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;
90 struct rb_node *active;
92 #define CFQ_RB_ROOT (struct cfq_rb_root) { .rb = RB_ROOT, .left = NULL, \
93 .count = 0, .min_vdisktime = 0, }
96 * Per process-grouping structure
101 /* various state flags, see below */
103 /* parent cfq_data */
104 struct cfq_data *cfqd;
105 /* service_tree member */
106 struct rb_node rb_node;
107 /* service_tree key */
108 unsigned long rb_key;
109 /* prio tree member */
110 struct rb_node p_node;
111 /* prio tree root we belong to, if any */
112 struct rb_root *p_root;
113 /* sorted list of pending requests */
114 struct rb_root sort_list;
115 /* if fifo isn't expired, next request to serve */
116 struct request *next_rq;
117 /* requests queued in sort_list */
119 /* currently allocated requests */
121 /* fifo list of requests in sort_list */
122 struct list_head fifo;
124 /* time when queue got scheduled in to dispatch first request. */
125 unsigned long dispatch_start;
126 unsigned int allocated_slice;
127 unsigned int slice_dispatch;
128 /* time when first request from queue completed and slice started. */
129 unsigned long slice_start;
130 unsigned long slice_end;
133 /* pending metadata requests */
135 /* number of requests that are on the dispatch list or inside driver */
138 /* io prio of this group */
139 unsigned short ioprio, org_ioprio;
140 unsigned short ioprio_class, org_ioprio_class;
145 sector_t last_request_pos;
147 struct cfq_rb_root *service_tree;
148 struct cfq_queue *new_cfqq;
149 struct cfq_group *cfqg;
150 struct cfq_group *orig_cfqg;
154 * First index in the service_trees.
155 * IDLE is handled separately, so it has negative index
164 * Second index in the service_trees.
168 SYNC_NOIDLE_WORKLOAD = 1,
172 /* This is per cgroup per device grouping structure */
174 /* group service_tree member */
175 struct rb_node rb_node;
177 /* group service_tree key */
182 /* number of cfqq currently on this group */
185 /* Per group busy queus average. Useful for workload slice calc. */
186 unsigned int busy_queues_avg[2];
188 * rr lists of queues with requests, onle rr for each priority class.
189 * Counts are embedded in the cfq_rb_root
191 struct cfq_rb_root service_trees[2][3];
192 struct cfq_rb_root service_tree_idle;
194 unsigned long saved_workload_slice;
195 enum wl_type_t saved_workload;
196 enum wl_prio_t saved_serving_prio;
197 struct blkio_group blkg;
198 #ifdef CONFIG_CFQ_GROUP_IOSCHED
199 struct hlist_node cfqd_node;
202 /* number of requests that are on the dispatch list or inside driver */
207 * Per block device queue structure
210 struct request_queue *queue;
211 /* Root service tree for cfq_groups */
212 struct cfq_rb_root grp_service_tree;
213 struct cfq_group root_group;
216 * The priority currently being served
218 enum wl_prio_t serving_prio;
219 enum wl_type_t serving_type;
220 unsigned long workload_expires;
221 struct cfq_group *serving_group;
222 bool noidle_tree_requires_idle;
225 * Each priority tree is sorted by next_request position. These
226 * trees are used when determining if two or more queues are
227 * interleaving requests (see cfq_close_cooperator).
229 struct rb_root prio_trees[CFQ_PRIO_LISTS];
231 unsigned int busy_queues;
237 * queue-depth detection
243 * -1 => indeterminate, (cfq will behave as if NCQ is present, to allow better detection)
244 * 1 => NCQ is present (hw_tag_est_depth is the estimated max depth)
247 int hw_tag_est_depth;
248 unsigned int hw_tag_samples;
251 * idle window management
253 struct timer_list idle_slice_timer;
254 struct work_struct unplug_work;
256 struct cfq_queue *active_queue;
257 struct cfq_io_context *active_cic;
260 * async queue for each priority case
262 struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
263 struct cfq_queue *async_idle_cfqq;
265 sector_t last_position;
268 * tunables, see top of file
270 unsigned int cfq_quantum;
271 unsigned int cfq_fifo_expire[2];
272 unsigned int cfq_back_penalty;
273 unsigned int cfq_back_max;
274 unsigned int cfq_slice[2];
275 unsigned int cfq_slice_async_rq;
276 unsigned int cfq_slice_idle;
277 unsigned int cfq_group_idle;
278 unsigned int cfq_latency;
279 unsigned int cfq_group_isolation;
281 unsigned int cic_index;
282 struct list_head cic_list;
285 * Fallback dummy cfqq for extreme OOM conditions
287 struct cfq_queue oom_cfqq;
289 unsigned long last_delayed_sync;
291 /* List of cfq groups being managed on this device*/
292 struct hlist_head cfqg_list;
296 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd);
298 static struct cfq_rb_root *service_tree_for(struct cfq_group *cfqg,
305 if (prio == IDLE_WORKLOAD)
306 return &cfqg->service_tree_idle;
308 return &cfqg->service_trees[prio][type];
311 enum cfqq_state_flags {
312 CFQ_CFQQ_FLAG_on_rr = 0, /* on round-robin busy list */
313 CFQ_CFQQ_FLAG_wait_request, /* waiting for a request */
314 CFQ_CFQQ_FLAG_must_dispatch, /* must be allowed a dispatch */
315 CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
316 CFQ_CFQQ_FLAG_fifo_expire, /* FIFO checked in this slice */
317 CFQ_CFQQ_FLAG_idle_window, /* slice idling enabled */
318 CFQ_CFQQ_FLAG_prio_changed, /* task priority has changed */
319 CFQ_CFQQ_FLAG_slice_new, /* no requests dispatched in slice */
320 CFQ_CFQQ_FLAG_sync, /* synchronous queue */
321 CFQ_CFQQ_FLAG_coop, /* cfqq is shared */
322 CFQ_CFQQ_FLAG_split_coop, /* shared cfqq will be splitted */
323 CFQ_CFQQ_FLAG_deep, /* sync cfqq experienced large depth */
324 CFQ_CFQQ_FLAG_wait_busy, /* Waiting for next request */
327 #define CFQ_CFQQ_FNS(name) \
328 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \
330 (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name); \
332 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \
334 (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \
336 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \
338 return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \
342 CFQ_CFQQ_FNS(wait_request);
343 CFQ_CFQQ_FNS(must_dispatch);
344 CFQ_CFQQ_FNS(must_alloc_slice);
345 CFQ_CFQQ_FNS(fifo_expire);
346 CFQ_CFQQ_FNS(idle_window);
347 CFQ_CFQQ_FNS(prio_changed);
348 CFQ_CFQQ_FNS(slice_new);
351 CFQ_CFQQ_FNS(split_coop);
353 CFQ_CFQQ_FNS(wait_busy);
356 #ifdef CONFIG_CFQ_GROUP_IOSCHED
357 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
358 blk_add_trace_msg((cfqd)->queue, "cfq%d%c %s " fmt, (cfqq)->pid, \
359 cfq_cfqq_sync((cfqq)) ? 'S' : 'A', \
360 blkg_path(&(cfqq)->cfqg->blkg), ##args);
362 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...) \
363 blk_add_trace_msg((cfqd)->queue, "%s " fmt, \
364 blkg_path(&(cfqg)->blkg), ##args); \
367 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
368 blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args)
369 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...) do {} while (0);
371 #define cfq_log(cfqd, fmt, args...) \
372 blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
374 /* Traverses through cfq group service trees */
375 #define for_each_cfqg_st(cfqg, i, j, st) \
376 for (i = 0; i <= IDLE_WORKLOAD; i++) \
377 for (j = 0, st = i < IDLE_WORKLOAD ? &cfqg->service_trees[i][j]\
378 : &cfqg->service_tree_idle; \
379 (i < IDLE_WORKLOAD && j <= SYNC_WORKLOAD) || \
380 (i == IDLE_WORKLOAD && j == 0); \
381 j++, st = i < IDLE_WORKLOAD ? \
382 &cfqg->service_trees[i][j]: NULL) \
385 static inline bool iops_mode(struct cfq_data *cfqd)
388 * If we are not idling on queues and it is a NCQ drive, parallel
389 * execution of requests is on and measuring time is not possible
390 * in most of the cases until and unless we drive shallower queue
391 * depths and that becomes a performance bottleneck. In such cases
392 * switch to start providing fairness in terms of number of IOs.
394 if (!cfqd->cfq_slice_idle && cfqd->hw_tag)
400 static inline enum wl_prio_t cfqq_prio(struct cfq_queue *cfqq)
402 if (cfq_class_idle(cfqq))
403 return IDLE_WORKLOAD;
404 if (cfq_class_rt(cfqq))
410 static enum wl_type_t cfqq_type(struct cfq_queue *cfqq)
412 if (!cfq_cfqq_sync(cfqq))
413 return ASYNC_WORKLOAD;
414 if (!cfq_cfqq_idle_window(cfqq))
415 return SYNC_NOIDLE_WORKLOAD;
416 return SYNC_WORKLOAD;
419 static inline int cfq_group_busy_queues_wl(enum wl_prio_t wl,
420 struct cfq_data *cfqd,
421 struct cfq_group *cfqg)
423 if (wl == IDLE_WORKLOAD)
424 return cfqg->service_tree_idle.count;
426 return cfqg->service_trees[wl][ASYNC_WORKLOAD].count
427 + cfqg->service_trees[wl][SYNC_NOIDLE_WORKLOAD].count
428 + cfqg->service_trees[wl][SYNC_WORKLOAD].count;
431 static inline int cfqg_busy_async_queues(struct cfq_data *cfqd,
432 struct cfq_group *cfqg)
434 return cfqg->service_trees[RT_WORKLOAD][ASYNC_WORKLOAD].count
435 + cfqg->service_trees[BE_WORKLOAD][ASYNC_WORKLOAD].count;
438 static void cfq_dispatch_insert(struct request_queue *, struct request *);
439 static struct cfq_queue *cfq_get_queue(struct cfq_data *, bool,
440 struct io_context *, gfp_t);
441 static struct cfq_io_context *cfq_cic_lookup(struct cfq_data *,
442 struct io_context *);
444 static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_context *cic,
447 return cic->cfqq[is_sync];
450 static inline void cic_set_cfqq(struct cfq_io_context *cic,
451 struct cfq_queue *cfqq, bool is_sync)
453 cic->cfqq[is_sync] = cfqq;
456 #define CIC_DEAD_KEY 1ul
457 #define CIC_DEAD_INDEX_SHIFT 1
459 static inline void *cfqd_dead_key(struct cfq_data *cfqd)
461 return (void *)(cfqd->cic_index << CIC_DEAD_INDEX_SHIFT | CIC_DEAD_KEY);
464 static inline struct cfq_data *cic_to_cfqd(struct cfq_io_context *cic)
466 struct cfq_data *cfqd = cic->key;
468 if (unlikely((unsigned long) cfqd & CIC_DEAD_KEY))
475 * We regard a request as SYNC, if it's either a read or has the SYNC bit
476 * set (in which case it could also be direct WRITE).
478 static inline bool cfq_bio_sync(struct bio *bio)
480 return bio_data_dir(bio) == READ || (bio->bi_rw & REQ_SYNC);
484 * scheduler run of queue, if there are requests pending and no one in the
485 * driver that will restart queueing
487 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
489 if (cfqd->busy_queues) {
490 cfq_log(cfqd, "schedule dispatch");
491 kblockd_schedule_work(cfqd->queue, &cfqd->unplug_work);
495 static int cfq_queue_empty(struct request_queue *q)
497 struct cfq_data *cfqd = q->elevator->elevator_data;
499 return !cfqd->rq_queued;
503 * Scale schedule slice based on io priority. Use the sync time slice only
504 * if a queue is marked sync and has sync io queued. A sync queue with async
505 * io only, should not get full sync slice length.
507 static inline int cfq_prio_slice(struct cfq_data *cfqd, bool sync,
510 const int base_slice = cfqd->cfq_slice[sync];
512 WARN_ON(prio >= IOPRIO_BE_NR);
514 return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
518 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
520 return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
523 static inline u64 cfq_scale_slice(unsigned long delta, struct cfq_group *cfqg)
525 u64 d = delta << CFQ_SERVICE_SHIFT;
527 d = d * BLKIO_WEIGHT_DEFAULT;
528 do_div(d, cfqg->weight);
532 static inline u64 max_vdisktime(u64 min_vdisktime, u64 vdisktime)
534 s64 delta = (s64)(vdisktime - min_vdisktime);
536 min_vdisktime = vdisktime;
538 return min_vdisktime;
541 static inline u64 min_vdisktime(u64 min_vdisktime, u64 vdisktime)
543 s64 delta = (s64)(vdisktime - min_vdisktime);
545 min_vdisktime = vdisktime;
547 return min_vdisktime;
550 static void update_min_vdisktime(struct cfq_rb_root *st)
552 u64 vdisktime = st->min_vdisktime;
553 struct cfq_group *cfqg;
556 cfqg = rb_entry_cfqg(st->active);
557 vdisktime = cfqg->vdisktime;
561 cfqg = rb_entry_cfqg(st->left);
562 vdisktime = min_vdisktime(vdisktime, cfqg->vdisktime);
565 st->min_vdisktime = max_vdisktime(st->min_vdisktime, vdisktime);
569 * get averaged number of queues of RT/BE priority.
570 * average is updated, with a formula that gives more weight to higher numbers,
571 * to quickly follows sudden increases and decrease slowly
574 static inline unsigned cfq_group_get_avg_queues(struct cfq_data *cfqd,
575 struct cfq_group *cfqg, bool rt)
577 unsigned min_q, max_q;
578 unsigned mult = cfq_hist_divisor - 1;
579 unsigned round = cfq_hist_divisor / 2;
580 unsigned busy = cfq_group_busy_queues_wl(rt, cfqd, cfqg);
582 min_q = min(cfqg->busy_queues_avg[rt], busy);
583 max_q = max(cfqg->busy_queues_avg[rt], busy);
584 cfqg->busy_queues_avg[rt] = (mult * max_q + min_q + round) /
586 return cfqg->busy_queues_avg[rt];
589 static inline unsigned
590 cfq_group_slice(struct cfq_data *cfqd, struct cfq_group *cfqg)
592 struct cfq_rb_root *st = &cfqd->grp_service_tree;
594 return cfq_target_latency * cfqg->weight / st->total_weight;
598 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
600 unsigned slice = cfq_prio_to_slice(cfqd, cfqq);
601 if (cfqd->cfq_latency) {
603 * interested queues (we consider only the ones with the same
604 * priority class in the cfq group)
606 unsigned iq = cfq_group_get_avg_queues(cfqd, cfqq->cfqg,
608 unsigned sync_slice = cfqd->cfq_slice[1];
609 unsigned expect_latency = sync_slice * iq;
610 unsigned group_slice = cfq_group_slice(cfqd, cfqq->cfqg);
612 if (expect_latency > group_slice) {
613 unsigned base_low_slice = 2 * cfqd->cfq_slice_idle;
614 /* scale low_slice according to IO priority
615 * and sync vs async */
617 min(slice, base_low_slice * slice / sync_slice);
618 /* the adapted slice value is scaled to fit all iqs
619 * into the target latency */
620 slice = max(slice * group_slice / expect_latency,
624 cfqq->slice_start = jiffies;
625 cfqq->slice_end = jiffies + slice;
626 cfqq->allocated_slice = slice;
627 cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
631 * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
632 * isn't valid until the first request from the dispatch is activated
633 * and the slice time set.
635 static inline bool cfq_slice_used(struct cfq_queue *cfqq)
637 if (cfq_cfqq_slice_new(cfqq))
639 if (time_before(jiffies, cfqq->slice_end))
646 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
647 * We choose the request that is closest to the head right now. Distance
648 * behind the head is penalized and only allowed to a certain extent.
650 static struct request *
651 cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2, sector_t last)
653 sector_t s1, s2, d1 = 0, d2 = 0;
654 unsigned long back_max;
655 #define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */
656 #define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */
657 unsigned wrap = 0; /* bit mask: requests behind the disk head? */
659 if (rq1 == NULL || rq1 == rq2)
664 if (rq_is_sync(rq1) && !rq_is_sync(rq2))
666 else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
668 if ((rq1->cmd_flags & REQ_META) && !(rq2->cmd_flags & REQ_META))
670 else if ((rq2->cmd_flags & REQ_META) &&
671 !(rq1->cmd_flags & REQ_META))
674 s1 = blk_rq_pos(rq1);
675 s2 = blk_rq_pos(rq2);
678 * by definition, 1KiB is 2 sectors
680 back_max = cfqd->cfq_back_max * 2;
683 * Strict one way elevator _except_ in the case where we allow
684 * short backward seeks which are biased as twice the cost of a
685 * similar forward seek.
689 else if (s1 + back_max >= last)
690 d1 = (last - s1) * cfqd->cfq_back_penalty;
692 wrap |= CFQ_RQ1_WRAP;
696 else if (s2 + back_max >= last)
697 d2 = (last - s2) * cfqd->cfq_back_penalty;
699 wrap |= CFQ_RQ2_WRAP;
701 /* Found required data */
704 * By doing switch() on the bit mask "wrap" we avoid having to
705 * check two variables for all permutations: --> faster!
708 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
724 case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
727 * Since both rqs are wrapped,
728 * start with the one that's further behind head
729 * (--> only *one* back seek required),
730 * since back seek takes more time than forward.
740 * The below is leftmost cache rbtree addon
742 static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
744 /* Service tree is empty */
749 root->left = rb_first(&root->rb);
752 return rb_entry(root->left, struct cfq_queue, rb_node);
757 static struct cfq_group *cfq_rb_first_group(struct cfq_rb_root *root)
760 root->left = rb_first(&root->rb);
763 return rb_entry_cfqg(root->left);
768 static void rb_erase_init(struct rb_node *n, struct rb_root *root)
774 static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
778 rb_erase_init(n, &root->rb);
783 * would be nice to take fifo expire time into account as well
785 static struct request *
786 cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
787 struct request *last)
789 struct rb_node *rbnext = rb_next(&last->rb_node);
790 struct rb_node *rbprev = rb_prev(&last->rb_node);
791 struct request *next = NULL, *prev = NULL;
793 BUG_ON(RB_EMPTY_NODE(&last->rb_node));
796 prev = rb_entry_rq(rbprev);
799 next = rb_entry_rq(rbnext);
801 rbnext = rb_first(&cfqq->sort_list);
802 if (rbnext && rbnext != &last->rb_node)
803 next = rb_entry_rq(rbnext);
806 return cfq_choose_req(cfqd, next, prev, blk_rq_pos(last));
809 static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
810 struct cfq_queue *cfqq)
813 * just an approximation, should be ok.
815 return (cfqq->cfqg->nr_cfqq - 1) * (cfq_prio_slice(cfqd, 1, 0) -
816 cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
820 cfqg_key(struct cfq_rb_root *st, struct cfq_group *cfqg)
822 return cfqg->vdisktime - st->min_vdisktime;
826 __cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
828 struct rb_node **node = &st->rb.rb_node;
829 struct rb_node *parent = NULL;
830 struct cfq_group *__cfqg;
831 s64 key = cfqg_key(st, cfqg);
834 while (*node != NULL) {
836 __cfqg = rb_entry_cfqg(parent);
838 if (key < cfqg_key(st, __cfqg))
839 node = &parent->rb_left;
841 node = &parent->rb_right;
847 st->left = &cfqg->rb_node;
849 rb_link_node(&cfqg->rb_node, parent, node);
850 rb_insert_color(&cfqg->rb_node, &st->rb);
854 cfq_group_service_tree_add(struct cfq_data *cfqd, struct cfq_group *cfqg)
856 struct cfq_rb_root *st = &cfqd->grp_service_tree;
857 struct cfq_group *__cfqg;
865 * Currently put the group at the end. Later implement something
866 * so that groups get lesser vtime based on their weights, so that
867 * if group does not loose all if it was not continously backlogged.
869 n = rb_last(&st->rb);
871 __cfqg = rb_entry_cfqg(n);
872 cfqg->vdisktime = __cfqg->vdisktime + CFQ_IDLE_DELAY;
874 cfqg->vdisktime = st->min_vdisktime;
876 __cfq_group_service_tree_add(st, cfqg);
878 st->total_weight += cfqg->weight;
882 cfq_group_service_tree_del(struct cfq_data *cfqd, struct cfq_group *cfqg)
884 struct cfq_rb_root *st = &cfqd->grp_service_tree;
886 if (st->active == &cfqg->rb_node)
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");
898 st->total_weight -= cfqg->weight;
899 if (!RB_EMPTY_NODE(&cfqg->rb_node))
900 cfq_rb_erase(&cfqg->rb_node, st);
901 cfqg->saved_workload_slice = 0;
902 cfq_blkiocg_update_dequeue_stats(&cfqg->blkg, 1);
905 static inline unsigned int cfq_cfqq_slice_usage(struct cfq_queue *cfqq)
907 unsigned int slice_used;
910 * Queue got expired before even a single request completed or
911 * got expired immediately after first request completion.
913 if (!cfqq->slice_start || cfqq->slice_start == jiffies) {
915 * Also charge the seek time incurred to the group, otherwise
916 * if there are mutiple queues in the group, each can dispatch
917 * a single request on seeky media and cause lots of seek time
918 * and group will never know it.
920 slice_used = max_t(unsigned, (jiffies - cfqq->dispatch_start),
923 slice_used = jiffies - cfqq->slice_start;
924 if (slice_used > cfqq->allocated_slice)
925 slice_used = cfqq->allocated_slice;
931 static void cfq_group_served(struct cfq_data *cfqd, struct cfq_group *cfqg,
932 struct cfq_queue *cfqq)
934 struct cfq_rb_root *st = &cfqd->grp_service_tree;
935 unsigned int used_sl, charge;
936 int nr_sync = cfqg->nr_cfqq - cfqg_busy_async_queues(cfqd, cfqg)
937 - cfqg->service_tree_idle.count;
940 used_sl = charge = cfq_cfqq_slice_usage(cfqq);
943 charge = cfqq->slice_dispatch;
944 else if (!cfq_cfqq_sync(cfqq) && !nr_sync)
945 charge = cfqq->allocated_slice;
947 /* Can't update vdisktime while group is on service tree */
948 cfq_rb_erase(&cfqg->rb_node, st);
949 cfqg->vdisktime += cfq_scale_slice(charge, cfqg);
950 __cfq_group_service_tree_add(st, cfqg);
952 /* This group is being expired. Save the context */
953 if (time_after(cfqd->workload_expires, jiffies)) {
954 cfqg->saved_workload_slice = cfqd->workload_expires
956 cfqg->saved_workload = cfqd->serving_type;
957 cfqg->saved_serving_prio = cfqd->serving_prio;
959 cfqg->saved_workload_slice = 0;
961 cfq_log_cfqg(cfqd, cfqg, "served: vt=%llu min_vt=%llu", cfqg->vdisktime,
963 cfq_log_cfqq(cfqq->cfqd, cfqq, "sl_used=%u disp=%u charge=%u iops=%u",
964 used_sl, cfqq->slice_dispatch, charge, iops_mode(cfqd));
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);
978 cfq_update_blkio_group_weight(struct blkio_group *blkg, unsigned int weight)
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);
1019 /* Add group onto cgroup list */
1020 sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
1021 cfq_blkiocg_add_blkio_group(blkcg, &cfqg->blkg, (void *)cfqd,
1022 MKDEV(major, minor));
1023 cfqg->weight = blkcg_get_weight(blkcg, cfqg->blkg.dev);
1025 /* Add group on cfqd list */
1026 hlist_add_head(&cfqg->cfqd_node, &cfqd->cfqg_list);
1033 * Search for the cfq group current task belongs to. If create = 1, then also
1034 * create the cfq group if it does not exist. request_queue lock must be held.
1036 static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd, int create)
1038 struct cgroup *cgroup;
1039 struct cfq_group *cfqg = NULL;
1042 cgroup = task_cgroup(current, blkio_subsys_id);
1043 cfqg = cfq_find_alloc_cfqg(cfqd, cgroup, create);
1044 if (!cfqg && create)
1045 cfqg = &cfqd->root_group;
1050 static inline struct cfq_group *cfq_ref_get_cfqg(struct cfq_group *cfqg)
1052 atomic_inc(&cfqg->ref);
1056 static void cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg)
1058 /* Currently, all async queues are mapped to root group */
1059 if (!cfq_cfqq_sync(cfqq))
1060 cfqg = &cfqq->cfqd->root_group;
1063 /* cfqq reference on cfqg */
1064 atomic_inc(&cfqq->cfqg->ref);
1067 static void cfq_put_cfqg(struct cfq_group *cfqg)
1069 struct cfq_rb_root *st;
1072 BUG_ON(atomic_read(&cfqg->ref) <= 0);
1073 if (!atomic_dec_and_test(&cfqg->ref))
1075 for_each_cfqg_st(cfqg, i, j, st)
1076 BUG_ON(!RB_EMPTY_ROOT(&st->rb) || st->active != NULL);
1080 static void cfq_destroy_cfqg(struct cfq_data *cfqd, struct cfq_group *cfqg)
1082 /* Something wrong if we are trying to remove same group twice */
1083 BUG_ON(hlist_unhashed(&cfqg->cfqd_node));
1085 hlist_del_init(&cfqg->cfqd_node);
1088 * Put the reference taken at the time of creation so that when all
1089 * queues are gone, group can be destroyed.
1094 static void cfq_release_cfq_groups(struct cfq_data *cfqd)
1096 struct hlist_node *pos, *n;
1097 struct cfq_group *cfqg;
1099 hlist_for_each_entry_safe(cfqg, pos, n, &cfqd->cfqg_list, cfqd_node) {
1101 * If cgroup removal path got to blk_group first and removed
1102 * it from cgroup list, then it will take care of destroying
1105 if (!cfq_blkiocg_del_blkio_group(&cfqg->blkg))
1106 cfq_destroy_cfqg(cfqd, cfqg);
1111 * Blk cgroup controller notification saying that blkio_group object is being
1112 * delinked as associated cgroup object is going away. That also means that
1113 * no new IO will come in this group. So get rid of this group as soon as
1114 * any pending IO in the group is finished.
1116 * This function is called under rcu_read_lock(). key is the rcu protected
1117 * pointer. That means "key" is a valid cfq_data pointer as long as we are rcu
1120 * "key" was fetched from blkio_group under blkio_cgroup->lock. That means
1121 * it should not be NULL as even if elevator was exiting, cgroup deltion
1122 * path got to it first.
1124 void cfq_unlink_blkio_group(void *key, struct blkio_group *blkg)
1126 unsigned long flags;
1127 struct cfq_data *cfqd = key;
1129 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1130 cfq_destroy_cfqg(cfqd, cfqg_of_blkg(blkg));
1131 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1134 #else /* GROUP_IOSCHED */
1135 static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd, int create)
1137 return &cfqd->root_group;
1140 static inline struct cfq_group *cfq_ref_get_cfqg(struct cfq_group *cfqg)
1146 cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) {
1150 static void cfq_release_cfq_groups(struct cfq_data *cfqd) {}
1151 static inline void cfq_put_cfqg(struct cfq_group *cfqg) {}
1153 #endif /* GROUP_IOSCHED */
1156 * The cfqd->service_trees holds all pending cfq_queue's that have
1157 * requests waiting to be processed. It is sorted in the order that
1158 * we will service the queues.
1160 static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1163 struct rb_node **p, *parent;
1164 struct cfq_queue *__cfqq;
1165 unsigned long rb_key;
1166 struct cfq_rb_root *service_tree;
1169 int group_changed = 0;
1171 #ifdef CONFIG_CFQ_GROUP_IOSCHED
1172 if (!cfqd->cfq_group_isolation
1173 && cfqq_type(cfqq) == SYNC_NOIDLE_WORKLOAD
1174 && cfqq->cfqg && cfqq->cfqg != &cfqd->root_group) {
1175 /* Move this cfq to root group */
1176 cfq_log_cfqq(cfqd, cfqq, "moving to root group");
1177 if (!RB_EMPTY_NODE(&cfqq->rb_node))
1178 cfq_group_service_tree_del(cfqd, cfqq->cfqg);
1179 cfqq->orig_cfqg = cfqq->cfqg;
1180 cfqq->cfqg = &cfqd->root_group;
1181 atomic_inc(&cfqd->root_group.ref);
1183 } else if (!cfqd->cfq_group_isolation
1184 && cfqq_type(cfqq) == SYNC_WORKLOAD && cfqq->orig_cfqg) {
1185 /* cfqq is sequential now needs to go to its original group */
1186 BUG_ON(cfqq->cfqg != &cfqd->root_group);
1187 if (!RB_EMPTY_NODE(&cfqq->rb_node))
1188 cfq_group_service_tree_del(cfqd, cfqq->cfqg);
1189 cfq_put_cfqg(cfqq->cfqg);
1190 cfqq->cfqg = cfqq->orig_cfqg;
1191 cfqq->orig_cfqg = NULL;
1193 cfq_log_cfqq(cfqd, cfqq, "moved to origin group");
1197 service_tree = service_tree_for(cfqq->cfqg, cfqq_prio(cfqq),
1199 if (cfq_class_idle(cfqq)) {
1200 rb_key = CFQ_IDLE_DELAY;
1201 parent = rb_last(&service_tree->rb);
1202 if (parent && parent != &cfqq->rb_node) {
1203 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1204 rb_key += __cfqq->rb_key;
1207 } else if (!add_front) {
1209 * Get our rb key offset. Subtract any residual slice
1210 * value carried from last service. A negative resid
1211 * count indicates slice overrun, and this should position
1212 * the next service time further away in the tree.
1214 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
1215 rb_key -= cfqq->slice_resid;
1216 cfqq->slice_resid = 0;
1219 __cfqq = cfq_rb_first(service_tree);
1220 rb_key += __cfqq ? __cfqq->rb_key : jiffies;
1223 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1226 * same position, nothing more to do
1228 if (rb_key == cfqq->rb_key &&
1229 cfqq->service_tree == service_tree)
1232 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1233 cfqq->service_tree = NULL;
1238 cfqq->service_tree = service_tree;
1239 p = &service_tree->rb.rb_node;
1244 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1247 * sort by key, that represents service time.
1249 if (time_before(rb_key, __cfqq->rb_key))
1252 n = &(*p)->rb_right;
1260 service_tree->left = &cfqq->rb_node;
1262 cfqq->rb_key = rb_key;
1263 rb_link_node(&cfqq->rb_node, parent, p);
1264 rb_insert_color(&cfqq->rb_node, &service_tree->rb);
1265 service_tree->count++;
1266 if ((add_front || !new_cfqq) && !group_changed)
1268 cfq_group_service_tree_add(cfqd, cfqq->cfqg);
1271 static struct cfq_queue *
1272 cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
1273 sector_t sector, struct rb_node **ret_parent,
1274 struct rb_node ***rb_link)
1276 struct rb_node **p, *parent;
1277 struct cfq_queue *cfqq = NULL;
1285 cfqq = rb_entry(parent, struct cfq_queue, p_node);
1288 * Sort strictly based on sector. Smallest to the left,
1289 * largest to the right.
1291 if (sector > blk_rq_pos(cfqq->next_rq))
1292 n = &(*p)->rb_right;
1293 else if (sector < blk_rq_pos(cfqq->next_rq))
1301 *ret_parent = parent;
1307 static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1309 struct rb_node **p, *parent;
1310 struct cfq_queue *__cfqq;
1313 rb_erase(&cfqq->p_node, cfqq->p_root);
1314 cfqq->p_root = NULL;
1317 if (cfq_class_idle(cfqq))
1322 cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
1323 __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
1324 blk_rq_pos(cfqq->next_rq), &parent, &p);
1326 rb_link_node(&cfqq->p_node, parent, p);
1327 rb_insert_color(&cfqq->p_node, cfqq->p_root);
1329 cfqq->p_root = NULL;
1333 * Update cfqq's position in the service tree.
1335 static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1338 * Resorting requires the cfqq to be on the RR list already.
1340 if (cfq_cfqq_on_rr(cfqq)) {
1341 cfq_service_tree_add(cfqd, cfqq, 0);
1342 cfq_prio_tree_add(cfqd, cfqq);
1347 * add to busy list of queues for service, trying to be fair in ordering
1348 * the pending list according to last request service
1350 static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1352 cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
1353 BUG_ON(cfq_cfqq_on_rr(cfqq));
1354 cfq_mark_cfqq_on_rr(cfqq);
1355 cfqd->busy_queues++;
1357 cfq_resort_rr_list(cfqd, cfqq);
1361 * Called when the cfqq no longer has requests pending, remove it from
1364 static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1366 cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
1367 BUG_ON(!cfq_cfqq_on_rr(cfqq));
1368 cfq_clear_cfqq_on_rr(cfqq);
1370 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1371 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1372 cfqq->service_tree = NULL;
1375 rb_erase(&cfqq->p_node, cfqq->p_root);
1376 cfqq->p_root = NULL;
1379 cfq_group_service_tree_del(cfqd, cfqq->cfqg);
1380 BUG_ON(!cfqd->busy_queues);
1381 cfqd->busy_queues--;
1385 * rb tree support functions
1387 static void cfq_del_rq_rb(struct request *rq)
1389 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1390 const int sync = rq_is_sync(rq);
1392 BUG_ON(!cfqq->queued[sync]);
1393 cfqq->queued[sync]--;
1395 elv_rb_del(&cfqq->sort_list, rq);
1397 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) {
1399 * Queue will be deleted from service tree when we actually
1400 * expire it later. Right now just remove it from prio tree
1404 rb_erase(&cfqq->p_node, cfqq->p_root);
1405 cfqq->p_root = NULL;
1410 static void cfq_add_rq_rb(struct request *rq)
1412 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1413 struct cfq_data *cfqd = cfqq->cfqd;
1414 struct request *__alias, *prev;
1416 cfqq->queued[rq_is_sync(rq)]++;
1419 * looks a little odd, but the first insert might return an alias.
1420 * if that happens, put the alias on the dispatch list
1422 while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL)
1423 cfq_dispatch_insert(cfqd->queue, __alias);
1425 if (!cfq_cfqq_on_rr(cfqq))
1426 cfq_add_cfqq_rr(cfqd, cfqq);
1429 * check if this request is a better next-serve candidate
1431 prev = cfqq->next_rq;
1432 cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq, cfqd->last_position);
1435 * adjust priority tree position, if ->next_rq changes
1437 if (prev != cfqq->next_rq)
1438 cfq_prio_tree_add(cfqd, cfqq);
1440 BUG_ON(!cfqq->next_rq);
1443 static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
1445 elv_rb_del(&cfqq->sort_list, rq);
1446 cfqq->queued[rq_is_sync(rq)]--;
1447 cfq_blkiocg_update_io_remove_stats(&(RQ_CFQG(rq))->blkg,
1448 rq_data_dir(rq), rq_is_sync(rq));
1450 cfq_blkiocg_update_io_add_stats(&(RQ_CFQG(rq))->blkg,
1451 &cfqq->cfqd->serving_group->blkg, rq_data_dir(rq),
1455 static struct request *
1456 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
1458 struct task_struct *tsk = current;
1459 struct cfq_io_context *cic;
1460 struct cfq_queue *cfqq;
1462 cic = cfq_cic_lookup(cfqd, tsk->io_context);
1466 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
1468 sector_t sector = bio->bi_sector + bio_sectors(bio);
1470 return elv_rb_find(&cfqq->sort_list, sector);
1476 static void cfq_activate_request(struct request_queue *q, struct request *rq)
1478 struct cfq_data *cfqd = q->elevator->elevator_data;
1480 cfqd->rq_in_driver++;
1481 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
1482 cfqd->rq_in_driver);
1484 cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
1487 static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
1489 struct cfq_data *cfqd = q->elevator->elevator_data;
1491 WARN_ON(!cfqd->rq_in_driver);
1492 cfqd->rq_in_driver--;
1493 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
1494 cfqd->rq_in_driver);
1497 static void cfq_remove_request(struct request *rq)
1499 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1501 if (cfqq->next_rq == rq)
1502 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
1504 list_del_init(&rq->queuelist);
1507 cfqq->cfqd->rq_queued--;
1508 cfq_blkiocg_update_io_remove_stats(&(RQ_CFQG(rq))->blkg,
1509 rq_data_dir(rq), rq_is_sync(rq));
1510 if (rq->cmd_flags & REQ_META) {
1511 WARN_ON(!cfqq->meta_pending);
1512 cfqq->meta_pending--;
1516 static int cfq_merge(struct request_queue *q, struct request **req,
1519 struct cfq_data *cfqd = q->elevator->elevator_data;
1520 struct request *__rq;
1522 __rq = cfq_find_rq_fmerge(cfqd, bio);
1523 if (__rq && elv_rq_merge_ok(__rq, bio)) {
1525 return ELEVATOR_FRONT_MERGE;
1528 return ELEVATOR_NO_MERGE;
1531 static void cfq_merged_request(struct request_queue *q, struct request *req,
1534 if (type == ELEVATOR_FRONT_MERGE) {
1535 struct cfq_queue *cfqq = RQ_CFQQ(req);
1537 cfq_reposition_rq_rb(cfqq, req);
1541 static void cfq_bio_merged(struct request_queue *q, struct request *req,
1544 cfq_blkiocg_update_io_merged_stats(&(RQ_CFQG(req))->blkg,
1545 bio_data_dir(bio), cfq_bio_sync(bio));
1549 cfq_merged_requests(struct request_queue *q, struct request *rq,
1550 struct request *next)
1552 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1554 * reposition in fifo if next is older than rq
1556 if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
1557 time_before(rq_fifo_time(next), rq_fifo_time(rq))) {
1558 list_move(&rq->queuelist, &next->queuelist);
1559 rq_set_fifo_time(rq, rq_fifo_time(next));
1562 if (cfqq->next_rq == next)
1564 cfq_remove_request(next);
1565 cfq_blkiocg_update_io_merged_stats(&(RQ_CFQG(rq))->blkg,
1566 rq_data_dir(next), rq_is_sync(next));
1569 static int cfq_allow_merge(struct request_queue *q, struct request *rq,
1572 struct cfq_data *cfqd = q->elevator->elevator_data;
1573 struct cfq_io_context *cic;
1574 struct cfq_queue *cfqq;
1577 * Disallow merge of a sync bio into an async request.
1579 if (cfq_bio_sync(bio) && !rq_is_sync(rq))
1583 * Lookup the cfqq that this bio will be queued with. Allow
1584 * merge only if rq is queued there.
1586 cic = cfq_cic_lookup(cfqd, current->io_context);
1590 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
1591 return cfqq == RQ_CFQQ(rq);
1594 static inline void cfq_del_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1596 del_timer(&cfqd->idle_slice_timer);
1597 cfq_blkiocg_update_idle_time_stats(&cfqq->cfqg->blkg);
1600 static void __cfq_set_active_queue(struct cfq_data *cfqd,
1601 struct cfq_queue *cfqq)
1604 cfq_log_cfqq(cfqd, cfqq, "set_active wl_prio:%d wl_type:%d",
1605 cfqd->serving_prio, cfqd->serving_type);
1606 cfq_blkiocg_update_avg_queue_size_stats(&cfqq->cfqg->blkg);
1607 cfqq->slice_start = 0;
1608 cfqq->dispatch_start = jiffies;
1609 cfqq->allocated_slice = 0;
1610 cfqq->slice_end = 0;
1611 cfqq->slice_dispatch = 0;
1613 cfq_clear_cfqq_wait_request(cfqq);
1614 cfq_clear_cfqq_must_dispatch(cfqq);
1615 cfq_clear_cfqq_must_alloc_slice(cfqq);
1616 cfq_clear_cfqq_fifo_expire(cfqq);
1617 cfq_mark_cfqq_slice_new(cfqq);
1619 cfq_del_timer(cfqd, cfqq);
1622 cfqd->active_queue = cfqq;
1626 * current cfqq expired its slice (or was too idle), select new one
1629 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1632 cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
1634 if (cfq_cfqq_wait_request(cfqq))
1635 cfq_del_timer(cfqd, cfqq);
1637 cfq_clear_cfqq_wait_request(cfqq);
1638 cfq_clear_cfqq_wait_busy(cfqq);
1641 * If this cfqq is shared between multiple processes, check to
1642 * make sure that those processes are still issuing I/Os within
1643 * the mean seek distance. If not, it may be time to break the
1644 * queues apart again.
1646 if (cfq_cfqq_coop(cfqq) && CFQQ_SEEKY(cfqq))
1647 cfq_mark_cfqq_split_coop(cfqq);
1650 * store what was left of this slice, if the queue idled/timed out
1652 if (timed_out && !cfq_cfqq_slice_new(cfqq)) {
1653 cfqq->slice_resid = cfqq->slice_end - jiffies;
1654 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
1657 cfq_group_served(cfqd, cfqq->cfqg, cfqq);
1659 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
1660 cfq_del_cfqq_rr(cfqd, cfqq);
1662 cfq_resort_rr_list(cfqd, cfqq);
1664 if (cfqq == cfqd->active_queue)
1665 cfqd->active_queue = NULL;
1667 if (&cfqq->cfqg->rb_node == cfqd->grp_service_tree.active)
1668 cfqd->grp_service_tree.active = NULL;
1670 if (cfqd->active_cic) {
1671 put_io_context(cfqd->active_cic->ioc);
1672 cfqd->active_cic = NULL;
1676 static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out)
1678 struct cfq_queue *cfqq = cfqd->active_queue;
1681 __cfq_slice_expired(cfqd, cfqq, timed_out);
1685 * Get next queue for service. Unless we have a queue preemption,
1686 * we'll simply select the first cfqq in the service tree.
1688 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
1690 struct cfq_rb_root *service_tree =
1691 service_tree_for(cfqd->serving_group, cfqd->serving_prio,
1692 cfqd->serving_type);
1694 if (!cfqd->rq_queued)
1697 /* There is nothing to dispatch */
1700 if (RB_EMPTY_ROOT(&service_tree->rb))
1702 return cfq_rb_first(service_tree);
1705 static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd)
1707 struct cfq_group *cfqg;
1708 struct cfq_queue *cfqq;
1710 struct cfq_rb_root *st;
1712 if (!cfqd->rq_queued)
1715 cfqg = cfq_get_next_cfqg(cfqd);
1719 for_each_cfqg_st(cfqg, i, j, st)
1720 if ((cfqq = cfq_rb_first(st)) != NULL)
1726 * Get and set a new active queue for service.
1728 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
1729 struct cfq_queue *cfqq)
1732 cfqq = cfq_get_next_queue(cfqd);
1734 __cfq_set_active_queue(cfqd, cfqq);
1738 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
1741 if (blk_rq_pos(rq) >= cfqd->last_position)
1742 return blk_rq_pos(rq) - cfqd->last_position;
1744 return cfqd->last_position - blk_rq_pos(rq);
1747 static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1750 return cfq_dist_from_last(cfqd, rq) <= CFQQ_CLOSE_THR;
1753 static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
1754 struct cfq_queue *cur_cfqq)
1756 struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
1757 struct rb_node *parent, *node;
1758 struct cfq_queue *__cfqq;
1759 sector_t sector = cfqd->last_position;
1761 if (RB_EMPTY_ROOT(root))
1765 * First, if we find a request starting at the end of the last
1766 * request, choose it.
1768 __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
1773 * If the exact sector wasn't found, the parent of the NULL leaf
1774 * will contain the closest sector.
1776 __cfqq = rb_entry(parent, struct cfq_queue, p_node);
1777 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1780 if (blk_rq_pos(__cfqq->next_rq) < sector)
1781 node = rb_next(&__cfqq->p_node);
1783 node = rb_prev(&__cfqq->p_node);
1787 __cfqq = rb_entry(node, struct cfq_queue, p_node);
1788 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1796 * cur_cfqq - passed in so that we don't decide that the current queue is
1797 * closely cooperating with itself.
1799 * So, basically we're assuming that that cur_cfqq has dispatched at least
1800 * one request, and that cfqd->last_position reflects a position on the disk
1801 * associated with the I/O issued by cur_cfqq. I'm not sure this is a valid
1804 static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
1805 struct cfq_queue *cur_cfqq)
1807 struct cfq_queue *cfqq;
1809 if (cfq_class_idle(cur_cfqq))
1811 if (!cfq_cfqq_sync(cur_cfqq))
1813 if (CFQQ_SEEKY(cur_cfqq))
1817 * Don't search priority tree if it's the only queue in the group.
1819 if (cur_cfqq->cfqg->nr_cfqq == 1)
1823 * We should notice if some of the queues are cooperating, eg
1824 * working closely on the same area of the disk. In that case,
1825 * we can group them together and don't waste time idling.
1827 cfqq = cfqq_close(cfqd, cur_cfqq);
1831 /* If new queue belongs to different cfq_group, don't choose it */
1832 if (cur_cfqq->cfqg != cfqq->cfqg)
1836 * It only makes sense to merge sync queues.
1838 if (!cfq_cfqq_sync(cfqq))
1840 if (CFQQ_SEEKY(cfqq))
1844 * Do not merge queues of different priority classes
1846 if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq))
1853 * Determine whether we should enforce idle window for this queue.
1856 static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1858 enum wl_prio_t prio = cfqq_prio(cfqq);
1859 struct cfq_rb_root *service_tree = cfqq->service_tree;
1861 BUG_ON(!service_tree);
1862 BUG_ON(!service_tree->count);
1864 if (!cfqd->cfq_slice_idle)
1867 /* We never do for idle class queues. */
1868 if (prio == IDLE_WORKLOAD)
1871 /* We do for queues that were marked with idle window flag. */
1872 if (cfq_cfqq_idle_window(cfqq) &&
1873 !(blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag))
1877 * Otherwise, we do only if they are the last ones
1878 * in their service tree.
1880 if (service_tree->count == 1 && cfq_cfqq_sync(cfqq))
1882 cfq_log_cfqq(cfqd, cfqq, "Not idling. st->count:%d",
1883 service_tree->count);
1887 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
1889 struct cfq_queue *cfqq = cfqd->active_queue;
1890 struct cfq_io_context *cic;
1891 unsigned long sl, group_idle = 0;
1894 * SSD device without seek penalty, disable idling. But only do so
1895 * for devices that support queuing, otherwise we still have a problem
1896 * with sync vs async workloads.
1898 if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
1901 WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
1902 WARN_ON(cfq_cfqq_slice_new(cfqq));
1905 * idle is disabled, either manually or by past process history
1907 if (!cfq_should_idle(cfqd, cfqq)) {
1908 /* no queue idling. Check for group idling */
1909 if (cfqd->cfq_group_idle)
1910 group_idle = cfqd->cfq_group_idle;
1916 * still active requests from this queue, don't idle
1918 if (cfqq->dispatched)
1922 * task has exited, don't wait
1924 cic = cfqd->active_cic;
1925 if (!cic || !atomic_read(&cic->ioc->nr_tasks))
1929 * If our average think time is larger than the remaining time
1930 * slice, then don't idle. This avoids overrunning the allotted
1933 if (sample_valid(cic->ttime_samples) &&
1934 (cfqq->slice_end - jiffies < cic->ttime_mean)) {
1935 cfq_log_cfqq(cfqd, cfqq, "Not idling. think_time:%d",
1940 /* There are other queues in the group, don't do group idle */
1941 if (group_idle && cfqq->cfqg->nr_cfqq > 1)
1944 cfq_mark_cfqq_wait_request(cfqq);
1947 sl = cfqd->cfq_group_idle;
1949 sl = cfqd->cfq_slice_idle;
1951 mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
1952 cfq_blkiocg_update_set_idle_time_stats(&cfqq->cfqg->blkg);
1953 cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu group_idle: %d", sl,
1954 group_idle ? 1 : 0);
1958 * Move request from internal lists to the request queue dispatch list.
1960 static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
1962 struct cfq_data *cfqd = q->elevator->elevator_data;
1963 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1965 cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
1967 cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
1968 cfq_remove_request(rq);
1970 (RQ_CFQG(rq))->dispatched++;
1971 elv_dispatch_sort(q, rq);
1973 cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]++;
1974 cfq_blkiocg_update_dispatch_stats(&cfqq->cfqg->blkg, blk_rq_bytes(rq),
1975 rq_data_dir(rq), rq_is_sync(rq));
1979 * return expired entry, or NULL to just start from scratch in rbtree
1981 static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
1983 struct request *rq = NULL;
1985 if (cfq_cfqq_fifo_expire(cfqq))
1988 cfq_mark_cfqq_fifo_expire(cfqq);
1990 if (list_empty(&cfqq->fifo))
1993 rq = rq_entry_fifo(cfqq->fifo.next);
1994 if (time_before(jiffies, rq_fifo_time(rq)))
1997 cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq);
2002 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2004 const int base_rq = cfqd->cfq_slice_async_rq;
2006 WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
2008 return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
2012 * Must be called with the queue_lock held.
2014 static int cfqq_process_refs(struct cfq_queue *cfqq)
2016 int process_refs, io_refs;
2018 io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE];
2019 process_refs = atomic_read(&cfqq->ref) - io_refs;
2020 BUG_ON(process_refs < 0);
2021 return process_refs;
2024 static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq)
2026 int process_refs, new_process_refs;
2027 struct cfq_queue *__cfqq;
2030 * If there are no process references on the new_cfqq, then it is
2031 * unsafe to follow the ->new_cfqq chain as other cfqq's in the
2032 * chain may have dropped their last reference (not just their
2033 * last process reference).
2035 if (!cfqq_process_refs(new_cfqq))
2038 /* Avoid a circular list and skip interim queue merges */
2039 while ((__cfqq = new_cfqq->new_cfqq)) {
2045 process_refs = cfqq_process_refs(cfqq);
2046 new_process_refs = cfqq_process_refs(new_cfqq);
2048 * If the process for the cfqq has gone away, there is no
2049 * sense in merging the queues.
2051 if (process_refs == 0 || new_process_refs == 0)
2055 * Merge in the direction of the lesser amount of work.
2057 if (new_process_refs >= process_refs) {
2058 cfqq->new_cfqq = new_cfqq;
2059 atomic_add(process_refs, &new_cfqq->ref);
2061 new_cfqq->new_cfqq = cfqq;
2062 atomic_add(new_process_refs, &cfqq->ref);
2066 static enum wl_type_t cfq_choose_wl(struct cfq_data *cfqd,
2067 struct cfq_group *cfqg, enum wl_prio_t prio)
2069 struct cfq_queue *queue;
2071 bool key_valid = false;
2072 unsigned long lowest_key = 0;
2073 enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD;
2075 for (i = 0; i <= SYNC_WORKLOAD; ++i) {
2076 /* select the one with lowest rb_key */
2077 queue = cfq_rb_first(service_tree_for(cfqg, prio, i));
2079 (!key_valid || time_before(queue->rb_key, lowest_key))) {
2080 lowest_key = queue->rb_key;
2089 static void choose_service_tree(struct cfq_data *cfqd, struct cfq_group *cfqg)
2093 struct cfq_rb_root *st;
2094 unsigned group_slice;
2097 cfqd->serving_prio = IDLE_WORKLOAD;
2098 cfqd->workload_expires = jiffies + 1;
2102 /* Choose next priority. RT > BE > IDLE */
2103 if (cfq_group_busy_queues_wl(RT_WORKLOAD, cfqd, cfqg))
2104 cfqd->serving_prio = RT_WORKLOAD;
2105 else if (cfq_group_busy_queues_wl(BE_WORKLOAD, cfqd, cfqg))
2106 cfqd->serving_prio = BE_WORKLOAD;
2108 cfqd->serving_prio = IDLE_WORKLOAD;
2109 cfqd->workload_expires = jiffies + 1;
2114 * For RT and BE, we have to choose also the type
2115 * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
2118 st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
2122 * check workload expiration, and that we still have other queues ready
2124 if (count && !time_after(jiffies, cfqd->workload_expires))
2127 /* otherwise select new workload type */
2128 cfqd->serving_type =
2129 cfq_choose_wl(cfqd, cfqg, cfqd->serving_prio);
2130 st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
2134 * the workload slice is computed as a fraction of target latency
2135 * proportional to the number of queues in that workload, over
2136 * all the queues in the same priority class
2138 group_slice = cfq_group_slice(cfqd, cfqg);
2140 slice = group_slice * count /
2141 max_t(unsigned, cfqg->busy_queues_avg[cfqd->serving_prio],
2142 cfq_group_busy_queues_wl(cfqd->serving_prio, cfqd, cfqg));
2144 if (cfqd->serving_type == ASYNC_WORKLOAD) {
2148 * Async queues are currently system wide. Just taking
2149 * proportion of queues with-in same group will lead to higher
2150 * async ratio system wide as generally root group is going
2151 * to have higher weight. A more accurate thing would be to
2152 * calculate system wide asnc/sync ratio.
2154 tmp = cfq_target_latency * cfqg_busy_async_queues(cfqd, cfqg);
2155 tmp = tmp/cfqd->busy_queues;
2156 slice = min_t(unsigned, slice, tmp);
2158 /* async workload slice is scaled down according to
2159 * the sync/async slice ratio. */
2160 slice = slice * cfqd->cfq_slice[0] / cfqd->cfq_slice[1];
2162 /* sync workload slice is at least 2 * cfq_slice_idle */
2163 slice = max(slice, 2 * cfqd->cfq_slice_idle);
2165 slice = max_t(unsigned, slice, CFQ_MIN_TT);
2166 cfq_log(cfqd, "workload slice:%d", slice);
2167 cfqd->workload_expires = jiffies + slice;
2168 cfqd->noidle_tree_requires_idle = false;
2171 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd)
2173 struct cfq_rb_root *st = &cfqd->grp_service_tree;
2174 struct cfq_group *cfqg;
2176 if (RB_EMPTY_ROOT(&st->rb))
2178 cfqg = cfq_rb_first_group(st);
2179 st->active = &cfqg->rb_node;
2180 update_min_vdisktime(st);
2184 static void cfq_choose_cfqg(struct cfq_data *cfqd)
2186 struct cfq_group *cfqg = cfq_get_next_cfqg(cfqd);
2188 cfqd->serving_group = cfqg;
2190 /* Restore the workload type data */
2191 if (cfqg->saved_workload_slice) {
2192 cfqd->workload_expires = jiffies + cfqg->saved_workload_slice;
2193 cfqd->serving_type = cfqg->saved_workload;
2194 cfqd->serving_prio = cfqg->saved_serving_prio;
2196 cfqd->workload_expires = jiffies - 1;
2198 choose_service_tree(cfqd, cfqg);
2202 * Select a queue for service. If we have a current active queue,
2203 * check whether to continue servicing it, or retrieve and set a new one.
2205 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
2207 struct cfq_queue *cfqq, *new_cfqq = NULL;
2209 cfqq = cfqd->active_queue;
2213 if (!cfqd->rq_queued)
2217 * We were waiting for group to get backlogged. Expire the queue
2219 if (cfq_cfqq_wait_busy(cfqq) && !RB_EMPTY_ROOT(&cfqq->sort_list))
2223 * The active queue has run out of time, expire it and select new.
2225 if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq)) {
2227 * If slice had not expired at the completion of last request
2228 * we might not have turned on wait_busy flag. Don't expire
2229 * the queue yet. Allow the group to get backlogged.
2231 * The very fact that we have used the slice, that means we
2232 * have been idling all along on this queue and it should be
2233 * ok to wait for this request to complete.
2235 if (cfqq->cfqg->nr_cfqq == 1 && RB_EMPTY_ROOT(&cfqq->sort_list)
2236 && cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2240 goto check_group_idle;
2244 * The active queue has requests and isn't expired, allow it to
2247 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
2251 * If another queue has a request waiting within our mean seek
2252 * distance, let it run. The expire code will check for close
2253 * cooperators and put the close queue at the front of the service
2254 * tree. If possible, merge the expiring queue with the new cfqq.
2256 new_cfqq = cfq_close_cooperator(cfqd, cfqq);
2258 if (!cfqq->new_cfqq)
2259 cfq_setup_merge(cfqq, new_cfqq);
2264 * No requests pending. If the active queue still has requests in
2265 * flight or is idling for a new request, allow either of these
2266 * conditions to happen (or time out) before selecting a new queue.
2268 if (timer_pending(&cfqd->idle_slice_timer)) {
2273 if (cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2279 * If group idle is enabled and there are requests dispatched from
2280 * this group, wait for requests to complete.
2283 if (cfqd->cfq_group_idle && cfqq->cfqg->nr_cfqq == 1
2284 && cfqq->cfqg->dispatched) {
2290 cfq_slice_expired(cfqd, 0);
2293 * Current queue expired. Check if we have to switch to a new
2297 cfq_choose_cfqg(cfqd);
2299 cfqq = cfq_set_active_queue(cfqd, new_cfqq);
2304 static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
2308 while (cfqq->next_rq) {
2309 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
2313 BUG_ON(!list_empty(&cfqq->fifo));
2315 /* By default cfqq is not expired if it is empty. Do it explicitly */
2316 __cfq_slice_expired(cfqq->cfqd, cfqq, 0);
2321 * Drain our current requests. Used for barriers and when switching
2322 * io schedulers on-the-fly.
2324 static int cfq_forced_dispatch(struct cfq_data *cfqd)
2326 struct cfq_queue *cfqq;
2329 /* Expire the timeslice of the current active queue first */
2330 cfq_slice_expired(cfqd, 0);
2331 while ((cfqq = cfq_get_next_queue_forced(cfqd)) != NULL) {
2332 __cfq_set_active_queue(cfqd, cfqq);
2333 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
2336 BUG_ON(cfqd->busy_queues);
2338 cfq_log(cfqd, "forced_dispatch=%d", dispatched);
2342 static inline bool cfq_slice_used_soon(struct cfq_data *cfqd,
2343 struct cfq_queue *cfqq)
2345 /* the queue hasn't finished any request, can't estimate */
2346 if (cfq_cfqq_slice_new(cfqq))
2348 if (time_after(jiffies + cfqd->cfq_slice_idle * cfqq->dispatched,
2355 static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2357 unsigned int max_dispatch;
2360 * Drain async requests before we start sync IO
2362 if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_flight[BLK_RW_ASYNC])
2366 * If this is an async queue and we have sync IO in flight, let it wait
2368 if (cfqd->rq_in_flight[BLK_RW_SYNC] && !cfq_cfqq_sync(cfqq))
2371 max_dispatch = max_t(unsigned int, cfqd->cfq_quantum / 2, 1);
2372 if (cfq_class_idle(cfqq))
2376 * Does this cfqq already have too much IO in flight?
2378 if (cfqq->dispatched >= max_dispatch) {
2380 * idle queue must always only have a single IO in flight
2382 if (cfq_class_idle(cfqq))
2386 * We have other queues, don't allow more IO from this one
2388 if (cfqd->busy_queues > 1 && cfq_slice_used_soon(cfqd, cfqq))
2392 * Sole queue user, no limit
2394 if (cfqd->busy_queues == 1)
2398 * Normally we start throttling cfqq when cfq_quantum/2
2399 * requests have been dispatched. But we can drive
2400 * deeper queue depths at the beginning of slice
2401 * subjected to upper limit of cfq_quantum.
2403 max_dispatch = cfqd->cfq_quantum;
2407 * Async queues must wait a bit before being allowed dispatch.
2408 * We also ramp up the dispatch depth gradually for async IO,
2409 * based on the last sync IO we serviced
2411 if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
2412 unsigned long last_sync = jiffies - cfqd->last_delayed_sync;
2415 depth = last_sync / cfqd->cfq_slice[1];
2416 if (!depth && !cfqq->dispatched)
2418 if (depth < max_dispatch)
2419 max_dispatch = depth;
2423 * If we're below the current max, allow a dispatch
2425 return cfqq->dispatched < max_dispatch;
2429 * Dispatch a request from cfqq, moving them to the request queue
2432 static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2436 BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
2438 if (!cfq_may_dispatch(cfqd, cfqq))
2442 * follow expired path, else get first next available
2444 rq = cfq_check_fifo(cfqq);
2449 * insert request into driver dispatch list
2451 cfq_dispatch_insert(cfqd->queue, rq);
2453 if (!cfqd->active_cic) {
2454 struct cfq_io_context *cic = RQ_CIC(rq);
2456 atomic_long_inc(&cic->ioc->refcount);
2457 cfqd->active_cic = cic;
2464 * Find the cfqq that we need to service and move a request from that to the
2467 static int cfq_dispatch_requests(struct request_queue *q, int force)
2469 struct cfq_data *cfqd = q->elevator->elevator_data;
2470 struct cfq_queue *cfqq;
2472 if (!cfqd->busy_queues)
2475 if (unlikely(force))
2476 return cfq_forced_dispatch(cfqd);
2478 cfqq = cfq_select_queue(cfqd);
2483 * Dispatch a request from this cfqq, if it is allowed
2485 if (!cfq_dispatch_request(cfqd, cfqq))
2488 cfqq->slice_dispatch++;
2489 cfq_clear_cfqq_must_dispatch(cfqq);
2492 * expire an async queue immediately if it has used up its slice. idle
2493 * queue always expire after 1 dispatch round.
2495 if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
2496 cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
2497 cfq_class_idle(cfqq))) {
2498 cfqq->slice_end = jiffies + 1;
2499 cfq_slice_expired(cfqd, 0);
2502 cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
2507 * task holds one reference to the queue, dropped when task exits. each rq
2508 * in-flight on this queue also holds a reference, dropped when rq is freed.
2510 * Each cfq queue took a reference on the parent group. Drop it now.
2511 * queue lock must be held here.
2513 static void cfq_put_queue(struct cfq_queue *cfqq)
2515 struct cfq_data *cfqd = cfqq->cfqd;
2516 struct cfq_group *cfqg, *orig_cfqg;
2518 BUG_ON(atomic_read(&cfqq->ref) <= 0);
2520 if (!atomic_dec_and_test(&cfqq->ref))
2523 cfq_log_cfqq(cfqd, cfqq, "put_queue");
2524 BUG_ON(rb_first(&cfqq->sort_list));
2525 BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
2527 orig_cfqg = cfqq->orig_cfqg;
2529 if (unlikely(cfqd->active_queue == cfqq)) {
2530 __cfq_slice_expired(cfqd, cfqq, 0);
2531 cfq_schedule_dispatch(cfqd);
2534 BUG_ON(cfq_cfqq_on_rr(cfqq));
2535 kmem_cache_free(cfq_pool, cfqq);
2538 cfq_put_cfqg(orig_cfqg);
2542 * Must always be called with the rcu_read_lock() held
2545 __call_for_each_cic(struct io_context *ioc,
2546 void (*func)(struct io_context *, struct cfq_io_context *))
2548 struct cfq_io_context *cic;
2549 struct hlist_node *n;
2551 hlist_for_each_entry_rcu(cic, n, &ioc->cic_list, cic_list)
2556 * Call func for each cic attached to this ioc.
2559 call_for_each_cic(struct io_context *ioc,
2560 void (*func)(struct io_context *, struct cfq_io_context *))
2563 __call_for_each_cic(ioc, func);
2567 static void cfq_cic_free_rcu(struct rcu_head *head)
2569 struct cfq_io_context *cic;
2571 cic = container_of(head, struct cfq_io_context, rcu_head);
2573 kmem_cache_free(cfq_ioc_pool, cic);
2574 elv_ioc_count_dec(cfq_ioc_count);
2578 * CFQ scheduler is exiting, grab exit lock and check
2579 * the pending io context count. If it hits zero,
2580 * complete ioc_gone and set it back to NULL
2582 spin_lock(&ioc_gone_lock);
2583 if (ioc_gone && !elv_ioc_count_read(cfq_ioc_count)) {
2587 spin_unlock(&ioc_gone_lock);
2591 static void cfq_cic_free(struct cfq_io_context *cic)
2593 call_rcu(&cic->rcu_head, cfq_cic_free_rcu);
2596 static void cic_free_func(struct io_context *ioc, struct cfq_io_context *cic)
2598 unsigned long flags;
2599 unsigned long dead_key = (unsigned long) cic->key;
2601 BUG_ON(!(dead_key & CIC_DEAD_KEY));
2603 spin_lock_irqsave(&ioc->lock, flags);
2604 radix_tree_delete(&ioc->radix_root, dead_key >> CIC_DEAD_INDEX_SHIFT);
2605 hlist_del_rcu(&cic->cic_list);
2606 spin_unlock_irqrestore(&ioc->lock, flags);
2612 * Must be called with rcu_read_lock() held or preemption otherwise disabled.
2613 * Only two callers of this - ->dtor() which is called with the rcu_read_lock(),
2614 * and ->trim() which is called with the task lock held
2616 static void cfq_free_io_context(struct io_context *ioc)
2619 * ioc->refcount is zero here, or we are called from elv_unregister(),
2620 * so no more cic's are allowed to be linked into this ioc. So it
2621 * should be ok to iterate over the known list, we will see all cic's
2622 * since no new ones are added.
2624 __call_for_each_cic(ioc, cic_free_func);
2627 static void cfq_put_cooperator(struct cfq_queue *cfqq)
2629 struct cfq_queue *__cfqq, *next;
2632 * If this queue was scheduled to merge with another queue, be
2633 * sure to drop the reference taken on that queue (and others in
2634 * the merge chain). See cfq_setup_merge and cfq_merge_cfqqs.
2636 __cfqq = cfqq->new_cfqq;
2638 if (__cfqq == cfqq) {
2639 WARN(1, "cfqq->new_cfqq loop detected\n");
2642 next = __cfqq->new_cfqq;
2643 cfq_put_queue(__cfqq);
2648 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2650 if (unlikely(cfqq == cfqd->active_queue)) {
2651 __cfq_slice_expired(cfqd, cfqq, 0);
2652 cfq_schedule_dispatch(cfqd);
2655 cfq_put_cooperator(cfqq);
2657 cfq_put_queue(cfqq);
2660 static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
2661 struct cfq_io_context *cic)
2663 struct io_context *ioc = cic->ioc;
2665 list_del_init(&cic->queue_list);
2668 * Make sure dead mark is seen for dead queues
2671 cic->key = cfqd_dead_key(cfqd);
2673 if (ioc->ioc_data == cic)
2674 rcu_assign_pointer(ioc->ioc_data, NULL);
2676 if (cic->cfqq[BLK_RW_ASYNC]) {
2677 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]);
2678 cic->cfqq[BLK_RW_ASYNC] = NULL;
2681 if (cic->cfqq[BLK_RW_SYNC]) {
2682 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]);
2683 cic->cfqq[BLK_RW_SYNC] = NULL;
2687 static void cfq_exit_single_io_context(struct io_context *ioc,
2688 struct cfq_io_context *cic)
2690 struct cfq_data *cfqd = cic_to_cfqd(cic);
2693 struct request_queue *q = cfqd->queue;
2694 unsigned long flags;
2696 spin_lock_irqsave(q->queue_lock, flags);
2699 * Ensure we get a fresh copy of the ->key to prevent
2700 * race between exiting task and queue
2702 smp_read_barrier_depends();
2703 if (cic->key == cfqd)
2704 __cfq_exit_single_io_context(cfqd, cic);
2706 spin_unlock_irqrestore(q->queue_lock, flags);
2711 * The process that ioc belongs to has exited, we need to clean up
2712 * and put the internal structures we have that belongs to that process.
2714 static void cfq_exit_io_context(struct io_context *ioc)
2716 call_for_each_cic(ioc, cfq_exit_single_io_context);
2719 static struct cfq_io_context *
2720 cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
2722 struct cfq_io_context *cic;
2724 cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO,
2727 cic->last_end_request = jiffies;
2728 INIT_LIST_HEAD(&cic->queue_list);
2729 INIT_HLIST_NODE(&cic->cic_list);
2730 cic->dtor = cfq_free_io_context;
2731 cic->exit = cfq_exit_io_context;
2732 elv_ioc_count_inc(cfq_ioc_count);
2738 static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc)
2740 struct task_struct *tsk = current;
2743 if (!cfq_cfqq_prio_changed(cfqq))
2746 ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio);
2747 switch (ioprio_class) {
2749 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
2750 case IOPRIO_CLASS_NONE:
2752 * no prio set, inherit CPU scheduling settings
2754 cfqq->ioprio = task_nice_ioprio(tsk);
2755 cfqq->ioprio_class = task_nice_ioclass(tsk);
2757 case IOPRIO_CLASS_RT:
2758 cfqq->ioprio = task_ioprio(ioc);
2759 cfqq->ioprio_class = IOPRIO_CLASS_RT;
2761 case IOPRIO_CLASS_BE:
2762 cfqq->ioprio = task_ioprio(ioc);
2763 cfqq->ioprio_class = IOPRIO_CLASS_BE;
2765 case IOPRIO_CLASS_IDLE:
2766 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
2768 cfq_clear_cfqq_idle_window(cfqq);
2773 * keep track of original prio settings in case we have to temporarily
2774 * elevate the priority of this queue
2776 cfqq->org_ioprio = cfqq->ioprio;
2777 cfqq->org_ioprio_class = cfqq->ioprio_class;
2778 cfq_clear_cfqq_prio_changed(cfqq);
2781 static void changed_ioprio(struct io_context *ioc, struct cfq_io_context *cic)
2783 struct cfq_data *cfqd = cic_to_cfqd(cic);
2784 struct cfq_queue *cfqq;
2785 unsigned long flags;
2787 if (unlikely(!cfqd))
2790 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2792 cfqq = cic->cfqq[BLK_RW_ASYNC];
2794 struct cfq_queue *new_cfqq;
2795 new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic->ioc,
2798 cic->cfqq[BLK_RW_ASYNC] = new_cfqq;
2799 cfq_put_queue(cfqq);
2803 cfqq = cic->cfqq[BLK_RW_SYNC];
2805 cfq_mark_cfqq_prio_changed(cfqq);
2807 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2810 static void cfq_ioc_set_ioprio(struct io_context *ioc)
2812 call_for_each_cic(ioc, changed_ioprio);
2813 ioc->ioprio_changed = 0;
2816 static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2817 pid_t pid, bool is_sync)
2819 RB_CLEAR_NODE(&cfqq->rb_node);
2820 RB_CLEAR_NODE(&cfqq->p_node);
2821 INIT_LIST_HEAD(&cfqq->fifo);
2823 atomic_set(&cfqq->ref, 0);
2826 cfq_mark_cfqq_prio_changed(cfqq);
2829 if (!cfq_class_idle(cfqq))
2830 cfq_mark_cfqq_idle_window(cfqq);
2831 cfq_mark_cfqq_sync(cfqq);
2836 #ifdef CONFIG_CFQ_GROUP_IOSCHED
2837 static void changed_cgroup(struct io_context *ioc, struct cfq_io_context *cic)
2839 struct cfq_queue *sync_cfqq = cic_to_cfqq(cic, 1);
2840 struct cfq_data *cfqd = cic_to_cfqd(cic);
2841 unsigned long flags;
2842 struct request_queue *q;
2844 if (unlikely(!cfqd))
2849 spin_lock_irqsave(q->queue_lock, flags);
2853 * Drop reference to sync queue. A new sync queue will be
2854 * assigned in new group upon arrival of a fresh request.
2856 cfq_log_cfqq(cfqd, sync_cfqq, "changed cgroup");
2857 cic_set_cfqq(cic, NULL, 1);
2858 cfq_put_queue(sync_cfqq);
2861 spin_unlock_irqrestore(q->queue_lock, flags);
2864 static void cfq_ioc_set_cgroup(struct io_context *ioc)
2866 call_for_each_cic(ioc, changed_cgroup);
2867 ioc->cgroup_changed = 0;
2869 #endif /* CONFIG_CFQ_GROUP_IOSCHED */
2871 static struct cfq_queue *
2872 cfq_find_alloc_queue(struct cfq_data *cfqd, bool is_sync,
2873 struct io_context *ioc, gfp_t gfp_mask)
2875 struct cfq_queue *cfqq, *new_cfqq = NULL;
2876 struct cfq_io_context *cic;
2877 struct cfq_group *cfqg;
2880 cfqg = cfq_get_cfqg(cfqd, 1);
2881 cic = cfq_cic_lookup(cfqd, ioc);
2882 /* cic always exists here */
2883 cfqq = cic_to_cfqq(cic, is_sync);
2886 * Always try a new alloc if we fell back to the OOM cfqq
2887 * originally, since it should just be a temporary situation.
2889 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
2894 } else if (gfp_mask & __GFP_WAIT) {
2895 spin_unlock_irq(cfqd->queue->queue_lock);
2896 new_cfqq = kmem_cache_alloc_node(cfq_pool,
2897 gfp_mask | __GFP_ZERO,
2899 spin_lock_irq(cfqd->queue->queue_lock);
2903 cfqq = kmem_cache_alloc_node(cfq_pool,
2904 gfp_mask | __GFP_ZERO,
2909 cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
2910 cfq_init_prio_data(cfqq, ioc);
2911 cfq_link_cfqq_cfqg(cfqq, cfqg);
2912 cfq_log_cfqq(cfqd, cfqq, "alloced");
2914 cfqq = &cfqd->oom_cfqq;
2918 kmem_cache_free(cfq_pool, new_cfqq);
2923 static struct cfq_queue **
2924 cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
2926 switch (ioprio_class) {
2927 case IOPRIO_CLASS_RT:
2928 return &cfqd->async_cfqq[0][ioprio];
2929 case IOPRIO_CLASS_BE:
2930 return &cfqd->async_cfqq[1][ioprio];
2931 case IOPRIO_CLASS_IDLE:
2932 return &cfqd->async_idle_cfqq;
2938 static struct cfq_queue *
2939 cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct io_context *ioc,
2942 const int ioprio = task_ioprio(ioc);
2943 const int ioprio_class = task_ioprio_class(ioc);
2944 struct cfq_queue **async_cfqq = NULL;
2945 struct cfq_queue *cfqq = NULL;
2948 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
2953 cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask);
2956 * pin the queue now that it's allocated, scheduler exit will prune it
2958 if (!is_sync && !(*async_cfqq)) {
2959 atomic_inc(&cfqq->ref);
2963 atomic_inc(&cfqq->ref);
2968 * We drop cfq io contexts lazily, so we may find a dead one.
2971 cfq_drop_dead_cic(struct cfq_data *cfqd, struct io_context *ioc,
2972 struct cfq_io_context *cic)
2974 unsigned long flags;
2976 WARN_ON(!list_empty(&cic->queue_list));
2977 BUG_ON(cic->key != cfqd_dead_key(cfqd));
2979 spin_lock_irqsave(&ioc->lock, flags);
2981 BUG_ON(ioc->ioc_data == cic);
2983 radix_tree_delete(&ioc->radix_root, cfqd->cic_index);
2984 hlist_del_rcu(&cic->cic_list);
2985 spin_unlock_irqrestore(&ioc->lock, flags);
2990 static struct cfq_io_context *
2991 cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc)
2993 struct cfq_io_context *cic;
2994 unsigned long flags;
3002 * we maintain a last-hit cache, to avoid browsing over the tree
3004 cic = rcu_dereference(ioc->ioc_data);
3005 if (cic && cic->key == cfqd) {
3011 cic = radix_tree_lookup(&ioc->radix_root, cfqd->cic_index);
3015 if (unlikely(cic->key != cfqd)) {
3016 cfq_drop_dead_cic(cfqd, ioc, cic);
3021 spin_lock_irqsave(&ioc->lock, flags);
3022 rcu_assign_pointer(ioc->ioc_data, cic);
3023 spin_unlock_irqrestore(&ioc->lock, flags);
3031 * Add cic into ioc, using cfqd as the search key. This enables us to lookup
3032 * the process specific cfq io context when entered from the block layer.
3033 * Also adds the cic to a per-cfqd list, used when this queue is removed.
3035 static int cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
3036 struct cfq_io_context *cic, gfp_t gfp_mask)
3038 unsigned long flags;
3041 ret = radix_tree_preload(gfp_mask);
3046 spin_lock_irqsave(&ioc->lock, flags);
3047 ret = radix_tree_insert(&ioc->radix_root,
3048 cfqd->cic_index, cic);
3050 hlist_add_head_rcu(&cic->cic_list, &ioc->cic_list);
3051 spin_unlock_irqrestore(&ioc->lock, flags);
3053 radix_tree_preload_end();
3056 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
3057 list_add(&cic->queue_list, &cfqd->cic_list);
3058 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
3063 printk(KERN_ERR "cfq: cic link failed!\n");
3069 * Setup general io context and cfq io context. There can be several cfq
3070 * io contexts per general io context, if this process is doing io to more
3071 * than one device managed by cfq.
3073 static struct cfq_io_context *
3074 cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
3076 struct io_context *ioc = NULL;
3077 struct cfq_io_context *cic;
3079 might_sleep_if(gfp_mask & __GFP_WAIT);
3081 ioc = get_io_context(gfp_mask, cfqd->queue->node);
3085 cic = cfq_cic_lookup(cfqd, ioc);
3089 cic = cfq_alloc_io_context(cfqd, gfp_mask);
3093 if (cfq_cic_link(cfqd, ioc, cic, gfp_mask))
3097 smp_read_barrier_depends();
3098 if (unlikely(ioc->ioprio_changed))
3099 cfq_ioc_set_ioprio(ioc);
3101 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3102 if (unlikely(ioc->cgroup_changed))
3103 cfq_ioc_set_cgroup(ioc);
3109 put_io_context(ioc);
3114 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
3116 unsigned long elapsed = jiffies - cic->last_end_request;
3117 unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
3119 cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
3120 cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
3121 cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
3125 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3129 sector_t n_sec = blk_rq_sectors(rq);
3130 if (cfqq->last_request_pos) {
3131 if (cfqq->last_request_pos < blk_rq_pos(rq))
3132 sdist = blk_rq_pos(rq) - cfqq->last_request_pos;
3134 sdist = cfqq->last_request_pos - blk_rq_pos(rq);
3137 cfqq->seek_history <<= 1;
3138 if (blk_queue_nonrot(cfqd->queue))
3139 cfqq->seek_history |= (n_sec < CFQQ_SECT_THR_NONROT);
3141 cfqq->seek_history |= (sdist > CFQQ_SEEK_THR);
3145 * Disable idle window if the process thinks too long or seeks so much that
3149 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3150 struct cfq_io_context *cic)
3152 int old_idle, enable_idle;
3155 * Don't idle for async or idle io prio class
3157 if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
3160 enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
3162 if (cfqq->queued[0] + cfqq->queued[1] >= 4)
3163 cfq_mark_cfqq_deep(cfqq);
3165 if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle ||
3166 (!cfq_cfqq_deep(cfqq) && CFQQ_SEEKY(cfqq)))
3168 else if (sample_valid(cic->ttime_samples)) {
3169 if (cic->ttime_mean > cfqd->cfq_slice_idle)
3175 if (old_idle != enable_idle) {
3176 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
3178 cfq_mark_cfqq_idle_window(cfqq);
3180 cfq_clear_cfqq_idle_window(cfqq);
3185 * Check if new_cfqq should preempt the currently active queue. Return 0 for
3186 * no or if we aren't sure, a 1 will cause a preempt.
3189 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
3192 struct cfq_queue *cfqq;
3194 cfqq = cfqd->active_queue;
3198 if (cfq_class_idle(new_cfqq))
3201 if (cfq_class_idle(cfqq))
3205 * Don't allow a non-RT request to preempt an ongoing RT cfqq timeslice.
3207 if (cfq_class_rt(cfqq) && !cfq_class_rt(new_cfqq))
3211 * if the new request is sync, but the currently running queue is
3212 * not, let the sync request have priority.
3214 if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
3217 if (new_cfqq->cfqg != cfqq->cfqg)
3220 if (cfq_slice_used(cfqq))
3223 /* Allow preemption only if we are idling on sync-noidle tree */
3224 if (cfqd->serving_type == SYNC_NOIDLE_WORKLOAD &&
3225 cfqq_type(new_cfqq) == SYNC_NOIDLE_WORKLOAD &&
3226 new_cfqq->service_tree->count == 2 &&
3227 RB_EMPTY_ROOT(&cfqq->sort_list))
3231 * So both queues are sync. Let the new request get disk time if
3232 * it's a metadata request and the current queue is doing regular IO.
3234 if ((rq->cmd_flags & REQ_META) && !cfqq->meta_pending)
3238 * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
3240 if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
3243 if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
3247 * if this request is as-good as one we would expect from the
3248 * current cfqq, let it preempt
3250 if (cfq_rq_close(cfqd, cfqq, rq))
3257 * cfqq preempts the active queue. if we allowed preempt with no slice left,
3258 * let it have half of its nominal slice.
3260 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3262 cfq_log_cfqq(cfqd, cfqq, "preempt");
3263 cfq_slice_expired(cfqd, 1);
3266 * Put the new queue at the front of the of the current list,
3267 * so we know that it will be selected next.
3269 BUG_ON(!cfq_cfqq_on_rr(cfqq));
3271 cfq_service_tree_add(cfqd, cfqq, 1);
3273 cfqq->slice_end = 0;
3274 cfq_mark_cfqq_slice_new(cfqq);
3278 * Called when a new fs request (rq) is added (to cfqq). Check if there's
3279 * something we should do about it
3282 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3285 struct cfq_io_context *cic = RQ_CIC(rq);
3288 if (rq->cmd_flags & REQ_META)
3289 cfqq->meta_pending++;
3291 cfq_update_io_thinktime(cfqd, cic);
3292 cfq_update_io_seektime(cfqd, cfqq, rq);
3293 cfq_update_idle_window(cfqd, cfqq, cic);
3295 cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
3297 if (cfqq == cfqd->active_queue) {
3299 * Remember that we saw a request from this process, but
3300 * don't start queuing just yet. Otherwise we risk seeing lots
3301 * of tiny requests, because we disrupt the normal plugging
3302 * and merging. If the request is already larger than a single
3303 * page, let it rip immediately. For that case we assume that
3304 * merging is already done. Ditto for a busy system that
3305 * has other work pending, don't risk delaying until the
3306 * idle timer unplug to continue working.
3308 if (cfq_cfqq_wait_request(cfqq)) {
3309 if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
3310 cfqd->busy_queues > 1) {
3311 cfq_del_timer(cfqd, cfqq);
3312 cfq_clear_cfqq_wait_request(cfqq);
3313 __blk_run_queue(cfqd->queue);
3315 cfq_blkiocg_update_idle_time_stats(
3317 cfq_mark_cfqq_must_dispatch(cfqq);
3320 } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
3322 * not the active queue - expire current slice if it is
3323 * idle and has expired it's mean thinktime or this new queue
3324 * has some old slice time left and is of higher priority or
3325 * this new queue is RT and the current one is BE
3327 cfq_preempt_queue(cfqd, cfqq);
3328 __blk_run_queue(cfqd->queue);
3332 static void cfq_insert_request(struct request_queue *q, struct request *rq)
3334 struct cfq_data *cfqd = q->elevator->elevator_data;
3335 struct cfq_queue *cfqq = RQ_CFQQ(rq);
3337 cfq_log_cfqq(cfqd, cfqq, "insert_request");
3338 cfq_init_prio_data(cfqq, RQ_CIC(rq)->ioc);
3340 rq_set_fifo_time(rq, jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)]);
3341 list_add_tail(&rq->queuelist, &cfqq->fifo);
3343 cfq_blkiocg_update_io_add_stats(&(RQ_CFQG(rq))->blkg,
3344 &cfqd->serving_group->blkg, rq_data_dir(rq),
3346 cfq_rq_enqueued(cfqd, cfqq, rq);
3350 * Update hw_tag based on peak queue depth over 50 samples under
3353 static void cfq_update_hw_tag(struct cfq_data *cfqd)
3355 struct cfq_queue *cfqq = cfqd->active_queue;
3357 if (cfqd->rq_in_driver > cfqd->hw_tag_est_depth)
3358 cfqd->hw_tag_est_depth = cfqd->rq_in_driver;
3360 if (cfqd->hw_tag == 1)
3363 if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
3364 cfqd->rq_in_driver <= CFQ_HW_QUEUE_MIN)
3368 * If active queue hasn't enough requests and can idle, cfq might not
3369 * dispatch sufficient requests to hardware. Don't zero hw_tag in this
3372 if (cfqq && cfq_cfqq_idle_window(cfqq) &&
3373 cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] <
3374 CFQ_HW_QUEUE_MIN && cfqd->rq_in_driver < CFQ_HW_QUEUE_MIN)
3377 if (cfqd->hw_tag_samples++ < 50)
3380 if (cfqd->hw_tag_est_depth >= CFQ_HW_QUEUE_MIN)
3386 static bool cfq_should_wait_busy(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3388 struct cfq_io_context *cic = cfqd->active_cic;
3390 /* If there are other queues in the group, don't wait */
3391 if (cfqq->cfqg->nr_cfqq > 1)
3394 if (cfq_slice_used(cfqq))
3397 /* if slice left is less than think time, wait busy */
3398 if (cic && sample_valid(cic->ttime_samples)
3399 && (cfqq->slice_end - jiffies < cic->ttime_mean))
3403 * If think times is less than a jiffy than ttime_mean=0 and above
3404 * will not be true. It might happen that slice has not expired yet
3405 * but will expire soon (4-5 ns) during select_queue(). To cover the
3406 * case where think time is less than a jiffy, mark the queue wait
3407 * busy if only 1 jiffy is left in the slice.
3409 if (cfqq->slice_end - jiffies == 1)
3415 static void cfq_completed_request(struct request_queue *q, struct request *rq)
3417 struct cfq_queue *cfqq = RQ_CFQQ(rq);
3418 struct cfq_data *cfqd = cfqq->cfqd;
3419 const int sync = rq_is_sync(rq);
3423 cfq_log_cfqq(cfqd, cfqq, "complete rqnoidle %d",
3424 !!(rq->cmd_flags & REQ_NOIDLE));
3426 cfq_update_hw_tag(cfqd);
3428 WARN_ON(!cfqd->rq_in_driver);
3429 WARN_ON(!cfqq->dispatched);
3430 cfqd->rq_in_driver--;
3432 (RQ_CFQG(rq))->dispatched--;
3433 cfq_blkiocg_update_completion_stats(&cfqq->cfqg->blkg,
3434 rq_start_time_ns(rq), rq_io_start_time_ns(rq),
3435 rq_data_dir(rq), rq_is_sync(rq));
3437 cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]--;
3440 RQ_CIC(rq)->last_end_request = now;
3441 if (!time_after(rq->start_time + cfqd->cfq_fifo_expire[1], now))
3442 cfqd->last_delayed_sync = now;
3446 * If this is the active queue, check if it needs to be expired,
3447 * or if we want to idle in case it has no pending requests.
3449 if (cfqd->active_queue == cfqq) {
3450 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
3452 if (cfq_cfqq_slice_new(cfqq)) {
3453 cfq_set_prio_slice(cfqd, cfqq);
3454 cfq_clear_cfqq_slice_new(cfqq);
3458 * Should we wait for next request to come in before we expire
3461 if (cfq_should_wait_busy(cfqd, cfqq)) {
3462 unsigned long extend_sl = cfqd->cfq_slice_idle;
3463 if (!cfqd->cfq_slice_idle)
3464 extend_sl = cfqd->cfq_group_idle;
3465 cfqq->slice_end = jiffies + extend_sl;
3466 cfq_mark_cfqq_wait_busy(cfqq);
3467 cfq_log_cfqq(cfqd, cfqq, "will busy wait");
3471 * Idling is not enabled on:
3473 * - idle-priority queues
3475 * - queues with still some requests queued
3476 * - when there is a close cooperator
3478 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
3479 cfq_slice_expired(cfqd, 1);
3480 else if (sync && cfqq_empty &&
3481 !cfq_close_cooperator(cfqd, cfqq)) {
3482 cfqd->noidle_tree_requires_idle |=
3483 !(rq->cmd_flags & REQ_NOIDLE);
3485 * Idling is enabled for SYNC_WORKLOAD.
3486 * SYNC_NOIDLE_WORKLOAD idles at the end of the tree
3487 * only if we processed at least one !REQ_NOIDLE request
3489 if (cfqd->serving_type == SYNC_WORKLOAD
3490 || cfqd->noidle_tree_requires_idle
3491 || cfqq->cfqg->nr_cfqq == 1)
3492 cfq_arm_slice_timer(cfqd);
3496 if (!cfqd->rq_in_driver)
3497 cfq_schedule_dispatch(cfqd);
3501 * we temporarily boost lower priority queues if they are holding fs exclusive
3502 * resources. they are boosted to normal prio (CLASS_BE/4)
3504 static void cfq_prio_boost(struct cfq_queue *cfqq)
3506 if (has_fs_excl()) {
3508 * boost idle prio on transactions that would lock out other
3509 * users of the filesystem
3511 if (cfq_class_idle(cfqq))
3512 cfqq->ioprio_class = IOPRIO_CLASS_BE;
3513 if (cfqq->ioprio > IOPRIO_NORM)
3514 cfqq->ioprio = IOPRIO_NORM;
3517 * unboost the queue (if needed)
3519 cfqq->ioprio_class = cfqq->org_ioprio_class;
3520 cfqq->ioprio = cfqq->org_ioprio;
3524 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
3526 if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
3527 cfq_mark_cfqq_must_alloc_slice(cfqq);
3528 return ELV_MQUEUE_MUST;
3531 return ELV_MQUEUE_MAY;
3534 static int cfq_may_queue(struct request_queue *q, int rw)
3536 struct cfq_data *cfqd = q->elevator->elevator_data;
3537 struct task_struct *tsk = current;
3538 struct cfq_io_context *cic;
3539 struct cfq_queue *cfqq;
3542 * don't force setup of a queue from here, as a call to may_queue
3543 * does not necessarily imply that a request actually will be queued.
3544 * so just lookup a possibly existing queue, or return 'may queue'
3547 cic = cfq_cic_lookup(cfqd, tsk->io_context);
3549 return ELV_MQUEUE_MAY;
3551 cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
3553 cfq_init_prio_data(cfqq, cic->ioc);
3554 cfq_prio_boost(cfqq);
3556 return __cfq_may_queue(cfqq);
3559 return ELV_MQUEUE_MAY;
3563 * queue lock held here
3565 static void cfq_put_request(struct request *rq)
3567 struct cfq_queue *cfqq = RQ_CFQQ(rq);
3570 const int rw = rq_data_dir(rq);
3572 BUG_ON(!cfqq->allocated[rw]);
3573 cfqq->allocated[rw]--;
3575 put_io_context(RQ_CIC(rq)->ioc);
3577 rq->elevator_private = NULL;
3578 rq->elevator_private2 = NULL;
3580 /* Put down rq reference on cfqg */
3581 cfq_put_cfqg(RQ_CFQG(rq));
3582 rq->elevator_private3 = NULL;
3584 cfq_put_queue(cfqq);
3588 static struct cfq_queue *
3589 cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_context *cic,
3590 struct cfq_queue *cfqq)
3592 cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq);
3593 cic_set_cfqq(cic, cfqq->new_cfqq, 1);
3594 cfq_mark_cfqq_coop(cfqq->new_cfqq);
3595 cfq_put_queue(cfqq);
3596 return cic_to_cfqq(cic, 1);
3600 * Returns NULL if a new cfqq should be allocated, or the old cfqq if this
3601 * was the last process referring to said cfqq.
3603 static struct cfq_queue *
3604 split_cfqq(struct cfq_io_context *cic, struct cfq_queue *cfqq)
3606 if (cfqq_process_refs(cfqq) == 1) {
3607 cfqq->pid = current->pid;
3608 cfq_clear_cfqq_coop(cfqq);
3609 cfq_clear_cfqq_split_coop(cfqq);
3613 cic_set_cfqq(cic, NULL, 1);
3615 cfq_put_cooperator(cfqq);
3617 cfq_put_queue(cfqq);
3621 * Allocate cfq data structures associated with this request.
3624 cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
3626 struct cfq_data *cfqd = q->elevator->elevator_data;
3627 struct cfq_io_context *cic;
3628 const int rw = rq_data_dir(rq);
3629 const bool is_sync = rq_is_sync(rq);
3630 struct cfq_queue *cfqq;
3631 unsigned long flags;
3633 might_sleep_if(gfp_mask & __GFP_WAIT);
3635 cic = cfq_get_io_context(cfqd, gfp_mask);
3637 spin_lock_irqsave(q->queue_lock, flags);
3643 cfqq = cic_to_cfqq(cic, is_sync);
3644 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
3645 cfqq = cfq_get_queue(cfqd, is_sync, cic->ioc, gfp_mask);
3646 cic_set_cfqq(cic, cfqq, is_sync);
3649 * If the queue was seeky for too long, break it apart.
3651 if (cfq_cfqq_coop(cfqq) && cfq_cfqq_split_coop(cfqq)) {
3652 cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq");
3653 cfqq = split_cfqq(cic, cfqq);
3659 * Check to see if this queue is scheduled to merge with
3660 * another, closely cooperating queue. The merging of
3661 * queues happens here as it must be done in process context.
3662 * The reference on new_cfqq was taken in merge_cfqqs.
3665 cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq);
3668 cfqq->allocated[rw]++;
3669 atomic_inc(&cfqq->ref);
3671 spin_unlock_irqrestore(q->queue_lock, flags);
3673 rq->elevator_private = cic;
3674 rq->elevator_private2 = cfqq;
3675 rq->elevator_private3 = cfq_ref_get_cfqg(cfqq->cfqg);
3680 put_io_context(cic->ioc);
3682 cfq_schedule_dispatch(cfqd);
3683 spin_unlock_irqrestore(q->queue_lock, flags);
3684 cfq_log(cfqd, "set_request fail");
3688 static void cfq_kick_queue(struct work_struct *work)
3690 struct cfq_data *cfqd =
3691 container_of(work, struct cfq_data, unplug_work);
3692 struct request_queue *q = cfqd->queue;
3694 spin_lock_irq(q->queue_lock);
3695 __blk_run_queue(cfqd->queue);
3696 spin_unlock_irq(q->queue_lock);
3700 * Timer running if the active_queue is currently idling inside its time slice
3702 static void cfq_idle_slice_timer(unsigned long data)
3704 struct cfq_data *cfqd = (struct cfq_data *) data;
3705 struct cfq_queue *cfqq;
3706 unsigned long flags;
3709 cfq_log(cfqd, "idle timer fired");
3711 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
3713 cfqq = cfqd->active_queue;
3718 * We saw a request before the queue expired, let it through
3720 if (cfq_cfqq_must_dispatch(cfqq))
3726 if (cfq_slice_used(cfqq))
3730 * only expire and reinvoke request handler, if there are
3731 * other queues with pending requests
3733 if (!cfqd->busy_queues)
3737 * not expired and it has a request pending, let it dispatch
3739 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3743 * Queue depth flag is reset only when the idle didn't succeed
3745 cfq_clear_cfqq_deep(cfqq);
3748 cfq_slice_expired(cfqd, timed_out);
3750 cfq_schedule_dispatch(cfqd);
3752 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
3755 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
3757 del_timer_sync(&cfqd->idle_slice_timer);
3758 cancel_work_sync(&cfqd->unplug_work);
3761 static void cfq_put_async_queues(struct cfq_data *cfqd)
3765 for (i = 0; i < IOPRIO_BE_NR; i++) {
3766 if (cfqd->async_cfqq[0][i])
3767 cfq_put_queue(cfqd->async_cfqq[0][i]);
3768 if (cfqd->async_cfqq[1][i])
3769 cfq_put_queue(cfqd->async_cfqq[1][i]);
3772 if (cfqd->async_idle_cfqq)
3773 cfq_put_queue(cfqd->async_idle_cfqq);
3776 static void cfq_cfqd_free(struct rcu_head *head)
3778 kfree(container_of(head, struct cfq_data, rcu));
3781 static void cfq_exit_queue(struct elevator_queue *e)
3783 struct cfq_data *cfqd = e->elevator_data;
3784 struct request_queue *q = cfqd->queue;
3786 cfq_shutdown_timer_wq(cfqd);
3788 spin_lock_irq(q->queue_lock);
3790 if (cfqd->active_queue)
3791 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
3793 while (!list_empty(&cfqd->cic_list)) {
3794 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
3795 struct cfq_io_context,
3798 __cfq_exit_single_io_context(cfqd, cic);
3801 cfq_put_async_queues(cfqd);
3802 cfq_release_cfq_groups(cfqd);
3803 cfq_blkiocg_del_blkio_group(&cfqd->root_group.blkg);
3805 spin_unlock_irq(q->queue_lock);
3807 cfq_shutdown_timer_wq(cfqd);
3809 spin_lock(&cic_index_lock);
3810 ida_remove(&cic_index_ida, cfqd->cic_index);
3811 spin_unlock(&cic_index_lock);
3813 /* Wait for cfqg->blkg->key accessors to exit their grace periods. */
3814 call_rcu(&cfqd->rcu, cfq_cfqd_free);
3817 static int cfq_alloc_cic_index(void)
3822 if (!ida_pre_get(&cic_index_ida, GFP_KERNEL))
3825 spin_lock(&cic_index_lock);
3826 error = ida_get_new(&cic_index_ida, &index);
3827 spin_unlock(&cic_index_lock);
3828 if (error && error != -EAGAIN)
3835 static void *cfq_init_queue(struct request_queue *q)
3837 struct cfq_data *cfqd;
3839 struct cfq_group *cfqg;
3840 struct cfq_rb_root *st;
3842 i = cfq_alloc_cic_index();
3846 cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
3850 cfqd->cic_index = i;
3852 /* Init root service tree */
3853 cfqd->grp_service_tree = CFQ_RB_ROOT;
3855 /* Init root group */
3856 cfqg = &cfqd->root_group;
3857 for_each_cfqg_st(cfqg, i, j, st)
3859 RB_CLEAR_NODE(&cfqg->rb_node);
3861 /* Give preference to root group over other groups */
3862 cfqg->weight = 2*BLKIO_WEIGHT_DEFAULT;
3864 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3866 * Take a reference to root group which we never drop. This is just
3867 * to make sure that cfq_put_cfqg() does not try to kfree root group
3869 atomic_set(&cfqg->ref, 1);
3871 cfq_blkiocg_add_blkio_group(&blkio_root_cgroup, &cfqg->blkg,
3876 * Not strictly needed (since RB_ROOT just clears the node and we
3877 * zeroed cfqd on alloc), but better be safe in case someone decides
3878 * to add magic to the rb code
3880 for (i = 0; i < CFQ_PRIO_LISTS; i++)
3881 cfqd->prio_trees[i] = RB_ROOT;
3884 * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
3885 * Grab a permanent reference to it, so that the normal code flow
3886 * will not attempt to free it.
3888 cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
3889 atomic_inc(&cfqd->oom_cfqq.ref);
3890 cfq_link_cfqq_cfqg(&cfqd->oom_cfqq, &cfqd->root_group);
3892 INIT_LIST_HEAD(&cfqd->cic_list);
3896 init_timer(&cfqd->idle_slice_timer);
3897 cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
3898 cfqd->idle_slice_timer.data = (unsigned long) cfqd;
3900 INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
3902 cfqd->cfq_quantum = cfq_quantum;
3903 cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
3904 cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
3905 cfqd->cfq_back_max = cfq_back_max;
3906 cfqd->cfq_back_penalty = cfq_back_penalty;
3907 cfqd->cfq_slice[0] = cfq_slice_async;
3908 cfqd->cfq_slice[1] = cfq_slice_sync;
3909 cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
3910 cfqd->cfq_slice_idle = cfq_slice_idle;
3911 cfqd->cfq_group_idle = cfq_group_idle;
3912 cfqd->cfq_latency = 1;
3913 cfqd->cfq_group_isolation = 0;
3916 * we optimistically start assuming sync ops weren't delayed in last
3917 * second, in order to have larger depth for async operations.
3919 cfqd->last_delayed_sync = jiffies - HZ;
3923 static void cfq_slab_kill(void)
3926 * Caller already ensured that pending RCU callbacks are completed,
3927 * so we should have no busy allocations at this point.
3930 kmem_cache_destroy(cfq_pool);
3932 kmem_cache_destroy(cfq_ioc_pool);
3935 static int __init cfq_slab_setup(void)
3937 cfq_pool = KMEM_CACHE(cfq_queue, 0);
3941 cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0);
3952 * sysfs parts below -->
3955 cfq_var_show(unsigned int var, char *page)
3957 return sprintf(page, "%d\n", var);
3961 cfq_var_store(unsigned int *var, const char *page, size_t count)
3963 char *p = (char *) page;
3965 *var = simple_strtoul(p, &p, 10);
3969 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
3970 static ssize_t __FUNC(struct elevator_queue *e, char *page) \
3972 struct cfq_data *cfqd = e->elevator_data; \
3973 unsigned int __data = __VAR; \
3975 __data = jiffies_to_msecs(__data); \
3976 return cfq_var_show(__data, (page)); \
3978 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
3979 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
3980 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
3981 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
3982 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
3983 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
3984 SHOW_FUNCTION(cfq_group_idle_show, cfqd->cfq_group_idle, 1);
3985 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
3986 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
3987 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
3988 SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
3989 SHOW_FUNCTION(cfq_group_isolation_show, cfqd->cfq_group_isolation, 0);
3990 #undef SHOW_FUNCTION
3992 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
3993 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
3995 struct cfq_data *cfqd = e->elevator_data; \
3996 unsigned int __data; \
3997 int ret = cfq_var_store(&__data, (page), count); \
3998 if (__data < (MIN)) \
4000 else if (__data > (MAX)) \
4003 *(__PTR) = msecs_to_jiffies(__data); \
4005 *(__PTR) = __data; \
4008 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
4009 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
4011 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
4013 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
4014 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
4016 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
4017 STORE_FUNCTION(cfq_group_idle_store, &cfqd->cfq_group_idle, 0, UINT_MAX, 1);
4018 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
4019 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
4020 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
4022 STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
4023 STORE_FUNCTION(cfq_group_isolation_store, &cfqd->cfq_group_isolation, 0, 1, 0);
4024 #undef STORE_FUNCTION
4026 #define CFQ_ATTR(name) \
4027 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
4029 static struct elv_fs_entry cfq_attrs[] = {
4031 CFQ_ATTR(fifo_expire_sync),
4032 CFQ_ATTR(fifo_expire_async),
4033 CFQ_ATTR(back_seek_max),
4034 CFQ_ATTR(back_seek_penalty),
4035 CFQ_ATTR(slice_sync),
4036 CFQ_ATTR(slice_async),
4037 CFQ_ATTR(slice_async_rq),
4038 CFQ_ATTR(slice_idle),
4039 CFQ_ATTR(group_idle),
4040 CFQ_ATTR(low_latency),
4041 CFQ_ATTR(group_isolation),
4045 static struct elevator_type iosched_cfq = {
4047 .elevator_merge_fn = cfq_merge,
4048 .elevator_merged_fn = cfq_merged_request,
4049 .elevator_merge_req_fn = cfq_merged_requests,
4050 .elevator_allow_merge_fn = cfq_allow_merge,
4051 .elevator_bio_merged_fn = cfq_bio_merged,
4052 .elevator_dispatch_fn = cfq_dispatch_requests,
4053 .elevator_add_req_fn = cfq_insert_request,
4054 .elevator_activate_req_fn = cfq_activate_request,
4055 .elevator_deactivate_req_fn = cfq_deactivate_request,
4056 .elevator_queue_empty_fn = cfq_queue_empty,
4057 .elevator_completed_req_fn = cfq_completed_request,
4058 .elevator_former_req_fn = elv_rb_former_request,
4059 .elevator_latter_req_fn = elv_rb_latter_request,
4060 .elevator_set_req_fn = cfq_set_request,
4061 .elevator_put_req_fn = cfq_put_request,
4062 .elevator_may_queue_fn = cfq_may_queue,
4063 .elevator_init_fn = cfq_init_queue,
4064 .elevator_exit_fn = cfq_exit_queue,
4065 .trim = cfq_free_io_context,
4067 .elevator_attrs = cfq_attrs,
4068 .elevator_name = "cfq",
4069 .elevator_owner = THIS_MODULE,
4072 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4073 static struct blkio_policy_type blkio_policy_cfq = {
4075 .blkio_unlink_group_fn = cfq_unlink_blkio_group,
4076 .blkio_update_group_weight_fn = cfq_update_blkio_group_weight,
4080 static struct blkio_policy_type blkio_policy_cfq;
4083 static int __init cfq_init(void)
4086 * could be 0 on HZ < 1000 setups
4088 if (!cfq_slice_async)
4089 cfq_slice_async = 1;
4090 if (!cfq_slice_idle)
4093 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4094 if (!cfq_group_idle)
4099 if (cfq_slab_setup())
4102 elv_register(&iosched_cfq);
4103 blkio_policy_register(&blkio_policy_cfq);
4108 static void __exit cfq_exit(void)
4110 DECLARE_COMPLETION_ONSTACK(all_gone);
4111 blkio_policy_unregister(&blkio_policy_cfq);
4112 elv_unregister(&iosched_cfq);
4113 ioc_gone = &all_gone;
4114 /* ioc_gone's update must be visible before reading ioc_count */
4118 * this also protects us from entering cfq_slab_kill() with
4119 * pending RCU callbacks
4121 if (elv_ioc_count_read(cfq_ioc_count))
4122 wait_for_completion(&all_gone);
4123 ida_destroy(&cic_index_ida);
4127 module_init(cfq_init);
4128 module_exit(cfq_exit);
4130 MODULE_AUTHOR("Jens Axboe");
4131 MODULE_LICENSE("GPL");
4132 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");