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 const int cfq_target_latency = HZ * 3/10; /* 300 ms */
34 static const int cfq_hist_divisor = 4;
37 * offset from end of service tree
39 #define CFQ_IDLE_DELAY (HZ / 5)
42 * below this threshold, we consider thinktime immediate
44 #define CFQ_MIN_TT (2)
46 #define CFQ_SLICE_SCALE (5)
47 #define CFQ_HW_QUEUE_MIN (5)
48 #define CFQ_SERVICE_SHIFT 12
50 #define CFQQ_SEEK_THR (sector_t)(8 * 100)
51 #define CFQQ_CLOSE_THR (sector_t)(8 * 1024)
52 #define CFQQ_SECT_THR_NONROT (sector_t)(2 * 32)
53 #define CFQQ_SEEKY(cfqq) (hweight32(cfqq->seek_history) > 32/8)
56 ((struct cfq_io_context *) (rq)->elevator_private)
57 #define RQ_CFQQ(rq) (struct cfq_queue *) ((rq)->elevator_private2)
58 #define RQ_CFQG(rq) (struct cfq_group *) ((rq)->elevator_private3)
60 static struct kmem_cache *cfq_pool;
61 static struct kmem_cache *cfq_ioc_pool;
63 static DEFINE_PER_CPU(unsigned long, cfq_ioc_count);
64 static struct completion *ioc_gone;
65 static DEFINE_SPINLOCK(ioc_gone_lock);
67 static DEFINE_SPINLOCK(cic_index_lock);
68 static DEFINE_IDA(cic_index_ida);
70 #define CFQ_PRIO_LISTS IOPRIO_BE_NR
71 #define cfq_class_idle(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
72 #define cfq_class_rt(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
74 #define sample_valid(samples) ((samples) > 80)
75 #define rb_entry_cfqg(node) rb_entry((node), struct cfq_group, rb_node)
78 * Most of our rbtree usage is for sorting with min extraction, so
79 * if we cache the leftmost node we don't have to walk down the tree
80 * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should
81 * move this into the elevator for the rq sorting as well.
87 unsigned total_weight;
89 struct rb_node *active;
91 #define CFQ_RB_ROOT (struct cfq_rb_root) { .rb = RB_ROOT, .left = NULL, \
92 .count = 0, .min_vdisktime = 0, }
95 * Per process-grouping structure
100 /* various state flags, see below */
102 /* parent cfq_data */
103 struct cfq_data *cfqd;
104 /* service_tree member */
105 struct rb_node rb_node;
106 /* service_tree key */
107 unsigned long rb_key;
108 /* prio tree member */
109 struct rb_node p_node;
110 /* prio tree root we belong to, if any */
111 struct rb_root *p_root;
112 /* sorted list of pending requests */
113 struct rb_root sort_list;
114 /* if fifo isn't expired, next request to serve */
115 struct request *next_rq;
116 /* requests queued in sort_list */
118 /* currently allocated requests */
120 /* fifo list of requests in sort_list */
121 struct list_head fifo;
123 /* time when queue got scheduled in to dispatch first request. */
124 unsigned long dispatch_start;
125 unsigned int allocated_slice;
126 unsigned int slice_dispatch;
127 /* time when first request from queue completed and slice started. */
128 unsigned long slice_start;
129 unsigned long slice_end;
132 /* pending metadata requests */
134 /* number of requests that are on the dispatch list or inside driver */
137 /* io prio of this group */
138 unsigned short ioprio, org_ioprio;
139 unsigned short ioprio_class, org_ioprio_class;
144 sector_t last_request_pos;
146 struct cfq_rb_root *service_tree;
147 struct cfq_queue *new_cfqq;
148 struct cfq_group *cfqg;
149 struct cfq_group *orig_cfqg;
153 * First index in the service_trees.
154 * IDLE is handled separately, so it has negative index
163 * Second index in the service_trees.
167 SYNC_NOIDLE_WORKLOAD = 1,
171 /* This is per cgroup per device grouping structure */
173 /* group service_tree member */
174 struct rb_node rb_node;
176 /* group service_tree key */
181 /* number of cfqq currently on this group */
184 /* Per group busy queus average. Useful for workload slice calc. */
185 unsigned int busy_queues_avg[2];
187 * rr lists of queues with requests, onle rr for each priority class.
188 * Counts are embedded in the cfq_rb_root
190 struct cfq_rb_root service_trees[2][3];
191 struct cfq_rb_root service_tree_idle;
193 unsigned long saved_workload_slice;
194 enum wl_type_t saved_workload;
195 enum wl_prio_t saved_serving_prio;
196 struct blkio_group blkg;
197 #ifdef CONFIG_CFQ_GROUP_IOSCHED
198 struct hlist_node cfqd_node;
204 * Per block device queue structure
207 struct request_queue *queue;
208 /* Root service tree for cfq_groups */
209 struct cfq_rb_root grp_service_tree;
210 struct cfq_group root_group;
213 * The priority currently being served
215 enum wl_prio_t serving_prio;
216 enum wl_type_t serving_type;
217 unsigned long workload_expires;
218 struct cfq_group *serving_group;
221 * Each priority tree is sorted by next_request position. These
222 * trees are used when determining if two or more queues are
223 * interleaving requests (see cfq_close_cooperator).
225 struct rb_root prio_trees[CFQ_PRIO_LISTS];
227 unsigned int busy_queues;
233 * queue-depth detection
239 * -1 => indeterminate, (cfq will behave as if NCQ is present, to allow better detection)
240 * 1 => NCQ is present (hw_tag_est_depth is the estimated max depth)
243 int hw_tag_est_depth;
244 unsigned int hw_tag_samples;
247 * idle window management
249 struct timer_list idle_slice_timer;
250 struct work_struct unplug_work;
252 struct cfq_queue *active_queue;
253 struct cfq_io_context *active_cic;
256 * async queue for each priority case
258 struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
259 struct cfq_queue *async_idle_cfqq;
261 sector_t last_position;
264 * tunables, see top of file
266 unsigned int cfq_quantum;
267 unsigned int cfq_fifo_expire[2];
268 unsigned int cfq_back_penalty;
269 unsigned int cfq_back_max;
270 unsigned int cfq_slice[2];
271 unsigned int cfq_slice_async_rq;
272 unsigned int cfq_slice_idle;
273 unsigned int cfq_latency;
274 unsigned int cfq_group_isolation;
276 unsigned int cic_index;
277 struct list_head cic_list;
280 * Fallback dummy cfqq for extreme OOM conditions
282 struct cfq_queue oom_cfqq;
284 unsigned long last_delayed_sync;
286 /* List of cfq groups being managed on this device*/
287 struct hlist_head cfqg_list;
291 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd);
293 static struct cfq_rb_root *service_tree_for(struct cfq_group *cfqg,
300 if (prio == IDLE_WORKLOAD)
301 return &cfqg->service_tree_idle;
303 return &cfqg->service_trees[prio][type];
306 enum cfqq_state_flags {
307 CFQ_CFQQ_FLAG_on_rr = 0, /* on round-robin busy list */
308 CFQ_CFQQ_FLAG_wait_request, /* waiting for a request */
309 CFQ_CFQQ_FLAG_must_dispatch, /* must be allowed a dispatch */
310 CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
311 CFQ_CFQQ_FLAG_fifo_expire, /* FIFO checked in this slice */
312 CFQ_CFQQ_FLAG_idle_window, /* slice idling enabled */
313 CFQ_CFQQ_FLAG_prio_changed, /* task priority has changed */
314 CFQ_CFQQ_FLAG_slice_new, /* no requests dispatched in slice */
315 CFQ_CFQQ_FLAG_sync, /* synchronous queue */
316 CFQ_CFQQ_FLAG_coop, /* cfqq is shared */
317 CFQ_CFQQ_FLAG_split_coop, /* shared cfqq will be splitted */
318 CFQ_CFQQ_FLAG_deep, /* sync cfqq experienced large depth */
319 CFQ_CFQQ_FLAG_wait_busy, /* Waiting for next request */
322 #define CFQ_CFQQ_FNS(name) \
323 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \
325 (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name); \
327 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \
329 (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \
331 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \
333 return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \
337 CFQ_CFQQ_FNS(wait_request);
338 CFQ_CFQQ_FNS(must_dispatch);
339 CFQ_CFQQ_FNS(must_alloc_slice);
340 CFQ_CFQQ_FNS(fifo_expire);
341 CFQ_CFQQ_FNS(idle_window);
342 CFQ_CFQQ_FNS(prio_changed);
343 CFQ_CFQQ_FNS(slice_new);
346 CFQ_CFQQ_FNS(split_coop);
348 CFQ_CFQQ_FNS(wait_busy);
351 #ifdef CONFIG_CFQ_GROUP_IOSCHED
352 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
353 blk_add_trace_msg((cfqd)->queue, "cfq%d%c %s " fmt, (cfqq)->pid, \
354 cfq_cfqq_sync((cfqq)) ? 'S' : 'A', \
355 blkg_path(&(cfqq)->cfqg->blkg), ##args);
357 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...) \
358 blk_add_trace_msg((cfqd)->queue, "%s " fmt, \
359 blkg_path(&(cfqg)->blkg), ##args); \
362 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
363 blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args)
364 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...) do {} while (0);
366 #define cfq_log(cfqd, fmt, args...) \
367 blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
369 /* Traverses through cfq group service trees */
370 #define for_each_cfqg_st(cfqg, i, j, st) \
371 for (i = 0; i <= IDLE_WORKLOAD; i++) \
372 for (j = 0, st = i < IDLE_WORKLOAD ? &cfqg->service_trees[i][j]\
373 : &cfqg->service_tree_idle; \
374 (i < IDLE_WORKLOAD && j <= SYNC_WORKLOAD) || \
375 (i == IDLE_WORKLOAD && j == 0); \
376 j++, st = i < IDLE_WORKLOAD ? \
377 &cfqg->service_trees[i][j]: NULL) \
380 static inline enum wl_prio_t cfqq_prio(struct cfq_queue *cfqq)
382 if (cfq_class_idle(cfqq))
383 return IDLE_WORKLOAD;
384 if (cfq_class_rt(cfqq))
390 static enum wl_type_t cfqq_type(struct cfq_queue *cfqq)
392 if (!cfq_cfqq_sync(cfqq))
393 return ASYNC_WORKLOAD;
394 if (!cfq_cfqq_idle_window(cfqq))
395 return SYNC_NOIDLE_WORKLOAD;
396 return SYNC_WORKLOAD;
399 static inline int cfq_group_busy_queues_wl(enum wl_prio_t wl,
400 struct cfq_data *cfqd,
401 struct cfq_group *cfqg)
403 if (wl == IDLE_WORKLOAD)
404 return cfqg->service_tree_idle.count;
406 return cfqg->service_trees[wl][ASYNC_WORKLOAD].count
407 + cfqg->service_trees[wl][SYNC_NOIDLE_WORKLOAD].count
408 + cfqg->service_trees[wl][SYNC_WORKLOAD].count;
411 static inline int cfqg_busy_async_queues(struct cfq_data *cfqd,
412 struct cfq_group *cfqg)
414 return cfqg->service_trees[RT_WORKLOAD][ASYNC_WORKLOAD].count
415 + cfqg->service_trees[BE_WORKLOAD][ASYNC_WORKLOAD].count;
418 static void cfq_dispatch_insert(struct request_queue *, struct request *);
419 static struct cfq_queue *cfq_get_queue(struct cfq_data *, bool,
420 struct io_context *, gfp_t);
421 static struct cfq_io_context *cfq_cic_lookup(struct cfq_data *,
422 struct io_context *);
424 static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_context *cic,
427 return cic->cfqq[is_sync];
430 static inline void cic_set_cfqq(struct cfq_io_context *cic,
431 struct cfq_queue *cfqq, bool is_sync)
433 cic->cfqq[is_sync] = cfqq;
436 #define CIC_DEAD_KEY 1ul
437 #define CIC_DEAD_INDEX_SHIFT 1
439 static inline void *cfqd_dead_key(struct cfq_data *cfqd)
441 return (void *)(cfqd->cic_index << CIC_DEAD_INDEX_SHIFT | CIC_DEAD_KEY);
444 static inline struct cfq_data *cic_to_cfqd(struct cfq_io_context *cic)
446 struct cfq_data *cfqd = cic->key;
448 if (unlikely((unsigned long) cfqd & CIC_DEAD_KEY))
455 * We regard a request as SYNC, if it's either a read or has the SYNC bit
456 * set (in which case it could also be direct WRITE).
458 static inline bool cfq_bio_sync(struct bio *bio)
460 return bio_data_dir(bio) == READ || (bio->bi_rw & REQ_SYNC);
464 * scheduler run of queue, if there are requests pending and no one in the
465 * driver that will restart queueing
467 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
469 if (cfqd->busy_queues) {
470 cfq_log(cfqd, "schedule dispatch");
471 kblockd_schedule_work(cfqd->queue, &cfqd->unplug_work);
475 static int cfq_queue_empty(struct request_queue *q)
477 struct cfq_data *cfqd = q->elevator->elevator_data;
479 return !cfqd->rq_queued;
483 * Scale schedule slice based on io priority. Use the sync time slice only
484 * if a queue is marked sync and has sync io queued. A sync queue with async
485 * io only, should not get full sync slice length.
487 static inline int cfq_prio_slice(struct cfq_data *cfqd, bool sync,
490 const int base_slice = cfqd->cfq_slice[sync];
492 WARN_ON(prio >= IOPRIO_BE_NR);
494 return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
498 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
500 return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
503 static inline u64 cfq_scale_slice(unsigned long delta, struct cfq_group *cfqg)
505 u64 d = delta << CFQ_SERVICE_SHIFT;
507 d = d * BLKIO_WEIGHT_DEFAULT;
508 do_div(d, cfqg->weight);
512 static inline u64 max_vdisktime(u64 min_vdisktime, u64 vdisktime)
514 s64 delta = (s64)(vdisktime - min_vdisktime);
516 min_vdisktime = vdisktime;
518 return min_vdisktime;
521 static inline u64 min_vdisktime(u64 min_vdisktime, u64 vdisktime)
523 s64 delta = (s64)(vdisktime - min_vdisktime);
525 min_vdisktime = vdisktime;
527 return min_vdisktime;
530 static void update_min_vdisktime(struct cfq_rb_root *st)
532 u64 vdisktime = st->min_vdisktime;
533 struct cfq_group *cfqg;
536 cfqg = rb_entry_cfqg(st->active);
537 vdisktime = cfqg->vdisktime;
541 cfqg = rb_entry_cfqg(st->left);
542 vdisktime = min_vdisktime(vdisktime, cfqg->vdisktime);
545 st->min_vdisktime = max_vdisktime(st->min_vdisktime, vdisktime);
549 * get averaged number of queues of RT/BE priority.
550 * average is updated, with a formula that gives more weight to higher numbers,
551 * to quickly follows sudden increases and decrease slowly
554 static inline unsigned cfq_group_get_avg_queues(struct cfq_data *cfqd,
555 struct cfq_group *cfqg, bool rt)
557 unsigned min_q, max_q;
558 unsigned mult = cfq_hist_divisor - 1;
559 unsigned round = cfq_hist_divisor / 2;
560 unsigned busy = cfq_group_busy_queues_wl(rt, cfqd, cfqg);
562 min_q = min(cfqg->busy_queues_avg[rt], busy);
563 max_q = max(cfqg->busy_queues_avg[rt], busy);
564 cfqg->busy_queues_avg[rt] = (mult * max_q + min_q + round) /
566 return cfqg->busy_queues_avg[rt];
569 static inline unsigned
570 cfq_group_slice(struct cfq_data *cfqd, struct cfq_group *cfqg)
572 struct cfq_rb_root *st = &cfqd->grp_service_tree;
574 return cfq_target_latency * cfqg->weight / st->total_weight;
578 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
580 unsigned slice = cfq_prio_to_slice(cfqd, cfqq);
581 if (cfqd->cfq_latency) {
583 * interested queues (we consider only the ones with the same
584 * priority class in the cfq group)
586 unsigned iq = cfq_group_get_avg_queues(cfqd, cfqq->cfqg,
588 unsigned sync_slice = cfqd->cfq_slice[1];
589 unsigned expect_latency = sync_slice * iq;
590 unsigned group_slice = cfq_group_slice(cfqd, cfqq->cfqg);
592 if (expect_latency > group_slice) {
593 unsigned base_low_slice = 2 * cfqd->cfq_slice_idle;
594 /* scale low_slice according to IO priority
595 * and sync vs async */
597 min(slice, base_low_slice * slice / sync_slice);
598 /* the adapted slice value is scaled to fit all iqs
599 * into the target latency */
600 slice = max(slice * group_slice / expect_latency,
604 cfqq->slice_start = jiffies;
605 cfqq->slice_end = jiffies + slice;
606 cfqq->allocated_slice = slice;
607 cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
611 * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
612 * isn't valid until the first request from the dispatch is activated
613 * and the slice time set.
615 static inline bool cfq_slice_used(struct cfq_queue *cfqq)
617 if (cfq_cfqq_slice_new(cfqq))
619 if (time_before(jiffies, cfqq->slice_end))
626 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
627 * We choose the request that is closest to the head right now. Distance
628 * behind the head is penalized and only allowed to a certain extent.
630 static struct request *
631 cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2, sector_t last)
633 sector_t s1, s2, d1 = 0, d2 = 0;
634 unsigned long back_max;
635 #define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */
636 #define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */
637 unsigned wrap = 0; /* bit mask: requests behind the disk head? */
639 if (rq1 == NULL || rq1 == rq2)
644 if (rq_is_sync(rq1) && !rq_is_sync(rq2))
646 else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
648 if ((rq1->cmd_flags & REQ_META) && !(rq2->cmd_flags & REQ_META))
650 else if ((rq2->cmd_flags & REQ_META) &&
651 !(rq1->cmd_flags & REQ_META))
654 s1 = blk_rq_pos(rq1);
655 s2 = blk_rq_pos(rq2);
658 * by definition, 1KiB is 2 sectors
660 back_max = cfqd->cfq_back_max * 2;
663 * Strict one way elevator _except_ in the case where we allow
664 * short backward seeks which are biased as twice the cost of a
665 * similar forward seek.
669 else if (s1 + back_max >= last)
670 d1 = (last - s1) * cfqd->cfq_back_penalty;
672 wrap |= CFQ_RQ1_WRAP;
676 else if (s2 + back_max >= last)
677 d2 = (last - s2) * cfqd->cfq_back_penalty;
679 wrap |= CFQ_RQ2_WRAP;
681 /* Found required data */
684 * By doing switch() on the bit mask "wrap" we avoid having to
685 * check two variables for all permutations: --> faster!
688 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
704 case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
707 * Since both rqs are wrapped,
708 * start with the one that's further behind head
709 * (--> only *one* back seek required),
710 * since back seek takes more time than forward.
720 * The below is leftmost cache rbtree addon
722 static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
724 /* Service tree is empty */
729 root->left = rb_first(&root->rb);
732 return rb_entry(root->left, struct cfq_queue, rb_node);
737 static struct cfq_group *cfq_rb_first_group(struct cfq_rb_root *root)
740 root->left = rb_first(&root->rb);
743 return rb_entry_cfqg(root->left);
748 static void rb_erase_init(struct rb_node *n, struct rb_root *root)
754 static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
758 rb_erase_init(n, &root->rb);
763 * would be nice to take fifo expire time into account as well
765 static struct request *
766 cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
767 struct request *last)
769 struct rb_node *rbnext = rb_next(&last->rb_node);
770 struct rb_node *rbprev = rb_prev(&last->rb_node);
771 struct request *next = NULL, *prev = NULL;
773 BUG_ON(RB_EMPTY_NODE(&last->rb_node));
776 prev = rb_entry_rq(rbprev);
779 next = rb_entry_rq(rbnext);
781 rbnext = rb_first(&cfqq->sort_list);
782 if (rbnext && rbnext != &last->rb_node)
783 next = rb_entry_rq(rbnext);
786 return cfq_choose_req(cfqd, next, prev, blk_rq_pos(last));
789 static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
790 struct cfq_queue *cfqq)
793 * just an approximation, should be ok.
795 return (cfqq->cfqg->nr_cfqq - 1) * (cfq_prio_slice(cfqd, 1, 0) -
796 cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
800 cfqg_key(struct cfq_rb_root *st, struct cfq_group *cfqg)
802 return cfqg->vdisktime - st->min_vdisktime;
806 __cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
808 struct rb_node **node = &st->rb.rb_node;
809 struct rb_node *parent = NULL;
810 struct cfq_group *__cfqg;
811 s64 key = cfqg_key(st, cfqg);
814 while (*node != NULL) {
816 __cfqg = rb_entry_cfqg(parent);
818 if (key < cfqg_key(st, __cfqg))
819 node = &parent->rb_left;
821 node = &parent->rb_right;
827 st->left = &cfqg->rb_node;
829 rb_link_node(&cfqg->rb_node, parent, node);
830 rb_insert_color(&cfqg->rb_node, &st->rb);
834 cfq_group_service_tree_add(struct cfq_data *cfqd, struct cfq_group *cfqg)
836 struct cfq_rb_root *st = &cfqd->grp_service_tree;
837 struct cfq_group *__cfqg;
845 * Currently put the group at the end. Later implement something
846 * so that groups get lesser vtime based on their weights, so that
847 * if group does not loose all if it was not continously backlogged.
849 n = rb_last(&st->rb);
851 __cfqg = rb_entry_cfqg(n);
852 cfqg->vdisktime = __cfqg->vdisktime + CFQ_IDLE_DELAY;
854 cfqg->vdisktime = st->min_vdisktime;
856 __cfq_group_service_tree_add(st, cfqg);
858 st->total_weight += cfqg->weight;
862 cfq_group_service_tree_del(struct cfq_data *cfqd, struct cfq_group *cfqg)
864 struct cfq_rb_root *st = &cfqd->grp_service_tree;
866 if (st->active == &cfqg->rb_node)
869 BUG_ON(cfqg->nr_cfqq < 1);
872 /* If there are other cfq queues under this group, don't delete it */
876 cfq_log_cfqg(cfqd, cfqg, "del_from_rr group");
878 st->total_weight -= cfqg->weight;
879 if (!RB_EMPTY_NODE(&cfqg->rb_node))
880 cfq_rb_erase(&cfqg->rb_node, st);
881 cfqg->saved_workload_slice = 0;
882 cfq_blkiocg_update_dequeue_stats(&cfqg->blkg, 1);
885 static inline unsigned int cfq_cfqq_slice_usage(struct cfq_queue *cfqq)
887 unsigned int slice_used;
890 * Queue got expired before even a single request completed or
891 * got expired immediately after first request completion.
893 if (!cfqq->slice_start || cfqq->slice_start == jiffies) {
895 * Also charge the seek time incurred to the group, otherwise
896 * if there are mutiple queues in the group, each can dispatch
897 * a single request on seeky media and cause lots of seek time
898 * and group will never know it.
900 slice_used = max_t(unsigned, (jiffies - cfqq->dispatch_start),
903 slice_used = jiffies - cfqq->slice_start;
904 if (slice_used > cfqq->allocated_slice)
905 slice_used = cfqq->allocated_slice;
908 cfq_log_cfqq(cfqq->cfqd, cfqq, "sl_used=%u", slice_used);
912 static void cfq_group_served(struct cfq_data *cfqd, struct cfq_group *cfqg,
913 struct cfq_queue *cfqq)
915 struct cfq_rb_root *st = &cfqd->grp_service_tree;
916 unsigned int used_sl, charge_sl;
917 int nr_sync = cfqg->nr_cfqq - cfqg_busy_async_queues(cfqd, cfqg)
918 - cfqg->service_tree_idle.count;
921 used_sl = charge_sl = cfq_cfqq_slice_usage(cfqq);
923 if (!cfq_cfqq_sync(cfqq) && !nr_sync)
924 charge_sl = cfqq->allocated_slice;
926 /* Can't update vdisktime while group is on service tree */
927 cfq_rb_erase(&cfqg->rb_node, st);
928 cfqg->vdisktime += cfq_scale_slice(charge_sl, cfqg);
929 __cfq_group_service_tree_add(st, cfqg);
931 /* This group is being expired. Save the context */
932 if (time_after(cfqd->workload_expires, jiffies)) {
933 cfqg->saved_workload_slice = cfqd->workload_expires
935 cfqg->saved_workload = cfqd->serving_type;
936 cfqg->saved_serving_prio = cfqd->serving_prio;
938 cfqg->saved_workload_slice = 0;
940 cfq_log_cfqg(cfqd, cfqg, "served: vt=%llu min_vt=%llu", cfqg->vdisktime,
942 cfq_blkiocg_update_timeslice_used(&cfqg->blkg, used_sl);
943 cfq_blkiocg_set_start_empty_time(&cfqg->blkg);
946 #ifdef CONFIG_CFQ_GROUP_IOSCHED
947 static inline struct cfq_group *cfqg_of_blkg(struct blkio_group *blkg)
950 return container_of(blkg, struct cfq_group, blkg);
954 void cfq_update_blkio_group_weight(void *key, struct blkio_group *blkg,
957 cfqg_of_blkg(blkg)->weight = weight;
960 static struct cfq_group *
961 cfq_find_alloc_cfqg(struct cfq_data *cfqd, struct cgroup *cgroup, int create)
963 struct blkio_cgroup *blkcg = cgroup_to_blkio_cgroup(cgroup);
964 struct cfq_group *cfqg = NULL;
967 struct cfq_rb_root *st;
968 struct backing_dev_info *bdi = &cfqd->queue->backing_dev_info;
969 unsigned int major, minor;
971 cfqg = cfqg_of_blkg(blkiocg_lookup_group(blkcg, key));
972 if (cfqg && !cfqg->blkg.dev && bdi->dev && dev_name(bdi->dev)) {
973 sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
974 cfqg->blkg.dev = MKDEV(major, minor);
980 cfqg = kzalloc_node(sizeof(*cfqg), GFP_ATOMIC, cfqd->queue->node);
984 for_each_cfqg_st(cfqg, i, j, st)
986 RB_CLEAR_NODE(&cfqg->rb_node);
989 * Take the initial reference that will be released on destroy
990 * This can be thought of a joint reference by cgroup and
991 * elevator which will be dropped by either elevator exit
992 * or cgroup deletion path depending on who is exiting first.
994 atomic_set(&cfqg->ref, 1);
996 /* Add group onto cgroup list */
997 sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
998 cfq_blkiocg_add_blkio_group(blkcg, &cfqg->blkg, (void *)cfqd,
999 MKDEV(major, minor));
1000 cfqg->weight = blkcg_get_weight(blkcg, cfqg->blkg.dev);
1002 /* Add group on cfqd list */
1003 hlist_add_head(&cfqg->cfqd_node, &cfqd->cfqg_list);
1010 * Search for the cfq group current task belongs to. If create = 1, then also
1011 * create the cfq group if it does not exist. request_queue lock must be held.
1013 static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd, int create)
1015 struct cgroup *cgroup;
1016 struct cfq_group *cfqg = NULL;
1019 cgroup = task_cgroup(current, blkio_subsys_id);
1020 cfqg = cfq_find_alloc_cfqg(cfqd, cgroup, create);
1021 if (!cfqg && create)
1022 cfqg = &cfqd->root_group;
1027 static inline struct cfq_group *cfq_ref_get_cfqg(struct cfq_group *cfqg)
1029 atomic_inc(&cfqg->ref);
1033 static void cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg)
1035 /* Currently, all async queues are mapped to root group */
1036 if (!cfq_cfqq_sync(cfqq))
1037 cfqg = &cfqq->cfqd->root_group;
1040 /* cfqq reference on cfqg */
1041 atomic_inc(&cfqq->cfqg->ref);
1044 static void cfq_put_cfqg(struct cfq_group *cfqg)
1046 struct cfq_rb_root *st;
1049 BUG_ON(atomic_read(&cfqg->ref) <= 0);
1050 if (!atomic_dec_and_test(&cfqg->ref))
1052 for_each_cfqg_st(cfqg, i, j, st)
1053 BUG_ON(!RB_EMPTY_ROOT(&st->rb) || st->active != NULL);
1057 static void cfq_destroy_cfqg(struct cfq_data *cfqd, struct cfq_group *cfqg)
1059 /* Something wrong if we are trying to remove same group twice */
1060 BUG_ON(hlist_unhashed(&cfqg->cfqd_node));
1062 hlist_del_init(&cfqg->cfqd_node);
1065 * Put the reference taken at the time of creation so that when all
1066 * queues are gone, group can be destroyed.
1071 static void cfq_release_cfq_groups(struct cfq_data *cfqd)
1073 struct hlist_node *pos, *n;
1074 struct cfq_group *cfqg;
1076 hlist_for_each_entry_safe(cfqg, pos, n, &cfqd->cfqg_list, cfqd_node) {
1078 * If cgroup removal path got to blk_group first and removed
1079 * it from cgroup list, then it will take care of destroying
1082 if (!cfq_blkiocg_del_blkio_group(&cfqg->blkg))
1083 cfq_destroy_cfqg(cfqd, cfqg);
1088 * Blk cgroup controller notification saying that blkio_group object is being
1089 * delinked as associated cgroup object is going away. That also means that
1090 * no new IO will come in this group. So get rid of this group as soon as
1091 * any pending IO in the group is finished.
1093 * This function is called under rcu_read_lock(). key is the rcu protected
1094 * pointer. That means "key" is a valid cfq_data pointer as long as we are rcu
1097 * "key" was fetched from blkio_group under blkio_cgroup->lock. That means
1098 * it should not be NULL as even if elevator was exiting, cgroup deltion
1099 * path got to it first.
1101 void cfq_unlink_blkio_group(void *key, struct blkio_group *blkg)
1103 unsigned long flags;
1104 struct cfq_data *cfqd = key;
1106 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1107 cfq_destroy_cfqg(cfqd, cfqg_of_blkg(blkg));
1108 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1111 #else /* GROUP_IOSCHED */
1112 static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd, int create)
1114 return &cfqd->root_group;
1117 static inline struct cfq_group *cfq_ref_get_cfqg(struct cfq_group *cfqg)
1123 cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) {
1127 static void cfq_release_cfq_groups(struct cfq_data *cfqd) {}
1128 static inline void cfq_put_cfqg(struct cfq_group *cfqg) {}
1130 #endif /* GROUP_IOSCHED */
1133 * The cfqd->service_trees holds all pending cfq_queue's that have
1134 * requests waiting to be processed. It is sorted in the order that
1135 * we will service the queues.
1137 static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1140 struct rb_node **p, *parent;
1141 struct cfq_queue *__cfqq;
1142 unsigned long rb_key;
1143 struct cfq_rb_root *service_tree;
1146 int group_changed = 0;
1148 #ifdef CONFIG_CFQ_GROUP_IOSCHED
1149 if (!cfqd->cfq_group_isolation
1150 && cfqq_type(cfqq) == SYNC_NOIDLE_WORKLOAD
1151 && cfqq->cfqg && cfqq->cfqg != &cfqd->root_group) {
1152 /* Move this cfq to root group */
1153 cfq_log_cfqq(cfqd, cfqq, "moving to root group");
1154 if (!RB_EMPTY_NODE(&cfqq->rb_node))
1155 cfq_group_service_tree_del(cfqd, cfqq->cfqg);
1156 cfqq->orig_cfqg = cfqq->cfqg;
1157 cfqq->cfqg = &cfqd->root_group;
1158 atomic_inc(&cfqd->root_group.ref);
1160 } else if (!cfqd->cfq_group_isolation
1161 && cfqq_type(cfqq) == SYNC_WORKLOAD && cfqq->orig_cfqg) {
1162 /* cfqq is sequential now needs to go to its original group */
1163 BUG_ON(cfqq->cfqg != &cfqd->root_group);
1164 if (!RB_EMPTY_NODE(&cfqq->rb_node))
1165 cfq_group_service_tree_del(cfqd, cfqq->cfqg);
1166 cfq_put_cfqg(cfqq->cfqg);
1167 cfqq->cfqg = cfqq->orig_cfqg;
1168 cfqq->orig_cfqg = NULL;
1170 cfq_log_cfqq(cfqd, cfqq, "moved to origin group");
1174 service_tree = service_tree_for(cfqq->cfqg, cfqq_prio(cfqq),
1176 if (cfq_class_idle(cfqq)) {
1177 rb_key = CFQ_IDLE_DELAY;
1178 parent = rb_last(&service_tree->rb);
1179 if (parent && parent != &cfqq->rb_node) {
1180 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1181 rb_key += __cfqq->rb_key;
1184 } else if (!add_front) {
1186 * Get our rb key offset. Subtract any residual slice
1187 * value carried from last service. A negative resid
1188 * count indicates slice overrun, and this should position
1189 * the next service time further away in the tree.
1191 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
1192 rb_key -= cfqq->slice_resid;
1193 cfqq->slice_resid = 0;
1196 __cfqq = cfq_rb_first(service_tree);
1197 rb_key += __cfqq ? __cfqq->rb_key : jiffies;
1200 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1203 * same position, nothing more to do
1205 if (rb_key == cfqq->rb_key &&
1206 cfqq->service_tree == service_tree)
1209 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1210 cfqq->service_tree = NULL;
1215 cfqq->service_tree = service_tree;
1216 p = &service_tree->rb.rb_node;
1221 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1224 * sort by key, that represents service time.
1226 if (time_before(rb_key, __cfqq->rb_key))
1229 n = &(*p)->rb_right;
1237 service_tree->left = &cfqq->rb_node;
1239 cfqq->rb_key = rb_key;
1240 rb_link_node(&cfqq->rb_node, parent, p);
1241 rb_insert_color(&cfqq->rb_node, &service_tree->rb);
1242 service_tree->count++;
1243 if ((add_front || !new_cfqq) && !group_changed)
1245 cfq_group_service_tree_add(cfqd, cfqq->cfqg);
1248 static struct cfq_queue *
1249 cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
1250 sector_t sector, struct rb_node **ret_parent,
1251 struct rb_node ***rb_link)
1253 struct rb_node **p, *parent;
1254 struct cfq_queue *cfqq = NULL;
1262 cfqq = rb_entry(parent, struct cfq_queue, p_node);
1265 * Sort strictly based on sector. Smallest to the left,
1266 * largest to the right.
1268 if (sector > blk_rq_pos(cfqq->next_rq))
1269 n = &(*p)->rb_right;
1270 else if (sector < blk_rq_pos(cfqq->next_rq))
1278 *ret_parent = parent;
1284 static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1286 struct rb_node **p, *parent;
1287 struct cfq_queue *__cfqq;
1290 rb_erase(&cfqq->p_node, cfqq->p_root);
1291 cfqq->p_root = NULL;
1294 if (cfq_class_idle(cfqq))
1299 cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
1300 __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
1301 blk_rq_pos(cfqq->next_rq), &parent, &p);
1303 rb_link_node(&cfqq->p_node, parent, p);
1304 rb_insert_color(&cfqq->p_node, cfqq->p_root);
1306 cfqq->p_root = NULL;
1310 * Update cfqq's position in the service tree.
1312 static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1315 * Resorting requires the cfqq to be on the RR list already.
1317 if (cfq_cfqq_on_rr(cfqq)) {
1318 cfq_service_tree_add(cfqd, cfqq, 0);
1319 cfq_prio_tree_add(cfqd, cfqq);
1324 * add to busy list of queues for service, trying to be fair in ordering
1325 * the pending list according to last request service
1327 static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1329 cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
1330 BUG_ON(cfq_cfqq_on_rr(cfqq));
1331 cfq_mark_cfqq_on_rr(cfqq);
1332 cfqd->busy_queues++;
1334 cfq_resort_rr_list(cfqd, cfqq);
1338 * Called when the cfqq no longer has requests pending, remove it from
1341 static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1343 cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
1344 BUG_ON(!cfq_cfqq_on_rr(cfqq));
1345 cfq_clear_cfqq_on_rr(cfqq);
1347 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1348 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1349 cfqq->service_tree = NULL;
1352 rb_erase(&cfqq->p_node, cfqq->p_root);
1353 cfqq->p_root = NULL;
1356 cfq_group_service_tree_del(cfqd, cfqq->cfqg);
1357 BUG_ON(!cfqd->busy_queues);
1358 cfqd->busy_queues--;
1362 * rb tree support functions
1364 static void cfq_del_rq_rb(struct request *rq)
1366 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1367 const int sync = rq_is_sync(rq);
1369 BUG_ON(!cfqq->queued[sync]);
1370 cfqq->queued[sync]--;
1372 elv_rb_del(&cfqq->sort_list, rq);
1374 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) {
1376 * Queue will be deleted from service tree when we actually
1377 * expire it later. Right now just remove it from prio tree
1381 rb_erase(&cfqq->p_node, cfqq->p_root);
1382 cfqq->p_root = NULL;
1387 static void cfq_add_rq_rb(struct request *rq)
1389 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1390 struct cfq_data *cfqd = cfqq->cfqd;
1391 struct request *__alias, *prev;
1393 cfqq->queued[rq_is_sync(rq)]++;
1396 * looks a little odd, but the first insert might return an alias.
1397 * if that happens, put the alias on the dispatch list
1399 while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL)
1400 cfq_dispatch_insert(cfqd->queue, __alias);
1402 if (!cfq_cfqq_on_rr(cfqq))
1403 cfq_add_cfqq_rr(cfqd, cfqq);
1406 * check if this request is a better next-serve candidate
1408 prev = cfqq->next_rq;
1409 cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq, cfqd->last_position);
1412 * adjust priority tree position, if ->next_rq changes
1414 if (prev != cfqq->next_rq)
1415 cfq_prio_tree_add(cfqd, cfqq);
1417 BUG_ON(!cfqq->next_rq);
1420 static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
1422 elv_rb_del(&cfqq->sort_list, rq);
1423 cfqq->queued[rq_is_sync(rq)]--;
1424 cfq_blkiocg_update_io_remove_stats(&(RQ_CFQG(rq))->blkg,
1425 rq_data_dir(rq), rq_is_sync(rq));
1427 cfq_blkiocg_update_io_add_stats(&(RQ_CFQG(rq))->blkg,
1428 &cfqq->cfqd->serving_group->blkg, rq_data_dir(rq),
1432 static struct request *
1433 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
1435 struct task_struct *tsk = current;
1436 struct cfq_io_context *cic;
1437 struct cfq_queue *cfqq;
1439 cic = cfq_cic_lookup(cfqd, tsk->io_context);
1443 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
1445 sector_t sector = bio->bi_sector + bio_sectors(bio);
1447 return elv_rb_find(&cfqq->sort_list, sector);
1453 static void cfq_activate_request(struct request_queue *q, struct request *rq)
1455 struct cfq_data *cfqd = q->elevator->elevator_data;
1457 cfqd->rq_in_driver++;
1458 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
1459 cfqd->rq_in_driver);
1461 cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
1464 static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
1466 struct cfq_data *cfqd = q->elevator->elevator_data;
1468 WARN_ON(!cfqd->rq_in_driver);
1469 cfqd->rq_in_driver--;
1470 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
1471 cfqd->rq_in_driver);
1474 static void cfq_remove_request(struct request *rq)
1476 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1478 if (cfqq->next_rq == rq)
1479 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
1481 list_del_init(&rq->queuelist);
1484 cfqq->cfqd->rq_queued--;
1485 cfq_blkiocg_update_io_remove_stats(&(RQ_CFQG(rq))->blkg,
1486 rq_data_dir(rq), rq_is_sync(rq));
1487 if (rq->cmd_flags & REQ_META) {
1488 WARN_ON(!cfqq->meta_pending);
1489 cfqq->meta_pending--;
1493 static int cfq_merge(struct request_queue *q, struct request **req,
1496 struct cfq_data *cfqd = q->elevator->elevator_data;
1497 struct request *__rq;
1499 __rq = cfq_find_rq_fmerge(cfqd, bio);
1500 if (__rq && elv_rq_merge_ok(__rq, bio)) {
1502 return ELEVATOR_FRONT_MERGE;
1505 return ELEVATOR_NO_MERGE;
1508 static void cfq_merged_request(struct request_queue *q, struct request *req,
1511 if (type == ELEVATOR_FRONT_MERGE) {
1512 struct cfq_queue *cfqq = RQ_CFQQ(req);
1514 cfq_reposition_rq_rb(cfqq, req);
1518 static void cfq_bio_merged(struct request_queue *q, struct request *req,
1521 cfq_blkiocg_update_io_merged_stats(&(RQ_CFQG(req))->blkg,
1522 bio_data_dir(bio), cfq_bio_sync(bio));
1526 cfq_merged_requests(struct request_queue *q, struct request *rq,
1527 struct request *next)
1529 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1531 * reposition in fifo if next is older than rq
1533 if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
1534 time_before(rq_fifo_time(next), rq_fifo_time(rq))) {
1535 list_move(&rq->queuelist, &next->queuelist);
1536 rq_set_fifo_time(rq, rq_fifo_time(next));
1539 if (cfqq->next_rq == next)
1541 cfq_remove_request(next);
1542 cfq_blkiocg_update_io_merged_stats(&(RQ_CFQG(rq))->blkg,
1543 rq_data_dir(next), rq_is_sync(next));
1546 static int cfq_allow_merge(struct request_queue *q, struct request *rq,
1549 struct cfq_data *cfqd = q->elevator->elevator_data;
1550 struct cfq_io_context *cic;
1551 struct cfq_queue *cfqq;
1554 * Disallow merge of a sync bio into an async request.
1556 if (cfq_bio_sync(bio) && !rq_is_sync(rq))
1560 * Lookup the cfqq that this bio will be queued with. Allow
1561 * merge only if rq is queued there.
1563 cic = cfq_cic_lookup(cfqd, current->io_context);
1567 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
1568 return cfqq == RQ_CFQQ(rq);
1571 static inline void cfq_del_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1573 del_timer(&cfqd->idle_slice_timer);
1574 cfq_blkiocg_update_idle_time_stats(&cfqq->cfqg->blkg);
1577 static void __cfq_set_active_queue(struct cfq_data *cfqd,
1578 struct cfq_queue *cfqq)
1581 cfq_log_cfqq(cfqd, cfqq, "set_active wl_prio:%d wl_type:%d",
1582 cfqd->serving_prio, cfqd->serving_type);
1583 cfq_blkiocg_update_avg_queue_size_stats(&cfqq->cfqg->blkg);
1584 cfqq->slice_start = 0;
1585 cfqq->dispatch_start = jiffies;
1586 cfqq->allocated_slice = 0;
1587 cfqq->slice_end = 0;
1588 cfqq->slice_dispatch = 0;
1590 cfq_clear_cfqq_wait_request(cfqq);
1591 cfq_clear_cfqq_must_dispatch(cfqq);
1592 cfq_clear_cfqq_must_alloc_slice(cfqq);
1593 cfq_clear_cfqq_fifo_expire(cfqq);
1594 cfq_mark_cfqq_slice_new(cfqq);
1596 cfq_del_timer(cfqd, cfqq);
1599 cfqd->active_queue = cfqq;
1603 * current cfqq expired its slice (or was too idle), select new one
1606 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1609 cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
1611 if (cfq_cfqq_wait_request(cfqq))
1612 cfq_del_timer(cfqd, cfqq);
1614 cfq_clear_cfqq_wait_request(cfqq);
1615 cfq_clear_cfqq_wait_busy(cfqq);
1618 * If this cfqq is shared between multiple processes, check to
1619 * make sure that those processes are still issuing I/Os within
1620 * the mean seek distance. If not, it may be time to break the
1621 * queues apart again.
1623 if (cfq_cfqq_coop(cfqq) && CFQQ_SEEKY(cfqq))
1624 cfq_mark_cfqq_split_coop(cfqq);
1627 * store what was left of this slice, if the queue idled/timed out
1629 if (timed_out && !cfq_cfqq_slice_new(cfqq)) {
1630 cfqq->slice_resid = cfqq->slice_end - jiffies;
1631 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
1634 cfq_group_served(cfqd, cfqq->cfqg, cfqq);
1636 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
1637 cfq_del_cfqq_rr(cfqd, cfqq);
1639 cfq_resort_rr_list(cfqd, cfqq);
1641 if (cfqq == cfqd->active_queue)
1642 cfqd->active_queue = NULL;
1644 if (&cfqq->cfqg->rb_node == cfqd->grp_service_tree.active)
1645 cfqd->grp_service_tree.active = NULL;
1647 if (cfqd->active_cic) {
1648 put_io_context(cfqd->active_cic->ioc);
1649 cfqd->active_cic = NULL;
1653 static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out)
1655 struct cfq_queue *cfqq = cfqd->active_queue;
1658 __cfq_slice_expired(cfqd, cfqq, timed_out);
1662 * Get next queue for service. Unless we have a queue preemption,
1663 * we'll simply select the first cfqq in the service tree.
1665 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
1667 struct cfq_rb_root *service_tree =
1668 service_tree_for(cfqd->serving_group, cfqd->serving_prio,
1669 cfqd->serving_type);
1671 if (!cfqd->rq_queued)
1674 /* There is nothing to dispatch */
1677 if (RB_EMPTY_ROOT(&service_tree->rb))
1679 return cfq_rb_first(service_tree);
1682 static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd)
1684 struct cfq_group *cfqg;
1685 struct cfq_queue *cfqq;
1687 struct cfq_rb_root *st;
1689 if (!cfqd->rq_queued)
1692 cfqg = cfq_get_next_cfqg(cfqd);
1696 for_each_cfqg_st(cfqg, i, j, st)
1697 if ((cfqq = cfq_rb_first(st)) != NULL)
1703 * Get and set a new active queue for service.
1705 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
1706 struct cfq_queue *cfqq)
1709 cfqq = cfq_get_next_queue(cfqd);
1711 __cfq_set_active_queue(cfqd, cfqq);
1715 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
1718 if (blk_rq_pos(rq) >= cfqd->last_position)
1719 return blk_rq_pos(rq) - cfqd->last_position;
1721 return cfqd->last_position - blk_rq_pos(rq);
1724 static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1727 return cfq_dist_from_last(cfqd, rq) <= CFQQ_CLOSE_THR;
1730 static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
1731 struct cfq_queue *cur_cfqq)
1733 struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
1734 struct rb_node *parent, *node;
1735 struct cfq_queue *__cfqq;
1736 sector_t sector = cfqd->last_position;
1738 if (RB_EMPTY_ROOT(root))
1742 * First, if we find a request starting at the end of the last
1743 * request, choose it.
1745 __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
1750 * If the exact sector wasn't found, the parent of the NULL leaf
1751 * will contain the closest sector.
1753 __cfqq = rb_entry(parent, struct cfq_queue, p_node);
1754 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1757 if (blk_rq_pos(__cfqq->next_rq) < sector)
1758 node = rb_next(&__cfqq->p_node);
1760 node = rb_prev(&__cfqq->p_node);
1764 __cfqq = rb_entry(node, struct cfq_queue, p_node);
1765 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1773 * cur_cfqq - passed in so that we don't decide that the current queue is
1774 * closely cooperating with itself.
1776 * So, basically we're assuming that that cur_cfqq has dispatched at least
1777 * one request, and that cfqd->last_position reflects a position on the disk
1778 * associated with the I/O issued by cur_cfqq. I'm not sure this is a valid
1781 static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
1782 struct cfq_queue *cur_cfqq)
1784 struct cfq_queue *cfqq;
1786 if (cfq_class_idle(cur_cfqq))
1788 if (!cfq_cfqq_sync(cur_cfqq))
1790 if (CFQQ_SEEKY(cur_cfqq))
1794 * Don't search priority tree if it's the only queue in the group.
1796 if (cur_cfqq->cfqg->nr_cfqq == 1)
1800 * We should notice if some of the queues are cooperating, eg
1801 * working closely on the same area of the disk. In that case,
1802 * we can group them together and don't waste time idling.
1804 cfqq = cfqq_close(cfqd, cur_cfqq);
1808 /* If new queue belongs to different cfq_group, don't choose it */
1809 if (cur_cfqq->cfqg != cfqq->cfqg)
1813 * It only makes sense to merge sync queues.
1815 if (!cfq_cfqq_sync(cfqq))
1817 if (CFQQ_SEEKY(cfqq))
1821 * Do not merge queues of different priority classes
1823 if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq))
1830 * Determine whether we should enforce idle window for this queue.
1833 static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1835 enum wl_prio_t prio = cfqq_prio(cfqq);
1836 struct cfq_rb_root *service_tree = cfqq->service_tree;
1838 BUG_ON(!service_tree);
1839 BUG_ON(!service_tree->count);
1841 /* We never do for idle class queues. */
1842 if (prio == IDLE_WORKLOAD)
1845 /* We do for queues that were marked with idle window flag. */
1846 if (cfq_cfqq_idle_window(cfqq) &&
1847 !(blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag))
1851 * Otherwise, we do only if they are the last ones
1852 * in their service tree.
1854 if (service_tree->count == 1 && cfq_cfqq_sync(cfqq))
1856 cfq_log_cfqq(cfqd, cfqq, "Not idling. st->count:%d",
1857 service_tree->count);
1861 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
1863 struct cfq_queue *cfqq = cfqd->active_queue;
1864 struct cfq_io_context *cic;
1868 * SSD device without seek penalty, disable idling. But only do so
1869 * for devices that support queuing, otherwise we still have a problem
1870 * with sync vs async workloads.
1872 if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
1875 WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
1876 WARN_ON(cfq_cfqq_slice_new(cfqq));
1879 * idle is disabled, either manually or by past process history
1881 if (!cfqd->cfq_slice_idle || !cfq_should_idle(cfqd, cfqq))
1885 * still active requests from this queue, don't idle
1887 if (cfqq->dispatched)
1891 * task has exited, don't wait
1893 cic = cfqd->active_cic;
1894 if (!cic || !atomic_read(&cic->ioc->nr_tasks))
1898 * If our average think time is larger than the remaining time
1899 * slice, then don't idle. This avoids overrunning the allotted
1902 if (sample_valid(cic->ttime_samples) &&
1903 (cfqq->slice_end - jiffies < cic->ttime_mean)) {
1904 cfq_log_cfqq(cfqd, cfqq, "Not idling. think_time:%d",
1909 cfq_mark_cfqq_wait_request(cfqq);
1911 sl = cfqd->cfq_slice_idle;
1913 mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
1914 cfq_blkiocg_update_set_idle_time_stats(&cfqq->cfqg->blkg);
1915 cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu", sl);
1919 * Move request from internal lists to the request queue dispatch list.
1921 static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
1923 struct cfq_data *cfqd = q->elevator->elevator_data;
1924 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1926 cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
1928 cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
1929 cfq_remove_request(rq);
1931 elv_dispatch_sort(q, rq);
1933 cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]++;
1934 cfq_blkiocg_update_dispatch_stats(&cfqq->cfqg->blkg, blk_rq_bytes(rq),
1935 rq_data_dir(rq), rq_is_sync(rq));
1939 * return expired entry, or NULL to just start from scratch in rbtree
1941 static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
1943 struct request *rq = NULL;
1945 if (cfq_cfqq_fifo_expire(cfqq))
1948 cfq_mark_cfqq_fifo_expire(cfqq);
1950 if (list_empty(&cfqq->fifo))
1953 rq = rq_entry_fifo(cfqq->fifo.next);
1954 if (time_before(jiffies, rq_fifo_time(rq)))
1957 cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq);
1962 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1964 const int base_rq = cfqd->cfq_slice_async_rq;
1966 WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
1968 return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
1972 * Must be called with the queue_lock held.
1974 static int cfqq_process_refs(struct cfq_queue *cfqq)
1976 int process_refs, io_refs;
1978 io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE];
1979 process_refs = atomic_read(&cfqq->ref) - io_refs;
1980 BUG_ON(process_refs < 0);
1981 return process_refs;
1984 static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq)
1986 int process_refs, new_process_refs;
1987 struct cfq_queue *__cfqq;
1990 * If there are no process references on the new_cfqq, then it is
1991 * unsafe to follow the ->new_cfqq chain as other cfqq's in the
1992 * chain may have dropped their last reference (not just their
1993 * last process reference).
1995 if (!cfqq_process_refs(new_cfqq))
1998 /* Avoid a circular list and skip interim queue merges */
1999 while ((__cfqq = new_cfqq->new_cfqq)) {
2005 process_refs = cfqq_process_refs(cfqq);
2006 new_process_refs = cfqq_process_refs(new_cfqq);
2008 * If the process for the cfqq has gone away, there is no
2009 * sense in merging the queues.
2011 if (process_refs == 0 || new_process_refs == 0)
2015 * Merge in the direction of the lesser amount of work.
2017 if (new_process_refs >= process_refs) {
2018 cfqq->new_cfqq = new_cfqq;
2019 atomic_add(process_refs, &new_cfqq->ref);
2021 new_cfqq->new_cfqq = cfqq;
2022 atomic_add(new_process_refs, &cfqq->ref);
2026 static enum wl_type_t cfq_choose_wl(struct cfq_data *cfqd,
2027 struct cfq_group *cfqg, enum wl_prio_t prio)
2029 struct cfq_queue *queue;
2031 bool key_valid = false;
2032 unsigned long lowest_key = 0;
2033 enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD;
2035 for (i = 0; i <= SYNC_WORKLOAD; ++i) {
2036 /* select the one with lowest rb_key */
2037 queue = cfq_rb_first(service_tree_for(cfqg, prio, i));
2039 (!key_valid || time_before(queue->rb_key, lowest_key))) {
2040 lowest_key = queue->rb_key;
2049 static void choose_service_tree(struct cfq_data *cfqd, struct cfq_group *cfqg)
2053 struct cfq_rb_root *st;
2054 unsigned group_slice;
2057 cfqd->serving_prio = IDLE_WORKLOAD;
2058 cfqd->workload_expires = jiffies + 1;
2062 /* Choose next priority. RT > BE > IDLE */
2063 if (cfq_group_busy_queues_wl(RT_WORKLOAD, cfqd, cfqg))
2064 cfqd->serving_prio = RT_WORKLOAD;
2065 else if (cfq_group_busy_queues_wl(BE_WORKLOAD, cfqd, cfqg))
2066 cfqd->serving_prio = BE_WORKLOAD;
2068 cfqd->serving_prio = IDLE_WORKLOAD;
2069 cfqd->workload_expires = jiffies + 1;
2074 * For RT and BE, we have to choose also the type
2075 * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
2078 st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
2082 * check workload expiration, and that we still have other queues ready
2084 if (count && !time_after(jiffies, cfqd->workload_expires))
2087 /* otherwise select new workload type */
2088 cfqd->serving_type =
2089 cfq_choose_wl(cfqd, cfqg, cfqd->serving_prio);
2090 st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
2094 * the workload slice is computed as a fraction of target latency
2095 * proportional to the number of queues in that workload, over
2096 * all the queues in the same priority class
2098 group_slice = cfq_group_slice(cfqd, cfqg);
2100 slice = group_slice * count /
2101 max_t(unsigned, cfqg->busy_queues_avg[cfqd->serving_prio],
2102 cfq_group_busy_queues_wl(cfqd->serving_prio, cfqd, cfqg));
2104 if (cfqd->serving_type == ASYNC_WORKLOAD) {
2108 * Async queues are currently system wide. Just taking
2109 * proportion of queues with-in same group will lead to higher
2110 * async ratio system wide as generally root group is going
2111 * to have higher weight. A more accurate thing would be to
2112 * calculate system wide asnc/sync ratio.
2114 tmp = cfq_target_latency * cfqg_busy_async_queues(cfqd, cfqg);
2115 tmp = tmp/cfqd->busy_queues;
2116 slice = min_t(unsigned, slice, tmp);
2118 /* async workload slice is scaled down according to
2119 * the sync/async slice ratio. */
2120 slice = slice * cfqd->cfq_slice[0] / cfqd->cfq_slice[1];
2122 /* sync workload slice is at least 2 * cfq_slice_idle */
2123 slice = max(slice, 2 * cfqd->cfq_slice_idle);
2125 slice = max_t(unsigned, slice, CFQ_MIN_TT);
2126 cfq_log(cfqd, "workload slice:%d", slice);
2127 cfqd->workload_expires = jiffies + slice;
2130 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd)
2132 struct cfq_rb_root *st = &cfqd->grp_service_tree;
2133 struct cfq_group *cfqg;
2135 if (RB_EMPTY_ROOT(&st->rb))
2137 cfqg = cfq_rb_first_group(st);
2138 st->active = &cfqg->rb_node;
2139 update_min_vdisktime(st);
2143 static void cfq_choose_cfqg(struct cfq_data *cfqd)
2145 struct cfq_group *cfqg = cfq_get_next_cfqg(cfqd);
2147 cfqd->serving_group = cfqg;
2149 /* Restore the workload type data */
2150 if (cfqg->saved_workload_slice) {
2151 cfqd->workload_expires = jiffies + cfqg->saved_workload_slice;
2152 cfqd->serving_type = cfqg->saved_workload;
2153 cfqd->serving_prio = cfqg->saved_serving_prio;
2155 cfqd->workload_expires = jiffies - 1;
2157 choose_service_tree(cfqd, cfqg);
2161 * Select a queue for service. If we have a current active queue,
2162 * check whether to continue servicing it, or retrieve and set a new one.
2164 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
2166 struct cfq_queue *cfqq, *new_cfqq = NULL;
2168 cfqq = cfqd->active_queue;
2172 if (!cfqd->rq_queued)
2176 * We were waiting for group to get backlogged. Expire the queue
2178 if (cfq_cfqq_wait_busy(cfqq) && !RB_EMPTY_ROOT(&cfqq->sort_list))
2182 * The active queue has run out of time, expire it and select new.
2184 if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq)) {
2186 * If slice had not expired at the completion of last request
2187 * we might not have turned on wait_busy flag. Don't expire
2188 * the queue yet. Allow the group to get backlogged.
2190 * The very fact that we have used the slice, that means we
2191 * have been idling all along on this queue and it should be
2192 * ok to wait for this request to complete.
2194 if (cfqq->cfqg->nr_cfqq == 1 && RB_EMPTY_ROOT(&cfqq->sort_list)
2195 && cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2203 * The active queue has requests and isn't expired, allow it to
2206 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
2210 * If another queue has a request waiting within our mean seek
2211 * distance, let it run. The expire code will check for close
2212 * cooperators and put the close queue at the front of the service
2213 * tree. If possible, merge the expiring queue with the new cfqq.
2215 new_cfqq = cfq_close_cooperator(cfqd, cfqq);
2217 if (!cfqq->new_cfqq)
2218 cfq_setup_merge(cfqq, new_cfqq);
2223 * No requests pending. If the active queue still has requests in
2224 * flight or is idling for a new request, allow either of these
2225 * conditions to happen (or time out) before selecting a new queue.
2227 if (timer_pending(&cfqd->idle_slice_timer) ||
2228 (cfqq->dispatched && cfq_should_idle(cfqd, cfqq))) {
2234 cfq_slice_expired(cfqd, 0);
2237 * Current queue expired. Check if we have to switch to a new
2241 cfq_choose_cfqg(cfqd);
2243 cfqq = cfq_set_active_queue(cfqd, new_cfqq);
2248 static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
2252 while (cfqq->next_rq) {
2253 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
2257 BUG_ON(!list_empty(&cfqq->fifo));
2259 /* By default cfqq is not expired if it is empty. Do it explicitly */
2260 __cfq_slice_expired(cfqq->cfqd, cfqq, 0);
2265 * Drain our current requests. Used for barriers and when switching
2266 * io schedulers on-the-fly.
2268 static int cfq_forced_dispatch(struct cfq_data *cfqd)
2270 struct cfq_queue *cfqq;
2273 /* Expire the timeslice of the current active queue first */
2274 cfq_slice_expired(cfqd, 0);
2275 while ((cfqq = cfq_get_next_queue_forced(cfqd)) != NULL) {
2276 __cfq_set_active_queue(cfqd, cfqq);
2277 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
2280 BUG_ON(cfqd->busy_queues);
2282 cfq_log(cfqd, "forced_dispatch=%d", dispatched);
2286 static inline bool cfq_slice_used_soon(struct cfq_data *cfqd,
2287 struct cfq_queue *cfqq)
2289 /* the queue hasn't finished any request, can't estimate */
2290 if (cfq_cfqq_slice_new(cfqq))
2292 if (time_after(jiffies + cfqd->cfq_slice_idle * cfqq->dispatched,
2299 static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2301 unsigned int max_dispatch;
2304 * Drain async requests before we start sync IO
2306 if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_flight[BLK_RW_ASYNC])
2310 * If this is an async queue and we have sync IO in flight, let it wait
2312 if (cfqd->rq_in_flight[BLK_RW_SYNC] && !cfq_cfqq_sync(cfqq))
2315 max_dispatch = max_t(unsigned int, cfqd->cfq_quantum / 2, 1);
2316 if (cfq_class_idle(cfqq))
2320 * Does this cfqq already have too much IO in flight?
2322 if (cfqq->dispatched >= max_dispatch) {
2324 * idle queue must always only have a single IO in flight
2326 if (cfq_class_idle(cfqq))
2330 * We have other queues, don't allow more IO from this one
2332 if (cfqd->busy_queues > 1 && cfq_slice_used_soon(cfqd, cfqq))
2336 * Sole queue user, no limit
2338 if (cfqd->busy_queues == 1)
2342 * Normally we start throttling cfqq when cfq_quantum/2
2343 * requests have been dispatched. But we can drive
2344 * deeper queue depths at the beginning of slice
2345 * subjected to upper limit of cfq_quantum.
2347 max_dispatch = cfqd->cfq_quantum;
2351 * Async queues must wait a bit before being allowed dispatch.
2352 * We also ramp up the dispatch depth gradually for async IO,
2353 * based on the last sync IO we serviced
2355 if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
2356 unsigned long last_sync = jiffies - cfqd->last_delayed_sync;
2359 depth = last_sync / cfqd->cfq_slice[1];
2360 if (!depth && !cfqq->dispatched)
2362 if (depth < max_dispatch)
2363 max_dispatch = depth;
2367 * If we're below the current max, allow a dispatch
2369 return cfqq->dispatched < max_dispatch;
2373 * Dispatch a request from cfqq, moving them to the request queue
2376 static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2380 BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
2382 if (!cfq_may_dispatch(cfqd, cfqq))
2386 * follow expired path, else get first next available
2388 rq = cfq_check_fifo(cfqq);
2393 * insert request into driver dispatch list
2395 cfq_dispatch_insert(cfqd->queue, rq);
2397 if (!cfqd->active_cic) {
2398 struct cfq_io_context *cic = RQ_CIC(rq);
2400 atomic_long_inc(&cic->ioc->refcount);
2401 cfqd->active_cic = cic;
2408 * Find the cfqq that we need to service and move a request from that to the
2411 static int cfq_dispatch_requests(struct request_queue *q, int force)
2413 struct cfq_data *cfqd = q->elevator->elevator_data;
2414 struct cfq_queue *cfqq;
2416 if (!cfqd->busy_queues)
2419 if (unlikely(force))
2420 return cfq_forced_dispatch(cfqd);
2422 cfqq = cfq_select_queue(cfqd);
2427 * Dispatch a request from this cfqq, if it is allowed
2429 if (!cfq_dispatch_request(cfqd, cfqq))
2432 cfqq->slice_dispatch++;
2433 cfq_clear_cfqq_must_dispatch(cfqq);
2436 * expire an async queue immediately if it has used up its slice. idle
2437 * queue always expire after 1 dispatch round.
2439 if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
2440 cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
2441 cfq_class_idle(cfqq))) {
2442 cfqq->slice_end = jiffies + 1;
2443 cfq_slice_expired(cfqd, 0);
2446 cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
2451 * task holds one reference to the queue, dropped when task exits. each rq
2452 * in-flight on this queue also holds a reference, dropped when rq is freed.
2454 * Each cfq queue took a reference on the parent group. Drop it now.
2455 * queue lock must be held here.
2457 static void cfq_put_queue(struct cfq_queue *cfqq)
2459 struct cfq_data *cfqd = cfqq->cfqd;
2460 struct cfq_group *cfqg, *orig_cfqg;
2462 BUG_ON(atomic_read(&cfqq->ref) <= 0);
2464 if (!atomic_dec_and_test(&cfqq->ref))
2467 cfq_log_cfqq(cfqd, cfqq, "put_queue");
2468 BUG_ON(rb_first(&cfqq->sort_list));
2469 BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
2471 orig_cfqg = cfqq->orig_cfqg;
2473 if (unlikely(cfqd->active_queue == cfqq)) {
2474 __cfq_slice_expired(cfqd, cfqq, 0);
2475 cfq_schedule_dispatch(cfqd);
2478 BUG_ON(cfq_cfqq_on_rr(cfqq));
2479 kmem_cache_free(cfq_pool, cfqq);
2482 cfq_put_cfqg(orig_cfqg);
2486 * Must always be called with the rcu_read_lock() held
2489 __call_for_each_cic(struct io_context *ioc,
2490 void (*func)(struct io_context *, struct cfq_io_context *))
2492 struct cfq_io_context *cic;
2493 struct hlist_node *n;
2495 hlist_for_each_entry_rcu(cic, n, &ioc->cic_list, cic_list)
2500 * Call func for each cic attached to this ioc.
2503 call_for_each_cic(struct io_context *ioc,
2504 void (*func)(struct io_context *, struct cfq_io_context *))
2507 __call_for_each_cic(ioc, func);
2511 static void cfq_cic_free_rcu(struct rcu_head *head)
2513 struct cfq_io_context *cic;
2515 cic = container_of(head, struct cfq_io_context, rcu_head);
2517 kmem_cache_free(cfq_ioc_pool, cic);
2518 elv_ioc_count_dec(cfq_ioc_count);
2522 * CFQ scheduler is exiting, grab exit lock and check
2523 * the pending io context count. If it hits zero,
2524 * complete ioc_gone and set it back to NULL
2526 spin_lock(&ioc_gone_lock);
2527 if (ioc_gone && !elv_ioc_count_read(cfq_ioc_count)) {
2531 spin_unlock(&ioc_gone_lock);
2535 static void cfq_cic_free(struct cfq_io_context *cic)
2537 call_rcu(&cic->rcu_head, cfq_cic_free_rcu);
2540 static void cic_free_func(struct io_context *ioc, struct cfq_io_context *cic)
2542 unsigned long flags;
2543 unsigned long dead_key = (unsigned long) cic->key;
2545 BUG_ON(!(dead_key & CIC_DEAD_KEY));
2547 spin_lock_irqsave(&ioc->lock, flags);
2548 radix_tree_delete(&ioc->radix_root, dead_key >> CIC_DEAD_INDEX_SHIFT);
2549 hlist_del_rcu(&cic->cic_list);
2550 spin_unlock_irqrestore(&ioc->lock, flags);
2556 * Must be called with rcu_read_lock() held or preemption otherwise disabled.
2557 * Only two callers of this - ->dtor() which is called with the rcu_read_lock(),
2558 * and ->trim() which is called with the task lock held
2560 static void cfq_free_io_context(struct io_context *ioc)
2563 * ioc->refcount is zero here, or we are called from elv_unregister(),
2564 * so no more cic's are allowed to be linked into this ioc. So it
2565 * should be ok to iterate over the known list, we will see all cic's
2566 * since no new ones are added.
2568 __call_for_each_cic(ioc, cic_free_func);
2571 static void cfq_put_cooperator(struct cfq_queue *cfqq)
2573 struct cfq_queue *__cfqq, *next;
2576 * If this queue was scheduled to merge with another queue, be
2577 * sure to drop the reference taken on that queue (and others in
2578 * the merge chain). See cfq_setup_merge and cfq_merge_cfqqs.
2580 __cfqq = cfqq->new_cfqq;
2582 if (__cfqq == cfqq) {
2583 WARN(1, "cfqq->new_cfqq loop detected\n");
2586 next = __cfqq->new_cfqq;
2587 cfq_put_queue(__cfqq);
2592 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2594 if (unlikely(cfqq == cfqd->active_queue)) {
2595 __cfq_slice_expired(cfqd, cfqq, 0);
2596 cfq_schedule_dispatch(cfqd);
2599 cfq_put_cooperator(cfqq);
2601 cfq_put_queue(cfqq);
2604 static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
2605 struct cfq_io_context *cic)
2607 struct io_context *ioc = cic->ioc;
2609 list_del_init(&cic->queue_list);
2612 * Make sure dead mark is seen for dead queues
2615 cic->key = cfqd_dead_key(cfqd);
2617 if (ioc->ioc_data == cic)
2618 rcu_assign_pointer(ioc->ioc_data, NULL);
2620 if (cic->cfqq[BLK_RW_ASYNC]) {
2621 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]);
2622 cic->cfqq[BLK_RW_ASYNC] = NULL;
2625 if (cic->cfqq[BLK_RW_SYNC]) {
2626 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]);
2627 cic->cfqq[BLK_RW_SYNC] = NULL;
2631 static void cfq_exit_single_io_context(struct io_context *ioc,
2632 struct cfq_io_context *cic)
2634 struct cfq_data *cfqd = cic_to_cfqd(cic);
2637 struct request_queue *q = cfqd->queue;
2638 unsigned long flags;
2640 spin_lock_irqsave(q->queue_lock, flags);
2643 * Ensure we get a fresh copy of the ->key to prevent
2644 * race between exiting task and queue
2646 smp_read_barrier_depends();
2647 if (cic->key == cfqd)
2648 __cfq_exit_single_io_context(cfqd, cic);
2650 spin_unlock_irqrestore(q->queue_lock, flags);
2655 * The process that ioc belongs to has exited, we need to clean up
2656 * and put the internal structures we have that belongs to that process.
2658 static void cfq_exit_io_context(struct io_context *ioc)
2660 call_for_each_cic(ioc, cfq_exit_single_io_context);
2663 static struct cfq_io_context *
2664 cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
2666 struct cfq_io_context *cic;
2668 cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO,
2671 cic->last_end_request = jiffies;
2672 INIT_LIST_HEAD(&cic->queue_list);
2673 INIT_HLIST_NODE(&cic->cic_list);
2674 cic->dtor = cfq_free_io_context;
2675 cic->exit = cfq_exit_io_context;
2676 elv_ioc_count_inc(cfq_ioc_count);
2682 static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc)
2684 struct task_struct *tsk = current;
2687 if (!cfq_cfqq_prio_changed(cfqq))
2690 ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio);
2691 switch (ioprio_class) {
2693 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
2694 case IOPRIO_CLASS_NONE:
2696 * no prio set, inherit CPU scheduling settings
2698 cfqq->ioprio = task_nice_ioprio(tsk);
2699 cfqq->ioprio_class = task_nice_ioclass(tsk);
2701 case IOPRIO_CLASS_RT:
2702 cfqq->ioprio = task_ioprio(ioc);
2703 cfqq->ioprio_class = IOPRIO_CLASS_RT;
2705 case IOPRIO_CLASS_BE:
2706 cfqq->ioprio = task_ioprio(ioc);
2707 cfqq->ioprio_class = IOPRIO_CLASS_BE;
2709 case IOPRIO_CLASS_IDLE:
2710 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
2712 cfq_clear_cfqq_idle_window(cfqq);
2717 * keep track of original prio settings in case we have to temporarily
2718 * elevate the priority of this queue
2720 cfqq->org_ioprio = cfqq->ioprio;
2721 cfqq->org_ioprio_class = cfqq->ioprio_class;
2722 cfq_clear_cfqq_prio_changed(cfqq);
2725 static void changed_ioprio(struct io_context *ioc, struct cfq_io_context *cic)
2727 struct cfq_data *cfqd = cic_to_cfqd(cic);
2728 struct cfq_queue *cfqq;
2729 unsigned long flags;
2731 if (unlikely(!cfqd))
2734 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2736 cfqq = cic->cfqq[BLK_RW_ASYNC];
2738 struct cfq_queue *new_cfqq;
2739 new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic->ioc,
2742 cic->cfqq[BLK_RW_ASYNC] = new_cfqq;
2743 cfq_put_queue(cfqq);
2747 cfqq = cic->cfqq[BLK_RW_SYNC];
2749 cfq_mark_cfqq_prio_changed(cfqq);
2751 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2754 static void cfq_ioc_set_ioprio(struct io_context *ioc)
2756 call_for_each_cic(ioc, changed_ioprio);
2757 ioc->ioprio_changed = 0;
2760 static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2761 pid_t pid, bool is_sync)
2763 RB_CLEAR_NODE(&cfqq->rb_node);
2764 RB_CLEAR_NODE(&cfqq->p_node);
2765 INIT_LIST_HEAD(&cfqq->fifo);
2767 atomic_set(&cfqq->ref, 0);
2770 cfq_mark_cfqq_prio_changed(cfqq);
2773 if (!cfq_class_idle(cfqq))
2774 cfq_mark_cfqq_idle_window(cfqq);
2775 cfq_mark_cfqq_sync(cfqq);
2780 #ifdef CONFIG_CFQ_GROUP_IOSCHED
2781 static void changed_cgroup(struct io_context *ioc, struct cfq_io_context *cic)
2783 struct cfq_queue *sync_cfqq = cic_to_cfqq(cic, 1);
2784 struct cfq_data *cfqd = cic_to_cfqd(cic);
2785 unsigned long flags;
2786 struct request_queue *q;
2788 if (unlikely(!cfqd))
2793 spin_lock_irqsave(q->queue_lock, flags);
2797 * Drop reference to sync queue. A new sync queue will be
2798 * assigned in new group upon arrival of a fresh request.
2800 cfq_log_cfqq(cfqd, sync_cfqq, "changed cgroup");
2801 cic_set_cfqq(cic, NULL, 1);
2802 cfq_put_queue(sync_cfqq);
2805 spin_unlock_irqrestore(q->queue_lock, flags);
2808 static void cfq_ioc_set_cgroup(struct io_context *ioc)
2810 call_for_each_cic(ioc, changed_cgroup);
2811 ioc->cgroup_changed = 0;
2813 #endif /* CONFIG_CFQ_GROUP_IOSCHED */
2815 static struct cfq_queue *
2816 cfq_find_alloc_queue(struct cfq_data *cfqd, bool is_sync,
2817 struct io_context *ioc, gfp_t gfp_mask)
2819 struct cfq_queue *cfqq, *new_cfqq = NULL;
2820 struct cfq_io_context *cic;
2821 struct cfq_group *cfqg;
2824 cfqg = cfq_get_cfqg(cfqd, 1);
2825 cic = cfq_cic_lookup(cfqd, ioc);
2826 /* cic always exists here */
2827 cfqq = cic_to_cfqq(cic, is_sync);
2830 * Always try a new alloc if we fell back to the OOM cfqq
2831 * originally, since it should just be a temporary situation.
2833 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
2838 } else if (gfp_mask & __GFP_WAIT) {
2839 spin_unlock_irq(cfqd->queue->queue_lock);
2840 new_cfqq = kmem_cache_alloc_node(cfq_pool,
2841 gfp_mask | __GFP_ZERO,
2843 spin_lock_irq(cfqd->queue->queue_lock);
2847 cfqq = kmem_cache_alloc_node(cfq_pool,
2848 gfp_mask | __GFP_ZERO,
2853 cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
2854 cfq_init_prio_data(cfqq, ioc);
2855 cfq_link_cfqq_cfqg(cfqq, cfqg);
2856 cfq_log_cfqq(cfqd, cfqq, "alloced");
2858 cfqq = &cfqd->oom_cfqq;
2862 kmem_cache_free(cfq_pool, new_cfqq);
2867 static struct cfq_queue **
2868 cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
2870 switch (ioprio_class) {
2871 case IOPRIO_CLASS_RT:
2872 return &cfqd->async_cfqq[0][ioprio];
2873 case IOPRIO_CLASS_BE:
2874 return &cfqd->async_cfqq[1][ioprio];
2875 case IOPRIO_CLASS_IDLE:
2876 return &cfqd->async_idle_cfqq;
2882 static struct cfq_queue *
2883 cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct io_context *ioc,
2886 const int ioprio = task_ioprio(ioc);
2887 const int ioprio_class = task_ioprio_class(ioc);
2888 struct cfq_queue **async_cfqq = NULL;
2889 struct cfq_queue *cfqq = NULL;
2892 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
2897 cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask);
2900 * pin the queue now that it's allocated, scheduler exit will prune it
2902 if (!is_sync && !(*async_cfqq)) {
2903 atomic_inc(&cfqq->ref);
2907 atomic_inc(&cfqq->ref);
2912 * We drop cfq io contexts lazily, so we may find a dead one.
2915 cfq_drop_dead_cic(struct cfq_data *cfqd, struct io_context *ioc,
2916 struct cfq_io_context *cic)
2918 unsigned long flags;
2920 WARN_ON(!list_empty(&cic->queue_list));
2921 BUG_ON(cic->key != cfqd_dead_key(cfqd));
2923 spin_lock_irqsave(&ioc->lock, flags);
2925 BUG_ON(ioc->ioc_data == cic);
2927 radix_tree_delete(&ioc->radix_root, cfqd->cic_index);
2928 hlist_del_rcu(&cic->cic_list);
2929 spin_unlock_irqrestore(&ioc->lock, flags);
2934 static struct cfq_io_context *
2935 cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc)
2937 struct cfq_io_context *cic;
2938 unsigned long flags;
2946 * we maintain a last-hit cache, to avoid browsing over the tree
2948 cic = rcu_dereference(ioc->ioc_data);
2949 if (cic && cic->key == cfqd) {
2955 cic = radix_tree_lookup(&ioc->radix_root, cfqd->cic_index);
2959 if (unlikely(cic->key != cfqd)) {
2960 cfq_drop_dead_cic(cfqd, ioc, cic);
2965 spin_lock_irqsave(&ioc->lock, flags);
2966 rcu_assign_pointer(ioc->ioc_data, cic);
2967 spin_unlock_irqrestore(&ioc->lock, flags);
2975 * Add cic into ioc, using cfqd as the search key. This enables us to lookup
2976 * the process specific cfq io context when entered from the block layer.
2977 * Also adds the cic to a per-cfqd list, used when this queue is removed.
2979 static int cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
2980 struct cfq_io_context *cic, gfp_t gfp_mask)
2982 unsigned long flags;
2985 ret = radix_tree_preload(gfp_mask);
2990 spin_lock_irqsave(&ioc->lock, flags);
2991 ret = radix_tree_insert(&ioc->radix_root,
2992 cfqd->cic_index, cic);
2994 hlist_add_head_rcu(&cic->cic_list, &ioc->cic_list);
2995 spin_unlock_irqrestore(&ioc->lock, flags);
2997 radix_tree_preload_end();
3000 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
3001 list_add(&cic->queue_list, &cfqd->cic_list);
3002 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
3007 printk(KERN_ERR "cfq: cic link failed!\n");
3013 * Setup general io context and cfq io context. There can be several cfq
3014 * io contexts per general io context, if this process is doing io to more
3015 * than one device managed by cfq.
3017 static struct cfq_io_context *
3018 cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
3020 struct io_context *ioc = NULL;
3021 struct cfq_io_context *cic;
3023 might_sleep_if(gfp_mask & __GFP_WAIT);
3025 ioc = get_io_context(gfp_mask, cfqd->queue->node);
3029 cic = cfq_cic_lookup(cfqd, ioc);
3033 cic = cfq_alloc_io_context(cfqd, gfp_mask);
3037 if (cfq_cic_link(cfqd, ioc, cic, gfp_mask))
3041 smp_read_barrier_depends();
3042 if (unlikely(ioc->ioprio_changed))
3043 cfq_ioc_set_ioprio(ioc);
3045 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3046 if (unlikely(ioc->cgroup_changed))
3047 cfq_ioc_set_cgroup(ioc);
3053 put_io_context(ioc);
3058 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
3060 unsigned long elapsed = jiffies - cic->last_end_request;
3061 unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
3063 cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
3064 cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
3065 cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
3069 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3073 sector_t n_sec = blk_rq_sectors(rq);
3074 if (cfqq->last_request_pos) {
3075 if (cfqq->last_request_pos < blk_rq_pos(rq))
3076 sdist = blk_rq_pos(rq) - cfqq->last_request_pos;
3078 sdist = cfqq->last_request_pos - blk_rq_pos(rq);
3081 cfqq->seek_history <<= 1;
3082 if (blk_queue_nonrot(cfqd->queue))
3083 cfqq->seek_history |= (n_sec < CFQQ_SECT_THR_NONROT);
3085 cfqq->seek_history |= (sdist > CFQQ_SEEK_THR);
3089 * Disable idle window if the process thinks too long or seeks so much that
3093 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3094 struct cfq_io_context *cic)
3096 int old_idle, enable_idle;
3099 * Don't idle for async or idle io prio class
3101 if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
3104 enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
3106 if (cfqq->queued[0] + cfqq->queued[1] >= 4)
3107 cfq_mark_cfqq_deep(cfqq);
3109 if (cfqq->next_rq && (cfqq->next_rq->cmd_flags & REQ_NOIDLE))
3111 else if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle ||
3112 (!cfq_cfqq_deep(cfqq) && CFQQ_SEEKY(cfqq)))
3114 else if (sample_valid(cic->ttime_samples)) {
3115 if (cic->ttime_mean > cfqd->cfq_slice_idle)
3121 if (old_idle != enable_idle) {
3122 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
3124 cfq_mark_cfqq_idle_window(cfqq);
3126 cfq_clear_cfqq_idle_window(cfqq);
3131 * Check if new_cfqq should preempt the currently active queue. Return 0 for
3132 * no or if we aren't sure, a 1 will cause a preempt.
3135 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
3138 struct cfq_queue *cfqq;
3140 cfqq = cfqd->active_queue;
3144 if (cfq_class_idle(new_cfqq))
3147 if (cfq_class_idle(cfqq))
3151 * Don't allow a non-RT request to preempt an ongoing RT cfqq timeslice.
3153 if (cfq_class_rt(cfqq) && !cfq_class_rt(new_cfqq))
3157 * if the new request is sync, but the currently running queue is
3158 * not, let the sync request have priority.
3160 if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
3163 if (new_cfqq->cfqg != cfqq->cfqg)
3166 if (cfq_slice_used(cfqq))
3169 /* Allow preemption only if we are idling on sync-noidle tree */
3170 if (cfqd->serving_type == SYNC_NOIDLE_WORKLOAD &&
3171 cfqq_type(new_cfqq) == SYNC_NOIDLE_WORKLOAD &&
3172 new_cfqq->service_tree->count == 2 &&
3173 RB_EMPTY_ROOT(&cfqq->sort_list))
3177 * So both queues are sync. Let the new request get disk time if
3178 * it's a metadata request and the current queue is doing regular IO.
3180 if ((rq->cmd_flags & REQ_META) && !cfqq->meta_pending)
3184 * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
3186 if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
3189 if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
3193 * if this request is as-good as one we would expect from the
3194 * current cfqq, let it preempt
3196 if (cfq_rq_close(cfqd, cfqq, rq))
3203 * cfqq preempts the active queue. if we allowed preempt with no slice left,
3204 * let it have half of its nominal slice.
3206 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3208 cfq_log_cfqq(cfqd, cfqq, "preempt");
3209 cfq_slice_expired(cfqd, 1);
3212 * Put the new queue at the front of the of the current list,
3213 * so we know that it will be selected next.
3215 BUG_ON(!cfq_cfqq_on_rr(cfqq));
3217 cfq_service_tree_add(cfqd, cfqq, 1);
3219 cfqq->slice_end = 0;
3220 cfq_mark_cfqq_slice_new(cfqq);
3224 * Called when a new fs request (rq) is added (to cfqq). Check if there's
3225 * something we should do about it
3228 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3231 struct cfq_io_context *cic = RQ_CIC(rq);
3234 if (rq->cmd_flags & REQ_META)
3235 cfqq->meta_pending++;
3237 cfq_update_io_thinktime(cfqd, cic);
3238 cfq_update_io_seektime(cfqd, cfqq, rq);
3239 cfq_update_idle_window(cfqd, cfqq, cic);
3241 cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
3243 if (cfqq == cfqd->active_queue) {
3245 * Remember that we saw a request from this process, but
3246 * don't start queuing just yet. Otherwise we risk seeing lots
3247 * of tiny requests, because we disrupt the normal plugging
3248 * and merging. If the request is already larger than a single
3249 * page, let it rip immediately. For that case we assume that
3250 * merging is already done. Ditto for a busy system that
3251 * has other work pending, don't risk delaying until the
3252 * idle timer unplug to continue working.
3254 if (cfq_cfqq_wait_request(cfqq)) {
3255 if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
3256 cfqd->busy_queues > 1) {
3257 cfq_del_timer(cfqd, cfqq);
3258 cfq_clear_cfqq_wait_request(cfqq);
3259 __blk_run_queue(cfqd->queue);
3261 cfq_blkiocg_update_idle_time_stats(
3263 cfq_mark_cfqq_must_dispatch(cfqq);
3266 } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
3268 * not the active queue - expire current slice if it is
3269 * idle and has expired it's mean thinktime or this new queue
3270 * has some old slice time left and is of higher priority or
3271 * this new queue is RT and the current one is BE
3273 cfq_preempt_queue(cfqd, cfqq);
3274 __blk_run_queue(cfqd->queue);
3278 static void cfq_insert_request(struct request_queue *q, struct request *rq)
3280 struct cfq_data *cfqd = q->elevator->elevator_data;
3281 struct cfq_queue *cfqq = RQ_CFQQ(rq);
3283 cfq_log_cfqq(cfqd, cfqq, "insert_request");
3284 cfq_init_prio_data(cfqq, RQ_CIC(rq)->ioc);
3286 rq_set_fifo_time(rq, jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)]);
3287 list_add_tail(&rq->queuelist, &cfqq->fifo);
3289 cfq_blkiocg_update_io_add_stats(&(RQ_CFQG(rq))->blkg,
3290 &cfqd->serving_group->blkg, rq_data_dir(rq),
3292 cfq_rq_enqueued(cfqd, cfqq, rq);
3296 * Update hw_tag based on peak queue depth over 50 samples under
3299 static void cfq_update_hw_tag(struct cfq_data *cfqd)
3301 struct cfq_queue *cfqq = cfqd->active_queue;
3303 if (cfqd->rq_in_driver > cfqd->hw_tag_est_depth)
3304 cfqd->hw_tag_est_depth = cfqd->rq_in_driver;
3306 if (cfqd->hw_tag == 1)
3309 if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
3310 cfqd->rq_in_driver <= CFQ_HW_QUEUE_MIN)
3314 * If active queue hasn't enough requests and can idle, cfq might not
3315 * dispatch sufficient requests to hardware. Don't zero hw_tag in this
3318 if (cfqq && cfq_cfqq_idle_window(cfqq) &&
3319 cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] <
3320 CFQ_HW_QUEUE_MIN && cfqd->rq_in_driver < CFQ_HW_QUEUE_MIN)
3323 if (cfqd->hw_tag_samples++ < 50)
3326 if (cfqd->hw_tag_est_depth >= CFQ_HW_QUEUE_MIN)
3332 static bool cfq_should_wait_busy(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3334 struct cfq_io_context *cic = cfqd->active_cic;
3336 /* If there are other queues in the group, don't wait */
3337 if (cfqq->cfqg->nr_cfqq > 1)
3340 if (cfq_slice_used(cfqq))
3343 /* if slice left is less than think time, wait busy */
3344 if (cic && sample_valid(cic->ttime_samples)
3345 && (cfqq->slice_end - jiffies < cic->ttime_mean))
3349 * If think times is less than a jiffy than ttime_mean=0 and above
3350 * will not be true. It might happen that slice has not expired yet
3351 * but will expire soon (4-5 ns) during select_queue(). To cover the
3352 * case where think time is less than a jiffy, mark the queue wait
3353 * busy if only 1 jiffy is left in the slice.
3355 if (cfqq->slice_end - jiffies == 1)
3361 static void cfq_completed_request(struct request_queue *q, struct request *rq)
3363 struct cfq_queue *cfqq = RQ_CFQQ(rq);
3364 struct cfq_data *cfqd = cfqq->cfqd;
3365 const int sync = rq_is_sync(rq);
3369 cfq_log_cfqq(cfqd, cfqq, "complete rqnoidle %d",
3370 !!(rq->cmd_flags & REQ_NOIDLE));
3372 cfq_update_hw_tag(cfqd);
3374 WARN_ON(!cfqd->rq_in_driver);
3375 WARN_ON(!cfqq->dispatched);
3376 cfqd->rq_in_driver--;
3378 cfq_blkiocg_update_completion_stats(&cfqq->cfqg->blkg,
3379 rq_start_time_ns(rq), rq_io_start_time_ns(rq),
3380 rq_data_dir(rq), rq_is_sync(rq));
3382 cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]--;
3385 RQ_CIC(rq)->last_end_request = now;
3386 if (!time_after(rq->start_time + cfqd->cfq_fifo_expire[1], now))
3387 cfqd->last_delayed_sync = now;
3391 * If this is the active queue, check if it needs to be expired,
3392 * or if we want to idle in case it has no pending requests.
3394 if (cfqd->active_queue == cfqq) {
3395 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
3397 if (cfq_cfqq_slice_new(cfqq)) {
3398 cfq_set_prio_slice(cfqd, cfqq);
3399 cfq_clear_cfqq_slice_new(cfqq);
3403 * Should we wait for next request to come in before we expire
3406 if (cfq_should_wait_busy(cfqd, cfqq)) {
3407 cfqq->slice_end = jiffies + cfqd->cfq_slice_idle;
3408 cfq_mark_cfqq_wait_busy(cfqq);
3409 cfq_log_cfqq(cfqd, cfqq, "will busy wait");
3413 * Idling is not enabled on:
3415 * - idle-priority queues
3417 * - queues with still some requests queued
3418 * - when there is a close cooperator
3420 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
3421 cfq_slice_expired(cfqd, 1);
3422 else if (sync && cfqq_empty &&
3423 !cfq_close_cooperator(cfqd, cfqq)) {
3424 cfq_arm_slice_timer(cfqd);
3428 if (!cfqd->rq_in_driver)
3429 cfq_schedule_dispatch(cfqd);
3433 * we temporarily boost lower priority queues if they are holding fs exclusive
3434 * resources. they are boosted to normal prio (CLASS_BE/4)
3436 static void cfq_prio_boost(struct cfq_queue *cfqq)
3438 if (has_fs_excl()) {
3440 * boost idle prio on transactions that would lock out other
3441 * users of the filesystem
3443 if (cfq_class_idle(cfqq))
3444 cfqq->ioprio_class = IOPRIO_CLASS_BE;
3445 if (cfqq->ioprio > IOPRIO_NORM)
3446 cfqq->ioprio = IOPRIO_NORM;
3449 * unboost the queue (if needed)
3451 cfqq->ioprio_class = cfqq->org_ioprio_class;
3452 cfqq->ioprio = cfqq->org_ioprio;
3456 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
3458 if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
3459 cfq_mark_cfqq_must_alloc_slice(cfqq);
3460 return ELV_MQUEUE_MUST;
3463 return ELV_MQUEUE_MAY;
3466 static int cfq_may_queue(struct request_queue *q, int rw)
3468 struct cfq_data *cfqd = q->elevator->elevator_data;
3469 struct task_struct *tsk = current;
3470 struct cfq_io_context *cic;
3471 struct cfq_queue *cfqq;
3474 * don't force setup of a queue from here, as a call to may_queue
3475 * does not necessarily imply that a request actually will be queued.
3476 * so just lookup a possibly existing queue, or return 'may queue'
3479 cic = cfq_cic_lookup(cfqd, tsk->io_context);
3481 return ELV_MQUEUE_MAY;
3483 cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
3485 cfq_init_prio_data(cfqq, cic->ioc);
3486 cfq_prio_boost(cfqq);
3488 return __cfq_may_queue(cfqq);
3491 return ELV_MQUEUE_MAY;
3495 * queue lock held here
3497 static void cfq_put_request(struct request *rq)
3499 struct cfq_queue *cfqq = RQ_CFQQ(rq);
3502 const int rw = rq_data_dir(rq);
3504 BUG_ON(!cfqq->allocated[rw]);
3505 cfqq->allocated[rw]--;
3507 put_io_context(RQ_CIC(rq)->ioc);
3509 rq->elevator_private = NULL;
3510 rq->elevator_private2 = NULL;
3512 /* Put down rq reference on cfqg */
3513 cfq_put_cfqg(RQ_CFQG(rq));
3514 rq->elevator_private3 = NULL;
3516 cfq_put_queue(cfqq);
3520 static struct cfq_queue *
3521 cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_context *cic,
3522 struct cfq_queue *cfqq)
3524 cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq);
3525 cic_set_cfqq(cic, cfqq->new_cfqq, 1);
3526 cfq_mark_cfqq_coop(cfqq->new_cfqq);
3527 cfq_put_queue(cfqq);
3528 return cic_to_cfqq(cic, 1);
3532 * Returns NULL if a new cfqq should be allocated, or the old cfqq if this
3533 * was the last process referring to said cfqq.
3535 static struct cfq_queue *
3536 split_cfqq(struct cfq_io_context *cic, struct cfq_queue *cfqq)
3538 if (cfqq_process_refs(cfqq) == 1) {
3539 cfqq->pid = current->pid;
3540 cfq_clear_cfqq_coop(cfqq);
3541 cfq_clear_cfqq_split_coop(cfqq);
3545 cic_set_cfqq(cic, NULL, 1);
3547 cfq_put_cooperator(cfqq);
3549 cfq_put_queue(cfqq);
3553 * Allocate cfq data structures associated with this request.
3556 cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
3558 struct cfq_data *cfqd = q->elevator->elevator_data;
3559 struct cfq_io_context *cic;
3560 const int rw = rq_data_dir(rq);
3561 const bool is_sync = rq_is_sync(rq);
3562 struct cfq_queue *cfqq;
3563 unsigned long flags;
3565 might_sleep_if(gfp_mask & __GFP_WAIT);
3567 cic = cfq_get_io_context(cfqd, gfp_mask);
3569 spin_lock_irqsave(q->queue_lock, flags);
3575 cfqq = cic_to_cfqq(cic, is_sync);
3576 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
3577 cfqq = cfq_get_queue(cfqd, is_sync, cic->ioc, gfp_mask);
3578 cic_set_cfqq(cic, cfqq, is_sync);
3581 * If the queue was seeky for too long, break it apart.
3583 if (cfq_cfqq_coop(cfqq) && cfq_cfqq_split_coop(cfqq)) {
3584 cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq");
3585 cfqq = split_cfqq(cic, cfqq);
3591 * Check to see if this queue is scheduled to merge with
3592 * another, closely cooperating queue. The merging of
3593 * queues happens here as it must be done in process context.
3594 * The reference on new_cfqq was taken in merge_cfqqs.
3597 cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq);
3600 cfqq->allocated[rw]++;
3601 atomic_inc(&cfqq->ref);
3603 spin_unlock_irqrestore(q->queue_lock, flags);
3605 rq->elevator_private = cic;
3606 rq->elevator_private2 = cfqq;
3607 rq->elevator_private3 = cfq_ref_get_cfqg(cfqq->cfqg);
3612 put_io_context(cic->ioc);
3614 cfq_schedule_dispatch(cfqd);
3615 spin_unlock_irqrestore(q->queue_lock, flags);
3616 cfq_log(cfqd, "set_request fail");
3620 static void cfq_kick_queue(struct work_struct *work)
3622 struct cfq_data *cfqd =
3623 container_of(work, struct cfq_data, unplug_work);
3624 struct request_queue *q = cfqd->queue;
3626 spin_lock_irq(q->queue_lock);
3627 __blk_run_queue(cfqd->queue);
3628 spin_unlock_irq(q->queue_lock);
3632 * Timer running if the active_queue is currently idling inside its time slice
3634 static void cfq_idle_slice_timer(unsigned long data)
3636 struct cfq_data *cfqd = (struct cfq_data *) data;
3637 struct cfq_queue *cfqq;
3638 unsigned long flags;
3641 cfq_log(cfqd, "idle timer fired");
3643 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
3645 cfqq = cfqd->active_queue;
3650 * We saw a request before the queue expired, let it through
3652 if (cfq_cfqq_must_dispatch(cfqq))
3658 if (cfq_slice_used(cfqq))
3662 * only expire and reinvoke request handler, if there are
3663 * other queues with pending requests
3665 if (!cfqd->busy_queues)
3669 * not expired and it has a request pending, let it dispatch
3671 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3675 * Queue depth flag is reset only when the idle didn't succeed
3677 cfq_clear_cfqq_deep(cfqq);
3680 cfq_slice_expired(cfqd, timed_out);
3682 cfq_schedule_dispatch(cfqd);
3684 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
3687 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
3689 del_timer_sync(&cfqd->idle_slice_timer);
3690 cancel_work_sync(&cfqd->unplug_work);
3693 static void cfq_put_async_queues(struct cfq_data *cfqd)
3697 for (i = 0; i < IOPRIO_BE_NR; i++) {
3698 if (cfqd->async_cfqq[0][i])
3699 cfq_put_queue(cfqd->async_cfqq[0][i]);
3700 if (cfqd->async_cfqq[1][i])
3701 cfq_put_queue(cfqd->async_cfqq[1][i]);
3704 if (cfqd->async_idle_cfqq)
3705 cfq_put_queue(cfqd->async_idle_cfqq);
3708 static void cfq_cfqd_free(struct rcu_head *head)
3710 kfree(container_of(head, struct cfq_data, rcu));
3713 static void cfq_exit_queue(struct elevator_queue *e)
3715 struct cfq_data *cfqd = e->elevator_data;
3716 struct request_queue *q = cfqd->queue;
3718 cfq_shutdown_timer_wq(cfqd);
3720 spin_lock_irq(q->queue_lock);
3722 if (cfqd->active_queue)
3723 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
3725 while (!list_empty(&cfqd->cic_list)) {
3726 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
3727 struct cfq_io_context,
3730 __cfq_exit_single_io_context(cfqd, cic);
3733 cfq_put_async_queues(cfqd);
3734 cfq_release_cfq_groups(cfqd);
3735 cfq_blkiocg_del_blkio_group(&cfqd->root_group.blkg);
3737 spin_unlock_irq(q->queue_lock);
3739 cfq_shutdown_timer_wq(cfqd);
3741 spin_lock(&cic_index_lock);
3742 ida_remove(&cic_index_ida, cfqd->cic_index);
3743 spin_unlock(&cic_index_lock);
3745 /* Wait for cfqg->blkg->key accessors to exit their grace periods. */
3746 call_rcu(&cfqd->rcu, cfq_cfqd_free);
3749 static int cfq_alloc_cic_index(void)
3754 if (!ida_pre_get(&cic_index_ida, GFP_KERNEL))
3757 spin_lock(&cic_index_lock);
3758 error = ida_get_new(&cic_index_ida, &index);
3759 spin_unlock(&cic_index_lock);
3760 if (error && error != -EAGAIN)
3767 static void *cfq_init_queue(struct request_queue *q)
3769 struct cfq_data *cfqd;
3771 struct cfq_group *cfqg;
3772 struct cfq_rb_root *st;
3774 i = cfq_alloc_cic_index();
3778 cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
3782 cfqd->cic_index = i;
3784 /* Init root service tree */
3785 cfqd->grp_service_tree = CFQ_RB_ROOT;
3787 /* Init root group */
3788 cfqg = &cfqd->root_group;
3789 for_each_cfqg_st(cfqg, i, j, st)
3791 RB_CLEAR_NODE(&cfqg->rb_node);
3793 /* Give preference to root group over other groups */
3794 cfqg->weight = 2*BLKIO_WEIGHT_DEFAULT;
3796 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3798 * Take a reference to root group which we never drop. This is just
3799 * to make sure that cfq_put_cfqg() does not try to kfree root group
3801 atomic_set(&cfqg->ref, 1);
3803 cfq_blkiocg_add_blkio_group(&blkio_root_cgroup, &cfqg->blkg,
3808 * Not strictly needed (since RB_ROOT just clears the node and we
3809 * zeroed cfqd on alloc), but better be safe in case someone decides
3810 * to add magic to the rb code
3812 for (i = 0; i < CFQ_PRIO_LISTS; i++)
3813 cfqd->prio_trees[i] = RB_ROOT;
3816 * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
3817 * Grab a permanent reference to it, so that the normal code flow
3818 * will not attempt to free it.
3820 cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
3821 atomic_inc(&cfqd->oom_cfqq.ref);
3822 cfq_link_cfqq_cfqg(&cfqd->oom_cfqq, &cfqd->root_group);
3824 INIT_LIST_HEAD(&cfqd->cic_list);
3828 init_timer(&cfqd->idle_slice_timer);
3829 cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
3830 cfqd->idle_slice_timer.data = (unsigned long) cfqd;
3832 INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
3834 cfqd->cfq_quantum = cfq_quantum;
3835 cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
3836 cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
3837 cfqd->cfq_back_max = cfq_back_max;
3838 cfqd->cfq_back_penalty = cfq_back_penalty;
3839 cfqd->cfq_slice[0] = cfq_slice_async;
3840 cfqd->cfq_slice[1] = cfq_slice_sync;
3841 cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
3842 cfqd->cfq_slice_idle = cfq_slice_idle;
3843 cfqd->cfq_latency = 1;
3844 cfqd->cfq_group_isolation = 0;
3847 * we optimistically start assuming sync ops weren't delayed in last
3848 * second, in order to have larger depth for async operations.
3850 cfqd->last_delayed_sync = jiffies - HZ;
3854 static void cfq_slab_kill(void)
3857 * Caller already ensured that pending RCU callbacks are completed,
3858 * so we should have no busy allocations at this point.
3861 kmem_cache_destroy(cfq_pool);
3863 kmem_cache_destroy(cfq_ioc_pool);
3866 static int __init cfq_slab_setup(void)
3868 cfq_pool = KMEM_CACHE(cfq_queue, 0);
3872 cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0);
3883 * sysfs parts below -->
3886 cfq_var_show(unsigned int var, char *page)
3888 return sprintf(page, "%d\n", var);
3892 cfq_var_store(unsigned int *var, const char *page, size_t count)
3894 char *p = (char *) page;
3896 *var = simple_strtoul(p, &p, 10);
3900 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
3901 static ssize_t __FUNC(struct elevator_queue *e, char *page) \
3903 struct cfq_data *cfqd = e->elevator_data; \
3904 unsigned int __data = __VAR; \
3906 __data = jiffies_to_msecs(__data); \
3907 return cfq_var_show(__data, (page)); \
3909 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
3910 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
3911 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
3912 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
3913 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
3914 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
3915 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
3916 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
3917 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
3918 SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
3919 SHOW_FUNCTION(cfq_group_isolation_show, cfqd->cfq_group_isolation, 0);
3920 #undef SHOW_FUNCTION
3922 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
3923 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
3925 struct cfq_data *cfqd = e->elevator_data; \
3926 unsigned int __data; \
3927 int ret = cfq_var_store(&__data, (page), count); \
3928 if (__data < (MIN)) \
3930 else if (__data > (MAX)) \
3933 *(__PTR) = msecs_to_jiffies(__data); \
3935 *(__PTR) = __data; \
3938 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
3939 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
3941 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
3943 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
3944 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
3946 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
3947 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
3948 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
3949 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
3951 STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
3952 STORE_FUNCTION(cfq_group_isolation_store, &cfqd->cfq_group_isolation, 0, 1, 0);
3953 #undef STORE_FUNCTION
3955 #define CFQ_ATTR(name) \
3956 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
3958 static struct elv_fs_entry cfq_attrs[] = {
3960 CFQ_ATTR(fifo_expire_sync),
3961 CFQ_ATTR(fifo_expire_async),
3962 CFQ_ATTR(back_seek_max),
3963 CFQ_ATTR(back_seek_penalty),
3964 CFQ_ATTR(slice_sync),
3965 CFQ_ATTR(slice_async),
3966 CFQ_ATTR(slice_async_rq),
3967 CFQ_ATTR(slice_idle),
3968 CFQ_ATTR(low_latency),
3969 CFQ_ATTR(group_isolation),
3973 static struct elevator_type iosched_cfq = {
3975 .elevator_merge_fn = cfq_merge,
3976 .elevator_merged_fn = cfq_merged_request,
3977 .elevator_merge_req_fn = cfq_merged_requests,
3978 .elevator_allow_merge_fn = cfq_allow_merge,
3979 .elevator_bio_merged_fn = cfq_bio_merged,
3980 .elevator_dispatch_fn = cfq_dispatch_requests,
3981 .elevator_add_req_fn = cfq_insert_request,
3982 .elevator_activate_req_fn = cfq_activate_request,
3983 .elevator_deactivate_req_fn = cfq_deactivate_request,
3984 .elevator_queue_empty_fn = cfq_queue_empty,
3985 .elevator_completed_req_fn = cfq_completed_request,
3986 .elevator_former_req_fn = elv_rb_former_request,
3987 .elevator_latter_req_fn = elv_rb_latter_request,
3988 .elevator_set_req_fn = cfq_set_request,
3989 .elevator_put_req_fn = cfq_put_request,
3990 .elevator_may_queue_fn = cfq_may_queue,
3991 .elevator_init_fn = cfq_init_queue,
3992 .elevator_exit_fn = cfq_exit_queue,
3993 .trim = cfq_free_io_context,
3995 .elevator_attrs = cfq_attrs,
3996 .elevator_name = "cfq",
3997 .elevator_owner = THIS_MODULE,
4000 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4001 static struct blkio_policy_type blkio_policy_cfq = {
4003 .blkio_unlink_group_fn = cfq_unlink_blkio_group,
4004 .blkio_update_group_weight_fn = cfq_update_blkio_group_weight,
4006 .plid = BLKIO_POLICY_PROP,
4009 static struct blkio_policy_type blkio_policy_cfq;
4012 static int __init cfq_init(void)
4015 * could be 0 on HZ < 1000 setups
4017 if (!cfq_slice_async)
4018 cfq_slice_async = 1;
4019 if (!cfq_slice_idle)
4022 if (cfq_slab_setup())
4025 elv_register(&iosched_cfq);
4026 blkio_policy_register(&blkio_policy_cfq);
4031 static void __exit cfq_exit(void)
4033 DECLARE_COMPLETION_ONSTACK(all_gone);
4034 blkio_policy_unregister(&blkio_policy_cfq);
4035 elv_unregister(&iosched_cfq);
4036 ioc_gone = &all_gone;
4037 /* ioc_gone's update must be visible before reading ioc_count */
4041 * this also protects us from entering cfq_slab_kill() with
4042 * pending RCU callbacks
4044 if (elv_ioc_count_read(cfq_ioc_count))
4045 wait_for_completion(&all_gone);
4046 ida_destroy(&cic_index_ida);
4050 module_init(cfq_init);
4051 module_exit(cfq_exit);
4053 MODULE_AUTHOR("Jens Axboe");
4054 MODULE_LICENSE("GPL");
4055 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");