry_cfqg(st->left);
521 vdisktime = min_vdisktime(vdisktime, cfqg->vdisktime);
524 st->min_vdisktime = max_vdisktime(st->min_vdisktime, vdisktime);
528 * get averaged number of queues of RT/BE priority.
529 * average is updated, with a formula that gives more weight to higher numbers,
530 * to quickly follows sudden increases and decrease slowly
533 static inline unsigned cfq_group_get_avg_queues(struct cfq_data *cfqd,
534 struct cfq_group *cfqg, bool rt)
536 unsigned min_q, max_q;
537 unsigned mult = cfq_hist_divisor - 1;
538 unsigned round = cfq_hist_divisor / 2;
539 unsigned busy = cfq_group_busy_queues_wl(rt, cfqd, cfqg);
541 min_q = min(cfqg->busy_queues_avg[rt], busy);
542 max_q = max(cfqg->busy_queues_avg[rt], busy);
543 cfqg->busy_queues_avg[rt] = (mult * max_q + min_q + round) /
545 return cfqg->busy_queues_avg[rt];
548 static inline unsigned
549 cfq_group_slice(struct cfq_data *cfqd, struct cfq_group *cfqg)
551 struct cfq_rb_root *st = &cfqd->grp_service_tree;
553 return cfq_target_latency * cfqg->weight / st->total_weight;
557 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
559 unsigned slice = cfq_prio_to_slice(cfqd, cfqq);
560 if (cfqd->cfq_latency) {
562 * interested queues (we consider only the ones with the same
563 * priority class in the cfq group)
565 unsigned iq = cfq_group_get_avg_queues(cfqd, cfqq->cfqg,
567 unsigned sync_slice = cfqd->cfq_slice[1];
568 unsigned expect_latency = sync_slice * iq;
569 unsigned group_slice = cfq_group_slice(cfqd, cfqq->cfqg);
571 if (expect_latency > group_slice) {
572 unsigned base_low_slice = 2 * cfqd->cfq_slice_idle;
573 /* scale low_slice according to IO priority
574 * and sync vs async */
576 min(slice, base_low_slice * slice / sync_slice);
577 /* the adapted slice value is scaled to fit all iqs
578 * into the target latency */
579 slice = max(slice * group_slice / expect_latency,
583 cfqq->slice_start = jiffies;
584 cfqq->slice_end = jiffies + slice;
585 cfqq->allocated_slice = slice;
586 cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
590 * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
591 * isn't valid until the first request from the dispatch is activated
592 * and the slice time set.
594 static inline bool cfq_slice_used(struct cfq_queue *cfqq)
596 if (cfq_cfqq_slice_new(cfqq))
598 if (time_before(jiffies, cfqq->slice_end))
605 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
606 * We choose the request that is closest to the head right now. Distance
607 * behind the head is penalized and only allowed to a certain extent.
609 static struct request *
610 cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2, sector_t last)
612 sector_t s1, s2, d1 = 0, d2 = 0;
613 unsigned long back_max;
614 #define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */
615 #define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */
616 unsigned wrap = 0; /* bit mask: requests behind the disk head? */
618 if (rq1 == NULL || rq1 == rq2)
623 if (rq_is_sync(rq1) && !rq_is_sync(rq2))
625 else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
627 if (rq_is_meta(rq1) && !rq_is_meta(rq2))
629 else if (rq_is_meta(rq2) && !rq_is_meta(rq1))
632 s1 = blk_rq_pos(rq1);
633 s2 = blk_rq_pos(rq2);
636 * by definition, 1KiB is 2 sectors
638 back_max = cfqd->cfq_back_max * 2;
641 * Strict one way elevator _except_ in the case where we allow
642 * short backward seeks which are biased as twice the cost of a
643 * similar forward seek.
647 else if (s1 + back_max >= last)
648 d1 = (last - s1) * cfqd->cfq_back_penalty;
650 wrap |= CFQ_RQ1_WRAP;
654 else if (s2 + back_max >= last)
655 d2 = (last - s2) * cfqd->cfq_back_penalty;
657 wrap |= CFQ_RQ2_WRAP;
659 /* Found required data */
662 * By doing switch() on the bit mask "wrap" we avoid having to
663 * check two variables for all permutations: --> faster!
666 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
682 case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
685 * Since both rqs are wrapped,
686 * start with the one that's further behind head
687 * (--> only *one* back seek required),
688 * since back seek takes more time than forward.
698 * The below is leftmost cache rbtree addon
700 static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
702 /* Service tree is empty */
707 root->left = rb_first(&root->rb);
710 return rb_entry(root->left, struct cfq_queue, rb_node);
715 static struct cfq_group *cfq_rb_first_group(struct cfq_rb_root *root)
718 root->left = rb_first(&root->rb);
721 return rb_entry_cfqg(root->left);
726 static void rb_erase_init(struct rb_node *n, struct rb_root *root)
732 static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
736 rb_erase_init(n, &root->rb);
741 * would be nice to take fifo expire time into account as well
743 static struct request *
744 cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
745 struct request *last)
747 struct rb_node *rbnext = rb_next(&last->rb_node);
748 struct rb_node *rbprev = rb_prev(&last->rb_node);
749 struct request *next = NULL, *prev = NULL;
751 BUG_ON(RB_EMPTY_NODE(&last->rb_node));
754 prev = rb_entry_rq(rbprev);
757 next = rb_entry_rq(rbnext);
759 rbnext = rb_first(&cfqq->sort_list);
760 if (rbnext && rbnext != &last->rb_node)
761 next = rb_entry_rq(rbnext);
764 return cfq_choose_req(cfqd, next, prev, blk_rq_pos(last));
767 static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
768 struct cfq_queue *cfqq)
771 * just an approximation, should be ok.
773 return (cfqq->cfqg->nr_cfqq - 1) * (cfq_prio_slice(cfqd, 1, 0) -
774 cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
778 cfqg_key(struct cfq_rb_root *st, struct cfq_group *cfqg)
780 return cfqg->vdisktime - st->min_vdisktime;
784 __cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
786 struct rb_node **node = &st->rb.rb_node;
787 struct rb_node *parent = NULL;
788 struct cfq_group *__cfqg;
789 s64 key = cfqg_key(st, cfqg);
792 while (*node != NULL) {
794 __cfqg = rb_entry_cfqg(parent);
796 if (key < cfqg_key(st, __cfqg))
797 node = &parent->rb_left;
799 node = &parent->rb_right;
805 st->left = &cfqg->rb_node;
807 rb_link_node(&cfqg->rb_node, parent, node);
808 rb_insert_color(&cfqg->rb_node, &st->rb);
812 cfq_group_service_tree_add(struct cfq_data *cfqd, struct cfq_group *cfqg)
814 struct cfq_rb_root *st = &cfqd->grp_service_tree;
815 struct cfq_group *__cfqg;
823 * Currently put the group at the end. Later implement something
824 * so that groups get lesser vtime based on their weights, so that
825 * if group does not loose all if it was not continously backlogged.
827 n = rb_last(&st->rb);
829 __cfqg = rb_entry_cfqg(n);
830 cfqg->vdisktime = __cfqg->vdisktime + CFQ_IDLE_DELAY;
832 cfqg->vdisktime = st->min_vdisktime;
834 __cfq_group_service_tree_add(st, cfqg);
836 st->total_weight += cfqg->weight;
840 cfq_group_service_tree_del(struct cfq_data *cfqd, struct cfq_group *cfqg)
842 struct cfq_rb_root *st = &cfqd->grp_service_tree;
844 if (st->active == &cfqg->rb_node)
847 BUG_ON(cfqg->nr_cfqq < 1);
850 /* If there are other cfq queues under this group, don't delete it */
854 cfq_log_cfqg(cfqd, cfqg, "del_from_rr group");
856 st->total_weight -= cfqg->weight;
857 if (!RB_EMPTY_NODE(&cfqg->rb_node))
858 cfq_rb_erase(&cfqg->rb_node, st);
859 cfqg->saved_workload_slice = 0;
860 blkiocg_update_dequeue_stats(&cfqg->blkg, 1);
863 static inline unsigned int cfq_cfqq_slice_usage(struct cfq_queue *cfqq)
865 unsigned int slice_used;
868 * Queue got expired before even a single request completed or
869 * got expired immediately after first request completion.
871 if (!cfqq->slice_start || cfqq->slice_start == jiffies) {
873 * Also charge the seek time incurred to the group, otherwise
874 * if there are mutiple queues in the group, each can dispatch
875 * a single request on seeky media and cause lots of seek time
876 * and group will never know it.
878 slice_used = max_t(unsigned, (jiffies - cfqq->dispatch_start),
881 slice_used = jiffies - cfqq->slice_start;
882 if (slice_used > cfqq->allocated_slice)
883 slice_used = cfqq->allocated_slice;
886 cfq_log_cfqq(cfqq->cfqd, cfqq, "sl_used=%u", slice_used);
890 static void cfq_group_served(struct cfq_data *cfqd, struct cfq_group *cfqg,
891 struct cfq_queue *cfqq, bool forced)
893 struct cfq_rb_root *st = &cfqd->grp_service_tree;
894 unsigned int used_sl, charge_sl;
895 int nr_sync = cfqg->nr_cfqq - cfqg_busy_async_queues(cfqd, cfqg)
896 - cfqg->service_tree_idle.count;
899 used_sl = charge_sl = cfq_cfqq_slice_usage(cfqq);
901 if (!cfq_cfqq_sync(cfqq) && !nr_sync)
902 charge_sl = cfqq->allocated_slice;
904 /* Can't update vdisktime while group is on service tree */
905 cfq_rb_erase(&cfqg->rb_node, st);
906 cfqg->vdisktime += cfq_scale_slice(charge_sl, cfqg);
907 __cfq_group_service_tree_add(st, cfqg);
909 /* This group is being expired. Save the context */
910 if (time_after(cfqd->workload_expires, jiffies)) {
911 cfqg->saved_workload_slice = cfqd->workload_expires
913 cfqg->saved_workload = cfqd->serving_type;
914 cfqg->saved_serving_prio = cfqd->serving_prio;
916 cfqg->saved_workload_slice = 0;
918 cfq_log_cfqg(cfqd, cfqg, "served: vt=%llu min_vt=%llu", cfqg->vdisktime,
920 blkiocg_update_timeslice_used(&cfqg->blkg, used_sl);
921 blkiocg_set_start_empty_time(&cfqg->blkg, forced);
924 #ifdef CONFIG_CFQ_GROUP_IOSCHED
925 static inline struct cfq_group *cfqg_of_blkg(struct blkio_group *blkg)
928 return container_of(blkg, struct cfq_group, blkg);
933 cfq_update_blkio_group_weight(struct blkio_group *blkg, unsigned int weight)
935 cfqg_of_blkg(blkg)->weight = weight;
938 static struct cfq_group *
939 cfq_find_alloc_cfqg(struct cfq_data *cfqd, struct cgroup *cgroup, int create)
941 struct blkio_cgroup *blkcg = cgroup_to_blkio_cgroup(cgroup);
942 struct cfq_group *cfqg = NULL;
945 struct cfq_rb_root *st;
946 struct backing_dev_info *bdi = &cfqd->queue->backing_dev_info;
947 unsigned int major, minor;
949 cfqg = cfqg_of_blkg(blkiocg_lookup_group(blkcg, key));
950 if (cfqg && !cfqg->blkg.dev && bdi->dev && dev_name(bdi->dev)) {
951 sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
952 cfqg->blkg.dev = MKDEV(major, minor);
958 cfqg = kzalloc_node(sizeof(*cfqg), GFP_ATOMIC, cfqd->queue->node);
962 for_each_cfqg_st(cfqg, i, j, st)
964 RB_CLEAR_NODE(&cfqg->rb_node);
967 * Take the initial reference that will be released on destroy
968 * This can be thought of a joint reference by cgroup and
969 * elevator which will be dropped by either elevator exit
970 * or cgroup deletion path depending on who is exiting first.
972 atomic_set(&cfqg->ref, 1);
974 /* Add group onto cgroup list */
975 sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
976 blkiocg_add_blkio_group(blkcg, &cfqg->blkg, (void *)cfqd,
977 MKDEV(major, minor));
978 cfqg->weight = blkcg_get_weight(blkcg, cfqg->blkg.dev);
980 /* Add group on cfqd list */
981 hlist_add_head(&cfqg->cfqd_node, &cfqd->cfqg_list);
988 * Search for the cfq group current task belongs to. If create = 1, then also
989 * create the cfq group if it does not exist. request_queue lock must be held.
991 static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd, int create)
993 struct cgroup *cgroup;
994 struct cfq_group *cfqg = NULL;
997 cgroup = task_cgroup(current, blkio_subsys_id);
998 cfqg = cfq_find_alloc_cfqg(cfqd, cgroup, create);
1000 cfqg = &cfqd->root_group;
1005 static inline struct cfq_group *cfq_ref_get_cfqg(struct cfq_group *cfqg)
1007 atomic_inc(&cfqg->ref);
1011 static void cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg)
1013 /* Currently, all async queues are mapped to root group */
1014 if (!cfq_cfqq_sync(cfqq))
1015 cfqg = &cfqq->cfqd->root_group;
1018 /* cfqq reference on cfqg */
1019 atomic_inc(&cfqq->cfqg->ref);
1022 static void cfq_put_cfqg(struct cfq_group *cfqg)
1024 struct cfq_rb_root *st;
1027 BUG_ON(atomic_read(&cfqg->ref) <= 0);
1028 if (!atomic_dec_and_test(&cfqg->ref))
1030 for_each_cfqg_st(cfqg, i, j, st)
1031 BUG_ON(!RB_EMPTY_ROOT(&st->rb) || st->active != NULL);
1035 static void cfq_destroy_cfqg(struct cfq_data *cfqd, struct cfq_group *cfqg)
1037 /* Something wrong if we are trying to remove same group twice */
1038 BUG_ON(hlist_unhashed(&cfqg->cfqd_node));
1040 hlist_del_init(&cfqg->cfqd_node);
1043 * Put the reference taken at the time of creation so that when all
1044 * queues are gone, group can be destroyed.
1049 static void cfq_release_cfq_groups(struct cfq_data *cfqd)
1051 struct hlist_node *pos, *n;
1052 struct cfq_group *cfqg;
1054 hlist_for_each_entry_safe(cfqg, pos, n, &cfqd->cfqg_list, cfqd_node) {
1056 * If cgroup removal path got to blk_group first and removed
1057 * it from cgroup list, then it will take care of destroying
1060 if (!blkiocg_del_blkio_group(&cfqg->blkg))
1061 cfq_destroy_cfqg(cfqd, cfqg);
1066 * Blk cgroup controller notification saying that blkio_group object is being
1067 * delinked as associated cgroup object is going away. That also means that
1068 * no new IO will come in this group. So get rid of this group as soon as
1069 * any pending IO in the group is finished.
1071 * This function is called under rcu_read_lock(). key is the rcu protected
1072 * pointer. That means "key" is a valid cfq_data pointer as long as we are rcu
1075 * "key" was fetched from blkio_group under blkio_cgroup->lock. That means
1076 * it should not be NULL as even if elevator was exiting, cgroup deltion
1077 * path got to it first.
1079 void cfq_unlink_blkio_group(void *key, struct blkio_group *blkg)
1081 unsigned long flags;
1082 struct cfq_data *cfqd = key;
1084 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1085 cfq_destroy_cfqg(cfqd, cfqg_of_blkg(blkg));
1086 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1089 #else /* GROUP_IOSCHED */
1090 static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd, int create)
1092 return &cfqd->root_group;
1095 static inline struct cfq_group *cfq_ref_get_cfqg(struct cfq_group *cfqg)
1101 cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) {
1105 static void cfq_release_cfq_groups(struct cfq_data *cfqd) {}
1106 static inline void cfq_put_cfqg(struct cfq_group *cfqg) {}
1108 #endif /* GROUP_IOSCHED */
1111 * The cfqd->service_trees holds all pending cfq_queue's that have
1112 * requests waiting to be processed. It is sorted in the order that
1113 * we will service the queues.
1115 static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1118 struct rb_node **p, *parent;
1119 struct cfq_queue *__cfqq;
1120 unsigned long rb_key;
1121 struct cfq_rb_root *service_tree;
1124 int group_changed = 0;
1126 #ifdef CONFIG_CFQ_GROUP_IOSCHED
1127 if (!cfqd->cfq_group_isolation
1128 && cfqq_type(cfqq) == SYNC_NOIDLE_WORKLOAD
1129 && cfqq->cfqg && cfqq->cfqg != &cfqd->root_group) {
1130 /* Move this cfq to root group */
1131 cfq_log_cfqq(cfqd, cfqq, "moving to root group");
1132 if (!RB_EMPTY_NODE(&cfqq->rb_node))
1133 cfq_group_service_tree_del(cfqd, cfqq->cfqg);
1134 cfqq->orig_cfqg = cfqq->cfqg;
1135 cfqq->cfqg = &cfqd->root_group;
1136 atomic_inc(&cfqd->root_group.ref);
1138 } else if (!cfqd->cfq_group_isolation
1139 && cfqq_type(cfqq) == SYNC_WORKLOAD && cfqq->orig_cfqg) {
1140 /* cfqq is sequential now needs to go to its original group */
1141 BUG_ON(cfqq->cfqg != &cfqd->root_group);
1142 if (!RB_EMPTY_NODE(&cfqq->rb_node))
1143 cfq_group_service_tree_del(cfqd, cfqq->cfqg);
1144 cfq_put_cfqg(cfqq->cfqg);
1145 cfqq->cfqg = cfqq->orig_cfqg;
1146 cfqq->orig_cfqg = NULL;
1148 cfq_log_cfqq(cfqd, cfqq, "moved to origin group");
1152 service_tree = service_tree_for(cfqq->cfqg, cfqq_prio(cfqq),
1154 if (cfq_class_idle(cfqq)) {
1155 rb_key = CFQ_IDLE_DELAY;
1156 parent = rb_last(&service_tree->rb);
1157 if (parent && parent != &cfqq->rb_node) {
1158 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1159 rb_key += __cfqq->rb_key;
1162 } else if (!add_front) {
1164 * Get our rb key offset. Subtract any residual slice
1165 * value carried from last service. A negative resid
1166 * count indicates slice overrun, and this should position
1167 * the next service time further away in the tree.
1169 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
1170 rb_key -= cfqq->slice_resid;
1171 cfqq->slice_resid = 0;
1174 __cfqq = cfq_rb_first(service_tree);
1175 rb_key += __cfqq ? __cfqq->rb_key : jiffies;
1178 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1181 * same position, nothing more to do
1183 if (rb_key == cfqq->rb_key &&
1184 cfqq->service_tree == service_tree)
1187 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1188 cfqq->service_tree = NULL;
1193 cfqq->service_tree = service_tree;
1194 p = &service_tree->rb.rb_node;
1199 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1202 * sort by key, that represents service time.
1204 if (time_before(rb_key, __cfqq->rb_key))
1207 n = &(*p)->rb_right;
1215 service_tree->left = &cfqq->rb_node;
1217 cfqq->rb_key = rb_key;
1218 rb_link_node(&cfqq->rb_node, parent, p);
1219 rb_insert_color(&cfqq->rb_node, &service_tree->rb);
1220 service_tree->count++;
1221 if ((add_front || !new_cfqq) && !group_changed)
1223 cfq_group_service_tree_add(cfqd, cfqq->cfqg);
1226 static struct cfq_queue *
1227 cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
1228 sector_t sector, struct rb_node **ret_parent,
1229 struct rb_node ***rb_link)
1231 struct rb_node **p, *parent;
1232 struct cfq_queue *cfqq = NULL;
1240 cfqq = rb_entry(parent, struct cfq_queue, p_node);
1243 * Sort strictly based on sector. Smallest to the left,
1244 * largest to the right.
1246 if (sector > blk_rq_pos(cfqq->next_rq))
1247 n = &(*p)->rb_right;
1248 else if (sector < blk_rq_pos(cfqq->next_rq))
1256 *ret_parent = parent;
1262 static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1264 struct rb_node **p, *parent;
1265 struct cfq_queue *__cfqq;
1268 rb_erase(&cfqq->p_node, cfqq->p_root);
1269 cfqq->p_root = NULL;
1272 if (cfq_class_idle(cfqq))
1277 cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
1278 __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
1279 blk_rq_pos(cfqq->next_rq), &parent, &p);
1281 rb_link_node(&cfqq->p_node, parent, p);
1282 rb_insert_color(&cfqq->p_node, cfqq->p_root);
1284 cfqq->p_root = NULL;
1288 * Update cfqq's position in the service tree.
1290 static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1293 * Resorting requires the cfqq to be on the RR list already.
1295 if (cfq_cfqq_on_rr(cfqq)) {
1296 cfq_service_tree_add(cfqd, cfqq, 0);
1297 cfq_prio_tree_add(cfqd, cfqq);
1302 * add to busy list of queues for service, trying to be fair in ordering
1303 * the pending list according to last request service
1305 static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1307 cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
1308 BUG_ON(cfq_cfqq_on_rr(cfqq));
1309 cfq_mark_cfqq_on_rr(cfqq);
1310 cfqd->busy_queues++;
1312 cfq_resort_rr_list(cfqd, cfqq);
1316 * Called when the cfqq no longer has requests pending, remove it from
1319 static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1321 cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
1322 BUG_ON(!cfq_cfqq_on_rr(cfqq));
1323 cfq_clear_cfqq_on_rr(cfqq);
1325 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1326 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1327 cfqq->service_tree = NULL;
1330 rb_erase(&cfqq->p_node, cfqq->p_root);
1331 cfqq->p_root = NULL;
1334 cfq_group_service_tree_del(cfqd, cfqq->cfqg);
1335 BUG_ON(!cfqd->busy_queues);
1336 cfqd->busy_queues--;
1340 * rb tree support functions
1342 static void cfq_del_rq_rb(struct request *rq)
1344 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1345 const int sync = rq_is_sync(rq);
1347 BUG_ON(!cfqq->queued[sync]);
1348 cfqq->queued[sync]--;
1350 elv_rb_del(&cfqq->sort_list, rq);
1352 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) {
1354 * Queue will be deleted from service tree when we actually
1355 * expire it later. Right now just remove it from prio tree
1359 rb_erase(&cfqq->p_node, cfqq->p_root);
1360 cfqq->p_root = NULL;
1365 static void cfq_add_rq_rb(struct request *rq)
1367 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1368 struct cfq_data *cfqd = cfqq->cfqd;
1369 struct request *__alias, *prev;
1371 cfqq->queued[rq_is_sync(rq)]++;
1374 * looks a little odd, but the first insert might return an alias.
1375 * if that happens, put the alias on the dispatch list
1377 while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL)
1378 cfq_dispatch_insert(cfqd->queue, __alias);
1380 if (!cfq_cfqq_on_rr(cfqq))
1381 cfq_add_cfqq_rr(cfqd, cfqq);
1384 * check if this request is a better next-serve candidate
1386 prev = cfqq->next_rq;
1387 cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq, cfqd->last_position);
1390 * adjust priority tree position, if ->next_rq changes
1392 if (prev != cfqq->next_rq)
1393 cfq_prio_tree_add(cfqd, cfqq);
1395 BUG_ON(!cfqq->next_rq);
1398 static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
1400 elv_rb_del(&cfqq->sort_list, rq);
1401 cfqq->queued[rq_is_sync(rq)]--;
1402 blkiocg_update_io_remove_stats(&(RQ_CFQG(rq))->blkg, rq_data_dir(rq),
1405 blkiocg_update_io_add_stats(&(RQ_CFQG(rq))->blkg,
1406 &cfqq->cfqd->serving_group->blkg, rq_data_dir(rq),
1410 static struct request *
1411 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
1413 struct task_struct *tsk = current;
1414 struct cfq_io_context *cic;
1415 struct cfq_queue *cfqq;
1417 cic = cfq_cic_lookup(cfqd, tsk->io_context);
1421 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
1423 sector_t sector = bio->bi_sector + bio_sectors(bio);
1425 return elv_rb_find(&cfqq->sort_list, sector);
1431 static void cfq_activate_request(struct request_queue *q, struct request *rq)
1433 struct cfq_data *cfqd = q->elevator->elevator_data;
1435 cfqd->rq_in_driver++;
1436 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
1437 cfqd->rq_in_driver);
1439 cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
1442 static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
1444 struct cfq_data *cfqd = q->elevator->elevator_data;
1446 WARN_ON(!cfqd->rq_in_driver);
1447 cfqd->rq_in_driver--;
1448 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
1449 cfqd->rq_in_driver);
1452 static void cfq_remove_request(struct request *rq)
1454 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1456 if (cfqq->next_rq == rq)
1457 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
1459 list_del_init(&rq->queuelist);
1462 cfqq->cfqd->rq_queued--;
1463 blkiocg_update_io_remove_stats(&(RQ_CFQG(rq))->blkg, rq_data_dir(rq),
1465 if (rq_is_meta(rq)) {
1466 WARN_ON(!cfqq->meta_pending);
1467 cfqq->meta_pending--;
1471 static int cfq_merge(struct request_queue *q, struct request **req,
1474 struct cfq_data *cfqd = q->elevator->elevator_data;
1475 struct request *__rq;
1477 __rq = cfq_find_rq_fmerge(cfqd, bio);
1478 if (__rq && elv_rq_merge_ok(__rq, bio)) {
1480 return ELEVATOR_FRONT_MERGE;
1483 return ELEVATOR_NO_MERGE;
1486 static void cfq_merged_request(struct request_queue *q, struct request *req,
1489 if (type == ELEVATOR_FRONT_MERGE) {
1490 struct cfq_queue *cfqq = RQ_CFQQ(req);
1492 cfq_reposition_rq_rb(cfqq, req);
1496 static void cfq_bio_merged(struct request_queue *q, struct request *req,
1499 blkiocg_update_io_merged_stats(&(RQ_CFQG(req))->blkg, bio_data_dir(bio),
1504 cfq_merged_requests(struct request_queue *q, struct request *rq,
1505 struct request *next)
1507 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1509 * reposition in fifo if next is older than rq
1511 if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
1512 time_before(rq_fifo_time(next), rq_fifo_time(rq))) {
1513 list_move(&rq->queuelist, &next->queuelist);
1514 rq_set_fifo_time(rq, rq_fifo_time(next));
1517 if (cfqq->next_rq == next)
1519 cfq_remove_request(next);
1520 blkiocg_update_io_merged_stats(&(RQ_CFQG(rq))->blkg, rq_data_dir(next),
1524 static int cfq_allow_merge(struct request_queue *q, struct request *rq,
1527 struct cfq_data *cfqd = q->elevator->elevator_data;
1528 struct cfq_io_context *cic;
1529 struct cfq_queue *cfqq;
1532 * Disallow merge of a sync bio into an async request.
1534 if (cfq_bio_sync(bio) && !rq_is_sync(rq))
1538 * Lookup the cfqq that this bio will be queued with. Allow
1539 * merge only if rq is queued there.
1541 cic = cfq_cic_lookup(cfqd, current->io_context);
1545 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
1546 return cfqq == RQ_CFQQ(rq);
1549 static inline void cfq_del_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1551 del_timer(&cfqd->idle_slice_timer);
1552 blkiocg_update_idle_time_stats(&cfqq->cfqg->blkg);
1555 static void __cfq_set_active_queue(struct cfq_data *cfqd,
1556 struct cfq_queue *cfqq)
1559 cfq_log_cfqq(cfqd, cfqq, "set_active wl_prio:%d wl_type:%d",
1560 cfqd->serving_prio, cfqd->serving_type);
1561 blkiocg_update_avg_queue_size_stats(&cfqq->cfqg->blkg);
1562 cfqq->slice_start = 0;
1563 cfqq->dispatch_start = jiffies;
1564 cfqq->allocated_slice = 0;
1565 cfqq->slice_end = 0;
1566 cfqq->slice_dispatch = 0;
1568 cfq_clear_cfqq_wait_request(cfqq);
1569 cfq_clear_cfqq_must_dispatch(cfqq);
1570 cfq_clear_cfqq_must_alloc_slice(cfqq);
1571 cfq_clear_cfqq_fifo_expire(cfqq);
1572 cfq_mark_cfqq_slice_new(cfqq);
1574 cfq_del_timer(cfqd, cfqq);
1577 cfqd->active_queue = cfqq;
1581 * current cfqq expired its slice (or was too idle), select new one
1584 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1585 bool timed_out, bool forced)
1587 cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
1589 if (cfq_cfqq_wait_request(cfqq))
1590 cfq_del_timer(cfqd, cfqq);
1592 cfq_clear_cfqq_wait_request(cfqq);
1593 cfq_clear_cfqq_wait_busy(cfqq);
1596 * If this cfqq is shared between multiple processes, check to
1597 * make sure that those processes are still issuing I/Os within
1598 * the mean seek distance. If not, it may be time to break the
1599 * queues apart again.
1601 if (cfq_cfqq_coop(cfqq) && CFQQ_SEEKY(cfqq))
1602 cfq_mark_cfqq_split_coop(cfqq);
1605 * store what was left of this slice, if the queue idled/timed out
1607 if (timed_out && !cfq_cfqq_slice_new(cfqq)) {
1608 cfqq->slice_resid = cfqq->slice_end - jiffies;
1609 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
1612 cfq_group_served(cfqd, cfqq->cfqg, cfqq, forced);
1614 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
1615 cfq_del_cfqq_rr(cfqd, cfqq);
1617 cfq_resort_rr_list(cfqd, cfqq);
1619 if (cfqq == cfqd->active_queue)
1620 cfqd->active_queue = NULL;
1622 if (&cfqq->cfqg->rb_node == cfqd->grp_service_tree.active)
1623 cfqd->grp_service_tree.active = NULL;
1625 if (cfqd->active_cic) {
1626 put_io_context(cfqd->active_cic->ioc);
1627 cfqd->active_cic = NULL;
1631 static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out,
1634 struct cfq_queue *cfqq = cfqd->active_queue;
1637 __cfq_slice_expired(cfqd, cfqq, timed_out, forced);
1641 * Get next queue for service. Unless we have a queue preemption,
1642 * we'll simply select the first cfqq in the service tree.
1644 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
1646 struct cfq_rb_root *service_tree =
1647 service_tree_for(cfqd->serving_group, cfqd->serving_prio,
1648 cfqd->serving_type);
1650 if (!cfqd->rq_queued)
1653 /* There is nothing to dispatch */
1656 if (RB_EMPTY_ROOT(&service_tree->rb))
1658 return cfq_rb_first(service_tree);
1661 static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd)
1663 struct cfq_group *cfqg;
1664 struct cfq_queue *cfqq;
1666 struct cfq_rb_root *st;
1668 if (!cfqd->rq_queued)
1671 cfqg = cfq_get_next_cfqg(cfqd);
1675 for_each_cfqg_st(cfqg, i, j, st)
1676 if ((cfqq = cfq_rb_first(st)) != NULL)
1682 * Get and set a new active queue for service.
1684 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
1685 struct cfq_queue *cfqq)
1688 cfqq = cfq_get_next_queue(cfqd);
1690 __cfq_set_active_queue(cfqd, cfqq);
1694 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
1697 if (blk_rq_pos(rq) >= cfqd->last_position)
1698 return blk_rq_pos(rq) - cfqd->last_position;
1700 return cfqd->last_position - blk_rq_pos(rq);
1703 static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1706 return cfq_dist_from_last(cfqd, rq) <= CFQQ_CLOSE_THR;
1709 static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
1710 struct cfq_queue *cur_cfqq)
1712 struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
1713 struct rb_node *parent, *node;
1714 struct cfq_queue *__cfqq;
1715 sector_t sector = cfqd->last_position;
1717 if (RB_EMPTY_ROOT(root))
1721 * First, if we find a request starting at the end of the last
1722 * request, choose it.
1724 __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
1729 * If the exact sector wasn't found, the parent of the NULL leaf
1730 * will contain the closest sector.
1732 __cfqq = rb_entry(parent, struct cfq_queue, p_node);
1733 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1736 if (blk_rq_pos(__cfqq->next_rq) < sector)
1737 node = rb_next(&__cfqq->p_node);
1739 node = rb_prev(&__cfqq->p_node);
1743 __cfqq = rb_entry(node, struct cfq_queue, p_node);
1744 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1752 * cur_cfqq - passed in so that we don't decide that the current queue is
1753 * closely cooperating with itself.
1755 * So, basically we're assuming that that cur_cfqq has dispatched at least
1756 * one request, and that cfqd->last_position reflects a position on the disk
1757 * associated with the I/O issued by cur_cfqq. I'm not sure this is a valid
1760 static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
1761 struct cfq_queue *cur_cfqq)
1763 struct cfq_queue *cfqq;
1765 if (cfq_class_idle(cur_cfqq))
1767 if (!cfq_cfqq_sync(cur_cfqq))
1769 if (CFQQ_SEEKY(cur_cfqq))
1773 * Don't search priority tree if it's the only queue in the group.
1775 if (cur_cfqq->cfqg->nr_cfqq == 1)
1779 * We should notice if some of the queues are cooperating, eg
1780 * working closely on the same area of the disk. In that case,
1781 * we can group them together and don't waste time idling.
1783 cfqq = cfqq_close(cfqd, cur_cfqq);
1787 /* If new queue belongs to different cfq_group, don't choose it */
1788 if (cur_cfqq->cfqg != cfqq->cfqg)
1792 * It only makes sense to merge sync queues.
1794 if (!cfq_cfqq_sync(cfqq))
1796 if (CFQQ_SEEKY(cfqq))
1800 * Do not merge queues of different priority classes
1802 if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq))
1809 * Determine whether we should enforce idle window for this queue.
1812 static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1814 enum wl_prio_t prio = cfqq_prio(cfqq);
1815 struct cfq_rb_root *service_tree = cfqq->service_tree;
1817 BUG_ON(!service_tree);
1818 BUG_ON(!service_tree->count);
1820 /* We never do for idle class queues. */
1821 if (prio == IDLE_WORKLOAD)
1824 /* We do for queues that were marked with idle window flag. */
1825 if (cfq_cfqq_idle_window(cfqq) &&
1826 !(blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag))
1830 * Otherwise, we do only if they are the last ones
1831 * in their service tree.
1833 if (service_tree->count == 1 && cfq_cfqq_sync(cfqq))
1835 cfq_log_cfqq(cfqd, cfqq, "Not idling. st->count:%d",
1836 service_tree->count);
1840 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
1842 struct cfq_queue *cfqq = cfqd->active_queue;
1843 struct cfq_io_context *cic;
1847 * SSD device without seek penalty, disable idling. But only do so
1848 * for devices that support queuing, otherwise we still have a problem
1849 * with sync vs async workloads.
1851 if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
1854 WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
1855 WARN_ON(cfq_cfqq_slice_new(cfqq));
1858 * idle is disabled, either manually or by past process history
1860 if (!cfqd->cfq_slice_idle || !cfq_should_idle(cfqd, cfqq))
1864 * still active requests from this queue, don't idle
1866 if (cfqq->dispatched)
1870 * task has exited, don't wait
1872 cic = cfqd->active_cic;
1873 if (!cic || !atomic_read(&cic->ioc->nr_tasks))
1877 * If our average think time is larger than the remaining time
1878 * slice, then don't idle. This avoids overrunning the allotted
1881 if (sample_valid(cic->ttime_samples) &&
1882 (cfqq->slice_end - jiffies < cic->ttime_mean)) {
1883 cfq_log_cfqq(cfqd, cfqq, "Not idling. think_time:%d",
1888 cfq_mark_cfqq_wait_request(cfqq);
1890 sl = cfqd->cfq_slice_idle;
1892 mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
1893 blkiocg_update_set_idle_time_stats(&cfqq->cfqg->blkg);
1894 cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu", sl);
1898 * Move request from internal lists to the request queue dispatch list.
1900 static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
1902 struct cfq_data *cfqd = q->elevator->elevator_data;
1903 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1905 cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
1907 cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
1908 cfq_remove_request(rq);
1910 elv_dispatch_sort(q, rq);
1912 cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]++;
1913 blkiocg_update_dispatch_stats(&cfqq->cfqg->blkg, blk_rq_bytes(rq),
1914 rq_data_dir(rq), rq_is_sync(rq));
1918 * return expired entry, or NULL to just start from scratch in rbtree
1920 static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
1922 struct request *rq = NULL;
1924 if (cfq_cfqq_fifo_expire(cfqq))
1927 cfq_mark_cfqq_fifo_expire(cfqq);
1929 if (list_empty(&cfqq->fifo))
1932 rq = rq_entry_fifo(cfqq->fifo.next);
1933 if (time_before(jiffies, rq_fifo_time(rq)))
1936 cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq);
1941 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1943 const int base_rq = cfqd->cfq_slice_async_rq;
1945 WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
1947 return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
1951 * Must be called with the queue_lock held.
1953 static int cfqq_process_refs(struct cfq_queue *cfqq)
1955 int process_refs, io_refs;
1957 io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE];
1958 process_refs = atomic_read(&cfqq->ref) - io_refs;
1959 BUG_ON(process_refs < 0);
1960 return process_refs;
1963 static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq)
1965 int process_refs, new_process_refs;
1966 struct cfq_queue *__cfqq;
1968 /* Avoid a circular list and skip interim queue merges */
1969 while ((__cfqq = new_cfqq->new_cfqq)) {
1975 process_refs = cfqq_process_refs(cfqq);
1977 * If the process for the cfqq has gone away, there is no
1978 * sense in merging the queues.
1980 if (process_refs == 0)
1984 * Merge in the direction of the lesser amount of work.
1986 new_process_refs = cfqq_process_refs(new_cfqq);
1987 if (new_process_refs >= process_refs) {
1988 cfqq->new_cfqq = new_cfqq;
1989 atomic_add(process_refs, &new_cfqq->ref);
1991 new_cfqq->new_cfqq = cfqq;
1992 atomic_add(new_process_refs, &cfqq->ref);
1996 static enum wl_type_t cfq_choose_wl(struct cfq_data *cfqd,
1997 struct cfq_group *cfqg, enum wl_prio_t prio)
1999 struct cfq_queue *queue;
2001 bool key_valid = false;
2002 unsigned long lowest_key = 0;
2003 enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD;
2005 for (i = 0; i <= SYNC_WORKLOAD; ++i) {
2006 /* select the one with lowest rb_key */
2007 queue = cfq_rb_first(service_tree_for(cfqg, prio, i));
2009 (!key_valid || time_before(queue->rb_key, lowest_key))) {
2010 lowest_key = queue->rb_key;
2019 static void choose_service_tree(struct cfq_data *cfqd, struct cfq_group *cfqg)
2023 struct cfq_rb_root *st;
2024 unsigned group_slice;
2027 cfqd->serving_prio = IDLE_WORKLOAD;
2028 cfqd->workload_expires = jiffies + 1;
2032 /* Choose next priority. RT > BE > IDLE */
2033 if (cfq_group_busy_queues_wl(RT_WORKLOAD, cfqd, cfqg))
2034 cfqd->serving_prio = RT_WORKLOAD;
2035 else if (cfq_group_busy_queues_wl(BE_WORKLOAD, cfqd, cfqg))
2036 cfqd->serving_prio = BE_WORKLOAD;
2038 cfqd->serving_prio = IDLE_WORKLOAD;
2039 cfqd->workload_expires = jiffies + 1;
2044 * For RT and BE, we have to choose also the type
2045 * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
2048 st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
2052 * check workload expiration, and that we still have other queues ready
2054 if (count && !time_after(jiffies, cfqd->workload_expires))
2057 /* otherwise select new workload type */
2058 cfqd->serving_type =
2059 cfq_choose_wl(cfqd, cfqg, cfqd->serving_prio);
2060 st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
2064 * the workload slice is computed as a fraction of target latency
2065 * proportional to the number of queues in that workload, over
2066 * all the queues in the same priority class
2068 group_slice = cfq_group_slice(cfqd, cfqg);
2070 slice = group_slice * count /
2071 max_t(unsigned, cfqg->busy_queues_avg[cfqd->serving_prio],
2072 cfq_group_busy_queues_wl(cfqd->serving_prio, cfqd, cfqg));
2074 if (cfqd->serving_type == ASYNC_WORKLOAD) {
2078 * Async queues are currently system wide. Just taking
2079 * proportion of queues with-in same group will lead to higher
2080 * async ratio system wide as generally root group is going
2081 * to have higher weight. A more accurate thing would be to
2082 * calculate system wide asnc/sync ratio.
2084 tmp = cfq_target_latency * cfqg_busy_async_queues(cfqd, cfqg);
2085 tmp = tmp/cfqd->busy_queues;
2086 slice = min_t(unsigned, slice, tmp);
2088 /* async workload slice is scaled down according to
2089 * the sync/async slice ratio. */
2090 slice = slice * cfqd->cfq_slice[0] / cfqd->cfq_slice[1];
2092 /* sync workload slice is at least 2 * cfq_slice_idle */
2093 slice = max(slice, 2 * cfqd->cfq_slice_idle);
2095 slice = max_t(unsigned, slice, CFQ_MIN_TT);
2096 cfq_log(cfqd, "workload slice:%d", slice);
2097 cfqd->workload_expires = jiffies + slice;
2098 cfqd->noidle_tree_requires_idle = false;
2101 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd)
2103 struct cfq_rb_root *st = &cfqd->grp_service_tree;
2104 struct cfq_group *cfqg;
2106 if (RB_EMPTY_ROOT(&st->rb))
2108 cfqg = cfq_rb_first_group(st);
2109 st->active = &cfqg->rb_node;
2110 update_min_vdisktime(st);
2114 static void cfq_choose_cfqg(struct cfq_data *cfqd)
2116 struct cfq_group *cfqg = cfq_get_next_cfqg(cfqd);
2118 cfqd->serving_group = cfqg;
2120 /* Restore the workload type data */
2121 if (cfqg->saved_workload_slice) {
2122 cfqd->workload_expires = jiffies + cfqg->saved_workload_slice;
2123 cfqd->serving_type = cfqg->saved_workload;
2124 cfqd->serving_prio = cfqg->saved_serving_prio;
2126 cfqd->workload_expires = jiffies - 1;
2128 choose_service_tree(cfqd, cfqg);
2132 * Select a queue for service. If we have a current active queue,
2133 * check whether to continue servicing it, or retrieve and set a new one.
2135 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
2137 struct cfq_queue *cfqq, *new_cfqq = NULL;
2139 cfqq = cfqd->active_queue;
2143 if (!cfqd->rq_queued)
2147 * We were waiting for group to get backlogged. Expire the queue
2149 if (cfq_cfqq_wait_busy(cfqq) && !RB_EMPTY_ROOT(&cfqq->sort_list))
2153 * The active queue has run out of time, expire it and select new.
2155 if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq)) {
2157 * If slice had not expired at the completion of last request
2158 * we might not have turned on wait_busy flag. Don't expire
2159 * the queue yet. Allow the group to get backlogged.
2161 * The very fact that we have used the slice, that means we
2162 * have been idling all along on this queue and it should be
2163 * ok to wait for this request to complete.
2165 if (cfqq->cfqg->nr_cfqq == 1 && RB_EMPTY_ROOT(&cfqq->sort_list)
2166 && cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2174 * The active queue has requests and isn't expired, allow it to
2177 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
2181 * If another queue has a request waiting within our mean seek
2182 * distance, let it run. The expire code will check for close
2183 * cooperators and put the close queue at the front of the service
2184 * tree. If possible, merge the expiring queue with the new cfqq.
2186 new_cfqq = cfq_close_cooperator(cfqd, cfqq);
2188 if (!cfqq->new_cfqq)
2189 cfq_setup_merge(cfqq, new_cfqq);
2194 * No requests pending. If the active queue still has requests in
2195 * flight or is idling for a new request, allow either of these
2196 * conditions to happen (or time out) before selecting a new queue.
2198 if (timer_pending(&cfqd->idle_slice_timer) ||
2199 (cfqq->dispatched && cfq_should_idle(cfqd, cfqq))) {
2205 cfq_slice_expired(cfqd, 0, false);
2208 * Current queue expired. Check if we have to switch to a new
2212 cfq_choose_cfqg(cfqd);
2214 cfqq = cfq_set_active_queue(cfqd, new_cfqq);
2219 static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
2223 while (cfqq->next_rq) {
2224 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
2228 BUG_ON(!list_empty(&cfqq->fifo));
2230 /* By default cfqq is not expired if it is empty. Do it explicitly */
2231 __cfq_slice_expired(cfqq->cfqd, cfqq, 0, true);
2236 * Drain our current requests. Used for barriers and when switching
2237 * io schedulers on-the-fly.
2239 static int cfq_forced_dispatch(struct cfq_data *cfqd)
2241 struct cfq_queue *cfqq;
2244 /* Expire the timeslice of the current active queue first */
2245 cfq_slice_expired(cfqd, 0, true);
2246 while ((cfqq = cfq_get_next_queue_forced(cfqd)) != NULL) {
2247 __cfq_set_active_queue(cfqd, cfqq);
2248 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
2251 BUG_ON(cfqd->busy_queues);
2253 cfq_log(cfqd, "forced_dispatch=%d", dispatched);
2257 static inline bool cfq_slice_used_soon(struct cfq_data *cfqd,
2258 struct cfq_queue *cfqq)
2260 /* the queue hasn't finished any request, can't estimate */
2261 if (cfq_cfqq_slice_new(cfqq))
2263 if (time_after(jiffies + cfqd->cfq_slice_idle * cfqq->dispatched,
2270 static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2272 unsigned int max_dispatch;
2275 * Drain async requests before we start sync IO
2277 if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_flight[BLK_RW_ASYNC])
2281 * If this is an async queue and we have sync IO in flight, let it wait
2283 if (cfqd->rq_in_flight[BLK_RW_SYNC] && !cfq_cfqq_sync(cfqq))
2286 max_dispatch = max_t(unsigned int, cfqd->cfq_quantum / 2, 1);
2287 if (cfq_class_idle(cfqq))
2291 * Does this cfqq already have too much IO in flight?
2293 if (cfqq->dispatched >= max_dispatch) {
2295 * idle queue must always only have a single IO in flight
2297 if (cfq_class_idle(cfqq))
2301 * We have other queues, don't allow more IO from this one
2303 if (cfqd->busy_queues > 1 && cfq_slice_used_soon(cfqd, cfqq))
2307 * Sole queue user, no limit
2309 if (cfqd->busy_queues == 1)
2313 * Normally we start throttling cfqq when cfq_quantum/2
2314 * requests have been dispatched. But we can drive
2315 * deeper queue depths at the beginning of slice
2316 * subjected to upper limit of cfq_quantum.
2318 max_dispatch = cfqd->cfq_quantum;
2322 * Async queues must wait a bit before being allowed dispatch.
2323 * We also ramp up the dispatch depth gradually for async IO,
2324 * based on the last sync IO we serviced
2326 if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
2327 unsigned long last_sync = jiffies - cfqd->last_delayed_sync;
2330 depth = last_sync / cfqd->cfq_slice[1];
2331 if (!depth && !cfqq->dispatched)
2333 if (depth < max_dispatch)
2334 max_dispatch = depth;
2338 * If we're below the current max, allow a dispatch
2340 return cfqq->dispatched < max_dispatch;
2344 * Dispatch a request from cfqq, moving them to the request queue
2347 static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2351 BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
2353 if (!cfq_may_dispatch(cfqd, cfqq))
2357 * follow expired path, else get first next available
2359 rq = cfq_check_fifo(cfqq);
2364 * insert request into driver dispatch list
2366 cfq_dispatch_insert(cfqd->queue, rq);
2368 if (!cfqd->active_cic) {
2369 struct cfq_io_context *cic = RQ_CIC(rq);
2371 atomic_long_inc(&cic->ioc->refcount);
2372 cfqd->active_cic = cic;
2379 * Find the cfqq that we need to service and move a request from that to the
2382 static int cfq_dispatch_requests(struct request_queue *q, int force)
2384 struct cfq_data *cfqd = q->elevator->elevator_data;
2385 struct cfq_queue *cfqq;
2387 if (!cfqd->busy_queues)
2390 if (unlikely(force))
2391 return cfq_forced_dispatch(cfqd);
2393 cfqq = cfq_select_queue(cfqd);
2398 * Dispatch a request from this cfqq, if it is allowed
2400 if (!cfq_dispatch_request(cfqd, cfqq))
2403 cfqq->slice_dispatch++;
2404 cfq_clear_cfqq_must_dispatch(cfqq);
2407 * expire an async queue immediately if it has used up its slice. idle
2408 * queue always expire after 1 dispatch round.
2410 if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
2411 cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
2412 cfq_class_idle(cfqq))) {
2413 cfqq->slice_end = jiffies + 1;
2414 cfq_slice_expired(cfqd, 0, false);
2417 cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
2422 * task holds one reference to the queue, dropped when task exits. each rq
2423 * in-flight on this queue also holds a reference, dropped when rq is freed.
2425 * Each cfq queue took a reference on the parent group. Drop it now.
2426 * queue lock must be held here.
2428 static void cfq_put_queue(struct cfq_queue *cfqq)
2430 struct cfq_data *cfqd = cfqq->cfqd;
2431 struct cfq_group *cfqg, *orig_cfqg;
2433 BUG_ON(atomic_read(&cfqq->ref) <= 0);
2435 if (!atomic_dec_and_test(&cfqq->ref))
2438 cfq_log_cfqq(cfqd, cfqq, "put_queue");
2439 BUG_ON(rb_first(&cfqq->sort_list));
2440 BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
2442 orig_cfqg = cfqq->orig_cfqg;
2444 if (unlikely(cfqd->active_queue == cfqq)) {
2445 __cfq_slice_expired(cfqd, cfqq, 0, false);
2446 cfq_schedule_dispatch(cfqd);
2449 BUG_ON(cfq_cfqq_on_rr(cfqq));
2450 kmem_cache_free(cfq_pool, cfqq);
2453 cfq_put_cfqg(orig_cfqg);
2457 * Must always be called with the rcu_read_lock() held
2460 __call_for_each_cic(struct io_context *ioc,
2461 void (*func)(struct io_context *, struct cfq_io_context *))
2463 struct cfq_io_context *cic;
2464 struct hlist_node *n;
2466 hlist_for_each_entry_rcu(cic, n, &ioc->cic_list, cic_list)
2471 * Call func for each cic attached to this ioc.
2474 call_for_each_cic(struct io_context *ioc,
2475 void (*func)(struct io_context *, struct cfq_io_context *))
2478 __call_for_each_cic(ioc, func);
2482 static void cfq_cic_free_rcu(struct rcu_head *head)
2484 struct cfq_io_context *cic;
2486 cic = container_of(head, struct cfq_io_context, rcu_head);
2488 kmem_cache_free(cfq_ioc_pool, cic);
2489 elv_ioc_count_dec(cfq_ioc_count);
2493 * CFQ scheduler is exiting, grab exit lock and check
2494 * the pending io context count. If it hits zero,
2495 * complete ioc_gone and set it back to NULL
2497 spin_lock(&ioc_gone_lock);
2498 if (ioc_gone && !elv_ioc_count_read(cfq_ioc_count)) {
2502 spin_unlock(&ioc_gone_lock);
2506 static void cfq_cic_free(struct cfq_io_context *cic)
2508 call_rcu(&cic->rcu_head, cfq_cic_free_rcu);
2511 static void cic_free_func(struct io_context *ioc, struct cfq_io_context *cic)
2513 unsigned long flags;
2515 BUG_ON(!cic->dead_key);
2517 spin_lock_irqsave(&ioc->lock, flags);
2518 radix_tree_delete(&ioc->radix_root, cic->dead_key);
2519 hlist_del_rcu(&cic->cic_list);
2520 spin_unlock_irqrestore(&ioc->lock, flags);
2526 * Must be called with rcu_read_lock() held or preemption otherwise disabled.
2527 * Only two callers of this - ->dtor() which is called with the rcu_read_lock(),
2528 * and ->trim() which is called with the task lock held
2530 static void cfq_free_io_context(struct io_context *ioc)
2533 * ioc->refcount is zero here, or we are called from elv_unregister(),
2534 * so no more cic's are allowed to be linked into this ioc. So it
2535 * should be ok to iterate over the known list, we will see all cic's
2536 * since no new ones are added.
2538 __call_for_each_cic(ioc, cic_free_func);
2541 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2543 struct cfq_queue *__cfqq, *next;
2545 if (unlikely(cfqq == cfqd->active_queue)) {
2546 __cfq_slice_expired(cfqd, cfqq, 0, false);
2547 cfq_schedule_dispatch(cfqd);
2551 * If this queue was scheduled to merge with another queue, be
2552 * sure to drop the reference taken on that queue (and others in
2553 * the merge chain). See cfq_setup_merge and cfq_merge_cfqqs.
2555 __cfqq = cfqq->new_cfqq;
2557 if (__cfqq == cfqq) {
2558 WARN(1, "cfqq->new_cfqq loop detected\n");
2561 next = __cfqq->new_cfqq;
2562 cfq_put_queue(__cfqq);
2566 cfq_put_queue(cfqq);
2569 static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
2570 struct cfq_io_context *cic)
2572 struct io_context *ioc = cic->ioc;
2574 list_del_init(&cic->queue_list);
2577 * Make sure key == NULL is seen for dead queues
2580 cic->dead_key = (unsigned long) cic->key;
2583 if (ioc->ioc_data == cic)
2584 rcu_assign_pointer(ioc->ioc_data, NULL);
2586 if (cic->cfqq[BLK_RW_ASYNC]) {
2587 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]);
2588 cic->cfqq[BLK_RW_ASYNC] = NULL;
2591 if (cic->cfqq[BLK_RW_SYNC]) {
2592 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]);
2593 cic->cfqq[BLK_RW_SYNC] = NULL;
2597 static void cfq_exit_single_io_context(struct io_context *ioc,
2598 struct cfq_io_context *cic)
2600 struct cfq_data *cfqd = cic->key;
2603 struct request_queue *q = cfqd->queue;
2604 unsigned long flags;
2606 spin_lock_irqsave(q->queue_lock, flags);
2609 * Ensure we get a fresh copy of the ->key to prevent
2610 * race between exiting task and queue
2612 smp_read_barrier_depends();
2614 __cfq_exit_single_io_context(cfqd, cic);
2616 spin_unlock_irqrestore(q->queue_lock, flags);
2621 * The process that ioc belongs to has exited, we need to clean up
2622 * and put the internal structures we have that belongs to that process.
2624 static void cfq_exit_io_context(struct io_context *ioc)
2626 call_for_each_cic(ioc, cfq_exit_single_io_context);
2629 static struct cfq_io_context *
2630 cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
2632 struct cfq_io_context *cic;
2634 cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO,
2637 cic->last_end_request = jiffies;
2638 INIT_LIST_HEAD(&cic->queue_list);
2639 INIT_HLIST_NODE(&cic->cic_list);
2640 cic->dtor = cfq_free_io_context;
2641 cic->exit = cfq_exit_io_context;
2642 elv_ioc_count_inc(cfq_ioc_count);
2648 static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc)
2650 struct task_struct *tsk = current;
2653 if (!cfq_cfqq_prio_changed(cfqq))
2656 ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio);
2657 switch (ioprio_class) {
2659 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
2660 case IOPRIO_CLASS_NONE:
2662 * no prio set, inherit CPU scheduling settings
2664 cfqq->ioprio = task_nice_ioprio(tsk);
2665 cfqq->ioprio_class = task_nice_ioclass(tsk);
2667 case IOPRIO_CLASS_RT:
2668 cfqq->ioprio = task_ioprio(ioc);
2669 cfqq->ioprio_class = IOPRIO_CLASS_RT;
2671 case IOPRIO_CLASS_BE:
2672 cfqq->ioprio = task_ioprio(ioc);
2673 cfqq->ioprio_class = IOPRIO_CLASS_BE;
2675 case IOPRIO_CLASS_IDLE:
2676 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
2678 cfq_clear_cfqq_idle_window(cfqq);
2683 * keep track of original prio settings in case we have to temporarily
2684 * elevate the priority of this queue
2686 cfqq->org_ioprio = cfqq->ioprio;
2687 cfqq->org_ioprio_class = cfqq->ioprio_class;
2688 cfq_clear_cfqq_prio_changed(cfqq);
2691 static void changed_ioprio(struct io_context *ioc, struct cfq_io_context *cic)
2693 struct cfq_data *cfqd = cic->key;
2694 struct cfq_queue *cfqq;
2695 unsigned long flags;
2697 if (unlikely(!cfqd))
2700 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2702 cfqq = cic->cfqq[BLK_RW_ASYNC];
2704 struct cfq_queue *new_cfqq;
2705 new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic->ioc,
2708 cic->cfqq[BLK_RW_ASYNC] = new_cfqq;
2709 cfq_put_queue(cfqq);
2713 cfqq = cic->cfqq[BLK_RW_SYNC];
2715 cfq_mark_cfqq_prio_changed(cfqq);
2717 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2720 static void cfq_ioc_set_ioprio(struct io_context *ioc)
2722 call_for_each_cic(ioc, changed_ioprio);
2723 ioc->ioprio_changed = 0;
2726 static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2727 pid_t pid, bool is_sync)
2729 RB_CLEAR_NODE(&cfqq->rb_node);
2730 RB_CLEAR_NODE(&cfqq->p_node);
2731 INIT_LIST_HEAD(&cfqq->fifo);
2733 atomic_set(&cfqq->ref, 0);
2736 cfq_mark_cfqq_prio_changed(cfqq);
2739 if (!cfq_class_idle(cfqq))
2740 cfq_mark_cfqq_idle_window(cfqq);
2741 cfq_mark_cfqq_sync(cfqq);
2746 #ifdef CONFIG_CFQ_GROUP_IOSCHED
2747 static void changed_cgroup(struct io_context *ioc, struct cfq_io_context *cic)
2749 struct cfq_queue *sync_cfqq = cic_to_cfqq(cic, 1);
2750 struct cfq_data *cfqd = cic->key;
2751 unsigned long flags;
2752 struct request_queue *q;
2754 if (unlikely(!cfqd))
2759 spin_lock_irqsave(q->queue_lock, flags);
2763 * Drop reference to sync queue. A new sync queue will be
2764 * assigned in new group upon arrival of a fresh request.
2766 cfq_log_cfqq(cfqd, sync_cfqq, "changed cgroup");
2767 cic_set_cfqq(cic, NULL, 1);
2768 cfq_put_queue(sync_cfqq);
2771 spin_unlock_irqrestore(q->queue_lock, flags);
2774 static void cfq_ioc_set_cgroup(struct io_context *ioc)
2776 call_for_each_cic(ioc, changed_cgroup);
2777 ioc->cgroup_changed = 0;
2779 #endif /* CONFIG_CFQ_GROUP_IOSCHED */
2781 static struct cfq_queue *
2782 cfq_find_alloc_queue(struct cfq_data *cfqd, bool is_sync,
2783 struct io_context *ioc, gfp_t gfp_mask)
2785 struct cfq_queue *cfqq, *new_cfqq = NULL;
2786 struct cfq_io_context *cic;
2787 struct cfq_group *cfqg;
2790 cfqg = cfq_get_cfqg(cfqd, 1);
2791 cic = cfq_cic_lookup(cfqd, ioc);
2792 /* cic always exists here */
2793 cfqq = cic_to_cfqq(cic, is_sync);
2796 * Always try a new alloc if we fell back to the OOM cfqq
2797 * originally, since it should just be a temporary situation.
2799 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
2804 } else if (gfp_mask & __GFP_WAIT) {
2805 spin_unlock_irq(cfqd->queue->queue_lock);
2806 new_cfqq = kmem_cache_alloc_node(cfq_pool,
2807 gfp_mask | __GFP_ZERO,
2809 spin_lock_irq(cfqd->queue->queue_lock);
2813 cfqq = kmem_cache_alloc_node(cfq_pool,
2814 gfp_mask | __GFP_ZERO,
2819 cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
2820 cfq_init_prio_data(cfqq, ioc);
2821 cfq_link_cfqq_cfqg(cfqq, cfqg);
2822 cfq_log_cfqq(cfqd, cfqq, "alloced");
2824 cfqq = &cfqd->oom_cfqq;
2828 kmem_cache_free(cfq_pool, new_cfqq);
2833 static struct cfq_queue **
2834 cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
2836 switch (ioprio_class) {
2837 case IOPRIO_CLASS_RT:
2838 return &cfqd->async_cfqq[0][ioprio];
2839 case IOPRIO_CLASS_BE:
2840 return &cfqd->async_cfqq[1][ioprio];
2841 case IOPRIO_CLASS_IDLE:
2842 return &cfqd->async_idle_cfqq;
2848 static struct cfq_queue *
2849 cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct io_context *ioc,
2852 const int ioprio = task_ioprio(ioc);
2853 const int ioprio_class = task_ioprio_class(ioc);
2854 struct cfq_queue **async_cfqq = NULL;
2855 struct cfq_queue *cfqq = NULL;
2858 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
2863 cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask);
2866 * pin the queue now that it's allocated, scheduler exit will prune it
2868 if (!is_sync && !(*async_cfqq)) {
2869 atomic_inc(&cfqq->ref);
2873 atomic_inc(&cfqq->ref);
2878 * We drop cfq io contexts lazily, so we may find a dead one.
2881 cfq_drop_dead_cic(struct cfq_data *cfqd, struct io_context *ioc,
2882 struct cfq_io_context *cic)
2884 unsigned long flags;
2886 WARN_ON(!list_empty(&cic->queue_list));
2888 spin_lock_irqsave(&ioc->lock, flags);
2890 BUG_ON(ioc->ioc_data == cic);
2892 radix_tree_delete(&ioc->radix_root, (unsigned long) cfqd);
2893 hlist_del_rcu(&cic->cic_list);
2894 spin_unlock_irqrestore(&ioc->lock, flags);
2899 static struct cfq_io_context *
2900 cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc)
2902 struct cfq_io_context *cic;
2903 unsigned long flags;
2912 * we maintain a last-hit cache, to avoid browsing over the tree
2914 cic = rcu_dereference(ioc->ioc_data);
2915 if (cic && cic->key == cfqd) {
2921 cic = radix_tree_lookup(&ioc->radix_root, (unsigned long) cfqd);
2925 /* ->key must be copied to avoid race with cfq_exit_queue() */
2928 cfq_drop_dead_cic(cfqd, ioc, cic);
2933 spin_lock_irqsave(&ioc->lock, flags);
2934 rcu_assign_pointer(ioc->ioc_data, cic);
2935 spin_unlock_irqrestore(&ioc->lock, flags);
2943 * Add cic into ioc, using cfqd as the search key. This enables us to lookup
2944 * the process specific cfq io context when entered from the block layer.
2945 * Also adds the cic to a per-cfqd list, used when this queue is removed.
2947 static int cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
2948 struct cfq_io_context *cic, gfp_t gfp_mask)
2950 unsigned long flags;
2953 ret = radix_tree_preload(gfp_mask);
2958 spin_lock_irqsave(&ioc->lock, flags);
2959 ret = radix_tree_insert(&ioc->radix_root,
2960 (unsigned long) cfqd, cic);
2962 hlist_add_head_rcu(&cic->cic_list, &ioc->cic_list);
2963 spin_unlock_irqrestore(&ioc->lock, flags);
2965 radix_tree_preload_end();
2968 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2969 list_add(&cic->queue_list, &cfqd->cic_list);
2970 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2975 printk(KERN_ERR "cfq: cic link failed!\n");
2981 * Setup general io context and cfq io context. There can be several cfq
2982 * io contexts per general io context, if this process is doing io to more
2983 * than one device managed by cfq.
2985 static struct cfq_io_context *
2986 cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
2988 struct io_context *ioc = NULL;
2989 struct cfq_io_context *cic;
2991 might_sleep_if(gfp_mask & __GFP_WAIT);
2993 ioc = get_io_context(gfp_mask, cfqd->queue->node);
2997 cic = cfq_cic_lookup(cfqd, ioc);
3001 cic = cfq_alloc_io_context(cfqd, gfp_mask);
3005 if (cfq_cic_link(cfqd, ioc, cic, gfp_mask))
3009 smp_read_barrier_depends();
3010 if (unlikely(ioc->ioprio_changed))
3011 cfq_ioc_set_ioprio(ioc);
3013 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3014 if (unlikely(ioc->cgroup_changed))
3015 cfq_ioc_set_cgroup(ioc);
3021 put_io_context(ioc);
3026 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
3028 unsigned long elapsed = jiffies - cic->last_end_request;
3029 unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
3031 cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
3032 cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
3033 cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
3037 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3041 sector_t n_sec = blk_rq_sectors(rq);
3042 if (cfqq->last_request_pos) {
3043 if (cfqq->last_request_pos < blk_rq_pos(rq))
3044 sdist = blk_rq_pos(rq) - cfqq->last_request_pos;
3046 sdist = cfqq->last_request_pos - blk_rq_pos(rq);
3049 cfqq->seek_history <<= 1;
3050 if (blk_queue_nonrot(cfqd->queue))
3051 cfqq->seek_history |= (n_sec < CFQQ_SECT_THR_NONROT);
3053 cfqq->seek_history |= (sdist > CFQQ_SEEK_THR);
3057 * Disable idle window if the process thinks too long or seeks so much that
3061 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3062 struct cfq_io_context *cic)
3064 int old_idle, enable_idle;
3067 * Don't idle for async or idle io prio class
3069 if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
3072 enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
3074 if (cfqq->queued[0] + cfqq->queued[1] >= 4)
3075 cfq_mark_cfqq_deep(cfqq);
3077 if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle ||
3078 (!cfq_cfqq_deep(cfqq) && CFQQ_SEEKY(cfqq)))
3080 else if (sample_valid(cic->ttime_samples)) {
3081 if (cic->ttime_mean > cfqd->cfq_slice_idle)
3087 if (old_idle != enable_idle) {
3088 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
3090 cfq_mark_cfqq_idle_window(cfqq);
3092 cfq_clear_cfqq_idle_window(cfqq);
3097 * Check if new_cfqq should preempt the currently active queue. Return 0 for
3098 * no or if we aren't sure, a 1 will cause a preempt.
3101 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
3104 struct cfq_queue *cfqq;
3106 cfqq = cfqd->active_queue;
3110 if (cfq_class_idle(new_cfqq))
3113 if (cfq_class_idle(cfqq))
3117 * Don't allow a non-RT request to preempt an ongoing RT cfqq timeslice.
3119 if (cfq_class_rt(cfqq) && !cfq_class_rt(new_cfqq))
3123 * if the new request is sync, but the currently running queue is
3124 * not, let the sync request have priority.
3126 if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
3129 if (new_cfqq->cfqg != cfqq->cfqg)
3132 if (cfq_slice_used(cfqq))
3135 /* Allow preemption only if we are idling on sync-noidle tree */
3136 if (cfqd->serving_type == SYNC_NOIDLE_WORKLOAD &&
3137 cfqq_type(new_cfqq) == SYNC_NOIDLE_WORKLOAD &&
3138 new_cfqq->service_tree->count == 2 &&
3139 RB_EMPTY_ROOT(&cfqq->sort_list))
3143 * So both queues are sync. Let the new request get disk time if
3144 * it's a metadata request and the current queue is doing regular IO.
3146 if (rq_is_meta(rq) && !cfqq->meta_pending)
3150 * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
3152 if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
3155 if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
3159 * if this request is as-good as one we would expect from the
3160 * current cfqq, let it preempt
3162 if (cfq_rq_close(cfqd, cfqq, rq))
3169 * cfqq preempts the active queue. if we allowed preempt with no slice left,
3170 * let it have half of its nominal slice.
3172 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3174 cfq_log_cfqq(cfqd, cfqq, "preempt");
3175 cfq_slice_expired(cfqd, 1, false);
3178 * Put the new queue at the front of the of the current list,
3179 * so we know that it will be selected next.
3181 BUG_ON(!cfq_cfqq_on_rr(cfqq));
3183 cfq_service_tree_add(cfqd, cfqq, 1);
3185 cfqq->slice_end = 0;
3186 cfq_mark_cfqq_slice_new(cfqq);
3190 * Called when a new fs request (rq) is added (to cfqq). Check if there's
3191 * something we should do about it
3194 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3197 struct cfq_io_context *cic = RQ_CIC(rq);
3201 cfqq->meta_pending++;
3203 cfq_update_io_thinktime(cfqd, cic);
3204 cfq_update_io_seektime(cfqd, cfqq, rq);
3205 cfq_update_idle_window(cfqd, cfqq, cic);
3207 cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
3209 if (cfqq == cfqd->active_queue) {
3211 * Remember that we saw a request from this process, but
3212 * don't start queuing just yet. Otherwise we risk seeing lots
3213 * of tiny requests, because we disrupt the normal plugging
3214 * and merging. If the request is already larger than a single
3215 * page, let it rip immediately. For that case we assume that
3216 * merging is already done. Ditto for a busy system that
3217 * has other work pending, don't risk delaying until the
3218 * idle timer unplug to continue working.
3220 if (cfq_cfqq_wait_request(cfqq)) {
3221 if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
3222 cfqd->busy_queues > 1) {
3223 cfq_del_timer(cfqd, cfqq);
3224 cfq_clear_cfqq_wait_request(cfqq);
3225 __blk_run_queue(cfqd->queue);
3227 blkiocg_update_idle_time_stats(
3229 cfq_mark_cfqq_must_dispatch(cfqq);
3232 } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
3234 * not the active queue - expire current slice if it is
3235 * idle and has expired it's mean thinktime or this new queue
3236 * has some old slice time left and is of higher priority or
3237 * this new queue is RT and the current one is BE
3239 cfq_preempt_queue(cfqd, cfqq);
3240 __blk_run_queue(cfqd->queue);
3244 static void cfq_insert_request(struct request_queue *q, struct request *rq)
3246 struct cfq_data *cfqd = q->elevator->elevator_data;
3247 struct cfq_queue *cfqq = RQ_CFQQ(rq);
3249 cfq_log_cfqq(cfqd, cfqq, "insert_request");
3250 cfq_init_prio_data(cfqq, RQ_CIC(rq)->ioc);
3252 rq_set_fifo_time(rq, jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)]);
3253 list_add_tail(&rq->queuelist, &cfqq->fifo);
3255 blkiocg_update_io_add_stats(&(RQ_CFQG(rq))->blkg,
3256 &cfqd->serving_group->blkg, rq_data_dir(rq),
3258 cfq_rq_enqueued(cfqd, cfqq, rq);
3262 * Update hw_tag based on peak queue depth over 50 samples under
3265 static void cfq_update_hw_tag(struct cfq_data *cfqd)
3267 struct cfq_queue *cfqq = cfqd->active_queue;
3269 if (cfqd->rq_in_driver > cfqd->hw_tag_est_depth)
3270 cfqd->hw_tag_est_depth = cfqd->rq_in_driver;
3272 if (cfqd->hw_tag == 1)
3275 if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
3276 cfqd->rq_in_driver <= CFQ_HW_QUEUE_MIN)
3280 * If active queue hasn't enough requests and can idle, cfq might not
3281 * dispatch sufficient requests to hardware. Don't zero hw_tag in this
3284 if (cfqq && cfq_cfqq_idle_window(cfqq) &&
3285 cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] <
3286 CFQ_HW_QUEUE_MIN && cfqd->rq_in_driver < CFQ_HW_QUEUE_MIN)
3289 if (cfqd->hw_tag_samples++ < 50)
3292 if (cfqd->hw_tag_est_depth >= CFQ_HW_QUEUE_MIN)
3298 static bool cfq_should_wait_busy(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3300 struct cfq_io_context *cic = cfqd->active_cic;
3302 /* If there are other queues in the group, don't wait */
3303 if (cfqq->cfqg->nr_cfqq > 1)
3306 if (cfq_slice_used(cfqq))
3309 /* if slice left is less than think time, wait busy */
3310 if (cic && sample_valid(cic->ttime_samples)
3311 && (cfqq->slice_end - jiffies < cic->ttime_mean))
3315 * If think times is less than a jiffy than ttime_mean=0 and above
3316 * will not be true. It might happen that slice has not expired yet
3317 * but will expire soon (4-5 ns) during select_queue(). To cover the
3318 * case where think time is less than a jiffy, mark the queue wait
3319 * busy if only 1 jiffy is left in the slice.
3321 if (cfqq->slice_end - jiffies == 1)
3327 static void cfq_completed_request(struct request_queue *q, struct request *rq)
3329 struct cfq_queue *cfqq = RQ_CFQQ(rq);
3330 struct cfq_data *cfqd = cfqq->cfqd;
3331 const int sync = rq_is_sync(rq);
3335 cfq_log_cfqq(cfqd, cfqq, "complete rqnoidle %d", !!rq_noidle(rq));
3337 cfq_update_hw_tag(cfqd);
3339 WARN_ON(!cfqd->rq_in_driver);
3340 WARN_ON(!cfqq->dispatched);
3341 cfqd->rq_in_driver--;
3343 blkiocg_update_completion_stats(&cfqq->cfqg->blkg, rq_start_time_ns(rq),
3344 rq_io_start_time_ns(rq), rq_data_dir(rq),
3347 cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]--;
3350 RQ_CIC(rq)->last_end_request = now;
3351 if (!time_after(rq->start_time + cfqd->cfq_fifo_expire[1], now))
3352 cfqd->last_delayed_sync = now;
3356 * If this is the active queue, check if it needs to be expired,
3357 * or if we want to idle in case it has no pending requests.
3359 if (cfqd->active_queue == cfqq) {
3360 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
3362 if (cfq_cfqq_slice_new(cfqq)) {
3363 cfq_set_prio_slice(cfqd, cfqq);
3364 cfq_clear_cfqq_slice_new(cfqq);
3368 * Should we wait for next request to come in before we expire
3371 if (cfq_should_wait_busy(cfqd, cfqq)) {
3372 cfqq->slice_end = jiffies + cfqd->cfq_slice_idle;
3373 cfq_mark_cfqq_wait_busy(cfqq);
3374 cfq_log_cfqq(cfqd, cfqq, "will busy wait");
3378 * Idling is not enabled on:
3380 * - idle-priority queues
3382 * - queues with still some requests queued
3383 * - when there is a close cooperator
3385 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
3386 cfq_slice_expired(cfqd, 1, false);
3387 else if (sync && cfqq_empty &&
3388 !cfq_close_cooperator(cfqd, cfqq)) {
3389 cfqd->noidle_tree_requires_idle |= !rq_noidle(rq);
3391 * Idling is enabled for SYNC_WORKLOAD.
3392 * SYNC_NOIDLE_WORKLOAD idles at the end of the tree
3393 * only if we processed at least one !rq_noidle request
3395 if (cfqd->serving_type == SYNC_WORKLOAD
3396 || cfqd->noidle_tree_requires_idle
3397 || cfqq->cfqg->nr_cfqq == 1)
3398 cfq_arm_slice_timer(cfqd);
3402 if (!cfqd->rq_in_driver)
3403 cfq_schedule_dispatch(cfqd);
3407 * we temporarily boost lower priority queues if they are holding fs exclusive
3408 * resources. they are boosted to normal prio (CLASS_BE/4)
3410 static void cfq_prio_boost(struct cfq_queue *cfqq)
3412 if (has_fs_excl()) {
3414 * boost idle prio on transactions that would lock out other
3415 * users of the filesystem
3417 if (cfq_class_idle(cfqq))
3418 cfqq->ioprio_class = IOPRIO_CLASS_BE;
3419 if (cfqq->ioprio > IOPRIO_NORM)
3420 cfqq->ioprio = IOPRIO_NORM;
3423 * unboost the queue (if needed)
3425 cfqq->ioprio_class = cfqq->org_ioprio_class;
3426 cfqq->ioprio = cfqq->org_ioprio;
3430 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
3432 if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
3433 cfq_mark_cfqq_must_alloc_slice(cfqq);
3434 return ELV_MQUEUE_MUST;
3437 return ELV_MQUEUE_MAY;
3440 static int cfq_may_queue(struct request_queue *q, int rw)
3442 struct cfq_data *cfqd = q->elevator->elevator_data;
3443 struct task_struct *tsk = current;
3444 struct cfq_io_context *cic;
3445 struct cfq_queue *cfqq;
3448 * don't force setup of a queue from here, as a call to may_queue
3449 * does not necessarily imply that a request actually will be queued.
3450 * so just lookup a possibly existing queue, or return 'may queue'
3453 cic = cfq_cic_lookup(cfqd, tsk->io_context);
3455 return ELV_MQUEUE_MAY;
3457 cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
3459 cfq_init_prio_data(cfqq, cic->ioc);
3460 cfq_prio_boost(cfqq);
3462 return __cfq_may_queue(cfqq);
3465 return ELV_MQUEUE_MAY;
3469 * queue lock held here
3471 static void cfq_put_request(struct request *rq)
3473 struct cfq_queue *cfqq = RQ_CFQQ(rq);
3476 const int rw = rq_data_dir(rq);
3478 BUG_ON(!cfqq->allocated[rw]);
3479 cfqq->allocated[rw]--;
3481 put_io_context(RQ_CIC(rq)->ioc);
3483 rq->elevator_private = NULL;
3484 rq->elevator_private2 = NULL;
3486 /* Put down rq reference on cfqg */
3487 cfq_put_cfqg(RQ_CFQG(rq));
3488 rq->elevator_private3 = NULL;
3490 cfq_put_queue(cfqq);
3494 static struct cfq_queue *
3495 cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_context *cic,
3496 struct cfq_queue *cfqq)
3498 cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq);
3499 cic_set_cfqq(cic, cfqq->new_cfqq, 1);
3500 cfq_mark_cfqq_coop(cfqq->new_cfqq);
3501 cfq_put_queue(cfqq);
3502 return cic_to_cfqq(cic, 1);
3506 * Returns NULL if a new cfqq should be allocated, or the old cfqq if this
3507 * was the last process referring to said cfqq.
3509 static struct cfq_queue *
3510 split_cfqq(struct cfq_io_context *cic, struct cfq_queue *cfqq)
3512 if (cfqq_process_refs(cfqq) == 1) {
3513 cfqq->pid = current->pid;
3514 cfq_clear_cfqq_coop(cfqq);
3515 cfq_clear_cfqq_split_coop(cfqq);
3519 cic_set_cfqq(cic, NULL, 1);
3520 cfq_put_queue(cfqq);
3524 * Allocate cfq data structures associated with this request.
3527 cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
3529 struct cfq_data *cfqd = q->elevator->elevator_data;
3530 struct cfq_io_context *cic;
3531 const int rw = rq_data_dir(rq);
3532 const bool is_sync = rq_is_sync(rq);
3533 struct cfq_queue *cfqq;
3534 unsigned long flags;
3536 might_sleep_if(gfp_mask & __GFP_WAIT);
3538 cic = cfq_get_io_context(cfqd, gfp_mask);
3540 spin_lock_irqsave(q->queue_lock, flags);
3546 cfqq = cic_to_cfqq(cic, is_sync);
3547 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
3548 cfqq = cfq_get_queue(cfqd, is_sync, cic->ioc, gfp_mask);
3549 cic_set_cfqq(cic, cfqq, is_sync);
3552 * If the queue was seeky for too long, break it apart.
3554 if (cfq_cfqq_coop(cfqq) && cfq_cfqq_split_coop(cfqq)) {
3555 cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq");
3556 cfqq = split_cfqq(cic, cfqq);
3562 * Check to see if this queue is scheduled to merge with
3563 * another, closely cooperating queue. The merging of
3564 * queues happens here as it must be done in process context.
3565 * The reference on new_cfqq was taken in merge_cfqqs.
3568 cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq);
3571 cfqq->allocated[rw]++;
3572 atomic_inc(&cfqq->ref);
3574 spin_unlock_irqrestore(q->queue_lock, flags);
3576 rq->elevator_private = cic;
3577 rq->elevator_private2 = cfqq;
3578 rq->elevator_private3 = cfq_ref_get_cfqg(cfqq->cfqg);
3583 put_io_context(cic->ioc);
3585 cfq_schedule_dispatch(cfqd);
3586 spin_unlock_irqrestore(q->queue_lock, flags);
3587 cfq_log(cfqd, "set_request fail");
3591 static void cfq_kick_queue(struct work_struct *work)
3593 struct cfq_data *cfqd =
3594 container_of(work, struct cfq_data, unplug_work);
3595 struct request_queue *q = cfqd->queue;
3597 spin_lock_irq(q->queue_lock);
3598 __blk_run_queue(cfqd->queue);
3599 spin_unlock_irq(q->queue_lock);
3603 * Timer running if the active_queue is currently idling inside its time slice
3605 static void cfq_idle_slice_timer(unsigned long data)
3607 struct cfq_data *cfqd = (struct cfq_data *) data;
3608 struct cfq_queue *cfqq;
3609 unsigned long flags;
3612 cfq_log(cfqd, "idle timer fired");
3614 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
3616 cfqq = cfqd->active_queue;
3621 * We saw a request before the queue expired, let it through
3623 if (cfq_cfqq_must_dispatch(cfqq))
3629 if (cfq_slice_used(cfqq))
3633 * only expire and reinvoke request handler, if there are
3634 * other queues with pending requests
3636 if (!cfqd->busy_queues)
3640 * not expired and it has a request pending, let it dispatch
3642 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3646 * Queue depth flag is reset only when the idle didn't succeed
3648 cfq_clear_cfqq_deep(cfqq);
3651 cfq_slice_expired(cfqd, timed_out, false);
3653 cfq_schedule_dispatch(cfqd);
3655 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
3658 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
3660 del_timer_sync(&cfqd->idle_slice_timer);
3661 cancel_work_sync(&cfqd->unplug_work);
3664 static void cfq_put_async_queues(struct cfq_data *cfqd)
3668 for (i = 0; i < IOPRIO_BE_NR; i++) {
3669 if (cfqd->async_cfqq[0][i])
3670 cfq_put_queue(cfqd->async_cfqq[0][i]);
3671 if (cfqd->async_cfqq[1][i])
3672 cfq_put_queue(cfqd->async_cfqq[1][i]);
3675 if (cfqd->async_idle_cfqq)
3676 cfq_put_queue(cfqd->async_idle_cfqq);
3679 static void cfq_cfqd_free(struct rcu_head *head)
3681 kfree(container_of(head, struct cfq_data, rcu));
3684 static void cfq_exit_queue(struct elevator_queue *e)
3686 struct cfq_data *cfqd = e->elevator_data;
3687 struct request_queue *q = cfqd->queue;
3689 cfq_shutdown_timer_wq(cfqd);
3691 spin_lock_irq(q->queue_lock);
3693 if (cfqd->active_queue)
3694 __cfq_slice_expired(cfqd, cfqd->active_queue, 0, false);
3696 while (!list_empty(&cfqd->cic_list)) {
3697 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
3698 struct cfq_io_context,
3701 __cfq_exit_single_io_context(cfqd, cic);
3704 cfq_put_async_queues(cfqd);
3705 cfq_release_cfq_groups(cfqd);
3706 blkiocg_del_blkio_group(&cfqd->root_group.blkg);
3708 spin_unlock_irq(q->queue_lock);
3710 cfq_shutdown_timer_wq(cfqd);
3712 /* Wait for cfqg->blkg->key accessors to exit their grace periods. */
3713 call_rcu(&cfqd->rcu, cfq_cfqd_free);
3716 static void *cfq_init_queue(struct request_queue *q)
3718 struct cfq_data *cfqd;
3720 struct cfq_group *cfqg;
3721 struct cfq_rb_root *st;
3723 cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
3727 /* Init root service tree */
3728 cfqd->grp_service_tree = CFQ_RB_ROOT;
3730 /* Init root group */
3731 cfqg = &cfqd->root_group;
3732 for_each_cfqg_st(cfqg, i, j, st)
3734 RB_CLEAR_NODE(&cfqg->rb_node);
3736 /* Give preference to root group over other groups */
3737 cfqg->weight = 2*BLKIO_WEIGHT_DEFAULT;
3739 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3741 * Take a reference to root group which we never drop. This is just
3742 * to make sure that cfq_put_cfqg() does not try to kfree root group
3744 atomic_set(&cfqg->ref, 1);
3745 blkiocg_add_blkio_group(&blkio_root_cgroup, &cfqg->blkg, (void *)cfqd,
3749 * Not strictly needed (since RB_ROOT just clears the node and we
3750 * zeroed cfqd on alloc), but better be safe in case someone decides
3751 * to add magic to the rb code
3753 for (i = 0; i < CFQ_PRIO_LISTS; i++)
3754 cfqd->prio_trees[i] = RB_ROOT;
3757 * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
3758 * Grab a permanent reference to it, so that the normal code flow
3759 * will not attempt to free it.
3761 cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
3762 atomic_inc(&cfqd->oom_cfqq.ref);
3763 cfq_link_cfqq_cfqg(&cfqd->oom_cfqq, &cfqd->root_group);
3765 INIT_LIST_HEAD(&cfqd->cic_list);
3769 init_timer(&cfqd->idle_slice_timer);
3770 cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
3771 cfqd->idle_slice_timer.data = (unsigned long) cfqd;
3773 INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
3775 cfqd->cfq_quantum = cfq_quantum;
3776 cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
3777 cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
3778 cfqd->cfq_back_max = cfq_back_max;
3779 cfqd->cfq_back_penalty = cfq_back_penalty;
3780 cfqd->cfq_slice[0] = cfq_slice_async;
3781 cfqd->cfq_slice[1] = cfq_slice_sync;
3782 cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
3783 cfqd->cfq_slice_idle = cfq_slice_idle;
3784 cfqd->cfq_latency = 1;
3785 cfqd->cfq_group_isolation = 0;
3788 * we optimistically start assuming sync ops weren't delayed in last
3789 * second, in order to have larger depth for async operations.
3791 cfqd->last_delayed_sync = jiffies - HZ;
3792 INIT_RCU_HEAD(&cfqd->rcu);
3796 static void cfq_slab_kill(void)
3799 * Caller already ensured that pending RCU callbacks are completed,
3800 * so we should have no busy allocations at this point.
3803 kmem_cache_destroy(cfq_pool);
3805 kmem_cache_destroy(cfq_ioc_pool);
3808 static int __init cfq_slab_setup(void)
3810 cfq_pool = KMEM_CACHE(cfq_queue, 0);
3814 cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0);
3825 * sysfs parts below -->
3828 cfq_var_show(unsigned int var, char *page)
3830 return sprintf(page, "%d\n", var);
3834 cfq_var_store(unsigned int *var, const char *page, size_t count)
3836 char *p = (char *) page;
3838 *var = simple_strtoul(p, &p, 10);
3842 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
3843 static ssize_t __FUNC(struct elevator_queue *e, char *page) \
3845 struct cfq_data *cfqd = e->elevator_data; \
3846 unsigned int __data = __VAR; \
3848 __data = jiffies_to_msecs(__data); \
3849 return cfq_var_show(__data, (page)); \
3851 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
3852 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
3853 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
3854 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
3855 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
3856 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
3857 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
3858 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
3859 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
3860 SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
3861 SHOW_FUNCTION(cfq_group_isolation_show, cfqd->cfq_group_isolation, 0);
3862 #undef SHOW_FUNCTION
3864 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
3865 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
3867 struct cfq_data *cfqd = e->elevator_data; \
3868 unsigned int __data; \
3869 int ret = cfq_var_store(&__data, (page), count); \
3870 if (__data < (MIN)) \
3872 else if (__data > (MAX)) \
3875 *(__PTR) = msecs_to_jiffies(__data); \
3877 *(__PTR) = __data; \
3880 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
3881 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
3883 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
3885 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
3886 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
3888 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
3889 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
3890 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
3891 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
3893 STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
3894 STORE_FUNCTION(cfq_group_isolation_store, &cfqd->cfq_group_isolation, 0, 1, 0);
3895 #undef STORE_FUNCTION
3897 #define CFQ_ATTR(name) \
3898 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
3900 static struct elv_fs_entry cfq_attrs[] = {
3902 CFQ_ATTR(fifo_expire_sync),
3903 CFQ_ATTR(fifo_expire_async),
3904 CFQ_ATTR(back_seek_max),
3905 CFQ_ATTR(back_seek_penalty),
3906 CFQ_ATTR(slice_sync),
3907 CFQ_ATTR(slice_async),
3908 CFQ_ATTR(slice_async_rq),
3909 CFQ_ATTR(slice_idle),
3910 CFQ_ATTR(low_latency),
3911 CFQ_ATTR(group_isolation),
3915 static struct elevator_type iosched_cfq = {
3917 .elevator_merge_fn = cfq_merge,
3918 .elevator_merged_fn = cfq_merged_request,
3919 .elevator_merge_req_fn = cfq_merged_requests,
3920 .elevator_allow_merge_fn = cfq_allow_merge,
3921 .elevator_bio_merged_fn = cfq_bio_merged,
3922 .elevator_dispatch_fn = cfq_dispatch_requests,
3923 .elevator_add_req_fn = cfq_insert_request,
3924 .elevator_activate_req_fn = cfq_activate_request,
3925 .elevator_deactivate_req_fn = cfq_deactivate_request,
3926 .elevator_queue_empty_fn = cfq_queue_empty,
3927 .elevator_completed_req_fn = cfq_completed_request,
3928 .elevator_former_req_fn = elv_rb_former_request,
3929 .elevator_latter_req_fn = elv_rb_latter_request,
3930 .elevator_set_req_fn = cfq_set_request,
3931 .elevator_put_req_fn = cfq_put_request,
3932 .elevator_may_queue_fn = cfq_may_queue,
3933 .elevator_init_fn = cfq_init_queue,
3934 .elevator_exit_fn = cfq_exit_queue,
3935 .trim = cfq_free_io_context,
3937 .elevator_attrs = cfq_attrs,
3938 .elevator_name = "cfq",
3939 .elevator_owner = THIS_MODULE,
3942 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3943 static struct blkio_policy_type blkio_policy_cfq = {
3945 .blkio_unlink_group_fn = cfq_unlink_blkio_group,
3946 .blkio_update_group_weight_fn = cfq_update_blkio_group_weight,
3950 static struct blkio_policy_type blkio_policy_cfq;
3953 static int __init cfq_init(void)
3956 * could be 0 on HZ < 1000 setups
3958 if (!cfq_slice_async)
3959 cfq_slice_async = 1;
3960 if (!cfq_slice_idle)
3963 if (cfq_slab_setup())
3966 elv_register(&iosched_cfq);
3967 blkio_policy_register(&blkio_policy_cfq);
3972 static void __exit cfq_exit(void)
3974 DECLARE_COMPLETION_ONSTACK(all_gone);
3975 blkio_policy_unregister(&blkio_policy_cfq);
3976 elv_unregister(&iosched_cfq);
3977 ioc_gone = &all_gone;
3978 /* ioc_gone's update must be visible before reading ioc_count */
3982 * this also protects us from entering cfq_slab_kill() with
3983 * pending RCU callbacks
3985 if (elv_ioc_count_read(cfq_ioc_count))
3986 wait_for_completion(&all_gone);
3990 module_init(cfq_init);
3991 module_exit(cfq_exit);
3993 MODULE_AUTHOR("Jens Axboe");
3994 MODULE_LICENSE("GPL");
3995 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");