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[0])
58 #define RQ_CFQQ(rq) (struct cfq_queue *) ((rq)->elevator_private[1])
59 #define RQ_CFQG(rq) (struct cfq_group *) ((rq)->elevator_private[2])
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 /* Number of sectors dispatched from queue in single dispatch round */
150 unsigned long nr_sectors;
154 * First index in the service_trees.
155 * IDLE is handled separately, so it has negative index
165 * Second index in the service_trees.
169 SYNC_NOIDLE_WORKLOAD = 1,
173 /* This is per cgroup per device grouping structure */
175 /* group service_tree member */
176 struct rb_node rb_node;
178 /* group service_tree key */
182 /* number of cfqq currently on this group */
186 * Per group busy queus average. Useful for workload slice calc. We
187 * create the array for each prio class but at run time it is used
188 * only for RT and BE class and slot for IDLE class remains unused.
189 * This is primarily done to avoid confusion and a gcc warning.
191 unsigned int busy_queues_avg[CFQ_PRIO_NR];
193 * rr lists of queues with requests. We maintain service trees for
194 * RT and BE classes. These trees are subdivided in subclasses
195 * of SYNC, SYNC_NOIDLE and ASYNC based on workload type. For IDLE
196 * class there is no subclassification and all the cfq queues go on
197 * a single tree service_tree_idle.
198 * Counts are embedded in the cfq_rb_root
200 struct cfq_rb_root service_trees[2][3];
201 struct cfq_rb_root service_tree_idle;
203 unsigned long saved_workload_slice;
204 enum wl_type_t saved_workload;
205 enum wl_prio_t saved_serving_prio;
206 struct blkio_group blkg;
207 #ifdef CONFIG_CFQ_GROUP_IOSCHED
208 struct hlist_node cfqd_node;
211 /* number of requests that are on the dispatch list or inside driver */
216 * Per block device queue structure
219 struct request_queue *queue;
220 /* Root service tree for cfq_groups */
221 struct cfq_rb_root grp_service_tree;
222 struct cfq_group root_group;
225 * The priority currently being served
227 enum wl_prio_t serving_prio;
228 enum wl_type_t serving_type;
229 unsigned long workload_expires;
230 struct cfq_group *serving_group;
233 * Each priority tree is sorted by next_request position. These
234 * trees are used when determining if two or more queues are
235 * interleaving requests (see cfq_close_cooperator).
237 struct rb_root prio_trees[CFQ_PRIO_LISTS];
239 unsigned int busy_queues;
240 unsigned int busy_sync_queues;
246 * queue-depth detection
252 * -1 => indeterminate, (cfq will behave as if NCQ is present, to allow better detection)
253 * 1 => NCQ is present (hw_tag_est_depth is the estimated max depth)
256 int hw_tag_est_depth;
257 unsigned int hw_tag_samples;
260 * idle window management
262 struct timer_list idle_slice_timer;
263 struct work_struct unplug_work;
265 struct cfq_queue *active_queue;
266 struct cfq_io_context *active_cic;
269 * async queue for each priority case
271 struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
272 struct cfq_queue *async_idle_cfqq;
274 sector_t last_position;
277 * tunables, see top of file
279 unsigned int cfq_quantum;
280 unsigned int cfq_fifo_expire[2];
281 unsigned int cfq_back_penalty;
282 unsigned int cfq_back_max;
283 unsigned int cfq_slice[2];
284 unsigned int cfq_slice_async_rq;
285 unsigned int cfq_slice_idle;
286 unsigned int cfq_group_idle;
287 unsigned int cfq_latency;
289 unsigned int cic_index;
290 struct list_head cic_list;
293 * Fallback dummy cfqq for extreme OOM conditions
295 struct cfq_queue oom_cfqq;
297 unsigned long last_delayed_sync;
299 /* List of cfq groups being managed on this device*/
300 struct hlist_head cfqg_list;
304 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd);
306 static struct cfq_rb_root *service_tree_for(struct cfq_group *cfqg,
313 if (prio == IDLE_WORKLOAD)
314 return &cfqg->service_tree_idle;
316 return &cfqg->service_trees[prio][type];
319 enum cfqq_state_flags {
320 CFQ_CFQQ_FLAG_on_rr = 0, /* on round-robin busy list */
321 CFQ_CFQQ_FLAG_wait_request, /* waiting for a request */
322 CFQ_CFQQ_FLAG_must_dispatch, /* must be allowed a dispatch */
323 CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
324 CFQ_CFQQ_FLAG_fifo_expire, /* FIFO checked in this slice */
325 CFQ_CFQQ_FLAG_idle_window, /* slice idling enabled */
326 CFQ_CFQQ_FLAG_prio_changed, /* task priority has changed */
327 CFQ_CFQQ_FLAG_slice_new, /* no requests dispatched in slice */
328 CFQ_CFQQ_FLAG_sync, /* synchronous queue */
329 CFQ_CFQQ_FLAG_coop, /* cfqq is shared */
330 CFQ_CFQQ_FLAG_split_coop, /* shared cfqq will be splitted */
331 CFQ_CFQQ_FLAG_deep, /* sync cfqq experienced large depth */
332 CFQ_CFQQ_FLAG_wait_busy, /* Waiting for next request */
335 #define CFQ_CFQQ_FNS(name) \
336 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \
338 (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name); \
340 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \
342 (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \
344 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \
346 return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \
350 CFQ_CFQQ_FNS(wait_request);
351 CFQ_CFQQ_FNS(must_dispatch);
352 CFQ_CFQQ_FNS(must_alloc_slice);
353 CFQ_CFQQ_FNS(fifo_expire);
354 CFQ_CFQQ_FNS(idle_window);
355 CFQ_CFQQ_FNS(prio_changed);
356 CFQ_CFQQ_FNS(slice_new);
359 CFQ_CFQQ_FNS(split_coop);
361 CFQ_CFQQ_FNS(wait_busy);
364 #ifdef CONFIG_CFQ_GROUP_IOSCHED
365 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
366 blk_add_trace_msg((cfqd)->queue, "cfq%d%c %s " fmt, (cfqq)->pid, \
367 cfq_cfqq_sync((cfqq)) ? 'S' : 'A', \
368 blkg_path(&(cfqq)->cfqg->blkg), ##args);
370 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...) \
371 blk_add_trace_msg((cfqd)->queue, "%s " fmt, \
372 blkg_path(&(cfqg)->blkg), ##args); \
375 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
376 blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args)
377 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...) do {} while (0);
379 #define cfq_log(cfqd, fmt, args...) \
380 blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
382 /* Traverses through cfq group service trees */
383 #define for_each_cfqg_st(cfqg, i, j, st) \
384 for (i = 0; i <= IDLE_WORKLOAD; i++) \
385 for (j = 0, st = i < IDLE_WORKLOAD ? &cfqg->service_trees[i][j]\
386 : &cfqg->service_tree_idle; \
387 (i < IDLE_WORKLOAD && j <= SYNC_WORKLOAD) || \
388 (i == IDLE_WORKLOAD && j == 0); \
389 j++, st = i < IDLE_WORKLOAD ? \
390 &cfqg->service_trees[i][j]: NULL) \
393 static inline bool iops_mode(struct cfq_data *cfqd)
396 * If we are not idling on queues and it is a NCQ drive, parallel
397 * execution of requests is on and measuring time is not possible
398 * in most of the cases until and unless we drive shallower queue
399 * depths and that becomes a performance bottleneck. In such cases
400 * switch to start providing fairness in terms of number of IOs.
402 if (!cfqd->cfq_slice_idle && cfqd->hw_tag)
408 static inline enum wl_prio_t cfqq_prio(struct cfq_queue *cfqq)
410 if (cfq_class_idle(cfqq))
411 return IDLE_WORKLOAD;
412 if (cfq_class_rt(cfqq))
418 static enum wl_type_t cfqq_type(struct cfq_queue *cfqq)
420 if (!cfq_cfqq_sync(cfqq))
421 return ASYNC_WORKLOAD;
422 if (!cfq_cfqq_idle_window(cfqq))
423 return SYNC_NOIDLE_WORKLOAD;
424 return SYNC_WORKLOAD;
427 static inline int cfq_group_busy_queues_wl(enum wl_prio_t wl,
428 struct cfq_data *cfqd,
429 struct cfq_group *cfqg)
431 if (wl == IDLE_WORKLOAD)
432 return cfqg->service_tree_idle.count;
434 return cfqg->service_trees[wl][ASYNC_WORKLOAD].count
435 + cfqg->service_trees[wl][SYNC_NOIDLE_WORKLOAD].count
436 + cfqg->service_trees[wl][SYNC_WORKLOAD].count;
439 static inline int cfqg_busy_async_queues(struct cfq_data *cfqd,
440 struct cfq_group *cfqg)
442 return cfqg->service_trees[RT_WORKLOAD][ASYNC_WORKLOAD].count
443 + cfqg->service_trees[BE_WORKLOAD][ASYNC_WORKLOAD].count;
446 static void cfq_dispatch_insert(struct request_queue *, struct request *);
447 static struct cfq_queue *cfq_get_queue(struct cfq_data *, bool,
448 struct io_context *, gfp_t);
449 static struct cfq_io_context *cfq_cic_lookup(struct cfq_data *,
450 struct io_context *);
452 static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_context *cic,
455 return cic->cfqq[is_sync];
458 static inline void cic_set_cfqq(struct cfq_io_context *cic,
459 struct cfq_queue *cfqq, bool is_sync)
461 cic->cfqq[is_sync] = cfqq;
464 #define CIC_DEAD_KEY 1ul
465 #define CIC_DEAD_INDEX_SHIFT 1
467 static inline void *cfqd_dead_key(struct cfq_data *cfqd)
469 return (void *)(cfqd->cic_index << CIC_DEAD_INDEX_SHIFT | CIC_DEAD_KEY);
472 static inline struct cfq_data *cic_to_cfqd(struct cfq_io_context *cic)
474 struct cfq_data *cfqd = cic->key;
476 if (unlikely((unsigned long) cfqd & CIC_DEAD_KEY))
483 * We regard a request as SYNC, if it's either a read or has the SYNC bit
484 * set (in which case it could also be direct WRITE).
486 static inline bool cfq_bio_sync(struct bio *bio)
488 return bio_data_dir(bio) == READ || (bio->bi_rw & REQ_SYNC);
492 * scheduler run of queue, if there are requests pending and no one in the
493 * driver that will restart queueing
495 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
497 if (cfqd->busy_queues) {
498 cfq_log(cfqd, "schedule dispatch");
499 kblockd_schedule_work(cfqd->queue, &cfqd->unplug_work);
504 * Scale schedule slice based on io priority. Use the sync time slice only
505 * if a queue is marked sync and has sync io queued. A sync queue with async
506 * io only, should not get full sync slice length.
508 static inline int cfq_prio_slice(struct cfq_data *cfqd, bool sync,
511 const int base_slice = cfqd->cfq_slice[sync];
513 WARN_ON(prio >= IOPRIO_BE_NR);
515 return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
519 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
521 return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
524 static inline u64 cfq_scale_slice(unsigned long delta, struct cfq_group *cfqg)
526 u64 d = delta << CFQ_SERVICE_SHIFT;
528 d = d * BLKIO_WEIGHT_DEFAULT;
529 do_div(d, cfqg->weight);
533 static inline u64 max_vdisktime(u64 min_vdisktime, u64 vdisktime)
535 s64 delta = (s64)(vdisktime - min_vdisktime);
537 min_vdisktime = vdisktime;
539 return min_vdisktime;
542 static inline u64 min_vdisktime(u64 min_vdisktime, u64 vdisktime)
544 s64 delta = (s64)(vdisktime - min_vdisktime);
546 min_vdisktime = vdisktime;
548 return min_vdisktime;
551 static void update_min_vdisktime(struct cfq_rb_root *st)
553 struct cfq_group *cfqg;
556 cfqg = rb_entry_cfqg(st->left);
557 st->min_vdisktime = max_vdisktime(st->min_vdisktime,
563 * get averaged number of queues of RT/BE priority.
564 * average is updated, with a formula that gives more weight to higher numbers,
565 * to quickly follows sudden increases and decrease slowly
568 static inline unsigned cfq_group_get_avg_queues(struct cfq_data *cfqd,
569 struct cfq_group *cfqg, bool rt)
571 unsigned min_q, max_q;
572 unsigned mult = cfq_hist_divisor - 1;
573 unsigned round = cfq_hist_divisor / 2;
574 unsigned busy = cfq_group_busy_queues_wl(rt, cfqd, cfqg);
576 min_q = min(cfqg->busy_queues_avg[rt], busy);
577 max_q = max(cfqg->busy_queues_avg[rt], busy);
578 cfqg->busy_queues_avg[rt] = (mult * max_q + min_q + round) /
580 return cfqg->busy_queues_avg[rt];
583 static inline unsigned
584 cfq_group_slice(struct cfq_data *cfqd, struct cfq_group *cfqg)
586 struct cfq_rb_root *st = &cfqd->grp_service_tree;
588 return cfq_target_latency * cfqg->weight / st->total_weight;
591 static inline unsigned
592 cfq_scaled_cfqq_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
594 unsigned slice = cfq_prio_to_slice(cfqd, cfqq);
595 if (cfqd->cfq_latency) {
597 * interested queues (we consider only the ones with the same
598 * priority class in the cfq group)
600 unsigned iq = cfq_group_get_avg_queues(cfqd, cfqq->cfqg,
602 unsigned sync_slice = cfqd->cfq_slice[1];
603 unsigned expect_latency = sync_slice * iq;
604 unsigned group_slice = cfq_group_slice(cfqd, cfqq->cfqg);
606 if (expect_latency > group_slice) {
607 unsigned base_low_slice = 2 * cfqd->cfq_slice_idle;
608 /* scale low_slice according to IO priority
609 * and sync vs async */
611 min(slice, base_low_slice * slice / sync_slice);
612 /* the adapted slice value is scaled to fit all iqs
613 * into the target latency */
614 slice = max(slice * group_slice / expect_latency,
622 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
624 unsigned slice = cfq_scaled_cfqq_slice(cfqd, cfqq);
626 cfqq->slice_start = jiffies;
627 cfqq->slice_end = jiffies + slice;
628 cfqq->allocated_slice = slice;
629 cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
633 * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
634 * isn't valid until the first request from the dispatch is activated
635 * and the slice time set.
637 static inline bool cfq_slice_used(struct cfq_queue *cfqq)
639 if (cfq_cfqq_slice_new(cfqq))
641 if (time_before(jiffies, cfqq->slice_end))
648 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
649 * We choose the request that is closest to the head right now. Distance
650 * behind the head is penalized and only allowed to a certain extent.
652 static struct request *
653 cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2, sector_t last)
655 sector_t s1, s2, d1 = 0, d2 = 0;
656 unsigned long back_max;
657 #define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */
658 #define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */
659 unsigned wrap = 0; /* bit mask: requests behind the disk head? */
661 if (rq1 == NULL || rq1 == rq2)
666 if (rq_is_sync(rq1) && !rq_is_sync(rq2))
668 else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
670 if ((rq1->cmd_flags & REQ_META) && !(rq2->cmd_flags & REQ_META))
672 else if ((rq2->cmd_flags & REQ_META) &&
673 !(rq1->cmd_flags & REQ_META))
676 s1 = blk_rq_pos(rq1);
677 s2 = blk_rq_pos(rq2);
680 * by definition, 1KiB is 2 sectors
682 back_max = cfqd->cfq_back_max * 2;
685 * Strict one way elevator _except_ in the case where we allow
686 * short backward seeks which are biased as twice the cost of a
687 * similar forward seek.
691 else if (s1 + back_max >= last)
692 d1 = (last - s1) * cfqd->cfq_back_penalty;
694 wrap |= CFQ_RQ1_WRAP;
698 else if (s2 + back_max >= last)
699 d2 = (last - s2) * cfqd->cfq_back_penalty;
701 wrap |= CFQ_RQ2_WRAP;
703 /* Found required data */
706 * By doing switch() on the bit mask "wrap" we avoid having to
707 * check two variables for all permutations: --> faster!
710 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
726 case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
729 * Since both rqs are wrapped,
730 * start with the one that's further behind head
731 * (--> only *one* back seek required),
732 * since back seek takes more time than forward.
742 * The below is leftmost cache rbtree addon
744 static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
746 /* Service tree is empty */
751 root->left = rb_first(&root->rb);
754 return rb_entry(root->left, struct cfq_queue, rb_node);
759 static struct cfq_group *cfq_rb_first_group(struct cfq_rb_root *root)
762 root->left = rb_first(&root->rb);
765 return rb_entry_cfqg(root->left);
770 static void rb_erase_init(struct rb_node *n, struct rb_root *root)
776 static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
780 rb_erase_init(n, &root->rb);
785 * would be nice to take fifo expire time into account as well
787 static struct request *
788 cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
789 struct request *last)
791 struct rb_node *rbnext = rb_next(&last->rb_node);
792 struct rb_node *rbprev = rb_prev(&last->rb_node);
793 struct request *next = NULL, *prev = NULL;
795 BUG_ON(RB_EMPTY_NODE(&last->rb_node));
798 prev = rb_entry_rq(rbprev);
801 next = rb_entry_rq(rbnext);
803 rbnext = rb_first(&cfqq->sort_list);
804 if (rbnext && rbnext != &last->rb_node)
805 next = rb_entry_rq(rbnext);
808 return cfq_choose_req(cfqd, next, prev, blk_rq_pos(last));
811 static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
812 struct cfq_queue *cfqq)
815 * just an approximation, should be ok.
817 return (cfqq->cfqg->nr_cfqq - 1) * (cfq_prio_slice(cfqd, 1, 0) -
818 cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
822 cfqg_key(struct cfq_rb_root *st, struct cfq_group *cfqg)
824 return cfqg->vdisktime - st->min_vdisktime;
828 __cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
830 struct rb_node **node = &st->rb.rb_node;
831 struct rb_node *parent = NULL;
832 struct cfq_group *__cfqg;
833 s64 key = cfqg_key(st, cfqg);
836 while (*node != NULL) {
838 __cfqg = rb_entry_cfqg(parent);
840 if (key < cfqg_key(st, __cfqg))
841 node = &parent->rb_left;
843 node = &parent->rb_right;
849 st->left = &cfqg->rb_node;
851 rb_link_node(&cfqg->rb_node, parent, node);
852 rb_insert_color(&cfqg->rb_node, &st->rb);
856 cfq_group_service_tree_add(struct cfq_data *cfqd, struct cfq_group *cfqg)
858 struct cfq_rb_root *st = &cfqd->grp_service_tree;
859 struct cfq_group *__cfqg;
863 if (!RB_EMPTY_NODE(&cfqg->rb_node))
867 * Currently put the group at the end. Later implement something
868 * so that groups get lesser vtime based on their weights, so that
869 * if group does not loose all if it was not continously backlogged.
871 n = rb_last(&st->rb);
873 __cfqg = rb_entry_cfqg(n);
874 cfqg->vdisktime = __cfqg->vdisktime + CFQ_IDLE_DELAY;
876 cfqg->vdisktime = st->min_vdisktime;
878 __cfq_group_service_tree_add(st, cfqg);
879 st->total_weight += cfqg->weight;
883 cfq_group_service_tree_del(struct cfq_data *cfqd, struct cfq_group *cfqg)
885 struct cfq_rb_root *st = &cfqd->grp_service_tree;
887 BUG_ON(cfqg->nr_cfqq < 1);
890 /* If there are other cfq queues under this group, don't delete it */
894 cfq_log_cfqg(cfqd, cfqg, "del_from_rr group");
895 st->total_weight -= cfqg->weight;
896 if (!RB_EMPTY_NODE(&cfqg->rb_node))
897 cfq_rb_erase(&cfqg->rb_node, st);
898 cfqg->saved_workload_slice = 0;
899 cfq_blkiocg_update_dequeue_stats(&cfqg->blkg, 1);
902 static inline unsigned int cfq_cfqq_slice_usage(struct cfq_queue *cfqq)
904 unsigned int slice_used;
907 * Queue got expired before even a single request completed or
908 * got expired immediately after first request completion.
910 if (!cfqq->slice_start || cfqq->slice_start == jiffies) {
912 * Also charge the seek time incurred to the group, otherwise
913 * if there are mutiple queues in the group, each can dispatch
914 * a single request on seeky media and cause lots of seek time
915 * and group will never know it.
917 slice_used = max_t(unsigned, (jiffies - cfqq->dispatch_start),
920 slice_used = jiffies - cfqq->slice_start;
921 if (slice_used > cfqq->allocated_slice)
922 slice_used = cfqq->allocated_slice;
928 static void cfq_group_served(struct cfq_data *cfqd, struct cfq_group *cfqg,
929 struct cfq_queue *cfqq)
931 struct cfq_rb_root *st = &cfqd->grp_service_tree;
932 unsigned int used_sl, charge;
933 int nr_sync = cfqg->nr_cfqq - cfqg_busy_async_queues(cfqd, cfqg)
934 - cfqg->service_tree_idle.count;
937 used_sl = charge = cfq_cfqq_slice_usage(cfqq);
940 charge = cfqq->slice_dispatch;
941 else if (!cfq_cfqq_sync(cfqq) && !nr_sync)
942 charge = cfqq->allocated_slice;
944 /* Can't update vdisktime while group is on service tree */
945 cfq_rb_erase(&cfqg->rb_node, st);
946 cfqg->vdisktime += cfq_scale_slice(charge, cfqg);
947 __cfq_group_service_tree_add(st, cfqg);
949 /* This group is being expired. Save the context */
950 if (time_after(cfqd->workload_expires, jiffies)) {
951 cfqg->saved_workload_slice = cfqd->workload_expires
953 cfqg->saved_workload = cfqd->serving_type;
954 cfqg->saved_serving_prio = cfqd->serving_prio;
956 cfqg->saved_workload_slice = 0;
958 cfq_log_cfqg(cfqd, cfqg, "served: vt=%llu min_vt=%llu", cfqg->vdisktime,
960 cfq_log_cfqq(cfqq->cfqd, cfqq, "sl_used=%u disp=%u charge=%u iops=%u"
961 " sect=%u", used_sl, cfqq->slice_dispatch, charge,
962 iops_mode(cfqd), cfqq->nr_sectors);
963 cfq_blkiocg_update_timeslice_used(&cfqg->blkg, used_sl);
964 cfq_blkiocg_set_start_empty_time(&cfqg->blkg);
967 #ifdef CONFIG_CFQ_GROUP_IOSCHED
968 static inline struct cfq_group *cfqg_of_blkg(struct blkio_group *blkg)
971 return container_of(blkg, struct cfq_group, blkg);
975 void cfq_update_blkio_group_weight(void *key, struct blkio_group *blkg,
978 cfqg_of_blkg(blkg)->weight = weight;
981 static struct cfq_group *
982 cfq_find_alloc_cfqg(struct cfq_data *cfqd, struct cgroup *cgroup, int create)
984 struct blkio_cgroup *blkcg = cgroup_to_blkio_cgroup(cgroup);
985 struct cfq_group *cfqg = NULL;
988 struct cfq_rb_root *st;
989 struct backing_dev_info *bdi = &cfqd->queue->backing_dev_info;
990 unsigned int major, minor;
992 cfqg = cfqg_of_blkg(blkiocg_lookup_group(blkcg, key));
993 if (cfqg && !cfqg->blkg.dev && bdi->dev && dev_name(bdi->dev)) {
994 sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
995 cfqg->blkg.dev = MKDEV(major, minor);
1001 cfqg = kzalloc_node(sizeof(*cfqg), GFP_ATOMIC, cfqd->queue->node);
1005 for_each_cfqg_st(cfqg, i, j, st)
1007 RB_CLEAR_NODE(&cfqg->rb_node);
1010 * Take the initial reference that will be released on destroy
1011 * This can be thought of a joint reference by cgroup and
1012 * elevator which will be dropped by either elevator exit
1013 * or cgroup deletion path depending on who is exiting first.
1018 * Add group onto cgroup list. It might happen that bdi->dev is
1019 * not initialized yet. Initialize this new group without major
1020 * and minor info and this info will be filled in once a new thread
1021 * comes for IO. See code above.
1024 sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
1025 cfq_blkiocg_add_blkio_group(blkcg, &cfqg->blkg, (void *)cfqd,
1026 MKDEV(major, minor));
1028 cfq_blkiocg_add_blkio_group(blkcg, &cfqg->blkg, (void *)cfqd,
1031 cfqg->weight = blkcg_get_weight(blkcg, cfqg->blkg.dev);
1033 /* Add group on cfqd list */
1034 hlist_add_head(&cfqg->cfqd_node, &cfqd->cfqg_list);
1041 * Search for the cfq group current task belongs to. If create = 1, then also
1042 * create the cfq group if it does not exist. request_queue lock must be held.
1044 static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd, int create)
1046 struct cgroup *cgroup;
1047 struct cfq_group *cfqg = NULL;
1050 cgroup = task_cgroup(current, blkio_subsys_id);
1051 cfqg = cfq_find_alloc_cfqg(cfqd, cgroup, create);
1052 if (!cfqg && create)
1053 cfqg = &cfqd->root_group;
1058 static inline struct cfq_group *cfq_ref_get_cfqg(struct cfq_group *cfqg)
1064 static void cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg)
1066 /* Currently, all async queues are mapped to root group */
1067 if (!cfq_cfqq_sync(cfqq))
1068 cfqg = &cfqq->cfqd->root_group;
1071 /* cfqq reference on cfqg */
1075 static void cfq_put_cfqg(struct cfq_group *cfqg)
1077 struct cfq_rb_root *st;
1080 BUG_ON(cfqg->ref <= 0);
1084 for_each_cfqg_st(cfqg, i, j, st)
1085 BUG_ON(!RB_EMPTY_ROOT(&st->rb));
1089 static void cfq_destroy_cfqg(struct cfq_data *cfqd, struct cfq_group *cfqg)
1091 /* Something wrong if we are trying to remove same group twice */
1092 BUG_ON(hlist_unhashed(&cfqg->cfqd_node));
1094 hlist_del_init(&cfqg->cfqd_node);
1097 * Put the reference taken at the time of creation so that when all
1098 * queues are gone, group can be destroyed.
1103 static void cfq_release_cfq_groups(struct cfq_data *cfqd)
1105 struct hlist_node *pos, *n;
1106 struct cfq_group *cfqg;
1108 hlist_for_each_entry_safe(cfqg, pos, n, &cfqd->cfqg_list, cfqd_node) {
1110 * If cgroup removal path got to blk_group first and removed
1111 * it from cgroup list, then it will take care of destroying
1114 if (!cfq_blkiocg_del_blkio_group(&cfqg->blkg))
1115 cfq_destroy_cfqg(cfqd, cfqg);
1120 * Blk cgroup controller notification saying that blkio_group object is being
1121 * delinked as associated cgroup object is going away. That also means that
1122 * no new IO will come in this group. So get rid of this group as soon as
1123 * any pending IO in the group is finished.
1125 * This function is called under rcu_read_lock(). key is the rcu protected
1126 * pointer. That means "key" is a valid cfq_data pointer as long as we are rcu
1129 * "key" was fetched from blkio_group under blkio_cgroup->lock. That means
1130 * it should not be NULL as even if elevator was exiting, cgroup deltion
1131 * path got to it first.
1133 void cfq_unlink_blkio_group(void *key, struct blkio_group *blkg)
1135 unsigned long flags;
1136 struct cfq_data *cfqd = key;
1138 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1139 cfq_destroy_cfqg(cfqd, cfqg_of_blkg(blkg));
1140 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1143 #else /* GROUP_IOSCHED */
1144 static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd, int create)
1146 return &cfqd->root_group;
1149 static inline struct cfq_group *cfq_ref_get_cfqg(struct cfq_group *cfqg)
1155 cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) {
1159 static void cfq_release_cfq_groups(struct cfq_data *cfqd) {}
1160 static inline void cfq_put_cfqg(struct cfq_group *cfqg) {}
1162 #endif /* GROUP_IOSCHED */
1165 * The cfqd->service_trees holds all pending cfq_queue's that have
1166 * requests waiting to be processed. It is sorted in the order that
1167 * we will service the queues.
1169 static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1172 struct rb_node **p, *parent;
1173 struct cfq_queue *__cfqq;
1174 unsigned long rb_key;
1175 struct cfq_rb_root *service_tree;
1178 int group_changed = 0;
1180 service_tree = service_tree_for(cfqq->cfqg, cfqq_prio(cfqq),
1182 if (cfq_class_idle(cfqq)) {
1183 rb_key = CFQ_IDLE_DELAY;
1184 parent = rb_last(&service_tree->rb);
1185 if (parent && parent != &cfqq->rb_node) {
1186 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1187 rb_key += __cfqq->rb_key;
1190 } else if (!add_front) {
1192 * Get our rb key offset. Subtract any residual slice
1193 * value carried from last service. A negative resid
1194 * count indicates slice overrun, and this should position
1195 * the next service time further away in the tree.
1197 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
1198 rb_key -= cfqq->slice_resid;
1199 cfqq->slice_resid = 0;
1202 __cfqq = cfq_rb_first(service_tree);
1203 rb_key += __cfqq ? __cfqq->rb_key : jiffies;
1206 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1209 * same position, nothing more to do
1211 if (rb_key == cfqq->rb_key &&
1212 cfqq->service_tree == service_tree)
1215 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1216 cfqq->service_tree = NULL;
1221 cfqq->service_tree = service_tree;
1222 p = &service_tree->rb.rb_node;
1227 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1230 * sort by key, that represents service time.
1232 if (time_before(rb_key, __cfqq->rb_key))
1235 n = &(*p)->rb_right;
1243 service_tree->left = &cfqq->rb_node;
1245 cfqq->rb_key = rb_key;
1246 rb_link_node(&cfqq->rb_node, parent, p);
1247 rb_insert_color(&cfqq->rb_node, &service_tree->rb);
1248 service_tree->count++;
1249 if ((add_front || !new_cfqq) && !group_changed)
1251 cfq_group_service_tree_add(cfqd, cfqq->cfqg);
1254 static struct cfq_queue *
1255 cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
1256 sector_t sector, struct rb_node **ret_parent,
1257 struct rb_node ***rb_link)
1259 struct rb_node **p, *parent;
1260 struct cfq_queue *cfqq = NULL;
1268 cfqq = rb_entry(parent, struct cfq_queue, p_node);
1271 * Sort strictly based on sector. Smallest to the left,
1272 * largest to the right.
1274 if (sector > blk_rq_pos(cfqq->next_rq))
1275 n = &(*p)->rb_right;
1276 else if (sector < blk_rq_pos(cfqq->next_rq))
1284 *ret_parent = parent;
1290 static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1292 struct rb_node **p, *parent;
1293 struct cfq_queue *__cfqq;
1296 rb_erase(&cfqq->p_node, cfqq->p_root);
1297 cfqq->p_root = NULL;
1300 if (cfq_class_idle(cfqq))
1305 cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
1306 __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
1307 blk_rq_pos(cfqq->next_rq), &parent, &p);
1309 rb_link_node(&cfqq->p_node, parent, p);
1310 rb_insert_color(&cfqq->p_node, cfqq->p_root);
1312 cfqq->p_root = NULL;
1316 * Update cfqq's position in the service tree.
1318 static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1321 * Resorting requires the cfqq to be on the RR list already.
1323 if (cfq_cfqq_on_rr(cfqq)) {
1324 cfq_service_tree_add(cfqd, cfqq, 0);
1325 cfq_prio_tree_add(cfqd, cfqq);
1330 * add to busy list of queues for service, trying to be fair in ordering
1331 * the pending list according to last request service
1333 static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1335 cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
1336 BUG_ON(cfq_cfqq_on_rr(cfqq));
1337 cfq_mark_cfqq_on_rr(cfqq);
1338 cfqd->busy_queues++;
1339 if (cfq_cfqq_sync(cfqq))
1340 cfqd->busy_sync_queues++;
1342 cfq_resort_rr_list(cfqd, cfqq);
1346 * Called when the cfqq no longer has requests pending, remove it from
1349 static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1351 cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
1352 BUG_ON(!cfq_cfqq_on_rr(cfqq));
1353 cfq_clear_cfqq_on_rr(cfqq);
1355 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1356 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1357 cfqq->service_tree = NULL;
1360 rb_erase(&cfqq->p_node, cfqq->p_root);
1361 cfqq->p_root = NULL;
1364 cfq_group_service_tree_del(cfqd, cfqq->cfqg);
1365 BUG_ON(!cfqd->busy_queues);
1366 cfqd->busy_queues--;
1367 if (cfq_cfqq_sync(cfqq))
1368 cfqd->busy_sync_queues--;
1372 * rb tree support functions
1374 static void cfq_del_rq_rb(struct request *rq)
1376 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1377 const int sync = rq_is_sync(rq);
1379 BUG_ON(!cfqq->queued[sync]);
1380 cfqq->queued[sync]--;
1382 elv_rb_del(&cfqq->sort_list, rq);
1384 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) {
1386 * Queue will be deleted from service tree when we actually
1387 * expire it later. Right now just remove it from prio tree
1391 rb_erase(&cfqq->p_node, cfqq->p_root);
1392 cfqq->p_root = NULL;
1397 static void cfq_add_rq_rb(struct request *rq)
1399 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1400 struct cfq_data *cfqd = cfqq->cfqd;
1401 struct request *__alias, *prev;
1403 cfqq->queued[rq_is_sync(rq)]++;
1406 * looks a little odd, but the first insert might return an alias.
1407 * if that happens, put the alias on the dispatch list
1409 while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL)
1410 cfq_dispatch_insert(cfqd->queue, __alias);
1412 if (!cfq_cfqq_on_rr(cfqq))
1413 cfq_add_cfqq_rr(cfqd, cfqq);
1416 * check if this request is a better next-serve candidate
1418 prev = cfqq->next_rq;
1419 cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq, cfqd->last_position);
1422 * adjust priority tree position, if ->next_rq changes
1424 if (prev != cfqq->next_rq)
1425 cfq_prio_tree_add(cfqd, cfqq);
1427 BUG_ON(!cfqq->next_rq);
1430 static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
1432 elv_rb_del(&cfqq->sort_list, rq);
1433 cfqq->queued[rq_is_sync(rq)]--;
1434 cfq_blkiocg_update_io_remove_stats(&(RQ_CFQG(rq))->blkg,
1435 rq_data_dir(rq), rq_is_sync(rq));
1437 cfq_blkiocg_update_io_add_stats(&(RQ_CFQG(rq))->blkg,
1438 &cfqq->cfqd->serving_group->blkg, rq_data_dir(rq),
1442 static struct request *
1443 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
1445 struct task_struct *tsk = current;
1446 struct cfq_io_context *cic;
1447 struct cfq_queue *cfqq;
1449 cic = cfq_cic_lookup(cfqd, tsk->io_context);
1453 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
1455 sector_t sector = bio->bi_sector + bio_sectors(bio);
1457 return elv_rb_find(&cfqq->sort_list, sector);
1463 static void cfq_activate_request(struct request_queue *q, struct request *rq)
1465 struct cfq_data *cfqd = q->elevator->elevator_data;
1467 cfqd->rq_in_driver++;
1468 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
1469 cfqd->rq_in_driver);
1471 cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
1474 static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
1476 struct cfq_data *cfqd = q->elevator->elevator_data;
1478 WARN_ON(!cfqd->rq_in_driver);
1479 cfqd->rq_in_driver--;
1480 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
1481 cfqd->rq_in_driver);
1484 static void cfq_remove_request(struct request *rq)
1486 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1488 if (cfqq->next_rq == rq)
1489 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
1491 list_del_init(&rq->queuelist);
1494 cfqq->cfqd->rq_queued--;
1495 cfq_blkiocg_update_io_remove_stats(&(RQ_CFQG(rq))->blkg,
1496 rq_data_dir(rq), rq_is_sync(rq));
1497 if (rq->cmd_flags & REQ_META) {
1498 WARN_ON(!cfqq->meta_pending);
1499 cfqq->meta_pending--;
1503 static int cfq_merge(struct request_queue *q, struct request **req,
1506 struct cfq_data *cfqd = q->elevator->elevator_data;
1507 struct request *__rq;
1509 __rq = cfq_find_rq_fmerge(cfqd, bio);
1510 if (__rq && elv_rq_merge_ok(__rq, bio)) {
1512 return ELEVATOR_FRONT_MERGE;
1515 return ELEVATOR_NO_MERGE;
1518 static void cfq_merged_request(struct request_queue *q, struct request *req,
1521 if (type == ELEVATOR_FRONT_MERGE) {
1522 struct cfq_queue *cfqq = RQ_CFQQ(req);
1524 cfq_reposition_rq_rb(cfqq, req);
1528 static void cfq_bio_merged(struct request_queue *q, struct request *req,
1531 cfq_blkiocg_update_io_merged_stats(&(RQ_CFQG(req))->blkg,
1532 bio_data_dir(bio), cfq_bio_sync(bio));
1536 cfq_merged_requests(struct request_queue *q, struct request *rq,
1537 struct request *next)
1539 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1541 * reposition in fifo if next is older than rq
1543 if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
1544 time_before(rq_fifo_time(next), rq_fifo_time(rq))) {
1545 list_move(&rq->queuelist, &next->queuelist);
1546 rq_set_fifo_time(rq, rq_fifo_time(next));
1549 if (cfqq->next_rq == next)
1551 cfq_remove_request(next);
1552 cfq_blkiocg_update_io_merged_stats(&(RQ_CFQG(rq))->blkg,
1553 rq_data_dir(next), rq_is_sync(next));
1556 static int cfq_allow_merge(struct request_queue *q, struct request *rq,
1559 struct cfq_data *cfqd = q->elevator->elevator_data;
1560 struct cfq_io_context *cic;
1561 struct cfq_queue *cfqq;
1564 * Disallow merge of a sync bio into an async request.
1566 if (cfq_bio_sync(bio) && !rq_is_sync(rq))
1570 * Lookup the cfqq that this bio will be queued with. Allow
1571 * merge only if rq is queued there.
1573 cic = cfq_cic_lookup(cfqd, current->io_context);
1577 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
1578 return cfqq == RQ_CFQQ(rq);
1581 static inline void cfq_del_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1583 del_timer(&cfqd->idle_slice_timer);
1584 cfq_blkiocg_update_idle_time_stats(&cfqq->cfqg->blkg);
1587 static void __cfq_set_active_queue(struct cfq_data *cfqd,
1588 struct cfq_queue *cfqq)
1591 cfq_log_cfqq(cfqd, cfqq, "set_active wl_prio:%d wl_type:%d",
1592 cfqd->serving_prio, cfqd->serving_type);
1593 cfq_blkiocg_update_avg_queue_size_stats(&cfqq->cfqg->blkg);
1594 cfqq->slice_start = 0;
1595 cfqq->dispatch_start = jiffies;
1596 cfqq->allocated_slice = 0;
1597 cfqq->slice_end = 0;
1598 cfqq->slice_dispatch = 0;
1599 cfqq->nr_sectors = 0;
1601 cfq_clear_cfqq_wait_request(cfqq);
1602 cfq_clear_cfqq_must_dispatch(cfqq);
1603 cfq_clear_cfqq_must_alloc_slice(cfqq);
1604 cfq_clear_cfqq_fifo_expire(cfqq);
1605 cfq_mark_cfqq_slice_new(cfqq);
1607 cfq_del_timer(cfqd, cfqq);
1610 cfqd->active_queue = cfqq;
1614 * current cfqq expired its slice (or was too idle), select new one
1617 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1620 cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
1622 if (cfq_cfqq_wait_request(cfqq))
1623 cfq_del_timer(cfqd, cfqq);
1625 cfq_clear_cfqq_wait_request(cfqq);
1626 cfq_clear_cfqq_wait_busy(cfqq);
1629 * If this cfqq is shared between multiple processes, check to
1630 * make sure that those processes are still issuing I/Os within
1631 * the mean seek distance. If not, it may be time to break the
1632 * queues apart again.
1634 if (cfq_cfqq_coop(cfqq) && CFQQ_SEEKY(cfqq))
1635 cfq_mark_cfqq_split_coop(cfqq);
1638 * store what was left of this slice, if the queue idled/timed out
1641 if (cfq_cfqq_slice_new(cfqq))
1642 cfqq->slice_resid = cfq_scaled_cfqq_slice(cfqd, cfqq);
1644 cfqq->slice_resid = cfqq->slice_end - jiffies;
1645 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
1648 cfq_group_served(cfqd, cfqq->cfqg, cfqq);
1650 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
1651 cfq_del_cfqq_rr(cfqd, cfqq);
1653 cfq_resort_rr_list(cfqd, cfqq);
1655 if (cfqq == cfqd->active_queue)
1656 cfqd->active_queue = NULL;
1658 if (cfqd->active_cic) {
1659 put_io_context(cfqd->active_cic->ioc);
1660 cfqd->active_cic = NULL;
1664 static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out)
1666 struct cfq_queue *cfqq = cfqd->active_queue;
1669 __cfq_slice_expired(cfqd, cfqq, timed_out);
1673 * Get next queue for service. Unless we have a queue preemption,
1674 * we'll simply select the first cfqq in the service tree.
1676 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
1678 struct cfq_rb_root *service_tree =
1679 service_tree_for(cfqd->serving_group, cfqd->serving_prio,
1680 cfqd->serving_type);
1682 if (!cfqd->rq_queued)
1685 /* There is nothing to dispatch */
1688 if (RB_EMPTY_ROOT(&service_tree->rb))
1690 return cfq_rb_first(service_tree);
1693 static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd)
1695 struct cfq_group *cfqg;
1696 struct cfq_queue *cfqq;
1698 struct cfq_rb_root *st;
1700 if (!cfqd->rq_queued)
1703 cfqg = cfq_get_next_cfqg(cfqd);
1707 for_each_cfqg_st(cfqg, i, j, st)
1708 if ((cfqq = cfq_rb_first(st)) != NULL)
1714 * Get and set a new active queue for service.
1716 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
1717 struct cfq_queue *cfqq)
1720 cfqq = cfq_get_next_queue(cfqd);
1722 __cfq_set_active_queue(cfqd, cfqq);
1726 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
1729 if (blk_rq_pos(rq) >= cfqd->last_position)
1730 return blk_rq_pos(rq) - cfqd->last_position;
1732 return cfqd->last_position - blk_rq_pos(rq);
1735 static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1738 return cfq_dist_from_last(cfqd, rq) <= CFQQ_CLOSE_THR;
1741 static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
1742 struct cfq_queue *cur_cfqq)
1744 struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
1745 struct rb_node *parent, *node;
1746 struct cfq_queue *__cfqq;
1747 sector_t sector = cfqd->last_position;
1749 if (RB_EMPTY_ROOT(root))
1753 * First, if we find a request starting at the end of the last
1754 * request, choose it.
1756 __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
1761 * If the exact sector wasn't found, the parent of the NULL leaf
1762 * will contain the closest sector.
1764 __cfqq = rb_entry(parent, struct cfq_queue, p_node);
1765 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1768 if (blk_rq_pos(__cfqq->next_rq) < sector)
1769 node = rb_next(&__cfqq->p_node);
1771 node = rb_prev(&__cfqq->p_node);
1775 __cfqq = rb_entry(node, struct cfq_queue, p_node);
1776 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1784 * cur_cfqq - passed in so that we don't decide that the current queue is
1785 * closely cooperating with itself.
1787 * So, basically we're assuming that that cur_cfqq has dispatched at least
1788 * one request, and that cfqd->last_position reflects a position on the disk
1789 * associated with the I/O issued by cur_cfqq. I'm not sure this is a valid
1792 static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
1793 struct cfq_queue *cur_cfqq)
1795 struct cfq_queue *cfqq;
1797 if (cfq_class_idle(cur_cfqq))
1799 if (!cfq_cfqq_sync(cur_cfqq))
1801 if (CFQQ_SEEKY(cur_cfqq))
1805 * Don't search priority tree if it's the only queue in the group.
1807 if (cur_cfqq->cfqg->nr_cfqq == 1)
1811 * We should notice if some of the queues are cooperating, eg
1812 * working closely on the same area of the disk. In that case,
1813 * we can group them together and don't waste time idling.
1815 cfqq = cfqq_close(cfqd, cur_cfqq);
1819 /* If new queue belongs to different cfq_group, don't choose it */
1820 if (cur_cfqq->cfqg != cfqq->cfqg)
1824 * It only makes sense to merge sync queues.
1826 if (!cfq_cfqq_sync(cfqq))
1828 if (CFQQ_SEEKY(cfqq))
1832 * Do not merge queues of different priority classes
1834 if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq))
1841 * Determine whether we should enforce idle window for this queue.
1844 static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1846 enum wl_prio_t prio = cfqq_prio(cfqq);
1847 struct cfq_rb_root *service_tree = cfqq->service_tree;
1849 BUG_ON(!service_tree);
1850 BUG_ON(!service_tree->count);
1852 if (!cfqd->cfq_slice_idle)
1855 /* We never do for idle class queues. */
1856 if (prio == IDLE_WORKLOAD)
1859 /* We do for queues that were marked with idle window flag. */
1860 if (cfq_cfqq_idle_window(cfqq) &&
1861 !(blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag))
1865 * Otherwise, we do only if they are the last ones
1866 * in their service tree.
1868 if (service_tree->count == 1 && cfq_cfqq_sync(cfqq))
1870 cfq_log_cfqq(cfqd, cfqq, "Not idling. st->count:%d",
1871 service_tree->count);
1875 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
1877 struct cfq_queue *cfqq = cfqd->active_queue;
1878 struct cfq_io_context *cic;
1879 unsigned long sl, group_idle = 0;
1882 * SSD device without seek penalty, disable idling. But only do so
1883 * for devices that support queuing, otherwise we still have a problem
1884 * with sync vs async workloads.
1886 if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
1889 WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
1890 WARN_ON(cfq_cfqq_slice_new(cfqq));
1893 * idle is disabled, either manually or by past process history
1895 if (!cfq_should_idle(cfqd, cfqq)) {
1896 /* no queue idling. Check for group idling */
1897 if (cfqd->cfq_group_idle)
1898 group_idle = cfqd->cfq_group_idle;
1904 * still active requests from this queue, don't idle
1906 if (cfqq->dispatched)
1910 * task has exited, don't wait
1912 cic = cfqd->active_cic;
1913 if (!cic || !atomic_read(&cic->ioc->nr_tasks))
1917 * If our average think time is larger than the remaining time
1918 * slice, then don't idle. This avoids overrunning the allotted
1921 if (sample_valid(cic->ttime_samples) &&
1922 (cfqq->slice_end - jiffies < cic->ttime_mean)) {
1923 cfq_log_cfqq(cfqd, cfqq, "Not idling. think_time:%d",
1928 /* There are other queues in the group, don't do group idle */
1929 if (group_idle && cfqq->cfqg->nr_cfqq > 1)
1932 cfq_mark_cfqq_wait_request(cfqq);
1935 sl = cfqd->cfq_group_idle;
1937 sl = cfqd->cfq_slice_idle;
1939 mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
1940 cfq_blkiocg_update_set_idle_time_stats(&cfqq->cfqg->blkg);
1941 cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu group_idle: %d", sl,
1942 group_idle ? 1 : 0);
1946 * Move request from internal lists to the request queue dispatch list.
1948 static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
1950 struct cfq_data *cfqd = q->elevator->elevator_data;
1951 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1953 cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
1955 cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
1956 cfq_remove_request(rq);
1958 (RQ_CFQG(rq))->dispatched++;
1959 elv_dispatch_sort(q, rq);
1961 cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]++;
1962 cfqq->nr_sectors += blk_rq_sectors(rq);
1963 cfq_blkiocg_update_dispatch_stats(&cfqq->cfqg->blkg, blk_rq_bytes(rq),
1964 rq_data_dir(rq), rq_is_sync(rq));
1968 * return expired entry, or NULL to just start from scratch in rbtree
1970 static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
1972 struct request *rq = NULL;
1974 if (cfq_cfqq_fifo_expire(cfqq))
1977 cfq_mark_cfqq_fifo_expire(cfqq);
1979 if (list_empty(&cfqq->fifo))
1982 rq = rq_entry_fifo(cfqq->fifo.next);
1983 if (time_before(jiffies, rq_fifo_time(rq)))
1986 cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq);
1991 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1993 const int base_rq = cfqd->cfq_slice_async_rq;
1995 WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
1997 return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
2001 * Must be called with the queue_lock held.
2003 static int cfqq_process_refs(struct cfq_queue *cfqq)
2005 int process_refs, io_refs;
2007 io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE];
2008 process_refs = cfqq->ref - io_refs;
2009 BUG_ON(process_refs < 0);
2010 return process_refs;
2013 static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq)
2015 int process_refs, new_process_refs;
2016 struct cfq_queue *__cfqq;
2019 * If there are no process references on the new_cfqq, then it is
2020 * unsafe to follow the ->new_cfqq chain as other cfqq's in the
2021 * chain may have dropped their last reference (not just their
2022 * last process reference).
2024 if (!cfqq_process_refs(new_cfqq))
2027 /* Avoid a circular list and skip interim queue merges */
2028 while ((__cfqq = new_cfqq->new_cfqq)) {
2034 process_refs = cfqq_process_refs(cfqq);
2035 new_process_refs = cfqq_process_refs(new_cfqq);
2037 * If the process for the cfqq has gone away, there is no
2038 * sense in merging the queues.
2040 if (process_refs == 0 || new_process_refs == 0)
2044 * Merge in the direction of the lesser amount of work.
2046 if (new_process_refs >= process_refs) {
2047 cfqq->new_cfqq = new_cfqq;
2048 new_cfqq->ref += process_refs;
2050 new_cfqq->new_cfqq = cfqq;
2051 cfqq->ref += new_process_refs;
2055 static enum wl_type_t cfq_choose_wl(struct cfq_data *cfqd,
2056 struct cfq_group *cfqg, enum wl_prio_t prio)
2058 struct cfq_queue *queue;
2060 bool key_valid = false;
2061 unsigned long lowest_key = 0;
2062 enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD;
2064 for (i = 0; i <= SYNC_WORKLOAD; ++i) {
2065 /* select the one with lowest rb_key */
2066 queue = cfq_rb_first(service_tree_for(cfqg, prio, i));
2068 (!key_valid || time_before(queue->rb_key, lowest_key))) {
2069 lowest_key = queue->rb_key;
2078 static void choose_service_tree(struct cfq_data *cfqd, struct cfq_group *cfqg)
2082 struct cfq_rb_root *st;
2083 unsigned group_slice;
2084 enum wl_prio_t original_prio = cfqd->serving_prio;
2086 /* Choose next priority. RT > BE > IDLE */
2087 if (cfq_group_busy_queues_wl(RT_WORKLOAD, cfqd, cfqg))
2088 cfqd->serving_prio = RT_WORKLOAD;
2089 else if (cfq_group_busy_queues_wl(BE_WORKLOAD, cfqd, cfqg))
2090 cfqd->serving_prio = BE_WORKLOAD;
2092 cfqd->serving_prio = IDLE_WORKLOAD;
2093 cfqd->workload_expires = jiffies + 1;
2097 if (original_prio != cfqd->serving_prio)
2101 * For RT and BE, we have to choose also the type
2102 * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
2105 st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
2109 * check workload expiration, and that we still have other queues ready
2111 if (count && !time_after(jiffies, cfqd->workload_expires))
2115 /* otherwise select new workload type */
2116 cfqd->serving_type =
2117 cfq_choose_wl(cfqd, cfqg, cfqd->serving_prio);
2118 st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
2122 * the workload slice is computed as a fraction of target latency
2123 * proportional to the number of queues in that workload, over
2124 * all the queues in the same priority class
2126 group_slice = cfq_group_slice(cfqd, cfqg);
2128 slice = group_slice * count /
2129 max_t(unsigned, cfqg->busy_queues_avg[cfqd->serving_prio],
2130 cfq_group_busy_queues_wl(cfqd->serving_prio, cfqd, cfqg));
2132 if (cfqd->serving_type == ASYNC_WORKLOAD) {
2136 * Async queues are currently system wide. Just taking
2137 * proportion of queues with-in same group will lead to higher
2138 * async ratio system wide as generally root group is going
2139 * to have higher weight. A more accurate thing would be to
2140 * calculate system wide asnc/sync ratio.
2142 tmp = cfq_target_latency * cfqg_busy_async_queues(cfqd, cfqg);
2143 tmp = tmp/cfqd->busy_queues;
2144 slice = min_t(unsigned, slice, tmp);
2146 /* async workload slice is scaled down according to
2147 * the sync/async slice ratio. */
2148 slice = slice * cfqd->cfq_slice[0] / cfqd->cfq_slice[1];
2150 /* sync workload slice is at least 2 * cfq_slice_idle */
2151 slice = max(slice, 2 * cfqd->cfq_slice_idle);
2153 slice = max_t(unsigned, slice, CFQ_MIN_TT);
2154 cfq_log(cfqd, "workload slice:%d", slice);
2155 cfqd->workload_expires = jiffies + slice;
2158 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd)
2160 struct cfq_rb_root *st = &cfqd->grp_service_tree;
2161 struct cfq_group *cfqg;
2163 if (RB_EMPTY_ROOT(&st->rb))
2165 cfqg = cfq_rb_first_group(st);
2166 update_min_vdisktime(st);
2170 static void cfq_choose_cfqg(struct cfq_data *cfqd)
2172 struct cfq_group *cfqg = cfq_get_next_cfqg(cfqd);
2174 cfqd->serving_group = cfqg;
2176 /* Restore the workload type data */
2177 if (cfqg->saved_workload_slice) {
2178 cfqd->workload_expires = jiffies + cfqg->saved_workload_slice;
2179 cfqd->serving_type = cfqg->saved_workload;
2180 cfqd->serving_prio = cfqg->saved_serving_prio;
2182 cfqd->workload_expires = jiffies - 1;
2184 choose_service_tree(cfqd, cfqg);
2188 * Select a queue for service. If we have a current active queue,
2189 * check whether to continue servicing it, or retrieve and set a new one.
2191 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
2193 struct cfq_queue *cfqq, *new_cfqq = NULL;
2195 cfqq = cfqd->active_queue;
2199 if (!cfqd->rq_queued)
2203 * We were waiting for group to get backlogged. Expire the queue
2205 if (cfq_cfqq_wait_busy(cfqq) && !RB_EMPTY_ROOT(&cfqq->sort_list))
2209 * The active queue has run out of time, expire it and select new.
2211 if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq)) {
2213 * If slice had not expired at the completion of last request
2214 * we might not have turned on wait_busy flag. Don't expire
2215 * the queue yet. Allow the group to get backlogged.
2217 * The very fact that we have used the slice, that means we
2218 * have been idling all along on this queue and it should be
2219 * ok to wait for this request to complete.
2221 if (cfqq->cfqg->nr_cfqq == 1 && RB_EMPTY_ROOT(&cfqq->sort_list)
2222 && cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2226 goto check_group_idle;
2230 * The active queue has requests and isn't expired, allow it to
2233 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
2237 * If another queue has a request waiting within our mean seek
2238 * distance, let it run. The expire code will check for close
2239 * cooperators and put the close queue at the front of the service
2240 * tree. If possible, merge the expiring queue with the new cfqq.
2242 new_cfqq = cfq_close_cooperator(cfqd, cfqq);
2244 if (!cfqq->new_cfqq)
2245 cfq_setup_merge(cfqq, new_cfqq);
2250 * No requests pending. If the active queue still has requests in
2251 * flight or is idling for a new request, allow either of these
2252 * conditions to happen (or time out) before selecting a new queue.
2254 if (timer_pending(&cfqd->idle_slice_timer)) {
2260 * This is a deep seek queue, but the device is much faster than
2261 * the queue can deliver, don't idle
2263 if (CFQQ_SEEKY(cfqq) && cfq_cfqq_idle_window(cfqq) &&
2264 (cfq_cfqq_slice_new(cfqq) ||
2265 (cfqq->slice_end - jiffies > jiffies - cfqq->slice_start))) {
2266 cfq_clear_cfqq_deep(cfqq);
2267 cfq_clear_cfqq_idle_window(cfqq);
2270 if (cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2276 * If group idle is enabled and there are requests dispatched from
2277 * this group, wait for requests to complete.
2280 if (cfqd->cfq_group_idle && cfqq->cfqg->nr_cfqq == 1
2281 && cfqq->cfqg->dispatched) {
2287 cfq_slice_expired(cfqd, 0);
2290 * Current queue expired. Check if we have to switch to a new
2294 cfq_choose_cfqg(cfqd);
2296 cfqq = cfq_set_active_queue(cfqd, new_cfqq);
2301 static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
2305 while (cfqq->next_rq) {
2306 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
2310 BUG_ON(!list_empty(&cfqq->fifo));
2312 /* By default cfqq is not expired if it is empty. Do it explicitly */
2313 __cfq_slice_expired(cfqq->cfqd, cfqq, 0);
2318 * Drain our current requests. Used for barriers and when switching
2319 * io schedulers on-the-fly.
2321 static int cfq_forced_dispatch(struct cfq_data *cfqd)
2323 struct cfq_queue *cfqq;
2326 /* Expire the timeslice of the current active queue first */
2327 cfq_slice_expired(cfqd, 0);
2328 while ((cfqq = cfq_get_next_queue_forced(cfqd)) != NULL) {
2329 __cfq_set_active_queue(cfqd, cfqq);
2330 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
2333 BUG_ON(cfqd->busy_queues);
2335 cfq_log(cfqd, "forced_dispatch=%d", dispatched);
2339 static inline bool cfq_slice_used_soon(struct cfq_data *cfqd,
2340 struct cfq_queue *cfqq)
2342 /* the queue hasn't finished any request, can't estimate */
2343 if (cfq_cfqq_slice_new(cfqq))
2345 if (time_after(jiffies + cfqd->cfq_slice_idle * cfqq->dispatched,
2352 static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2354 unsigned int max_dispatch;
2357 * Drain async requests before we start sync IO
2359 if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_flight[BLK_RW_ASYNC])
2363 * If this is an async queue and we have sync IO in flight, let it wait
2365 if (cfqd->rq_in_flight[BLK_RW_SYNC] && !cfq_cfqq_sync(cfqq))
2368 max_dispatch = max_t(unsigned int, cfqd->cfq_quantum / 2, 1);
2369 if (cfq_class_idle(cfqq))
2373 * Does this cfqq already have too much IO in flight?
2375 if (cfqq->dispatched >= max_dispatch) {
2376 bool promote_sync = false;
2378 * idle queue must always only have a single IO in flight
2380 if (cfq_class_idle(cfqq))
2384 * If there is only one sync queue, and its think time is
2385 * small, we can ignore async queue here and give the sync
2386 * queue no dispatch limit. The reason is a sync queue can
2387 * preempt async queue, limiting the sync queue doesn't make
2388 * sense. This is useful for aiostress test.
2390 if (cfq_cfqq_sync(cfqq) && cfqd->busy_sync_queues == 1) {
2391 struct cfq_io_context *cic = RQ_CIC(cfqq->next_rq);
2393 if (sample_valid(cic->ttime_samples) &&
2394 cic->ttime_mean < cfqd->cfq_slice_idle)
2395 promote_sync = true;
2399 * We have other queues, don't allow more IO from this one
2401 if (cfqd->busy_queues > 1 && cfq_slice_used_soon(cfqd, cfqq) &&
2406 * Sole queue user, no limit
2408 if (cfqd->busy_queues == 1 || promote_sync)
2412 * Normally we start throttling cfqq when cfq_quantum/2
2413 * requests have been dispatched. But we can drive
2414 * deeper queue depths at the beginning of slice
2415 * subjected to upper limit of cfq_quantum.
2417 max_dispatch = cfqd->cfq_quantum;
2421 * Async queues must wait a bit before being allowed dispatch.
2422 * We also ramp up the dispatch depth gradually for async IO,
2423 * based on the last sync IO we serviced
2425 if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
2426 unsigned long last_sync = jiffies - cfqd->last_delayed_sync;
2429 depth = last_sync / cfqd->cfq_slice[1];
2430 if (!depth && !cfqq->dispatched)
2432 if (depth < max_dispatch)
2433 max_dispatch = depth;
2437 * If we're below the current max, allow a dispatch
2439 return cfqq->dispatched < max_dispatch;
2443 * Dispatch a request from cfqq, moving them to the request queue
2446 static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2450 BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
2452 if (!cfq_may_dispatch(cfqd, cfqq))
2456 * follow expired path, else get first next available
2458 rq = cfq_check_fifo(cfqq);
2463 * insert request into driver dispatch list
2465 cfq_dispatch_insert(cfqd->queue, rq);
2467 if (!cfqd->active_cic) {
2468 struct cfq_io_context *cic = RQ_CIC(rq);
2470 atomic_long_inc(&cic->ioc->refcount);
2471 cfqd->active_cic = cic;
2478 * Find the cfqq that we need to service and move a request from that to the
2481 static int cfq_dispatch_requests(struct request_queue *q, int force)
2483 struct cfq_data *cfqd = q->elevator->elevator_data;
2484 struct cfq_queue *cfqq;
2486 if (!cfqd->busy_queues)
2489 if (unlikely(force))
2490 return cfq_forced_dispatch(cfqd);
2492 cfqq = cfq_select_queue(cfqd);
2497 * Dispatch a request from this cfqq, if it is allowed
2499 if (!cfq_dispatch_request(cfqd, cfqq))
2502 cfqq->slice_dispatch++;
2503 cfq_clear_cfqq_must_dispatch(cfqq);
2506 * expire an async queue immediately if it has used up its slice. idle
2507 * queue always expire after 1 dispatch round.
2509 if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
2510 cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
2511 cfq_class_idle(cfqq))) {
2512 cfqq->slice_end = jiffies + 1;
2513 cfq_slice_expired(cfqd, 0);
2516 cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
2521 * task holds one reference to the queue, dropped when task exits. each rq
2522 * in-flight on this queue also holds a reference, dropped when rq is freed.
2524 * Each cfq queue took a reference on the parent group. Drop it now.
2525 * queue lock must be held here.
2527 static void cfq_put_queue(struct cfq_queue *cfqq)
2529 struct cfq_data *cfqd = cfqq->cfqd;
2530 struct cfq_group *cfqg;
2532 BUG_ON(cfqq->ref <= 0);
2538 cfq_log_cfqq(cfqd, cfqq, "put_queue");
2539 BUG_ON(rb_first(&cfqq->sort_list));
2540 BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
2543 if (unlikely(cfqd->active_queue == cfqq)) {
2544 __cfq_slice_expired(cfqd, cfqq, 0);
2545 cfq_schedule_dispatch(cfqd);
2548 BUG_ON(cfq_cfqq_on_rr(cfqq));
2549 kmem_cache_free(cfq_pool, cfqq);
2554 * Must always be called with the rcu_read_lock() held
2557 __call_for_each_cic(struct io_context *ioc,
2558 void (*func)(struct io_context *, struct cfq_io_context *))
2560 struct cfq_io_context *cic;
2561 struct hlist_node *n;
2563 hlist_for_each_entry_rcu(cic, n, &ioc->cic_list, cic_list)
2568 * Call func for each cic attached to this ioc.
2571 call_for_each_cic(struct io_context *ioc,
2572 void (*func)(struct io_context *, struct cfq_io_context *))
2575 __call_for_each_cic(ioc, func);
2579 static void cfq_cic_free_rcu(struct rcu_head *head)
2581 struct cfq_io_context *cic;
2583 cic = container_of(head, struct cfq_io_context, rcu_head);
2585 kmem_cache_free(cfq_ioc_pool, cic);
2586 elv_ioc_count_dec(cfq_ioc_count);
2590 * CFQ scheduler is exiting, grab exit lock and check
2591 * the pending io context count. If it hits zero,
2592 * complete ioc_gone and set it back to NULL
2594 spin_lock(&ioc_gone_lock);
2595 if (ioc_gone && !elv_ioc_count_read(cfq_ioc_count)) {
2599 spin_unlock(&ioc_gone_lock);
2603 static void cfq_cic_free(struct cfq_io_context *cic)
2605 call_rcu(&cic->rcu_head, cfq_cic_free_rcu);
2608 static void cic_free_func(struct io_context *ioc, struct cfq_io_context *cic)
2610 unsigned long flags;
2611 unsigned long dead_key = (unsigned long) cic->key;
2613 BUG_ON(!(dead_key & CIC_DEAD_KEY));
2615 spin_lock_irqsave(&ioc->lock, flags);
2616 radix_tree_delete(&ioc->radix_root, dead_key >> CIC_DEAD_INDEX_SHIFT);
2617 hlist_del_rcu(&cic->cic_list);
2618 spin_unlock_irqrestore(&ioc->lock, flags);
2624 * Must be called with rcu_read_lock() held or preemption otherwise disabled.
2625 * Only two callers of this - ->dtor() which is called with the rcu_read_lock(),
2626 * and ->trim() which is called with the task lock held
2628 static void cfq_free_io_context(struct io_context *ioc)
2631 * ioc->refcount is zero here, or we are called from elv_unregister(),
2632 * so no more cic's are allowed to be linked into this ioc. So it
2633 * should be ok to iterate over the known list, we will see all cic's
2634 * since no new ones are added.
2636 __call_for_each_cic(ioc, cic_free_func);
2639 static void cfq_put_cooperator(struct cfq_queue *cfqq)
2641 struct cfq_queue *__cfqq, *next;
2644 * If this queue was scheduled to merge with another queue, be
2645 * sure to drop the reference taken on that queue (and others in
2646 * the merge chain). See cfq_setup_merge and cfq_merge_cfqqs.
2648 __cfqq = cfqq->new_cfqq;
2650 if (__cfqq == cfqq) {
2651 WARN(1, "cfqq->new_cfqq loop detected\n");
2654 next = __cfqq->new_cfqq;
2655 cfq_put_queue(__cfqq);
2660 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2662 if (unlikely(cfqq == cfqd->active_queue)) {
2663 __cfq_slice_expired(cfqd, cfqq, 0);
2664 cfq_schedule_dispatch(cfqd);
2667 cfq_put_cooperator(cfqq);
2669 cfq_put_queue(cfqq);
2672 static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
2673 struct cfq_io_context *cic)
2675 struct io_context *ioc = cic->ioc;
2677 list_del_init(&cic->queue_list);
2680 * Make sure dead mark is seen for dead queues
2683 cic->key = cfqd_dead_key(cfqd);
2685 if (ioc->ioc_data == cic)
2686 rcu_assign_pointer(ioc->ioc_data, NULL);
2688 if (cic->cfqq[BLK_RW_ASYNC]) {
2689 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]);
2690 cic->cfqq[BLK_RW_ASYNC] = NULL;
2693 if (cic->cfqq[BLK_RW_SYNC]) {
2694 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]);
2695 cic->cfqq[BLK_RW_SYNC] = NULL;
2699 static void cfq_exit_single_io_context(struct io_context *ioc,
2700 struct cfq_io_context *cic)
2702 struct cfq_data *cfqd = cic_to_cfqd(cic);
2705 struct request_queue *q = cfqd->queue;
2706 unsigned long flags;
2708 spin_lock_irqsave(q->queue_lock, flags);
2711 * Ensure we get a fresh copy of the ->key to prevent
2712 * race between exiting task and queue
2714 smp_read_barrier_depends();
2715 if (cic->key == cfqd)
2716 __cfq_exit_single_io_context(cfqd, cic);
2718 spin_unlock_irqrestore(q->queue_lock, flags);
2723 * The process that ioc belongs to has exited, we need to clean up
2724 * and put the internal structures we have that belongs to that process.
2726 static void cfq_exit_io_context(struct io_context *ioc)
2728 call_for_each_cic(ioc, cfq_exit_single_io_context);
2731 static struct cfq_io_context *
2732 cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
2734 struct cfq_io_context *cic;
2736 cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO,
2739 cic->last_end_request = jiffies;
2740 INIT_LIST_HEAD(&cic->queue_list);
2741 INIT_HLIST_NODE(&cic->cic_list);
2742 cic->dtor = cfq_free_io_context;
2743 cic->exit = cfq_exit_io_context;
2744 elv_ioc_count_inc(cfq_ioc_count);
2750 static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc)
2752 struct task_struct *tsk = current;
2755 if (!cfq_cfqq_prio_changed(cfqq))
2758 ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio);
2759 switch (ioprio_class) {
2761 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
2762 case IOPRIO_CLASS_NONE:
2764 * no prio set, inherit CPU scheduling settings
2766 cfqq->ioprio = task_nice_ioprio(tsk);
2767 cfqq->ioprio_class = task_nice_ioclass(tsk);
2769 case IOPRIO_CLASS_RT:
2770 cfqq->ioprio = task_ioprio(ioc);
2771 cfqq->ioprio_class = IOPRIO_CLASS_RT;
2773 case IOPRIO_CLASS_BE:
2774 cfqq->ioprio = task_ioprio(ioc);
2775 cfqq->ioprio_class = IOPRIO_CLASS_BE;
2777 case IOPRIO_CLASS_IDLE:
2778 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
2780 cfq_clear_cfqq_idle_window(cfqq);
2785 * keep track of original prio settings in case we have to temporarily
2786 * elevate the priority of this queue
2788 cfqq->org_ioprio = cfqq->ioprio;
2789 cfqq->org_ioprio_class = cfqq->ioprio_class;
2790 cfq_clear_cfqq_prio_changed(cfqq);
2793 static void changed_ioprio(struct io_context *ioc, struct cfq_io_context *cic)
2795 struct cfq_data *cfqd = cic_to_cfqd(cic);
2796 struct cfq_queue *cfqq;
2797 unsigned long flags;
2799 if (unlikely(!cfqd))
2802 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2804 cfqq = cic->cfqq[BLK_RW_ASYNC];
2806 struct cfq_queue *new_cfqq;
2807 new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic->ioc,
2810 cic->cfqq[BLK_RW_ASYNC] = new_cfqq;
2811 cfq_put_queue(cfqq);
2815 cfqq = cic->cfqq[BLK_RW_SYNC];
2817 cfq_mark_cfqq_prio_changed(cfqq);
2819 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2822 static void cfq_ioc_set_ioprio(struct io_context *ioc)
2824 call_for_each_cic(ioc, changed_ioprio);
2825 ioc->ioprio_changed = 0;
2828 static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2829 pid_t pid, bool is_sync)
2831 RB_CLEAR_NODE(&cfqq->rb_node);
2832 RB_CLEAR_NODE(&cfqq->p_node);
2833 INIT_LIST_HEAD(&cfqq->fifo);
2838 cfq_mark_cfqq_prio_changed(cfqq);
2841 if (!cfq_class_idle(cfqq))
2842 cfq_mark_cfqq_idle_window(cfqq);
2843 cfq_mark_cfqq_sync(cfqq);
2848 #ifdef CONFIG_CFQ_GROUP_IOSCHED
2849 static void changed_cgroup(struct io_context *ioc, struct cfq_io_context *cic)
2851 struct cfq_queue *sync_cfqq = cic_to_cfqq(cic, 1);
2852 struct cfq_data *cfqd = cic_to_cfqd(cic);
2853 unsigned long flags;
2854 struct request_queue *q;
2856 if (unlikely(!cfqd))
2861 spin_lock_irqsave(q->queue_lock, flags);
2865 * Drop reference to sync queue. A new sync queue will be
2866 * assigned in new group upon arrival of a fresh request.
2868 cfq_log_cfqq(cfqd, sync_cfqq, "changed cgroup");
2869 cic_set_cfqq(cic, NULL, 1);
2870 cfq_put_queue(sync_cfqq);
2873 spin_unlock_irqrestore(q->queue_lock, flags);
2876 static void cfq_ioc_set_cgroup(struct io_context *ioc)
2878 call_for_each_cic(ioc, changed_cgroup);
2879 ioc->cgroup_changed = 0;
2881 #endif /* CONFIG_CFQ_GROUP_IOSCHED */
2883 static struct cfq_queue *
2884 cfq_find_alloc_queue(struct cfq_data *cfqd, bool is_sync,
2885 struct io_context *ioc, gfp_t gfp_mask)
2887 struct cfq_queue *cfqq, *new_cfqq = NULL;
2888 struct cfq_io_context *cic;
2889 struct cfq_group *cfqg;
2892 cfqg = cfq_get_cfqg(cfqd, 1);
2893 cic = cfq_cic_lookup(cfqd, ioc);
2894 /* cic always exists here */
2895 cfqq = cic_to_cfqq(cic, is_sync);
2898 * Always try a new alloc if we fell back to the OOM cfqq
2899 * originally, since it should just be a temporary situation.
2901 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
2906 } else if (gfp_mask & __GFP_WAIT) {
2907 spin_unlock_irq(cfqd->queue->queue_lock);
2908 new_cfqq = kmem_cache_alloc_node(cfq_pool,
2909 gfp_mask | __GFP_ZERO,
2911 spin_lock_irq(cfqd->queue->queue_lock);
2915 cfqq = kmem_cache_alloc_node(cfq_pool,
2916 gfp_mask | __GFP_ZERO,
2921 cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
2922 cfq_init_prio_data(cfqq, ioc);
2923 cfq_link_cfqq_cfqg(cfqq, cfqg);
2924 cfq_log_cfqq(cfqd, cfqq, "alloced");
2926 cfqq = &cfqd->oom_cfqq;
2930 kmem_cache_free(cfq_pool, new_cfqq);
2935 static struct cfq_queue **
2936 cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
2938 switch (ioprio_class) {
2939 case IOPRIO_CLASS_RT:
2940 return &cfqd->async_cfqq[0][ioprio];
2941 case IOPRIO_CLASS_BE:
2942 return &cfqd->async_cfqq[1][ioprio];
2943 case IOPRIO_CLASS_IDLE:
2944 return &cfqd->async_idle_cfqq;
2950 static struct cfq_queue *
2951 cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct io_context *ioc,
2954 const int ioprio = task_ioprio(ioc);
2955 const int ioprio_class = task_ioprio_class(ioc);
2956 struct cfq_queue **async_cfqq = NULL;
2957 struct cfq_queue *cfqq = NULL;
2960 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
2965 cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask);
2968 * pin the queue now that it's allocated, scheduler exit will prune it
2970 if (!is_sync && !(*async_cfqq)) {
2980 * We drop cfq io contexts lazily, so we may find a dead one.
2983 cfq_drop_dead_cic(struct cfq_data *cfqd, struct io_context *ioc,
2984 struct cfq_io_context *cic)
2986 unsigned long flags;
2988 WARN_ON(!list_empty(&cic->queue_list));
2989 BUG_ON(cic->key != cfqd_dead_key(cfqd));
2991 spin_lock_irqsave(&ioc->lock, flags);
2993 BUG_ON(ioc->ioc_data == cic);
2995 radix_tree_delete(&ioc->radix_root, cfqd->cic_index);
2996 hlist_del_rcu(&cic->cic_list);
2997 spin_unlock_irqrestore(&ioc->lock, flags);
3002 static struct cfq_io_context *
3003 cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc)
3005 struct cfq_io_context *cic;
3006 unsigned long flags;
3014 * we maintain a last-hit cache, to avoid browsing over the tree
3016 cic = rcu_dereference(ioc->ioc_data);
3017 if (cic && cic->key == cfqd) {
3023 cic = radix_tree_lookup(&ioc->radix_root, cfqd->cic_index);
3027 if (unlikely(cic->key != cfqd)) {
3028 cfq_drop_dead_cic(cfqd, ioc, cic);
3033 spin_lock_irqsave(&ioc->lock, flags);
3034 rcu_assign_pointer(ioc->ioc_data, cic);
3035 spin_unlock_irqrestore(&ioc->lock, flags);
3043 * Add cic into ioc, using cfqd as the search key. This enables us to lookup
3044 * the process specific cfq io context when entered from the block layer.
3045 * Also adds the cic to a per-cfqd list, used when this queue is removed.
3047 static int cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
3048 struct cfq_io_context *cic, gfp_t gfp_mask)
3050 unsigned long flags;
3053 ret = radix_tree_preload(gfp_mask);
3058 spin_lock_irqsave(&ioc->lock, flags);
3059 ret = radix_tree_insert(&ioc->radix_root,
3060 cfqd->cic_index, cic);
3062 hlist_add_head_rcu(&cic->cic_list, &ioc->cic_list);
3063 spin_unlock_irqrestore(&ioc->lock, flags);
3065 radix_tree_preload_end();
3068 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
3069 list_add(&cic->queue_list, &cfqd->cic_list);
3070 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
3075 printk(KERN_ERR "cfq: cic link failed!\n");
3081 * Setup general io context and cfq io context. There can be several cfq
3082 * io contexts per general io context, if this process is doing io to more
3083 * than one device managed by cfq.
3085 static struct cfq_io_context *
3086 cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
3088 struct io_context *ioc = NULL;
3089 struct cfq_io_context *cic;
3091 might_sleep_if(gfp_mask & __GFP_WAIT);
3093 ioc = get_io_context(gfp_mask, cfqd->queue->node);
3097 cic = cfq_cic_lookup(cfqd, ioc);
3101 cic = cfq_alloc_io_context(cfqd, gfp_mask);
3105 if (cfq_cic_link(cfqd, ioc, cic, gfp_mask))
3109 smp_read_barrier_depends();
3110 if (unlikely(ioc->ioprio_changed))
3111 cfq_ioc_set_ioprio(ioc);
3113 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3114 if (unlikely(ioc->cgroup_changed))
3115 cfq_ioc_set_cgroup(ioc);
3121 put_io_context(ioc);
3126 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
3128 unsigned long elapsed = jiffies - cic->last_end_request;
3129 unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
3131 cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
3132 cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
3133 cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
3137 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3141 sector_t n_sec = blk_rq_sectors(rq);
3142 if (cfqq->last_request_pos) {
3143 if (cfqq->last_request_pos < blk_rq_pos(rq))
3144 sdist = blk_rq_pos(rq) - cfqq->last_request_pos;
3146 sdist = cfqq->last_request_pos - blk_rq_pos(rq);
3149 cfqq->seek_history <<= 1;
3150 if (blk_queue_nonrot(cfqd->queue))
3151 cfqq->seek_history |= (n_sec < CFQQ_SECT_THR_NONROT);
3153 cfqq->seek_history |= (sdist > CFQQ_SEEK_THR);
3157 * Disable idle window if the process thinks too long or seeks so much that
3161 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3162 struct cfq_io_context *cic)
3164 int old_idle, enable_idle;
3167 * Don't idle for async or idle io prio class
3169 if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
3172 enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
3174 if (cfqq->queued[0] + cfqq->queued[1] >= 4)
3175 cfq_mark_cfqq_deep(cfqq);
3177 if (cfqq->next_rq && (cfqq->next_rq->cmd_flags & REQ_NOIDLE))
3179 else if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle ||
3180 (!cfq_cfqq_deep(cfqq) && CFQQ_SEEKY(cfqq)))
3182 else if (sample_valid(cic->ttime_samples)) {
3183 if (cic->ttime_mean > cfqd->cfq_slice_idle)
3189 if (old_idle != enable_idle) {
3190 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
3192 cfq_mark_cfqq_idle_window(cfqq);
3194 cfq_clear_cfqq_idle_window(cfqq);
3199 * Check if new_cfqq should preempt the currently active queue. Return 0 for
3200 * no or if we aren't sure, a 1 will cause a preempt.
3203 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
3206 struct cfq_queue *cfqq;
3208 cfqq = cfqd->active_queue;
3212 if (cfq_class_idle(new_cfqq))
3215 if (cfq_class_idle(cfqq))
3219 * Don't allow a non-RT request to preempt an ongoing RT cfqq timeslice.
3221 if (cfq_class_rt(cfqq) && !cfq_class_rt(new_cfqq))
3225 * if the new request is sync, but the currently running queue is
3226 * not, let the sync request have priority.
3228 if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
3231 if (new_cfqq->cfqg != cfqq->cfqg)
3234 if (cfq_slice_used(cfqq))
3237 /* Allow preemption only if we are idling on sync-noidle tree */
3238 if (cfqd->serving_type == SYNC_NOIDLE_WORKLOAD &&
3239 cfqq_type(new_cfqq) == SYNC_NOIDLE_WORKLOAD &&
3240 new_cfqq->service_tree->count == 2 &&
3241 RB_EMPTY_ROOT(&cfqq->sort_list))
3245 * So both queues are sync. Let the new request get disk time if
3246 * it's a metadata request and the current queue is doing regular IO.
3248 if ((rq->cmd_flags & REQ_META) && !cfqq->meta_pending)
3252 * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
3254 if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
3257 /* An idle queue should not be idle now for some reason */
3258 if (RB_EMPTY_ROOT(&cfqq->sort_list) && !cfq_should_idle(cfqd, cfqq))
3261 if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
3265 * if this request is as-good as one we would expect from the
3266 * current cfqq, let it preempt
3268 if (cfq_rq_close(cfqd, cfqq, rq))
3275 * cfqq preempts the active queue. if we allowed preempt with no slice left,
3276 * let it have half of its nominal slice.
3278 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3280 struct cfq_queue *old_cfqq = cfqd->active_queue;
3282 cfq_log_cfqq(cfqd, cfqq, "preempt");
3283 cfq_slice_expired(cfqd, 1);
3286 * workload type is changed, don't save slice, otherwise preempt
3289 if (cfqq_type(old_cfqq) != cfqq_type(cfqq))
3290 cfqq->cfqg->saved_workload_slice = 0;
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, false);
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, false);
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 the queue already has requests, don't wait */
3418 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3421 /* If there are other queues in the group, don't wait */
3422 if (cfqq->cfqg->nr_cfqq > 1)
3425 if (cfq_slice_used(cfqq))
3428 /* if slice left is less than think time, wait busy */
3429 if (cic && sample_valid(cic->ttime_samples)
3430 && (cfqq->slice_end - jiffies < cic->ttime_mean))
3434 * If think times is less than a jiffy than ttime_mean=0 and above
3435 * will not be true. It might happen that slice has not expired yet
3436 * but will expire soon (4-5 ns) during select_queue(). To cover the
3437 * case where think time is less than a jiffy, mark the queue wait
3438 * busy if only 1 jiffy is left in the slice.
3440 if (cfqq->slice_end - jiffies == 1)
3446 static void cfq_completed_request(struct request_queue *q, struct request *rq)
3448 struct cfq_queue *cfqq = RQ_CFQQ(rq);
3449 struct cfq_data *cfqd = cfqq->cfqd;
3450 const int sync = rq_is_sync(rq);
3454 cfq_log_cfqq(cfqd, cfqq, "complete rqnoidle %d",
3455 !!(rq->cmd_flags & REQ_NOIDLE));
3457 cfq_update_hw_tag(cfqd);
3459 WARN_ON(!cfqd->rq_in_driver);
3460 WARN_ON(!cfqq->dispatched);
3461 cfqd->rq_in_driver--;
3463 (RQ_CFQG(rq))->dispatched--;
3464 cfq_blkiocg_update_completion_stats(&cfqq->cfqg->blkg,
3465 rq_start_time_ns(rq), rq_io_start_time_ns(rq),
3466 rq_data_dir(rq), rq_is_sync(rq));
3468 cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]--;
3471 RQ_CIC(rq)->last_end_request = now;
3472 if (!time_after(rq->start_time + cfqd->cfq_fifo_expire[1], now))
3473 cfqd->last_delayed_sync = now;
3477 * If this is the active queue, check if it needs to be expired,
3478 * or if we want to idle in case it has no pending requests.
3480 if (cfqd->active_queue == cfqq) {
3481 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
3483 if (cfq_cfqq_slice_new(cfqq)) {
3484 cfq_set_prio_slice(cfqd, cfqq);
3485 cfq_clear_cfqq_slice_new(cfqq);
3489 * Should we wait for next request to come in before we expire
3492 if (cfq_should_wait_busy(cfqd, cfqq)) {
3493 unsigned long extend_sl = cfqd->cfq_slice_idle;
3494 if (!cfqd->cfq_slice_idle)
3495 extend_sl = cfqd->cfq_group_idle;
3496 cfqq->slice_end = jiffies + extend_sl;
3497 cfq_mark_cfqq_wait_busy(cfqq);
3498 cfq_log_cfqq(cfqd, cfqq, "will busy wait");
3502 * Idling is not enabled on:
3504 * - idle-priority queues
3506 * - queues with still some requests queued
3507 * - when there is a close cooperator
3509 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
3510 cfq_slice_expired(cfqd, 1);
3511 else if (sync && cfqq_empty &&
3512 !cfq_close_cooperator(cfqd, cfqq)) {
3513 cfq_arm_slice_timer(cfqd);
3517 if (!cfqd->rq_in_driver)
3518 cfq_schedule_dispatch(cfqd);
3522 * we temporarily boost lower priority queues if they are holding fs exclusive
3523 * resources. they are boosted to normal prio (CLASS_BE/4)
3525 static void cfq_prio_boost(struct cfq_queue *cfqq)
3527 if (has_fs_excl()) {
3529 * boost idle prio on transactions that would lock out other
3530 * users of the filesystem
3532 if (cfq_class_idle(cfqq))
3533 cfqq->ioprio_class = IOPRIO_CLASS_BE;
3534 if (cfqq->ioprio > IOPRIO_NORM)
3535 cfqq->ioprio = IOPRIO_NORM;
3538 * unboost the queue (if needed)
3540 cfqq->ioprio_class = cfqq->org_ioprio_class;
3541 cfqq->ioprio = cfqq->org_ioprio;
3545 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
3547 if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
3548 cfq_mark_cfqq_must_alloc_slice(cfqq);
3549 return ELV_MQUEUE_MUST;
3552 return ELV_MQUEUE_MAY;
3555 static int cfq_may_queue(struct request_queue *q, int rw)
3557 struct cfq_data *cfqd = q->elevator->elevator_data;
3558 struct task_struct *tsk = current;
3559 struct cfq_io_context *cic;
3560 struct cfq_queue *cfqq;
3563 * don't force setup of a queue from here, as a call to may_queue
3564 * does not necessarily imply that a request actually will be queued.
3565 * so just lookup a possibly existing queue, or return 'may queue'
3568 cic = cfq_cic_lookup(cfqd, tsk->io_context);
3570 return ELV_MQUEUE_MAY;
3572 cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
3574 cfq_init_prio_data(cfqq, cic->ioc);
3575 cfq_prio_boost(cfqq);
3577 return __cfq_may_queue(cfqq);
3580 return ELV_MQUEUE_MAY;
3584 * queue lock held here
3586 static void cfq_put_request(struct request *rq)
3588 struct cfq_queue *cfqq = RQ_CFQQ(rq);
3591 const int rw = rq_data_dir(rq);
3593 BUG_ON(!cfqq->allocated[rw]);
3594 cfqq->allocated[rw]--;
3596 put_io_context(RQ_CIC(rq)->ioc);
3598 rq->elevator_private[0] = NULL;
3599 rq->elevator_private[1] = NULL;
3601 /* Put down rq reference on cfqg */
3602 cfq_put_cfqg(RQ_CFQG(rq));
3603 rq->elevator_private[2] = NULL;
3605 cfq_put_queue(cfqq);
3609 static struct cfq_queue *
3610 cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_context *cic,
3611 struct cfq_queue *cfqq)
3613 cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq);
3614 cic_set_cfqq(cic, cfqq->new_cfqq, 1);
3615 cfq_mark_cfqq_coop(cfqq->new_cfqq);
3616 cfq_put_queue(cfqq);
3617 return cic_to_cfqq(cic, 1);
3621 * Returns NULL if a new cfqq should be allocated, or the old cfqq if this
3622 * was the last process referring to said cfqq.
3624 static struct cfq_queue *
3625 split_cfqq(struct cfq_io_context *cic, struct cfq_queue *cfqq)
3627 if (cfqq_process_refs(cfqq) == 1) {
3628 cfqq->pid = current->pid;
3629 cfq_clear_cfqq_coop(cfqq);
3630 cfq_clear_cfqq_split_coop(cfqq);
3634 cic_set_cfqq(cic, NULL, 1);
3636 cfq_put_cooperator(cfqq);
3638 cfq_put_queue(cfqq);
3642 * Allocate cfq data structures associated with this request.
3645 cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
3647 struct cfq_data *cfqd = q->elevator->elevator_data;
3648 struct cfq_io_context *cic;
3649 const int rw = rq_data_dir(rq);
3650 const bool is_sync = rq_is_sync(rq);
3651 struct cfq_queue *cfqq;
3652 unsigned long flags;
3654 might_sleep_if(gfp_mask & __GFP_WAIT);
3656 cic = cfq_get_io_context(cfqd, gfp_mask);
3658 spin_lock_irqsave(q->queue_lock, flags);
3664 cfqq = cic_to_cfqq(cic, is_sync);
3665 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
3666 cfqq = cfq_get_queue(cfqd, is_sync, cic->ioc, gfp_mask);
3667 cic_set_cfqq(cic, cfqq, is_sync);
3670 * If the queue was seeky for too long, break it apart.
3672 if (cfq_cfqq_coop(cfqq) && cfq_cfqq_split_coop(cfqq)) {
3673 cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq");
3674 cfqq = split_cfqq(cic, cfqq);
3680 * Check to see if this queue is scheduled to merge with
3681 * another, closely cooperating queue. The merging of
3682 * queues happens here as it must be done in process context.
3683 * The reference on new_cfqq was taken in merge_cfqqs.
3686 cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq);
3689 cfqq->allocated[rw]++;
3692 rq->elevator_private[0] = cic;
3693 rq->elevator_private[1] = cfqq;
3694 rq->elevator_private[2] = cfq_ref_get_cfqg(cfqq->cfqg);
3695 spin_unlock_irqrestore(q->queue_lock, flags);
3700 put_io_context(cic->ioc);
3702 cfq_schedule_dispatch(cfqd);
3703 spin_unlock_irqrestore(q->queue_lock, flags);
3704 cfq_log(cfqd, "set_request fail");
3708 static void cfq_kick_queue(struct work_struct *work)
3710 struct cfq_data *cfqd =
3711 container_of(work, struct cfq_data, unplug_work);
3712 struct request_queue *q = cfqd->queue;
3714 spin_lock_irq(q->queue_lock);
3715 __blk_run_queue(cfqd->queue, false);
3716 spin_unlock_irq(q->queue_lock);
3720 * Timer running if the active_queue is currently idling inside its time slice
3722 static void cfq_idle_slice_timer(unsigned long data)
3724 struct cfq_data *cfqd = (struct cfq_data *) data;
3725 struct cfq_queue *cfqq;
3726 unsigned long flags;
3729 cfq_log(cfqd, "idle timer fired");
3731 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
3733 cfqq = cfqd->active_queue;
3738 * We saw a request before the queue expired, let it through
3740 if (cfq_cfqq_must_dispatch(cfqq))
3746 if (cfq_slice_used(cfqq))
3750 * only expire and reinvoke request handler, if there are
3751 * other queues with pending requests
3753 if (!cfqd->busy_queues)
3757 * not expired and it has a request pending, let it dispatch
3759 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3763 * Queue depth flag is reset only when the idle didn't succeed
3765 cfq_clear_cfqq_deep(cfqq);
3768 cfq_slice_expired(cfqd, timed_out);
3770 cfq_schedule_dispatch(cfqd);
3772 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
3775 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
3777 del_timer_sync(&cfqd->idle_slice_timer);
3778 cancel_work_sync(&cfqd->unplug_work);
3781 static void cfq_put_async_queues(struct cfq_data *cfqd)
3785 for (i = 0; i < IOPRIO_BE_NR; i++) {
3786 if (cfqd->async_cfqq[0][i])
3787 cfq_put_queue(cfqd->async_cfqq[0][i]);
3788 if (cfqd->async_cfqq[1][i])
3789 cfq_put_queue(cfqd->async_cfqq[1][i]);
3792 if (cfqd->async_idle_cfqq)
3793 cfq_put_queue(cfqd->async_idle_cfqq);
3796 static void cfq_cfqd_free(struct rcu_head *head)
3798 kfree(container_of(head, struct cfq_data, rcu));
3801 static void cfq_exit_queue(struct elevator_queue *e)
3803 struct cfq_data *cfqd = e->elevator_data;
3804 struct request_queue *q = cfqd->queue;
3806 cfq_shutdown_timer_wq(cfqd);
3808 spin_lock_irq(q->queue_lock);
3810 if (cfqd->active_queue)
3811 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
3813 while (!list_empty(&cfqd->cic_list)) {
3814 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
3815 struct cfq_io_context,
3818 __cfq_exit_single_io_context(cfqd, cic);
3821 cfq_put_async_queues(cfqd);
3822 cfq_release_cfq_groups(cfqd);
3823 cfq_blkiocg_del_blkio_group(&cfqd->root_group.blkg);
3825 spin_unlock_irq(q->queue_lock);
3827 cfq_shutdown_timer_wq(cfqd);
3829 spin_lock(&cic_index_lock);
3830 ida_remove(&cic_index_ida, cfqd->cic_index);
3831 spin_unlock(&cic_index_lock);
3833 /* Wait for cfqg->blkg->key accessors to exit their grace periods. */
3834 call_rcu(&cfqd->rcu, cfq_cfqd_free);
3837 static int cfq_alloc_cic_index(void)
3842 if (!ida_pre_get(&cic_index_ida, GFP_KERNEL))
3845 spin_lock(&cic_index_lock);
3846 error = ida_get_new(&cic_index_ida, &index);
3847 spin_unlock(&cic_index_lock);
3848 if (error && error != -EAGAIN)
3855 static void *cfq_init_queue(struct request_queue *q)
3857 struct cfq_data *cfqd;
3859 struct cfq_group *cfqg;
3860 struct cfq_rb_root *st;
3862 i = cfq_alloc_cic_index();
3866 cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
3871 * Don't need take queue_lock in the routine, since we are
3872 * initializing the ioscheduler, and nobody is using cfqd
3874 cfqd->cic_index = i;
3876 /* Init root service tree */
3877 cfqd->grp_service_tree = CFQ_RB_ROOT;
3879 /* Init root group */
3880 cfqg = &cfqd->root_group;
3881 for_each_cfqg_st(cfqg, i, j, st)
3883 RB_CLEAR_NODE(&cfqg->rb_node);
3885 /* Give preference to root group over other groups */
3886 cfqg->weight = 2*BLKIO_WEIGHT_DEFAULT;
3888 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3890 * Take a reference to root group which we never drop. This is just
3891 * to make sure that cfq_put_cfqg() does not try to kfree root group
3895 cfq_blkiocg_add_blkio_group(&blkio_root_cgroup, &cfqg->blkg,
3900 * Not strictly needed (since RB_ROOT just clears the node and we
3901 * zeroed cfqd on alloc), but better be safe in case someone decides
3902 * to add magic to the rb code
3904 for (i = 0; i < CFQ_PRIO_LISTS; i++)
3905 cfqd->prio_trees[i] = RB_ROOT;
3908 * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
3909 * Grab a permanent reference to it, so that the normal code flow
3910 * will not attempt to free it.
3912 cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
3913 cfqd->oom_cfqq.ref++;
3914 cfq_link_cfqq_cfqg(&cfqd->oom_cfqq, &cfqd->root_group);
3916 INIT_LIST_HEAD(&cfqd->cic_list);
3920 init_timer(&cfqd->idle_slice_timer);
3921 cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
3922 cfqd->idle_slice_timer.data = (unsigned long) cfqd;
3924 INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
3926 cfqd->cfq_quantum = cfq_quantum;
3927 cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
3928 cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
3929 cfqd->cfq_back_max = cfq_back_max;
3930 cfqd->cfq_back_penalty = cfq_back_penalty;
3931 cfqd->cfq_slice[0] = cfq_slice_async;
3932 cfqd->cfq_slice[1] = cfq_slice_sync;
3933 cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
3934 cfqd->cfq_slice_idle = cfq_slice_idle;
3935 cfqd->cfq_group_idle = cfq_group_idle;
3936 cfqd->cfq_latency = 1;
3939 * we optimistically start assuming sync ops weren't delayed in last
3940 * second, in order to have larger depth for async operations.
3942 cfqd->last_delayed_sync = jiffies - HZ;
3946 static void cfq_slab_kill(void)
3949 * Caller already ensured that pending RCU callbacks are completed,
3950 * so we should have no busy allocations at this point.
3953 kmem_cache_destroy(cfq_pool);
3955 kmem_cache_destroy(cfq_ioc_pool);
3958 static int __init cfq_slab_setup(void)
3960 cfq_pool = KMEM_CACHE(cfq_queue, 0);
3964 cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0);
3975 * sysfs parts below -->
3978 cfq_var_show(unsigned int var, char *page)
3980 return sprintf(page, "%d\n", var);
3984 cfq_var_store(unsigned int *var, const char *page, size_t count)
3986 char *p = (char *) page;
3988 *var = simple_strtoul(p, &p, 10);
3992 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
3993 static ssize_t __FUNC(struct elevator_queue *e, char *page) \
3995 struct cfq_data *cfqd = e->elevator_data; \
3996 unsigned int __data = __VAR; \
3998 __data = jiffies_to_msecs(__data); \
3999 return cfq_var_show(__data, (page)); \
4001 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
4002 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
4003 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
4004 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
4005 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
4006 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
4007 SHOW_FUNCTION(cfq_group_idle_show, cfqd->cfq_group_idle, 1);
4008 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
4009 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
4010 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
4011 SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
4012 #undef SHOW_FUNCTION
4014 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
4015 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
4017 struct cfq_data *cfqd = e->elevator_data; \
4018 unsigned int __data; \
4019 int ret = cfq_var_store(&__data, (page), count); \
4020 if (__data < (MIN)) \
4022 else if (__data > (MAX)) \
4025 *(__PTR) = msecs_to_jiffies(__data); \
4027 *(__PTR) = __data; \
4030 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
4031 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
4033 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
4035 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
4036 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
4038 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
4039 STORE_FUNCTION(cfq_group_idle_store, &cfqd->cfq_group_idle, 0, UINT_MAX, 1);
4040 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
4041 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
4042 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
4044 STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
4045 #undef STORE_FUNCTION
4047 #define CFQ_ATTR(name) \
4048 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
4050 static struct elv_fs_entry cfq_attrs[] = {
4052 CFQ_ATTR(fifo_expire_sync),
4053 CFQ_ATTR(fifo_expire_async),
4054 CFQ_ATTR(back_seek_max),
4055 CFQ_ATTR(back_seek_penalty),
4056 CFQ_ATTR(slice_sync),
4057 CFQ_ATTR(slice_async),
4058 CFQ_ATTR(slice_async_rq),
4059 CFQ_ATTR(slice_idle),
4060 CFQ_ATTR(group_idle),
4061 CFQ_ATTR(low_latency),
4065 static struct elevator_type iosched_cfq = {
4067 .elevator_merge_fn = cfq_merge,
4068 .elevator_merged_fn = cfq_merged_request,
4069 .elevator_merge_req_fn = cfq_merged_requests,
4070 .elevator_allow_merge_fn = cfq_allow_merge,
4071 .elevator_bio_merged_fn = cfq_bio_merged,
4072 .elevator_dispatch_fn = cfq_dispatch_requests,
4073 .elevator_add_req_fn = cfq_insert_request,
4074 .elevator_activate_req_fn = cfq_activate_request,
4075 .elevator_deactivate_req_fn = cfq_deactivate_request,
4076 .elevator_completed_req_fn = cfq_completed_request,
4077 .elevator_former_req_fn = elv_rb_former_request,
4078 .elevator_latter_req_fn = elv_rb_latter_request,
4079 .elevator_set_req_fn = cfq_set_request,
4080 .elevator_put_req_fn = cfq_put_request,
4081 .elevator_may_queue_fn = cfq_may_queue,
4082 .elevator_init_fn = cfq_init_queue,
4083 .elevator_exit_fn = cfq_exit_queue,
4084 .trim = cfq_free_io_context,
4086 .elevator_attrs = cfq_attrs,
4087 .elevator_name = "cfq",
4088 .elevator_owner = THIS_MODULE,
4091 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4092 static struct blkio_policy_type blkio_policy_cfq = {
4094 .blkio_unlink_group_fn = cfq_unlink_blkio_group,
4095 .blkio_update_group_weight_fn = cfq_update_blkio_group_weight,
4097 .plid = BLKIO_POLICY_PROP,
4100 static struct blkio_policy_type blkio_policy_cfq;
4103 static int __init cfq_init(void)
4106 * could be 0 on HZ < 1000 setups
4108 if (!cfq_slice_async)
4109 cfq_slice_async = 1;
4110 if (!cfq_slice_idle)
4113 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4114 if (!cfq_group_idle)
4119 if (cfq_slab_setup())
4122 elv_register(&iosched_cfq);
4123 blkio_policy_register(&blkio_policy_cfq);
4128 static void __exit cfq_exit(void)
4130 DECLARE_COMPLETION_ONSTACK(all_gone);
4131 blkio_policy_unregister(&blkio_policy_cfq);
4132 elv_unregister(&iosched_cfq);
4133 ioc_gone = &all_gone;
4134 /* ioc_gone's update must be visible before reading ioc_count */
4138 * this also protects us from entering cfq_slab_kill() with
4139 * pending RCU callbacks
4141 if (elv_ioc_count_read(cfq_ioc_count))
4142 wait_for_completion(&all_gone);
4143 ida_destroy(&cic_index_ida);
4147 module_init(cfq_init);
4148 module_exit(cfq_exit);
4150 MODULE_AUTHOR("Jens Axboe");
4151 MODULE_LICENSE("GPL");
4152 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");