2 * CFQ, or complete fairness queueing, disk scheduler.
4 * Based on ideas from a previously unfinished io
5 * scheduler (round robin per-process disk scheduling) and Andrea Arcangeli.
7 * Copyright (C) 2003 Jens Axboe <axboe@kernel.dk>
9 #include <linux/module.h>
10 #include <linux/slab.h>
11 #include <linux/blkdev.h>
12 #include <linux/elevator.h>
13 #include <linux/jiffies.h>
14 #include <linux/rbtree.h>
15 #include <linux/ioprio.h>
16 #include <linux/blktrace_api.h>
22 /* max queue in one round of service */
23 static const int cfq_quantum = 8;
24 static const int cfq_fifo_expire[2] = { HZ / 4, HZ / 8 };
25 /* maximum backwards seek, in KiB */
26 static const int cfq_back_max = 16 * 1024;
27 /* penalty of a backwards seek */
28 static const int cfq_back_penalty = 2;
29 static const int cfq_slice_sync = HZ / 10;
30 static int cfq_slice_async = HZ / 25;
31 static const int cfq_slice_async_rq = 2;
32 static int cfq_slice_idle = HZ / 125;
33 static int cfq_group_idle = HZ / 125;
34 static const int cfq_target_latency = HZ * 3/10; /* 300 ms */
35 static const int cfq_hist_divisor = 4;
38 * offset from end of service tree
40 #define CFQ_IDLE_DELAY (HZ / 5)
43 * below this threshold, we consider thinktime immediate
45 #define CFQ_MIN_TT (2)
47 #define CFQ_SLICE_SCALE (5)
48 #define CFQ_HW_QUEUE_MIN (5)
49 #define CFQ_SERVICE_SHIFT 12
51 #define CFQQ_SEEK_THR (sector_t)(8 * 100)
52 #define CFQQ_CLOSE_THR (sector_t)(8 * 1024)
53 #define CFQQ_SECT_THR_NONROT (sector_t)(2 * 32)
54 #define CFQQ_SEEKY(cfqq) (hweight32(cfqq->seek_history) > 32/8)
57 ((struct cfq_io_context *) (rq)->elevator_private)
58 #define RQ_CFQQ(rq) (struct cfq_queue *) ((rq)->elevator_private2)
59 #define RQ_CFQG(rq) (struct cfq_group *) ((rq)->elevator_private3)
61 static struct kmem_cache *cfq_pool;
62 static struct kmem_cache *cfq_ioc_pool;
64 static DEFINE_PER_CPU(unsigned long, cfq_ioc_count);
65 static struct completion *ioc_gone;
66 static DEFINE_SPINLOCK(ioc_gone_lock);
68 static DEFINE_SPINLOCK(cic_index_lock);
69 static DEFINE_IDA(cic_index_ida);
71 #define CFQ_PRIO_LISTS IOPRIO_BE_NR
72 #define cfq_class_idle(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
73 #define cfq_class_rt(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
75 #define sample_valid(samples) ((samples) > 80)
76 #define rb_entry_cfqg(node) rb_entry((node), struct cfq_group, rb_node)
79 * Most of our rbtree usage is for sorting with min extraction, so
80 * if we cache the leftmost node we don't have to walk down the tree
81 * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should
82 * move this into the elevator for the rq sorting as well.
88 unsigned total_weight;
91 #define CFQ_RB_ROOT (struct cfq_rb_root) { .rb = RB_ROOT, .left = NULL, \
92 .count = 0, .min_vdisktime = 0, }
95 * Per process-grouping structure
100 /* various state flags, see below */
102 /* parent cfq_data */
103 struct cfq_data *cfqd;
104 /* service_tree member */
105 struct rb_node rb_node;
106 /* service_tree key */
107 unsigned long rb_key;
108 /* prio tree member */
109 struct rb_node p_node;
110 /* prio tree root we belong to, if any */
111 struct rb_root *p_root;
112 /* sorted list of pending requests */
113 struct rb_root sort_list;
114 /* if fifo isn't expired, next request to serve */
115 struct request *next_rq;
116 /* requests queued in sort_list */
118 /* currently allocated requests */
120 /* fifo list of requests in sort_list */
121 struct list_head fifo;
123 /* time when queue got scheduled in to dispatch first request. */
124 unsigned long dispatch_start;
125 unsigned int allocated_slice;
126 unsigned int slice_dispatch;
127 /* time when first request from queue completed and slice started. */
128 unsigned long slice_start;
129 unsigned long slice_end;
132 /* pending metadata requests */
134 /* number of requests that are on the dispatch list or inside driver */
137 /* io prio of this group */
138 unsigned short ioprio, org_ioprio;
139 unsigned short ioprio_class, org_ioprio_class;
144 sector_t last_request_pos;
146 struct cfq_rb_root *service_tree;
147 struct cfq_queue *new_cfqq;
148 struct cfq_group *cfqg;
149 struct cfq_group *orig_cfqg;
150 /* Number of sectors dispatched from queue in single dispatch round */
151 unsigned long nr_sectors;
155 * First index in the service_trees.
156 * IDLE is handled separately, so it has negative index
166 * Second index in the service_trees.
170 SYNC_NOIDLE_WORKLOAD = 1,
174 /* This is per cgroup per device grouping structure */
176 /* group service_tree member */
177 struct rb_node rb_node;
179 /* group service_tree key */
183 /* number of cfqq currently on this group */
187 * Per group busy queus average. Useful for workload slice calc. We
188 * create the array for each prio class but at run time it is used
189 * only for RT and BE class and slot for IDLE class remains unused.
190 * This is primarily done to avoid confusion and a gcc warning.
192 unsigned int busy_queues_avg[CFQ_PRIO_NR];
194 * rr lists of queues with requests. We maintain service trees for
195 * RT and BE classes. These trees are subdivided in subclasses
196 * of SYNC, SYNC_NOIDLE and ASYNC based on workload type. For IDLE
197 * class there is no subclassification and all the cfq queues go on
198 * a single tree service_tree_idle.
199 * Counts are embedded in the cfq_rb_root
201 struct cfq_rb_root service_trees[2][3];
202 struct cfq_rb_root service_tree_idle;
204 unsigned long saved_workload_slice;
205 enum wl_type_t saved_workload;
206 enum wl_prio_t saved_serving_prio;
207 struct blkio_group blkg;
208 #ifdef CONFIG_CFQ_GROUP_IOSCHED
209 struct hlist_node cfqd_node;
212 /* number of requests that are on the dispatch list or inside driver */
217 * Per block device queue structure
220 struct request_queue *queue;
221 /* Root service tree for cfq_groups */
222 struct cfq_rb_root grp_service_tree;
223 struct cfq_group root_group;
226 * The priority currently being served
228 enum wl_prio_t serving_prio;
229 enum wl_type_t serving_type;
230 unsigned long workload_expires;
231 struct cfq_group *serving_group;
234 * Each priority tree is sorted by next_request position. These
235 * trees are used when determining if two or more queues are
236 * interleaving requests (see cfq_close_cooperator).
238 struct rb_root prio_trees[CFQ_PRIO_LISTS];
240 unsigned int busy_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;
288 unsigned int cfq_group_isolation;
290 unsigned int cic_index;
291 struct list_head cic_list;
294 * Fallback dummy cfqq for extreme OOM conditions
296 struct cfq_queue oom_cfqq;
298 unsigned long last_delayed_sync;
300 /* List of cfq groups being managed on this device*/
301 struct hlist_head cfqg_list;
305 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd);
307 static struct cfq_rb_root *service_tree_for(struct cfq_group *cfqg,
314 if (prio == IDLE_WORKLOAD)
315 return &cfqg->service_tree_idle;
317 return &cfqg->service_trees[prio][type];
320 enum cfqq_state_flags {
321 CFQ_CFQQ_FLAG_on_rr = 0, /* on round-robin busy list */
322 CFQ_CFQQ_FLAG_wait_request, /* waiting for a request */
323 CFQ_CFQQ_FLAG_must_dispatch, /* must be allowed a dispatch */
324 CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
325 CFQ_CFQQ_FLAG_fifo_expire, /* FIFO checked in this slice */
326 CFQ_CFQQ_FLAG_idle_window, /* slice idling enabled */
327 CFQ_CFQQ_FLAG_prio_changed, /* task priority has changed */
328 CFQ_CFQQ_FLAG_slice_new, /* no requests dispatched in slice */
329 CFQ_CFQQ_FLAG_sync, /* synchronous queue */
330 CFQ_CFQQ_FLAG_coop, /* cfqq is shared */
331 CFQ_CFQQ_FLAG_split_coop, /* shared cfqq will be splitted */
332 CFQ_CFQQ_FLAG_deep, /* sync cfqq experienced large depth */
333 CFQ_CFQQ_FLAG_wait_busy, /* Waiting for next request */
336 #define CFQ_CFQQ_FNS(name) \
337 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \
339 (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name); \
341 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \
343 (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \
345 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \
347 return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \
351 CFQ_CFQQ_FNS(wait_request);
352 CFQ_CFQQ_FNS(must_dispatch);
353 CFQ_CFQQ_FNS(must_alloc_slice);
354 CFQ_CFQQ_FNS(fifo_expire);
355 CFQ_CFQQ_FNS(idle_window);
356 CFQ_CFQQ_FNS(prio_changed);
357 CFQ_CFQQ_FNS(slice_new);
360 CFQ_CFQQ_FNS(split_coop);
362 CFQ_CFQQ_FNS(wait_busy);
365 #ifdef CONFIG_CFQ_GROUP_IOSCHED
366 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
367 blk_add_trace_msg((cfqd)->queue, "cfq%d%c %s " fmt, (cfqq)->pid, \
368 cfq_cfqq_sync((cfqq)) ? 'S' : 'A', \
369 blkg_path(&(cfqq)->cfqg->blkg), ##args);
371 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...) \
372 blk_add_trace_msg((cfqd)->queue, "%s " fmt, \
373 blkg_path(&(cfqg)->blkg), ##args); \
376 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
377 blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args)
378 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...) do {} while (0);
380 #define cfq_log(cfqd, fmt, args...) \
381 blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
383 /* Traverses through cfq group service trees */
384 #define for_each_cfqg_st(cfqg, i, j, st) \
385 for (i = 0; i <= IDLE_WORKLOAD; i++) \
386 for (j = 0, st = i < IDLE_WORKLOAD ? &cfqg->service_trees[i][j]\
387 : &cfqg->service_tree_idle; \
388 (i < IDLE_WORKLOAD && j <= SYNC_WORKLOAD) || \
389 (i == IDLE_WORKLOAD && j == 0); \
390 j++, st = i < IDLE_WORKLOAD ? \
391 &cfqg->service_trees[i][j]: NULL) \
394 static inline bool iops_mode(struct cfq_data *cfqd)
397 * If we are not idling on queues and it is a NCQ drive, parallel
398 * execution of requests is on and measuring time is not possible
399 * in most of the cases until and unless we drive shallower queue
400 * depths and that becomes a performance bottleneck. In such cases
401 * switch to start providing fairness in terms of number of IOs.
403 if (!cfqd->cfq_slice_idle && cfqd->hw_tag)
409 static inline enum wl_prio_t cfqq_prio(struct cfq_queue *cfqq)
411 if (cfq_class_idle(cfqq))
412 return IDLE_WORKLOAD;
413 if (cfq_class_rt(cfqq))
419 static enum wl_type_t cfqq_type(struct cfq_queue *cfqq)
421 if (!cfq_cfqq_sync(cfqq))
422 return ASYNC_WORKLOAD;
423 if (!cfq_cfqq_idle_window(cfqq))
424 return SYNC_NOIDLE_WORKLOAD;
425 return SYNC_WORKLOAD;
428 static inline int cfq_group_busy_queues_wl(enum wl_prio_t wl,
429 struct cfq_data *cfqd,
430 struct cfq_group *cfqg)
432 if (wl == IDLE_WORKLOAD)
433 return cfqg->service_tree_idle.count;
435 return cfqg->service_trees[wl][ASYNC_WORKLOAD].count
436 + cfqg->service_trees[wl][SYNC_NOIDLE_WORKLOAD].count
437 + cfqg->service_trees[wl][SYNC_WORKLOAD].count;
440 static inline int cfqg_busy_async_queues(struct cfq_data *cfqd,
441 struct cfq_group *cfqg)
443 return cfqg->service_trees[RT_WORKLOAD][ASYNC_WORKLOAD].count
444 + cfqg->service_trees[BE_WORKLOAD][ASYNC_WORKLOAD].count;
447 static void cfq_dispatch_insert(struct request_queue *, struct request *);
448 static struct cfq_queue *cfq_get_queue(struct cfq_data *, bool,
449 struct io_context *, gfp_t);
450 static struct cfq_io_context *cfq_cic_lookup(struct cfq_data *,
451 struct io_context *);
453 static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_context *cic,
456 return cic->cfqq[is_sync];
459 static inline void cic_set_cfqq(struct cfq_io_context *cic,
460 struct cfq_queue *cfqq, bool is_sync)
462 cic->cfqq[is_sync] = cfqq;
465 #define CIC_DEAD_KEY 1ul
466 #define CIC_DEAD_INDEX_SHIFT 1
468 static inline void *cfqd_dead_key(struct cfq_data *cfqd)
470 return (void *)(cfqd->cic_index << CIC_DEAD_INDEX_SHIFT | CIC_DEAD_KEY);
473 static inline struct cfq_data *cic_to_cfqd(struct cfq_io_context *cic)
475 struct cfq_data *cfqd = cic->key;
477 if (unlikely((unsigned long) cfqd & CIC_DEAD_KEY))
484 * We regard a request as SYNC, if it's either a read or has the SYNC bit
485 * set (in which case it could also be direct WRITE).
487 static inline bool cfq_bio_sync(struct bio *bio)
489 return bio_data_dir(bio) == READ || (bio->bi_rw & REQ_SYNC);
493 * scheduler run of queue, if there are requests pending and no one in the
494 * driver that will restart queueing
496 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
498 if (cfqd->busy_queues) {
499 cfq_log(cfqd, "schedule dispatch");
500 kblockd_schedule_work(cfqd->queue, &cfqd->unplug_work);
504 static int cfq_queue_empty(struct request_queue *q)
506 struct cfq_data *cfqd = q->elevator->elevator_data;
508 return !cfqd->rq_queued;
512 * Scale schedule slice based on io priority. Use the sync time slice only
513 * if a queue is marked sync and has sync io queued. A sync queue with async
514 * io only, should not get full sync slice length.
516 static inline int cfq_prio_slice(struct cfq_data *cfqd, bool sync,
519 const int base_slice = cfqd->cfq_slice[sync];
521 WARN_ON(prio >= IOPRIO_BE_NR);
523 return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
527 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
529 return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
532 static inline u64 cfq_scale_slice(unsigned long delta, struct cfq_group *cfqg)
534 u64 d = delta << CFQ_SERVICE_SHIFT;
536 d = d * BLKIO_WEIGHT_DEFAULT;
537 do_div(d, cfqg->weight);
541 static inline u64 max_vdisktime(u64 min_vdisktime, u64 vdisktime)
543 s64 delta = (s64)(vdisktime - min_vdisktime);
545 min_vdisktime = vdisktime;
547 return min_vdisktime;
550 static inline u64 min_vdisktime(u64 min_vdisktime, u64 vdisktime)
552 s64 delta = (s64)(vdisktime - min_vdisktime);
554 min_vdisktime = vdisktime;
556 return min_vdisktime;
559 static void update_min_vdisktime(struct cfq_rb_root *st)
561 u64 vdisktime = st->min_vdisktime;
562 struct cfq_group *cfqg;
565 cfqg = rb_entry_cfqg(st->left);
566 vdisktime = min_vdisktime(vdisktime, cfqg->vdisktime);
569 st->min_vdisktime = max_vdisktime(st->min_vdisktime, vdisktime);
573 * get averaged number of queues of RT/BE priority.
574 * average is updated, with a formula that gives more weight to higher numbers,
575 * to quickly follows sudden increases and decrease slowly
578 static inline unsigned cfq_group_get_avg_queues(struct cfq_data *cfqd,
579 struct cfq_group *cfqg, bool rt)
581 unsigned min_q, max_q;
582 unsigned mult = cfq_hist_divisor - 1;
583 unsigned round = cfq_hist_divisor / 2;
584 unsigned busy = cfq_group_busy_queues_wl(rt, cfqd, cfqg);
586 min_q = min(cfqg->busy_queues_avg[rt], busy);
587 max_q = max(cfqg->busy_queues_avg[rt], busy);
588 cfqg->busy_queues_avg[rt] = (mult * max_q + min_q + round) /
590 return cfqg->busy_queues_avg[rt];
593 static inline unsigned
594 cfq_group_slice(struct cfq_data *cfqd, struct cfq_group *cfqg)
596 struct cfq_rb_root *st = &cfqd->grp_service_tree;
598 return cfq_target_latency * cfqg->weight / st->total_weight;
602 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
604 unsigned slice = cfq_prio_to_slice(cfqd, cfqq);
605 if (cfqd->cfq_latency) {
607 * interested queues (we consider only the ones with the same
608 * priority class in the cfq group)
610 unsigned iq = cfq_group_get_avg_queues(cfqd, cfqq->cfqg,
612 unsigned sync_slice = cfqd->cfq_slice[1];
613 unsigned expect_latency = sync_slice * iq;
614 unsigned group_slice = cfq_group_slice(cfqd, cfqq->cfqg);
616 if (expect_latency > group_slice) {
617 unsigned base_low_slice = 2 * cfqd->cfq_slice_idle;
618 /* scale low_slice according to IO priority
619 * and sync vs async */
621 min(slice, base_low_slice * slice / sync_slice);
622 /* the adapted slice value is scaled to fit all iqs
623 * into the target latency */
624 slice = max(slice * group_slice / expect_latency,
628 cfqq->slice_start = jiffies;
629 cfqq->slice_end = jiffies + slice;
630 cfqq->allocated_slice = slice;
631 cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
635 * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
636 * isn't valid until the first request from the dispatch is activated
637 * and the slice time set.
639 static inline bool cfq_slice_used(struct cfq_queue *cfqq)
641 if (cfq_cfqq_slice_new(cfqq))
643 if (time_before(jiffies, cfqq->slice_end))
650 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
651 * We choose the request that is closest to the head right now. Distance
652 * behind the head is penalized and only allowed to a certain extent.
654 static struct request *
655 cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2, sector_t last)
657 sector_t s1, s2, d1 = 0, d2 = 0;
658 unsigned long back_max;
659 #define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */
660 #define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */
661 unsigned wrap = 0; /* bit mask: requests behind the disk head? */
663 if (rq1 == NULL || rq1 == rq2)
668 if (rq_is_sync(rq1) && !rq_is_sync(rq2))
670 else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
672 if ((rq1->cmd_flags & REQ_META) && !(rq2->cmd_flags & REQ_META))
674 else if ((rq2->cmd_flags & REQ_META) &&
675 !(rq1->cmd_flags & REQ_META))
678 s1 = blk_rq_pos(rq1);
679 s2 = blk_rq_pos(rq2);
682 * by definition, 1KiB is 2 sectors
684 back_max = cfqd->cfq_back_max * 2;
687 * Strict one way elevator _except_ in the case where we allow
688 * short backward seeks which are biased as twice the cost of a
689 * similar forward seek.
693 else if (s1 + back_max >= last)
694 d1 = (last - s1) * cfqd->cfq_back_penalty;
696 wrap |= CFQ_RQ1_WRAP;
700 else if (s2 + back_max >= last)
701 d2 = (last - s2) * cfqd->cfq_back_penalty;
703 wrap |= CFQ_RQ2_WRAP;
705 /* Found required data */
708 * By doing switch() on the bit mask "wrap" we avoid having to
709 * check two variables for all permutations: --> faster!
712 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
728 case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
731 * Since both rqs are wrapped,
732 * start with the one that's further behind head
733 * (--> only *one* back seek required),
734 * since back seek takes more time than forward.
744 * The below is leftmost cache rbtree addon
746 static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
748 /* Service tree is empty */
753 root->left = rb_first(&root->rb);
756 return rb_entry(root->left, struct cfq_queue, rb_node);
761 static struct cfq_group *cfq_rb_first_group(struct cfq_rb_root *root)
764 root->left = rb_first(&root->rb);
767 return rb_entry_cfqg(root->left);
772 static void rb_erase_init(struct rb_node *n, struct rb_root *root)
778 static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
782 rb_erase_init(n, &root->rb);
787 * would be nice to take fifo expire time into account as well
789 static struct request *
790 cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
791 struct request *last)
793 struct rb_node *rbnext = rb_next(&last->rb_node);
794 struct rb_node *rbprev = rb_prev(&last->rb_node);
795 struct request *next = NULL, *prev = NULL;
797 BUG_ON(RB_EMPTY_NODE(&last->rb_node));
800 prev = rb_entry_rq(rbprev);
803 next = rb_entry_rq(rbnext);
805 rbnext = rb_first(&cfqq->sort_list);
806 if (rbnext && rbnext != &last->rb_node)
807 next = rb_entry_rq(rbnext);
810 return cfq_choose_req(cfqd, next, prev, blk_rq_pos(last));
813 static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
814 struct cfq_queue *cfqq)
817 * just an approximation, should be ok.
819 return (cfqq->cfqg->nr_cfqq - 1) * (cfq_prio_slice(cfqd, 1, 0) -
820 cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
824 cfqg_key(struct cfq_rb_root *st, struct cfq_group *cfqg)
826 return cfqg->vdisktime - st->min_vdisktime;
830 __cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
832 struct rb_node **node = &st->rb.rb_node;
833 struct rb_node *parent = NULL;
834 struct cfq_group *__cfqg;
835 s64 key = cfqg_key(st, cfqg);
838 while (*node != NULL) {
840 __cfqg = rb_entry_cfqg(parent);
842 if (key < cfqg_key(st, __cfqg))
843 node = &parent->rb_left;
845 node = &parent->rb_right;
851 st->left = &cfqg->rb_node;
853 rb_link_node(&cfqg->rb_node, parent, node);
854 rb_insert_color(&cfqg->rb_node, &st->rb);
858 cfq_group_service_tree_add(struct cfq_data *cfqd, struct cfq_group *cfqg)
860 struct cfq_rb_root *st = &cfqd->grp_service_tree;
861 struct cfq_group *__cfqg;
865 if (!RB_EMPTY_NODE(&cfqg->rb_node))
869 * Currently put the group at the end. Later implement something
870 * so that groups get lesser vtime based on their weights, so that
871 * if group does not loose all if it was not continously backlogged.
873 n = rb_last(&st->rb);
875 __cfqg = rb_entry_cfqg(n);
876 cfqg->vdisktime = __cfqg->vdisktime + CFQ_IDLE_DELAY;
878 cfqg->vdisktime = st->min_vdisktime;
880 __cfq_group_service_tree_add(st, cfqg);
881 st->total_weight += cfqg->weight;
885 cfq_group_service_tree_del(struct cfq_data *cfqd, struct cfq_group *cfqg)
887 struct cfq_rb_root *st = &cfqd->grp_service_tree;
889 BUG_ON(cfqg->nr_cfqq < 1);
892 /* If there are other cfq queues under this group, don't delete it */
896 cfq_log_cfqg(cfqd, cfqg, "del_from_rr group");
897 st->total_weight -= cfqg->weight;
898 if (!RB_EMPTY_NODE(&cfqg->rb_node))
899 cfq_rb_erase(&cfqg->rb_node, st);
900 cfqg->saved_workload_slice = 0;
901 cfq_blkiocg_update_dequeue_stats(&cfqg->blkg, 1);
904 static inline unsigned int cfq_cfqq_slice_usage(struct cfq_queue *cfqq)
906 unsigned int slice_used;
909 * Queue got expired before even a single request completed or
910 * got expired immediately after first request completion.
912 if (!cfqq->slice_start || cfqq->slice_start == jiffies) {
914 * Also charge the seek time incurred to the group, otherwise
915 * if there are mutiple queues in the group, each can dispatch
916 * a single request on seeky media and cause lots of seek time
917 * and group will never know it.
919 slice_used = max_t(unsigned, (jiffies - cfqq->dispatch_start),
922 slice_used = jiffies - cfqq->slice_start;
923 if (slice_used > cfqq->allocated_slice)
924 slice_used = cfqq->allocated_slice;
930 static void cfq_group_served(struct cfq_data *cfqd, struct cfq_group *cfqg,
931 struct cfq_queue *cfqq)
933 struct cfq_rb_root *st = &cfqd->grp_service_tree;
934 unsigned int used_sl, charge;
935 int nr_sync = cfqg->nr_cfqq - cfqg_busy_async_queues(cfqd, cfqg)
936 - cfqg->service_tree_idle.count;
939 used_sl = charge = cfq_cfqq_slice_usage(cfqq);
942 charge = cfqq->slice_dispatch;
943 else if (!cfq_cfqq_sync(cfqq) && !nr_sync)
944 charge = cfqq->allocated_slice;
946 /* Can't update vdisktime while group is on service tree */
947 cfq_rb_erase(&cfqg->rb_node, st);
948 cfqg->vdisktime += cfq_scale_slice(charge, cfqg);
949 __cfq_group_service_tree_add(st, cfqg);
951 /* This group is being expired. Save the context */
952 if (time_after(cfqd->workload_expires, jiffies)) {
953 cfqg->saved_workload_slice = cfqd->workload_expires
955 cfqg->saved_workload = cfqd->serving_type;
956 cfqg->saved_serving_prio = cfqd->serving_prio;
958 cfqg->saved_workload_slice = 0;
960 cfq_log_cfqg(cfqd, cfqg, "served: vt=%llu min_vt=%llu", cfqg->vdisktime,
962 cfq_log_cfqq(cfqq->cfqd, cfqq, "sl_used=%u disp=%u charge=%u iops=%u"
963 " sect=%u", used_sl, cfqq->slice_dispatch, charge,
964 iops_mode(cfqd), cfqq->nr_sectors);
965 cfq_blkiocg_update_timeslice_used(&cfqg->blkg, used_sl);
966 cfq_blkiocg_set_start_empty_time(&cfqg->blkg);
969 #ifdef CONFIG_CFQ_GROUP_IOSCHED
970 static inline struct cfq_group *cfqg_of_blkg(struct blkio_group *blkg)
973 return container_of(blkg, struct cfq_group, blkg);
977 void cfq_update_blkio_group_weight(void *key, struct blkio_group *blkg,
980 cfqg_of_blkg(blkg)->weight = weight;
983 static struct cfq_group *
984 cfq_find_alloc_cfqg(struct cfq_data *cfqd, struct cgroup *cgroup, int create)
986 struct blkio_cgroup *blkcg = cgroup_to_blkio_cgroup(cgroup);
987 struct cfq_group *cfqg = NULL;
990 struct cfq_rb_root *st;
991 struct backing_dev_info *bdi = &cfqd->queue->backing_dev_info;
992 unsigned int major, minor;
994 cfqg = cfqg_of_blkg(blkiocg_lookup_group(blkcg, key));
995 if (cfqg && !cfqg->blkg.dev && bdi->dev && dev_name(bdi->dev)) {
996 sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
997 cfqg->blkg.dev = MKDEV(major, minor);
1000 if (cfqg || !create)
1003 cfqg = kzalloc_node(sizeof(*cfqg), GFP_ATOMIC, cfqd->queue->node);
1007 for_each_cfqg_st(cfqg, i, j, st)
1009 RB_CLEAR_NODE(&cfqg->rb_node);
1012 * Take the initial reference that will be released on destroy
1013 * This can be thought of a joint reference by cgroup and
1014 * elevator which will be dropped by either elevator exit
1015 * or cgroup deletion path depending on who is exiting first.
1017 atomic_set(&cfqg->ref, 1);
1020 * Add group onto cgroup list. It might happen that bdi->dev is
1021 * not initiliazed yet. Initialize this new group without major
1022 * and minor info and this info will be filled in once a new thread
1023 * comes for IO. See code above.
1026 sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
1027 cfq_blkiocg_add_blkio_group(blkcg, &cfqg->blkg, (void *)cfqd,
1028 MKDEV(major, minor));
1030 cfq_blkiocg_add_blkio_group(blkcg, &cfqg->blkg, (void *)cfqd,
1033 cfqg->weight = blkcg_get_weight(blkcg, cfqg->blkg.dev);
1035 /* Add group on cfqd list */
1036 hlist_add_head(&cfqg->cfqd_node, &cfqd->cfqg_list);
1043 * Search for the cfq group current task belongs to. If create = 1, then also
1044 * create the cfq group if it does not exist. request_queue lock must be held.
1046 static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd, int create)
1048 struct cgroup *cgroup;
1049 struct cfq_group *cfqg = NULL;
1052 cgroup = task_cgroup(current, blkio_subsys_id);
1053 cfqg = cfq_find_alloc_cfqg(cfqd, cgroup, create);
1054 if (!cfqg && create)
1055 cfqg = &cfqd->root_group;
1060 static inline struct cfq_group *cfq_ref_get_cfqg(struct cfq_group *cfqg)
1062 atomic_inc(&cfqg->ref);
1066 static void cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg)
1068 /* Currently, all async queues are mapped to root group */
1069 if (!cfq_cfqq_sync(cfqq))
1070 cfqg = &cfqq->cfqd->root_group;
1073 /* cfqq reference on cfqg */
1074 atomic_inc(&cfqq->cfqg->ref);
1077 static void cfq_put_cfqg(struct cfq_group *cfqg)
1079 struct cfq_rb_root *st;
1082 BUG_ON(atomic_read(&cfqg->ref) <= 0);
1083 if (!atomic_dec_and_test(&cfqg->ref))
1085 for_each_cfqg_st(cfqg, i, j, st)
1086 BUG_ON(!RB_EMPTY_ROOT(&st->rb));
1090 static void cfq_destroy_cfqg(struct cfq_data *cfqd, struct cfq_group *cfqg)
1092 /* Something wrong if we are trying to remove same group twice */
1093 BUG_ON(hlist_unhashed(&cfqg->cfqd_node));
1095 hlist_del_init(&cfqg->cfqd_node);
1098 * Put the reference taken at the time of creation so that when all
1099 * queues are gone, group can be destroyed.
1104 static void cfq_release_cfq_groups(struct cfq_data *cfqd)
1106 struct hlist_node *pos, *n;
1107 struct cfq_group *cfqg;
1109 hlist_for_each_entry_safe(cfqg, pos, n, &cfqd->cfqg_list, cfqd_node) {
1111 * If cgroup removal path got to blk_group first and removed
1112 * it from cgroup list, then it will take care of destroying
1115 if (!cfq_blkiocg_del_blkio_group(&cfqg->blkg))
1116 cfq_destroy_cfqg(cfqd, cfqg);
1121 * Blk cgroup controller notification saying that blkio_group object is being
1122 * delinked as associated cgroup object is going away. That also means that
1123 * no new IO will come in this group. So get rid of this group as soon as
1124 * any pending IO in the group is finished.
1126 * This function is called under rcu_read_lock(). key is the rcu protected
1127 * pointer. That means "key" is a valid cfq_data pointer as long as we are rcu
1130 * "key" was fetched from blkio_group under blkio_cgroup->lock. That means
1131 * it should not be NULL as even if elevator was exiting, cgroup deltion
1132 * path got to it first.
1134 void cfq_unlink_blkio_group(void *key, struct blkio_group *blkg)
1136 unsigned long flags;
1137 struct cfq_data *cfqd = key;
1139 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1140 cfq_destroy_cfqg(cfqd, cfqg_of_blkg(blkg));
1141 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1144 #else /* GROUP_IOSCHED */
1145 static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd, int create)
1147 return &cfqd->root_group;
1150 static inline struct cfq_group *cfq_ref_get_cfqg(struct cfq_group *cfqg)
1156 cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) {
1160 static void cfq_release_cfq_groups(struct cfq_data *cfqd) {}
1161 static inline void cfq_put_cfqg(struct cfq_group *cfqg) {}
1163 #endif /* GROUP_IOSCHED */
1166 * The cfqd->service_trees holds all pending cfq_queue's that have
1167 * requests waiting to be processed. It is sorted in the order that
1168 * we will service the queues.
1170 static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1173 struct rb_node **p, *parent;
1174 struct cfq_queue *__cfqq;
1175 unsigned long rb_key;
1176 struct cfq_rb_root *service_tree;
1179 int group_changed = 0;
1181 #ifdef CONFIG_CFQ_GROUP_IOSCHED
1182 if (!cfqd->cfq_group_isolation
1183 && cfqq_type(cfqq) == SYNC_NOIDLE_WORKLOAD
1184 && cfqq->cfqg && cfqq->cfqg != &cfqd->root_group) {
1185 /* Move this cfq to root group */
1186 cfq_log_cfqq(cfqd, cfqq, "moving to root group");
1187 if (!RB_EMPTY_NODE(&cfqq->rb_node))
1188 cfq_group_service_tree_del(cfqd, cfqq->cfqg);
1189 cfqq->orig_cfqg = cfqq->cfqg;
1190 cfqq->cfqg = &cfqd->root_group;
1191 atomic_inc(&cfqd->root_group.ref);
1193 } else if (!cfqd->cfq_group_isolation
1194 && cfqq_type(cfqq) == SYNC_WORKLOAD && cfqq->orig_cfqg) {
1195 /* cfqq is sequential now needs to go to its original group */
1196 BUG_ON(cfqq->cfqg != &cfqd->root_group);
1197 if (!RB_EMPTY_NODE(&cfqq->rb_node))
1198 cfq_group_service_tree_del(cfqd, cfqq->cfqg);
1199 cfq_put_cfqg(cfqq->cfqg);
1200 cfqq->cfqg = cfqq->orig_cfqg;
1201 cfqq->orig_cfqg = NULL;
1203 cfq_log_cfqq(cfqd, cfqq, "moved to origin group");
1207 service_tree = service_tree_for(cfqq->cfqg, cfqq_prio(cfqq),
1209 if (cfq_class_idle(cfqq)) {
1210 rb_key = CFQ_IDLE_DELAY;
1211 parent = rb_last(&service_tree->rb);
1212 if (parent && parent != &cfqq->rb_node) {
1213 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1214 rb_key += __cfqq->rb_key;
1217 } else if (!add_front) {
1219 * Get our rb key offset. Subtract any residual slice
1220 * value carried from last service. A negative resid
1221 * count indicates slice overrun, and this should position
1222 * the next service time further away in the tree.
1224 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
1225 rb_key -= cfqq->slice_resid;
1226 cfqq->slice_resid = 0;
1229 __cfqq = cfq_rb_first(service_tree);
1230 rb_key += __cfqq ? __cfqq->rb_key : jiffies;
1233 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1236 * same position, nothing more to do
1238 if (rb_key == cfqq->rb_key &&
1239 cfqq->service_tree == service_tree)
1242 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1243 cfqq->service_tree = NULL;
1248 cfqq->service_tree = service_tree;
1249 p = &service_tree->rb.rb_node;
1254 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1257 * sort by key, that represents service time.
1259 if (time_before(rb_key, __cfqq->rb_key))
1262 n = &(*p)->rb_right;
1270 service_tree->left = &cfqq->rb_node;
1272 cfqq->rb_key = rb_key;
1273 rb_link_node(&cfqq->rb_node, parent, p);
1274 rb_insert_color(&cfqq->rb_node, &service_tree->rb);
1275 service_tree->count++;
1276 if ((add_front || !new_cfqq) && !group_changed)
1278 cfq_group_service_tree_add(cfqd, cfqq->cfqg);
1281 static struct cfq_queue *
1282 cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
1283 sector_t sector, struct rb_node **ret_parent,
1284 struct rb_node ***rb_link)
1286 struct rb_node **p, *parent;
1287 struct cfq_queue *cfqq = NULL;
1295 cfqq = rb_entry(parent, struct cfq_queue, p_node);
1298 * Sort strictly based on sector. Smallest to the left,
1299 * largest to the right.
1301 if (sector > blk_rq_pos(cfqq->next_rq))
1302 n = &(*p)->rb_right;
1303 else if (sector < blk_rq_pos(cfqq->next_rq))
1311 *ret_parent = parent;
1317 static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1319 struct rb_node **p, *parent;
1320 struct cfq_queue *__cfqq;
1323 rb_erase(&cfqq->p_node, cfqq->p_root);
1324 cfqq->p_root = NULL;
1327 if (cfq_class_idle(cfqq))
1332 cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
1333 __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
1334 blk_rq_pos(cfqq->next_rq), &parent, &p);
1336 rb_link_node(&cfqq->p_node, parent, p);
1337 rb_insert_color(&cfqq->p_node, cfqq->p_root);
1339 cfqq->p_root = NULL;
1343 * Update cfqq's position in the service tree.
1345 static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1348 * Resorting requires the cfqq to be on the RR list already.
1350 if (cfq_cfqq_on_rr(cfqq)) {
1351 cfq_service_tree_add(cfqd, cfqq, 0);
1352 cfq_prio_tree_add(cfqd, cfqq);
1357 * add to busy list of queues for service, trying to be fair in ordering
1358 * the pending list according to last request service
1360 static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1362 cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
1363 BUG_ON(cfq_cfqq_on_rr(cfqq));
1364 cfq_mark_cfqq_on_rr(cfqq);
1365 cfqd->busy_queues++;
1367 cfq_resort_rr_list(cfqd, cfqq);
1371 * Called when the cfqq no longer has requests pending, remove it from
1374 static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1376 cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
1377 BUG_ON(!cfq_cfqq_on_rr(cfqq));
1378 cfq_clear_cfqq_on_rr(cfqq);
1380 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1381 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1382 cfqq->service_tree = NULL;
1385 rb_erase(&cfqq->p_node, cfqq->p_root);
1386 cfqq->p_root = NULL;
1389 cfq_group_service_tree_del(cfqd, cfqq->cfqg);
1390 BUG_ON(!cfqd->busy_queues);
1391 cfqd->busy_queues--;
1395 * rb tree support functions
1397 static void cfq_del_rq_rb(struct request *rq)
1399 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1400 const int sync = rq_is_sync(rq);
1402 BUG_ON(!cfqq->queued[sync]);
1403 cfqq->queued[sync]--;
1405 elv_rb_del(&cfqq->sort_list, rq);
1407 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) {
1409 * Queue will be deleted from service tree when we actually
1410 * expire it later. Right now just remove it from prio tree
1414 rb_erase(&cfqq->p_node, cfqq->p_root);
1415 cfqq->p_root = NULL;
1420 static void cfq_add_rq_rb(struct request *rq)
1422 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1423 struct cfq_data *cfqd = cfqq->cfqd;
1424 struct request *__alias, *prev;
1426 cfqq->queued[rq_is_sync(rq)]++;
1429 * looks a little odd, but the first insert might return an alias.
1430 * if that happens, put the alias on the dispatch list
1432 while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL)
1433 cfq_dispatch_insert(cfqd->queue, __alias);
1435 if (!cfq_cfqq_on_rr(cfqq))
1436 cfq_add_cfqq_rr(cfqd, cfqq);
1439 * check if this request is a better next-serve candidate
1441 prev = cfqq->next_rq;
1442 cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq, cfqd->last_position);
1445 * adjust priority tree position, if ->next_rq changes
1447 if (prev != cfqq->next_rq)
1448 cfq_prio_tree_add(cfqd, cfqq);
1450 BUG_ON(!cfqq->next_rq);
1453 static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
1455 elv_rb_del(&cfqq->sort_list, rq);
1456 cfqq->queued[rq_is_sync(rq)]--;
1457 cfq_blkiocg_update_io_remove_stats(&(RQ_CFQG(rq))->blkg,
1458 rq_data_dir(rq), rq_is_sync(rq));
1460 cfq_blkiocg_update_io_add_stats(&(RQ_CFQG(rq))->blkg,
1461 &cfqq->cfqd->serving_group->blkg, rq_data_dir(rq),
1465 static struct request *
1466 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
1468 struct task_struct *tsk = current;
1469 struct cfq_io_context *cic;
1470 struct cfq_queue *cfqq;
1472 cic = cfq_cic_lookup(cfqd, tsk->io_context);
1476 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
1478 sector_t sector = bio->bi_sector + bio_sectors(bio);
1480 return elv_rb_find(&cfqq->sort_list, sector);
1486 static void cfq_activate_request(struct request_queue *q, struct request *rq)
1488 struct cfq_data *cfqd = q->elevator->elevator_data;
1490 cfqd->rq_in_driver++;
1491 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
1492 cfqd->rq_in_driver);
1494 cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
1497 static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
1499 struct cfq_data *cfqd = q->elevator->elevator_data;
1501 WARN_ON(!cfqd->rq_in_driver);
1502 cfqd->rq_in_driver--;
1503 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
1504 cfqd->rq_in_driver);
1507 static void cfq_remove_request(struct request *rq)
1509 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1511 if (cfqq->next_rq == rq)
1512 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
1514 list_del_init(&rq->queuelist);
1517 cfqq->cfqd->rq_queued--;
1518 cfq_blkiocg_update_io_remove_stats(&(RQ_CFQG(rq))->blkg,
1519 rq_data_dir(rq), rq_is_sync(rq));
1520 if (rq->cmd_flags & REQ_META) {
1521 WARN_ON(!cfqq->meta_pending);
1522 cfqq->meta_pending--;
1526 static int cfq_merge(struct request_queue *q, struct request **req,
1529 struct cfq_data *cfqd = q->elevator->elevator_data;
1530 struct request *__rq;
1532 __rq = cfq_find_rq_fmerge(cfqd, bio);
1533 if (__rq && elv_rq_merge_ok(__rq, bio)) {
1535 return ELEVATOR_FRONT_MERGE;
1538 return ELEVATOR_NO_MERGE;
1541 static void cfq_merged_request(struct request_queue *q, struct request *req,
1544 if (type == ELEVATOR_FRONT_MERGE) {
1545 struct cfq_queue *cfqq = RQ_CFQQ(req);
1547 cfq_reposition_rq_rb(cfqq, req);
1551 static void cfq_bio_merged(struct request_queue *q, struct request *req,
1554 cfq_blkiocg_update_io_merged_stats(&(RQ_CFQG(req))->blkg,
1555 bio_data_dir(bio), cfq_bio_sync(bio));
1559 cfq_merged_requests(struct request_queue *q, struct request *rq,
1560 struct request *next)
1562 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1564 * reposition in fifo if next is older than rq
1566 if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
1567 time_before(rq_fifo_time(next), rq_fifo_time(rq))) {
1568 list_move(&rq->queuelist, &next->queuelist);
1569 rq_set_fifo_time(rq, rq_fifo_time(next));
1572 if (cfqq->next_rq == next)
1574 cfq_remove_request(next);
1575 cfq_blkiocg_update_io_merged_stats(&(RQ_CFQG(rq))->blkg,
1576 rq_data_dir(next), rq_is_sync(next));
1579 static int cfq_allow_merge(struct request_queue *q, struct request *rq,
1582 struct cfq_data *cfqd = q->elevator->elevator_data;
1583 struct cfq_io_context *cic;
1584 struct cfq_queue *cfqq;
1587 * Disallow merge of a sync bio into an async request.
1589 if (cfq_bio_sync(bio) && !rq_is_sync(rq))
1593 * Lookup the cfqq that this bio will be queued with. Allow
1594 * merge only if rq is queued there.
1596 cic = cfq_cic_lookup(cfqd, current->io_context);
1600 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
1601 return cfqq == RQ_CFQQ(rq);
1604 static inline void cfq_del_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1606 del_timer(&cfqd->idle_slice_timer);
1607 cfq_blkiocg_update_idle_time_stats(&cfqq->cfqg->blkg);
1610 static void __cfq_set_active_queue(struct cfq_data *cfqd,
1611 struct cfq_queue *cfqq)
1614 cfq_log_cfqq(cfqd, cfqq, "set_active wl_prio:%d wl_type:%d",
1615 cfqd->serving_prio, cfqd->serving_type);
1616 cfq_blkiocg_update_avg_queue_size_stats(&cfqq->cfqg->blkg);
1617 cfqq->slice_start = 0;
1618 cfqq->dispatch_start = jiffies;
1619 cfqq->allocated_slice = 0;
1620 cfqq->slice_end = 0;
1621 cfqq->slice_dispatch = 0;
1622 cfqq->nr_sectors = 0;
1624 cfq_clear_cfqq_wait_request(cfqq);
1625 cfq_clear_cfqq_must_dispatch(cfqq);
1626 cfq_clear_cfqq_must_alloc_slice(cfqq);
1627 cfq_clear_cfqq_fifo_expire(cfqq);
1628 cfq_mark_cfqq_slice_new(cfqq);
1630 cfq_del_timer(cfqd, cfqq);
1633 cfqd->active_queue = cfqq;
1637 * current cfqq expired its slice (or was too idle), select new one
1640 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1643 cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
1645 if (cfq_cfqq_wait_request(cfqq))
1646 cfq_del_timer(cfqd, cfqq);
1648 cfq_clear_cfqq_wait_request(cfqq);
1649 cfq_clear_cfqq_wait_busy(cfqq);
1652 * If this cfqq is shared between multiple processes, check to
1653 * make sure that those processes are still issuing I/Os within
1654 * the mean seek distance. If not, it may be time to break the
1655 * queues apart again.
1657 if (cfq_cfqq_coop(cfqq) && CFQQ_SEEKY(cfqq))
1658 cfq_mark_cfqq_split_coop(cfqq);
1661 * store what was left of this slice, if the queue idled/timed out
1663 if (timed_out && !cfq_cfqq_slice_new(cfqq)) {
1664 cfqq->slice_resid = cfqq->slice_end - jiffies;
1665 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
1668 cfq_group_served(cfqd, cfqq->cfqg, cfqq);
1670 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
1671 cfq_del_cfqq_rr(cfqd, cfqq);
1673 cfq_resort_rr_list(cfqd, cfqq);
1675 if (cfqq == cfqd->active_queue)
1676 cfqd->active_queue = NULL;
1678 if (cfqd->active_cic) {
1679 put_io_context(cfqd->active_cic->ioc);
1680 cfqd->active_cic = NULL;
1684 static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out)
1686 struct cfq_queue *cfqq = cfqd->active_queue;
1689 __cfq_slice_expired(cfqd, cfqq, timed_out);
1693 * Get next queue for service. Unless we have a queue preemption,
1694 * we'll simply select the first cfqq in the service tree.
1696 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
1698 struct cfq_rb_root *service_tree =
1699 service_tree_for(cfqd->serving_group, cfqd->serving_prio,
1700 cfqd->serving_type);
1702 if (!cfqd->rq_queued)
1705 /* There is nothing to dispatch */
1708 if (RB_EMPTY_ROOT(&service_tree->rb))
1710 return cfq_rb_first(service_tree);
1713 static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd)
1715 struct cfq_group *cfqg;
1716 struct cfq_queue *cfqq;
1718 struct cfq_rb_root *st;
1720 if (!cfqd->rq_queued)
1723 cfqg = cfq_get_next_cfqg(cfqd);
1727 for_each_cfqg_st(cfqg, i, j, st)
1728 if ((cfqq = cfq_rb_first(st)) != NULL)
1734 * Get and set a new active queue for service.
1736 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
1737 struct cfq_queue *cfqq)
1740 cfqq = cfq_get_next_queue(cfqd);
1742 __cfq_set_active_queue(cfqd, cfqq);
1746 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
1749 if (blk_rq_pos(rq) >= cfqd->last_position)
1750 return blk_rq_pos(rq) - cfqd->last_position;
1752 return cfqd->last_position - blk_rq_pos(rq);
1755 static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1758 return cfq_dist_from_last(cfqd, rq) <= CFQQ_CLOSE_THR;
1761 static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
1762 struct cfq_queue *cur_cfqq)
1764 struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
1765 struct rb_node *parent, *node;
1766 struct cfq_queue *__cfqq;
1767 sector_t sector = cfqd->last_position;
1769 if (RB_EMPTY_ROOT(root))
1773 * First, if we find a request starting at the end of the last
1774 * request, choose it.
1776 __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
1781 * If the exact sector wasn't found, the parent of the NULL leaf
1782 * will contain the closest sector.
1784 __cfqq = rb_entry(parent, struct cfq_queue, p_node);
1785 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1788 if (blk_rq_pos(__cfqq->next_rq) < sector)
1789 node = rb_next(&__cfqq->p_node);
1791 node = rb_prev(&__cfqq->p_node);
1795 __cfqq = rb_entry(node, struct cfq_queue, p_node);
1796 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1804 * cur_cfqq - passed in so that we don't decide that the current queue is
1805 * closely cooperating with itself.
1807 * So, basically we're assuming that that cur_cfqq has dispatched at least
1808 * one request, and that cfqd->last_position reflects a position on the disk
1809 * associated with the I/O issued by cur_cfqq. I'm not sure this is a valid
1812 static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
1813 struct cfq_queue *cur_cfqq)
1815 struct cfq_queue *cfqq;
1817 if (cfq_class_idle(cur_cfqq))
1819 if (!cfq_cfqq_sync(cur_cfqq))
1821 if (CFQQ_SEEKY(cur_cfqq))
1825 * Don't search priority tree if it's the only queue in the group.
1827 if (cur_cfqq->cfqg->nr_cfqq == 1)
1831 * We should notice if some of the queues are cooperating, eg
1832 * working closely on the same area of the disk. In that case,
1833 * we can group them together and don't waste time idling.
1835 cfqq = cfqq_close(cfqd, cur_cfqq);
1839 /* If new queue belongs to different cfq_group, don't choose it */
1840 if (cur_cfqq->cfqg != cfqq->cfqg)
1844 * It only makes sense to merge sync queues.
1846 if (!cfq_cfqq_sync(cfqq))
1848 if (CFQQ_SEEKY(cfqq))
1852 * Do not merge queues of different priority classes
1854 if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq))
1861 * Determine whether we should enforce idle window for this queue.
1864 static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1866 enum wl_prio_t prio = cfqq_prio(cfqq);
1867 struct cfq_rb_root *service_tree = cfqq->service_tree;
1869 BUG_ON(!service_tree);
1870 BUG_ON(!service_tree->count);
1872 if (!cfqd->cfq_slice_idle)
1875 /* We never do for idle class queues. */
1876 if (prio == IDLE_WORKLOAD)
1879 /* We do for queues that were marked with idle window flag. */
1880 if (cfq_cfqq_idle_window(cfqq) &&
1881 !(blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag))
1885 * Otherwise, we do only if they are the last ones
1886 * in their service tree.
1888 if (service_tree->count == 1 && cfq_cfqq_sync(cfqq))
1890 cfq_log_cfqq(cfqd, cfqq, "Not idling. st->count:%d",
1891 service_tree->count);
1895 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
1897 struct cfq_queue *cfqq = cfqd->active_queue;
1898 struct cfq_io_context *cic;
1899 unsigned long sl, group_idle = 0;
1902 * SSD device without seek penalty, disable idling. But only do so
1903 * for devices that support queuing, otherwise we still have a problem
1904 * with sync vs async workloads.
1906 if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
1909 WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
1910 WARN_ON(cfq_cfqq_slice_new(cfqq));
1913 * idle is disabled, either manually or by past process history
1915 if (!cfq_should_idle(cfqd, cfqq)) {
1916 /* no queue idling. Check for group idling */
1917 if (cfqd->cfq_group_idle)
1918 group_idle = cfqd->cfq_group_idle;
1924 * still active requests from this queue, don't idle
1926 if (cfqq->dispatched)
1930 * task has exited, don't wait
1932 cic = cfqd->active_cic;
1933 if (!cic || !atomic_read(&cic->ioc->nr_tasks))
1937 * If our average think time is larger than the remaining time
1938 * slice, then don't idle. This avoids overrunning the allotted
1941 if (sample_valid(cic->ttime_samples) &&
1942 (cfqq->slice_end - jiffies < cic->ttime_mean)) {
1943 cfq_log_cfqq(cfqd, cfqq, "Not idling. think_time:%d",
1948 /* There are other queues in the group, don't do group idle */
1949 if (group_idle && cfqq->cfqg->nr_cfqq > 1)
1952 cfq_mark_cfqq_wait_request(cfqq);
1955 sl = cfqd->cfq_group_idle;
1957 sl = cfqd->cfq_slice_idle;
1959 mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
1960 cfq_blkiocg_update_set_idle_time_stats(&cfqq->cfqg->blkg);
1961 cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu group_idle: %d", sl,
1962 group_idle ? 1 : 0);
1966 * Move request from internal lists to the request queue dispatch list.
1968 static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
1970 struct cfq_data *cfqd = q->elevator->elevator_data;
1971 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1973 cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
1975 cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
1976 cfq_remove_request(rq);
1978 (RQ_CFQG(rq))->dispatched++;
1979 elv_dispatch_sort(q, rq);
1981 cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]++;
1982 cfqq->nr_sectors += blk_rq_sectors(rq);
1983 cfq_blkiocg_update_dispatch_stats(&cfqq->cfqg->blkg, blk_rq_bytes(rq),
1984 rq_data_dir(rq), rq_is_sync(rq));
1988 * return expired entry, or NULL to just start from scratch in rbtree
1990 static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
1992 struct request *rq = NULL;
1994 if (cfq_cfqq_fifo_expire(cfqq))
1997 cfq_mark_cfqq_fifo_expire(cfqq);
1999 if (list_empty(&cfqq->fifo))
2002 rq = rq_entry_fifo(cfqq->fifo.next);
2003 if (time_before(jiffies, rq_fifo_time(rq)))
2006 cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq);
2011 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2013 const int base_rq = cfqd->cfq_slice_async_rq;
2015 WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
2017 return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
2021 * Must be called with the queue_lock held.
2023 static int cfqq_process_refs(struct cfq_queue *cfqq)
2025 int process_refs, io_refs;
2027 io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE];
2028 process_refs = cfqq->ref - io_refs;
2029 BUG_ON(process_refs < 0);
2030 return process_refs;
2033 static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq)
2035 int process_refs, new_process_refs;
2036 struct cfq_queue *__cfqq;
2039 * If there are no process references on the new_cfqq, then it is
2040 * unsafe to follow the ->new_cfqq chain as other cfqq's in the
2041 * chain may have dropped their last reference (not just their
2042 * last process reference).
2044 if (!cfqq_process_refs(new_cfqq))
2047 /* Avoid a circular list and skip interim queue merges */
2048 while ((__cfqq = new_cfqq->new_cfqq)) {
2054 process_refs = cfqq_process_refs(cfqq);
2055 new_process_refs = cfqq_process_refs(new_cfqq);
2057 * If the process for the cfqq has gone away, there is no
2058 * sense in merging the queues.
2060 if (process_refs == 0 || new_process_refs == 0)
2064 * Merge in the direction of the lesser amount of work.
2066 if (new_process_refs >= process_refs) {
2067 cfqq->new_cfqq = new_cfqq;
2068 new_cfqq->ref += process_refs;
2070 new_cfqq->new_cfqq = cfqq;
2071 cfqq->ref += new_process_refs;
2075 static enum wl_type_t cfq_choose_wl(struct cfq_data *cfqd,
2076 struct cfq_group *cfqg, enum wl_prio_t prio)
2078 struct cfq_queue *queue;
2080 bool key_valid = false;
2081 unsigned long lowest_key = 0;
2082 enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD;
2084 for (i = 0; i <= SYNC_WORKLOAD; ++i) {
2085 /* select the one with lowest rb_key */
2086 queue = cfq_rb_first(service_tree_for(cfqg, prio, i));
2088 (!key_valid || time_before(queue->rb_key, lowest_key))) {
2089 lowest_key = queue->rb_key;
2098 static void choose_service_tree(struct cfq_data *cfqd, struct cfq_group *cfqg)
2102 struct cfq_rb_root *st;
2103 unsigned group_slice;
2104 enum wl_prio_t original_prio = cfqd->serving_prio;
2106 /* Choose next priority. RT > BE > IDLE */
2107 if (cfq_group_busy_queues_wl(RT_WORKLOAD, cfqd, cfqg))
2108 cfqd->serving_prio = RT_WORKLOAD;
2109 else if (cfq_group_busy_queues_wl(BE_WORKLOAD, cfqd, cfqg))
2110 cfqd->serving_prio = BE_WORKLOAD;
2112 cfqd->serving_prio = IDLE_WORKLOAD;
2113 cfqd->workload_expires = jiffies + 1;
2117 if (original_prio != cfqd->serving_prio)
2121 * For RT and BE, we have to choose also the type
2122 * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
2125 st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
2129 * check workload expiration, and that we still have other queues ready
2131 if (count && !time_after(jiffies, cfqd->workload_expires))
2135 /* otherwise select new workload type */
2136 cfqd->serving_type =
2137 cfq_choose_wl(cfqd, cfqg, cfqd->serving_prio);
2138 st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
2142 * the workload slice is computed as a fraction of target latency
2143 * proportional to the number of queues in that workload, over
2144 * all the queues in the same priority class
2146 group_slice = cfq_group_slice(cfqd, cfqg);
2148 slice = group_slice * count /
2149 max_t(unsigned, cfqg->busy_queues_avg[cfqd->serving_prio],
2150 cfq_group_busy_queues_wl(cfqd->serving_prio, cfqd, cfqg));
2152 if (cfqd->serving_type == ASYNC_WORKLOAD) {
2156 * Async queues are currently system wide. Just taking
2157 * proportion of queues with-in same group will lead to higher
2158 * async ratio system wide as generally root group is going
2159 * to have higher weight. A more accurate thing would be to
2160 * calculate system wide asnc/sync ratio.
2162 tmp = cfq_target_latency * cfqg_busy_async_queues(cfqd, cfqg);
2163 tmp = tmp/cfqd->busy_queues;
2164 slice = min_t(unsigned, slice, tmp);
2166 /* async workload slice is scaled down according to
2167 * the sync/async slice ratio. */
2168 slice = slice * cfqd->cfq_slice[0] / cfqd->cfq_slice[1];
2170 /* sync workload slice is at least 2 * cfq_slice_idle */
2171 slice = max(slice, 2 * cfqd->cfq_slice_idle);
2173 slice = max_t(unsigned, slice, CFQ_MIN_TT);
2174 cfq_log(cfqd, "workload slice:%d", slice);
2175 cfqd->workload_expires = jiffies + slice;
2178 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd)
2180 struct cfq_rb_root *st = &cfqd->grp_service_tree;
2181 struct cfq_group *cfqg;
2183 if (RB_EMPTY_ROOT(&st->rb))
2185 cfqg = cfq_rb_first_group(st);
2186 update_min_vdisktime(st);
2190 static void cfq_choose_cfqg(struct cfq_data *cfqd)
2192 struct cfq_group *cfqg = cfq_get_next_cfqg(cfqd);
2194 cfqd->serving_group = cfqg;
2196 /* Restore the workload type data */
2197 if (cfqg->saved_workload_slice) {
2198 cfqd->workload_expires = jiffies + cfqg->saved_workload_slice;
2199 cfqd->serving_type = cfqg->saved_workload;
2200 cfqd->serving_prio = cfqg->saved_serving_prio;
2202 cfqd->workload_expires = jiffies - 1;
2204 choose_service_tree(cfqd, cfqg);
2208 * Select a queue for service. If we have a current active queue,
2209 * check whether to continue servicing it, or retrieve and set a new one.
2211 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
2213 struct cfq_queue *cfqq, *new_cfqq = NULL;
2215 cfqq = cfqd->active_queue;
2219 if (!cfqd->rq_queued)
2223 * We were waiting for group to get backlogged. Expire the queue
2225 if (cfq_cfqq_wait_busy(cfqq) && !RB_EMPTY_ROOT(&cfqq->sort_list))
2229 * The active queue has run out of time, expire it and select new.
2231 if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq)) {
2233 * If slice had not expired at the completion of last request
2234 * we might not have turned on wait_busy flag. Don't expire
2235 * the queue yet. Allow the group to get backlogged.
2237 * The very fact that we have used the slice, that means we
2238 * have been idling all along on this queue and it should be
2239 * ok to wait for this request to complete.
2241 if (cfqq->cfqg->nr_cfqq == 1 && RB_EMPTY_ROOT(&cfqq->sort_list)
2242 && cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2246 goto check_group_idle;
2250 * The active queue has requests and isn't expired, allow it to
2253 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
2257 * If another queue has a request waiting within our mean seek
2258 * distance, let it run. The expire code will check for close
2259 * cooperators and put the close queue at the front of the service
2260 * tree. If possible, merge the expiring queue with the new cfqq.
2262 new_cfqq = cfq_close_cooperator(cfqd, cfqq);
2264 if (!cfqq->new_cfqq)
2265 cfq_setup_merge(cfqq, new_cfqq);
2270 * No requests pending. If the active queue still has requests in
2271 * flight or is idling for a new request, allow either of these
2272 * conditions to happen (or time out) before selecting a new queue.
2274 if (timer_pending(&cfqd->idle_slice_timer)) {
2280 * This is a deep seek queue, but the device is much faster than
2281 * the queue can deliver, don't idle
2283 if (CFQQ_SEEKY(cfqq) && cfq_cfqq_idle_window(cfqq) &&
2284 (cfq_cfqq_slice_new(cfqq) ||
2285 (cfqq->slice_end - jiffies > jiffies - cfqq->slice_start))) {
2286 cfq_clear_cfqq_deep(cfqq);
2287 cfq_clear_cfqq_idle_window(cfqq);
2290 if (cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2296 * If group idle is enabled and there are requests dispatched from
2297 * this group, wait for requests to complete.
2300 if (cfqd->cfq_group_idle && cfqq->cfqg->nr_cfqq == 1
2301 && cfqq->cfqg->dispatched) {
2307 cfq_slice_expired(cfqd, 0);
2310 * Current queue expired. Check if we have to switch to a new
2314 cfq_choose_cfqg(cfqd);
2316 cfqq = cfq_set_active_queue(cfqd, new_cfqq);
2321 static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
2325 while (cfqq->next_rq) {
2326 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
2330 BUG_ON(!list_empty(&cfqq->fifo));
2332 /* By default cfqq is not expired if it is empty. Do it explicitly */
2333 __cfq_slice_expired(cfqq->cfqd, cfqq, 0);
2338 * Drain our current requests. Used for barriers and when switching
2339 * io schedulers on-the-fly.
2341 static int cfq_forced_dispatch(struct cfq_data *cfqd)
2343 struct cfq_queue *cfqq;
2346 /* Expire the timeslice of the current active queue first */
2347 cfq_slice_expired(cfqd, 0);
2348 while ((cfqq = cfq_get_next_queue_forced(cfqd)) != NULL) {
2349 __cfq_set_active_queue(cfqd, cfqq);
2350 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
2353 BUG_ON(cfqd->busy_queues);
2355 cfq_log(cfqd, "forced_dispatch=%d", dispatched);
2359 static inline bool cfq_slice_used_soon(struct cfq_data *cfqd,
2360 struct cfq_queue *cfqq)
2362 /* the queue hasn't finished any request, can't estimate */
2363 if (cfq_cfqq_slice_new(cfqq))
2365 if (time_after(jiffies + cfqd->cfq_slice_idle * cfqq->dispatched,
2372 static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2374 unsigned int max_dispatch;
2377 * Drain async requests before we start sync IO
2379 if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_flight[BLK_RW_ASYNC])
2383 * If this is an async queue and we have sync IO in flight, let it wait
2385 if (cfqd->rq_in_flight[BLK_RW_SYNC] && !cfq_cfqq_sync(cfqq))
2388 max_dispatch = max_t(unsigned int, cfqd->cfq_quantum / 2, 1);
2389 if (cfq_class_idle(cfqq))
2393 * Does this cfqq already have too much IO in flight?
2395 if (cfqq->dispatched >= max_dispatch) {
2397 * idle queue must always only have a single IO in flight
2399 if (cfq_class_idle(cfqq))
2403 * We have other queues, don't allow more IO from this one
2405 if (cfqd->busy_queues > 1 && cfq_slice_used_soon(cfqd, cfqq))
2409 * Sole queue user, no limit
2411 if (cfqd->busy_queues == 1)
2415 * Normally we start throttling cfqq when cfq_quantum/2
2416 * requests have been dispatched. But we can drive
2417 * deeper queue depths at the beginning of slice
2418 * subjected to upper limit of cfq_quantum.
2420 max_dispatch = cfqd->cfq_quantum;
2424 * Async queues must wait a bit before being allowed dispatch.
2425 * We also ramp up the dispatch depth gradually for async IO,
2426 * based on the last sync IO we serviced
2428 if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
2429 unsigned long last_sync = jiffies - cfqd->last_delayed_sync;
2432 depth = last_sync / cfqd->cfq_slice[1];
2433 if (!depth && !cfqq->dispatched)
2435 if (depth < max_dispatch)
2436 max_dispatch = depth;
2440 * If we're below the current max, allow a dispatch
2442 return cfqq->dispatched < max_dispatch;
2446 * Dispatch a request from cfqq, moving them to the request queue
2449 static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2453 BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
2455 if (!cfq_may_dispatch(cfqd, cfqq))
2459 * follow expired path, else get first next available
2461 rq = cfq_check_fifo(cfqq);
2466 * insert request into driver dispatch list
2468 cfq_dispatch_insert(cfqd->queue, rq);
2470 if (!cfqd->active_cic) {
2471 struct cfq_io_context *cic = RQ_CIC(rq);
2473 atomic_long_inc(&cic->ioc->refcount);
2474 cfqd->active_cic = cic;
2481 * Find the cfqq that we need to service and move a request from that to the
2484 static int cfq_dispatch_requests(struct request_queue *q, int force)
2486 struct cfq_data *cfqd = q->elevator->elevator_data;
2487 struct cfq_queue *cfqq;
2489 if (!cfqd->busy_queues)
2492 if (unlikely(force))
2493 return cfq_forced_dispatch(cfqd);
2495 cfqq = cfq_select_queue(cfqd);
2500 * Dispatch a request from this cfqq, if it is allowed
2502 if (!cfq_dispatch_request(cfqd, cfqq))
2505 cfqq->slice_dispatch++;
2506 cfq_clear_cfqq_must_dispatch(cfqq);
2509 * expire an async queue immediately if it has used up its slice. idle
2510 * queue always expire after 1 dispatch round.
2512 if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
2513 cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
2514 cfq_class_idle(cfqq))) {
2515 cfqq->slice_end = jiffies + 1;
2516 cfq_slice_expired(cfqd, 0);
2519 cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
2524 * task holds one reference to the queue, dropped when task exits. each rq
2525 * in-flight on this queue also holds a reference, dropped when rq is freed.
2527 * Each cfq queue took a reference on the parent group. Drop it now.
2528 * queue lock must be held here.
2530 static void cfq_put_queue(struct cfq_queue *cfqq)
2532 struct cfq_data *cfqd = cfqq->cfqd;
2533 struct cfq_group *cfqg, *orig_cfqg;
2535 BUG_ON(cfqq->ref <= 0);
2541 cfq_log_cfqq(cfqd, cfqq, "put_queue");
2542 BUG_ON(rb_first(&cfqq->sort_list));
2543 BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
2545 orig_cfqg = cfqq->orig_cfqg;
2547 if (unlikely(cfqd->active_queue == cfqq)) {
2548 __cfq_slice_expired(cfqd, cfqq, 0);
2549 cfq_schedule_dispatch(cfqd);
2552 BUG_ON(cfq_cfqq_on_rr(cfqq));
2553 kmem_cache_free(cfq_pool, cfqq);
2556 cfq_put_cfqg(orig_cfqg);
2560 * Must always be called with the rcu_read_lock() held
2563 __call_for_each_cic(struct io_context *ioc,
2564 void (*func)(struct io_context *, struct cfq_io_context *))
2566 struct cfq_io_context *cic;
2567 struct hlist_node *n;
2569 hlist_for_each_entry_rcu(cic, n, &ioc->cic_list, cic_list)
2574 * Call func for each cic attached to this ioc.
2577 call_for_each_cic(struct io_context *ioc,
2578 void (*func)(struct io_context *, struct cfq_io_context *))
2581 __call_for_each_cic(ioc, func);
2585 static void cfq_cic_free_rcu(struct rcu_head *head)
2587 struct cfq_io_context *cic;
2589 cic = container_of(head, struct cfq_io_context, rcu_head);
2591 kmem_cache_free(cfq_ioc_pool, cic);
2592 elv_ioc_count_dec(cfq_ioc_count);
2596 * CFQ scheduler is exiting, grab exit lock and check
2597 * the pending io context count. If it hits zero,
2598 * complete ioc_gone and set it back to NULL
2600 spin_lock(&ioc_gone_lock);
2601 if (ioc_gone && !elv_ioc_count_read(cfq_ioc_count)) {
2605 spin_unlock(&ioc_gone_lock);
2609 static void cfq_cic_free(struct cfq_io_context *cic)
2611 call_rcu(&cic->rcu_head, cfq_cic_free_rcu);
2614 static void cic_free_func(struct io_context *ioc, struct cfq_io_context *cic)
2616 unsigned long flags;
2617 unsigned long dead_key = (unsigned long) cic->key;
2619 BUG_ON(!(dead_key & CIC_DEAD_KEY));
2621 spin_lock_irqsave(&ioc->lock, flags);
2622 radix_tree_delete(&ioc->radix_root, dead_key >> CIC_DEAD_INDEX_SHIFT);
2623 hlist_del_rcu(&cic->cic_list);
2624 spin_unlock_irqrestore(&ioc->lock, flags);
2630 * Must be called with rcu_read_lock() held or preemption otherwise disabled.
2631 * Only two callers of this - ->dtor() which is called with the rcu_read_lock(),
2632 * and ->trim() which is called with the task lock held
2634 static void cfq_free_io_context(struct io_context *ioc)
2637 * ioc->refcount is zero here, or we are called from elv_unregister(),
2638 * so no more cic's are allowed to be linked into this ioc. So it
2639 * should be ok to iterate over the known list, we will see all cic's
2640 * since no new ones are added.
2642 __call_for_each_cic(ioc, cic_free_func);
2645 static void cfq_put_cooperator(struct cfq_queue *cfqq)
2647 struct cfq_queue *__cfqq, *next;
2650 * If this queue was scheduled to merge with another queue, be
2651 * sure to drop the reference taken on that queue (and others in
2652 * the merge chain). See cfq_setup_merge and cfq_merge_cfqqs.
2654 __cfqq = cfqq->new_cfqq;
2656 if (__cfqq == cfqq) {
2657 WARN(1, "cfqq->new_cfqq loop detected\n");
2660 next = __cfqq->new_cfqq;
2661 cfq_put_queue(__cfqq);
2666 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2668 if (unlikely(cfqq == cfqd->active_queue)) {
2669 __cfq_slice_expired(cfqd, cfqq, 0);
2670 cfq_schedule_dispatch(cfqd);
2673 cfq_put_cooperator(cfqq);
2675 cfq_put_queue(cfqq);
2678 static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
2679 struct cfq_io_context *cic)
2681 struct io_context *ioc = cic->ioc;
2683 list_del_init(&cic->queue_list);
2686 * Make sure dead mark is seen for dead queues
2689 cic->key = cfqd_dead_key(cfqd);
2691 if (ioc->ioc_data == cic)
2692 rcu_assign_pointer(ioc->ioc_data, NULL);
2694 if (cic->cfqq[BLK_RW_ASYNC]) {
2695 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]);
2696 cic->cfqq[BLK_RW_ASYNC] = NULL;
2699 if (cic->cfqq[BLK_RW_SYNC]) {
2700 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]);
2701 cic->cfqq[BLK_RW_SYNC] = NULL;
2705 static void cfq_exit_single_io_context(struct io_context *ioc,
2706 struct cfq_io_context *cic)
2708 struct cfq_data *cfqd = cic_to_cfqd(cic);
2711 struct request_queue *q = cfqd->queue;
2712 unsigned long flags;
2714 spin_lock_irqsave(q->queue_lock, flags);
2717 * Ensure we get a fresh copy of the ->key to prevent
2718 * race between exiting task and queue
2720 smp_read_barrier_depends();
2721 if (cic->key == cfqd)
2722 __cfq_exit_single_io_context(cfqd, cic);
2724 spin_unlock_irqrestore(q->queue_lock, flags);
2729 * The process that ioc belongs to has exited, we need to clean up
2730 * and put the internal structures we have that belongs to that process.
2732 static void cfq_exit_io_context(struct io_context *ioc)
2734 call_for_each_cic(ioc, cfq_exit_single_io_context);
2737 static struct cfq_io_context *
2738 cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
2740 struct cfq_io_context *cic;
2742 cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO,
2745 cic->last_end_request = jiffies;
2746 INIT_LIST_HEAD(&cic->queue_list);
2747 INIT_HLIST_NODE(&cic->cic_list);
2748 cic->dtor = cfq_free_io_context;
2749 cic->exit = cfq_exit_io_context;
2750 elv_ioc_count_inc(cfq_ioc_count);
2756 static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc)
2758 struct task_struct *tsk = current;
2761 if (!cfq_cfqq_prio_changed(cfqq))
2764 ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio);
2765 switch (ioprio_class) {
2767 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
2768 case IOPRIO_CLASS_NONE:
2770 * no prio set, inherit CPU scheduling settings
2772 cfqq->ioprio = task_nice_ioprio(tsk);
2773 cfqq->ioprio_class = task_nice_ioclass(tsk);
2775 case IOPRIO_CLASS_RT:
2776 cfqq->ioprio = task_ioprio(ioc);
2777 cfqq->ioprio_class = IOPRIO_CLASS_RT;
2779 case IOPRIO_CLASS_BE:
2780 cfqq->ioprio = task_ioprio(ioc);
2781 cfqq->ioprio_class = IOPRIO_CLASS_BE;
2783 case IOPRIO_CLASS_IDLE:
2784 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
2786 cfq_clear_cfqq_idle_window(cfqq);
2791 * keep track of original prio settings in case we have to temporarily
2792 * elevate the priority of this queue
2794 cfqq->org_ioprio = cfqq->ioprio;
2795 cfqq->org_ioprio_class = cfqq->ioprio_class;
2796 cfq_clear_cfqq_prio_changed(cfqq);
2799 static void changed_ioprio(struct io_context *ioc, struct cfq_io_context *cic)
2801 struct cfq_data *cfqd = cic_to_cfqd(cic);
2802 struct cfq_queue *cfqq;
2803 unsigned long flags;
2805 if (unlikely(!cfqd))
2808 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2810 cfqq = cic->cfqq[BLK_RW_ASYNC];
2812 struct cfq_queue *new_cfqq;
2813 new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic->ioc,
2816 cic->cfqq[BLK_RW_ASYNC] = new_cfqq;
2817 cfq_put_queue(cfqq);
2821 cfqq = cic->cfqq[BLK_RW_SYNC];
2823 cfq_mark_cfqq_prio_changed(cfqq);
2825 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2828 static void cfq_ioc_set_ioprio(struct io_context *ioc)
2830 call_for_each_cic(ioc, changed_ioprio);
2831 ioc->ioprio_changed = 0;
2834 static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2835 pid_t pid, bool is_sync)
2837 RB_CLEAR_NODE(&cfqq->rb_node);
2838 RB_CLEAR_NODE(&cfqq->p_node);
2839 INIT_LIST_HEAD(&cfqq->fifo);
2844 cfq_mark_cfqq_prio_changed(cfqq);
2847 if (!cfq_class_idle(cfqq))
2848 cfq_mark_cfqq_idle_window(cfqq);
2849 cfq_mark_cfqq_sync(cfqq);
2854 #ifdef CONFIG_CFQ_GROUP_IOSCHED
2855 static void changed_cgroup(struct io_context *ioc, struct cfq_io_context *cic)
2857 struct cfq_queue *sync_cfqq = cic_to_cfqq(cic, 1);
2858 struct cfq_data *cfqd = cic_to_cfqd(cic);
2859 unsigned long flags;
2860 struct request_queue *q;
2862 if (unlikely(!cfqd))
2867 spin_lock_irqsave(q->queue_lock, flags);
2871 * Drop reference to sync queue. A new sync queue will be
2872 * assigned in new group upon arrival of a fresh request.
2874 cfq_log_cfqq(cfqd, sync_cfqq, "changed cgroup");
2875 cic_set_cfqq(cic, NULL, 1);
2876 cfq_put_queue(sync_cfqq);
2879 spin_unlock_irqrestore(q->queue_lock, flags);
2882 static void cfq_ioc_set_cgroup(struct io_context *ioc)
2884 call_for_each_cic(ioc, changed_cgroup);
2885 ioc->cgroup_changed = 0;
2887 #endif /* CONFIG_CFQ_GROUP_IOSCHED */
2889 static struct cfq_queue *
2890 cfq_find_alloc_queue(struct cfq_data *cfqd, bool is_sync,
2891 struct io_context *ioc, gfp_t gfp_mask)
2893 struct cfq_queue *cfqq, *new_cfqq = NULL;
2894 struct cfq_io_context *cic;
2895 struct cfq_group *cfqg;
2898 cfqg = cfq_get_cfqg(cfqd, 1);
2899 cic = cfq_cic_lookup(cfqd, ioc);
2900 /* cic always exists here */
2901 cfqq = cic_to_cfqq(cic, is_sync);
2904 * Always try a new alloc if we fell back to the OOM cfqq
2905 * originally, since it should just be a temporary situation.
2907 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
2912 } else if (gfp_mask & __GFP_WAIT) {
2913 spin_unlock_irq(cfqd->queue->queue_lock);
2914 new_cfqq = kmem_cache_alloc_node(cfq_pool,
2915 gfp_mask | __GFP_ZERO,
2917 spin_lock_irq(cfqd->queue->queue_lock);
2921 cfqq = kmem_cache_alloc_node(cfq_pool,
2922 gfp_mask | __GFP_ZERO,
2927 cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
2928 cfq_init_prio_data(cfqq, ioc);
2929 cfq_link_cfqq_cfqg(cfqq, cfqg);
2930 cfq_log_cfqq(cfqd, cfqq, "alloced");
2932 cfqq = &cfqd->oom_cfqq;
2936 kmem_cache_free(cfq_pool, new_cfqq);
2941 static struct cfq_queue **
2942 cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
2944 switch (ioprio_class) {
2945 case IOPRIO_CLASS_RT:
2946 return &cfqd->async_cfqq[0][ioprio];
2947 case IOPRIO_CLASS_BE:
2948 return &cfqd->async_cfqq[1][ioprio];
2949 case IOPRIO_CLASS_IDLE:
2950 return &cfqd->async_idle_cfqq;
2956 static struct cfq_queue *
2957 cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct io_context *ioc,
2960 const int ioprio = task_ioprio(ioc);
2961 const int ioprio_class = task_ioprio_class(ioc);
2962 struct cfq_queue **async_cfqq = NULL;
2963 struct cfq_queue *cfqq = NULL;
2966 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
2971 cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask);
2974 * pin the queue now that it's allocated, scheduler exit will prune it
2976 if (!is_sync && !(*async_cfqq)) {
2986 * We drop cfq io contexts lazily, so we may find a dead one.
2989 cfq_drop_dead_cic(struct cfq_data *cfqd, struct io_context *ioc,
2990 struct cfq_io_context *cic)
2992 unsigned long flags;
2994 WARN_ON(!list_empty(&cic->queue_list));
2995 BUG_ON(cic->key != cfqd_dead_key(cfqd));
2997 spin_lock_irqsave(&ioc->lock, flags);
2999 BUG_ON(ioc->ioc_data == cic);
3001 radix_tree_delete(&ioc->radix_root, cfqd->cic_index);
3002 hlist_del_rcu(&cic->cic_list);
3003 spin_unlock_irqrestore(&ioc->lock, flags);
3008 static struct cfq_io_context *
3009 cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc)
3011 struct cfq_io_context *cic;
3012 unsigned long flags;
3020 * we maintain a last-hit cache, to avoid browsing over the tree
3022 cic = rcu_dereference(ioc->ioc_data);
3023 if (cic && cic->key == cfqd) {
3029 cic = radix_tree_lookup(&ioc->radix_root, cfqd->cic_index);
3033 if (unlikely(cic->key != cfqd)) {
3034 cfq_drop_dead_cic(cfqd, ioc, cic);
3039 spin_lock_irqsave(&ioc->lock, flags);
3040 rcu_assign_pointer(ioc->ioc_data, cic);
3041 spin_unlock_irqrestore(&ioc->lock, flags);
3049 * Add cic into ioc, using cfqd as the search key. This enables us to lookup
3050 * the process specific cfq io context when entered from the block layer.
3051 * Also adds the cic to a per-cfqd list, used when this queue is removed.
3053 static int cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
3054 struct cfq_io_context *cic, gfp_t gfp_mask)
3056 unsigned long flags;
3059 ret = radix_tree_preload(gfp_mask);
3064 spin_lock_irqsave(&ioc->lock, flags);
3065 ret = radix_tree_insert(&ioc->radix_root,
3066 cfqd->cic_index, cic);
3068 hlist_add_head_rcu(&cic->cic_list, &ioc->cic_list);
3069 spin_unlock_irqrestore(&ioc->lock, flags);
3071 radix_tree_preload_end();
3074 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
3075 list_add(&cic->queue_list, &cfqd->cic_list);
3076 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
3081 printk(KERN_ERR "cfq: cic link failed!\n");
3087 * Setup general io context and cfq io context. There can be several cfq
3088 * io contexts per general io context, if this process is doing io to more
3089 * than one device managed by cfq.
3091 static struct cfq_io_context *
3092 cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
3094 struct io_context *ioc = NULL;
3095 struct cfq_io_context *cic;
3097 might_sleep_if(gfp_mask & __GFP_WAIT);
3099 ioc = get_io_context(gfp_mask, cfqd->queue->node);
3103 cic = cfq_cic_lookup(cfqd, ioc);
3107 cic = cfq_alloc_io_context(cfqd, gfp_mask);
3111 if (cfq_cic_link(cfqd, ioc, cic, gfp_mask))
3115 smp_read_barrier_depends();
3116 if (unlikely(ioc->ioprio_changed))
3117 cfq_ioc_set_ioprio(ioc);
3119 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3120 if (unlikely(ioc->cgroup_changed))
3121 cfq_ioc_set_cgroup(ioc);
3127 put_io_context(ioc);
3132 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
3134 unsigned long elapsed = jiffies - cic->last_end_request;
3135 unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
3137 cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
3138 cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
3139 cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
3143 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3147 sector_t n_sec = blk_rq_sectors(rq);
3148 if (cfqq->last_request_pos) {
3149 if (cfqq->last_request_pos < blk_rq_pos(rq))
3150 sdist = blk_rq_pos(rq) - cfqq->last_request_pos;
3152 sdist = cfqq->last_request_pos - blk_rq_pos(rq);
3155 cfqq->seek_history <<= 1;
3156 if (blk_queue_nonrot(cfqd->queue))
3157 cfqq->seek_history |= (n_sec < CFQQ_SECT_THR_NONROT);
3159 cfqq->seek_history |= (sdist > CFQQ_SEEK_THR);
3163 * Disable idle window if the process thinks too long or seeks so much that
3167 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3168 struct cfq_io_context *cic)
3170 int old_idle, enable_idle;
3173 * Don't idle for async or idle io prio class
3175 if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
3178 enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
3180 if (cfqq->queued[0] + cfqq->queued[1] >= 4)
3181 cfq_mark_cfqq_deep(cfqq);
3183 if (cfqq->next_rq && (cfqq->next_rq->cmd_flags & REQ_NOIDLE))
3185 else if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle ||
3186 (!cfq_cfqq_deep(cfqq) && CFQQ_SEEKY(cfqq)))
3188 else if (sample_valid(cic->ttime_samples)) {
3189 if (cic->ttime_mean > cfqd->cfq_slice_idle)
3195 if (old_idle != enable_idle) {
3196 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
3198 cfq_mark_cfqq_idle_window(cfqq);
3200 cfq_clear_cfqq_idle_window(cfqq);
3205 * Check if new_cfqq should preempt the currently active queue. Return 0 for
3206 * no or if we aren't sure, a 1 will cause a preempt.
3209 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
3212 struct cfq_queue *cfqq;
3214 cfqq = cfqd->active_queue;
3218 if (cfq_class_idle(new_cfqq))
3221 if (cfq_class_idle(cfqq))
3225 * Don't allow a non-RT request to preempt an ongoing RT cfqq timeslice.
3227 if (cfq_class_rt(cfqq) && !cfq_class_rt(new_cfqq))
3231 * if the new request is sync, but the currently running queue is
3232 * not, let the sync request have priority.
3234 if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
3237 if (new_cfqq->cfqg != cfqq->cfqg)
3240 if (cfq_slice_used(cfqq))
3243 /* Allow preemption only if we are idling on sync-noidle tree */
3244 if (cfqd->serving_type == SYNC_NOIDLE_WORKLOAD &&
3245 cfqq_type(new_cfqq) == SYNC_NOIDLE_WORKLOAD &&
3246 new_cfqq->service_tree->count == 2 &&
3247 RB_EMPTY_ROOT(&cfqq->sort_list))
3251 * So both queues are sync. Let the new request get disk time if
3252 * it's a metadata request and the current queue is doing regular IO.
3254 if ((rq->cmd_flags & REQ_META) && !cfqq->meta_pending)
3258 * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
3260 if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
3263 /* An idle queue should not be idle now for some reason */
3264 if (RB_EMPTY_ROOT(&cfqq->sort_list) && !cfq_should_idle(cfqd, cfqq))
3267 if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
3271 * if this request is as-good as one we would expect from the
3272 * current cfqq, let it preempt
3274 if (cfq_rq_close(cfqd, cfqq, rq))
3281 * cfqq preempts the active queue. if we allowed preempt with no slice left,
3282 * let it have half of its nominal slice.
3284 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3286 cfq_log_cfqq(cfqd, cfqq, "preempt");
3287 cfq_slice_expired(cfqd, 1);
3290 * Put the new queue at the front of the of the current list,
3291 * so we know that it will be selected next.
3293 BUG_ON(!cfq_cfqq_on_rr(cfqq));
3295 cfq_service_tree_add(cfqd, cfqq, 1);
3297 cfqq->slice_end = 0;
3298 cfq_mark_cfqq_slice_new(cfqq);
3302 * Called when a new fs request (rq) is added (to cfqq). Check if there's
3303 * something we should do about it
3306 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3309 struct cfq_io_context *cic = RQ_CIC(rq);
3312 if (rq->cmd_flags & REQ_META)
3313 cfqq->meta_pending++;
3315 cfq_update_io_thinktime(cfqd, cic);
3316 cfq_update_io_seektime(cfqd, cfqq, rq);
3317 cfq_update_idle_window(cfqd, cfqq, cic);
3319 cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
3321 if (cfqq == cfqd->active_queue) {
3323 * Remember that we saw a request from this process, but
3324 * don't start queuing just yet. Otherwise we risk seeing lots
3325 * of tiny requests, because we disrupt the normal plugging
3326 * and merging. If the request is already larger than a single
3327 * page, let it rip immediately. For that case we assume that
3328 * merging is already done. Ditto for a busy system that
3329 * has other work pending, don't risk delaying until the
3330 * idle timer unplug to continue working.
3332 if (cfq_cfqq_wait_request(cfqq)) {
3333 if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
3334 cfqd->busy_queues > 1) {
3335 cfq_del_timer(cfqd, cfqq);
3336 cfq_clear_cfqq_wait_request(cfqq);
3337 __blk_run_queue(cfqd->queue);
3339 cfq_blkiocg_update_idle_time_stats(
3341 cfq_mark_cfqq_must_dispatch(cfqq);
3344 } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
3346 * not the active queue - expire current slice if it is
3347 * idle and has expired it's mean thinktime or this new queue
3348 * has some old slice time left and is of higher priority or
3349 * this new queue is RT and the current one is BE
3351 cfq_preempt_queue(cfqd, cfqq);
3352 __blk_run_queue(cfqd->queue);
3356 static void cfq_insert_request(struct request_queue *q, struct request *rq)
3358 struct cfq_data *cfqd = q->elevator->elevator_data;
3359 struct cfq_queue *cfqq = RQ_CFQQ(rq);
3361 cfq_log_cfqq(cfqd, cfqq, "insert_request");
3362 cfq_init_prio_data(cfqq, RQ_CIC(rq)->ioc);
3364 rq_set_fifo_time(rq, jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)]);
3365 list_add_tail(&rq->queuelist, &cfqq->fifo);
3367 cfq_blkiocg_update_io_add_stats(&(RQ_CFQG(rq))->blkg,
3368 &cfqd->serving_group->blkg, rq_data_dir(rq),
3370 cfq_rq_enqueued(cfqd, cfqq, rq);
3374 * Update hw_tag based on peak queue depth over 50 samples under
3377 static void cfq_update_hw_tag(struct cfq_data *cfqd)
3379 struct cfq_queue *cfqq = cfqd->active_queue;
3381 if (cfqd->rq_in_driver > cfqd->hw_tag_est_depth)
3382 cfqd->hw_tag_est_depth = cfqd->rq_in_driver;
3384 if (cfqd->hw_tag == 1)
3387 if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
3388 cfqd->rq_in_driver <= CFQ_HW_QUEUE_MIN)
3392 * If active queue hasn't enough requests and can idle, cfq might not
3393 * dispatch sufficient requests to hardware. Don't zero hw_tag in this
3396 if (cfqq && cfq_cfqq_idle_window(cfqq) &&
3397 cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] <
3398 CFQ_HW_QUEUE_MIN && cfqd->rq_in_driver < CFQ_HW_QUEUE_MIN)
3401 if (cfqd->hw_tag_samples++ < 50)
3404 if (cfqd->hw_tag_est_depth >= CFQ_HW_QUEUE_MIN)
3410 static bool cfq_should_wait_busy(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3412 struct cfq_io_context *cic = cfqd->active_cic;
3414 /* If there are other queues in the group, don't wait */
3415 if (cfqq->cfqg->nr_cfqq > 1)
3418 if (cfq_slice_used(cfqq))
3421 /* if slice left is less than think time, wait busy */
3422 if (cic && sample_valid(cic->ttime_samples)
3423 && (cfqq->slice_end - jiffies < cic->ttime_mean))
3427 * If think times is less than a jiffy than ttime_mean=0 and above
3428 * will not be true. It might happen that slice has not expired yet
3429 * but will expire soon (4-5 ns) during select_queue(). To cover the
3430 * case where think time is less than a jiffy, mark the queue wait
3431 * busy if only 1 jiffy is left in the slice.
3433 if (cfqq->slice_end - jiffies == 1)
3439 static void cfq_completed_request(struct request_queue *q, struct request *rq)
3441 struct cfq_queue *cfqq = RQ_CFQQ(rq);
3442 struct cfq_data *cfqd = cfqq->cfqd;
3443 const int sync = rq_is_sync(rq);
3447 cfq_log_cfqq(cfqd, cfqq, "complete rqnoidle %d",
3448 !!(rq->cmd_flags & REQ_NOIDLE));
3450 cfq_update_hw_tag(cfqd);
3452 WARN_ON(!cfqd->rq_in_driver);
3453 WARN_ON(!cfqq->dispatched);
3454 cfqd->rq_in_driver--;
3456 (RQ_CFQG(rq))->dispatched--;
3457 cfq_blkiocg_update_completion_stats(&cfqq->cfqg->blkg,
3458 rq_start_time_ns(rq), rq_io_start_time_ns(rq),
3459 rq_data_dir(rq), rq_is_sync(rq));
3461 cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]--;
3464 RQ_CIC(rq)->last_end_request = now;
3465 if (!time_after(rq->start_time + cfqd->cfq_fifo_expire[1], now))
3466 cfqd->last_delayed_sync = now;
3470 * If this is the active queue, check if it needs to be expired,
3471 * or if we want to idle in case it has no pending requests.
3473 if (cfqd->active_queue == cfqq) {
3474 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
3476 if (cfq_cfqq_slice_new(cfqq)) {
3477 cfq_set_prio_slice(cfqd, cfqq);
3478 cfq_clear_cfqq_slice_new(cfqq);
3482 * Should we wait for next request to come in before we expire
3485 if (cfq_should_wait_busy(cfqd, cfqq)) {
3486 unsigned long extend_sl = cfqd->cfq_slice_idle;
3487 if (!cfqd->cfq_slice_idle)
3488 extend_sl = cfqd->cfq_group_idle;
3489 cfqq->slice_end = jiffies + extend_sl;
3490 cfq_mark_cfqq_wait_busy(cfqq);
3491 cfq_log_cfqq(cfqd, cfqq, "will busy wait");
3495 * Idling is not enabled on:
3497 * - idle-priority queues
3499 * - queues with still some requests queued
3500 * - when there is a close cooperator
3502 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
3503 cfq_slice_expired(cfqd, 1);
3504 else if (sync && cfqq_empty &&
3505 !cfq_close_cooperator(cfqd, cfqq)) {
3506 cfq_arm_slice_timer(cfqd);
3510 if (!cfqd->rq_in_driver)
3511 cfq_schedule_dispatch(cfqd);
3515 * we temporarily boost lower priority queues if they are holding fs exclusive
3516 * resources. they are boosted to normal prio (CLASS_BE/4)
3518 static void cfq_prio_boost(struct cfq_queue *cfqq)
3520 if (has_fs_excl()) {
3522 * boost idle prio on transactions that would lock out other
3523 * users of the filesystem
3525 if (cfq_class_idle(cfqq))
3526 cfqq->ioprio_class = IOPRIO_CLASS_BE;
3527 if (cfqq->ioprio > IOPRIO_NORM)
3528 cfqq->ioprio = IOPRIO_NORM;
3531 * unboost the queue (if needed)
3533 cfqq->ioprio_class = cfqq->org_ioprio_class;
3534 cfqq->ioprio = cfqq->org_ioprio;
3538 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
3540 if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
3541 cfq_mark_cfqq_must_alloc_slice(cfqq);
3542 return ELV_MQUEUE_MUST;
3545 return ELV_MQUEUE_MAY;
3548 static int cfq_may_queue(struct request_queue *q, int rw)
3550 struct cfq_data *cfqd = q->elevator->elevator_data;
3551 struct task_struct *tsk = current;
3552 struct cfq_io_context *cic;
3553 struct cfq_queue *cfqq;
3556 * don't force setup of a queue from here, as a call to may_queue
3557 * does not necessarily imply that a request actually will be queued.
3558 * so just lookup a possibly existing queue, or return 'may queue'
3561 cic = cfq_cic_lookup(cfqd, tsk->io_context);
3563 return ELV_MQUEUE_MAY;
3565 cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
3567 cfq_init_prio_data(cfqq, cic->ioc);
3568 cfq_prio_boost(cfqq);
3570 return __cfq_may_queue(cfqq);
3573 return ELV_MQUEUE_MAY;
3577 * queue lock held here
3579 static void cfq_put_request(struct request *rq)
3581 struct cfq_queue *cfqq = RQ_CFQQ(rq);
3584 const int rw = rq_data_dir(rq);
3586 BUG_ON(!cfqq->allocated[rw]);
3587 cfqq->allocated[rw]--;
3589 put_io_context(RQ_CIC(rq)->ioc);
3591 rq->elevator_private = NULL;
3592 rq->elevator_private2 = NULL;
3594 /* Put down rq reference on cfqg */
3595 cfq_put_cfqg(RQ_CFQG(rq));
3596 rq->elevator_private3 = NULL;
3598 cfq_put_queue(cfqq);
3602 static struct cfq_queue *
3603 cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_context *cic,
3604 struct cfq_queue *cfqq)
3606 cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq);
3607 cic_set_cfqq(cic, cfqq->new_cfqq, 1);
3608 cfq_mark_cfqq_coop(cfqq->new_cfqq);
3609 cfq_put_queue(cfqq);
3610 return cic_to_cfqq(cic, 1);
3614 * Returns NULL if a new cfqq should be allocated, or the old cfqq if this
3615 * was the last process referring to said cfqq.
3617 static struct cfq_queue *
3618 split_cfqq(struct cfq_io_context *cic, struct cfq_queue *cfqq)
3620 if (cfqq_process_refs(cfqq) == 1) {
3621 cfqq->pid = current->pid;
3622 cfq_clear_cfqq_coop(cfqq);
3623 cfq_clear_cfqq_split_coop(cfqq);
3627 cic_set_cfqq(cic, NULL, 1);
3629 cfq_put_cooperator(cfqq);
3631 cfq_put_queue(cfqq);
3635 * Allocate cfq data structures associated with this request.
3638 cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
3640 struct cfq_data *cfqd = q->elevator->elevator_data;
3641 struct cfq_io_context *cic;
3642 const int rw = rq_data_dir(rq);
3643 const bool is_sync = rq_is_sync(rq);
3644 struct cfq_queue *cfqq;
3645 unsigned long flags;
3647 might_sleep_if(gfp_mask & __GFP_WAIT);
3649 cic = cfq_get_io_context(cfqd, gfp_mask);
3651 spin_lock_irqsave(q->queue_lock, flags);
3657 cfqq = cic_to_cfqq(cic, is_sync);
3658 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
3659 cfqq = cfq_get_queue(cfqd, is_sync, cic->ioc, gfp_mask);
3660 cic_set_cfqq(cic, cfqq, is_sync);
3663 * If the queue was seeky for too long, break it apart.
3665 if (cfq_cfqq_coop(cfqq) && cfq_cfqq_split_coop(cfqq)) {
3666 cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq");
3667 cfqq = split_cfqq(cic, cfqq);
3673 * Check to see if this queue is scheduled to merge with
3674 * another, closely cooperating queue. The merging of
3675 * queues happens here as it must be done in process context.
3676 * The reference on new_cfqq was taken in merge_cfqqs.
3679 cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq);
3682 cfqq->allocated[rw]++;
3685 spin_unlock_irqrestore(q->queue_lock, flags);
3687 rq->elevator_private = cic;
3688 rq->elevator_private2 = cfqq;
3689 rq->elevator_private3 = cfq_ref_get_cfqg(cfqq->cfqg);
3694 put_io_context(cic->ioc);
3696 cfq_schedule_dispatch(cfqd);
3697 spin_unlock_irqrestore(q->queue_lock, flags);
3698 cfq_log(cfqd, "set_request fail");
3702 static void cfq_kick_queue(struct work_struct *work)
3704 struct cfq_data *cfqd =
3705 container_of(work, struct cfq_data, unplug_work);
3706 struct request_queue *q = cfqd->queue;
3708 spin_lock_irq(q->queue_lock);
3709 __blk_run_queue(cfqd->queue);
3710 spin_unlock_irq(q->queue_lock);
3714 * Timer running if the active_queue is currently idling inside its time slice
3716 static void cfq_idle_slice_timer(unsigned long data)
3718 struct cfq_data *cfqd = (struct cfq_data *) data;
3719 struct cfq_queue *cfqq;
3720 unsigned long flags;
3723 cfq_log(cfqd, "idle timer fired");
3725 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
3727 cfqq = cfqd->active_queue;
3732 * We saw a request before the queue expired, let it through
3734 if (cfq_cfqq_must_dispatch(cfqq))
3740 if (cfq_slice_used(cfqq))
3744 * only expire and reinvoke request handler, if there are
3745 * other queues with pending requests
3747 if (!cfqd->busy_queues)
3751 * not expired and it has a request pending, let it dispatch
3753 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3757 * Queue depth flag is reset only when the idle didn't succeed
3759 cfq_clear_cfqq_deep(cfqq);
3762 cfq_slice_expired(cfqd, timed_out);
3764 cfq_schedule_dispatch(cfqd);
3766 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
3769 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
3771 del_timer_sync(&cfqd->idle_slice_timer);
3772 cancel_work_sync(&cfqd->unplug_work);
3775 static void cfq_put_async_queues(struct cfq_data *cfqd)
3779 for (i = 0; i < IOPRIO_BE_NR; i++) {
3780 if (cfqd->async_cfqq[0][i])
3781 cfq_put_queue(cfqd->async_cfqq[0][i]);
3782 if (cfqd->async_cfqq[1][i])
3783 cfq_put_queue(cfqd->async_cfqq[1][i]);
3786 if (cfqd->async_idle_cfqq)
3787 cfq_put_queue(cfqd->async_idle_cfqq);
3790 static void cfq_cfqd_free(struct rcu_head *head)
3792 kfree(container_of(head, struct cfq_data, rcu));
3795 static void cfq_exit_queue(struct elevator_queue *e)
3797 struct cfq_data *cfqd = e->elevator_data;
3798 struct request_queue *q = cfqd->queue;
3800 cfq_shutdown_timer_wq(cfqd);
3802 spin_lock_irq(q->queue_lock);
3804 if (cfqd->active_queue)
3805 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
3807 while (!list_empty(&cfqd->cic_list)) {
3808 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
3809 struct cfq_io_context,
3812 __cfq_exit_single_io_context(cfqd, cic);
3815 cfq_put_async_queues(cfqd);
3816 cfq_release_cfq_groups(cfqd);
3817 cfq_blkiocg_del_blkio_group(&cfqd->root_group.blkg);
3819 spin_unlock_irq(q->queue_lock);
3821 cfq_shutdown_timer_wq(cfqd);
3823 spin_lock(&cic_index_lock);
3824 ida_remove(&cic_index_ida, cfqd->cic_index);
3825 spin_unlock(&cic_index_lock);
3827 /* Wait for cfqg->blkg->key accessors to exit their grace periods. */
3828 call_rcu(&cfqd->rcu, cfq_cfqd_free);
3831 static int cfq_alloc_cic_index(void)
3836 if (!ida_pre_get(&cic_index_ida, GFP_KERNEL))
3839 spin_lock(&cic_index_lock);
3840 error = ida_get_new(&cic_index_ida, &index);
3841 spin_unlock(&cic_index_lock);
3842 if (error && error != -EAGAIN)
3849 static void *cfq_init_queue(struct request_queue *q)
3851 struct cfq_data *cfqd;
3853 struct cfq_group *cfqg;
3854 struct cfq_rb_root *st;
3856 i = cfq_alloc_cic_index();
3860 cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
3865 * Don't need take queue_lock in the routine, since we are
3866 * initializing the ioscheduler, and nobody is using cfqd
3868 cfqd->cic_index = i;
3870 /* Init root service tree */
3871 cfqd->grp_service_tree = CFQ_RB_ROOT;
3873 /* Init root group */
3874 cfqg = &cfqd->root_group;
3875 for_each_cfqg_st(cfqg, i, j, st)
3877 RB_CLEAR_NODE(&cfqg->rb_node);
3879 /* Give preference to root group over other groups */
3880 cfqg->weight = 2*BLKIO_WEIGHT_DEFAULT;
3882 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3884 * Take a reference to root group which we never drop. This is just
3885 * to make sure that cfq_put_cfqg() does not try to kfree root group
3887 atomic_set(&cfqg->ref, 1);
3889 cfq_blkiocg_add_blkio_group(&blkio_root_cgroup, &cfqg->blkg,
3894 * Not strictly needed (since RB_ROOT just clears the node and we
3895 * zeroed cfqd on alloc), but better be safe in case someone decides
3896 * to add magic to the rb code
3898 for (i = 0; i < CFQ_PRIO_LISTS; i++)
3899 cfqd->prio_trees[i] = RB_ROOT;
3902 * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
3903 * Grab a permanent reference to it, so that the normal code flow
3904 * will not attempt to free it.
3906 cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
3907 cfqd->oom_cfqq.ref++;
3908 cfq_link_cfqq_cfqg(&cfqd->oom_cfqq, &cfqd->root_group);
3910 INIT_LIST_HEAD(&cfqd->cic_list);
3914 init_timer(&cfqd->idle_slice_timer);
3915 cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
3916 cfqd->idle_slice_timer.data = (unsigned long) cfqd;
3918 INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
3920 cfqd->cfq_quantum = cfq_quantum;
3921 cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
3922 cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
3923 cfqd->cfq_back_max = cfq_back_max;
3924 cfqd->cfq_back_penalty = cfq_back_penalty;
3925 cfqd->cfq_slice[0] = cfq_slice_async;
3926 cfqd->cfq_slice[1] = cfq_slice_sync;
3927 cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
3928 cfqd->cfq_slice_idle = cfq_slice_idle;
3929 cfqd->cfq_group_idle = cfq_group_idle;
3930 cfqd->cfq_latency = 1;
3931 cfqd->cfq_group_isolation = 0;
3934 * we optimistically start assuming sync ops weren't delayed in last
3935 * second, in order to have larger depth for async operations.
3937 cfqd->last_delayed_sync = jiffies - HZ;
3941 static void cfq_slab_kill(void)
3944 * Caller already ensured that pending RCU callbacks are completed,
3945 * so we should have no busy allocations at this point.
3948 kmem_cache_destroy(cfq_pool);
3950 kmem_cache_destroy(cfq_ioc_pool);
3953 static int __init cfq_slab_setup(void)
3955 cfq_pool = KMEM_CACHE(cfq_queue, 0);
3959 cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0);
3970 * sysfs parts below -->
3973 cfq_var_show(unsigned int var, char *page)
3975 return sprintf(page, "%d\n", var);
3979 cfq_var_store(unsigned int *var, const char *page, size_t count)
3981 char *p = (char *) page;
3983 *var = simple_strtoul(p, &p, 10);
3987 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
3988 static ssize_t __FUNC(struct elevator_queue *e, char *page) \
3990 struct cfq_data *cfqd = e->elevator_data; \
3991 unsigned int __data = __VAR; \
3993 __data = jiffies_to_msecs(__data); \
3994 return cfq_var_show(__data, (page)); \
3996 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
3997 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
3998 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
3999 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
4000 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
4001 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
4002 SHOW_FUNCTION(cfq_group_idle_show, cfqd->cfq_group_idle, 1);
4003 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
4004 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
4005 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
4006 SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
4007 SHOW_FUNCTION(cfq_group_isolation_show, cfqd->cfq_group_isolation, 0);
4008 #undef SHOW_FUNCTION
4010 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
4011 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
4013 struct cfq_data *cfqd = e->elevator_data; \
4014 unsigned int __data; \
4015 int ret = cfq_var_store(&__data, (page), count); \
4016 if (__data < (MIN)) \
4018 else if (__data > (MAX)) \
4021 *(__PTR) = msecs_to_jiffies(__data); \
4023 *(__PTR) = __data; \
4026 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
4027 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
4029 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
4031 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
4032 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
4034 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
4035 STORE_FUNCTION(cfq_group_idle_store, &cfqd->cfq_group_idle, 0, UINT_MAX, 1);
4036 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
4037 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
4038 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
4040 STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
4041 STORE_FUNCTION(cfq_group_isolation_store, &cfqd->cfq_group_isolation, 0, 1, 0);
4042 #undef STORE_FUNCTION
4044 #define CFQ_ATTR(name) \
4045 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
4047 static struct elv_fs_entry cfq_attrs[] = {
4049 CFQ_ATTR(fifo_expire_sync),
4050 CFQ_ATTR(fifo_expire_async),
4051 CFQ_ATTR(back_seek_max),
4052 CFQ_ATTR(back_seek_penalty),
4053 CFQ_ATTR(slice_sync),
4054 CFQ_ATTR(slice_async),
4055 CFQ_ATTR(slice_async_rq),
4056 CFQ_ATTR(slice_idle),
4057 CFQ_ATTR(group_idle),
4058 CFQ_ATTR(low_latency),
4059 CFQ_ATTR(group_isolation),
4063 static struct elevator_type iosched_cfq = {
4065 .elevator_merge_fn = cfq_merge,
4066 .elevator_merged_fn = cfq_merged_request,
4067 .elevator_merge_req_fn = cfq_merged_requests,
4068 .elevator_allow_merge_fn = cfq_allow_merge,
4069 .elevator_bio_merged_fn = cfq_bio_merged,
4070 .elevator_dispatch_fn = cfq_dispatch_requests,
4071 .elevator_add_req_fn = cfq_insert_request,
4072 .elevator_activate_req_fn = cfq_activate_request,
4073 .elevator_deactivate_req_fn = cfq_deactivate_request,
4074 .elevator_queue_empty_fn = cfq_queue_empty,
4075 .elevator_completed_req_fn = cfq_completed_request,
4076 .elevator_former_req_fn = elv_rb_former_request,
4077 .elevator_latter_req_fn = elv_rb_latter_request,
4078 .elevator_set_req_fn = cfq_set_request,
4079 .elevator_put_req_fn = cfq_put_request,
4080 .elevator_may_queue_fn = cfq_may_queue,
4081 .elevator_init_fn = cfq_init_queue,
4082 .elevator_exit_fn = cfq_exit_queue,
4083 .trim = cfq_free_io_context,
4085 .elevator_attrs = cfq_attrs,
4086 .elevator_name = "cfq",
4087 .elevator_owner = THIS_MODULE,
4090 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4091 static struct blkio_policy_type blkio_policy_cfq = {
4093 .blkio_unlink_group_fn = cfq_unlink_blkio_group,
4094 .blkio_update_group_weight_fn = cfq_update_blkio_group_weight,
4096 .plid = BLKIO_POLICY_PROP,
4099 static struct blkio_policy_type blkio_policy_cfq;
4102 static int __init cfq_init(void)
4105 * could be 0 on HZ < 1000 setups
4107 if (!cfq_slice_async)
4108 cfq_slice_async = 1;
4109 if (!cfq_slice_idle)
4112 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4113 if (!cfq_group_idle)
4118 if (cfq_slab_setup())
4121 elv_register(&iosched_cfq);
4122 blkio_policy_register(&blkio_policy_cfq);
4127 static void __exit cfq_exit(void)
4129 DECLARE_COMPLETION_ONSTACK(all_gone);
4130 blkio_policy_unregister(&blkio_policy_cfq);
4131 elv_unregister(&iosched_cfq);
4132 ioc_gone = &all_gone;
4133 /* ioc_gone's update must be visible before reading ioc_count */
4137 * this also protects us from entering cfq_slab_kill() with
4138 * pending RCU callbacks
4140 if (elv_ioc_count_read(cfq_ioc_count))
4141 wait_for_completion(&all_gone);
4142 ida_destroy(&cic_index_ida);
4146 module_init(cfq_init);
4147 module_exit(cfq_exit);
4149 MODULE_AUTHOR("Jens Axboe");
4150 MODULE_LICENSE("GPL");
4151 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");