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>
22 /* max queue in one round of service */
23 static const int cfq_quantum = 8;
24 static const int cfq_fifo_expire[2] = { HZ / 4, HZ / 8 };
25 /* maximum backwards seek, in KiB */
26 static const int cfq_back_max = 16 * 1024;
27 /* penalty of a backwards seek */
28 static const int cfq_back_penalty = 2;
29 static const int cfq_slice_sync = HZ / 10;
30 static int cfq_slice_async = HZ / 25;
31 static const int cfq_slice_async_rq = 2;
32 static int cfq_slice_idle = HZ / 125;
33 static int cfq_group_idle = HZ / 125;
34 static const int cfq_target_latency = HZ * 3/10; /* 300 ms */
35 static const int cfq_hist_divisor = 4;
38 * offset from end of service tree
40 #define CFQ_IDLE_DELAY (HZ / 5)
43 * below this threshold, we consider thinktime immediate
45 #define CFQ_MIN_TT (2)
47 #define CFQ_SLICE_SCALE (5)
48 #define CFQ_HW_QUEUE_MIN (5)
49 #define CFQ_SERVICE_SHIFT 12
51 #define CFQQ_SEEK_THR (sector_t)(8 * 100)
52 #define CFQQ_CLOSE_THR (sector_t)(8 * 1024)
53 #define CFQQ_SECT_THR_NONROT (sector_t)(2 * 32)
54 #define CFQQ_SEEKY(cfqq) (hweight32(cfqq->seek_history) > 32/8)
57 ((struct cfq_io_context *) (rq)->elevator_private)
58 #define RQ_CFQQ(rq) (struct cfq_queue *) ((rq)->elevator_private2)
59 #define RQ_CFQG(rq) (struct cfq_group *) ((rq)->elevator_private3)
61 static struct kmem_cache *cfq_pool;
62 static struct kmem_cache *cfq_ioc_pool;
64 static DEFINE_PER_CPU(unsigned long, cfq_ioc_count);
65 static struct completion *ioc_gone;
66 static DEFINE_SPINLOCK(ioc_gone_lock);
68 static DEFINE_SPINLOCK(cic_index_lock);
69 static DEFINE_IDA(cic_index_ida);
71 #define CFQ_PRIO_LISTS IOPRIO_BE_NR
72 #define cfq_class_idle(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
73 #define cfq_class_rt(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
75 #define sample_valid(samples) ((samples) > 80)
76 #define rb_entry_cfqg(node) rb_entry((node), struct cfq_group, rb_node)
79 * Most of our rbtree usage is for sorting with min extraction, so
80 * if we cache the leftmost node we don't have to walk down the tree
81 * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should
82 * move this into the elevator for the rq sorting as well.
88 unsigned total_weight;
91 #define CFQ_RB_ROOT (struct cfq_rb_root) { .rb = RB_ROOT, .left = NULL, \
92 .count = 0, .min_vdisktime = 0, }
95 * Per process-grouping structure
100 /* various state flags, see below */
102 /* parent cfq_data */
103 struct cfq_data *cfqd;
104 /* service_tree member */
105 struct rb_node rb_node;
106 /* service_tree key */
107 unsigned long rb_key;
108 /* prio tree member */
109 struct rb_node p_node;
110 /* prio tree root we belong to, if any */
111 struct rb_root *p_root;
112 /* sorted list of pending requests */
113 struct rb_root sort_list;
114 /* if fifo isn't expired, next request to serve */
115 struct request *next_rq;
116 /* requests queued in sort_list */
118 /* currently allocated requests */
120 /* fifo list of requests in sort_list */
121 struct list_head fifo;
123 /* time when queue got scheduled in to dispatch first request. */
124 unsigned long dispatch_start;
125 unsigned int allocated_slice;
126 unsigned int slice_dispatch;
127 /* time when first request from queue completed and slice started. */
128 unsigned long slice_start;
129 unsigned long slice_end;
132 /* pending metadata requests */
134 /* number of requests that are on the dispatch list or inside driver */
137 /* io prio of this group */
138 unsigned short ioprio, org_ioprio;
139 unsigned short ioprio_class, org_ioprio_class;
144 sector_t last_request_pos;
146 struct cfq_rb_root *service_tree;
147 struct cfq_queue *new_cfqq;
148 struct cfq_group *cfqg;
149 struct cfq_group *orig_cfqg;
150 /* Number of sectors dispatched from queue in single dispatch round */
151 unsigned long nr_sectors;
155 * First index in the service_trees.
156 * IDLE is handled separately, so it has negative index
166 * Second index in the service_trees.
170 SYNC_NOIDLE_WORKLOAD = 1,
174 /* This is per cgroup per device grouping structure */
176 /* group service_tree member */
177 struct rb_node rb_node;
179 /* group service_tree key */
184 /* number of cfqq currently on this group */
188 * Per group busy queus average. Useful for workload slice calc. We
189 * create the array for each prio class but at run time it is used
190 * only for RT and BE class and slot for IDLE class remains unused.
191 * This is primarily done to avoid confusion and a gcc warning.
193 unsigned int busy_queues_avg[CFQ_PRIO_NR];
195 * rr lists of queues with requests. We maintain service trees for
196 * RT and BE classes. These trees are subdivided in subclasses
197 * of SYNC, SYNC_NOIDLE and ASYNC based on workload type. For IDLE
198 * class there is no subclassification and all the cfq queues go on
199 * a single tree service_tree_idle.
200 * Counts are embedded in the cfq_rb_root
202 struct cfq_rb_root service_trees[2][3];
203 struct cfq_rb_root service_tree_idle;
205 unsigned long saved_workload_slice;
206 enum wl_type_t saved_workload;
207 enum wl_prio_t saved_serving_prio;
208 struct blkio_group blkg;
209 #ifdef CONFIG_CFQ_GROUP_IOSCHED
210 struct hlist_node cfqd_node;
213 /* number of requests that are on the dispatch list or inside driver */
218 * Per block device queue structure
221 struct request_queue *queue;
222 /* Root service tree for cfq_groups */
223 struct cfq_rb_root grp_service_tree;
224 struct cfq_group root_group;
227 * The priority currently being served
229 enum wl_prio_t serving_prio;
230 enum wl_type_t serving_type;
231 unsigned long workload_expires;
232 struct cfq_group *serving_group;
235 * Each priority tree is sorted by next_request position. These
236 * trees are used when determining if two or more queues are
237 * interleaving requests (see cfq_close_cooperator).
239 struct rb_root prio_trees[CFQ_PRIO_LISTS];
241 unsigned int busy_queues;
247 * queue-depth detection
253 * -1 => indeterminate, (cfq will behave as if NCQ is present, to allow better detection)
254 * 1 => NCQ is present (hw_tag_est_depth is the estimated max depth)
257 int hw_tag_est_depth;
258 unsigned int hw_tag_samples;
261 * idle window management
263 struct timer_list idle_slice_timer;
264 struct work_struct unplug_work;
266 struct cfq_queue *active_queue;
267 struct cfq_io_context *active_cic;
270 * async queue for each priority case
272 struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
273 struct cfq_queue *async_idle_cfqq;
275 sector_t last_position;
278 * tunables, see top of file
280 unsigned int cfq_quantum;
281 unsigned int cfq_fifo_expire[2];
282 unsigned int cfq_back_penalty;
283 unsigned int cfq_back_max;
284 unsigned int cfq_slice[2];
285 unsigned int cfq_slice_async_rq;
286 unsigned int cfq_slice_idle;
287 unsigned int cfq_group_idle;
288 unsigned int cfq_latency;
289 unsigned int cfq_group_isolation;
291 unsigned int cic_index;
292 struct list_head cic_list;
295 * Fallback dummy cfqq for extreme OOM conditions
297 struct cfq_queue oom_cfqq;
299 unsigned long last_delayed_sync;
301 /* List of cfq groups being managed on this device*/
302 struct hlist_head cfqg_list;
306 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd);
308 static struct cfq_rb_root *service_tree_for(struct cfq_group *cfqg,
315 if (prio == IDLE_WORKLOAD)
316 return &cfqg->service_tree_idle;
318 return &cfqg->service_trees[prio][type];
321 enum cfqq_state_flags {
322 CFQ_CFQQ_FLAG_on_rr = 0, /* on round-robin busy list */
323 CFQ_CFQQ_FLAG_wait_request, /* waiting for a request */
324 CFQ_CFQQ_FLAG_must_dispatch, /* must be allowed a dispatch */
325 CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
326 CFQ_CFQQ_FLAG_fifo_expire, /* FIFO checked in this slice */
327 CFQ_CFQQ_FLAG_idle_window, /* slice idling enabled */
328 CFQ_CFQQ_FLAG_prio_changed, /* task priority has changed */
329 CFQ_CFQQ_FLAG_slice_new, /* no requests dispatched in slice */
330 CFQ_CFQQ_FLAG_sync, /* synchronous queue */
331 CFQ_CFQQ_FLAG_coop, /* cfqq is shared */
332 CFQ_CFQQ_FLAG_split_coop, /* shared cfqq will be splitted */
333 CFQ_CFQQ_FLAG_deep, /* sync cfqq experienced large depth */
334 CFQ_CFQQ_FLAG_wait_busy, /* Waiting for next request */
337 #define CFQ_CFQQ_FNS(name) \
338 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \
340 (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name); \
342 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \
344 (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \
346 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \
348 return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \
352 CFQ_CFQQ_FNS(wait_request);
353 CFQ_CFQQ_FNS(must_dispatch);
354 CFQ_CFQQ_FNS(must_alloc_slice);
355 CFQ_CFQQ_FNS(fifo_expire);
356 CFQ_CFQQ_FNS(idle_window);
357 CFQ_CFQQ_FNS(prio_changed);
358 CFQ_CFQQ_FNS(slice_new);
361 CFQ_CFQQ_FNS(split_coop);
363 CFQ_CFQQ_FNS(wait_busy);
366 #ifdef CONFIG_CFQ_GROUP_IOSCHED
367 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
368 blk_add_trace_msg((cfqd)->queue, "cfq%d%c %s " fmt, (cfqq)->pid, \
369 cfq_cfqq_sync((cfqq)) ? 'S' : 'A', \
370 blkg_path(&(cfqq)->cfqg->blkg), ##args);
372 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...) \
373 blk_add_trace_msg((cfqd)->queue, "%s " fmt, \
374 blkg_path(&(cfqg)->blkg), ##args); \
377 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
378 blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args)
379 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...) do {} while (0);
381 #define cfq_log(cfqd, fmt, args...) \
382 blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
384 /* Traverses through cfq group service trees */
385 #define for_each_cfqg_st(cfqg, i, j, st) \
386 for (i = 0; i <= IDLE_WORKLOAD; i++) \
387 for (j = 0, st = i < IDLE_WORKLOAD ? &cfqg->service_trees[i][j]\
388 : &cfqg->service_tree_idle; \
389 (i < IDLE_WORKLOAD && j <= SYNC_WORKLOAD) || \
390 (i == IDLE_WORKLOAD && j == 0); \
391 j++, st = i < IDLE_WORKLOAD ? \
392 &cfqg->service_trees[i][j]: NULL) \
395 static inline bool iops_mode(struct cfq_data *cfqd)
398 * If we are not idling on queues and it is a NCQ drive, parallel
399 * execution of requests is on and measuring time is not possible
400 * in most of the cases until and unless we drive shallower queue
401 * depths and that becomes a performance bottleneck. In such cases
402 * switch to start providing fairness in terms of number of IOs.
404 if (!cfqd->cfq_slice_idle && cfqd->hw_tag)
410 static inline enum wl_prio_t cfqq_prio(struct cfq_queue *cfqq)
412 if (cfq_class_idle(cfqq))
413 return IDLE_WORKLOAD;
414 if (cfq_class_rt(cfqq))
420 static enum wl_type_t cfqq_type(struct cfq_queue *cfqq)
422 if (!cfq_cfqq_sync(cfqq))
423 return ASYNC_WORKLOAD;
424 if (!cfq_cfqq_idle_window(cfqq))
425 return SYNC_NOIDLE_WORKLOAD;
426 return SYNC_WORKLOAD;
429 static inline int cfq_group_busy_queues_wl(enum wl_prio_t wl,
430 struct cfq_data *cfqd,
431 struct cfq_group *cfqg)
433 if (wl == IDLE_WORKLOAD)
434 return cfqg->service_tree_idle.count;
436 return cfqg->service_trees[wl][ASYNC_WORKLOAD].count
437 + cfqg->service_trees[wl][SYNC_NOIDLE_WORKLOAD].count
438 + cfqg->service_trees[wl][SYNC_WORKLOAD].count;
441 static inline int cfqg_busy_async_queues(struct cfq_data *cfqd,
442 struct cfq_group *cfqg)
444 return cfqg->service_trees[RT_WORKLOAD][ASYNC_WORKLOAD].count
445 + cfqg->service_trees[BE_WORKLOAD][ASYNC_WORKLOAD].count;
448 static void cfq_dispatch_insert(struct request_queue *, struct request *);
449 static struct cfq_queue *cfq_get_queue(struct cfq_data *, bool,
450 struct io_context *, gfp_t);
451 static struct cfq_io_context *cfq_cic_lookup(struct cfq_data *,
452 struct io_context *);
454 static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_context *cic,
457 return cic->cfqq[is_sync];
460 static inline void cic_set_cfqq(struct cfq_io_context *cic,
461 struct cfq_queue *cfqq, bool is_sync)
463 cic->cfqq[is_sync] = cfqq;
466 #define CIC_DEAD_KEY 1ul
467 #define CIC_DEAD_INDEX_SHIFT 1
469 static inline void *cfqd_dead_key(struct cfq_data *cfqd)
471 return (void *)(cfqd->cic_index << CIC_DEAD_INDEX_SHIFT | CIC_DEAD_KEY);
474 static inline struct cfq_data *cic_to_cfqd(struct cfq_io_context *cic)
476 struct cfq_data *cfqd = cic->key;
478 if (unlikely((unsigned long) cfqd & CIC_DEAD_KEY))
485 * We regard a request as SYNC, if it's either a read or has the SYNC bit
486 * set (in which case it could also be direct WRITE).
488 static inline bool cfq_bio_sync(struct bio *bio)
490 return bio_data_dir(bio) == READ || (bio->bi_rw & REQ_SYNC);
494 * scheduler run of queue, if there are requests pending and no one in the
495 * driver that will restart queueing
497 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
499 if (cfqd->busy_queues) {
500 cfq_log(cfqd, "schedule dispatch");
501 kblockd_schedule_work(cfqd->queue, &cfqd->unplug_work);
505 static int cfq_queue_empty(struct request_queue *q)
507 struct cfq_data *cfqd = q->elevator->elevator_data;
509 return !cfqd->rq_queued;
513 * Scale schedule slice based on io priority. Use the sync time slice only
514 * if a queue is marked sync and has sync io queued. A sync queue with async
515 * io only, should not get full sync slice length.
517 static inline int cfq_prio_slice(struct cfq_data *cfqd, bool sync,
520 const int base_slice = cfqd->cfq_slice[sync];
522 WARN_ON(prio >= IOPRIO_BE_NR);
524 return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
528 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
530 return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
533 static inline u64 cfq_scale_slice(unsigned long delta, struct cfq_group *cfqg)
535 u64 d = delta << CFQ_SERVICE_SHIFT;
537 d = d * BLKIO_WEIGHT_DEFAULT;
538 do_div(d, cfqg->weight);
542 static inline u64 max_vdisktime(u64 min_vdisktime, u64 vdisktime)
544 s64 delta = (s64)(vdisktime - min_vdisktime);
546 min_vdisktime = vdisktime;
548 return min_vdisktime;
551 static inline u64 min_vdisktime(u64 min_vdisktime, u64 vdisktime)
553 s64 delta = (s64)(vdisktime - min_vdisktime);
555 min_vdisktime = vdisktime;
557 return min_vdisktime;
560 static void update_min_vdisktime(struct cfq_rb_root *st)
562 u64 vdisktime = st->min_vdisktime;
563 struct cfq_group *cfqg;
566 cfqg = rb_entry_cfqg(st->left);
567 vdisktime = min_vdisktime(vdisktime, cfqg->vdisktime);
570 st->min_vdisktime = max_vdisktime(st->min_vdisktime, vdisktime);
574 * get averaged number of queues of RT/BE priority.
575 * average is updated, with a formula that gives more weight to higher numbers,
576 * to quickly follows sudden increases and decrease slowly
579 static inline unsigned cfq_group_get_avg_queues(struct cfq_data *cfqd,
580 struct cfq_group *cfqg, bool rt)
582 unsigned min_q, max_q;
583 unsigned mult = cfq_hist_divisor - 1;
584 unsigned round = cfq_hist_divisor / 2;
585 unsigned busy = cfq_group_busy_queues_wl(rt, cfqd, cfqg);
587 min_q = min(cfqg->busy_queues_avg[rt], busy);
588 max_q = max(cfqg->busy_queues_avg[rt], busy);
589 cfqg->busy_queues_avg[rt] = (mult * max_q + min_q + round) /
591 return cfqg->busy_queues_avg[rt];
594 static inline unsigned
595 cfq_group_slice(struct cfq_data *cfqd, struct cfq_group *cfqg)
597 struct cfq_rb_root *st = &cfqd->grp_service_tree;
599 return cfq_target_latency * cfqg->weight / st->total_weight;
603 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
605 unsigned slice = cfq_prio_to_slice(cfqd, cfqq);
606 if (cfqd->cfq_latency) {
608 * interested queues (we consider only the ones with the same
609 * priority class in the cfq group)
611 unsigned iq = cfq_group_get_avg_queues(cfqd, cfqq->cfqg,
613 unsigned sync_slice = cfqd->cfq_slice[1];
614 unsigned expect_latency = sync_slice * iq;
615 unsigned group_slice = cfq_group_slice(cfqd, cfqq->cfqg);
617 if (expect_latency > group_slice) {
618 unsigned base_low_slice = 2 * cfqd->cfq_slice_idle;
619 /* scale low_slice according to IO priority
620 * and sync vs async */
622 min(slice, base_low_slice * slice / sync_slice);
623 /* the adapted slice value is scaled to fit all iqs
624 * into the target latency */
625 slice = max(slice * group_slice / expect_latency,
629 cfqq->slice_start = jiffies;
630 cfqq->slice_end = jiffies + slice;
631 cfqq->allocated_slice = slice;
632 cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
636 * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
637 * isn't valid until the first request from the dispatch is activated
638 * and the slice time set.
640 static inline bool cfq_slice_used(struct cfq_queue *cfqq)
642 if (cfq_cfqq_slice_new(cfqq))
644 if (time_before(jiffies, cfqq->slice_end))
651 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
652 * We choose the request that is closest to the head right now. Distance
653 * behind the head is penalized and only allowed to a certain extent.
655 static struct request *
656 cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2, sector_t last)
658 sector_t s1, s2, d1 = 0, d2 = 0;
659 unsigned long back_max;
660 #define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */
661 #define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */
662 unsigned wrap = 0; /* bit mask: requests behind the disk head? */
664 if (rq1 == NULL || rq1 == rq2)
669 if (rq_is_sync(rq1) && !rq_is_sync(rq2))
671 else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
673 if ((rq1->cmd_flags & REQ_META) && !(rq2->cmd_flags & REQ_META))
675 else if ((rq2->cmd_flags & REQ_META) &&
676 !(rq1->cmd_flags & REQ_META))
679 s1 = blk_rq_pos(rq1);
680 s2 = blk_rq_pos(rq2);
683 * by definition, 1KiB is 2 sectors
685 back_max = cfqd->cfq_back_max * 2;
688 * Strict one way elevator _except_ in the case where we allow
689 * short backward seeks which are biased as twice the cost of a
690 * similar forward seek.
694 else if (s1 + back_max >= last)
695 d1 = (last - s1) * cfqd->cfq_back_penalty;
697 wrap |= CFQ_RQ1_WRAP;
701 else if (s2 + back_max >= last)
702 d2 = (last - s2) * cfqd->cfq_back_penalty;
704 wrap |= CFQ_RQ2_WRAP;
706 /* Found required data */
709 * By doing switch() on the bit mask "wrap" we avoid having to
710 * check two variables for all permutations: --> faster!
713 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
729 case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
732 * Since both rqs are wrapped,
733 * start with the one that's further behind head
734 * (--> only *one* back seek required),
735 * since back seek takes more time than forward.
745 * The below is leftmost cache rbtree addon
747 static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
749 /* Service tree is empty */
754 root->left = rb_first(&root->rb);
757 return rb_entry(root->left, struct cfq_queue, rb_node);
762 static struct cfq_group *cfq_rb_first_group(struct cfq_rb_root *root)
765 root->left = rb_first(&root->rb);
768 return rb_entry_cfqg(root->left);
773 static void rb_erase_init(struct rb_node *n, struct rb_root *root)
779 static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
783 rb_erase_init(n, &root->rb);
788 * would be nice to take fifo expire time into account as well
790 static struct request *
791 cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
792 struct request *last)
794 struct rb_node *rbnext = rb_next(&last->rb_node);
795 struct rb_node *rbprev = rb_prev(&last->rb_node);
796 struct request *next = NULL, *prev = NULL;
798 BUG_ON(RB_EMPTY_NODE(&last->rb_node));
801 prev = rb_entry_rq(rbprev);
804 next = rb_entry_rq(rbnext);
806 rbnext = rb_first(&cfqq->sort_list);
807 if (rbnext && rbnext != &last->rb_node)
808 next = rb_entry_rq(rbnext);
811 return cfq_choose_req(cfqd, next, prev, blk_rq_pos(last));
814 static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
815 struct cfq_queue *cfqq)
818 * just an approximation, should be ok.
820 return (cfqq->cfqg->nr_cfqq - 1) * (cfq_prio_slice(cfqd, 1, 0) -
821 cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
825 cfqg_key(struct cfq_rb_root *st, struct cfq_group *cfqg)
827 return cfqg->vdisktime - st->min_vdisktime;
831 __cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
833 struct rb_node **node = &st->rb.rb_node;
834 struct rb_node *parent = NULL;
835 struct cfq_group *__cfqg;
836 s64 key = cfqg_key(st, cfqg);
839 while (*node != NULL) {
841 __cfqg = rb_entry_cfqg(parent);
843 if (key < cfqg_key(st, __cfqg))
844 node = &parent->rb_left;
846 node = &parent->rb_right;
852 st->left = &cfqg->rb_node;
854 rb_link_node(&cfqg->rb_node, parent, node);
855 rb_insert_color(&cfqg->rb_node, &st->rb);
859 cfq_group_service_tree_add(struct cfq_data *cfqd, struct cfq_group *cfqg)
861 struct cfq_rb_root *st = &cfqd->grp_service_tree;
862 struct cfq_group *__cfqg;
870 * Currently put the group at the end. Later implement something
871 * so that groups get lesser vtime based on their weights, so that
872 * if group does not loose all if it was not continously backlogged.
874 n = rb_last(&st->rb);
876 __cfqg = rb_entry_cfqg(n);
877 cfqg->vdisktime = __cfqg->vdisktime + CFQ_IDLE_DELAY;
879 cfqg->vdisktime = st->min_vdisktime;
881 __cfq_group_service_tree_add(st, cfqg);
883 st->total_weight += cfqg->weight;
887 cfq_group_service_tree_del(struct cfq_data *cfqd, struct cfq_group *cfqg)
889 struct cfq_rb_root *st = &cfqd->grp_service_tree;
891 BUG_ON(cfqg->nr_cfqq < 1);
894 /* If there are other cfq queues under this group, don't delete it */
898 cfq_log_cfqg(cfqd, cfqg, "del_from_rr group");
900 st->total_weight -= cfqg->weight;
901 if (!RB_EMPTY_NODE(&cfqg->rb_node))
902 cfq_rb_erase(&cfqg->rb_node, st);
903 cfqg->saved_workload_slice = 0;
904 cfq_blkiocg_update_dequeue_stats(&cfqg->blkg, 1);
907 static inline unsigned int cfq_cfqq_slice_usage(struct cfq_queue *cfqq)
909 unsigned int slice_used;
912 * Queue got expired before even a single request completed or
913 * got expired immediately after first request completion.
915 if (!cfqq->slice_start || cfqq->slice_start == jiffies) {
917 * Also charge the seek time incurred to the group, otherwise
918 * if there are mutiple queues in the group, each can dispatch
919 * a single request on seeky media and cause lots of seek time
920 * and group will never know it.
922 slice_used = max_t(unsigned, (jiffies - cfqq->dispatch_start),
925 slice_used = jiffies - cfqq->slice_start;
926 if (slice_used > cfqq->allocated_slice)
927 slice_used = cfqq->allocated_slice;
933 static void cfq_group_served(struct cfq_data *cfqd, struct cfq_group *cfqg,
934 struct cfq_queue *cfqq)
936 struct cfq_rb_root *st = &cfqd->grp_service_tree;
937 unsigned int used_sl, charge;
938 int nr_sync = cfqg->nr_cfqq - cfqg_busy_async_queues(cfqd, cfqg)
939 - cfqg->service_tree_idle.count;
942 used_sl = charge = cfq_cfqq_slice_usage(cfqq);
945 charge = cfqq->slice_dispatch;
946 else if (!cfq_cfqq_sync(cfqq) && !nr_sync)
947 charge = cfqq->allocated_slice;
949 /* Can't update vdisktime while group is on service tree */
950 cfq_rb_erase(&cfqg->rb_node, st);
951 cfqg->vdisktime += cfq_scale_slice(charge, cfqg);
952 __cfq_group_service_tree_add(st, cfqg);
954 /* This group is being expired. Save the context */
955 if (time_after(cfqd->workload_expires, jiffies)) {
956 cfqg->saved_workload_slice = cfqd->workload_expires
958 cfqg->saved_workload = cfqd->serving_type;
959 cfqg->saved_serving_prio = cfqd->serving_prio;
961 cfqg->saved_workload_slice = 0;
963 cfq_log_cfqg(cfqd, cfqg, "served: vt=%llu min_vt=%llu", cfqg->vdisktime,
965 cfq_log_cfqq(cfqq->cfqd, cfqq, "sl_used=%u disp=%u charge=%u iops=%u"
966 " sect=%u", used_sl, cfqq->slice_dispatch, charge,
967 iops_mode(cfqd), cfqq->nr_sectors);
968 cfq_blkiocg_update_timeslice_used(&cfqg->blkg, used_sl);
969 cfq_blkiocg_set_start_empty_time(&cfqg->blkg);
972 #ifdef CONFIG_CFQ_GROUP_IOSCHED
973 static inline struct cfq_group *cfqg_of_blkg(struct blkio_group *blkg)
976 return container_of(blkg, struct cfq_group, blkg);
980 void cfq_update_blkio_group_weight(void *key, struct blkio_group *blkg,
983 cfqg_of_blkg(blkg)->weight = weight;
986 static struct cfq_group *
987 cfq_find_alloc_cfqg(struct cfq_data *cfqd, struct cgroup *cgroup, int create)
989 struct blkio_cgroup *blkcg = cgroup_to_blkio_cgroup(cgroup);
990 struct cfq_group *cfqg = NULL;
993 struct cfq_rb_root *st;
994 struct backing_dev_info *bdi = &cfqd->queue->backing_dev_info;
995 unsigned int major, minor;
997 cfqg = cfqg_of_blkg(blkiocg_lookup_group(blkcg, key));
998 if (cfqg && !cfqg->blkg.dev && bdi->dev && dev_name(bdi->dev)) {
999 sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
1000 cfqg->blkg.dev = MKDEV(major, minor);
1003 if (cfqg || !create)
1006 cfqg = kzalloc_node(sizeof(*cfqg), GFP_ATOMIC, cfqd->queue->node);
1010 for_each_cfqg_st(cfqg, i, j, st)
1012 RB_CLEAR_NODE(&cfqg->rb_node);
1015 * Take the initial reference that will be released on destroy
1016 * This can be thought of a joint reference by cgroup and
1017 * elevator which will be dropped by either elevator exit
1018 * or cgroup deletion path depending on who is exiting first.
1020 atomic_set(&cfqg->ref, 1);
1023 * Add group onto cgroup list. It might happen that bdi->dev is
1024 * not initiliazed yet. Initialize this new group without major
1025 * and minor info and this info will be filled in once a new thread
1026 * comes for IO. See code above.
1029 sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
1030 cfq_blkiocg_add_blkio_group(blkcg, &cfqg->blkg, (void *)cfqd,
1031 MKDEV(major, minor));
1033 cfq_blkiocg_add_blkio_group(blkcg, &cfqg->blkg, (void *)cfqd,
1036 cfqg->weight = blkcg_get_weight(blkcg, cfqg->blkg.dev);
1038 /* Add group on cfqd list */
1039 hlist_add_head(&cfqg->cfqd_node, &cfqd->cfqg_list);
1046 * Search for the cfq group current task belongs to. If create = 1, then also
1047 * create the cfq group if it does not exist. request_queue lock must be held.
1049 static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd, int create)
1051 struct cgroup *cgroup;
1052 struct cfq_group *cfqg = NULL;
1055 cgroup = task_cgroup(current, blkio_subsys_id);
1056 cfqg = cfq_find_alloc_cfqg(cfqd, cgroup, create);
1057 if (!cfqg && create)
1058 cfqg = &cfqd->root_group;
1063 static inline struct cfq_group *cfq_ref_get_cfqg(struct cfq_group *cfqg)
1065 atomic_inc(&cfqg->ref);
1069 static void cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg)
1071 /* Currently, all async queues are mapped to root group */
1072 if (!cfq_cfqq_sync(cfqq))
1073 cfqg = &cfqq->cfqd->root_group;
1076 /* cfqq reference on cfqg */
1077 atomic_inc(&cfqq->cfqg->ref);
1080 static void cfq_put_cfqg(struct cfq_group *cfqg)
1082 struct cfq_rb_root *st;
1085 BUG_ON(atomic_read(&cfqg->ref) <= 0);
1086 if (!atomic_dec_and_test(&cfqg->ref))
1088 for_each_cfqg_st(cfqg, i, j, st)
1089 BUG_ON(!RB_EMPTY_ROOT(&st->rb));
1093 static void cfq_destroy_cfqg(struct cfq_data *cfqd, struct cfq_group *cfqg)
1095 /* Something wrong if we are trying to remove same group twice */
1096 BUG_ON(hlist_unhashed(&cfqg->cfqd_node));
1098 hlist_del_init(&cfqg->cfqd_node);
1101 * Put the reference taken at the time of creation so that when all
1102 * queues are gone, group can be destroyed.
1107 static void cfq_release_cfq_groups(struct cfq_data *cfqd)
1109 struct hlist_node *pos, *n;
1110 struct cfq_group *cfqg;
1112 hlist_for_each_entry_safe(cfqg, pos, n, &cfqd->cfqg_list, cfqd_node) {
1114 * If cgroup removal path got to blk_group first and removed
1115 * it from cgroup list, then it will take care of destroying
1118 if (!cfq_blkiocg_del_blkio_group(&cfqg->blkg))
1119 cfq_destroy_cfqg(cfqd, cfqg);
1124 * Blk cgroup controller notification saying that blkio_group object is being
1125 * delinked as associated cgroup object is going away. That also means that
1126 * no new IO will come in this group. So get rid of this group as soon as
1127 * any pending IO in the group is finished.
1129 * This function is called under rcu_read_lock(). key is the rcu protected
1130 * pointer. That means "key" is a valid cfq_data pointer as long as we are rcu
1133 * "key" was fetched from blkio_group under blkio_cgroup->lock. That means
1134 * it should not be NULL as even if elevator was exiting, cgroup deltion
1135 * path got to it first.
1137 void cfq_unlink_blkio_group(void *key, struct blkio_group *blkg)
1139 unsigned long flags;
1140 struct cfq_data *cfqd = key;
1142 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1143 cfq_destroy_cfqg(cfqd, cfqg_of_blkg(blkg));
1144 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1147 #else /* GROUP_IOSCHED */
1148 static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd, int create)
1150 return &cfqd->root_group;
1153 static inline struct cfq_group *cfq_ref_get_cfqg(struct cfq_group *cfqg)
1159 cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) {
1163 static void cfq_release_cfq_groups(struct cfq_data *cfqd) {}
1164 static inline void cfq_put_cfqg(struct cfq_group *cfqg) {}
1166 #endif /* GROUP_IOSCHED */
1169 * The cfqd->service_trees holds all pending cfq_queue's that have
1170 * requests waiting to be processed. It is sorted in the order that
1171 * we will service the queues.
1173 static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1176 struct rb_node **p, *parent;
1177 struct cfq_queue *__cfqq;
1178 unsigned long rb_key;
1179 struct cfq_rb_root *service_tree;
1182 int group_changed = 0;
1184 #ifdef CONFIG_CFQ_GROUP_IOSCHED
1185 if (!cfqd->cfq_group_isolation
1186 && cfqq_type(cfqq) == SYNC_NOIDLE_WORKLOAD
1187 && cfqq->cfqg && cfqq->cfqg != &cfqd->root_group) {
1188 /* Move this cfq to root group */
1189 cfq_log_cfqq(cfqd, cfqq, "moving to root group");
1190 if (!RB_EMPTY_NODE(&cfqq->rb_node))
1191 cfq_group_service_tree_del(cfqd, cfqq->cfqg);
1192 cfqq->orig_cfqg = cfqq->cfqg;
1193 cfqq->cfqg = &cfqd->root_group;
1194 atomic_inc(&cfqd->root_group.ref);
1196 } else if (!cfqd->cfq_group_isolation
1197 && cfqq_type(cfqq) == SYNC_WORKLOAD && cfqq->orig_cfqg) {
1198 /* cfqq is sequential now needs to go to its original group */
1199 BUG_ON(cfqq->cfqg != &cfqd->root_group);
1200 if (!RB_EMPTY_NODE(&cfqq->rb_node))
1201 cfq_group_service_tree_del(cfqd, cfqq->cfqg);
1202 cfq_put_cfqg(cfqq->cfqg);
1203 cfqq->cfqg = cfqq->orig_cfqg;
1204 cfqq->orig_cfqg = NULL;
1206 cfq_log_cfqq(cfqd, cfqq, "moved to origin group");
1210 service_tree = service_tree_for(cfqq->cfqg, cfqq_prio(cfqq),
1212 if (cfq_class_idle(cfqq)) {
1213 rb_key = CFQ_IDLE_DELAY;
1214 parent = rb_last(&service_tree->rb);
1215 if (parent && parent != &cfqq->rb_node) {
1216 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1217 rb_key += __cfqq->rb_key;
1220 } else if (!add_front) {
1222 * Get our rb key offset. Subtract any residual slice
1223 * value carried from last service. A negative resid
1224 * count indicates slice overrun, and this should position
1225 * the next service time further away in the tree.
1227 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
1228 rb_key -= cfqq->slice_resid;
1229 cfqq->slice_resid = 0;
1232 __cfqq = cfq_rb_first(service_tree);
1233 rb_key += __cfqq ? __cfqq->rb_key : jiffies;
1236 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1239 * same position, nothing more to do
1241 if (rb_key == cfqq->rb_key &&
1242 cfqq->service_tree == service_tree)
1245 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1246 cfqq->service_tree = NULL;
1251 cfqq->service_tree = service_tree;
1252 p = &service_tree->rb.rb_node;
1257 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1260 * sort by key, that represents service time.
1262 if (time_before(rb_key, __cfqq->rb_key))
1265 n = &(*p)->rb_right;
1273 service_tree->left = &cfqq->rb_node;
1275 cfqq->rb_key = rb_key;
1276 rb_link_node(&cfqq->rb_node, parent, p);
1277 rb_insert_color(&cfqq->rb_node, &service_tree->rb);
1278 service_tree->count++;
1279 if ((add_front || !new_cfqq) && !group_changed)
1281 cfq_group_service_tree_add(cfqd, cfqq->cfqg);
1284 static struct cfq_queue *
1285 cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
1286 sector_t sector, struct rb_node **ret_parent,
1287 struct rb_node ***rb_link)
1289 struct rb_node **p, *parent;
1290 struct cfq_queue *cfqq = NULL;
1298 cfqq = rb_entry(parent, struct cfq_queue, p_node);
1301 * Sort strictly based on sector. Smallest to the left,
1302 * largest to the right.
1304 if (sector > blk_rq_pos(cfqq->next_rq))
1305 n = &(*p)->rb_right;
1306 else if (sector < blk_rq_pos(cfqq->next_rq))
1314 *ret_parent = parent;
1320 static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1322 struct rb_node **p, *parent;
1323 struct cfq_queue *__cfqq;
1326 rb_erase(&cfqq->p_node, cfqq->p_root);
1327 cfqq->p_root = NULL;
1330 if (cfq_class_idle(cfqq))
1335 cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
1336 __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
1337 blk_rq_pos(cfqq->next_rq), &parent, &p);
1339 rb_link_node(&cfqq->p_node, parent, p);
1340 rb_insert_color(&cfqq->p_node, cfqq->p_root);
1342 cfqq->p_root = NULL;
1346 * Update cfqq's position in the service tree.
1348 static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1351 * Resorting requires the cfqq to be on the RR list already.
1353 if (cfq_cfqq_on_rr(cfqq)) {
1354 cfq_service_tree_add(cfqd, cfqq, 0);
1355 cfq_prio_tree_add(cfqd, cfqq);
1360 * add to busy list of queues for service, trying to be fair in ordering
1361 * the pending list according to last request service
1363 static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1365 cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
1366 BUG_ON(cfq_cfqq_on_rr(cfqq));
1367 cfq_mark_cfqq_on_rr(cfqq);
1368 cfqd->busy_queues++;
1370 cfq_resort_rr_list(cfqd, cfqq);
1374 * Called when the cfqq no longer has requests pending, remove it from
1377 static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1379 cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
1380 BUG_ON(!cfq_cfqq_on_rr(cfqq));
1381 cfq_clear_cfqq_on_rr(cfqq);
1383 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1384 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1385 cfqq->service_tree = NULL;
1388 rb_erase(&cfqq->p_node, cfqq->p_root);
1389 cfqq->p_root = NULL;
1392 cfq_group_service_tree_del(cfqd, cfqq->cfqg);
1393 BUG_ON(!cfqd->busy_queues);
1394 cfqd->busy_queues--;
1398 * rb tree support functions
1400 static void cfq_del_rq_rb(struct request *rq)
1402 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1403 const int sync = rq_is_sync(rq);
1405 BUG_ON(!cfqq->queued[sync]);
1406 cfqq->queued[sync]--;
1408 elv_rb_del(&cfqq->sort_list, rq);
1410 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) {
1412 * Queue will be deleted from service tree when we actually
1413 * expire it later. Right now just remove it from prio tree
1417 rb_erase(&cfqq->p_node, cfqq->p_root);
1418 cfqq->p_root = NULL;
1423 static void cfq_add_rq_rb(struct request *rq)
1425 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1426 struct cfq_data *cfqd = cfqq->cfqd;
1427 struct request *__alias, *prev;
1429 cfqq->queued[rq_is_sync(rq)]++;
1432 * looks a little odd, but the first insert might return an alias.
1433 * if that happens, put the alias on the dispatch list
1435 while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL)
1436 cfq_dispatch_insert(cfqd->queue, __alias);
1438 if (!cfq_cfqq_on_rr(cfqq))
1439 cfq_add_cfqq_rr(cfqd, cfqq);
1442 * check if this request is a better next-serve candidate
1444 prev = cfqq->next_rq;
1445 cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq, cfqd->last_position);
1448 * adjust priority tree position, if ->next_rq changes
1450 if (prev != cfqq->next_rq)
1451 cfq_prio_tree_add(cfqd, cfqq);
1453 BUG_ON(!cfqq->next_rq);
1456 static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
1458 elv_rb_del(&cfqq->sort_list, rq);
1459 cfqq->queued[rq_is_sync(rq)]--;
1460 cfq_blkiocg_update_io_remove_stats(&(RQ_CFQG(rq))->blkg,
1461 rq_data_dir(rq), rq_is_sync(rq));
1463 cfq_blkiocg_update_io_add_stats(&(RQ_CFQG(rq))->blkg,
1464 &cfqq->cfqd->serving_group->blkg, rq_data_dir(rq),
1468 static struct request *
1469 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
1471 struct task_struct *tsk = current;
1472 struct cfq_io_context *cic;
1473 struct cfq_queue *cfqq;
1475 cic = cfq_cic_lookup(cfqd, tsk->io_context);
1479 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
1481 sector_t sector = bio->bi_sector + bio_sectors(bio);
1483 return elv_rb_find(&cfqq->sort_list, sector);
1489 static void cfq_activate_request(struct request_queue *q, struct request *rq)
1491 struct cfq_data *cfqd = q->elevator->elevator_data;
1493 cfqd->rq_in_driver++;
1494 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
1495 cfqd->rq_in_driver);
1497 cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
1500 static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
1502 struct cfq_data *cfqd = q->elevator->elevator_data;
1504 WARN_ON(!cfqd->rq_in_driver);
1505 cfqd->rq_in_driver--;
1506 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
1507 cfqd->rq_in_driver);
1510 static void cfq_remove_request(struct request *rq)
1512 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1514 if (cfqq->next_rq == rq)
1515 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
1517 list_del_init(&rq->queuelist);
1520 cfqq->cfqd->rq_queued--;
1521 cfq_blkiocg_update_io_remove_stats(&(RQ_CFQG(rq))->blkg,
1522 rq_data_dir(rq), rq_is_sync(rq));
1523 if (rq->cmd_flags & REQ_META) {
1524 WARN_ON(!cfqq->meta_pending);
1525 cfqq->meta_pending--;
1529 static int cfq_merge(struct request_queue *q, struct request **req,
1532 struct cfq_data *cfqd = q->elevator->elevator_data;
1533 struct request *__rq;
1535 __rq = cfq_find_rq_fmerge(cfqd, bio);
1536 if (__rq && elv_rq_merge_ok(__rq, bio)) {
1538 return ELEVATOR_FRONT_MERGE;
1541 return ELEVATOR_NO_MERGE;
1544 static void cfq_merged_request(struct request_queue *q, struct request *req,
1547 if (type == ELEVATOR_FRONT_MERGE) {
1548 struct cfq_queue *cfqq = RQ_CFQQ(req);
1550 cfq_reposition_rq_rb(cfqq, req);
1554 static void cfq_bio_merged(struct request_queue *q, struct request *req,
1557 cfq_blkiocg_update_io_merged_stats(&(RQ_CFQG(req))->blkg,
1558 bio_data_dir(bio), cfq_bio_sync(bio));
1562 cfq_merged_requests(struct request_queue *q, struct request *rq,
1563 struct request *next)
1565 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1567 * reposition in fifo if next is older than rq
1569 if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
1570 time_before(rq_fifo_time(next), rq_fifo_time(rq))) {
1571 list_move(&rq->queuelist, &next->queuelist);
1572 rq_set_fifo_time(rq, rq_fifo_time(next));
1575 if (cfqq->next_rq == next)
1577 cfq_remove_request(next);
1578 cfq_blkiocg_update_io_merged_stats(&(RQ_CFQG(rq))->blkg,
1579 rq_data_dir(next), rq_is_sync(next));
1582 static int cfq_allow_merge(struct request_queue *q, struct request *rq,
1585 struct cfq_data *cfqd = q->elevator->elevator_data;
1586 struct cfq_io_context *cic;
1587 struct cfq_queue *cfqq;
1590 * Disallow merge of a sync bio into an async request.
1592 if (cfq_bio_sync(bio) && !rq_is_sync(rq))
1596 * Lookup the cfqq that this bio will be queued with. Allow
1597 * merge only if rq is queued there.
1599 cic = cfq_cic_lookup(cfqd, current->io_context);
1603 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
1604 return cfqq == RQ_CFQQ(rq);
1607 static inline void cfq_del_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1609 del_timer(&cfqd->idle_slice_timer);
1610 cfq_blkiocg_update_idle_time_stats(&cfqq->cfqg->blkg);
1613 static void __cfq_set_active_queue(struct cfq_data *cfqd,
1614 struct cfq_queue *cfqq)
1617 cfq_log_cfqq(cfqd, cfqq, "set_active wl_prio:%d wl_type:%d",
1618 cfqd->serving_prio, cfqd->serving_type);
1619 cfq_blkiocg_update_avg_queue_size_stats(&cfqq->cfqg->blkg);
1620 cfqq->slice_start = 0;
1621 cfqq->dispatch_start = jiffies;
1622 cfqq->allocated_slice = 0;
1623 cfqq->slice_end = 0;
1624 cfqq->slice_dispatch = 0;
1625 cfqq->nr_sectors = 0;
1627 cfq_clear_cfqq_wait_request(cfqq);
1628 cfq_clear_cfqq_must_dispatch(cfqq);
1629 cfq_clear_cfqq_must_alloc_slice(cfqq);
1630 cfq_clear_cfqq_fifo_expire(cfqq);
1631 cfq_mark_cfqq_slice_new(cfqq);
1633 cfq_del_timer(cfqd, cfqq);
1636 cfqd->active_queue = cfqq;
1640 * current cfqq expired its slice (or was too idle), select new one
1643 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1646 cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
1648 if (cfq_cfqq_wait_request(cfqq))
1649 cfq_del_timer(cfqd, cfqq);
1651 cfq_clear_cfqq_wait_request(cfqq);
1652 cfq_clear_cfqq_wait_busy(cfqq);
1655 * If this cfqq is shared between multiple processes, check to
1656 * make sure that those processes are still issuing I/Os within
1657 * the mean seek distance. If not, it may be time to break the
1658 * queues apart again.
1660 if (cfq_cfqq_coop(cfqq) && CFQQ_SEEKY(cfqq))
1661 cfq_mark_cfqq_split_coop(cfqq);
1664 * store what was left of this slice, if the queue idled/timed out
1666 if (timed_out && !cfq_cfqq_slice_new(cfqq)) {
1667 cfqq->slice_resid = cfqq->slice_end - jiffies;
1668 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
1671 cfq_group_served(cfqd, cfqq->cfqg, cfqq);
1673 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
1674 cfq_del_cfqq_rr(cfqd, cfqq);
1676 cfq_resort_rr_list(cfqd, cfqq);
1678 if (cfqq == cfqd->active_queue)
1679 cfqd->active_queue = NULL;
1681 if (cfqd->active_cic) {
1682 put_io_context(cfqd->active_cic->ioc);
1683 cfqd->active_cic = NULL;
1687 static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out)
1689 struct cfq_queue *cfqq = cfqd->active_queue;
1692 __cfq_slice_expired(cfqd, cfqq, timed_out);
1696 * Get next queue for service. Unless we have a queue preemption,
1697 * we'll simply select the first cfqq in the service tree.
1699 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
1701 struct cfq_rb_root *service_tree =
1702 service_tree_for(cfqd->serving_group, cfqd->serving_prio,
1703 cfqd->serving_type);
1705 if (!cfqd->rq_queued)
1708 /* There is nothing to dispatch */
1711 if (RB_EMPTY_ROOT(&service_tree->rb))
1713 return cfq_rb_first(service_tree);
1716 static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd)
1718 struct cfq_group *cfqg;
1719 struct cfq_queue *cfqq;
1721 struct cfq_rb_root *st;
1723 if (!cfqd->rq_queued)
1726 cfqg = cfq_get_next_cfqg(cfqd);
1730 for_each_cfqg_st(cfqg, i, j, st)
1731 if ((cfqq = cfq_rb_first(st)) != NULL)
1737 * Get and set a new active queue for service.
1739 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
1740 struct cfq_queue *cfqq)
1743 cfqq = cfq_get_next_queue(cfqd);
1745 __cfq_set_active_queue(cfqd, cfqq);
1749 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
1752 if (blk_rq_pos(rq) >= cfqd->last_position)
1753 return blk_rq_pos(rq) - cfqd->last_position;
1755 return cfqd->last_position - blk_rq_pos(rq);
1758 static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1761 return cfq_dist_from_last(cfqd, rq) <= CFQQ_CLOSE_THR;
1764 static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
1765 struct cfq_queue *cur_cfqq)
1767 struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
1768 struct rb_node *parent, *node;
1769 struct cfq_queue *__cfqq;
1770 sector_t sector = cfqd->last_position;
1772 if (RB_EMPTY_ROOT(root))
1776 * First, if we find a request starting at the end of the last
1777 * request, choose it.
1779 __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
1784 * If the exact sector wasn't found, the parent of the NULL leaf
1785 * will contain the closest sector.
1787 __cfqq = rb_entry(parent, struct cfq_queue, p_node);
1788 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1791 if (blk_rq_pos(__cfqq->next_rq) < sector)
1792 node = rb_next(&__cfqq->p_node);
1794 node = rb_prev(&__cfqq->p_node);
1798 __cfqq = rb_entry(node, struct cfq_queue, p_node);
1799 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1807 * cur_cfqq - passed in so that we don't decide that the current queue is
1808 * closely cooperating with itself.
1810 * So, basically we're assuming that that cur_cfqq has dispatched at least
1811 * one request, and that cfqd->last_position reflects a position on the disk
1812 * associated with the I/O issued by cur_cfqq. I'm not sure this is a valid
1815 static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
1816 struct cfq_queue *cur_cfqq)
1818 struct cfq_queue *cfqq;
1820 if (cfq_class_idle(cur_cfqq))
1822 if (!cfq_cfqq_sync(cur_cfqq))
1824 if (CFQQ_SEEKY(cur_cfqq))
1828 * Don't search priority tree if it's the only queue in the group.
1830 if (cur_cfqq->cfqg->nr_cfqq == 1)
1834 * We should notice if some of the queues are cooperating, eg
1835 * working closely on the same area of the disk. In that case,
1836 * we can group them together and don't waste time idling.
1838 cfqq = cfqq_close(cfqd, cur_cfqq);
1842 /* If new queue belongs to different cfq_group, don't choose it */
1843 if (cur_cfqq->cfqg != cfqq->cfqg)
1847 * It only makes sense to merge sync queues.
1849 if (!cfq_cfqq_sync(cfqq))
1851 if (CFQQ_SEEKY(cfqq))
1855 * Do not merge queues of different priority classes
1857 if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq))
1864 * Determine whether we should enforce idle window for this queue.
1867 static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1869 enum wl_prio_t prio = cfqq_prio(cfqq);
1870 struct cfq_rb_root *service_tree = cfqq->service_tree;
1872 BUG_ON(!service_tree);
1873 BUG_ON(!service_tree->count);
1875 if (!cfqd->cfq_slice_idle)
1878 /* We never do for idle class queues. */
1879 if (prio == IDLE_WORKLOAD)
1882 /* We do for queues that were marked with idle window flag. */
1883 if (cfq_cfqq_idle_window(cfqq) &&
1884 !(blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag))
1888 * Otherwise, we do only if they are the last ones
1889 * in their service tree.
1891 if (service_tree->count == 1 && cfq_cfqq_sync(cfqq))
1893 cfq_log_cfqq(cfqd, cfqq, "Not idling. st->count:%d",
1894 service_tree->count);
1898 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
1900 struct cfq_queue *cfqq = cfqd->active_queue;
1901 struct cfq_io_context *cic;
1902 unsigned long sl, group_idle = 0;
1905 * SSD device without seek penalty, disable idling. But only do so
1906 * for devices that support queuing, otherwise we still have a problem
1907 * with sync vs async workloads.
1909 if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
1912 WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
1913 WARN_ON(cfq_cfqq_slice_new(cfqq));
1916 * idle is disabled, either manually or by past process history
1918 if (!cfq_should_idle(cfqd, cfqq)) {
1919 /* no queue idling. Check for group idling */
1920 if (cfqd->cfq_group_idle)
1921 group_idle = cfqd->cfq_group_idle;
1927 * still active requests from this queue, don't idle
1929 if (cfqq->dispatched)
1933 * task has exited, don't wait
1935 cic = cfqd->active_cic;
1936 if (!cic || !atomic_read(&cic->ioc->nr_tasks))
1940 * If our average think time is larger than the remaining time
1941 * slice, then don't idle. This avoids overrunning the allotted
1944 if (sample_valid(cic->ttime_samples) &&
1945 (cfqq->slice_end - jiffies < cic->ttime_mean)) {
1946 cfq_log_cfqq(cfqd, cfqq, "Not idling. think_time:%d",
1951 /* There are other queues in the group, don't do group idle */
1952 if (group_idle && cfqq->cfqg->nr_cfqq > 1)
1955 cfq_mark_cfqq_wait_request(cfqq);
1958 sl = cfqd->cfq_group_idle;
1960 sl = cfqd->cfq_slice_idle;
1962 mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
1963 cfq_blkiocg_update_set_idle_time_stats(&cfqq->cfqg->blkg);
1964 cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu group_idle: %d", sl,
1965 group_idle ? 1 : 0);
1969 * Move request from internal lists to the request queue dispatch list.
1971 static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
1973 struct cfq_data *cfqd = q->elevator->elevator_data;
1974 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1976 cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
1978 cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
1979 cfq_remove_request(rq);
1981 (RQ_CFQG(rq))->dispatched++;
1982 elv_dispatch_sort(q, rq);
1984 cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]++;
1985 cfqq->nr_sectors += blk_rq_sectors(rq);
1986 cfq_blkiocg_update_dispatch_stats(&cfqq->cfqg->blkg, blk_rq_bytes(rq),
1987 rq_data_dir(rq), rq_is_sync(rq));
1991 * return expired entry, or NULL to just start from scratch in rbtree
1993 static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
1995 struct request *rq = NULL;
1997 if (cfq_cfqq_fifo_expire(cfqq))
2000 cfq_mark_cfqq_fifo_expire(cfqq);
2002 if (list_empty(&cfqq->fifo))
2005 rq = rq_entry_fifo(cfqq->fifo.next);
2006 if (time_before(jiffies, rq_fifo_time(rq)))
2009 cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq);
2014 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2016 const int base_rq = cfqd->cfq_slice_async_rq;
2018 WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
2020 return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
2024 * Must be called with the queue_lock held.
2026 static int cfqq_process_refs(struct cfq_queue *cfqq)
2028 int process_refs, io_refs;
2030 io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE];
2031 process_refs = atomic_read(&cfqq->ref) - io_refs;
2032 BUG_ON(process_refs < 0);
2033 return process_refs;
2036 static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq)
2038 int process_refs, new_process_refs;
2039 struct cfq_queue *__cfqq;
2042 * If there are no process references on the new_cfqq, then it is
2043 * unsafe to follow the ->new_cfqq chain as other cfqq's in the
2044 * chain may have dropped their last reference (not just their
2045 * last process reference).
2047 if (!cfqq_process_refs(new_cfqq))
2050 /* Avoid a circular list and skip interim queue merges */
2051 while ((__cfqq = new_cfqq->new_cfqq)) {
2057 process_refs = cfqq_process_refs(cfqq);
2058 new_process_refs = cfqq_process_refs(new_cfqq);
2060 * If the process for the cfqq has gone away, there is no
2061 * sense in merging the queues.
2063 if (process_refs == 0 || new_process_refs == 0)
2067 * Merge in the direction of the lesser amount of work.
2069 if (new_process_refs >= process_refs) {
2070 cfqq->new_cfqq = new_cfqq;
2071 atomic_add(process_refs, &new_cfqq->ref);
2073 new_cfqq->new_cfqq = cfqq;
2074 atomic_add(new_process_refs, &cfqq->ref);
2078 static enum wl_type_t cfq_choose_wl(struct cfq_data *cfqd,
2079 struct cfq_group *cfqg, enum wl_prio_t prio)
2081 struct cfq_queue *queue;
2083 bool key_valid = false;
2084 unsigned long lowest_key = 0;
2085 enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD;
2087 for (i = 0; i <= SYNC_WORKLOAD; ++i) {
2088 /* select the one with lowest rb_key */
2089 queue = cfq_rb_first(service_tree_for(cfqg, prio, i));
2091 (!key_valid || time_before(queue->rb_key, lowest_key))) {
2092 lowest_key = queue->rb_key;
2101 static void choose_service_tree(struct cfq_data *cfqd, struct cfq_group *cfqg)
2105 struct cfq_rb_root *st;
2106 unsigned group_slice;
2109 cfqd->serving_prio = IDLE_WORKLOAD;
2110 cfqd->workload_expires = jiffies + 1;
2114 /* Choose next priority. RT > BE > IDLE */
2115 if (cfq_group_busy_queues_wl(RT_WORKLOAD, cfqd, cfqg))
2116 cfqd->serving_prio = RT_WORKLOAD;
2117 else if (cfq_group_busy_queues_wl(BE_WORKLOAD, cfqd, cfqg))
2118 cfqd->serving_prio = BE_WORKLOAD;
2120 cfqd->serving_prio = IDLE_WORKLOAD;
2121 cfqd->workload_expires = jiffies + 1;
2126 * For RT and BE, we have to choose also the type
2127 * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
2130 st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
2134 * check workload expiration, and that we still have other queues ready
2136 if (count && !time_after(jiffies, cfqd->workload_expires))
2139 /* otherwise select new workload type */
2140 cfqd->serving_type =
2141 cfq_choose_wl(cfqd, cfqg, cfqd->serving_prio);
2142 st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
2146 * the workload slice is computed as a fraction of target latency
2147 * proportional to the number of queues in that workload, over
2148 * all the queues in the same priority class
2150 group_slice = cfq_group_slice(cfqd, cfqg);
2152 slice = group_slice * count /
2153 max_t(unsigned, cfqg->busy_queues_avg[cfqd->serving_prio],
2154 cfq_group_busy_queues_wl(cfqd->serving_prio, cfqd, cfqg));
2156 if (cfqd->serving_type == ASYNC_WORKLOAD) {
2160 * Async queues are currently system wide. Just taking
2161 * proportion of queues with-in same group will lead to higher
2162 * async ratio system wide as generally root group is going
2163 * to have higher weight. A more accurate thing would be to
2164 * calculate system wide asnc/sync ratio.
2166 tmp = cfq_target_latency * cfqg_busy_async_queues(cfqd, cfqg);
2167 tmp = tmp/cfqd->busy_queues;
2168 slice = min_t(unsigned, slice, tmp);
2170 /* async workload slice is scaled down according to
2171 * the sync/async slice ratio. */
2172 slice = slice * cfqd->cfq_slice[0] / cfqd->cfq_slice[1];
2174 /* sync workload slice is at least 2 * cfq_slice_idle */
2175 slice = max(slice, 2 * cfqd->cfq_slice_idle);
2177 slice = max_t(unsigned, slice, CFQ_MIN_TT);
2178 cfq_log(cfqd, "workload slice:%d", slice);
2179 cfqd->workload_expires = jiffies + slice;
2182 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd)
2184 struct cfq_rb_root *st = &cfqd->grp_service_tree;
2185 struct cfq_group *cfqg;
2187 if (RB_EMPTY_ROOT(&st->rb))
2189 cfqg = cfq_rb_first_group(st);
2190 update_min_vdisktime(st);
2194 static void cfq_choose_cfqg(struct cfq_data *cfqd)
2196 struct cfq_group *cfqg = cfq_get_next_cfqg(cfqd);
2198 cfqd->serving_group = cfqg;
2200 /* Restore the workload type data */
2201 if (cfqg->saved_workload_slice) {
2202 cfqd->workload_expires = jiffies + cfqg->saved_workload_slice;
2203 cfqd->serving_type = cfqg->saved_workload;
2204 cfqd->serving_prio = cfqg->saved_serving_prio;
2206 cfqd->workload_expires = jiffies - 1;
2208 choose_service_tree(cfqd, cfqg);
2212 * Select a queue for service. If we have a current active queue,
2213 * check whether to continue servicing it, or retrieve and set a new one.
2215 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
2217 struct cfq_queue *cfqq, *new_cfqq = NULL;
2219 cfqq = cfqd->active_queue;
2223 if (!cfqd->rq_queued)
2227 * We were waiting for group to get backlogged. Expire the queue
2229 if (cfq_cfqq_wait_busy(cfqq) && !RB_EMPTY_ROOT(&cfqq->sort_list))
2233 * The active queue has run out of time, expire it and select new.
2235 if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq)) {
2237 * If slice had not expired at the completion of last request
2238 * we might not have turned on wait_busy flag. Don't expire
2239 * the queue yet. Allow the group to get backlogged.
2241 * The very fact that we have used the slice, that means we
2242 * have been idling all along on this queue and it should be
2243 * ok to wait for this request to complete.
2245 if (cfqq->cfqg->nr_cfqq == 1 && RB_EMPTY_ROOT(&cfqq->sort_list)
2246 && cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2250 goto check_group_idle;
2254 * The active queue has requests and isn't expired, allow it to
2257 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
2261 * If another queue has a request waiting within our mean seek
2262 * distance, let it run. The expire code will check for close
2263 * cooperators and put the close queue at the front of the service
2264 * tree. If possible, merge the expiring queue with the new cfqq.
2266 new_cfqq = cfq_close_cooperator(cfqd, cfqq);
2268 if (!cfqq->new_cfqq)
2269 cfq_setup_merge(cfqq, new_cfqq);
2274 * No requests pending. If the active queue still has requests in
2275 * flight or is idling for a new request, allow either of these
2276 * conditions to happen (or time out) before selecting a new queue.
2278 if (timer_pending(&cfqd->idle_slice_timer)) {
2284 * This is a deep seek queue, but the device is much faster than
2285 * the queue can deliver, don't idle
2287 if (CFQQ_SEEKY(cfqq) && cfq_cfqq_idle_window(cfqq) &&
2288 (cfq_cfqq_slice_new(cfqq) ||
2289 (cfqq->slice_end - jiffies > jiffies - cfqq->slice_start))) {
2290 cfq_clear_cfqq_deep(cfqq);
2291 cfq_clear_cfqq_idle_window(cfqq);
2294 if (cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2300 * If group idle is enabled and there are requests dispatched from
2301 * this group, wait for requests to complete.
2304 if (cfqd->cfq_group_idle && cfqq->cfqg->nr_cfqq == 1
2305 && cfqq->cfqg->dispatched) {
2311 cfq_slice_expired(cfqd, 0);
2314 * Current queue expired. Check if we have to switch to a new
2318 cfq_choose_cfqg(cfqd);
2320 cfqq = cfq_set_active_queue(cfqd, new_cfqq);
2325 static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
2329 while (cfqq->next_rq) {
2330 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
2334 BUG_ON(!list_empty(&cfqq->fifo));
2336 /* By default cfqq is not expired if it is empty. Do it explicitly */
2337 __cfq_slice_expired(cfqq->cfqd, cfqq, 0);
2342 * Drain our current requests. Used for barriers and when switching
2343 * io schedulers on-the-fly.
2345 static int cfq_forced_dispatch(struct cfq_data *cfqd)
2347 struct cfq_queue *cfqq;
2350 /* Expire the timeslice of the current active queue first */
2351 cfq_slice_expired(cfqd, 0);
2352 while ((cfqq = cfq_get_next_queue_forced(cfqd)) != NULL) {
2353 __cfq_set_active_queue(cfqd, cfqq);
2354 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
2357 BUG_ON(cfqd->busy_queues);
2359 cfq_log(cfqd, "forced_dispatch=%d", dispatched);
2363 static inline bool cfq_slice_used_soon(struct cfq_data *cfqd,
2364 struct cfq_queue *cfqq)
2366 /* the queue hasn't finished any request, can't estimate */
2367 if (cfq_cfqq_slice_new(cfqq))
2369 if (time_after(jiffies + cfqd->cfq_slice_idle * cfqq->dispatched,
2376 static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2378 unsigned int max_dispatch;
2381 * Drain async requests before we start sync IO
2383 if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_flight[BLK_RW_ASYNC])
2387 * If this is an async queue and we have sync IO in flight, let it wait
2389 if (cfqd->rq_in_flight[BLK_RW_SYNC] && !cfq_cfqq_sync(cfqq))
2392 max_dispatch = max_t(unsigned int, cfqd->cfq_quantum / 2, 1);
2393 if (cfq_class_idle(cfqq))
2397 * Does this cfqq already have too much IO in flight?
2399 if (cfqq->dispatched >= max_dispatch) {
2401 * idle queue must always only have a single IO in flight
2403 if (cfq_class_idle(cfqq))
2407 * We have other queues, don't allow more IO from this one
2409 if (cfqd->busy_queues > 1 && cfq_slice_used_soon(cfqd, cfqq))
2413 * Sole queue user, no limit
2415 if (cfqd->busy_queues == 1)
2419 * Normally we start throttling cfqq when cfq_quantum/2
2420 * requests have been dispatched. But we can drive
2421 * deeper queue depths at the beginning of slice
2422 * subjected to upper limit of cfq_quantum.
2424 max_dispatch = cfqd->cfq_quantum;
2428 * Async queues must wait a bit before being allowed dispatch.
2429 * We also ramp up the dispatch depth gradually for async IO,
2430 * based on the last sync IO we serviced
2432 if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
2433 unsigned long last_sync = jiffies - cfqd->last_delayed_sync;
2436 depth = last_sync / cfqd->cfq_slice[1];
2437 if (!depth && !cfqq->dispatched)
2439 if (depth < max_dispatch)
2440 max_dispatch = depth;
2444 * If we're below the current max, allow a dispatch
2446 return cfqq->dispatched < max_dispatch;
2450 * Dispatch a request from cfqq, moving them to the request queue
2453 static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2457 BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
2459 if (!cfq_may_dispatch(cfqd, cfqq))
2463 * follow expired path, else get first next available
2465 rq = cfq_check_fifo(cfqq);
2470 * insert request into driver dispatch list
2472 cfq_dispatch_insert(cfqd->queue, rq);
2474 if (!cfqd->active_cic) {
2475 struct cfq_io_context *cic = RQ_CIC(rq);
2477 atomic_long_inc(&cic->ioc->refcount);
2478 cfqd->active_cic = cic;
2485 * Find the cfqq that we need to service and move a request from that to the
2488 static int cfq_dispatch_requests(struct request_queue *q, int force)
2490 struct cfq_data *cfqd = q->elevator->elevator_data;
2491 struct cfq_queue *cfqq;
2493 if (!cfqd->busy_queues)
2496 if (unlikely(force))
2497 return cfq_forced_dispatch(cfqd);
2499 cfqq = cfq_select_queue(cfqd);
2504 * Dispatch a request from this cfqq, if it is allowed
2506 if (!cfq_dispatch_request(cfqd, cfqq))
2509 cfqq->slice_dispatch++;
2510 cfq_clear_cfqq_must_dispatch(cfqq);
2513 * expire an async queue immediately if it has used up its slice. idle
2514 * queue always expire after 1 dispatch round.
2516 if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
2517 cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
2518 cfq_class_idle(cfqq))) {
2519 cfqq->slice_end = jiffies + 1;
2520 cfq_slice_expired(cfqd, 0);
2523 cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
2528 * task holds one reference to the queue, dropped when task exits. each rq
2529 * in-flight on this queue also holds a reference, dropped when rq is freed.
2531 * Each cfq queue took a reference on the parent group. Drop it now.
2532 * queue lock must be held here.
2534 static void cfq_put_queue(struct cfq_queue *cfqq)
2536 struct cfq_data *cfqd = cfqq->cfqd;
2537 struct cfq_group *cfqg, *orig_cfqg;
2539 BUG_ON(atomic_read(&cfqq->ref) <= 0);
2541 if (!atomic_dec_and_test(&cfqq->ref))
2544 cfq_log_cfqq(cfqd, cfqq, "put_queue");
2545 BUG_ON(rb_first(&cfqq->sort_list));
2546 BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
2548 orig_cfqg = cfqq->orig_cfqg;
2550 if (unlikely(cfqd->active_queue == cfqq)) {
2551 __cfq_slice_expired(cfqd, cfqq, 0);
2552 cfq_schedule_dispatch(cfqd);
2555 BUG_ON(cfq_cfqq_on_rr(cfqq));
2556 kmem_cache_free(cfq_pool, cfqq);
2559 cfq_put_cfqg(orig_cfqg);
2563 * Must always be called with the rcu_read_lock() held
2566 __call_for_each_cic(struct io_context *ioc,
2567 void (*func)(struct io_context *, struct cfq_io_context *))
2569 struct cfq_io_context *cic;
2570 struct hlist_node *n;
2572 hlist_for_each_entry_rcu(cic, n, &ioc->cic_list, cic_list)
2577 * Call func for each cic attached to this ioc.
2580 call_for_each_cic(struct io_context *ioc,
2581 void (*func)(struct io_context *, struct cfq_io_context *))
2584 __call_for_each_cic(ioc, func);
2588 static void cfq_cic_free_rcu(struct rcu_head *head)
2590 struct cfq_io_context *cic;
2592 cic = container_of(head, struct cfq_io_context, rcu_head);
2594 kmem_cache_free(cfq_ioc_pool, cic);
2595 elv_ioc_count_dec(cfq_ioc_count);
2599 * CFQ scheduler is exiting, grab exit lock and check
2600 * the pending io context count. If it hits zero,
2601 * complete ioc_gone and set it back to NULL
2603 spin_lock(&ioc_gone_lock);
2604 if (ioc_gone && !elv_ioc_count_read(cfq_ioc_count)) {
2608 spin_unlock(&ioc_gone_lock);
2612 static void cfq_cic_free(struct cfq_io_context *cic)
2614 call_rcu(&cic->rcu_head, cfq_cic_free_rcu);
2617 static void cic_free_func(struct io_context *ioc, struct cfq_io_context *cic)
2619 unsigned long flags;
2620 unsigned long dead_key = (unsigned long) cic->key;
2622 BUG_ON(!(dead_key & CIC_DEAD_KEY));
2624 spin_lock_irqsave(&ioc->lock, flags);
2625 radix_tree_delete(&ioc->radix_root, dead_key >> CIC_DEAD_INDEX_SHIFT);
2626 hlist_del_rcu(&cic->cic_list);
2627 spin_unlock_irqrestore(&ioc->lock, flags);
2633 * Must be called with rcu_read_lock() held or preemption otherwise disabled.
2634 * Only two callers of this - ->dtor() which is called with the rcu_read_lock(),
2635 * and ->trim() which is called with the task lock held
2637 static void cfq_free_io_context(struct io_context *ioc)
2640 * ioc->refcount is zero here, or we are called from elv_unregister(),
2641 * so no more cic's are allowed to be linked into this ioc. So it
2642 * should be ok to iterate over the known list, we will see all cic's
2643 * since no new ones are added.
2645 __call_for_each_cic(ioc, cic_free_func);
2648 static void cfq_put_cooperator(struct cfq_queue *cfqq)
2650 struct cfq_queue *__cfqq, *next;
2653 * If this queue was scheduled to merge with another queue, be
2654 * sure to drop the reference taken on that queue (and others in
2655 * the merge chain). See cfq_setup_merge and cfq_merge_cfqqs.
2657 __cfqq = cfqq->new_cfqq;
2659 if (__cfqq == cfqq) {
2660 WARN(1, "cfqq->new_cfqq loop detected\n");
2663 next = __cfqq->new_cfqq;
2664 cfq_put_queue(__cfqq);
2669 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2671 if (unlikely(cfqq == cfqd->active_queue)) {
2672 __cfq_slice_expired(cfqd, cfqq, 0);
2673 cfq_schedule_dispatch(cfqd);
2676 cfq_put_cooperator(cfqq);
2678 cfq_put_queue(cfqq);
2681 static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
2682 struct cfq_io_context *cic)
2684 struct io_context *ioc = cic->ioc;
2686 list_del_init(&cic->queue_list);
2689 * Make sure dead mark is seen for dead queues
2692 cic->key = cfqd_dead_key(cfqd);
2694 if (ioc->ioc_data == cic)
2695 rcu_assign_pointer(ioc->ioc_data, NULL);
2697 if (cic->cfqq[BLK_RW_ASYNC]) {
2698 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]);
2699 cic->cfqq[BLK_RW_ASYNC] = NULL;
2702 if (cic->cfqq[BLK_RW_SYNC]) {
2703 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]);
2704 cic->cfqq[BLK_RW_SYNC] = NULL;
2708 static void cfq_exit_single_io_context(struct io_context *ioc,
2709 struct cfq_io_context *cic)
2711 struct cfq_data *cfqd = cic_to_cfqd(cic);
2714 struct request_queue *q = cfqd->queue;
2715 unsigned long flags;
2717 spin_lock_irqsave(q->queue_lock, flags);
2720 * Ensure we get a fresh copy of the ->key to prevent
2721 * race between exiting task and queue
2723 smp_read_barrier_depends();
2724 if (cic->key == cfqd)
2725 __cfq_exit_single_io_context(cfqd, cic);
2727 spin_unlock_irqrestore(q->queue_lock, flags);
2732 * The process that ioc belongs to has exited, we need to clean up
2733 * and put the internal structures we have that belongs to that process.
2735 static void cfq_exit_io_context(struct io_context *ioc)
2737 call_for_each_cic(ioc, cfq_exit_single_io_context);
2740 static struct cfq_io_context *
2741 cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
2743 struct cfq_io_context *cic;
2745 cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO,
2748 cic->last_end_request = jiffies;
2749 INIT_LIST_HEAD(&cic->queue_list);
2750 INIT_HLIST_NODE(&cic->cic_list);
2751 cic->dtor = cfq_free_io_context;
2752 cic->exit = cfq_exit_io_context;
2753 elv_ioc_count_inc(cfq_ioc_count);
2759 static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc)
2761 struct task_struct *tsk = current;
2764 if (!cfq_cfqq_prio_changed(cfqq))
2767 ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio);
2768 switch (ioprio_class) {
2770 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
2771 case IOPRIO_CLASS_NONE:
2773 * no prio set, inherit CPU scheduling settings
2775 cfqq->ioprio = task_nice_ioprio(tsk);
2776 cfqq->ioprio_class = task_nice_ioclass(tsk);
2778 case IOPRIO_CLASS_RT:
2779 cfqq->ioprio = task_ioprio(ioc);
2780 cfqq->ioprio_class = IOPRIO_CLASS_RT;
2782 case IOPRIO_CLASS_BE:
2783 cfqq->ioprio = task_ioprio(ioc);
2784 cfqq->ioprio_class = IOPRIO_CLASS_BE;
2786 case IOPRIO_CLASS_IDLE:
2787 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
2789 cfq_clear_cfqq_idle_window(cfqq);
2794 * keep track of original prio settings in case we have to temporarily
2795 * elevate the priority of this queue
2797 cfqq->org_ioprio = cfqq->ioprio;
2798 cfqq->org_ioprio_class = cfqq->ioprio_class;
2799 cfq_clear_cfqq_prio_changed(cfqq);
2802 static void changed_ioprio(struct io_context *ioc, struct cfq_io_context *cic)
2804 struct cfq_data *cfqd = cic_to_cfqd(cic);
2805 struct cfq_queue *cfqq;
2806 unsigned long flags;
2808 if (unlikely(!cfqd))
2811 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2813 cfqq = cic->cfqq[BLK_RW_ASYNC];
2815 struct cfq_queue *new_cfqq;
2816 new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic->ioc,
2819 cic->cfqq[BLK_RW_ASYNC] = new_cfqq;
2820 cfq_put_queue(cfqq);
2824 cfqq = cic->cfqq[BLK_RW_SYNC];
2826 cfq_mark_cfqq_prio_changed(cfqq);
2828 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2831 static void cfq_ioc_set_ioprio(struct io_context *ioc)
2833 call_for_each_cic(ioc, changed_ioprio);
2834 ioc->ioprio_changed = 0;
2837 static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2838 pid_t pid, bool is_sync)
2840 RB_CLEAR_NODE(&cfqq->rb_node);
2841 RB_CLEAR_NODE(&cfqq->p_node);
2842 INIT_LIST_HEAD(&cfqq->fifo);
2844 atomic_set(&cfqq->ref, 0);
2847 cfq_mark_cfqq_prio_changed(cfqq);
2850 if (!cfq_class_idle(cfqq))
2851 cfq_mark_cfqq_idle_window(cfqq);
2852 cfq_mark_cfqq_sync(cfqq);
2857 #ifdef CONFIG_CFQ_GROUP_IOSCHED
2858 static void changed_cgroup(struct io_context *ioc, struct cfq_io_context *cic)
2860 struct cfq_queue *sync_cfqq = cic_to_cfqq(cic, 1);
2861 struct cfq_data *cfqd = cic_to_cfqd(cic);
2862 unsigned long flags;
2863 struct request_queue *q;
2865 if (unlikely(!cfqd))
2870 spin_lock_irqsave(q->queue_lock, flags);
2874 * Drop reference to sync queue. A new sync queue will be
2875 * assigned in new group upon arrival of a fresh request.
2877 cfq_log_cfqq(cfqd, sync_cfqq, "changed cgroup");
2878 cic_set_cfqq(cic, NULL, 1);
2879 cfq_put_queue(sync_cfqq);
2882 spin_unlock_irqrestore(q->queue_lock, flags);
2885 static void cfq_ioc_set_cgroup(struct io_context *ioc)
2887 call_for_each_cic(ioc, changed_cgroup);
2888 ioc->cgroup_changed = 0;
2890 #endif /* CONFIG_CFQ_GROUP_IOSCHED */
2892 static struct cfq_queue *
2893 cfq_find_alloc_queue(struct cfq_data *cfqd, bool is_sync,
2894 struct io_context *ioc, gfp_t gfp_mask)
2896 struct cfq_queue *cfqq, *new_cfqq = NULL;
2897 struct cfq_io_context *cic;
2898 struct cfq_group *cfqg;
2901 cfqg = cfq_get_cfqg(cfqd, 1);
2902 cic = cfq_cic_lookup(cfqd, ioc);
2903 /* cic always exists here */
2904 cfqq = cic_to_cfqq(cic, is_sync);
2907 * Always try a new alloc if we fell back to the OOM cfqq
2908 * originally, since it should just be a temporary situation.
2910 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
2915 } else if (gfp_mask & __GFP_WAIT) {
2916 spin_unlock_irq(cfqd->queue->queue_lock);
2917 new_cfqq = kmem_cache_alloc_node(cfq_pool,
2918 gfp_mask | __GFP_ZERO,
2920 spin_lock_irq(cfqd->queue->queue_lock);
2924 cfqq = kmem_cache_alloc_node(cfq_pool,
2925 gfp_mask | __GFP_ZERO,
2930 cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
2931 cfq_init_prio_data(cfqq, ioc);
2932 cfq_link_cfqq_cfqg(cfqq, cfqg);
2933 cfq_log_cfqq(cfqd, cfqq, "alloced");
2935 cfqq = &cfqd->oom_cfqq;
2939 kmem_cache_free(cfq_pool, new_cfqq);
2944 static struct cfq_queue **
2945 cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
2947 switch (ioprio_class) {
2948 case IOPRIO_CLASS_RT:
2949 return &cfqd->async_cfqq[0][ioprio];
2950 case IOPRIO_CLASS_BE:
2951 return &cfqd->async_cfqq[1][ioprio];
2952 case IOPRIO_CLASS_IDLE:
2953 return &cfqd->async_idle_cfqq;
2959 static struct cfq_queue *
2960 cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct io_context *ioc,
2963 const int ioprio = task_ioprio(ioc);
2964 const int ioprio_class = task_ioprio_class(ioc);
2965 struct cfq_queue **async_cfqq = NULL;
2966 struct cfq_queue *cfqq = NULL;
2969 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
2974 cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask);
2977 * pin the queue now that it's allocated, scheduler exit will prune it
2979 if (!is_sync && !(*async_cfqq)) {
2980 atomic_inc(&cfqq->ref);
2984 atomic_inc(&cfqq->ref);
2989 * We drop cfq io contexts lazily, so we may find a dead one.
2992 cfq_drop_dead_cic(struct cfq_data *cfqd, struct io_context *ioc,
2993 struct cfq_io_context *cic)
2995 unsigned long flags;
2997 WARN_ON(!list_empty(&cic->queue_list));
2998 BUG_ON(cic->key != cfqd_dead_key(cfqd));
3000 spin_lock_irqsave(&ioc->lock, flags);
3002 BUG_ON(ioc->ioc_data == cic);
3004 radix_tree_delete(&ioc->radix_root, cfqd->cic_index);
3005 hlist_del_rcu(&cic->cic_list);
3006 spin_unlock_irqrestore(&ioc->lock, flags);
3011 static struct cfq_io_context *
3012 cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc)
3014 struct cfq_io_context *cic;
3015 unsigned long flags;
3023 * we maintain a last-hit cache, to avoid browsing over the tree
3025 cic = rcu_dereference(ioc->ioc_data);
3026 if (cic && cic->key == cfqd) {
3032 cic = radix_tree_lookup(&ioc->radix_root, cfqd->cic_index);
3036 if (unlikely(cic->key != cfqd)) {
3037 cfq_drop_dead_cic(cfqd, ioc, cic);
3042 spin_lock_irqsave(&ioc->lock, flags);
3043 rcu_assign_pointer(ioc->ioc_data, cic);
3044 spin_unlock_irqrestore(&ioc->lock, flags);
3052 * Add cic into ioc, using cfqd as the search key. This enables us to lookup
3053 * the process specific cfq io context when entered from the block layer.
3054 * Also adds the cic to a per-cfqd list, used when this queue is removed.
3056 static int cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
3057 struct cfq_io_context *cic, gfp_t gfp_mask)
3059 unsigned long flags;
3062 ret = radix_tree_preload(gfp_mask);
3067 spin_lock_irqsave(&ioc->lock, flags);
3068 ret = radix_tree_insert(&ioc->radix_root,
3069 cfqd->cic_index, cic);
3071 hlist_add_head_rcu(&cic->cic_list, &ioc->cic_list);
3072 spin_unlock_irqrestore(&ioc->lock, flags);
3074 radix_tree_preload_end();
3077 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
3078 list_add(&cic->queue_list, &cfqd->cic_list);
3079 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
3084 printk(KERN_ERR "cfq: cic link failed!\n");
3090 * Setup general io context and cfq io context. There can be several cfq
3091 * io contexts per general io context, if this process is doing io to more
3092 * than one device managed by cfq.
3094 static struct cfq_io_context *
3095 cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
3097 struct io_context *ioc = NULL;
3098 struct cfq_io_context *cic;
3100 might_sleep_if(gfp_mask & __GFP_WAIT);
3102 ioc = get_io_context(gfp_mask, cfqd->queue->node);
3106 cic = cfq_cic_lookup(cfqd, ioc);
3110 cic = cfq_alloc_io_context(cfqd, gfp_mask);
3114 if (cfq_cic_link(cfqd, ioc, cic, gfp_mask))
3118 smp_read_barrier_depends();
3119 if (unlikely(ioc->ioprio_changed))
3120 cfq_ioc_set_ioprio(ioc);
3122 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3123 if (unlikely(ioc->cgroup_changed))
3124 cfq_ioc_set_cgroup(ioc);
3130 put_io_context(ioc);
3135 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
3137 unsigned long elapsed = jiffies - cic->last_end_request;
3138 unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
3140 cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
3141 cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
3142 cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
3146 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3150 sector_t n_sec = blk_rq_sectors(rq);
3151 if (cfqq->last_request_pos) {
3152 if (cfqq->last_request_pos < blk_rq_pos(rq))
3153 sdist = blk_rq_pos(rq) - cfqq->last_request_pos;
3155 sdist = cfqq->last_request_pos - blk_rq_pos(rq);
3158 cfqq->seek_history <<= 1;
3159 if (blk_queue_nonrot(cfqd->queue))
3160 cfqq->seek_history |= (n_sec < CFQQ_SECT_THR_NONROT);
3162 cfqq->seek_history |= (sdist > CFQQ_SEEK_THR);
3166 * Disable idle window if the process thinks too long or seeks so much that
3170 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3171 struct cfq_io_context *cic)
3173 int old_idle, enable_idle;
3176 * Don't idle for async or idle io prio class
3178 if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
3181 enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
3183 if (cfqq->queued[0] + cfqq->queued[1] >= 4)
3184 cfq_mark_cfqq_deep(cfqq);
3186 if (cfqq->next_rq && (cfqq->next_rq->cmd_flags & REQ_NOIDLE))
3188 else if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle ||
3189 (!cfq_cfqq_deep(cfqq) && CFQQ_SEEKY(cfqq)))
3191 else if (sample_valid(cic->ttime_samples)) {
3192 if (cic->ttime_mean > cfqd->cfq_slice_idle)
3198 if (old_idle != enable_idle) {
3199 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
3201 cfq_mark_cfqq_idle_window(cfqq);
3203 cfq_clear_cfqq_idle_window(cfqq);
3208 * Check if new_cfqq should preempt the currently active queue. Return 0 for
3209 * no or if we aren't sure, a 1 will cause a preempt.
3212 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
3215 struct cfq_queue *cfqq;
3217 cfqq = cfqd->active_queue;
3221 if (cfq_class_idle(new_cfqq))
3224 if (cfq_class_idle(cfqq))
3228 * Don't allow a non-RT request to preempt an ongoing RT cfqq timeslice.
3230 if (cfq_class_rt(cfqq) && !cfq_class_rt(new_cfqq))
3234 * if the new request is sync, but the currently running queue is
3235 * not, let the sync request have priority.
3237 if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
3240 if (new_cfqq->cfqg != cfqq->cfqg)
3243 if (cfq_slice_used(cfqq))
3246 /* Allow preemption only if we are idling on sync-noidle tree */
3247 if (cfqd->serving_type == SYNC_NOIDLE_WORKLOAD &&
3248 cfqq_type(new_cfqq) == SYNC_NOIDLE_WORKLOAD &&
3249 new_cfqq->service_tree->count == 2 &&
3250 RB_EMPTY_ROOT(&cfqq->sort_list))
3254 * So both queues are sync. Let the new request get disk time if
3255 * it's a metadata request and the current queue is doing regular IO.
3257 if ((rq->cmd_flags & REQ_META) && !cfqq->meta_pending)
3261 * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
3263 if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
3266 /* An idle queue should not be idle now for some reason */
3267 if (RB_EMPTY_ROOT(&cfqq->sort_list) && !cfq_should_idle(cfqd, cfqq))
3270 if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
3274 * if this request is as-good as one we would expect from the
3275 * current cfqq, let it preempt
3277 if (cfq_rq_close(cfqd, cfqq, rq))
3284 * cfqq preempts the active queue. if we allowed preempt with no slice left,
3285 * let it have half of its nominal slice.
3287 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3289 cfq_log_cfqq(cfqd, cfqq, "preempt");
3290 cfq_slice_expired(cfqd, 1);
3293 * Put the new queue at the front of the of the current list,
3294 * so we know that it will be selected next.
3296 BUG_ON(!cfq_cfqq_on_rr(cfqq));
3298 cfq_service_tree_add(cfqd, cfqq, 1);
3300 cfqq->slice_end = 0;
3301 cfq_mark_cfqq_slice_new(cfqq);
3305 * Called when a new fs request (rq) is added (to cfqq). Check if there's
3306 * something we should do about it
3309 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3312 struct cfq_io_context *cic = RQ_CIC(rq);
3315 if (rq->cmd_flags & REQ_META)
3316 cfqq->meta_pending++;
3318 cfq_update_io_thinktime(cfqd, cic);
3319 cfq_update_io_seektime(cfqd, cfqq, rq);
3320 cfq_update_idle_window(cfqd, cfqq, cic);
3322 cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
3324 if (cfqq == cfqd->active_queue) {
3326 * Remember that we saw a request from this process, but
3327 * don't start queuing just yet. Otherwise we risk seeing lots
3328 * of tiny requests, because we disrupt the normal plugging
3329 * and merging. If the request is already larger than a single
3330 * page, let it rip immediately. For that case we assume that
3331 * merging is already done. Ditto for a busy system that
3332 * has other work pending, don't risk delaying until the
3333 * idle timer unplug to continue working.
3335 if (cfq_cfqq_wait_request(cfqq)) {
3336 if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
3337 cfqd->busy_queues > 1) {
3338 cfq_del_timer(cfqd, cfqq);
3339 cfq_clear_cfqq_wait_request(cfqq);
3340 __blk_run_queue(cfqd->queue);
3342 cfq_blkiocg_update_idle_time_stats(
3344 cfq_mark_cfqq_must_dispatch(cfqq);
3347 } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
3349 * not the active queue - expire current slice if it is
3350 * idle and has expired it's mean thinktime or this new queue
3351 * has some old slice time left and is of higher priority or
3352 * this new queue is RT and the current one is BE
3354 cfq_preempt_queue(cfqd, cfqq);
3355 __blk_run_queue(cfqd->queue);
3359 static void cfq_insert_request(struct request_queue *q, struct request *rq)
3361 struct cfq_data *cfqd = q->elevator->elevator_data;
3362 struct cfq_queue *cfqq = RQ_CFQQ(rq);
3364 cfq_log_cfqq(cfqd, cfqq, "insert_request");
3365 cfq_init_prio_data(cfqq, RQ_CIC(rq)->ioc);
3367 rq_set_fifo_time(rq, jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)]);
3368 list_add_tail(&rq->queuelist, &cfqq->fifo);
3370 cfq_blkiocg_update_io_add_stats(&(RQ_CFQG(rq))->blkg,
3371 &cfqd->serving_group->blkg, rq_data_dir(rq),
3373 cfq_rq_enqueued(cfqd, cfqq, rq);
3377 * Update hw_tag based on peak queue depth over 50 samples under
3380 static void cfq_update_hw_tag(struct cfq_data *cfqd)
3382 struct cfq_queue *cfqq = cfqd->active_queue;
3384 if (cfqd->rq_in_driver > cfqd->hw_tag_est_depth)
3385 cfqd->hw_tag_est_depth = cfqd->rq_in_driver;
3387 if (cfqd->hw_tag == 1)
3390 if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
3391 cfqd->rq_in_driver <= CFQ_HW_QUEUE_MIN)
3395 * If active queue hasn't enough requests and can idle, cfq might not
3396 * dispatch sufficient requests to hardware. Don't zero hw_tag in this
3399 if (cfqq && cfq_cfqq_idle_window(cfqq) &&
3400 cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] <
3401 CFQ_HW_QUEUE_MIN && cfqd->rq_in_driver < CFQ_HW_QUEUE_MIN)
3404 if (cfqd->hw_tag_samples++ < 50)
3407 if (cfqd->hw_tag_est_depth >= CFQ_HW_QUEUE_MIN)
3413 static bool cfq_should_wait_busy(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3415 struct cfq_io_context *cic = cfqd->active_cic;
3417 /* If there are other queues in the group, don't wait */
3418 if (cfqq->cfqg->nr_cfqq > 1)
3421 if (cfq_slice_used(cfqq))
3424 /* if slice left is less than think time, wait busy */
3425 if (cic && sample_valid(cic->ttime_samples)
3426 && (cfqq->slice_end - jiffies < cic->ttime_mean))
3430 * If think times is less than a jiffy than ttime_mean=0 and above
3431 * will not be true. It might happen that slice has not expired yet
3432 * but will expire soon (4-5 ns) during select_queue(). To cover the
3433 * case where think time is less than a jiffy, mark the queue wait
3434 * busy if only 1 jiffy is left in the slice.
3436 if (cfqq->slice_end - jiffies == 1)
3442 static void cfq_completed_request(struct request_queue *q, struct request *rq)
3444 struct cfq_queue *cfqq = RQ_CFQQ(rq);
3445 struct cfq_data *cfqd = cfqq->cfqd;
3446 const int sync = rq_is_sync(rq);
3450 cfq_log_cfqq(cfqd, cfqq, "complete rqnoidle %d",
3451 !!(rq->cmd_flags & REQ_NOIDLE));
3453 cfq_update_hw_tag(cfqd);
3455 WARN_ON(!cfqd->rq_in_driver);
3456 WARN_ON(!cfqq->dispatched);
3457 cfqd->rq_in_driver--;
3459 (RQ_CFQG(rq))->dispatched--;
3460 cfq_blkiocg_update_completion_stats(&cfqq->cfqg->blkg,
3461 rq_start_time_ns(rq), rq_io_start_time_ns(rq),
3462 rq_data_dir(rq), rq_is_sync(rq));
3464 cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]--;
3467 RQ_CIC(rq)->last_end_request = now;
3468 if (!time_after(rq->start_time + cfqd->cfq_fifo_expire[1], now))
3469 cfqd->last_delayed_sync = now;
3473 * If this is the active queue, check if it needs to be expired,
3474 * or if we want to idle in case it has no pending requests.
3476 if (cfqd->active_queue == cfqq) {
3477 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
3479 if (cfq_cfqq_slice_new(cfqq)) {
3480 cfq_set_prio_slice(cfqd, cfqq);
3481 cfq_clear_cfqq_slice_new(cfqq);
3485 * Should we wait for next request to come in before we expire
3488 if (cfq_should_wait_busy(cfqd, cfqq)) {
3489 unsigned long extend_sl = cfqd->cfq_slice_idle;
3490 if (!cfqd->cfq_slice_idle)
3491 extend_sl = cfqd->cfq_group_idle;
3492 cfqq->slice_end = jiffies + extend_sl;
3493 cfq_mark_cfqq_wait_busy(cfqq);
3494 cfq_log_cfqq(cfqd, cfqq, "will busy wait");
3498 * Idling is not enabled on:
3500 * - idle-priority queues
3502 * - queues with still some requests queued
3503 * - when there is a close cooperator
3505 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
3506 cfq_slice_expired(cfqd, 1);
3507 else if (sync && cfqq_empty &&
3508 !cfq_close_cooperator(cfqd, cfqq)) {
3509 cfq_arm_slice_timer(cfqd);
3513 if (!cfqd->rq_in_driver)
3514 cfq_schedule_dispatch(cfqd);
3518 * we temporarily boost lower priority queues if they are holding fs exclusive
3519 * resources. they are boosted to normal prio (CLASS_BE/4)
3521 static void cfq_prio_boost(struct cfq_queue *cfqq)
3523 if (has_fs_excl()) {
3525 * boost idle prio on transactions that would lock out other
3526 * users of the filesystem
3528 if (cfq_class_idle(cfqq))
3529 cfqq->ioprio_class = IOPRIO_CLASS_BE;
3530 if (cfqq->ioprio > IOPRIO_NORM)
3531 cfqq->ioprio = IOPRIO_NORM;
3534 * unboost the queue (if needed)
3536 cfqq->ioprio_class = cfqq->org_ioprio_class;
3537 cfqq->ioprio = cfqq->org_ioprio;
3541 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
3543 if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
3544 cfq_mark_cfqq_must_alloc_slice(cfqq);
3545 return ELV_MQUEUE_MUST;
3548 return ELV_MQUEUE_MAY;
3551 static int cfq_may_queue(struct request_queue *q, int rw)
3553 struct cfq_data *cfqd = q->elevator->elevator_data;
3554 struct task_struct *tsk = current;
3555 struct cfq_io_context *cic;
3556 struct cfq_queue *cfqq;
3559 * don't force setup of a queue from here, as a call to may_queue
3560 * does not necessarily imply that a request actually will be queued.
3561 * so just lookup a possibly existing queue, or return 'may queue'
3564 cic = cfq_cic_lookup(cfqd, tsk->io_context);
3566 return ELV_MQUEUE_MAY;
3568 cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
3570 cfq_init_prio_data(cfqq, cic->ioc);
3571 cfq_prio_boost(cfqq);
3573 return __cfq_may_queue(cfqq);
3576 return ELV_MQUEUE_MAY;
3580 * queue lock held here
3582 static void cfq_put_request(struct request *rq)
3584 struct cfq_queue *cfqq = RQ_CFQQ(rq);
3587 const int rw = rq_data_dir(rq);
3589 BUG_ON(!cfqq->allocated[rw]);
3590 cfqq->allocated[rw]--;
3592 put_io_context(RQ_CIC(rq)->ioc);
3594 rq->elevator_private = NULL;
3595 rq->elevator_private2 = NULL;
3597 /* Put down rq reference on cfqg */
3598 cfq_put_cfqg(RQ_CFQG(rq));
3599 rq->elevator_private3 = NULL;
3601 cfq_put_queue(cfqq);
3605 static struct cfq_queue *
3606 cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_context *cic,
3607 struct cfq_queue *cfqq)
3609 cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq);
3610 cic_set_cfqq(cic, cfqq->new_cfqq, 1);
3611 cfq_mark_cfqq_coop(cfqq->new_cfqq);
3612 cfq_put_queue(cfqq);
3613 return cic_to_cfqq(cic, 1);
3617 * Returns NULL if a new cfqq should be allocated, or the old cfqq if this
3618 * was the last process referring to said cfqq.
3620 static struct cfq_queue *
3621 split_cfqq(struct cfq_io_context *cic, struct cfq_queue *cfqq)
3623 if (cfqq_process_refs(cfqq) == 1) {
3624 cfqq->pid = current->pid;
3625 cfq_clear_cfqq_coop(cfqq);
3626 cfq_clear_cfqq_split_coop(cfqq);
3630 cic_set_cfqq(cic, NULL, 1);
3632 cfq_put_cooperator(cfqq);
3634 cfq_put_queue(cfqq);
3638 * Allocate cfq data structures associated with this request.
3641 cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
3643 struct cfq_data *cfqd = q->elevator->elevator_data;
3644 struct cfq_io_context *cic;
3645 const int rw = rq_data_dir(rq);
3646 const bool is_sync = rq_is_sync(rq);
3647 struct cfq_queue *cfqq;
3648 unsigned long flags;
3650 might_sleep_if(gfp_mask & __GFP_WAIT);
3652 cic = cfq_get_io_context(cfqd, gfp_mask);
3654 spin_lock_irqsave(q->queue_lock, flags);
3660 cfqq = cic_to_cfqq(cic, is_sync);
3661 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
3662 cfqq = cfq_get_queue(cfqd, is_sync, cic->ioc, gfp_mask);
3663 cic_set_cfqq(cic, cfqq, is_sync);
3666 * If the queue was seeky for too long, break it apart.
3668 if (cfq_cfqq_coop(cfqq) && cfq_cfqq_split_coop(cfqq)) {
3669 cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq");
3670 cfqq = split_cfqq(cic, cfqq);
3676 * Check to see if this queue is scheduled to merge with
3677 * another, closely cooperating queue. The merging of
3678 * queues happens here as it must be done in process context.
3679 * The reference on new_cfqq was taken in merge_cfqqs.
3682 cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq);
3685 cfqq->allocated[rw]++;
3686 atomic_inc(&cfqq->ref);
3688 spin_unlock_irqrestore(q->queue_lock, flags);
3690 rq->elevator_private = cic;
3691 rq->elevator_private2 = cfqq;
3692 rq->elevator_private3 = cfq_ref_get_cfqg(cfqq->cfqg);
3697 put_io_context(cic->ioc);
3699 cfq_schedule_dispatch(cfqd);
3700 spin_unlock_irqrestore(q->queue_lock, flags);
3701 cfq_log(cfqd, "set_request fail");
3705 static void cfq_kick_queue(struct work_struct *work)
3707 struct cfq_data *cfqd =
3708 container_of(work, struct cfq_data, unplug_work);
3709 struct request_queue *q = cfqd->queue;
3711 spin_lock_irq(q->queue_lock);
3712 __blk_run_queue(cfqd->queue);
3713 spin_unlock_irq(q->queue_lock);
3717 * Timer running if the active_queue is currently idling inside its time slice
3719 static void cfq_idle_slice_timer(unsigned long data)
3721 struct cfq_data *cfqd = (struct cfq_data *) data;
3722 struct cfq_queue *cfqq;
3723 unsigned long flags;
3726 cfq_log(cfqd, "idle timer fired");
3728 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
3730 cfqq = cfqd->active_queue;
3735 * We saw a request before the queue expired, let it through
3737 if (cfq_cfqq_must_dispatch(cfqq))
3743 if (cfq_slice_used(cfqq))
3747 * only expire and reinvoke request handler, if there are
3748 * other queues with pending requests
3750 if (!cfqd->busy_queues)
3754 * not expired and it has a request pending, let it dispatch
3756 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3760 * Queue depth flag is reset only when the idle didn't succeed
3762 cfq_clear_cfqq_deep(cfqq);
3765 cfq_slice_expired(cfqd, timed_out);
3767 cfq_schedule_dispatch(cfqd);
3769 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
3772 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
3774 del_timer_sync(&cfqd->idle_slice_timer);
3775 cancel_work_sync(&cfqd->unplug_work);
3778 static void cfq_put_async_queues(struct cfq_data *cfqd)
3782 for (i = 0; i < IOPRIO_BE_NR; i++) {
3783 if (cfqd->async_cfqq[0][i])
3784 cfq_put_queue(cfqd->async_cfqq[0][i]);
3785 if (cfqd->async_cfqq[1][i])
3786 cfq_put_queue(cfqd->async_cfqq[1][i]);
3789 if (cfqd->async_idle_cfqq)
3790 cfq_put_queue(cfqd->async_idle_cfqq);
3793 static void cfq_cfqd_free(struct rcu_head *head)
3795 kfree(container_of(head, struct cfq_data, rcu));
3798 static void cfq_exit_queue(struct elevator_queue *e)
3800 struct cfq_data *cfqd = e->elevator_data;
3801 struct request_queue *q = cfqd->queue;
3803 cfq_shutdown_timer_wq(cfqd);
3805 spin_lock_irq(q->queue_lock);
3807 if (cfqd->active_queue)
3808 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
3810 while (!list_empty(&cfqd->cic_list)) {
3811 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
3812 struct cfq_io_context,
3815 __cfq_exit_single_io_context(cfqd, cic);
3818 cfq_put_async_queues(cfqd);
3819 cfq_release_cfq_groups(cfqd);
3820 cfq_blkiocg_del_blkio_group(&cfqd->root_group.blkg);
3822 spin_unlock_irq(q->queue_lock);
3824 cfq_shutdown_timer_wq(cfqd);
3826 spin_lock(&cic_index_lock);
3827 ida_remove(&cic_index_ida, cfqd->cic_index);
3828 spin_unlock(&cic_index_lock);
3830 /* Wait for cfqg->blkg->key accessors to exit their grace periods. */
3831 call_rcu(&cfqd->rcu, cfq_cfqd_free);
3834 static int cfq_alloc_cic_index(void)
3839 if (!ida_pre_get(&cic_index_ida, GFP_KERNEL))
3842 spin_lock(&cic_index_lock);
3843 error = ida_get_new(&cic_index_ida, &index);
3844 spin_unlock(&cic_index_lock);
3845 if (error && error != -EAGAIN)
3852 static void *cfq_init_queue(struct request_queue *q)
3854 struct cfq_data *cfqd;
3856 struct cfq_group *cfqg;
3857 struct cfq_rb_root *st;
3859 i = cfq_alloc_cic_index();
3863 cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
3867 cfqd->cic_index = i;
3869 /* Init root service tree */
3870 cfqd->grp_service_tree = CFQ_RB_ROOT;
3872 /* Init root group */
3873 cfqg = &cfqd->root_group;
3874 for_each_cfqg_st(cfqg, i, j, st)
3876 RB_CLEAR_NODE(&cfqg->rb_node);
3878 /* Give preference to root group over other groups */
3879 cfqg->weight = 2*BLKIO_WEIGHT_DEFAULT;
3881 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3883 * Take a reference to root group which we never drop. This is just
3884 * to make sure that cfq_put_cfqg() does not try to kfree root group
3886 atomic_set(&cfqg->ref, 1);
3888 cfq_blkiocg_add_blkio_group(&blkio_root_cgroup, &cfqg->blkg,
3893 * Not strictly needed (since RB_ROOT just clears the node and we
3894 * zeroed cfqd on alloc), but better be safe in case someone decides
3895 * to add magic to the rb code
3897 for (i = 0; i < CFQ_PRIO_LISTS; i++)
3898 cfqd->prio_trees[i] = RB_ROOT;
3901 * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
3902 * Grab a permanent reference to it, so that the normal code flow
3903 * will not attempt to free it.
3905 cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
3906 atomic_inc(&cfqd->oom_cfqq.ref);
3907 cfq_link_cfqq_cfqg(&cfqd->oom_cfqq, &cfqd->root_group);
3909 INIT_LIST_HEAD(&cfqd->cic_list);
3913 init_timer(&cfqd->idle_slice_timer);
3914 cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
3915 cfqd->idle_slice_timer.data = (unsigned long) cfqd;
3917 INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
3919 cfqd->cfq_quantum = cfq_quantum;
3920 cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
3921 cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
3922 cfqd->cfq_back_max = cfq_back_max;
3923 cfqd->cfq_back_penalty = cfq_back_penalty;
3924 cfqd->cfq_slice[0] = cfq_slice_async;
3925 cfqd->cfq_slice[1] = cfq_slice_sync;
3926 cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
3927 cfqd->cfq_slice_idle = cfq_slice_idle;
3928 cfqd->cfq_group_idle = cfq_group_idle;
3929 cfqd->cfq_latency = 1;
3930 cfqd->cfq_group_isolation = 0;
3933 * we optimistically start assuming sync ops weren't delayed in last
3934 * second, in order to have larger depth for async operations.
3936 cfqd->last_delayed_sync = jiffies - HZ;
3940 static void cfq_slab_kill(void)
3943 * Caller already ensured that pending RCU callbacks are completed,
3944 * so we should have no busy allocations at this point.
3947 kmem_cache_destroy(cfq_pool);
3949 kmem_cache_destroy(cfq_ioc_pool);
3952 static int __init cfq_slab_setup(void)
3954 cfq_pool = KMEM_CACHE(cfq_queue, 0);
3958 cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0);
3969 * sysfs parts below -->
3972 cfq_var_show(unsigned int var, char *page)
3974 return sprintf(page, "%d\n", var);
3978 cfq_var_store(unsigned int *var, const char *page, size_t count)
3980 char *p = (char *) page;
3982 *var = simple_strtoul(p, &p, 10);
3986 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
3987 static ssize_t __FUNC(struct elevator_queue *e, char *page) \
3989 struct cfq_data *cfqd = e->elevator_data; \
3990 unsigned int __data = __VAR; \
3992 __data = jiffies_to_msecs(__data); \
3993 return cfq_var_show(__data, (page)); \
3995 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
3996 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
3997 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
3998 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
3999 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
4000 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
4001 SHOW_FUNCTION(cfq_group_idle_show, cfqd->cfq_group_idle, 1);
4002 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
4003 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
4004 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
4005 SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
4006 SHOW_FUNCTION(cfq_group_isolation_show, cfqd->cfq_group_isolation, 0);
4007 #undef SHOW_FUNCTION
4009 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
4010 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
4012 struct cfq_data *cfqd = e->elevator_data; \
4013 unsigned int __data; \
4014 int ret = cfq_var_store(&__data, (page), count); \
4015 if (__data < (MIN)) \
4017 else if (__data > (MAX)) \
4020 *(__PTR) = msecs_to_jiffies(__data); \
4022 *(__PTR) = __data; \
4025 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
4026 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
4028 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
4030 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
4031 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
4033 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
4034 STORE_FUNCTION(cfq_group_idle_store, &cfqd->cfq_group_idle, 0, UINT_MAX, 1);
4035 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
4036 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
4037 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
4039 STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
4040 STORE_FUNCTION(cfq_group_isolation_store, &cfqd->cfq_group_isolation, 0, 1, 0);
4041 #undef STORE_FUNCTION
4043 #define CFQ_ATTR(name) \
4044 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
4046 static struct elv_fs_entry cfq_attrs[] = {
4048 CFQ_ATTR(fifo_expire_sync),
4049 CFQ_ATTR(fifo_expire_async),
4050 CFQ_ATTR(back_seek_max),
4051 CFQ_ATTR(back_seek_penalty),
4052 CFQ_ATTR(slice_sync),
4053 CFQ_ATTR(slice_async),
4054 CFQ_ATTR(slice_async_rq),
4055 CFQ_ATTR(slice_idle),
4056 CFQ_ATTR(group_idle),
4057 CFQ_ATTR(low_latency),
4058 CFQ_ATTR(group_isolation),
4062 static struct elevator_type iosched_cfq = {
4064 .elevator_merge_fn = cfq_merge,
4065 .elevator_merged_fn = cfq_merged_request,
4066 .elevator_merge_req_fn = cfq_merged_requests,
4067 .elevator_allow_merge_fn = cfq_allow_merge,
4068 .elevator_bio_merged_fn = cfq_bio_merged,
4069 .elevator_dispatch_fn = cfq_dispatch_requests,
4070 .elevator_add_req_fn = cfq_insert_request,
4071 .elevator_activate_req_fn = cfq_activate_request,
4072 .elevator_deactivate_req_fn = cfq_deactivate_request,
4073 .elevator_queue_empty_fn = cfq_queue_empty,
4074 .elevator_completed_req_fn = cfq_completed_request,
4075 .elevator_former_req_fn = elv_rb_former_request,
4076 .elevator_latter_req_fn = elv_rb_latter_request,
4077 .elevator_set_req_fn = cfq_set_request,
4078 .elevator_put_req_fn = cfq_put_request,
4079 .elevator_may_queue_fn = cfq_may_queue,
4080 .elevator_init_fn = cfq_init_queue,
4081 .elevator_exit_fn = cfq_exit_queue,
4082 .trim = cfq_free_io_context,
4084 .elevator_attrs = cfq_attrs,
4085 .elevator_name = "cfq",
4086 .elevator_owner = THIS_MODULE,
4089 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4090 static struct blkio_policy_type blkio_policy_cfq = {
4092 .blkio_unlink_group_fn = cfq_unlink_blkio_group,
4093 .blkio_update_group_weight_fn = cfq_update_blkio_group_weight,
4095 .plid = BLKIO_POLICY_PROP,
4098 static struct blkio_policy_type blkio_policy_cfq;
4101 static int __init cfq_init(void)
4104 * could be 0 on HZ < 1000 setups
4106 if (!cfq_slice_async)
4107 cfq_slice_async = 1;
4108 if (!cfq_slice_idle)
4111 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4112 if (!cfq_group_idle)
4117 if (cfq_slab_setup())
4120 elv_register(&iosched_cfq);
4121 blkio_policy_register(&blkio_policy_cfq);
4126 static void __exit cfq_exit(void)
4128 DECLARE_COMPLETION_ONSTACK(all_gone);
4129 blkio_policy_unregister(&blkio_policy_cfq);
4130 elv_unregister(&iosched_cfq);
4131 ioc_gone = &all_gone;
4132 /* ioc_gone's update must be visible before reading ioc_count */
4136 * this also protects us from entering cfq_slab_kill() with
4137 * pending RCU callbacks
4139 if (elv_ioc_count_read(cfq_ioc_count))
4140 wait_for_completion(&all_gone);
4141 ida_destroy(&cic_index_ida);
4145 module_init(cfq_init);
4146 module_exit(cfq_exit);
4148 MODULE_AUTHOR("Jens Axboe");
4149 MODULE_LICENSE("GPL");
4150 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");