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>
20 static struct blkio_policy_type blkio_policy_cfq;
25 /* max queue in one round of service */
26 static const int cfq_quantum = 8;
27 static const int cfq_fifo_expire[2] = { HZ / 4, HZ / 8 };
28 /* maximum backwards seek, in KiB */
29 static const int cfq_back_max = 16 * 1024;
30 /* penalty of a backwards seek */
31 static const int cfq_back_penalty = 2;
32 static const int cfq_slice_sync = HZ / 10;
33 static int cfq_slice_async = HZ / 25;
34 static const int cfq_slice_async_rq = 2;
35 static int cfq_slice_idle = HZ / 125;
36 static int cfq_group_idle = HZ / 125;
37 static const int cfq_target_latency = HZ * 3/10; /* 300 ms */
38 static const int cfq_hist_divisor = 4;
41 * offset from end of service tree
43 #define CFQ_IDLE_DELAY (HZ / 5)
46 * below this threshold, we consider thinktime immediate
48 #define CFQ_MIN_TT (2)
50 #define CFQ_SLICE_SCALE (5)
51 #define CFQ_HW_QUEUE_MIN (5)
52 #define CFQ_SERVICE_SHIFT 12
54 #define CFQQ_SEEK_THR (sector_t)(8 * 100)
55 #define CFQQ_CLOSE_THR (sector_t)(8 * 1024)
56 #define CFQQ_SECT_THR_NONROT (sector_t)(2 * 32)
57 #define CFQQ_SEEKY(cfqq) (hweight32(cfqq->seek_history) > 32/8)
59 #define RQ_CIC(rq) icq_to_cic((rq)->elv.icq)
60 #define RQ_CFQQ(rq) (struct cfq_queue *) ((rq)->elv.priv[0])
61 #define RQ_CFQG(rq) (struct cfq_group *) ((rq)->elv.priv[1])
63 static struct kmem_cache *cfq_pool;
65 #define CFQ_PRIO_LISTS IOPRIO_BE_NR
66 #define cfq_class_idle(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
67 #define cfq_class_rt(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
69 #define sample_valid(samples) ((samples) > 80)
70 #define rb_entry_cfqg(node) rb_entry((node), struct cfq_group, rb_node)
73 unsigned long last_end_request;
75 unsigned long ttime_total;
76 unsigned long ttime_samples;
77 unsigned long ttime_mean;
81 * Most of our rbtree usage is for sorting with min extraction, so
82 * if we cache the leftmost node we don't have to walk down the tree
83 * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should
84 * move this into the elevator for the rq sorting as well.
90 unsigned total_weight;
92 struct cfq_ttime ttime;
94 #define CFQ_RB_ROOT (struct cfq_rb_root) { .rb = RB_ROOT, \
95 .ttime = {.last_end_request = jiffies,},}
98 * Per process-grouping structure
101 /* reference count */
103 /* various state flags, see below */
105 /* parent cfq_data */
106 struct cfq_data *cfqd;
107 /* service_tree member */
108 struct rb_node rb_node;
109 /* service_tree key */
110 unsigned long rb_key;
111 /* prio tree member */
112 struct rb_node p_node;
113 /* prio tree root we belong to, if any */
114 struct rb_root *p_root;
115 /* sorted list of pending requests */
116 struct rb_root sort_list;
117 /* if fifo isn't expired, next request to serve */
118 struct request *next_rq;
119 /* requests queued in sort_list */
121 /* currently allocated requests */
123 /* fifo list of requests in sort_list */
124 struct list_head fifo;
126 /* time when queue got scheduled in to dispatch first request. */
127 unsigned long dispatch_start;
128 unsigned int allocated_slice;
129 unsigned int slice_dispatch;
130 /* time when first request from queue completed and slice started. */
131 unsigned long slice_start;
132 unsigned long slice_end;
135 /* pending priority requests */
137 /* number of requests that are on the dispatch list or inside driver */
140 /* io prio of this group */
141 unsigned short ioprio, org_ioprio;
142 unsigned short ioprio_class;
147 sector_t last_request_pos;
149 struct cfq_rb_root *service_tree;
150 struct cfq_queue *new_cfqq;
151 struct cfq_group *cfqg;
152 /* Number of sectors dispatched from queue in single dispatch round */
153 unsigned long nr_sectors;
157 * First index in the service_trees.
158 * IDLE is handled separately, so it has negative index
168 * Second index in the service_trees.
172 SYNC_NOIDLE_WORKLOAD = 1,
176 /* This is per cgroup per device grouping structure */
178 /* group service_tree member */
179 struct rb_node rb_node;
181 /* group service_tree key */
184 unsigned int new_weight;
187 /* number of cfqq currently on this group */
191 * Per group busy queues average. Useful for workload slice calc. We
192 * create the array for each prio class but at run time it is used
193 * only for RT and BE class and slot for IDLE class remains unused.
194 * This is primarily done to avoid confusion and a gcc warning.
196 unsigned int busy_queues_avg[CFQ_PRIO_NR];
198 * rr lists of queues with requests. We maintain service trees for
199 * RT and BE classes. These trees are subdivided in subclasses
200 * of SYNC, SYNC_NOIDLE and ASYNC based on workload type. For IDLE
201 * class there is no subclassification and all the cfq queues go on
202 * a single tree service_tree_idle.
203 * Counts are embedded in the cfq_rb_root
205 struct cfq_rb_root service_trees[2][3];
206 struct cfq_rb_root service_tree_idle;
208 unsigned long saved_workload_slice;
209 enum wl_type_t saved_workload;
210 enum wl_prio_t saved_serving_prio;
211 #ifdef CONFIG_CFQ_GROUP_IOSCHED
212 struct hlist_node cfqd_node;
215 /* number of requests that are on the dispatch list or inside driver */
217 struct cfq_ttime ttime;
221 struct io_cq icq; /* must be the first member */
222 struct cfq_queue *cfqq[2];
223 struct cfq_ttime ttime;
227 * Per block device queue structure
230 struct request_queue *queue;
231 /* Root service tree for cfq_groups */
232 struct cfq_rb_root grp_service_tree;
233 struct cfq_group *root_group;
236 * The priority currently being served
238 enum wl_prio_t serving_prio;
239 enum wl_type_t serving_type;
240 unsigned long workload_expires;
241 struct cfq_group *serving_group;
244 * Each priority tree is sorted by next_request position. These
245 * trees are used when determining if two or more queues are
246 * interleaving requests (see cfq_close_cooperator).
248 struct rb_root prio_trees[CFQ_PRIO_LISTS];
250 unsigned int busy_queues;
251 unsigned int busy_sync_queues;
257 * queue-depth detection
263 * -1 => indeterminate, (cfq will behave as if NCQ is present, to allow better detection)
264 * 1 => NCQ is present (hw_tag_est_depth is the estimated max depth)
267 int hw_tag_est_depth;
268 unsigned int hw_tag_samples;
271 * idle window management
273 struct timer_list idle_slice_timer;
274 struct work_struct unplug_work;
276 struct cfq_queue *active_queue;
277 struct cfq_io_cq *active_cic;
280 * async queue for each priority case
282 struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
283 struct cfq_queue *async_idle_cfqq;
285 sector_t last_position;
288 * tunables, see top of file
290 unsigned int cfq_quantum;
291 unsigned int cfq_fifo_expire[2];
292 unsigned int cfq_back_penalty;
293 unsigned int cfq_back_max;
294 unsigned int cfq_slice[2];
295 unsigned int cfq_slice_async_rq;
296 unsigned int cfq_slice_idle;
297 unsigned int cfq_group_idle;
298 unsigned int cfq_latency;
301 * Fallback dummy cfqq for extreme OOM conditions
303 struct cfq_queue oom_cfqq;
305 unsigned long last_delayed_sync;
307 /* List of cfq groups being managed on this device*/
308 struct hlist_head cfqg_list;
310 /* Number of groups which are on blkcg->blkg_list */
311 unsigned int nr_blkcg_linked_grps;
314 static inline struct cfq_group *blkg_to_cfqg(struct blkio_group *blkg)
316 return blkg_to_pdata(blkg, &blkio_policy_cfq);
319 static inline struct blkio_group *cfqg_to_blkg(struct cfq_group *cfqg)
321 return pdata_to_blkg(cfqg, &blkio_policy_cfq);
324 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd);
326 static struct cfq_rb_root *service_tree_for(struct cfq_group *cfqg,
333 if (prio == IDLE_WORKLOAD)
334 return &cfqg->service_tree_idle;
336 return &cfqg->service_trees[prio][type];
339 enum cfqq_state_flags {
340 CFQ_CFQQ_FLAG_on_rr = 0, /* on round-robin busy list */
341 CFQ_CFQQ_FLAG_wait_request, /* waiting for a request */
342 CFQ_CFQQ_FLAG_must_dispatch, /* must be allowed a dispatch */
343 CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
344 CFQ_CFQQ_FLAG_fifo_expire, /* FIFO checked in this slice */
345 CFQ_CFQQ_FLAG_idle_window, /* slice idling enabled */
346 CFQ_CFQQ_FLAG_prio_changed, /* task priority has changed */
347 CFQ_CFQQ_FLAG_slice_new, /* no requests dispatched in slice */
348 CFQ_CFQQ_FLAG_sync, /* synchronous queue */
349 CFQ_CFQQ_FLAG_coop, /* cfqq is shared */
350 CFQ_CFQQ_FLAG_split_coop, /* shared cfqq will be splitted */
351 CFQ_CFQQ_FLAG_deep, /* sync cfqq experienced large depth */
352 CFQ_CFQQ_FLAG_wait_busy, /* Waiting for next request */
355 #define CFQ_CFQQ_FNS(name) \
356 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \
358 (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name); \
360 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \
362 (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \
364 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \
366 return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \
370 CFQ_CFQQ_FNS(wait_request);
371 CFQ_CFQQ_FNS(must_dispatch);
372 CFQ_CFQQ_FNS(must_alloc_slice);
373 CFQ_CFQQ_FNS(fifo_expire);
374 CFQ_CFQQ_FNS(idle_window);
375 CFQ_CFQQ_FNS(prio_changed);
376 CFQ_CFQQ_FNS(slice_new);
379 CFQ_CFQQ_FNS(split_coop);
381 CFQ_CFQQ_FNS(wait_busy);
384 #ifdef CONFIG_CFQ_GROUP_IOSCHED
385 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
386 blk_add_trace_msg((cfqd)->queue, "cfq%d%c %s " fmt, (cfqq)->pid, \
387 cfq_cfqq_sync((cfqq)) ? 'S' : 'A', \
388 blkg_path(cfqg_to_blkg((cfqq)->cfqg)), ##args)
390 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...) \
391 blk_add_trace_msg((cfqd)->queue, "%s " fmt, \
392 blkg_path(cfqg_to_blkg((cfqg))), ##args) \
395 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
396 blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args)
397 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...) do {} while (0)
399 #define cfq_log(cfqd, fmt, args...) \
400 blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
402 /* Traverses through cfq group service trees */
403 #define for_each_cfqg_st(cfqg, i, j, st) \
404 for (i = 0; i <= IDLE_WORKLOAD; i++) \
405 for (j = 0, st = i < IDLE_WORKLOAD ? &cfqg->service_trees[i][j]\
406 : &cfqg->service_tree_idle; \
407 (i < IDLE_WORKLOAD && j <= SYNC_WORKLOAD) || \
408 (i == IDLE_WORKLOAD && j == 0); \
409 j++, st = i < IDLE_WORKLOAD ? \
410 &cfqg->service_trees[i][j]: NULL) \
412 static inline bool cfq_io_thinktime_big(struct cfq_data *cfqd,
413 struct cfq_ttime *ttime, bool group_idle)
416 if (!sample_valid(ttime->ttime_samples))
419 slice = cfqd->cfq_group_idle;
421 slice = cfqd->cfq_slice_idle;
422 return ttime->ttime_mean > slice;
425 static inline bool iops_mode(struct cfq_data *cfqd)
428 * If we are not idling on queues and it is a NCQ drive, parallel
429 * execution of requests is on and measuring time is not possible
430 * in most of the cases until and unless we drive shallower queue
431 * depths and that becomes a performance bottleneck. In such cases
432 * switch to start providing fairness in terms of number of IOs.
434 if (!cfqd->cfq_slice_idle && cfqd->hw_tag)
440 static inline enum wl_prio_t cfqq_prio(struct cfq_queue *cfqq)
442 if (cfq_class_idle(cfqq))
443 return IDLE_WORKLOAD;
444 if (cfq_class_rt(cfqq))
450 static enum wl_type_t cfqq_type(struct cfq_queue *cfqq)
452 if (!cfq_cfqq_sync(cfqq))
453 return ASYNC_WORKLOAD;
454 if (!cfq_cfqq_idle_window(cfqq))
455 return SYNC_NOIDLE_WORKLOAD;
456 return SYNC_WORKLOAD;
459 static inline int cfq_group_busy_queues_wl(enum wl_prio_t wl,
460 struct cfq_data *cfqd,
461 struct cfq_group *cfqg)
463 if (wl == IDLE_WORKLOAD)
464 return cfqg->service_tree_idle.count;
466 return cfqg->service_trees[wl][ASYNC_WORKLOAD].count
467 + cfqg->service_trees[wl][SYNC_NOIDLE_WORKLOAD].count
468 + cfqg->service_trees[wl][SYNC_WORKLOAD].count;
471 static inline int cfqg_busy_async_queues(struct cfq_data *cfqd,
472 struct cfq_group *cfqg)
474 return cfqg->service_trees[RT_WORKLOAD][ASYNC_WORKLOAD].count
475 + cfqg->service_trees[BE_WORKLOAD][ASYNC_WORKLOAD].count;
478 static void cfq_dispatch_insert(struct request_queue *, struct request *);
479 static struct cfq_queue *cfq_get_queue(struct cfq_data *, bool,
480 struct io_context *, gfp_t);
482 static inline struct cfq_io_cq *icq_to_cic(struct io_cq *icq)
484 /* cic->icq is the first member, %NULL will convert to %NULL */
485 return container_of(icq, struct cfq_io_cq, icq);
488 static inline struct cfq_io_cq *cfq_cic_lookup(struct cfq_data *cfqd,
489 struct io_context *ioc)
492 return icq_to_cic(ioc_lookup_icq(ioc, cfqd->queue));
496 static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_cq *cic, bool is_sync)
498 return cic->cfqq[is_sync];
501 static inline void cic_set_cfqq(struct cfq_io_cq *cic, struct cfq_queue *cfqq,
504 cic->cfqq[is_sync] = cfqq;
507 static inline struct cfq_data *cic_to_cfqd(struct cfq_io_cq *cic)
509 return cic->icq.q->elevator->elevator_data;
513 * We regard a request as SYNC, if it's either a read or has the SYNC bit
514 * set (in which case it could also be direct WRITE).
516 static inline bool cfq_bio_sync(struct bio *bio)
518 return bio_data_dir(bio) == READ || (bio->bi_rw & REQ_SYNC);
522 * scheduler run of queue, if there are requests pending and no one in the
523 * driver that will restart queueing
525 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
527 if (cfqd->busy_queues) {
528 cfq_log(cfqd, "schedule dispatch");
529 kblockd_schedule_work(cfqd->queue, &cfqd->unplug_work);
534 * Scale schedule slice based on io priority. Use the sync time slice only
535 * if a queue is marked sync and has sync io queued. A sync queue with async
536 * io only, should not get full sync slice length.
538 static inline int cfq_prio_slice(struct cfq_data *cfqd, bool sync,
541 const int base_slice = cfqd->cfq_slice[sync];
543 WARN_ON(prio >= IOPRIO_BE_NR);
545 return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
549 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
551 return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
554 static inline u64 cfq_scale_slice(unsigned long delta, struct cfq_group *cfqg)
556 u64 d = delta << CFQ_SERVICE_SHIFT;
558 d = d * BLKIO_WEIGHT_DEFAULT;
559 do_div(d, cfqg->weight);
563 static inline u64 max_vdisktime(u64 min_vdisktime, u64 vdisktime)
565 s64 delta = (s64)(vdisktime - min_vdisktime);
567 min_vdisktime = vdisktime;
569 return min_vdisktime;
572 static inline u64 min_vdisktime(u64 min_vdisktime, u64 vdisktime)
574 s64 delta = (s64)(vdisktime - min_vdisktime);
576 min_vdisktime = vdisktime;
578 return min_vdisktime;
581 static void update_min_vdisktime(struct cfq_rb_root *st)
583 struct cfq_group *cfqg;
586 cfqg = rb_entry_cfqg(st->left);
587 st->min_vdisktime = max_vdisktime(st->min_vdisktime,
593 * get averaged number of queues of RT/BE priority.
594 * average is updated, with a formula that gives more weight to higher numbers,
595 * to quickly follows sudden increases and decrease slowly
598 static inline unsigned cfq_group_get_avg_queues(struct cfq_data *cfqd,
599 struct cfq_group *cfqg, bool rt)
601 unsigned min_q, max_q;
602 unsigned mult = cfq_hist_divisor - 1;
603 unsigned round = cfq_hist_divisor / 2;
604 unsigned busy = cfq_group_busy_queues_wl(rt, cfqd, cfqg);
606 min_q = min(cfqg->busy_queues_avg[rt], busy);
607 max_q = max(cfqg->busy_queues_avg[rt], busy);
608 cfqg->busy_queues_avg[rt] = (mult * max_q + min_q + round) /
610 return cfqg->busy_queues_avg[rt];
613 static inline unsigned
614 cfq_group_slice(struct cfq_data *cfqd, struct cfq_group *cfqg)
616 struct cfq_rb_root *st = &cfqd->grp_service_tree;
618 return cfq_target_latency * cfqg->weight / st->total_weight;
621 static inline unsigned
622 cfq_scaled_cfqq_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
624 unsigned slice = cfq_prio_to_slice(cfqd, cfqq);
625 if (cfqd->cfq_latency) {
627 * interested queues (we consider only the ones with the same
628 * priority class in the cfq group)
630 unsigned iq = cfq_group_get_avg_queues(cfqd, cfqq->cfqg,
632 unsigned sync_slice = cfqd->cfq_slice[1];
633 unsigned expect_latency = sync_slice * iq;
634 unsigned group_slice = cfq_group_slice(cfqd, cfqq->cfqg);
636 if (expect_latency > group_slice) {
637 unsigned base_low_slice = 2 * cfqd->cfq_slice_idle;
638 /* scale low_slice according to IO priority
639 * and sync vs async */
641 min(slice, base_low_slice * slice / sync_slice);
642 /* the adapted slice value is scaled to fit all iqs
643 * into the target latency */
644 slice = max(slice * group_slice / expect_latency,
652 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
654 unsigned slice = cfq_scaled_cfqq_slice(cfqd, cfqq);
656 cfqq->slice_start = jiffies;
657 cfqq->slice_end = jiffies + slice;
658 cfqq->allocated_slice = slice;
659 cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
663 * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
664 * isn't valid until the first request from the dispatch is activated
665 * and the slice time set.
667 static inline bool cfq_slice_used(struct cfq_queue *cfqq)
669 if (cfq_cfqq_slice_new(cfqq))
671 if (time_before(jiffies, cfqq->slice_end))
678 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
679 * We choose the request that is closest to the head right now. Distance
680 * behind the head is penalized and only allowed to a certain extent.
682 static struct request *
683 cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2, sector_t last)
685 sector_t s1, s2, d1 = 0, d2 = 0;
686 unsigned long back_max;
687 #define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */
688 #define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */
689 unsigned wrap = 0; /* bit mask: requests behind the disk head? */
691 if (rq1 == NULL || rq1 == rq2)
696 if (rq_is_sync(rq1) != rq_is_sync(rq2))
697 return rq_is_sync(rq1) ? rq1 : rq2;
699 if ((rq1->cmd_flags ^ rq2->cmd_flags) & REQ_PRIO)
700 return rq1->cmd_flags & REQ_PRIO ? rq1 : rq2;
702 s1 = blk_rq_pos(rq1);
703 s2 = blk_rq_pos(rq2);
706 * by definition, 1KiB is 2 sectors
708 back_max = cfqd->cfq_back_max * 2;
711 * Strict one way elevator _except_ in the case where we allow
712 * short backward seeks which are biased as twice the cost of a
713 * similar forward seek.
717 else if (s1 + back_max >= last)
718 d1 = (last - s1) * cfqd->cfq_back_penalty;
720 wrap |= CFQ_RQ1_WRAP;
724 else if (s2 + back_max >= last)
725 d2 = (last - s2) * cfqd->cfq_back_penalty;
727 wrap |= CFQ_RQ2_WRAP;
729 /* Found required data */
732 * By doing switch() on the bit mask "wrap" we avoid having to
733 * check two variables for all permutations: --> faster!
736 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
752 case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
755 * Since both rqs are wrapped,
756 * start with the one that's further behind head
757 * (--> only *one* back seek required),
758 * since back seek takes more time than forward.
768 * The below is leftmost cache rbtree addon
770 static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
772 /* Service tree is empty */
777 root->left = rb_first(&root->rb);
780 return rb_entry(root->left, struct cfq_queue, rb_node);
785 static struct cfq_group *cfq_rb_first_group(struct cfq_rb_root *root)
788 root->left = rb_first(&root->rb);
791 return rb_entry_cfqg(root->left);
796 static void rb_erase_init(struct rb_node *n, struct rb_root *root)
802 static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
806 rb_erase_init(n, &root->rb);
811 * would be nice to take fifo expire time into account as well
813 static struct request *
814 cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
815 struct request *last)
817 struct rb_node *rbnext = rb_next(&last->rb_node);
818 struct rb_node *rbprev = rb_prev(&last->rb_node);
819 struct request *next = NULL, *prev = NULL;
821 BUG_ON(RB_EMPTY_NODE(&last->rb_node));
824 prev = rb_entry_rq(rbprev);
827 next = rb_entry_rq(rbnext);
829 rbnext = rb_first(&cfqq->sort_list);
830 if (rbnext && rbnext != &last->rb_node)
831 next = rb_entry_rq(rbnext);
834 return cfq_choose_req(cfqd, next, prev, blk_rq_pos(last));
837 static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
838 struct cfq_queue *cfqq)
841 * just an approximation, should be ok.
843 return (cfqq->cfqg->nr_cfqq - 1) * (cfq_prio_slice(cfqd, 1, 0) -
844 cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
848 cfqg_key(struct cfq_rb_root *st, struct cfq_group *cfqg)
850 return cfqg->vdisktime - st->min_vdisktime;
854 __cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
856 struct rb_node **node = &st->rb.rb_node;
857 struct rb_node *parent = NULL;
858 struct cfq_group *__cfqg;
859 s64 key = cfqg_key(st, cfqg);
862 while (*node != NULL) {
864 __cfqg = rb_entry_cfqg(parent);
866 if (key < cfqg_key(st, __cfqg))
867 node = &parent->rb_left;
869 node = &parent->rb_right;
875 st->left = &cfqg->rb_node;
877 rb_link_node(&cfqg->rb_node, parent, node);
878 rb_insert_color(&cfqg->rb_node, &st->rb);
882 cfq_update_group_weight(struct cfq_group *cfqg)
884 BUG_ON(!RB_EMPTY_NODE(&cfqg->rb_node));
885 if (cfqg->needs_update) {
886 cfqg->weight = cfqg->new_weight;
887 cfqg->needs_update = false;
892 cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
894 BUG_ON(!RB_EMPTY_NODE(&cfqg->rb_node));
896 cfq_update_group_weight(cfqg);
897 __cfq_group_service_tree_add(st, cfqg);
898 st->total_weight += cfqg->weight;
902 cfq_group_notify_queue_add(struct cfq_data *cfqd, struct cfq_group *cfqg)
904 struct cfq_rb_root *st = &cfqd->grp_service_tree;
905 struct cfq_group *__cfqg;
909 if (!RB_EMPTY_NODE(&cfqg->rb_node))
913 * Currently put the group at the end. Later implement something
914 * so that groups get lesser vtime based on their weights, so that
915 * if group does not loose all if it was not continuously backlogged.
917 n = rb_last(&st->rb);
919 __cfqg = rb_entry_cfqg(n);
920 cfqg->vdisktime = __cfqg->vdisktime + CFQ_IDLE_DELAY;
922 cfqg->vdisktime = st->min_vdisktime;
923 cfq_group_service_tree_add(st, cfqg);
927 cfq_group_service_tree_del(struct cfq_rb_root *st, struct cfq_group *cfqg)
929 st->total_weight -= cfqg->weight;
930 if (!RB_EMPTY_NODE(&cfqg->rb_node))
931 cfq_rb_erase(&cfqg->rb_node, st);
935 cfq_group_notify_queue_del(struct cfq_data *cfqd, struct cfq_group *cfqg)
937 struct cfq_rb_root *st = &cfqd->grp_service_tree;
939 BUG_ON(cfqg->nr_cfqq < 1);
942 /* If there are other cfq queues under this group, don't delete it */
946 cfq_log_cfqg(cfqd, cfqg, "del_from_rr group");
947 cfq_group_service_tree_del(st, cfqg);
948 cfqg->saved_workload_slice = 0;
949 cfq_blkiocg_update_dequeue_stats(cfqg_to_blkg(cfqg), 1);
952 static inline unsigned int cfq_cfqq_slice_usage(struct cfq_queue *cfqq,
953 unsigned int *unaccounted_time)
955 unsigned int slice_used;
958 * Queue got expired before even a single request completed or
959 * got expired immediately after first request completion.
961 if (!cfqq->slice_start || cfqq->slice_start == jiffies) {
963 * Also charge the seek time incurred to the group, otherwise
964 * if there are mutiple queues in the group, each can dispatch
965 * a single request on seeky media and cause lots of seek time
966 * and group will never know it.
968 slice_used = max_t(unsigned, (jiffies - cfqq->dispatch_start),
971 slice_used = jiffies - cfqq->slice_start;
972 if (slice_used > cfqq->allocated_slice) {
973 *unaccounted_time = slice_used - cfqq->allocated_slice;
974 slice_used = cfqq->allocated_slice;
976 if (time_after(cfqq->slice_start, cfqq->dispatch_start))
977 *unaccounted_time += cfqq->slice_start -
978 cfqq->dispatch_start;
984 static void cfq_group_served(struct cfq_data *cfqd, struct cfq_group *cfqg,
985 struct cfq_queue *cfqq)
987 struct cfq_rb_root *st = &cfqd->grp_service_tree;
988 unsigned int used_sl, charge, unaccounted_sl = 0;
989 int nr_sync = cfqg->nr_cfqq - cfqg_busy_async_queues(cfqd, cfqg)
990 - cfqg->service_tree_idle.count;
993 used_sl = charge = cfq_cfqq_slice_usage(cfqq, &unaccounted_sl);
996 charge = cfqq->slice_dispatch;
997 else if (!cfq_cfqq_sync(cfqq) && !nr_sync)
998 charge = cfqq->allocated_slice;
1000 /* Can't update vdisktime while group is on service tree */
1001 cfq_group_service_tree_del(st, cfqg);
1002 cfqg->vdisktime += cfq_scale_slice(charge, cfqg);
1003 /* If a new weight was requested, update now, off tree */
1004 cfq_group_service_tree_add(st, cfqg);
1006 /* This group is being expired. Save the context */
1007 if (time_after(cfqd->workload_expires, jiffies)) {
1008 cfqg->saved_workload_slice = cfqd->workload_expires
1010 cfqg->saved_workload = cfqd->serving_type;
1011 cfqg->saved_serving_prio = cfqd->serving_prio;
1013 cfqg->saved_workload_slice = 0;
1015 cfq_log_cfqg(cfqd, cfqg, "served: vt=%llu min_vt=%llu", cfqg->vdisktime,
1017 cfq_log_cfqq(cfqq->cfqd, cfqq,
1018 "sl_used=%u disp=%u charge=%u iops=%u sect=%lu",
1019 used_sl, cfqq->slice_dispatch, charge,
1020 iops_mode(cfqd), cfqq->nr_sectors);
1021 cfq_blkiocg_update_timeslice_used(cfqg_to_blkg(cfqg), used_sl,
1023 cfq_blkiocg_set_start_empty_time(cfqg_to_blkg(cfqg));
1027 * cfq_init_cfqg_base - initialize base part of a cfq_group
1028 * @cfqg: cfq_group to initialize
1030 * Initialize the base part which is used whether %CONFIG_CFQ_GROUP_IOSCHED
1031 * is enabled or not.
1033 static void cfq_init_cfqg_base(struct cfq_group *cfqg)
1035 struct cfq_rb_root *st;
1038 for_each_cfqg_st(cfqg, i, j, st)
1040 RB_CLEAR_NODE(&cfqg->rb_node);
1042 cfqg->ttime.last_end_request = jiffies;
1045 #ifdef CONFIG_CFQ_GROUP_IOSCHED
1046 static void cfq_update_blkio_group_weight(struct request_queue *q,
1047 struct blkio_group *blkg,
1048 unsigned int weight)
1050 struct cfq_group *cfqg = blkg_to_cfqg(blkg);
1052 cfqg->new_weight = weight;
1053 cfqg->needs_update = true;
1056 static void cfq_link_blkio_group(struct request_queue *q,
1057 struct blkio_group *blkg)
1059 struct cfq_data *cfqd = q->elevator->elevator_data;
1060 struct cfq_group *cfqg = blkg_to_cfqg(blkg);
1062 cfqd->nr_blkcg_linked_grps++;
1064 /* Add group on cfqd list */
1065 hlist_add_head(&cfqg->cfqd_node, &cfqd->cfqg_list);
1068 static void cfq_init_blkio_group(struct blkio_group *blkg)
1070 struct cfq_group *cfqg = blkg_to_cfqg(blkg);
1072 cfq_init_cfqg_base(cfqg);
1073 cfqg->weight = blkg->blkcg->weight;
1076 * Take the initial reference that will be released on destroy
1077 * This can be thought of a joint reference by cgroup and
1078 * elevator which will be dropped by either elevator exit
1079 * or cgroup deletion path depending on who is exiting first.
1085 * Search for the cfq group current task belongs to. request_queue lock must
1088 static struct cfq_group *cfq_lookup_create_cfqg(struct cfq_data *cfqd,
1089 struct blkio_cgroup *blkcg)
1091 struct request_queue *q = cfqd->queue;
1092 struct cfq_group *cfqg = NULL;
1094 /* avoid lookup for the common case where there's no blkio cgroup */
1095 if (blkcg == &blkio_root_cgroup) {
1096 cfqg = cfqd->root_group;
1098 struct blkio_group *blkg;
1100 blkg = blkg_lookup_create(blkcg, q, BLKIO_POLICY_PROP, false);
1102 cfqg = blkg_to_cfqg(blkg);
1108 static inline struct cfq_group *cfq_ref_get_cfqg(struct cfq_group *cfqg)
1114 static void cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg)
1116 /* Currently, all async queues are mapped to root group */
1117 if (!cfq_cfqq_sync(cfqq))
1118 cfqg = cfqq->cfqd->root_group;
1121 /* cfqq reference on cfqg */
1125 static void cfq_put_cfqg(struct cfq_group *cfqg)
1127 struct blkio_group *blkg = cfqg_to_blkg(cfqg);
1128 struct cfq_rb_root *st;
1131 BUG_ON(cfqg->ref <= 0);
1136 /* release the extra blkcg reference this blkg has been holding */
1137 css_put(&blkg->blkcg->css);
1139 for_each_cfqg_st(cfqg, i, j, st)
1140 BUG_ON(!RB_EMPTY_ROOT(&st->rb));
1141 free_percpu(blkg->stats_cpu);
1146 static void cfq_destroy_cfqg(struct cfq_data *cfqd, struct cfq_group *cfqg)
1148 /* Something wrong if we are trying to remove same group twice */
1149 BUG_ON(hlist_unhashed(&cfqg->cfqd_node));
1151 hlist_del_init(&cfqg->cfqd_node);
1153 BUG_ON(cfqd->nr_blkcg_linked_grps <= 0);
1154 cfqd->nr_blkcg_linked_grps--;
1157 * Put the reference taken at the time of creation so that when all
1158 * queues are gone, group can be destroyed.
1163 static bool cfq_release_cfq_groups(struct cfq_data *cfqd)
1165 struct hlist_node *pos, *n;
1166 struct cfq_group *cfqg;
1169 hlist_for_each_entry_safe(cfqg, pos, n, &cfqd->cfqg_list, cfqd_node) {
1171 * If cgroup removal path got to blk_group first and removed
1172 * it from cgroup list, then it will take care of destroying
1175 if (!cfq_blkiocg_del_blkio_group(cfqg_to_blkg(cfqg)))
1176 cfq_destroy_cfqg(cfqd, cfqg);
1184 * Blk cgroup controller notification saying that blkio_group object is being
1185 * delinked as associated cgroup object is going away. That also means that
1186 * no new IO will come in this group. So get rid of this group as soon as
1187 * any pending IO in the group is finished.
1189 * This function is called under rcu_read_lock(). key is the rcu protected
1190 * pointer. That means @q is a valid request_queue pointer as long as we
1191 * are rcu read lock.
1193 * @q was fetched from blkio_group under blkio_cgroup->lock. That means
1194 * it should not be NULL as even if elevator was exiting, cgroup deltion
1195 * path got to it first.
1197 static void cfq_unlink_blkio_group(struct request_queue *q,
1198 struct blkio_group *blkg)
1200 struct cfq_data *cfqd = q->elevator->elevator_data;
1201 unsigned long flags;
1203 spin_lock_irqsave(q->queue_lock, flags);
1204 cfq_destroy_cfqg(cfqd, blkg_to_cfqg(blkg));
1205 spin_unlock_irqrestore(q->queue_lock, flags);
1208 static struct elevator_type iosched_cfq;
1210 static bool cfq_clear_queue(struct request_queue *q)
1212 lockdep_assert_held(q->queue_lock);
1214 /* shoot down blkgs iff the current elevator is cfq */
1215 if (!q->elevator || q->elevator->type != &iosched_cfq)
1218 return cfq_release_cfq_groups(q->elevator->elevator_data);
1221 #else /* GROUP_IOSCHED */
1222 static struct cfq_group *cfq_lookup_create_cfqg(struct cfq_data *cfqd,
1223 struct blkio_cgroup *blkcg)
1225 return cfqd->root_group;
1228 static inline struct cfq_group *cfq_ref_get_cfqg(struct cfq_group *cfqg)
1234 cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) {
1238 static void cfq_release_cfq_groups(struct cfq_data *cfqd) {}
1239 static inline void cfq_put_cfqg(struct cfq_group *cfqg) {}
1241 #endif /* GROUP_IOSCHED */
1244 * The cfqd->service_trees holds all pending cfq_queue's that have
1245 * requests waiting to be processed. It is sorted in the order that
1246 * we will service the queues.
1248 static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1251 struct rb_node **p, *parent;
1252 struct cfq_queue *__cfqq;
1253 unsigned long rb_key;
1254 struct cfq_rb_root *service_tree;
1258 service_tree = service_tree_for(cfqq->cfqg, cfqq_prio(cfqq),
1260 if (cfq_class_idle(cfqq)) {
1261 rb_key = CFQ_IDLE_DELAY;
1262 parent = rb_last(&service_tree->rb);
1263 if (parent && parent != &cfqq->rb_node) {
1264 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1265 rb_key += __cfqq->rb_key;
1268 } else if (!add_front) {
1270 * Get our rb key offset. Subtract any residual slice
1271 * value carried from last service. A negative resid
1272 * count indicates slice overrun, and this should position
1273 * the next service time further away in the tree.
1275 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
1276 rb_key -= cfqq->slice_resid;
1277 cfqq->slice_resid = 0;
1280 __cfqq = cfq_rb_first(service_tree);
1281 rb_key += __cfqq ? __cfqq->rb_key : jiffies;
1284 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1287 * same position, nothing more to do
1289 if (rb_key == cfqq->rb_key &&
1290 cfqq->service_tree == service_tree)
1293 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1294 cfqq->service_tree = NULL;
1299 cfqq->service_tree = service_tree;
1300 p = &service_tree->rb.rb_node;
1305 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1308 * sort by key, that represents service time.
1310 if (time_before(rb_key, __cfqq->rb_key))
1313 n = &(*p)->rb_right;
1321 service_tree->left = &cfqq->rb_node;
1323 cfqq->rb_key = rb_key;
1324 rb_link_node(&cfqq->rb_node, parent, p);
1325 rb_insert_color(&cfqq->rb_node, &service_tree->rb);
1326 service_tree->count++;
1327 if (add_front || !new_cfqq)
1329 cfq_group_notify_queue_add(cfqd, cfqq->cfqg);
1332 static struct cfq_queue *
1333 cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
1334 sector_t sector, struct rb_node **ret_parent,
1335 struct rb_node ***rb_link)
1337 struct rb_node **p, *parent;
1338 struct cfq_queue *cfqq = NULL;
1346 cfqq = rb_entry(parent, struct cfq_queue, p_node);
1349 * Sort strictly based on sector. Smallest to the left,
1350 * largest to the right.
1352 if (sector > blk_rq_pos(cfqq->next_rq))
1353 n = &(*p)->rb_right;
1354 else if (sector < blk_rq_pos(cfqq->next_rq))
1362 *ret_parent = parent;
1368 static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1370 struct rb_node **p, *parent;
1371 struct cfq_queue *__cfqq;
1374 rb_erase(&cfqq->p_node, cfqq->p_root);
1375 cfqq->p_root = NULL;
1378 if (cfq_class_idle(cfqq))
1383 cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
1384 __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
1385 blk_rq_pos(cfqq->next_rq), &parent, &p);
1387 rb_link_node(&cfqq->p_node, parent, p);
1388 rb_insert_color(&cfqq->p_node, cfqq->p_root);
1390 cfqq->p_root = NULL;
1394 * Update cfqq's position in the service tree.
1396 static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1399 * Resorting requires the cfqq to be on the RR list already.
1401 if (cfq_cfqq_on_rr(cfqq)) {
1402 cfq_service_tree_add(cfqd, cfqq, 0);
1403 cfq_prio_tree_add(cfqd, cfqq);
1408 * add to busy list of queues for service, trying to be fair in ordering
1409 * the pending list according to last request service
1411 static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1413 cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
1414 BUG_ON(cfq_cfqq_on_rr(cfqq));
1415 cfq_mark_cfqq_on_rr(cfqq);
1416 cfqd->busy_queues++;
1417 if (cfq_cfqq_sync(cfqq))
1418 cfqd->busy_sync_queues++;
1420 cfq_resort_rr_list(cfqd, cfqq);
1424 * Called when the cfqq no longer has requests pending, remove it from
1427 static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1429 cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
1430 BUG_ON(!cfq_cfqq_on_rr(cfqq));
1431 cfq_clear_cfqq_on_rr(cfqq);
1433 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1434 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1435 cfqq->service_tree = NULL;
1438 rb_erase(&cfqq->p_node, cfqq->p_root);
1439 cfqq->p_root = NULL;
1442 cfq_group_notify_queue_del(cfqd, cfqq->cfqg);
1443 BUG_ON(!cfqd->busy_queues);
1444 cfqd->busy_queues--;
1445 if (cfq_cfqq_sync(cfqq))
1446 cfqd->busy_sync_queues--;
1450 * rb tree support functions
1452 static void cfq_del_rq_rb(struct request *rq)
1454 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1455 const int sync = rq_is_sync(rq);
1457 BUG_ON(!cfqq->queued[sync]);
1458 cfqq->queued[sync]--;
1460 elv_rb_del(&cfqq->sort_list, rq);
1462 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) {
1464 * Queue will be deleted from service tree when we actually
1465 * expire it later. Right now just remove it from prio tree
1469 rb_erase(&cfqq->p_node, cfqq->p_root);
1470 cfqq->p_root = NULL;
1475 static void cfq_add_rq_rb(struct request *rq)
1477 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1478 struct cfq_data *cfqd = cfqq->cfqd;
1479 struct request *prev;
1481 cfqq->queued[rq_is_sync(rq)]++;
1483 elv_rb_add(&cfqq->sort_list, rq);
1485 if (!cfq_cfqq_on_rr(cfqq))
1486 cfq_add_cfqq_rr(cfqd, cfqq);
1489 * check if this request is a better next-serve candidate
1491 prev = cfqq->next_rq;
1492 cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq, cfqd->last_position);
1495 * adjust priority tree position, if ->next_rq changes
1497 if (prev != cfqq->next_rq)
1498 cfq_prio_tree_add(cfqd, cfqq);
1500 BUG_ON(!cfqq->next_rq);
1503 static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
1505 elv_rb_del(&cfqq->sort_list, rq);
1506 cfqq->queued[rq_is_sync(rq)]--;
1507 cfq_blkiocg_update_io_remove_stats(cfqg_to_blkg(RQ_CFQG(rq)),
1508 rq_data_dir(rq), rq_is_sync(rq));
1510 cfq_blkiocg_update_io_add_stats(cfqg_to_blkg(RQ_CFQG(rq)),
1511 cfqg_to_blkg(cfqq->cfqd->serving_group),
1512 rq_data_dir(rq), rq_is_sync(rq));
1515 static struct request *
1516 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
1518 struct task_struct *tsk = current;
1519 struct cfq_io_cq *cic;
1520 struct cfq_queue *cfqq;
1522 cic = cfq_cic_lookup(cfqd, tsk->io_context);
1526 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
1528 sector_t sector = bio->bi_sector + bio_sectors(bio);
1530 return elv_rb_find(&cfqq->sort_list, sector);
1536 static void cfq_activate_request(struct request_queue *q, struct request *rq)
1538 struct cfq_data *cfqd = q->elevator->elevator_data;
1540 cfqd->rq_in_driver++;
1541 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
1542 cfqd->rq_in_driver);
1544 cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
1547 static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
1549 struct cfq_data *cfqd = q->elevator->elevator_data;
1551 WARN_ON(!cfqd->rq_in_driver);
1552 cfqd->rq_in_driver--;
1553 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
1554 cfqd->rq_in_driver);
1557 static void cfq_remove_request(struct request *rq)
1559 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1561 if (cfqq->next_rq == rq)
1562 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
1564 list_del_init(&rq->queuelist);
1567 cfqq->cfqd->rq_queued--;
1568 cfq_blkiocg_update_io_remove_stats(cfqg_to_blkg(RQ_CFQG(rq)),
1569 rq_data_dir(rq), rq_is_sync(rq));
1570 if (rq->cmd_flags & REQ_PRIO) {
1571 WARN_ON(!cfqq->prio_pending);
1572 cfqq->prio_pending--;
1576 static int cfq_merge(struct request_queue *q, struct request **req,
1579 struct cfq_data *cfqd = q->elevator->elevator_data;
1580 struct request *__rq;
1582 __rq = cfq_find_rq_fmerge(cfqd, bio);
1583 if (__rq && elv_rq_merge_ok(__rq, bio)) {
1585 return ELEVATOR_FRONT_MERGE;
1588 return ELEVATOR_NO_MERGE;
1591 static void cfq_merged_request(struct request_queue *q, struct request *req,
1594 if (type == ELEVATOR_FRONT_MERGE) {
1595 struct cfq_queue *cfqq = RQ_CFQQ(req);
1597 cfq_reposition_rq_rb(cfqq, req);
1601 static void cfq_bio_merged(struct request_queue *q, struct request *req,
1604 cfq_blkiocg_update_io_merged_stats(cfqg_to_blkg(RQ_CFQG(req)),
1605 bio_data_dir(bio), cfq_bio_sync(bio));
1609 cfq_merged_requests(struct request_queue *q, struct request *rq,
1610 struct request *next)
1612 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1613 struct cfq_data *cfqd = q->elevator->elevator_data;
1616 * reposition in fifo if next is older than rq
1618 if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
1619 time_before(rq_fifo_time(next), rq_fifo_time(rq))) {
1620 list_move(&rq->queuelist, &next->queuelist);
1621 rq_set_fifo_time(rq, rq_fifo_time(next));
1624 if (cfqq->next_rq == next)
1626 cfq_remove_request(next);
1627 cfq_blkiocg_update_io_merged_stats(cfqg_to_blkg(RQ_CFQG(rq)),
1628 rq_data_dir(next), rq_is_sync(next));
1630 cfqq = RQ_CFQQ(next);
1632 * all requests of this queue are merged to other queues, delete it
1633 * from the service tree. If it's the active_queue,
1634 * cfq_dispatch_requests() will choose to expire it or do idle
1636 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list) &&
1637 cfqq != cfqd->active_queue)
1638 cfq_del_cfqq_rr(cfqd, cfqq);
1641 static int cfq_allow_merge(struct request_queue *q, struct request *rq,
1644 struct cfq_data *cfqd = q->elevator->elevator_data;
1645 struct cfq_io_cq *cic;
1646 struct cfq_queue *cfqq;
1649 * Disallow merge of a sync bio into an async request.
1651 if (cfq_bio_sync(bio) && !rq_is_sync(rq))
1655 * Lookup the cfqq that this bio will be queued with and allow
1656 * merge only if rq is queued there.
1658 cic = cfq_cic_lookup(cfqd, current->io_context);
1662 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
1663 return cfqq == RQ_CFQQ(rq);
1666 static inline void cfq_del_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1668 del_timer(&cfqd->idle_slice_timer);
1669 cfq_blkiocg_update_idle_time_stats(cfqg_to_blkg(cfqq->cfqg));
1672 static void __cfq_set_active_queue(struct cfq_data *cfqd,
1673 struct cfq_queue *cfqq)
1676 cfq_log_cfqq(cfqd, cfqq, "set_active wl_prio:%d wl_type:%d",
1677 cfqd->serving_prio, cfqd->serving_type);
1678 cfq_blkiocg_update_avg_queue_size_stats(cfqg_to_blkg(cfqq->cfqg));
1679 cfqq->slice_start = 0;
1680 cfqq->dispatch_start = jiffies;
1681 cfqq->allocated_slice = 0;
1682 cfqq->slice_end = 0;
1683 cfqq->slice_dispatch = 0;
1684 cfqq->nr_sectors = 0;
1686 cfq_clear_cfqq_wait_request(cfqq);
1687 cfq_clear_cfqq_must_dispatch(cfqq);
1688 cfq_clear_cfqq_must_alloc_slice(cfqq);
1689 cfq_clear_cfqq_fifo_expire(cfqq);
1690 cfq_mark_cfqq_slice_new(cfqq);
1692 cfq_del_timer(cfqd, cfqq);
1695 cfqd->active_queue = cfqq;
1699 * current cfqq expired its slice (or was too idle), select new one
1702 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1705 cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
1707 if (cfq_cfqq_wait_request(cfqq))
1708 cfq_del_timer(cfqd, cfqq);
1710 cfq_clear_cfqq_wait_request(cfqq);
1711 cfq_clear_cfqq_wait_busy(cfqq);
1714 * If this cfqq is shared between multiple processes, check to
1715 * make sure that those processes are still issuing I/Os within
1716 * the mean seek distance. If not, it may be time to break the
1717 * queues apart again.
1719 if (cfq_cfqq_coop(cfqq) && CFQQ_SEEKY(cfqq))
1720 cfq_mark_cfqq_split_coop(cfqq);
1723 * store what was left of this slice, if the queue idled/timed out
1726 if (cfq_cfqq_slice_new(cfqq))
1727 cfqq->slice_resid = cfq_scaled_cfqq_slice(cfqd, cfqq);
1729 cfqq->slice_resid = cfqq->slice_end - jiffies;
1730 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
1733 cfq_group_served(cfqd, cfqq->cfqg, cfqq);
1735 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
1736 cfq_del_cfqq_rr(cfqd, cfqq);
1738 cfq_resort_rr_list(cfqd, cfqq);
1740 if (cfqq == cfqd->active_queue)
1741 cfqd->active_queue = NULL;
1743 if (cfqd->active_cic) {
1744 put_io_context(cfqd->active_cic->icq.ioc);
1745 cfqd->active_cic = NULL;
1749 static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out)
1751 struct cfq_queue *cfqq = cfqd->active_queue;
1754 __cfq_slice_expired(cfqd, cfqq, timed_out);
1758 * Get next queue for service. Unless we have a queue preemption,
1759 * we'll simply select the first cfqq in the service tree.
1761 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
1763 struct cfq_rb_root *service_tree =
1764 service_tree_for(cfqd->serving_group, cfqd->serving_prio,
1765 cfqd->serving_type);
1767 if (!cfqd->rq_queued)
1770 /* There is nothing to dispatch */
1773 if (RB_EMPTY_ROOT(&service_tree->rb))
1775 return cfq_rb_first(service_tree);
1778 static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd)
1780 struct cfq_group *cfqg;
1781 struct cfq_queue *cfqq;
1783 struct cfq_rb_root *st;
1785 if (!cfqd->rq_queued)
1788 cfqg = cfq_get_next_cfqg(cfqd);
1792 for_each_cfqg_st(cfqg, i, j, st)
1793 if ((cfqq = cfq_rb_first(st)) != NULL)
1799 * Get and set a new active queue for service.
1801 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
1802 struct cfq_queue *cfqq)
1805 cfqq = cfq_get_next_queue(cfqd);
1807 __cfq_set_active_queue(cfqd, cfqq);
1811 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
1814 if (blk_rq_pos(rq) >= cfqd->last_position)
1815 return blk_rq_pos(rq) - cfqd->last_position;
1817 return cfqd->last_position - blk_rq_pos(rq);
1820 static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1823 return cfq_dist_from_last(cfqd, rq) <= CFQQ_CLOSE_THR;
1826 static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
1827 struct cfq_queue *cur_cfqq)
1829 struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
1830 struct rb_node *parent, *node;
1831 struct cfq_queue *__cfqq;
1832 sector_t sector = cfqd->last_position;
1834 if (RB_EMPTY_ROOT(root))
1838 * First, if we find a request starting at the end of the last
1839 * request, choose it.
1841 __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
1846 * If the exact sector wasn't found, the parent of the NULL leaf
1847 * will contain the closest sector.
1849 __cfqq = rb_entry(parent, struct cfq_queue, p_node);
1850 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1853 if (blk_rq_pos(__cfqq->next_rq) < sector)
1854 node = rb_next(&__cfqq->p_node);
1856 node = rb_prev(&__cfqq->p_node);
1860 __cfqq = rb_entry(node, struct cfq_queue, p_node);
1861 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1869 * cur_cfqq - passed in so that we don't decide that the current queue is
1870 * closely cooperating with itself.
1872 * So, basically we're assuming that that cur_cfqq has dispatched at least
1873 * one request, and that cfqd->last_position reflects a position on the disk
1874 * associated with the I/O issued by cur_cfqq. I'm not sure this is a valid
1877 static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
1878 struct cfq_queue *cur_cfqq)
1880 struct cfq_queue *cfqq;
1882 if (cfq_class_idle(cur_cfqq))
1884 if (!cfq_cfqq_sync(cur_cfqq))
1886 if (CFQQ_SEEKY(cur_cfqq))
1890 * Don't search priority tree if it's the only queue in the group.
1892 if (cur_cfqq->cfqg->nr_cfqq == 1)
1896 * We should notice if some of the queues are cooperating, eg
1897 * working closely on the same area of the disk. In that case,
1898 * we can group them together and don't waste time idling.
1900 cfqq = cfqq_close(cfqd, cur_cfqq);
1904 /* If new queue belongs to different cfq_group, don't choose it */
1905 if (cur_cfqq->cfqg != cfqq->cfqg)
1909 * It only makes sense to merge sync queues.
1911 if (!cfq_cfqq_sync(cfqq))
1913 if (CFQQ_SEEKY(cfqq))
1917 * Do not merge queues of different priority classes
1919 if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq))
1926 * Determine whether we should enforce idle window for this queue.
1929 static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1931 enum wl_prio_t prio = cfqq_prio(cfqq);
1932 struct cfq_rb_root *service_tree = cfqq->service_tree;
1934 BUG_ON(!service_tree);
1935 BUG_ON(!service_tree->count);
1937 if (!cfqd->cfq_slice_idle)
1940 /* We never do for idle class queues. */
1941 if (prio == IDLE_WORKLOAD)
1944 /* We do for queues that were marked with idle window flag. */
1945 if (cfq_cfqq_idle_window(cfqq) &&
1946 !(blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag))
1950 * Otherwise, we do only if they are the last ones
1951 * in their service tree.
1953 if (service_tree->count == 1 && cfq_cfqq_sync(cfqq) &&
1954 !cfq_io_thinktime_big(cfqd, &service_tree->ttime, false))
1956 cfq_log_cfqq(cfqd, cfqq, "Not idling. st->count:%d",
1957 service_tree->count);
1961 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
1963 struct cfq_queue *cfqq = cfqd->active_queue;
1964 struct cfq_io_cq *cic;
1965 unsigned long sl, group_idle = 0;
1968 * SSD device without seek penalty, disable idling. But only do so
1969 * for devices that support queuing, otherwise we still have a problem
1970 * with sync vs async workloads.
1972 if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
1975 WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
1976 WARN_ON(cfq_cfqq_slice_new(cfqq));
1979 * idle is disabled, either manually or by past process history
1981 if (!cfq_should_idle(cfqd, cfqq)) {
1982 /* no queue idling. Check for group idling */
1983 if (cfqd->cfq_group_idle)
1984 group_idle = cfqd->cfq_group_idle;
1990 * still active requests from this queue, don't idle
1992 if (cfqq->dispatched)
1996 * task has exited, don't wait
1998 cic = cfqd->active_cic;
1999 if (!cic || !atomic_read(&cic->icq.ioc->nr_tasks))
2003 * If our average think time is larger than the remaining time
2004 * slice, then don't idle. This avoids overrunning the allotted
2007 if (sample_valid(cic->ttime.ttime_samples) &&
2008 (cfqq->slice_end - jiffies < cic->ttime.ttime_mean)) {
2009 cfq_log_cfqq(cfqd, cfqq, "Not idling. think_time:%lu",
2010 cic->ttime.ttime_mean);
2014 /* There are other queues in the group, don't do group idle */
2015 if (group_idle && cfqq->cfqg->nr_cfqq > 1)
2018 cfq_mark_cfqq_wait_request(cfqq);
2021 sl = cfqd->cfq_group_idle;
2023 sl = cfqd->cfq_slice_idle;
2025 mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
2026 cfq_blkiocg_update_set_idle_time_stats(cfqg_to_blkg(cfqq->cfqg));
2027 cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu group_idle: %d", sl,
2028 group_idle ? 1 : 0);
2032 * Move request from internal lists to the request queue dispatch list.
2034 static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
2036 struct cfq_data *cfqd = q->elevator->elevator_data;
2037 struct cfq_queue *cfqq = RQ_CFQQ(rq);
2039 cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
2041 cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
2042 cfq_remove_request(rq);
2044 (RQ_CFQG(rq))->dispatched++;
2045 elv_dispatch_sort(q, rq);
2047 cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]++;
2048 cfqq->nr_sectors += blk_rq_sectors(rq);
2049 cfq_blkiocg_update_dispatch_stats(cfqg_to_blkg(cfqq->cfqg),
2050 blk_rq_bytes(rq), rq_data_dir(rq),
2055 * return expired entry, or NULL to just start from scratch in rbtree
2057 static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
2059 struct request *rq = NULL;
2061 if (cfq_cfqq_fifo_expire(cfqq))
2064 cfq_mark_cfqq_fifo_expire(cfqq);
2066 if (list_empty(&cfqq->fifo))
2069 rq = rq_entry_fifo(cfqq->fifo.next);
2070 if (time_before(jiffies, rq_fifo_time(rq)))
2073 cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq);
2078 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2080 const int base_rq = cfqd->cfq_slice_async_rq;
2082 WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
2084 return 2 * base_rq * (IOPRIO_BE_NR - cfqq->ioprio);
2088 * Must be called with the queue_lock held.
2090 static int cfqq_process_refs(struct cfq_queue *cfqq)
2092 int process_refs, io_refs;
2094 io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE];
2095 process_refs = cfqq->ref - io_refs;
2096 BUG_ON(process_refs < 0);
2097 return process_refs;
2100 static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq)
2102 int process_refs, new_process_refs;
2103 struct cfq_queue *__cfqq;
2106 * If there are no process references on the new_cfqq, then it is
2107 * unsafe to follow the ->new_cfqq chain as other cfqq's in the
2108 * chain may have dropped their last reference (not just their
2109 * last process reference).
2111 if (!cfqq_process_refs(new_cfqq))
2114 /* Avoid a circular list and skip interim queue merges */
2115 while ((__cfqq = new_cfqq->new_cfqq)) {
2121 process_refs = cfqq_process_refs(cfqq);
2122 new_process_refs = cfqq_process_refs(new_cfqq);
2124 * If the process for the cfqq has gone away, there is no
2125 * sense in merging the queues.
2127 if (process_refs == 0 || new_process_refs == 0)
2131 * Merge in the direction of the lesser amount of work.
2133 if (new_process_refs >= process_refs) {
2134 cfqq->new_cfqq = new_cfqq;
2135 new_cfqq->ref += process_refs;
2137 new_cfqq->new_cfqq = cfqq;
2138 cfqq->ref += new_process_refs;
2142 static enum wl_type_t cfq_choose_wl(struct cfq_data *cfqd,
2143 struct cfq_group *cfqg, enum wl_prio_t prio)
2145 struct cfq_queue *queue;
2147 bool key_valid = false;
2148 unsigned long lowest_key = 0;
2149 enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD;
2151 for (i = 0; i <= SYNC_WORKLOAD; ++i) {
2152 /* select the one with lowest rb_key */
2153 queue = cfq_rb_first(service_tree_for(cfqg, prio, i));
2155 (!key_valid || time_before(queue->rb_key, lowest_key))) {
2156 lowest_key = queue->rb_key;
2165 static void choose_service_tree(struct cfq_data *cfqd, struct cfq_group *cfqg)
2169 struct cfq_rb_root *st;
2170 unsigned group_slice;
2171 enum wl_prio_t original_prio = cfqd->serving_prio;
2173 /* Choose next priority. RT > BE > IDLE */
2174 if (cfq_group_busy_queues_wl(RT_WORKLOAD, cfqd, cfqg))
2175 cfqd->serving_prio = RT_WORKLOAD;
2176 else if (cfq_group_busy_queues_wl(BE_WORKLOAD, cfqd, cfqg))
2177 cfqd->serving_prio = BE_WORKLOAD;
2179 cfqd->serving_prio = IDLE_WORKLOAD;
2180 cfqd->workload_expires = jiffies + 1;
2184 if (original_prio != cfqd->serving_prio)
2188 * For RT and BE, we have to choose also the type
2189 * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
2192 st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
2196 * check workload expiration, and that we still have other queues ready
2198 if (count && !time_after(jiffies, cfqd->workload_expires))
2202 /* otherwise select new workload type */
2203 cfqd->serving_type =
2204 cfq_choose_wl(cfqd, cfqg, cfqd->serving_prio);
2205 st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
2209 * the workload slice is computed as a fraction of target latency
2210 * proportional to the number of queues in that workload, over
2211 * all the queues in the same priority class
2213 group_slice = cfq_group_slice(cfqd, cfqg);
2215 slice = group_slice * count /
2216 max_t(unsigned, cfqg->busy_queues_avg[cfqd->serving_prio],
2217 cfq_group_busy_queues_wl(cfqd->serving_prio, cfqd, cfqg));
2219 if (cfqd->serving_type == ASYNC_WORKLOAD) {
2223 * Async queues are currently system wide. Just taking
2224 * proportion of queues with-in same group will lead to higher
2225 * async ratio system wide as generally root group is going
2226 * to have higher weight. A more accurate thing would be to
2227 * calculate system wide asnc/sync ratio.
2229 tmp = cfq_target_latency * cfqg_busy_async_queues(cfqd, cfqg);
2230 tmp = tmp/cfqd->busy_queues;
2231 slice = min_t(unsigned, slice, tmp);
2233 /* async workload slice is scaled down according to
2234 * the sync/async slice ratio. */
2235 slice = slice * cfqd->cfq_slice[0] / cfqd->cfq_slice[1];
2237 /* sync workload slice is at least 2 * cfq_slice_idle */
2238 slice = max(slice, 2 * cfqd->cfq_slice_idle);
2240 slice = max_t(unsigned, slice, CFQ_MIN_TT);
2241 cfq_log(cfqd, "workload slice:%d", slice);
2242 cfqd->workload_expires = jiffies + slice;
2245 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd)
2247 struct cfq_rb_root *st = &cfqd->grp_service_tree;
2248 struct cfq_group *cfqg;
2250 if (RB_EMPTY_ROOT(&st->rb))
2252 cfqg = cfq_rb_first_group(st);
2253 update_min_vdisktime(st);
2257 static void cfq_choose_cfqg(struct cfq_data *cfqd)
2259 struct cfq_group *cfqg = cfq_get_next_cfqg(cfqd);
2261 cfqd->serving_group = cfqg;
2263 /* Restore the workload type data */
2264 if (cfqg->saved_workload_slice) {
2265 cfqd->workload_expires = jiffies + cfqg->saved_workload_slice;
2266 cfqd->serving_type = cfqg->saved_workload;
2267 cfqd->serving_prio = cfqg->saved_serving_prio;
2269 cfqd->workload_expires = jiffies - 1;
2271 choose_service_tree(cfqd, cfqg);
2275 * Select a queue for service. If we have a current active queue,
2276 * check whether to continue servicing it, or retrieve and set a new one.
2278 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
2280 struct cfq_queue *cfqq, *new_cfqq = NULL;
2282 cfqq = cfqd->active_queue;
2286 if (!cfqd->rq_queued)
2290 * We were waiting for group to get backlogged. Expire the queue
2292 if (cfq_cfqq_wait_busy(cfqq) && !RB_EMPTY_ROOT(&cfqq->sort_list))
2296 * The active queue has run out of time, expire it and select new.
2298 if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq)) {
2300 * If slice had not expired at the completion of last request
2301 * we might not have turned on wait_busy flag. Don't expire
2302 * the queue yet. Allow the group to get backlogged.
2304 * The very fact that we have used the slice, that means we
2305 * have been idling all along on this queue and it should be
2306 * ok to wait for this request to complete.
2308 if (cfqq->cfqg->nr_cfqq == 1 && RB_EMPTY_ROOT(&cfqq->sort_list)
2309 && cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2313 goto check_group_idle;
2317 * The active queue has requests and isn't expired, allow it to
2320 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
2324 * If another queue has a request waiting within our mean seek
2325 * distance, let it run. The expire code will check for close
2326 * cooperators and put the close queue at the front of the service
2327 * tree. If possible, merge the expiring queue with the new cfqq.
2329 new_cfqq = cfq_close_cooperator(cfqd, cfqq);
2331 if (!cfqq->new_cfqq)
2332 cfq_setup_merge(cfqq, new_cfqq);
2337 * No requests pending. If the active queue still has requests in
2338 * flight or is idling for a new request, allow either of these
2339 * conditions to happen (or time out) before selecting a new queue.
2341 if (timer_pending(&cfqd->idle_slice_timer)) {
2347 * This is a deep seek queue, but the device is much faster than
2348 * the queue can deliver, don't idle
2350 if (CFQQ_SEEKY(cfqq) && cfq_cfqq_idle_window(cfqq) &&
2351 (cfq_cfqq_slice_new(cfqq) ||
2352 (cfqq->slice_end - jiffies > jiffies - cfqq->slice_start))) {
2353 cfq_clear_cfqq_deep(cfqq);
2354 cfq_clear_cfqq_idle_window(cfqq);
2357 if (cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2363 * If group idle is enabled and there are requests dispatched from
2364 * this group, wait for requests to complete.
2367 if (cfqd->cfq_group_idle && cfqq->cfqg->nr_cfqq == 1 &&
2368 cfqq->cfqg->dispatched &&
2369 !cfq_io_thinktime_big(cfqd, &cfqq->cfqg->ttime, true)) {
2375 cfq_slice_expired(cfqd, 0);
2378 * Current queue expired. Check if we have to switch to a new
2382 cfq_choose_cfqg(cfqd);
2384 cfqq = cfq_set_active_queue(cfqd, new_cfqq);
2389 static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
2393 while (cfqq->next_rq) {
2394 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
2398 BUG_ON(!list_empty(&cfqq->fifo));
2400 /* By default cfqq is not expired if it is empty. Do it explicitly */
2401 __cfq_slice_expired(cfqq->cfqd, cfqq, 0);
2406 * Drain our current requests. Used for barriers and when switching
2407 * io schedulers on-the-fly.
2409 static int cfq_forced_dispatch(struct cfq_data *cfqd)
2411 struct cfq_queue *cfqq;
2414 /* Expire the timeslice of the current active queue first */
2415 cfq_slice_expired(cfqd, 0);
2416 while ((cfqq = cfq_get_next_queue_forced(cfqd)) != NULL) {
2417 __cfq_set_active_queue(cfqd, cfqq);
2418 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
2421 BUG_ON(cfqd->busy_queues);
2423 cfq_log(cfqd, "forced_dispatch=%d", dispatched);
2427 static inline bool cfq_slice_used_soon(struct cfq_data *cfqd,
2428 struct cfq_queue *cfqq)
2430 /* the queue hasn't finished any request, can't estimate */
2431 if (cfq_cfqq_slice_new(cfqq))
2433 if (time_after(jiffies + cfqd->cfq_slice_idle * cfqq->dispatched,
2440 static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2442 unsigned int max_dispatch;
2445 * Drain async requests before we start sync IO
2447 if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_flight[BLK_RW_ASYNC])
2451 * If this is an async queue and we have sync IO in flight, let it wait
2453 if (cfqd->rq_in_flight[BLK_RW_SYNC] && !cfq_cfqq_sync(cfqq))
2456 max_dispatch = max_t(unsigned int, cfqd->cfq_quantum / 2, 1);
2457 if (cfq_class_idle(cfqq))
2461 * Does this cfqq already have too much IO in flight?
2463 if (cfqq->dispatched >= max_dispatch) {
2464 bool promote_sync = false;
2466 * idle queue must always only have a single IO in flight
2468 if (cfq_class_idle(cfqq))
2472 * If there is only one sync queue
2473 * we can ignore async queue here and give the sync
2474 * queue no dispatch limit. The reason is a sync queue can
2475 * preempt async queue, limiting the sync queue doesn't make
2476 * sense. This is useful for aiostress test.
2478 if (cfq_cfqq_sync(cfqq) && cfqd->busy_sync_queues == 1)
2479 promote_sync = true;
2482 * We have other queues, don't allow more IO from this one
2484 if (cfqd->busy_queues > 1 && cfq_slice_used_soon(cfqd, cfqq) &&
2489 * Sole queue user, no limit
2491 if (cfqd->busy_queues == 1 || promote_sync)
2495 * Normally we start throttling cfqq when cfq_quantum/2
2496 * requests have been dispatched. But we can drive
2497 * deeper queue depths at the beginning of slice
2498 * subjected to upper limit of cfq_quantum.
2500 max_dispatch = cfqd->cfq_quantum;
2504 * Async queues must wait a bit before being allowed dispatch.
2505 * We also ramp up the dispatch depth gradually for async IO,
2506 * based on the last sync IO we serviced
2508 if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
2509 unsigned long last_sync = jiffies - cfqd->last_delayed_sync;
2512 depth = last_sync / cfqd->cfq_slice[1];
2513 if (!depth && !cfqq->dispatched)
2515 if (depth < max_dispatch)
2516 max_dispatch = depth;
2520 * If we're below the current max, allow a dispatch
2522 return cfqq->dispatched < max_dispatch;
2526 * Dispatch a request from cfqq, moving them to the request queue
2529 static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2533 BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
2535 if (!cfq_may_dispatch(cfqd, cfqq))
2539 * follow expired path, else get first next available
2541 rq = cfq_check_fifo(cfqq);
2546 * insert request into driver dispatch list
2548 cfq_dispatch_insert(cfqd->queue, rq);
2550 if (!cfqd->active_cic) {
2551 struct cfq_io_cq *cic = RQ_CIC(rq);
2553 atomic_long_inc(&cic->icq.ioc->refcount);
2554 cfqd->active_cic = cic;
2561 * Find the cfqq that we need to service and move a request from that to the
2564 static int cfq_dispatch_requests(struct request_queue *q, int force)
2566 struct cfq_data *cfqd = q->elevator->elevator_data;
2567 struct cfq_queue *cfqq;
2569 if (!cfqd->busy_queues)
2572 if (unlikely(force))
2573 return cfq_forced_dispatch(cfqd);
2575 cfqq = cfq_select_queue(cfqd);
2580 * Dispatch a request from this cfqq, if it is allowed
2582 if (!cfq_dispatch_request(cfqd, cfqq))
2585 cfqq->slice_dispatch++;
2586 cfq_clear_cfqq_must_dispatch(cfqq);
2589 * expire an async queue immediately if it has used up its slice. idle
2590 * queue always expire after 1 dispatch round.
2592 if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
2593 cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
2594 cfq_class_idle(cfqq))) {
2595 cfqq->slice_end = jiffies + 1;
2596 cfq_slice_expired(cfqd, 0);
2599 cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
2604 * task holds one reference to the queue, dropped when task exits. each rq
2605 * in-flight on this queue also holds a reference, dropped when rq is freed.
2607 * Each cfq queue took a reference on the parent group. Drop it now.
2608 * queue lock must be held here.
2610 static void cfq_put_queue(struct cfq_queue *cfqq)
2612 struct cfq_data *cfqd = cfqq->cfqd;
2613 struct cfq_group *cfqg;
2615 BUG_ON(cfqq->ref <= 0);
2621 cfq_log_cfqq(cfqd, cfqq, "put_queue");
2622 BUG_ON(rb_first(&cfqq->sort_list));
2623 BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
2626 if (unlikely(cfqd->active_queue == cfqq)) {
2627 __cfq_slice_expired(cfqd, cfqq, 0);
2628 cfq_schedule_dispatch(cfqd);
2631 BUG_ON(cfq_cfqq_on_rr(cfqq));
2632 kmem_cache_free(cfq_pool, cfqq);
2636 static void cfq_put_cooperator(struct cfq_queue *cfqq)
2638 struct cfq_queue *__cfqq, *next;
2641 * If this queue was scheduled to merge with another queue, be
2642 * sure to drop the reference taken on that queue (and others in
2643 * the merge chain). See cfq_setup_merge and cfq_merge_cfqqs.
2645 __cfqq = cfqq->new_cfqq;
2647 if (__cfqq == cfqq) {
2648 WARN(1, "cfqq->new_cfqq loop detected\n");
2651 next = __cfqq->new_cfqq;
2652 cfq_put_queue(__cfqq);
2657 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2659 if (unlikely(cfqq == cfqd->active_queue)) {
2660 __cfq_slice_expired(cfqd, cfqq, 0);
2661 cfq_schedule_dispatch(cfqd);
2664 cfq_put_cooperator(cfqq);
2666 cfq_put_queue(cfqq);
2669 static void cfq_init_icq(struct io_cq *icq)
2671 struct cfq_io_cq *cic = icq_to_cic(icq);
2673 cic->ttime.last_end_request = jiffies;
2676 static void cfq_exit_icq(struct io_cq *icq)
2678 struct cfq_io_cq *cic = icq_to_cic(icq);
2679 struct cfq_data *cfqd = cic_to_cfqd(cic);
2681 if (cic->cfqq[BLK_RW_ASYNC]) {
2682 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]);
2683 cic->cfqq[BLK_RW_ASYNC] = NULL;
2686 if (cic->cfqq[BLK_RW_SYNC]) {
2687 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]);
2688 cic->cfqq[BLK_RW_SYNC] = NULL;
2692 static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc)
2694 struct task_struct *tsk = current;
2697 if (!cfq_cfqq_prio_changed(cfqq))
2700 ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio);
2701 switch (ioprio_class) {
2703 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
2704 case IOPRIO_CLASS_NONE:
2706 * no prio set, inherit CPU scheduling settings
2708 cfqq->ioprio = task_nice_ioprio(tsk);
2709 cfqq->ioprio_class = task_nice_ioclass(tsk);
2711 case IOPRIO_CLASS_RT:
2712 cfqq->ioprio = task_ioprio(ioc);
2713 cfqq->ioprio_class = IOPRIO_CLASS_RT;
2715 case IOPRIO_CLASS_BE:
2716 cfqq->ioprio = task_ioprio(ioc);
2717 cfqq->ioprio_class = IOPRIO_CLASS_BE;
2719 case IOPRIO_CLASS_IDLE:
2720 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
2722 cfq_clear_cfqq_idle_window(cfqq);
2727 * keep track of original prio settings in case we have to temporarily
2728 * elevate the priority of this queue
2730 cfqq->org_ioprio = cfqq->ioprio;
2731 cfq_clear_cfqq_prio_changed(cfqq);
2734 static void changed_ioprio(struct cfq_io_cq *cic)
2736 struct cfq_data *cfqd = cic_to_cfqd(cic);
2737 struct cfq_queue *cfqq;
2739 if (unlikely(!cfqd))
2742 cfqq = cic->cfqq[BLK_RW_ASYNC];
2744 struct cfq_queue *new_cfqq;
2745 new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic->icq.ioc,
2748 cic->cfqq[BLK_RW_ASYNC] = new_cfqq;
2749 cfq_put_queue(cfqq);
2753 cfqq = cic->cfqq[BLK_RW_SYNC];
2755 cfq_mark_cfqq_prio_changed(cfqq);
2758 static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2759 pid_t pid, bool is_sync)
2761 RB_CLEAR_NODE(&cfqq->rb_node);
2762 RB_CLEAR_NODE(&cfqq->p_node);
2763 INIT_LIST_HEAD(&cfqq->fifo);
2768 cfq_mark_cfqq_prio_changed(cfqq);
2771 if (!cfq_class_idle(cfqq))
2772 cfq_mark_cfqq_idle_window(cfqq);
2773 cfq_mark_cfqq_sync(cfqq);
2778 #ifdef CONFIG_CFQ_GROUP_IOSCHED
2779 static void changed_cgroup(struct cfq_io_cq *cic)
2781 struct cfq_queue *sync_cfqq = cic_to_cfqq(cic, 1);
2782 struct cfq_data *cfqd = cic_to_cfqd(cic);
2783 struct request_queue *q;
2785 if (unlikely(!cfqd))
2792 * Drop reference to sync queue. A new sync queue will be
2793 * assigned in new group upon arrival of a fresh request.
2795 cfq_log_cfqq(cfqd, sync_cfqq, "changed cgroup");
2796 cic_set_cfqq(cic, NULL, 1);
2797 cfq_put_queue(sync_cfqq);
2800 #endif /* CONFIG_CFQ_GROUP_IOSCHED */
2802 static struct cfq_queue *
2803 cfq_find_alloc_queue(struct cfq_data *cfqd, bool is_sync,
2804 struct io_context *ioc, gfp_t gfp_mask)
2806 struct blkio_cgroup *blkcg;
2807 struct cfq_queue *cfqq, *new_cfqq = NULL;
2808 struct cfq_io_cq *cic;
2809 struct cfq_group *cfqg;
2814 blkcg = task_blkio_cgroup(current);
2816 cfqg = cfq_lookup_create_cfqg(cfqd, blkcg);
2818 cic = cfq_cic_lookup(cfqd, ioc);
2819 /* cic always exists here */
2820 cfqq = cic_to_cfqq(cic, is_sync);
2823 * Always try a new alloc if we fell back to the OOM cfqq
2824 * originally, since it should just be a temporary situation.
2826 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
2831 } else if (gfp_mask & __GFP_WAIT) {
2833 spin_unlock_irq(cfqd->queue->queue_lock);
2834 new_cfqq = kmem_cache_alloc_node(cfq_pool,
2835 gfp_mask | __GFP_ZERO,
2837 spin_lock_irq(cfqd->queue->queue_lock);
2841 cfqq = kmem_cache_alloc_node(cfq_pool,
2842 gfp_mask | __GFP_ZERO,
2847 cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
2848 cfq_init_prio_data(cfqq, ioc);
2849 cfq_link_cfqq_cfqg(cfqq, cfqg);
2850 cfq_log_cfqq(cfqd, cfqq, "alloced");
2852 cfqq = &cfqd->oom_cfqq;
2856 kmem_cache_free(cfq_pool, new_cfqq);
2862 static struct cfq_queue **
2863 cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
2865 switch (ioprio_class) {
2866 case IOPRIO_CLASS_RT:
2867 return &cfqd->async_cfqq[0][ioprio];
2868 case IOPRIO_CLASS_BE:
2869 return &cfqd->async_cfqq[1][ioprio];
2870 case IOPRIO_CLASS_IDLE:
2871 return &cfqd->async_idle_cfqq;
2877 static struct cfq_queue *
2878 cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct io_context *ioc,
2881 const int ioprio = task_ioprio(ioc);
2882 const int ioprio_class = task_ioprio_class(ioc);
2883 struct cfq_queue **async_cfqq = NULL;
2884 struct cfq_queue *cfqq = NULL;
2887 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
2892 cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask);
2895 * pin the queue now that it's allocated, scheduler exit will prune it
2897 if (!is_sync && !(*async_cfqq)) {
2907 __cfq_update_io_thinktime(struct cfq_ttime *ttime, unsigned long slice_idle)
2909 unsigned long elapsed = jiffies - ttime->last_end_request;
2910 elapsed = min(elapsed, 2UL * slice_idle);
2912 ttime->ttime_samples = (7*ttime->ttime_samples + 256) / 8;
2913 ttime->ttime_total = (7*ttime->ttime_total + 256*elapsed) / 8;
2914 ttime->ttime_mean = (ttime->ttime_total + 128) / ttime->ttime_samples;
2918 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2919 struct cfq_io_cq *cic)
2921 if (cfq_cfqq_sync(cfqq)) {
2922 __cfq_update_io_thinktime(&cic->ttime, cfqd->cfq_slice_idle);
2923 __cfq_update_io_thinktime(&cfqq->service_tree->ttime,
2924 cfqd->cfq_slice_idle);
2926 #ifdef CONFIG_CFQ_GROUP_IOSCHED
2927 __cfq_update_io_thinktime(&cfqq->cfqg->ttime, cfqd->cfq_group_idle);
2932 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2936 sector_t n_sec = blk_rq_sectors(rq);
2937 if (cfqq->last_request_pos) {
2938 if (cfqq->last_request_pos < blk_rq_pos(rq))
2939 sdist = blk_rq_pos(rq) - cfqq->last_request_pos;
2941 sdist = cfqq->last_request_pos - blk_rq_pos(rq);
2944 cfqq->seek_history <<= 1;
2945 if (blk_queue_nonrot(cfqd->queue))
2946 cfqq->seek_history |= (n_sec < CFQQ_SECT_THR_NONROT);
2948 cfqq->seek_history |= (sdist > CFQQ_SEEK_THR);
2952 * Disable idle window if the process thinks too long or seeks so much that
2956 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2957 struct cfq_io_cq *cic)
2959 int old_idle, enable_idle;
2962 * Don't idle for async or idle io prio class
2964 if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
2967 enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
2969 if (cfqq->queued[0] + cfqq->queued[1] >= 4)
2970 cfq_mark_cfqq_deep(cfqq);
2972 if (cfqq->next_rq && (cfqq->next_rq->cmd_flags & REQ_NOIDLE))
2974 else if (!atomic_read(&cic->icq.ioc->nr_tasks) ||
2975 !cfqd->cfq_slice_idle ||
2976 (!cfq_cfqq_deep(cfqq) && CFQQ_SEEKY(cfqq)))
2978 else if (sample_valid(cic->ttime.ttime_samples)) {
2979 if (cic->ttime.ttime_mean > cfqd->cfq_slice_idle)
2985 if (old_idle != enable_idle) {
2986 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
2988 cfq_mark_cfqq_idle_window(cfqq);
2990 cfq_clear_cfqq_idle_window(cfqq);
2995 * Check if new_cfqq should preempt the currently active queue. Return 0 for
2996 * no or if we aren't sure, a 1 will cause a preempt.
2999 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
3002 struct cfq_queue *cfqq;
3004 cfqq = cfqd->active_queue;
3008 if (cfq_class_idle(new_cfqq))
3011 if (cfq_class_idle(cfqq))
3015 * Don't allow a non-RT request to preempt an ongoing RT cfqq timeslice.
3017 if (cfq_class_rt(cfqq) && !cfq_class_rt(new_cfqq))
3021 * if the new request is sync, but the currently running queue is
3022 * not, let the sync request have priority.
3024 if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
3027 if (new_cfqq->cfqg != cfqq->cfqg)
3030 if (cfq_slice_used(cfqq))
3033 /* Allow preemption only if we are idling on sync-noidle tree */
3034 if (cfqd->serving_type == SYNC_NOIDLE_WORKLOAD &&
3035 cfqq_type(new_cfqq) == SYNC_NOIDLE_WORKLOAD &&
3036 new_cfqq->service_tree->count == 2 &&
3037 RB_EMPTY_ROOT(&cfqq->sort_list))
3041 * So both queues are sync. Let the new request get disk time if
3042 * it's a metadata request and the current queue is doing regular IO.
3044 if ((rq->cmd_flags & REQ_PRIO) && !cfqq->prio_pending)
3048 * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
3050 if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
3053 /* An idle queue should not be idle now for some reason */
3054 if (RB_EMPTY_ROOT(&cfqq->sort_list) && !cfq_should_idle(cfqd, cfqq))
3057 if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
3061 * if this request is as-good as one we would expect from the
3062 * current cfqq, let it preempt
3064 if (cfq_rq_close(cfqd, cfqq, rq))
3071 * cfqq preempts the active queue. if we allowed preempt with no slice left,
3072 * let it have half of its nominal slice.
3074 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3076 enum wl_type_t old_type = cfqq_type(cfqd->active_queue);
3078 cfq_log_cfqq(cfqd, cfqq, "preempt");
3079 cfq_slice_expired(cfqd, 1);
3082 * workload type is changed, don't save slice, otherwise preempt
3085 if (old_type != cfqq_type(cfqq))
3086 cfqq->cfqg->saved_workload_slice = 0;
3089 * Put the new queue at the front of the of the current list,
3090 * so we know that it will be selected next.
3092 BUG_ON(!cfq_cfqq_on_rr(cfqq));
3094 cfq_service_tree_add(cfqd, cfqq, 1);
3096 cfqq->slice_end = 0;
3097 cfq_mark_cfqq_slice_new(cfqq);
3101 * Called when a new fs request (rq) is added (to cfqq). Check if there's
3102 * something we should do about it
3105 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3108 struct cfq_io_cq *cic = RQ_CIC(rq);
3111 if (rq->cmd_flags & REQ_PRIO)
3112 cfqq->prio_pending++;
3114 cfq_update_io_thinktime(cfqd, cfqq, cic);
3115 cfq_update_io_seektime(cfqd, cfqq, rq);
3116 cfq_update_idle_window(cfqd, cfqq, cic);
3118 cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
3120 if (cfqq == cfqd->active_queue) {
3122 * Remember that we saw a request from this process, but
3123 * don't start queuing just yet. Otherwise we risk seeing lots
3124 * of tiny requests, because we disrupt the normal plugging
3125 * and merging. If the request is already larger than a single
3126 * page, let it rip immediately. For that case we assume that
3127 * merging is already done. Ditto for a busy system that
3128 * has other work pending, don't risk delaying until the
3129 * idle timer unplug to continue working.
3131 if (cfq_cfqq_wait_request(cfqq)) {
3132 if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
3133 cfqd->busy_queues > 1) {
3134 cfq_del_timer(cfqd, cfqq);
3135 cfq_clear_cfqq_wait_request(cfqq);
3136 __blk_run_queue(cfqd->queue);
3138 cfq_blkiocg_update_idle_time_stats(
3139 cfqg_to_blkg(cfqq->cfqg));
3140 cfq_mark_cfqq_must_dispatch(cfqq);
3143 } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
3145 * not the active queue - expire current slice if it is
3146 * idle and has expired it's mean thinktime or this new queue
3147 * has some old slice time left and is of higher priority or
3148 * this new queue is RT and the current one is BE
3150 cfq_preempt_queue(cfqd, cfqq);
3151 __blk_run_queue(cfqd->queue);
3155 static void cfq_insert_request(struct request_queue *q, struct request *rq)
3157 struct cfq_data *cfqd = q->elevator->elevator_data;
3158 struct cfq_queue *cfqq = RQ_CFQQ(rq);
3160 cfq_log_cfqq(cfqd, cfqq, "insert_request");
3161 cfq_init_prio_data(cfqq, RQ_CIC(rq)->icq.ioc);
3163 rq_set_fifo_time(rq, jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)]);
3164 list_add_tail(&rq->queuelist, &cfqq->fifo);
3166 cfq_blkiocg_update_io_add_stats(cfqg_to_blkg(RQ_CFQG(rq)),
3167 cfqg_to_blkg(cfqd->serving_group),
3168 rq_data_dir(rq), rq_is_sync(rq));
3169 cfq_rq_enqueued(cfqd, cfqq, rq);
3173 * Update hw_tag based on peak queue depth over 50 samples under
3176 static void cfq_update_hw_tag(struct cfq_data *cfqd)
3178 struct cfq_queue *cfqq = cfqd->active_queue;
3180 if (cfqd->rq_in_driver > cfqd->hw_tag_est_depth)
3181 cfqd->hw_tag_est_depth = cfqd->rq_in_driver;
3183 if (cfqd->hw_tag == 1)
3186 if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
3187 cfqd->rq_in_driver <= CFQ_HW_QUEUE_MIN)
3191 * If active queue hasn't enough requests and can idle, cfq might not
3192 * dispatch sufficient requests to hardware. Don't zero hw_tag in this
3195 if (cfqq && cfq_cfqq_idle_window(cfqq) &&
3196 cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] <
3197 CFQ_HW_QUEUE_MIN && cfqd->rq_in_driver < CFQ_HW_QUEUE_MIN)
3200 if (cfqd->hw_tag_samples++ < 50)
3203 if (cfqd->hw_tag_est_depth >= CFQ_HW_QUEUE_MIN)
3209 static bool cfq_should_wait_busy(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3211 struct cfq_io_cq *cic = cfqd->active_cic;
3213 /* If the queue already has requests, don't wait */
3214 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3217 /* If there are other queues in the group, don't wait */
3218 if (cfqq->cfqg->nr_cfqq > 1)
3221 /* the only queue in the group, but think time is big */
3222 if (cfq_io_thinktime_big(cfqd, &cfqq->cfqg->ttime, true))
3225 if (cfq_slice_used(cfqq))
3228 /* if slice left is less than think time, wait busy */
3229 if (cic && sample_valid(cic->ttime.ttime_samples)
3230 && (cfqq->slice_end - jiffies < cic->ttime.ttime_mean))
3234 * If think times is less than a jiffy than ttime_mean=0 and above
3235 * will not be true. It might happen that slice has not expired yet
3236 * but will expire soon (4-5 ns) during select_queue(). To cover the
3237 * case where think time is less than a jiffy, mark the queue wait
3238 * busy if only 1 jiffy is left in the slice.
3240 if (cfqq->slice_end - jiffies == 1)
3246 static void cfq_completed_request(struct request_queue *q, struct request *rq)
3248 struct cfq_queue *cfqq = RQ_CFQQ(rq);
3249 struct cfq_data *cfqd = cfqq->cfqd;
3250 const int sync = rq_is_sync(rq);
3254 cfq_log_cfqq(cfqd, cfqq, "complete rqnoidle %d",
3255 !!(rq->cmd_flags & REQ_NOIDLE));
3257 cfq_update_hw_tag(cfqd);
3259 WARN_ON(!cfqd->rq_in_driver);
3260 WARN_ON(!cfqq->dispatched);
3261 cfqd->rq_in_driver--;
3263 (RQ_CFQG(rq))->dispatched--;
3264 cfq_blkiocg_update_completion_stats(cfqg_to_blkg(cfqq->cfqg),
3265 rq_start_time_ns(rq), rq_io_start_time_ns(rq),
3266 rq_data_dir(rq), rq_is_sync(rq));
3268 cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]--;
3271 struct cfq_rb_root *service_tree;
3273 RQ_CIC(rq)->ttime.last_end_request = now;
3275 if (cfq_cfqq_on_rr(cfqq))
3276 service_tree = cfqq->service_tree;
3278 service_tree = service_tree_for(cfqq->cfqg,
3279 cfqq_prio(cfqq), cfqq_type(cfqq));
3280 service_tree->ttime.last_end_request = now;
3281 if (!time_after(rq->start_time + cfqd->cfq_fifo_expire[1], now))
3282 cfqd->last_delayed_sync = now;
3285 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3286 cfqq->cfqg->ttime.last_end_request = now;
3290 * If this is the active queue, check if it needs to be expired,
3291 * or if we want to idle in case it has no pending requests.
3293 if (cfqd->active_queue == cfqq) {
3294 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
3296 if (cfq_cfqq_slice_new(cfqq)) {
3297 cfq_set_prio_slice(cfqd, cfqq);
3298 cfq_clear_cfqq_slice_new(cfqq);
3302 * Should we wait for next request to come in before we expire
3305 if (cfq_should_wait_busy(cfqd, cfqq)) {
3306 unsigned long extend_sl = cfqd->cfq_slice_idle;
3307 if (!cfqd->cfq_slice_idle)
3308 extend_sl = cfqd->cfq_group_idle;
3309 cfqq->slice_end = jiffies + extend_sl;
3310 cfq_mark_cfqq_wait_busy(cfqq);
3311 cfq_log_cfqq(cfqd, cfqq, "will busy wait");
3315 * Idling is not enabled on:
3317 * - idle-priority queues
3319 * - queues with still some requests queued
3320 * - when there is a close cooperator
3322 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
3323 cfq_slice_expired(cfqd, 1);
3324 else if (sync && cfqq_empty &&
3325 !cfq_close_cooperator(cfqd, cfqq)) {
3326 cfq_arm_slice_timer(cfqd);
3330 if (!cfqd->rq_in_driver)
3331 cfq_schedule_dispatch(cfqd);
3334 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
3336 if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
3337 cfq_mark_cfqq_must_alloc_slice(cfqq);
3338 return ELV_MQUEUE_MUST;
3341 return ELV_MQUEUE_MAY;
3344 static int cfq_may_queue(struct request_queue *q, int rw)
3346 struct cfq_data *cfqd = q->elevator->elevator_data;
3347 struct task_struct *tsk = current;
3348 struct cfq_io_cq *cic;
3349 struct cfq_queue *cfqq;
3352 * don't force setup of a queue from here, as a call to may_queue
3353 * does not necessarily imply that a request actually will be queued.
3354 * so just lookup a possibly existing queue, or return 'may queue'
3357 cic = cfq_cic_lookup(cfqd, tsk->io_context);
3359 return ELV_MQUEUE_MAY;
3361 cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
3363 cfq_init_prio_data(cfqq, cic->icq.ioc);
3365 return __cfq_may_queue(cfqq);
3368 return ELV_MQUEUE_MAY;
3372 * queue lock held here
3374 static void cfq_put_request(struct request *rq)
3376 struct cfq_queue *cfqq = RQ_CFQQ(rq);
3379 const int rw = rq_data_dir(rq);
3381 BUG_ON(!cfqq->allocated[rw]);
3382 cfqq->allocated[rw]--;
3384 /* Put down rq reference on cfqg */
3385 cfq_put_cfqg(RQ_CFQG(rq));
3386 rq->elv.priv[0] = NULL;
3387 rq->elv.priv[1] = NULL;
3389 cfq_put_queue(cfqq);
3393 static struct cfq_queue *
3394 cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_cq *cic,
3395 struct cfq_queue *cfqq)
3397 cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq);
3398 cic_set_cfqq(cic, cfqq->new_cfqq, 1);
3399 cfq_mark_cfqq_coop(cfqq->new_cfqq);
3400 cfq_put_queue(cfqq);
3401 return cic_to_cfqq(cic, 1);
3405 * Returns NULL if a new cfqq should be allocated, or the old cfqq if this
3406 * was the last process referring to said cfqq.
3408 static struct cfq_queue *
3409 split_cfqq(struct cfq_io_cq *cic, struct cfq_queue *cfqq)
3411 if (cfqq_process_refs(cfqq) == 1) {
3412 cfqq->pid = current->pid;
3413 cfq_clear_cfqq_coop(cfqq);
3414 cfq_clear_cfqq_split_coop(cfqq);
3418 cic_set_cfqq(cic, NULL, 1);
3420 cfq_put_cooperator(cfqq);
3422 cfq_put_queue(cfqq);
3426 * Allocate cfq data structures associated with this request.
3429 cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
3431 struct cfq_data *cfqd = q->elevator->elevator_data;
3432 struct cfq_io_cq *cic = icq_to_cic(rq->elv.icq);
3433 const int rw = rq_data_dir(rq);
3434 const bool is_sync = rq_is_sync(rq);
3435 struct cfq_queue *cfqq;
3436 unsigned int changed;
3438 might_sleep_if(gfp_mask & __GFP_WAIT);
3440 spin_lock_irq(q->queue_lock);
3442 /* handle changed notifications */
3443 changed = icq_get_changed(&cic->icq);
3444 if (unlikely(changed & ICQ_IOPRIO_CHANGED))
3445 changed_ioprio(cic);
3446 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3447 if (unlikely(changed & ICQ_CGROUP_CHANGED))
3448 changed_cgroup(cic);
3452 cfqq = cic_to_cfqq(cic, is_sync);
3453 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
3454 cfqq = cfq_get_queue(cfqd, is_sync, cic->icq.ioc, gfp_mask);
3455 cic_set_cfqq(cic, cfqq, is_sync);
3458 * If the queue was seeky for too long, break it apart.
3460 if (cfq_cfqq_coop(cfqq) && cfq_cfqq_split_coop(cfqq)) {
3461 cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq");
3462 cfqq = split_cfqq(cic, cfqq);
3468 * Check to see if this queue is scheduled to merge with
3469 * another, closely cooperating queue. The merging of
3470 * queues happens here as it must be done in process context.
3471 * The reference on new_cfqq was taken in merge_cfqqs.
3474 cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq);
3477 cfqq->allocated[rw]++;
3480 rq->elv.priv[0] = cfqq;
3481 rq->elv.priv[1] = cfq_ref_get_cfqg(cfqq->cfqg);
3482 spin_unlock_irq(q->queue_lock);
3486 static void cfq_kick_queue(struct work_struct *work)
3488 struct cfq_data *cfqd =
3489 container_of(work, struct cfq_data, unplug_work);
3490 struct request_queue *q = cfqd->queue;
3492 spin_lock_irq(q->queue_lock);
3493 __blk_run_queue(cfqd->queue);
3494 spin_unlock_irq(q->queue_lock);
3498 * Timer running if the active_queue is currently idling inside its time slice
3500 static void cfq_idle_slice_timer(unsigned long data)
3502 struct cfq_data *cfqd = (struct cfq_data *) data;
3503 struct cfq_queue *cfqq;
3504 unsigned long flags;
3507 cfq_log(cfqd, "idle timer fired");
3509 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
3511 cfqq = cfqd->active_queue;
3516 * We saw a request before the queue expired, let it through
3518 if (cfq_cfqq_must_dispatch(cfqq))
3524 if (cfq_slice_used(cfqq))
3528 * only expire and reinvoke request handler, if there are
3529 * other queues with pending requests
3531 if (!cfqd->busy_queues)
3535 * not expired and it has a request pending, let it dispatch
3537 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3541 * Queue depth flag is reset only when the idle didn't succeed
3543 cfq_clear_cfqq_deep(cfqq);
3546 cfq_slice_expired(cfqd, timed_out);
3548 cfq_schedule_dispatch(cfqd);
3550 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
3553 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
3555 del_timer_sync(&cfqd->idle_slice_timer);
3556 cancel_work_sync(&cfqd->unplug_work);
3559 static void cfq_put_async_queues(struct cfq_data *cfqd)
3563 for (i = 0; i < IOPRIO_BE_NR; i++) {
3564 if (cfqd->async_cfqq[0][i])
3565 cfq_put_queue(cfqd->async_cfqq[0][i]);
3566 if (cfqd->async_cfqq[1][i])
3567 cfq_put_queue(cfqd->async_cfqq[1][i]);
3570 if (cfqd->async_idle_cfqq)
3571 cfq_put_queue(cfqd->async_idle_cfqq);
3574 static void cfq_exit_queue(struct elevator_queue *e)
3576 struct cfq_data *cfqd = e->elevator_data;
3577 struct request_queue *q = cfqd->queue;
3580 cfq_shutdown_timer_wq(cfqd);
3582 spin_lock_irq(q->queue_lock);
3584 if (cfqd->active_queue)
3585 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
3587 cfq_put_async_queues(cfqd);
3588 cfq_release_cfq_groups(cfqd);
3591 * If there are groups which we could not unlink from blkcg list,
3592 * wait for a rcu period for them to be freed.
3594 if (cfqd->nr_blkcg_linked_grps)
3597 spin_unlock_irq(q->queue_lock);
3599 cfq_shutdown_timer_wq(cfqd);
3602 * Wait for cfqg->blkg->key accessors to exit their grace periods.
3603 * Do this wait only if there are other unlinked groups out
3604 * there. This can happen if cgroup deletion path claimed the
3605 * responsibility of cleaning up a group before queue cleanup code
3608 * Do not call synchronize_rcu() unconditionally as there are drivers
3609 * which create/delete request queue hundreds of times during scan/boot
3610 * and synchronize_rcu() can take significant time and slow down boot.
3615 #ifndef CONFIG_CFQ_GROUP_IOSCHED
3616 kfree(cfqd->root_group);
3621 static int cfq_init_queue(struct request_queue *q)
3623 struct cfq_data *cfqd;
3624 struct blkio_group *blkg __maybe_unused;
3627 cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
3632 q->elevator->elevator_data = cfqd;
3634 /* Init root service tree */
3635 cfqd->grp_service_tree = CFQ_RB_ROOT;
3637 /* Init root group and prefer root group over other groups by default */
3638 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3640 spin_lock_irq(q->queue_lock);
3642 blkg = blkg_lookup_create(&blkio_root_cgroup, q, BLKIO_POLICY_PROP,
3645 cfqd->root_group = blkg_to_cfqg(blkg);
3647 spin_unlock_irq(q->queue_lock);
3650 cfqd->root_group = kzalloc_node(sizeof(*cfqd->root_group),
3651 GFP_KERNEL, cfqd->queue->node);
3652 if (cfqd->root_group)
3653 cfq_init_cfqg_base(cfqd->root_group);
3655 if (!cfqd->root_group) {
3660 cfqd->root_group->weight = 2*BLKIO_WEIGHT_DEFAULT;
3663 * Not strictly needed (since RB_ROOT just clears the node and we
3664 * zeroed cfqd on alloc), but better be safe in case someone decides
3665 * to add magic to the rb code
3667 for (i = 0; i < CFQ_PRIO_LISTS; i++)
3668 cfqd->prio_trees[i] = RB_ROOT;
3671 * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
3672 * Grab a permanent reference to it, so that the normal code flow
3673 * will not attempt to free it. oom_cfqq is linked to root_group
3674 * but shouldn't hold a reference as it'll never be unlinked. Lose
3675 * the reference from linking right away.
3677 cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
3678 cfqd->oom_cfqq.ref++;
3679 cfq_link_cfqq_cfqg(&cfqd->oom_cfqq, cfqd->root_group);
3680 cfq_put_cfqg(cfqd->root_group);
3682 init_timer(&cfqd->idle_slice_timer);
3683 cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
3684 cfqd->idle_slice_timer.data = (unsigned long) cfqd;
3686 INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
3688 cfqd->cfq_quantum = cfq_quantum;
3689 cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
3690 cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
3691 cfqd->cfq_back_max = cfq_back_max;
3692 cfqd->cfq_back_penalty = cfq_back_penalty;
3693 cfqd->cfq_slice[0] = cfq_slice_async;
3694 cfqd->cfq_slice[1] = cfq_slice_sync;
3695 cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
3696 cfqd->cfq_slice_idle = cfq_slice_idle;
3697 cfqd->cfq_group_idle = cfq_group_idle;
3698 cfqd->cfq_latency = 1;
3701 * we optimistically start assuming sync ops weren't delayed in last
3702 * second, in order to have larger depth for async operations.
3704 cfqd->last_delayed_sync = jiffies - HZ;
3709 * sysfs parts below -->
3712 cfq_var_show(unsigned int var, char *page)
3714 return sprintf(page, "%d\n", var);
3718 cfq_var_store(unsigned int *var, const char *page, size_t count)
3720 char *p = (char *) page;
3722 *var = simple_strtoul(p, &p, 10);
3726 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
3727 static ssize_t __FUNC(struct elevator_queue *e, char *page) \
3729 struct cfq_data *cfqd = e->elevator_data; \
3730 unsigned int __data = __VAR; \
3732 __data = jiffies_to_msecs(__data); \
3733 return cfq_var_show(__data, (page)); \
3735 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
3736 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
3737 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
3738 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
3739 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
3740 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
3741 SHOW_FUNCTION(cfq_group_idle_show, cfqd->cfq_group_idle, 1);
3742 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
3743 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
3744 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
3745 SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
3746 #undef SHOW_FUNCTION
3748 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
3749 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
3751 struct cfq_data *cfqd = e->elevator_data; \
3752 unsigned int __data; \
3753 int ret = cfq_var_store(&__data, (page), count); \
3754 if (__data < (MIN)) \
3756 else if (__data > (MAX)) \
3759 *(__PTR) = msecs_to_jiffies(__data); \
3761 *(__PTR) = __data; \
3764 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
3765 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
3767 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
3769 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
3770 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
3772 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
3773 STORE_FUNCTION(cfq_group_idle_store, &cfqd->cfq_group_idle, 0, UINT_MAX, 1);
3774 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
3775 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
3776 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
3778 STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
3779 #undef STORE_FUNCTION
3781 #define CFQ_ATTR(name) \
3782 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
3784 static struct elv_fs_entry cfq_attrs[] = {
3786 CFQ_ATTR(fifo_expire_sync),
3787 CFQ_ATTR(fifo_expire_async),
3788 CFQ_ATTR(back_seek_max),
3789 CFQ_ATTR(back_seek_penalty),
3790 CFQ_ATTR(slice_sync),
3791 CFQ_ATTR(slice_async),
3792 CFQ_ATTR(slice_async_rq),
3793 CFQ_ATTR(slice_idle),
3794 CFQ_ATTR(group_idle),
3795 CFQ_ATTR(low_latency),
3799 static struct elevator_type iosched_cfq = {
3801 .elevator_merge_fn = cfq_merge,
3802 .elevator_merged_fn = cfq_merged_request,
3803 .elevator_merge_req_fn = cfq_merged_requests,
3804 .elevator_allow_merge_fn = cfq_allow_merge,
3805 .elevator_bio_merged_fn = cfq_bio_merged,
3806 .elevator_dispatch_fn = cfq_dispatch_requests,
3807 .elevator_add_req_fn = cfq_insert_request,
3808 .elevator_activate_req_fn = cfq_activate_request,
3809 .elevator_deactivate_req_fn = cfq_deactivate_request,
3810 .elevator_completed_req_fn = cfq_completed_request,
3811 .elevator_former_req_fn = elv_rb_former_request,
3812 .elevator_latter_req_fn = elv_rb_latter_request,
3813 .elevator_init_icq_fn = cfq_init_icq,
3814 .elevator_exit_icq_fn = cfq_exit_icq,
3815 .elevator_set_req_fn = cfq_set_request,
3816 .elevator_put_req_fn = cfq_put_request,
3817 .elevator_may_queue_fn = cfq_may_queue,
3818 .elevator_init_fn = cfq_init_queue,
3819 .elevator_exit_fn = cfq_exit_queue,
3821 .icq_size = sizeof(struct cfq_io_cq),
3822 .icq_align = __alignof__(struct cfq_io_cq),
3823 .elevator_attrs = cfq_attrs,
3824 .elevator_name = "cfq",
3825 .elevator_owner = THIS_MODULE,
3828 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3829 static struct blkio_policy_type blkio_policy_cfq = {
3831 .blkio_init_group_fn = cfq_init_blkio_group,
3832 .blkio_link_group_fn = cfq_link_blkio_group,
3833 .blkio_unlink_group_fn = cfq_unlink_blkio_group,
3834 .blkio_clear_queue_fn = cfq_clear_queue,
3835 .blkio_update_group_weight_fn = cfq_update_blkio_group_weight,
3837 .plid = BLKIO_POLICY_PROP,
3838 .pdata_size = sizeof(struct cfq_group),
3842 static int __init cfq_init(void)
3847 * could be 0 on HZ < 1000 setups
3849 if (!cfq_slice_async)
3850 cfq_slice_async = 1;
3851 if (!cfq_slice_idle)
3854 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3855 if (!cfq_group_idle)
3860 cfq_pool = KMEM_CACHE(cfq_queue, 0);
3864 ret = elv_register(&iosched_cfq);
3866 kmem_cache_destroy(cfq_pool);
3870 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3871 blkio_policy_register(&blkio_policy_cfq);
3876 static void __exit cfq_exit(void)
3878 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3879 blkio_policy_unregister(&blkio_policy_cfq);
3881 elv_unregister(&iosched_cfq);
3882 kmem_cache_destroy(cfq_pool);
3885 module_init(cfq_init);
3886 module_exit(cfq_exit);
3888 MODULE_AUTHOR("Jens Axboe");
3889 MODULE_LICENSE("GPL");
3890 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");