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/blkdev.h>
11 #include <linux/elevator.h>
12 #include <linux/jiffies.h>
13 #include <linux/rbtree.h>
14 #include <linux/ioprio.h>
15 #include <linux/blktrace_api.h>
16 #include "blk-cgroup.h"
21 /* max queue in one round of service */
22 static const int cfq_quantum = 4;
23 static const int cfq_fifo_expire[2] = { HZ / 4, HZ / 8 };
24 /* maximum backwards seek, in KiB */
25 static const int cfq_back_max = 16 * 1024;
26 /* penalty of a backwards seek */
27 static const int cfq_back_penalty = 2;
28 static const int cfq_slice_sync = HZ / 10;
29 static int cfq_slice_async = HZ / 25;
30 static const int cfq_slice_async_rq = 2;
31 static int cfq_slice_idle = HZ / 125;
32 static const int cfq_target_latency = HZ * 3/10; /* 300 ms */
33 static const int cfq_hist_divisor = 4;
36 * offset from end of service tree
38 #define CFQ_IDLE_DELAY (HZ / 5)
41 * below this threshold, we consider thinktime immediate
43 #define CFQ_MIN_TT (2)
46 * Allow merged cfqqs to perform this amount of seeky I/O before
47 * deciding to break the queues up again.
49 #define CFQQ_COOP_TOUT (HZ)
51 #define CFQ_SLICE_SCALE (5)
52 #define CFQ_HW_QUEUE_MIN (5)
53 #define CFQ_SERVICE_SHIFT 12
56 ((struct cfq_io_context *) (rq)->elevator_private)
57 #define RQ_CFQQ(rq) (struct cfq_queue *) ((rq)->elevator_private2)
59 static struct kmem_cache *cfq_pool;
60 static struct kmem_cache *cfq_ioc_pool;
62 static DEFINE_PER_CPU(unsigned long, cfq_ioc_count);
63 static struct completion *ioc_gone;
64 static DEFINE_SPINLOCK(ioc_gone_lock);
66 #define CFQ_PRIO_LISTS IOPRIO_BE_NR
67 #define cfq_class_idle(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
68 #define cfq_class_rt(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
70 #define sample_valid(samples) ((samples) > 80)
71 #define rb_entry_cfqg(node) rb_entry((node), struct cfq_group, rb_node)
74 * Most of our rbtree usage is for sorting with min extraction, so
75 * if we cache the leftmost node we don't have to walk down the tree
76 * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should
77 * move this into the elevator for the rq sorting as well.
84 struct rb_node *active;
86 #define CFQ_RB_ROOT (struct cfq_rb_root) { RB_ROOT, NULL, 0, 0, }
89 * Per process-grouping structure
94 /* various state flags, see below */
97 struct cfq_data *cfqd;
98 /* service_tree member */
99 struct rb_node rb_node;
100 /* service_tree key */
101 unsigned long rb_key;
102 /* prio tree member */
103 struct rb_node p_node;
104 /* prio tree root we belong to, if any */
105 struct rb_root *p_root;
106 /* sorted list of pending requests */
107 struct rb_root sort_list;
108 /* if fifo isn't expired, next request to serve */
109 struct request *next_rq;
110 /* requests queued in sort_list */
112 /* currently allocated requests */
114 /* fifo list of requests in sort_list */
115 struct list_head fifo;
117 unsigned long slice_end;
119 unsigned int slice_dispatch;
121 /* pending metadata requests */
123 /* number of requests that are on the dispatch list or inside driver */
126 /* io prio of this group */
127 unsigned short ioprio, org_ioprio;
128 unsigned short ioprio_class, org_ioprio_class;
130 unsigned int seek_samples;
133 sector_t last_request_pos;
134 unsigned long seeky_start;
138 struct cfq_rb_root *service_tree;
139 struct cfq_queue *new_cfqq;
140 struct cfq_group *cfqg;
144 * First index in the service_trees.
145 * IDLE is handled separately, so it has negative index
154 * Second index in the service_trees.
158 SYNC_NOIDLE_WORKLOAD = 1,
162 /* This is per cgroup per device grouping structure */
164 /* group service_tree member */
165 struct rb_node rb_node;
167 /* group service_tree key */
172 /* number of cfqq currently on this group */
176 * rr lists of queues with requests, onle rr for each priority class.
177 * Counts are embedded in the cfq_rb_root
179 struct cfq_rb_root service_trees[2][3];
180 struct cfq_rb_root service_tree_idle;
184 * Per block device queue structure
187 struct request_queue *queue;
188 /* Root service tree for cfq_groups */
189 struct cfq_rb_root grp_service_tree;
190 struct cfq_group root_group;
193 * The priority currently being served
195 enum wl_prio_t serving_prio;
196 enum wl_type_t serving_type;
197 unsigned long workload_expires;
198 struct cfq_group *serving_group;
199 bool noidle_tree_requires_idle;
202 * Each priority tree is sorted by next_request position. These
203 * trees are used when determining if two or more queues are
204 * interleaving requests (see cfq_close_cooperator).
206 struct rb_root prio_trees[CFQ_PRIO_LISTS];
208 unsigned int busy_queues;
209 unsigned int busy_queues_avg[2];
215 * queue-depth detection
221 * -1 => indeterminate, (cfq will behave as if NCQ is present, to allow better detection)
222 * 1 => NCQ is present (hw_tag_est_depth is the estimated max depth)
225 int hw_tag_est_depth;
226 unsigned int hw_tag_samples;
229 * idle window management
231 struct timer_list idle_slice_timer;
232 struct work_struct unplug_work;
234 struct cfq_queue *active_queue;
235 struct cfq_io_context *active_cic;
238 * async queue for each priority case
240 struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
241 struct cfq_queue *async_idle_cfqq;
243 sector_t last_position;
246 * tunables, see top of file
248 unsigned int cfq_quantum;
249 unsigned int cfq_fifo_expire[2];
250 unsigned int cfq_back_penalty;
251 unsigned int cfq_back_max;
252 unsigned int cfq_slice[2];
253 unsigned int cfq_slice_async_rq;
254 unsigned int cfq_slice_idle;
255 unsigned int cfq_latency;
257 struct list_head cic_list;
260 * Fallback dummy cfqq for extreme OOM conditions
262 struct cfq_queue oom_cfqq;
264 unsigned long last_end_sync_rq;
267 static struct cfq_rb_root *service_tree_for(struct cfq_group *cfqg,
270 struct cfq_data *cfqd)
275 if (prio == IDLE_WORKLOAD)
276 return &cfqg->service_tree_idle;
278 return &cfqg->service_trees[prio][type];
281 enum cfqq_state_flags {
282 CFQ_CFQQ_FLAG_on_rr = 0, /* on round-robin busy list */
283 CFQ_CFQQ_FLAG_wait_request, /* waiting for a request */
284 CFQ_CFQQ_FLAG_must_dispatch, /* must be allowed a dispatch */
285 CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
286 CFQ_CFQQ_FLAG_fifo_expire, /* FIFO checked in this slice */
287 CFQ_CFQQ_FLAG_idle_window, /* slice idling enabled */
288 CFQ_CFQQ_FLAG_prio_changed, /* task priority has changed */
289 CFQ_CFQQ_FLAG_slice_new, /* no requests dispatched in slice */
290 CFQ_CFQQ_FLAG_sync, /* synchronous queue */
291 CFQ_CFQQ_FLAG_coop, /* cfqq is shared */
292 CFQ_CFQQ_FLAG_deep, /* sync cfqq experienced large depth */
295 #define CFQ_CFQQ_FNS(name) \
296 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \
298 (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name); \
300 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \
302 (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \
304 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \
306 return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \
310 CFQ_CFQQ_FNS(wait_request);
311 CFQ_CFQQ_FNS(must_dispatch);
312 CFQ_CFQQ_FNS(must_alloc_slice);
313 CFQ_CFQQ_FNS(fifo_expire);
314 CFQ_CFQQ_FNS(idle_window);
315 CFQ_CFQQ_FNS(prio_changed);
316 CFQ_CFQQ_FNS(slice_new);
322 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
323 blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args)
324 #define cfq_log(cfqd, fmt, args...) \
325 blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
327 /* Traverses through cfq group service trees */
328 #define for_each_cfqg_st(cfqg, i, j, st) \
329 for (i = 0; i <= IDLE_WORKLOAD; i++) \
330 for (j = 0, st = i < IDLE_WORKLOAD ? &cfqg->service_trees[i][j]\
331 : &cfqg->service_tree_idle; \
332 (i < IDLE_WORKLOAD && j <= SYNC_WORKLOAD) || \
333 (i == IDLE_WORKLOAD && j == 0); \
334 j++, st = i < IDLE_WORKLOAD ? \
335 &cfqg->service_trees[i][j]: NULL) \
338 static inline enum wl_prio_t cfqq_prio(struct cfq_queue *cfqq)
340 if (cfq_class_idle(cfqq))
341 return IDLE_WORKLOAD;
342 if (cfq_class_rt(cfqq))
348 static enum wl_type_t cfqq_type(struct cfq_queue *cfqq)
350 if (!cfq_cfqq_sync(cfqq))
351 return ASYNC_WORKLOAD;
352 if (!cfq_cfqq_idle_window(cfqq))
353 return SYNC_NOIDLE_WORKLOAD;
354 return SYNC_WORKLOAD;
357 static inline int cfq_busy_queues_wl(enum wl_prio_t wl, struct cfq_data *cfqd)
359 struct cfq_group *cfqg = &cfqd->root_group;
361 if (wl == IDLE_WORKLOAD)
362 return cfqg->service_tree_idle.count;
364 return cfqg->service_trees[wl][ASYNC_WORKLOAD].count
365 + cfqg->service_trees[wl][SYNC_NOIDLE_WORKLOAD].count
366 + cfqg->service_trees[wl][SYNC_WORKLOAD].count;
369 static void cfq_dispatch_insert(struct request_queue *, struct request *);
370 static struct cfq_queue *cfq_get_queue(struct cfq_data *, bool,
371 struct io_context *, gfp_t);
372 static struct cfq_io_context *cfq_cic_lookup(struct cfq_data *,
373 struct io_context *);
375 static inline int rq_in_driver(struct cfq_data *cfqd)
377 return cfqd->rq_in_driver[0] + cfqd->rq_in_driver[1];
380 static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_context *cic,
383 return cic->cfqq[is_sync];
386 static inline void cic_set_cfqq(struct cfq_io_context *cic,
387 struct cfq_queue *cfqq, bool is_sync)
389 cic->cfqq[is_sync] = cfqq;
393 * We regard a request as SYNC, if it's either a read or has the SYNC bit
394 * set (in which case it could also be direct WRITE).
396 static inline bool cfq_bio_sync(struct bio *bio)
398 return bio_data_dir(bio) == READ || bio_rw_flagged(bio, BIO_RW_SYNCIO);
402 * scheduler run of queue, if there are requests pending and no one in the
403 * driver that will restart queueing
405 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
407 if (cfqd->busy_queues) {
408 cfq_log(cfqd, "schedule dispatch");
409 kblockd_schedule_work(cfqd->queue, &cfqd->unplug_work);
413 static int cfq_queue_empty(struct request_queue *q)
415 struct cfq_data *cfqd = q->elevator->elevator_data;
417 return !cfqd->rq_queued;
421 * Scale schedule slice based on io priority. Use the sync time slice only
422 * if a queue is marked sync and has sync io queued. A sync queue with async
423 * io only, should not get full sync slice length.
425 static inline int cfq_prio_slice(struct cfq_data *cfqd, bool sync,
428 const int base_slice = cfqd->cfq_slice[sync];
430 WARN_ON(prio >= IOPRIO_BE_NR);
432 return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
436 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
438 return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
441 static inline u64 cfq_scale_slice(unsigned long delta, struct cfq_group *cfqg)
443 u64 d = delta << CFQ_SERVICE_SHIFT;
445 d = d * BLKIO_WEIGHT_DEFAULT;
446 do_div(d, cfqg->weight);
450 static inline u64 max_vdisktime(u64 min_vdisktime, u64 vdisktime)
452 s64 delta = (s64)(vdisktime - min_vdisktime);
454 min_vdisktime = vdisktime;
456 return min_vdisktime;
459 static inline u64 min_vdisktime(u64 min_vdisktime, u64 vdisktime)
461 s64 delta = (s64)(vdisktime - min_vdisktime);
463 min_vdisktime = vdisktime;
465 return min_vdisktime;
468 static void update_min_vdisktime(struct cfq_rb_root *st)
470 u64 vdisktime = st->min_vdisktime;
471 struct cfq_group *cfqg;
474 cfqg = rb_entry_cfqg(st->active);
475 vdisktime = cfqg->vdisktime;
479 cfqg = rb_entry_cfqg(st->left);
480 vdisktime = min_vdisktime(vdisktime, cfqg->vdisktime);
483 st->min_vdisktime = max_vdisktime(st->min_vdisktime, vdisktime);
487 * get averaged number of queues of RT/BE priority.
488 * average is updated, with a formula that gives more weight to higher numbers,
489 * to quickly follows sudden increases and decrease slowly
492 static inline unsigned cfq_get_avg_queues(struct cfq_data *cfqd, bool rt)
494 unsigned min_q, max_q;
495 unsigned mult = cfq_hist_divisor - 1;
496 unsigned round = cfq_hist_divisor / 2;
497 unsigned busy = cfq_busy_queues_wl(rt, cfqd);
499 min_q = min(cfqd->busy_queues_avg[rt], busy);
500 max_q = max(cfqd->busy_queues_avg[rt], busy);
501 cfqd->busy_queues_avg[rt] = (mult * max_q + min_q + round) /
503 return cfqd->busy_queues_avg[rt];
507 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
509 unsigned slice = cfq_prio_to_slice(cfqd, cfqq);
510 if (cfqd->cfq_latency) {
511 /* interested queues (we consider only the ones with the same
513 unsigned iq = cfq_get_avg_queues(cfqd, cfq_class_rt(cfqq));
514 unsigned sync_slice = cfqd->cfq_slice[1];
515 unsigned expect_latency = sync_slice * iq;
516 if (expect_latency > cfq_target_latency) {
517 unsigned base_low_slice = 2 * cfqd->cfq_slice_idle;
518 /* scale low_slice according to IO priority
519 * and sync vs async */
521 min(slice, base_low_slice * slice / sync_slice);
522 /* the adapted slice value is scaled to fit all iqs
523 * into the target latency */
524 slice = max(slice * cfq_target_latency / expect_latency,
528 cfqq->slice_end = jiffies + slice;
529 cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
533 * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
534 * isn't valid until the first request from the dispatch is activated
535 * and the slice time set.
537 static inline bool cfq_slice_used(struct cfq_queue *cfqq)
539 if (cfq_cfqq_slice_new(cfqq))
541 if (time_before(jiffies, cfqq->slice_end))
548 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
549 * We choose the request that is closest to the head right now. Distance
550 * behind the head is penalized and only allowed to a certain extent.
552 static struct request *
553 cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2, sector_t last)
555 sector_t s1, s2, d1 = 0, d2 = 0;
556 unsigned long back_max;
557 #define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */
558 #define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */
559 unsigned wrap = 0; /* bit mask: requests behind the disk head? */
561 if (rq1 == NULL || rq1 == rq2)
566 if (rq_is_sync(rq1) && !rq_is_sync(rq2))
568 else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
570 if (rq_is_meta(rq1) && !rq_is_meta(rq2))
572 else if (rq_is_meta(rq2) && !rq_is_meta(rq1))
575 s1 = blk_rq_pos(rq1);
576 s2 = blk_rq_pos(rq2);
579 * by definition, 1KiB is 2 sectors
581 back_max = cfqd->cfq_back_max * 2;
584 * Strict one way elevator _except_ in the case where we allow
585 * short backward seeks which are biased as twice the cost of a
586 * similar forward seek.
590 else if (s1 + back_max >= last)
591 d1 = (last - s1) * cfqd->cfq_back_penalty;
593 wrap |= CFQ_RQ1_WRAP;
597 else if (s2 + back_max >= last)
598 d2 = (last - s2) * cfqd->cfq_back_penalty;
600 wrap |= CFQ_RQ2_WRAP;
602 /* Found required data */
605 * By doing switch() on the bit mask "wrap" we avoid having to
606 * check two variables for all permutations: --> faster!
609 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
625 case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
628 * Since both rqs are wrapped,
629 * start with the one that's further behind head
630 * (--> only *one* back seek required),
631 * since back seek takes more time than forward.
641 * The below is leftmost cache rbtree addon
643 static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
645 /* Service tree is empty */
650 root->left = rb_first(&root->rb);
653 return rb_entry(root->left, struct cfq_queue, rb_node);
658 static struct cfq_group *cfq_rb_first_group(struct cfq_rb_root *root)
661 root->left = rb_first(&root->rb);
664 return rb_entry_cfqg(root->left);
669 static void rb_erase_init(struct rb_node *n, struct rb_root *root)
675 static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
679 rb_erase_init(n, &root->rb);
684 * would be nice to take fifo expire time into account as well
686 static struct request *
687 cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
688 struct request *last)
690 struct rb_node *rbnext = rb_next(&last->rb_node);
691 struct rb_node *rbprev = rb_prev(&last->rb_node);
692 struct request *next = NULL, *prev = NULL;
694 BUG_ON(RB_EMPTY_NODE(&last->rb_node));
697 prev = rb_entry_rq(rbprev);
700 next = rb_entry_rq(rbnext);
702 rbnext = rb_first(&cfqq->sort_list);
703 if (rbnext && rbnext != &last->rb_node)
704 next = rb_entry_rq(rbnext);
707 return cfq_choose_req(cfqd, next, prev, blk_rq_pos(last));
710 static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
711 struct cfq_queue *cfqq)
714 * just an approximation, should be ok.
716 return (cfqq->cfqg->nr_cfqq - 1) * (cfq_prio_slice(cfqd, 1, 0) -
717 cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
721 cfqg_key(struct cfq_rb_root *st, struct cfq_group *cfqg)
723 return cfqg->vdisktime - st->min_vdisktime;
727 __cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
729 struct rb_node **node = &st->rb.rb_node;
730 struct rb_node *parent = NULL;
731 struct cfq_group *__cfqg;
732 s64 key = cfqg_key(st, cfqg);
735 while (*node != NULL) {
737 __cfqg = rb_entry_cfqg(parent);
739 if (key < cfqg_key(st, __cfqg))
740 node = &parent->rb_left;
742 node = &parent->rb_right;
748 st->left = &cfqg->rb_node;
750 rb_link_node(&cfqg->rb_node, parent, node);
751 rb_insert_color(&cfqg->rb_node, &st->rb);
755 cfq_group_service_tree_add(struct cfq_data *cfqd, struct cfq_group *cfqg)
757 struct cfq_rb_root *st = &cfqd->grp_service_tree;
758 struct cfq_group *__cfqg;
766 * Currently put the group at the end. Later implement something
767 * so that groups get lesser vtime based on their weights, so that
768 * if group does not loose all if it was not continously backlogged.
770 n = rb_last(&st->rb);
772 __cfqg = rb_entry_cfqg(n);
773 cfqg->vdisktime = __cfqg->vdisktime + CFQ_IDLE_DELAY;
775 cfqg->vdisktime = st->min_vdisktime;
777 __cfq_group_service_tree_add(st, cfqg);
782 cfq_group_service_tree_del(struct cfq_data *cfqd, struct cfq_group *cfqg)
784 struct cfq_rb_root *st = &cfqd->grp_service_tree;
786 if (st->active == &cfqg->rb_node)
789 BUG_ON(cfqg->nr_cfqq < 1);
792 /* If there are other cfq queues under this group, don't delete it */
797 if (!RB_EMPTY_NODE(&cfqg->rb_node))
798 cfq_rb_erase(&cfqg->rb_node, st);
802 * The cfqd->service_trees holds all pending cfq_queue's that have
803 * requests waiting to be processed. It is sorted in the order that
804 * we will service the queues.
806 static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
809 struct rb_node **p, *parent;
810 struct cfq_queue *__cfqq;
811 unsigned long rb_key;
812 struct cfq_rb_root *service_tree;
815 service_tree = service_tree_for(cfqq->cfqg, cfqq_prio(cfqq),
816 cfqq_type(cfqq), cfqd);
817 if (cfq_class_idle(cfqq)) {
818 rb_key = CFQ_IDLE_DELAY;
819 parent = rb_last(&service_tree->rb);
820 if (parent && parent != &cfqq->rb_node) {
821 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
822 rb_key += __cfqq->rb_key;
825 } else if (!add_front) {
827 * Get our rb key offset. Subtract any residual slice
828 * value carried from last service. A negative resid
829 * count indicates slice overrun, and this should position
830 * the next service time further away in the tree.
832 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
833 rb_key -= cfqq->slice_resid;
834 cfqq->slice_resid = 0;
837 __cfqq = cfq_rb_first(service_tree);
838 rb_key += __cfqq ? __cfqq->rb_key : jiffies;
841 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
843 * same position, nothing more to do
845 if (rb_key == cfqq->rb_key &&
846 cfqq->service_tree == service_tree)
849 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
850 cfqq->service_tree = NULL;
855 cfqq->service_tree = service_tree;
856 p = &service_tree->rb.rb_node;
861 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
864 * sort by key, that represents service time.
866 if (time_before(rb_key, __cfqq->rb_key))
877 service_tree->left = &cfqq->rb_node;
879 cfqq->rb_key = rb_key;
880 rb_link_node(&cfqq->rb_node, parent, p);
881 rb_insert_color(&cfqq->rb_node, &service_tree->rb);
882 service_tree->count++;
883 cfq_group_service_tree_add(cfqd, cfqq->cfqg);
886 static struct cfq_queue *
887 cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
888 sector_t sector, struct rb_node **ret_parent,
889 struct rb_node ***rb_link)
891 struct rb_node **p, *parent;
892 struct cfq_queue *cfqq = NULL;
900 cfqq = rb_entry(parent, struct cfq_queue, p_node);
903 * Sort strictly based on sector. Smallest to the left,
904 * largest to the right.
906 if (sector > blk_rq_pos(cfqq->next_rq))
908 else if (sector < blk_rq_pos(cfqq->next_rq))
916 *ret_parent = parent;
922 static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
924 struct rb_node **p, *parent;
925 struct cfq_queue *__cfqq;
928 rb_erase(&cfqq->p_node, cfqq->p_root);
932 if (cfq_class_idle(cfqq))
937 cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
938 __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
939 blk_rq_pos(cfqq->next_rq), &parent, &p);
941 rb_link_node(&cfqq->p_node, parent, p);
942 rb_insert_color(&cfqq->p_node, cfqq->p_root);
948 * Update cfqq's position in the service tree.
950 static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
953 * Resorting requires the cfqq to be on the RR list already.
955 if (cfq_cfqq_on_rr(cfqq)) {
956 cfq_service_tree_add(cfqd, cfqq, 0);
957 cfq_prio_tree_add(cfqd, cfqq);
962 * add to busy list of queues for service, trying to be fair in ordering
963 * the pending list according to last request service
965 static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
967 cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
968 BUG_ON(cfq_cfqq_on_rr(cfqq));
969 cfq_mark_cfqq_on_rr(cfqq);
972 cfq_resort_rr_list(cfqd, cfqq);
976 * Called when the cfqq no longer has requests pending, remove it from
979 static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
981 cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
982 BUG_ON(!cfq_cfqq_on_rr(cfqq));
983 cfq_clear_cfqq_on_rr(cfqq);
985 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
986 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
987 cfqq->service_tree = NULL;
990 rb_erase(&cfqq->p_node, cfqq->p_root);
994 cfq_group_service_tree_del(cfqd, cfqq->cfqg);
995 BUG_ON(!cfqd->busy_queues);
1000 * rb tree support functions
1002 static void cfq_del_rq_rb(struct request *rq)
1004 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1005 const int sync = rq_is_sync(rq);
1007 BUG_ON(!cfqq->queued[sync]);
1008 cfqq->queued[sync]--;
1010 elv_rb_del(&cfqq->sort_list, rq);
1012 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) {
1014 * Queue will be deleted from service tree when we actually
1015 * expire it later. Right now just remove it from prio tree
1019 rb_erase(&cfqq->p_node, cfqq->p_root);
1020 cfqq->p_root = NULL;
1025 static void cfq_add_rq_rb(struct request *rq)
1027 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1028 struct cfq_data *cfqd = cfqq->cfqd;
1029 struct request *__alias, *prev;
1031 cfqq->queued[rq_is_sync(rq)]++;
1034 * looks a little odd, but the first insert might return an alias.
1035 * if that happens, put the alias on the dispatch list
1037 while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL)
1038 cfq_dispatch_insert(cfqd->queue, __alias);
1040 if (!cfq_cfqq_on_rr(cfqq))
1041 cfq_add_cfqq_rr(cfqd, cfqq);
1044 * check if this request is a better next-serve candidate
1046 prev = cfqq->next_rq;
1047 cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq, cfqd->last_position);
1050 * adjust priority tree position, if ->next_rq changes
1052 if (prev != cfqq->next_rq)
1053 cfq_prio_tree_add(cfqd, cfqq);
1055 BUG_ON(!cfqq->next_rq);
1058 static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
1060 elv_rb_del(&cfqq->sort_list, rq);
1061 cfqq->queued[rq_is_sync(rq)]--;
1065 static struct request *
1066 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
1068 struct task_struct *tsk = current;
1069 struct cfq_io_context *cic;
1070 struct cfq_queue *cfqq;
1072 cic = cfq_cic_lookup(cfqd, tsk->io_context);
1076 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
1078 sector_t sector = bio->bi_sector + bio_sectors(bio);
1080 return elv_rb_find(&cfqq->sort_list, sector);
1086 static void cfq_activate_request(struct request_queue *q, struct request *rq)
1088 struct cfq_data *cfqd = q->elevator->elevator_data;
1090 cfqd->rq_in_driver[rq_is_sync(rq)]++;
1091 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
1092 rq_in_driver(cfqd));
1094 cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
1097 static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
1099 struct cfq_data *cfqd = q->elevator->elevator_data;
1100 const int sync = rq_is_sync(rq);
1102 WARN_ON(!cfqd->rq_in_driver[sync]);
1103 cfqd->rq_in_driver[sync]--;
1104 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
1105 rq_in_driver(cfqd));
1108 static void cfq_remove_request(struct request *rq)
1110 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1112 if (cfqq->next_rq == rq)
1113 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
1115 list_del_init(&rq->queuelist);
1118 cfqq->cfqd->rq_queued--;
1119 if (rq_is_meta(rq)) {
1120 WARN_ON(!cfqq->meta_pending);
1121 cfqq->meta_pending--;
1125 static int cfq_merge(struct request_queue *q, struct request **req,
1128 struct cfq_data *cfqd = q->elevator->elevator_data;
1129 struct request *__rq;
1131 __rq = cfq_find_rq_fmerge(cfqd, bio);
1132 if (__rq && elv_rq_merge_ok(__rq, bio)) {
1134 return ELEVATOR_FRONT_MERGE;
1137 return ELEVATOR_NO_MERGE;
1140 static void cfq_merged_request(struct request_queue *q, struct request *req,
1143 if (type == ELEVATOR_FRONT_MERGE) {
1144 struct cfq_queue *cfqq = RQ_CFQQ(req);
1146 cfq_reposition_rq_rb(cfqq, req);
1151 cfq_merged_requests(struct request_queue *q, struct request *rq,
1152 struct request *next)
1154 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1156 * reposition in fifo if next is older than rq
1158 if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
1159 time_before(rq_fifo_time(next), rq_fifo_time(rq))) {
1160 list_move(&rq->queuelist, &next->queuelist);
1161 rq_set_fifo_time(rq, rq_fifo_time(next));
1164 if (cfqq->next_rq == next)
1166 cfq_remove_request(next);
1169 static int cfq_allow_merge(struct request_queue *q, struct request *rq,
1172 struct cfq_data *cfqd = q->elevator->elevator_data;
1173 struct cfq_io_context *cic;
1174 struct cfq_queue *cfqq;
1177 * Disallow merge of a sync bio into an async request.
1179 if (cfq_bio_sync(bio) && !rq_is_sync(rq))
1183 * Lookup the cfqq that this bio will be queued with. Allow
1184 * merge only if rq is queued there.
1186 cic = cfq_cic_lookup(cfqd, current->io_context);
1190 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
1191 return cfqq == RQ_CFQQ(rq);
1194 static void __cfq_set_active_queue(struct cfq_data *cfqd,
1195 struct cfq_queue *cfqq)
1198 cfq_log_cfqq(cfqd, cfqq, "set_active");
1199 cfqq->slice_end = 0;
1200 cfqq->slice_dispatch = 0;
1202 cfq_clear_cfqq_wait_request(cfqq);
1203 cfq_clear_cfqq_must_dispatch(cfqq);
1204 cfq_clear_cfqq_must_alloc_slice(cfqq);
1205 cfq_clear_cfqq_fifo_expire(cfqq);
1206 cfq_mark_cfqq_slice_new(cfqq);
1208 del_timer(&cfqd->idle_slice_timer);
1211 cfqd->active_queue = cfqq;
1215 * current cfqq expired its slice (or was too idle), select new one
1218 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1221 cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
1223 if (cfq_cfqq_wait_request(cfqq))
1224 del_timer(&cfqd->idle_slice_timer);
1226 cfq_clear_cfqq_wait_request(cfqq);
1229 * store what was left of this slice, if the queue idled/timed out
1231 if (timed_out && !cfq_cfqq_slice_new(cfqq)) {
1232 cfqq->slice_resid = cfqq->slice_end - jiffies;
1233 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
1236 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
1237 cfq_del_cfqq_rr(cfqd, cfqq);
1239 cfq_resort_rr_list(cfqd, cfqq);
1241 if (cfqq == cfqd->active_queue)
1242 cfqd->active_queue = NULL;
1244 if (cfqd->active_cic) {
1245 put_io_context(cfqd->active_cic->ioc);
1246 cfqd->active_cic = NULL;
1250 static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out)
1252 struct cfq_queue *cfqq = cfqd->active_queue;
1255 __cfq_slice_expired(cfqd, cfqq, timed_out);
1259 * Get next queue for service. Unless we have a queue preemption,
1260 * we'll simply select the first cfqq in the service tree.
1262 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
1264 struct cfq_rb_root *service_tree =
1265 service_tree_for(cfqd->serving_group, cfqd->serving_prio,
1266 cfqd->serving_type, cfqd);
1268 if (!cfqd->rq_queued)
1271 /* There is nothing to dispatch */
1274 if (RB_EMPTY_ROOT(&service_tree->rb))
1276 return cfq_rb_first(service_tree);
1279 static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd)
1281 struct cfq_group *cfqg = &cfqd->root_group;
1282 struct cfq_queue *cfqq;
1284 struct cfq_rb_root *st;
1286 if (!cfqd->rq_queued)
1289 for_each_cfqg_st(cfqg, i, j, st)
1290 if ((cfqq = cfq_rb_first(st)) != NULL)
1296 * Get and set a new active queue for service.
1298 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
1299 struct cfq_queue *cfqq)
1302 cfqq = cfq_get_next_queue(cfqd);
1304 __cfq_set_active_queue(cfqd, cfqq);
1308 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
1311 if (blk_rq_pos(rq) >= cfqd->last_position)
1312 return blk_rq_pos(rq) - cfqd->last_position;
1314 return cfqd->last_position - blk_rq_pos(rq);
1317 #define CFQQ_SEEK_THR 8 * 1024
1318 #define CFQQ_SEEKY(cfqq) ((cfqq)->seek_mean > CFQQ_SEEK_THR)
1320 static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1323 sector_t sdist = cfqq->seek_mean;
1325 if (!sample_valid(cfqq->seek_samples))
1326 sdist = CFQQ_SEEK_THR;
1328 return cfq_dist_from_last(cfqd, rq) <= sdist;
1331 static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
1332 struct cfq_queue *cur_cfqq)
1334 struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
1335 struct rb_node *parent, *node;
1336 struct cfq_queue *__cfqq;
1337 sector_t sector = cfqd->last_position;
1339 if (RB_EMPTY_ROOT(root))
1343 * First, if we find a request starting at the end of the last
1344 * request, choose it.
1346 __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
1351 * If the exact sector wasn't found, the parent of the NULL leaf
1352 * will contain the closest sector.
1354 __cfqq = rb_entry(parent, struct cfq_queue, p_node);
1355 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1358 if (blk_rq_pos(__cfqq->next_rq) < sector)
1359 node = rb_next(&__cfqq->p_node);
1361 node = rb_prev(&__cfqq->p_node);
1365 __cfqq = rb_entry(node, struct cfq_queue, p_node);
1366 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1374 * cur_cfqq - passed in so that we don't decide that the current queue is
1375 * closely cooperating with itself.
1377 * So, basically we're assuming that that cur_cfqq has dispatched at least
1378 * one request, and that cfqd->last_position reflects a position on the disk
1379 * associated with the I/O issued by cur_cfqq. I'm not sure this is a valid
1382 static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
1383 struct cfq_queue *cur_cfqq)
1385 struct cfq_queue *cfqq;
1387 if (!cfq_cfqq_sync(cur_cfqq))
1389 if (CFQQ_SEEKY(cur_cfqq))
1393 * We should notice if some of the queues are cooperating, eg
1394 * working closely on the same area of the disk. In that case,
1395 * we can group them together and don't waste time idling.
1397 cfqq = cfqq_close(cfqd, cur_cfqq);
1402 * It only makes sense to merge sync queues.
1404 if (!cfq_cfqq_sync(cfqq))
1406 if (CFQQ_SEEKY(cfqq))
1410 * Do not merge queues of different priority classes
1412 if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq))
1419 * Determine whether we should enforce idle window for this queue.
1422 static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1424 enum wl_prio_t prio = cfqq_prio(cfqq);
1425 struct cfq_rb_root *service_tree = cfqq->service_tree;
1427 BUG_ON(!service_tree);
1428 BUG_ON(!service_tree->count);
1430 /* We never do for idle class queues. */
1431 if (prio == IDLE_WORKLOAD)
1434 /* We do for queues that were marked with idle window flag. */
1435 if (cfq_cfqq_idle_window(cfqq))
1439 * Otherwise, we do only if they are the last ones
1440 * in their service tree.
1442 return service_tree->count == 1;
1445 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
1447 struct cfq_queue *cfqq = cfqd->active_queue;
1448 struct cfq_io_context *cic;
1452 * SSD device without seek penalty, disable idling. But only do so
1453 * for devices that support queuing, otherwise we still have a problem
1454 * with sync vs async workloads.
1456 if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
1459 WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
1460 WARN_ON(cfq_cfqq_slice_new(cfqq));
1463 * idle is disabled, either manually or by past process history
1465 if (!cfqd->cfq_slice_idle || !cfq_should_idle(cfqd, cfqq))
1469 * still active requests from this queue, don't idle
1471 if (cfqq->dispatched)
1475 * task has exited, don't wait
1477 cic = cfqd->active_cic;
1478 if (!cic || !atomic_read(&cic->ioc->nr_tasks))
1482 * If our average think time is larger than the remaining time
1483 * slice, then don't idle. This avoids overrunning the allotted
1486 if (sample_valid(cic->ttime_samples) &&
1487 (cfqq->slice_end - jiffies < cic->ttime_mean))
1490 cfq_mark_cfqq_wait_request(cfqq);
1492 sl = cfqd->cfq_slice_idle;
1494 mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
1495 cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu", sl);
1499 * Move request from internal lists to the request queue dispatch list.
1501 static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
1503 struct cfq_data *cfqd = q->elevator->elevator_data;
1504 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1506 cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
1508 cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
1509 cfq_remove_request(rq);
1511 elv_dispatch_sort(q, rq);
1513 if (cfq_cfqq_sync(cfqq))
1514 cfqd->sync_flight++;
1518 * return expired entry, or NULL to just start from scratch in rbtree
1520 static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
1522 struct request *rq = NULL;
1524 if (cfq_cfqq_fifo_expire(cfqq))
1527 cfq_mark_cfqq_fifo_expire(cfqq);
1529 if (list_empty(&cfqq->fifo))
1532 rq = rq_entry_fifo(cfqq->fifo.next);
1533 if (time_before(jiffies, rq_fifo_time(rq)))
1536 cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq);
1541 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1543 const int base_rq = cfqd->cfq_slice_async_rq;
1545 WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
1547 return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
1551 * Must be called with the queue_lock held.
1553 static int cfqq_process_refs(struct cfq_queue *cfqq)
1555 int process_refs, io_refs;
1557 io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE];
1558 process_refs = atomic_read(&cfqq->ref) - io_refs;
1559 BUG_ON(process_refs < 0);
1560 return process_refs;
1563 static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq)
1565 int process_refs, new_process_refs;
1566 struct cfq_queue *__cfqq;
1568 /* Avoid a circular list and skip interim queue merges */
1569 while ((__cfqq = new_cfqq->new_cfqq)) {
1575 process_refs = cfqq_process_refs(cfqq);
1577 * If the process for the cfqq has gone away, there is no
1578 * sense in merging the queues.
1580 if (process_refs == 0)
1584 * Merge in the direction of the lesser amount of work.
1586 new_process_refs = cfqq_process_refs(new_cfqq);
1587 if (new_process_refs >= process_refs) {
1588 cfqq->new_cfqq = new_cfqq;
1589 atomic_add(process_refs, &new_cfqq->ref);
1591 new_cfqq->new_cfqq = cfqq;
1592 atomic_add(new_process_refs, &cfqq->ref);
1596 static enum wl_type_t cfq_choose_wl(struct cfq_data *cfqd,
1597 struct cfq_group *cfqg, enum wl_prio_t prio,
1600 struct cfq_queue *queue;
1602 bool key_valid = false;
1603 unsigned long lowest_key = 0;
1604 enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD;
1608 * When priorities switched, we prefer starting
1609 * from SYNC_NOIDLE (first choice), or just SYNC
1612 if (service_tree_for(cfqg, prio, cur_best, cfqd)->count)
1614 cur_best = SYNC_WORKLOAD;
1615 if (service_tree_for(cfqg, prio, cur_best, cfqd)->count)
1618 return ASYNC_WORKLOAD;
1621 for (i = 0; i < 3; ++i) {
1622 /* otherwise, select the one with lowest rb_key */
1623 queue = cfq_rb_first(service_tree_for(cfqg, prio, i, cfqd));
1625 (!key_valid || time_before(queue->rb_key, lowest_key))) {
1626 lowest_key = queue->rb_key;
1635 static void choose_service_tree(struct cfq_data *cfqd, struct cfq_group *cfqg)
1637 enum wl_prio_t previous_prio = cfqd->serving_prio;
1641 struct cfq_rb_root *st;
1644 cfqd->serving_prio = IDLE_WORKLOAD;
1645 cfqd->workload_expires = jiffies + 1;
1649 /* Choose next priority. RT > BE > IDLE */
1650 if (cfq_busy_queues_wl(RT_WORKLOAD, cfqd))
1651 cfqd->serving_prio = RT_WORKLOAD;
1652 else if (cfq_busy_queues_wl(BE_WORKLOAD, cfqd))
1653 cfqd->serving_prio = BE_WORKLOAD;
1655 cfqd->serving_prio = IDLE_WORKLOAD;
1656 cfqd->workload_expires = jiffies + 1;
1661 * For RT and BE, we have to choose also the type
1662 * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
1665 prio_changed = (cfqd->serving_prio != previous_prio);
1666 st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type,
1671 * If priority didn't change, check workload expiration,
1672 * and that we still have other queues ready
1674 if (!prio_changed && count &&
1675 !time_after(jiffies, cfqd->workload_expires))
1678 /* otherwise select new workload type */
1679 cfqd->serving_type =
1680 cfq_choose_wl(cfqd, cfqg, cfqd->serving_prio, prio_changed);
1681 st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type,
1686 * the workload slice is computed as a fraction of target latency
1687 * proportional to the number of queues in that workload, over
1688 * all the queues in the same priority class
1690 slice = cfq_target_latency * count /
1691 max_t(unsigned, cfqd->busy_queues_avg[cfqd->serving_prio],
1692 cfq_busy_queues_wl(cfqd->serving_prio, cfqd));
1694 if (cfqd->serving_type == ASYNC_WORKLOAD)
1695 /* async workload slice is scaled down according to
1696 * the sync/async slice ratio. */
1697 slice = slice * cfqd->cfq_slice[0] / cfqd->cfq_slice[1];
1699 /* sync workload slice is at least 2 * cfq_slice_idle */
1700 slice = max(slice, 2 * cfqd->cfq_slice_idle);
1702 slice = max_t(unsigned, slice, CFQ_MIN_TT);
1703 cfqd->workload_expires = jiffies + slice;
1704 cfqd->noidle_tree_requires_idle = false;
1707 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd)
1709 struct cfq_rb_root *st = &cfqd->grp_service_tree;
1710 struct cfq_group *cfqg;
1712 if (RB_EMPTY_ROOT(&st->rb))
1714 cfqg = cfq_rb_first_group(st);
1715 st->active = &cfqg->rb_node;
1716 update_min_vdisktime(st);
1720 static void cfq_choose_cfqg(struct cfq_data *cfqd)
1722 struct cfq_group *cfqg = cfq_get_next_cfqg(cfqd);
1724 cfqd->serving_group = cfqg;
1725 choose_service_tree(cfqd, cfqg);
1729 * Select a queue for service. If we have a current active queue,
1730 * check whether to continue servicing it, or retrieve and set a new one.
1732 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
1734 struct cfq_queue *cfqq, *new_cfqq = NULL;
1736 cfqq = cfqd->active_queue;
1740 if (!cfqd->rq_queued)
1743 * The active queue has run out of time, expire it and select new.
1745 if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq))
1749 * The active queue has requests and isn't expired, allow it to
1752 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
1756 * If another queue has a request waiting within our mean seek
1757 * distance, let it run. The expire code will check for close
1758 * cooperators and put the close queue at the front of the service
1759 * tree. If possible, merge the expiring queue with the new cfqq.
1761 new_cfqq = cfq_close_cooperator(cfqd, cfqq);
1763 if (!cfqq->new_cfqq)
1764 cfq_setup_merge(cfqq, new_cfqq);
1769 * No requests pending. If the active queue still has requests in
1770 * flight or is idling for a new request, allow either of these
1771 * conditions to happen (or time out) before selecting a new queue.
1773 if (timer_pending(&cfqd->idle_slice_timer) ||
1774 (cfqq->dispatched && cfq_should_idle(cfqd, cfqq))) {
1780 cfq_slice_expired(cfqd, 0);
1783 * Current queue expired. Check if we have to switch to a new
1787 cfq_choose_cfqg(cfqd);
1789 cfqq = cfq_set_active_queue(cfqd, new_cfqq);
1794 static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
1798 while (cfqq->next_rq) {
1799 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
1803 BUG_ON(!list_empty(&cfqq->fifo));
1805 /* By default cfqq is not expired if it is empty. Do it explicitly */
1806 __cfq_slice_expired(cfqq->cfqd, cfqq, 0);
1811 * Drain our current requests. Used for barriers and when switching
1812 * io schedulers on-the-fly.
1814 static int cfq_forced_dispatch(struct cfq_data *cfqd)
1816 struct cfq_queue *cfqq;
1819 while ((cfqq = cfq_get_next_queue_forced(cfqd)) != NULL)
1820 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
1822 cfq_slice_expired(cfqd, 0);
1823 BUG_ON(cfqd->busy_queues);
1825 cfq_log(cfqd, "forced_dispatch=%d", dispatched);
1829 static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1831 unsigned int max_dispatch;
1834 * Drain async requests before we start sync IO
1836 if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_driver[BLK_RW_ASYNC])
1840 * If this is an async queue and we have sync IO in flight, let it wait
1842 if (cfqd->sync_flight && !cfq_cfqq_sync(cfqq))
1845 max_dispatch = cfqd->cfq_quantum;
1846 if (cfq_class_idle(cfqq))
1850 * Does this cfqq already have too much IO in flight?
1852 if (cfqq->dispatched >= max_dispatch) {
1854 * idle queue must always only have a single IO in flight
1856 if (cfq_class_idle(cfqq))
1860 * We have other queues, don't allow more IO from this one
1862 if (cfqd->busy_queues > 1)
1866 * Sole queue user, no limit
1872 * Async queues must wait a bit before being allowed dispatch.
1873 * We also ramp up the dispatch depth gradually for async IO,
1874 * based on the last sync IO we serviced
1876 if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
1877 unsigned long last_sync = jiffies - cfqd->last_end_sync_rq;
1880 depth = last_sync / cfqd->cfq_slice[1];
1881 if (!depth && !cfqq->dispatched)
1883 if (depth < max_dispatch)
1884 max_dispatch = depth;
1888 * If we're below the current max, allow a dispatch
1890 return cfqq->dispatched < max_dispatch;
1894 * Dispatch a request from cfqq, moving them to the request queue
1897 static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1901 BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
1903 if (!cfq_may_dispatch(cfqd, cfqq))
1907 * follow expired path, else get first next available
1909 rq = cfq_check_fifo(cfqq);
1914 * insert request into driver dispatch list
1916 cfq_dispatch_insert(cfqd->queue, rq);
1918 if (!cfqd->active_cic) {
1919 struct cfq_io_context *cic = RQ_CIC(rq);
1921 atomic_long_inc(&cic->ioc->refcount);
1922 cfqd->active_cic = cic;
1929 * Find the cfqq that we need to service and move a request from that to the
1932 static int cfq_dispatch_requests(struct request_queue *q, int force)
1934 struct cfq_data *cfqd = q->elevator->elevator_data;
1935 struct cfq_queue *cfqq;
1937 if (!cfqd->busy_queues)
1940 if (unlikely(force))
1941 return cfq_forced_dispatch(cfqd);
1943 cfqq = cfq_select_queue(cfqd);
1948 * Dispatch a request from this cfqq, if it is allowed
1950 if (!cfq_dispatch_request(cfqd, cfqq))
1953 cfqq->slice_dispatch++;
1954 cfq_clear_cfqq_must_dispatch(cfqq);
1957 * expire an async queue immediately if it has used up its slice. idle
1958 * queue always expire after 1 dispatch round.
1960 if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
1961 cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
1962 cfq_class_idle(cfqq))) {
1963 cfqq->slice_end = jiffies + 1;
1964 cfq_slice_expired(cfqd, 0);
1967 cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
1972 * task holds one reference to the queue, dropped when task exits. each rq
1973 * in-flight on this queue also holds a reference, dropped when rq is freed.
1975 * queue lock must be held here.
1977 static void cfq_put_queue(struct cfq_queue *cfqq)
1979 struct cfq_data *cfqd = cfqq->cfqd;
1981 BUG_ON(atomic_read(&cfqq->ref) <= 0);
1983 if (!atomic_dec_and_test(&cfqq->ref))
1986 cfq_log_cfqq(cfqd, cfqq, "put_queue");
1987 BUG_ON(rb_first(&cfqq->sort_list));
1988 BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
1990 if (unlikely(cfqd->active_queue == cfqq)) {
1991 __cfq_slice_expired(cfqd, cfqq, 0);
1992 cfq_schedule_dispatch(cfqd);
1995 BUG_ON(cfq_cfqq_on_rr(cfqq));
1996 kmem_cache_free(cfq_pool, cfqq);
2000 * Must always be called with the rcu_read_lock() held
2003 __call_for_each_cic(struct io_context *ioc,
2004 void (*func)(struct io_context *, struct cfq_io_context *))
2006 struct cfq_io_context *cic;
2007 struct hlist_node *n;
2009 hlist_for_each_entry_rcu(cic, n, &ioc->cic_list, cic_list)
2014 * Call func for each cic attached to this ioc.
2017 call_for_each_cic(struct io_context *ioc,
2018 void (*func)(struct io_context *, struct cfq_io_context *))
2021 __call_for_each_cic(ioc, func);
2025 static void cfq_cic_free_rcu(struct rcu_head *head)
2027 struct cfq_io_context *cic;
2029 cic = container_of(head, struct cfq_io_context, rcu_head);
2031 kmem_cache_free(cfq_ioc_pool, cic);
2032 elv_ioc_count_dec(cfq_ioc_count);
2036 * CFQ scheduler is exiting, grab exit lock and check
2037 * the pending io context count. If it hits zero,
2038 * complete ioc_gone and set it back to NULL
2040 spin_lock(&ioc_gone_lock);
2041 if (ioc_gone && !elv_ioc_count_read(cfq_ioc_count)) {
2045 spin_unlock(&ioc_gone_lock);
2049 static void cfq_cic_free(struct cfq_io_context *cic)
2051 call_rcu(&cic->rcu_head, cfq_cic_free_rcu);
2054 static void cic_free_func(struct io_context *ioc, struct cfq_io_context *cic)
2056 unsigned long flags;
2058 BUG_ON(!cic->dead_key);
2060 spin_lock_irqsave(&ioc->lock, flags);
2061 radix_tree_delete(&ioc->radix_root, cic->dead_key);
2062 hlist_del_rcu(&cic->cic_list);
2063 spin_unlock_irqrestore(&ioc->lock, flags);
2069 * Must be called with rcu_read_lock() held or preemption otherwise disabled.
2070 * Only two callers of this - ->dtor() which is called with the rcu_read_lock(),
2071 * and ->trim() which is called with the task lock held
2073 static void cfq_free_io_context(struct io_context *ioc)
2076 * ioc->refcount is zero here, or we are called from elv_unregister(),
2077 * so no more cic's are allowed to be linked into this ioc. So it
2078 * should be ok to iterate over the known list, we will see all cic's
2079 * since no new ones are added.
2081 __call_for_each_cic(ioc, cic_free_func);
2084 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2086 struct cfq_queue *__cfqq, *next;
2088 if (unlikely(cfqq == cfqd->active_queue)) {
2089 __cfq_slice_expired(cfqd, cfqq, 0);
2090 cfq_schedule_dispatch(cfqd);
2094 * If this queue was scheduled to merge with another queue, be
2095 * sure to drop the reference taken on that queue (and others in
2096 * the merge chain). See cfq_setup_merge and cfq_merge_cfqqs.
2098 __cfqq = cfqq->new_cfqq;
2100 if (__cfqq == cfqq) {
2101 WARN(1, "cfqq->new_cfqq loop detected\n");
2104 next = __cfqq->new_cfqq;
2105 cfq_put_queue(__cfqq);
2109 cfq_put_queue(cfqq);
2112 static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
2113 struct cfq_io_context *cic)
2115 struct io_context *ioc = cic->ioc;
2117 list_del_init(&cic->queue_list);
2120 * Make sure key == NULL is seen for dead queues
2123 cic->dead_key = (unsigned long) cic->key;
2126 if (ioc->ioc_data == cic)
2127 rcu_assign_pointer(ioc->ioc_data, NULL);
2129 if (cic->cfqq[BLK_RW_ASYNC]) {
2130 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]);
2131 cic->cfqq[BLK_RW_ASYNC] = NULL;
2134 if (cic->cfqq[BLK_RW_SYNC]) {
2135 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]);
2136 cic->cfqq[BLK_RW_SYNC] = NULL;
2140 static void cfq_exit_single_io_context(struct io_context *ioc,
2141 struct cfq_io_context *cic)
2143 struct cfq_data *cfqd = cic->key;
2146 struct request_queue *q = cfqd->queue;
2147 unsigned long flags;
2149 spin_lock_irqsave(q->queue_lock, flags);
2152 * Ensure we get a fresh copy of the ->key to prevent
2153 * race between exiting task and queue
2155 smp_read_barrier_depends();
2157 __cfq_exit_single_io_context(cfqd, cic);
2159 spin_unlock_irqrestore(q->queue_lock, flags);
2164 * The process that ioc belongs to has exited, we need to clean up
2165 * and put the internal structures we have that belongs to that process.
2167 static void cfq_exit_io_context(struct io_context *ioc)
2169 call_for_each_cic(ioc, cfq_exit_single_io_context);
2172 static struct cfq_io_context *
2173 cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
2175 struct cfq_io_context *cic;
2177 cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO,
2180 cic->last_end_request = jiffies;
2181 INIT_LIST_HEAD(&cic->queue_list);
2182 INIT_HLIST_NODE(&cic->cic_list);
2183 cic->dtor = cfq_free_io_context;
2184 cic->exit = cfq_exit_io_context;
2185 elv_ioc_count_inc(cfq_ioc_count);
2191 static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc)
2193 struct task_struct *tsk = current;
2196 if (!cfq_cfqq_prio_changed(cfqq))
2199 ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio);
2200 switch (ioprio_class) {
2202 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
2203 case IOPRIO_CLASS_NONE:
2205 * no prio set, inherit CPU scheduling settings
2207 cfqq->ioprio = task_nice_ioprio(tsk);
2208 cfqq->ioprio_class = task_nice_ioclass(tsk);
2210 case IOPRIO_CLASS_RT:
2211 cfqq->ioprio = task_ioprio(ioc);
2212 cfqq->ioprio_class = IOPRIO_CLASS_RT;
2214 case IOPRIO_CLASS_BE:
2215 cfqq->ioprio = task_ioprio(ioc);
2216 cfqq->ioprio_class = IOPRIO_CLASS_BE;
2218 case IOPRIO_CLASS_IDLE:
2219 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
2221 cfq_clear_cfqq_idle_window(cfqq);
2226 * keep track of original prio settings in case we have to temporarily
2227 * elevate the priority of this queue
2229 cfqq->org_ioprio = cfqq->ioprio;
2230 cfqq->org_ioprio_class = cfqq->ioprio_class;
2231 cfq_clear_cfqq_prio_changed(cfqq);
2234 static void changed_ioprio(struct io_context *ioc, struct cfq_io_context *cic)
2236 struct cfq_data *cfqd = cic->key;
2237 struct cfq_queue *cfqq;
2238 unsigned long flags;
2240 if (unlikely(!cfqd))
2243 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2245 cfqq = cic->cfqq[BLK_RW_ASYNC];
2247 struct cfq_queue *new_cfqq;
2248 new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic->ioc,
2251 cic->cfqq[BLK_RW_ASYNC] = new_cfqq;
2252 cfq_put_queue(cfqq);
2256 cfqq = cic->cfqq[BLK_RW_SYNC];
2258 cfq_mark_cfqq_prio_changed(cfqq);
2260 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2263 static void cfq_ioc_set_ioprio(struct io_context *ioc)
2265 call_for_each_cic(ioc, changed_ioprio);
2266 ioc->ioprio_changed = 0;
2269 static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2270 pid_t pid, bool is_sync)
2272 RB_CLEAR_NODE(&cfqq->rb_node);
2273 RB_CLEAR_NODE(&cfqq->p_node);
2274 INIT_LIST_HEAD(&cfqq->fifo);
2276 atomic_set(&cfqq->ref, 0);
2279 cfq_mark_cfqq_prio_changed(cfqq);
2282 if (!cfq_class_idle(cfqq))
2283 cfq_mark_cfqq_idle_window(cfqq);
2284 cfq_mark_cfqq_sync(cfqq);
2289 static void cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg)
2294 static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd, int create)
2296 return &cfqd->root_group;
2299 static struct cfq_queue *
2300 cfq_find_alloc_queue(struct cfq_data *cfqd, bool is_sync,
2301 struct io_context *ioc, gfp_t gfp_mask)
2303 struct cfq_queue *cfqq, *new_cfqq = NULL;
2304 struct cfq_io_context *cic;
2305 struct cfq_group *cfqg;
2308 cfqg = cfq_get_cfqg(cfqd, 1);
2309 cic = cfq_cic_lookup(cfqd, ioc);
2310 /* cic always exists here */
2311 cfqq = cic_to_cfqq(cic, is_sync);
2314 * Always try a new alloc if we fell back to the OOM cfqq
2315 * originally, since it should just be a temporary situation.
2317 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
2322 } else if (gfp_mask & __GFP_WAIT) {
2323 spin_unlock_irq(cfqd->queue->queue_lock);
2324 new_cfqq = kmem_cache_alloc_node(cfq_pool,
2325 gfp_mask | __GFP_ZERO,
2327 spin_lock_irq(cfqd->queue->queue_lock);
2331 cfqq = kmem_cache_alloc_node(cfq_pool,
2332 gfp_mask | __GFP_ZERO,
2337 cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
2338 cfq_init_prio_data(cfqq, ioc);
2339 cfq_link_cfqq_cfqg(cfqq, cfqg);
2340 cfq_log_cfqq(cfqd, cfqq, "alloced");
2342 cfqq = &cfqd->oom_cfqq;
2346 kmem_cache_free(cfq_pool, new_cfqq);
2351 static struct cfq_queue **
2352 cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
2354 switch (ioprio_class) {
2355 case IOPRIO_CLASS_RT:
2356 return &cfqd->async_cfqq[0][ioprio];
2357 case IOPRIO_CLASS_BE:
2358 return &cfqd->async_cfqq[1][ioprio];
2359 case IOPRIO_CLASS_IDLE:
2360 return &cfqd->async_idle_cfqq;
2366 static struct cfq_queue *
2367 cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct io_context *ioc,
2370 const int ioprio = task_ioprio(ioc);
2371 const int ioprio_class = task_ioprio_class(ioc);
2372 struct cfq_queue **async_cfqq = NULL;
2373 struct cfq_queue *cfqq = NULL;
2376 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
2381 cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask);
2384 * pin the queue now that it's allocated, scheduler exit will prune it
2386 if (!is_sync && !(*async_cfqq)) {
2387 atomic_inc(&cfqq->ref);
2391 atomic_inc(&cfqq->ref);
2396 * We drop cfq io contexts lazily, so we may find a dead one.
2399 cfq_drop_dead_cic(struct cfq_data *cfqd, struct io_context *ioc,
2400 struct cfq_io_context *cic)
2402 unsigned long flags;
2404 WARN_ON(!list_empty(&cic->queue_list));
2406 spin_lock_irqsave(&ioc->lock, flags);
2408 BUG_ON(ioc->ioc_data == cic);
2410 radix_tree_delete(&ioc->radix_root, (unsigned long) cfqd);
2411 hlist_del_rcu(&cic->cic_list);
2412 spin_unlock_irqrestore(&ioc->lock, flags);
2417 static struct cfq_io_context *
2418 cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc)
2420 struct cfq_io_context *cic;
2421 unsigned long flags;
2430 * we maintain a last-hit cache, to avoid browsing over the tree
2432 cic = rcu_dereference(ioc->ioc_data);
2433 if (cic && cic->key == cfqd) {
2439 cic = radix_tree_lookup(&ioc->radix_root, (unsigned long) cfqd);
2443 /* ->key must be copied to avoid race with cfq_exit_queue() */
2446 cfq_drop_dead_cic(cfqd, ioc, cic);
2451 spin_lock_irqsave(&ioc->lock, flags);
2452 rcu_assign_pointer(ioc->ioc_data, cic);
2453 spin_unlock_irqrestore(&ioc->lock, flags);
2461 * Add cic into ioc, using cfqd as the search key. This enables us to lookup
2462 * the process specific cfq io context when entered from the block layer.
2463 * Also adds the cic to a per-cfqd list, used when this queue is removed.
2465 static int cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
2466 struct cfq_io_context *cic, gfp_t gfp_mask)
2468 unsigned long flags;
2471 ret = radix_tree_preload(gfp_mask);
2476 spin_lock_irqsave(&ioc->lock, flags);
2477 ret = radix_tree_insert(&ioc->radix_root,
2478 (unsigned long) cfqd, cic);
2480 hlist_add_head_rcu(&cic->cic_list, &ioc->cic_list);
2481 spin_unlock_irqrestore(&ioc->lock, flags);
2483 radix_tree_preload_end();
2486 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2487 list_add(&cic->queue_list, &cfqd->cic_list);
2488 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2493 printk(KERN_ERR "cfq: cic link failed!\n");
2499 * Setup general io context and cfq io context. There can be several cfq
2500 * io contexts per general io context, if this process is doing io to more
2501 * than one device managed by cfq.
2503 static struct cfq_io_context *
2504 cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
2506 struct io_context *ioc = NULL;
2507 struct cfq_io_context *cic;
2509 might_sleep_if(gfp_mask & __GFP_WAIT);
2511 ioc = get_io_context(gfp_mask, cfqd->queue->node);
2515 cic = cfq_cic_lookup(cfqd, ioc);
2519 cic = cfq_alloc_io_context(cfqd, gfp_mask);
2523 if (cfq_cic_link(cfqd, ioc, cic, gfp_mask))
2527 smp_read_barrier_depends();
2528 if (unlikely(ioc->ioprio_changed))
2529 cfq_ioc_set_ioprio(ioc);
2535 put_io_context(ioc);
2540 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
2542 unsigned long elapsed = jiffies - cic->last_end_request;
2543 unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
2545 cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
2546 cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
2547 cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
2551 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2557 if (!cfqq->last_request_pos)
2559 else if (cfqq->last_request_pos < blk_rq_pos(rq))
2560 sdist = blk_rq_pos(rq) - cfqq->last_request_pos;
2562 sdist = cfqq->last_request_pos - blk_rq_pos(rq);
2565 * Don't allow the seek distance to get too large from the
2566 * odd fragment, pagein, etc
2568 if (cfqq->seek_samples <= 60) /* second&third seek */
2569 sdist = min(sdist, (cfqq->seek_mean * 4) + 2*1024*1024);
2571 sdist = min(sdist, (cfqq->seek_mean * 4) + 2*1024*64);
2573 cfqq->seek_samples = (7*cfqq->seek_samples + 256) / 8;
2574 cfqq->seek_total = (7*cfqq->seek_total + (u64)256*sdist) / 8;
2575 total = cfqq->seek_total + (cfqq->seek_samples/2);
2576 do_div(total, cfqq->seek_samples);
2577 cfqq->seek_mean = (sector_t)total;
2580 * If this cfqq is shared between multiple processes, check to
2581 * make sure that those processes are still issuing I/Os within
2582 * the mean seek distance. If not, it may be time to break the
2583 * queues apart again.
2585 if (cfq_cfqq_coop(cfqq)) {
2586 if (CFQQ_SEEKY(cfqq) && !cfqq->seeky_start)
2587 cfqq->seeky_start = jiffies;
2588 else if (!CFQQ_SEEKY(cfqq))
2589 cfqq->seeky_start = 0;
2594 * Disable idle window if the process thinks too long or seeks so much that
2598 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2599 struct cfq_io_context *cic)
2601 int old_idle, enable_idle;
2604 * Don't idle for async or idle io prio class
2606 if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
2609 enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
2611 if (cfqq->queued[0] + cfqq->queued[1] >= 4)
2612 cfq_mark_cfqq_deep(cfqq);
2614 if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle ||
2615 (!cfq_cfqq_deep(cfqq) && sample_valid(cfqq->seek_samples)
2616 && CFQQ_SEEKY(cfqq)))
2618 else if (sample_valid(cic->ttime_samples)) {
2619 if (cic->ttime_mean > cfqd->cfq_slice_idle)
2625 if (old_idle != enable_idle) {
2626 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
2628 cfq_mark_cfqq_idle_window(cfqq);
2630 cfq_clear_cfqq_idle_window(cfqq);
2635 * Check if new_cfqq should preempt the currently active queue. Return 0 for
2636 * no or if we aren't sure, a 1 will cause a preempt.
2639 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
2642 struct cfq_queue *cfqq;
2644 cfqq = cfqd->active_queue;
2648 if (cfq_slice_used(cfqq))
2651 if (cfq_class_idle(new_cfqq))
2654 if (cfq_class_idle(cfqq))
2657 /* Allow preemption only if we are idling on sync-noidle tree */
2658 if (cfqd->serving_type == SYNC_NOIDLE_WORKLOAD &&
2659 cfqq_type(new_cfqq) == SYNC_NOIDLE_WORKLOAD &&
2660 new_cfqq->service_tree->count == 2 &&
2661 RB_EMPTY_ROOT(&cfqq->sort_list))
2665 * if the new request is sync, but the currently running queue is
2666 * not, let the sync request have priority.
2668 if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
2672 * So both queues are sync. Let the new request get disk time if
2673 * it's a metadata request and the current queue is doing regular IO.
2675 if (rq_is_meta(rq) && !cfqq->meta_pending)
2679 * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
2681 if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
2684 if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
2688 * if this request is as-good as one we would expect from the
2689 * current cfqq, let it preempt
2691 if (cfq_rq_close(cfqd, cfqq, rq))
2698 * cfqq preempts the active queue. if we allowed preempt with no slice left,
2699 * let it have half of its nominal slice.
2701 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2703 cfq_log_cfqq(cfqd, cfqq, "preempt");
2704 cfq_slice_expired(cfqd, 1);
2707 * Put the new queue at the front of the of the current list,
2708 * so we know that it will be selected next.
2710 BUG_ON(!cfq_cfqq_on_rr(cfqq));
2712 cfq_service_tree_add(cfqd, cfqq, 1);
2714 cfqq->slice_end = 0;
2715 cfq_mark_cfqq_slice_new(cfqq);
2719 * Called when a new fs request (rq) is added (to cfqq). Check if there's
2720 * something we should do about it
2723 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2726 struct cfq_io_context *cic = RQ_CIC(rq);
2730 cfqq->meta_pending++;
2732 cfq_update_io_thinktime(cfqd, cic);
2733 cfq_update_io_seektime(cfqd, cfqq, rq);
2734 cfq_update_idle_window(cfqd, cfqq, cic);
2736 cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
2738 if (cfqq == cfqd->active_queue) {
2740 * Remember that we saw a request from this process, but
2741 * don't start queuing just yet. Otherwise we risk seeing lots
2742 * of tiny requests, because we disrupt the normal plugging
2743 * and merging. If the request is already larger than a single
2744 * page, let it rip immediately. For that case we assume that
2745 * merging is already done. Ditto for a busy system that
2746 * has other work pending, don't risk delaying until the
2747 * idle timer unplug to continue working.
2749 if (cfq_cfqq_wait_request(cfqq)) {
2750 if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
2751 cfqd->busy_queues > 1) {
2752 del_timer(&cfqd->idle_slice_timer);
2753 __blk_run_queue(cfqd->queue);
2755 cfq_mark_cfqq_must_dispatch(cfqq);
2757 } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
2759 * not the active queue - expire current slice if it is
2760 * idle and has expired it's mean thinktime or this new queue
2761 * has some old slice time left and is of higher priority or
2762 * this new queue is RT and the current one is BE
2764 cfq_preempt_queue(cfqd, cfqq);
2765 __blk_run_queue(cfqd->queue);
2769 static void cfq_insert_request(struct request_queue *q, struct request *rq)
2771 struct cfq_data *cfqd = q->elevator->elevator_data;
2772 struct cfq_queue *cfqq = RQ_CFQQ(rq);
2774 cfq_log_cfqq(cfqd, cfqq, "insert_request");
2775 cfq_init_prio_data(cfqq, RQ_CIC(rq)->ioc);
2777 rq_set_fifo_time(rq, jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)]);
2778 list_add_tail(&rq->queuelist, &cfqq->fifo);
2781 cfq_rq_enqueued(cfqd, cfqq, rq);
2785 * Update hw_tag based on peak queue depth over 50 samples under
2788 static void cfq_update_hw_tag(struct cfq_data *cfqd)
2790 struct cfq_queue *cfqq = cfqd->active_queue;
2792 if (rq_in_driver(cfqd) > cfqd->hw_tag_est_depth)
2793 cfqd->hw_tag_est_depth = rq_in_driver(cfqd);
2795 if (cfqd->hw_tag == 1)
2798 if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
2799 rq_in_driver(cfqd) <= CFQ_HW_QUEUE_MIN)
2803 * If active queue hasn't enough requests and can idle, cfq might not
2804 * dispatch sufficient requests to hardware. Don't zero hw_tag in this
2807 if (cfqq && cfq_cfqq_idle_window(cfqq) &&
2808 cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] <
2809 CFQ_HW_QUEUE_MIN && rq_in_driver(cfqd) < CFQ_HW_QUEUE_MIN)
2812 if (cfqd->hw_tag_samples++ < 50)
2815 if (cfqd->hw_tag_est_depth >= CFQ_HW_QUEUE_MIN)
2821 static void cfq_completed_request(struct request_queue *q, struct request *rq)
2823 struct cfq_queue *cfqq = RQ_CFQQ(rq);
2824 struct cfq_data *cfqd = cfqq->cfqd;
2825 const int sync = rq_is_sync(rq);
2829 cfq_log_cfqq(cfqd, cfqq, "complete");
2831 cfq_update_hw_tag(cfqd);
2833 WARN_ON(!cfqd->rq_in_driver[sync]);
2834 WARN_ON(!cfqq->dispatched);
2835 cfqd->rq_in_driver[sync]--;
2838 if (cfq_cfqq_sync(cfqq))
2839 cfqd->sync_flight--;
2842 RQ_CIC(rq)->last_end_request = now;
2843 cfqd->last_end_sync_rq = now;
2847 * If this is the active queue, check if it needs to be expired,
2848 * or if we want to idle in case it has no pending requests.
2850 if (cfqd->active_queue == cfqq) {
2851 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
2853 if (cfq_cfqq_slice_new(cfqq)) {
2854 cfq_set_prio_slice(cfqd, cfqq);
2855 cfq_clear_cfqq_slice_new(cfqq);
2858 * Idling is not enabled on:
2860 * - idle-priority queues
2862 * - queues with still some requests queued
2863 * - when there is a close cooperator
2865 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
2866 cfq_slice_expired(cfqd, 1);
2867 else if (sync && cfqq_empty &&
2868 !cfq_close_cooperator(cfqd, cfqq)) {
2869 cfqd->noidle_tree_requires_idle |= !rq_noidle(rq);
2871 * Idling is enabled for SYNC_WORKLOAD.
2872 * SYNC_NOIDLE_WORKLOAD idles at the end of the tree
2873 * only if we processed at least one !rq_noidle request
2875 if (cfqd->serving_type == SYNC_WORKLOAD
2876 || cfqd->noidle_tree_requires_idle)
2877 cfq_arm_slice_timer(cfqd);
2881 if (!rq_in_driver(cfqd))
2882 cfq_schedule_dispatch(cfqd);
2886 * we temporarily boost lower priority queues if they are holding fs exclusive
2887 * resources. they are boosted to normal prio (CLASS_BE/4)
2889 static void cfq_prio_boost(struct cfq_queue *cfqq)
2891 if (has_fs_excl()) {
2893 * boost idle prio on transactions that would lock out other
2894 * users of the filesystem
2896 if (cfq_class_idle(cfqq))
2897 cfqq->ioprio_class = IOPRIO_CLASS_BE;
2898 if (cfqq->ioprio > IOPRIO_NORM)
2899 cfqq->ioprio = IOPRIO_NORM;
2902 * unboost the queue (if needed)
2904 cfqq->ioprio_class = cfqq->org_ioprio_class;
2905 cfqq->ioprio = cfqq->org_ioprio;
2909 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
2911 if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
2912 cfq_mark_cfqq_must_alloc_slice(cfqq);
2913 return ELV_MQUEUE_MUST;
2916 return ELV_MQUEUE_MAY;
2919 static int cfq_may_queue(struct request_queue *q, int rw)
2921 struct cfq_data *cfqd = q->elevator->elevator_data;
2922 struct task_struct *tsk = current;
2923 struct cfq_io_context *cic;
2924 struct cfq_queue *cfqq;
2927 * don't force setup of a queue from here, as a call to may_queue
2928 * does not necessarily imply that a request actually will be queued.
2929 * so just lookup a possibly existing queue, or return 'may queue'
2932 cic = cfq_cic_lookup(cfqd, tsk->io_context);
2934 return ELV_MQUEUE_MAY;
2936 cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
2938 cfq_init_prio_data(cfqq, cic->ioc);
2939 cfq_prio_boost(cfqq);
2941 return __cfq_may_queue(cfqq);
2944 return ELV_MQUEUE_MAY;
2948 * queue lock held here
2950 static void cfq_put_request(struct request *rq)
2952 struct cfq_queue *cfqq = RQ_CFQQ(rq);
2955 const int rw = rq_data_dir(rq);
2957 BUG_ON(!cfqq->allocated[rw]);
2958 cfqq->allocated[rw]--;
2960 put_io_context(RQ_CIC(rq)->ioc);
2962 rq->elevator_private = NULL;
2963 rq->elevator_private2 = NULL;
2965 cfq_put_queue(cfqq);
2969 static struct cfq_queue *
2970 cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_context *cic,
2971 struct cfq_queue *cfqq)
2973 cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq);
2974 cic_set_cfqq(cic, cfqq->new_cfqq, 1);
2975 cfq_mark_cfqq_coop(cfqq->new_cfqq);
2976 cfq_put_queue(cfqq);
2977 return cic_to_cfqq(cic, 1);
2980 static int should_split_cfqq(struct cfq_queue *cfqq)
2982 if (cfqq->seeky_start &&
2983 time_after(jiffies, cfqq->seeky_start + CFQQ_COOP_TOUT))
2989 * Returns NULL if a new cfqq should be allocated, or the old cfqq if this
2990 * was the last process referring to said cfqq.
2992 static struct cfq_queue *
2993 split_cfqq(struct cfq_io_context *cic, struct cfq_queue *cfqq)
2995 if (cfqq_process_refs(cfqq) == 1) {
2996 cfqq->seeky_start = 0;
2997 cfqq->pid = current->pid;
2998 cfq_clear_cfqq_coop(cfqq);
3002 cic_set_cfqq(cic, NULL, 1);
3003 cfq_put_queue(cfqq);
3007 * Allocate cfq data structures associated with this request.
3010 cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
3012 struct cfq_data *cfqd = q->elevator->elevator_data;
3013 struct cfq_io_context *cic;
3014 const int rw = rq_data_dir(rq);
3015 const bool is_sync = rq_is_sync(rq);
3016 struct cfq_queue *cfqq;
3017 unsigned long flags;
3019 might_sleep_if(gfp_mask & __GFP_WAIT);
3021 cic = cfq_get_io_context(cfqd, gfp_mask);
3023 spin_lock_irqsave(q->queue_lock, flags);
3029 cfqq = cic_to_cfqq(cic, is_sync);
3030 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
3031 cfqq = cfq_get_queue(cfqd, is_sync, cic->ioc, gfp_mask);
3032 cic_set_cfqq(cic, cfqq, is_sync);
3035 * If the queue was seeky for too long, break it apart.
3037 if (cfq_cfqq_coop(cfqq) && should_split_cfqq(cfqq)) {
3038 cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq");
3039 cfqq = split_cfqq(cic, cfqq);
3045 * Check to see if this queue is scheduled to merge with
3046 * another, closely cooperating queue. The merging of
3047 * queues happens here as it must be done in process context.
3048 * The reference on new_cfqq was taken in merge_cfqqs.
3051 cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq);
3054 cfqq->allocated[rw]++;
3055 atomic_inc(&cfqq->ref);
3057 spin_unlock_irqrestore(q->queue_lock, flags);
3059 rq->elevator_private = cic;
3060 rq->elevator_private2 = cfqq;
3065 put_io_context(cic->ioc);
3067 cfq_schedule_dispatch(cfqd);
3068 spin_unlock_irqrestore(q->queue_lock, flags);
3069 cfq_log(cfqd, "set_request fail");
3073 static void cfq_kick_queue(struct work_struct *work)
3075 struct cfq_data *cfqd =
3076 container_of(work, struct cfq_data, unplug_work);
3077 struct request_queue *q = cfqd->queue;
3079 spin_lock_irq(q->queue_lock);
3080 __blk_run_queue(cfqd->queue);
3081 spin_unlock_irq(q->queue_lock);
3085 * Timer running if the active_queue is currently idling inside its time slice
3087 static void cfq_idle_slice_timer(unsigned long data)
3089 struct cfq_data *cfqd = (struct cfq_data *) data;
3090 struct cfq_queue *cfqq;
3091 unsigned long flags;
3094 cfq_log(cfqd, "idle timer fired");
3096 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
3098 cfqq = cfqd->active_queue;
3103 * We saw a request before the queue expired, let it through
3105 if (cfq_cfqq_must_dispatch(cfqq))
3111 if (cfq_slice_used(cfqq))
3115 * only expire and reinvoke request handler, if there are
3116 * other queues with pending requests
3118 if (!cfqd->busy_queues)
3122 * not expired and it has a request pending, let it dispatch
3124 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3128 * Queue depth flag is reset only when the idle didn't succeed
3130 cfq_clear_cfqq_deep(cfqq);
3133 cfq_slice_expired(cfqd, timed_out);
3135 cfq_schedule_dispatch(cfqd);
3137 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
3140 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
3142 del_timer_sync(&cfqd->idle_slice_timer);
3143 cancel_work_sync(&cfqd->unplug_work);
3146 static void cfq_put_async_queues(struct cfq_data *cfqd)
3150 for (i = 0; i < IOPRIO_BE_NR; i++) {
3151 if (cfqd->async_cfqq[0][i])
3152 cfq_put_queue(cfqd->async_cfqq[0][i]);
3153 if (cfqd->async_cfqq[1][i])
3154 cfq_put_queue(cfqd->async_cfqq[1][i]);
3157 if (cfqd->async_idle_cfqq)
3158 cfq_put_queue(cfqd->async_idle_cfqq);
3161 static void cfq_exit_queue(struct elevator_queue *e)
3163 struct cfq_data *cfqd = e->elevator_data;
3164 struct request_queue *q = cfqd->queue;
3166 cfq_shutdown_timer_wq(cfqd);
3168 spin_lock_irq(q->queue_lock);
3170 if (cfqd->active_queue)
3171 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
3173 while (!list_empty(&cfqd->cic_list)) {
3174 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
3175 struct cfq_io_context,
3178 __cfq_exit_single_io_context(cfqd, cic);
3181 cfq_put_async_queues(cfqd);
3183 spin_unlock_irq(q->queue_lock);
3185 cfq_shutdown_timer_wq(cfqd);
3190 static void *cfq_init_queue(struct request_queue *q)
3192 struct cfq_data *cfqd;
3194 struct cfq_group *cfqg;
3195 struct cfq_rb_root *st;
3197 cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
3201 /* Init root service tree */
3202 cfqd->grp_service_tree = CFQ_RB_ROOT;
3204 /* Init root group */
3205 cfqg = &cfqd->root_group;
3206 for_each_cfqg_st(cfqg, i, j, st)
3208 RB_CLEAR_NODE(&cfqg->rb_node);
3210 /* Give preference to root group over other groups */
3211 cfqg->weight = 2*BLKIO_WEIGHT_DEFAULT;
3214 * Not strictly needed (since RB_ROOT just clears the node and we
3215 * zeroed cfqd on alloc), but better be safe in case someone decides
3216 * to add magic to the rb code
3218 for (i = 0; i < CFQ_PRIO_LISTS; i++)
3219 cfqd->prio_trees[i] = RB_ROOT;
3222 * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
3223 * Grab a permanent reference to it, so that the normal code flow
3224 * will not attempt to free it.
3226 cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
3227 atomic_inc(&cfqd->oom_cfqq.ref);
3228 cfq_link_cfqq_cfqg(&cfqd->oom_cfqq, &cfqd->root_group);
3230 INIT_LIST_HEAD(&cfqd->cic_list);
3234 init_timer(&cfqd->idle_slice_timer);
3235 cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
3236 cfqd->idle_slice_timer.data = (unsigned long) cfqd;
3238 INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
3240 cfqd->cfq_quantum = cfq_quantum;
3241 cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
3242 cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
3243 cfqd->cfq_back_max = cfq_back_max;
3244 cfqd->cfq_back_penalty = cfq_back_penalty;
3245 cfqd->cfq_slice[0] = cfq_slice_async;
3246 cfqd->cfq_slice[1] = cfq_slice_sync;
3247 cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
3248 cfqd->cfq_slice_idle = cfq_slice_idle;
3249 cfqd->cfq_latency = 1;
3251 cfqd->last_end_sync_rq = jiffies;
3255 static void cfq_slab_kill(void)
3258 * Caller already ensured that pending RCU callbacks are completed,
3259 * so we should have no busy allocations at this point.
3262 kmem_cache_destroy(cfq_pool);
3264 kmem_cache_destroy(cfq_ioc_pool);
3267 static int __init cfq_slab_setup(void)
3269 cfq_pool = KMEM_CACHE(cfq_queue, 0);
3273 cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0);
3284 * sysfs parts below -->
3287 cfq_var_show(unsigned int var, char *page)
3289 return sprintf(page, "%d\n", var);
3293 cfq_var_store(unsigned int *var, const char *page, size_t count)
3295 char *p = (char *) page;
3297 *var = simple_strtoul(p, &p, 10);
3301 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
3302 static ssize_t __FUNC(struct elevator_queue *e, char *page) \
3304 struct cfq_data *cfqd = e->elevator_data; \
3305 unsigned int __data = __VAR; \
3307 __data = jiffies_to_msecs(__data); \
3308 return cfq_var_show(__data, (page)); \
3310 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
3311 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
3312 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
3313 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
3314 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
3315 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
3316 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
3317 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
3318 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
3319 SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
3320 #undef SHOW_FUNCTION
3322 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
3323 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
3325 struct cfq_data *cfqd = e->elevator_data; \
3326 unsigned int __data; \
3327 int ret = cfq_var_store(&__data, (page), count); \
3328 if (__data < (MIN)) \
3330 else if (__data > (MAX)) \
3333 *(__PTR) = msecs_to_jiffies(__data); \
3335 *(__PTR) = __data; \
3338 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
3339 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
3341 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
3343 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
3344 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
3346 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
3347 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
3348 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
3349 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
3351 STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
3352 #undef STORE_FUNCTION
3354 #define CFQ_ATTR(name) \
3355 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
3357 static struct elv_fs_entry cfq_attrs[] = {
3359 CFQ_ATTR(fifo_expire_sync),
3360 CFQ_ATTR(fifo_expire_async),
3361 CFQ_ATTR(back_seek_max),
3362 CFQ_ATTR(back_seek_penalty),
3363 CFQ_ATTR(slice_sync),
3364 CFQ_ATTR(slice_async),
3365 CFQ_ATTR(slice_async_rq),
3366 CFQ_ATTR(slice_idle),
3367 CFQ_ATTR(low_latency),
3371 static struct elevator_type iosched_cfq = {
3373 .elevator_merge_fn = cfq_merge,
3374 .elevator_merged_fn = cfq_merged_request,
3375 .elevator_merge_req_fn = cfq_merged_requests,
3376 .elevator_allow_merge_fn = cfq_allow_merge,
3377 .elevator_dispatch_fn = cfq_dispatch_requests,
3378 .elevator_add_req_fn = cfq_insert_request,
3379 .elevator_activate_req_fn = cfq_activate_request,
3380 .elevator_deactivate_req_fn = cfq_deactivate_request,
3381 .elevator_queue_empty_fn = cfq_queue_empty,
3382 .elevator_completed_req_fn = cfq_completed_request,
3383 .elevator_former_req_fn = elv_rb_former_request,
3384 .elevator_latter_req_fn = elv_rb_latter_request,
3385 .elevator_set_req_fn = cfq_set_request,
3386 .elevator_put_req_fn = cfq_put_request,
3387 .elevator_may_queue_fn = cfq_may_queue,
3388 .elevator_init_fn = cfq_init_queue,
3389 .elevator_exit_fn = cfq_exit_queue,
3390 .trim = cfq_free_io_context,
3392 .elevator_attrs = cfq_attrs,
3393 .elevator_name = "cfq",
3394 .elevator_owner = THIS_MODULE,
3397 static int __init cfq_init(void)
3400 * could be 0 on HZ < 1000 setups
3402 if (!cfq_slice_async)
3403 cfq_slice_async = 1;
3404 if (!cfq_slice_idle)
3407 if (cfq_slab_setup())
3410 elv_register(&iosched_cfq);
3415 static void __exit cfq_exit(void)
3417 DECLARE_COMPLETION_ONSTACK(all_gone);
3418 elv_unregister(&iosched_cfq);
3419 ioc_gone = &all_gone;
3420 /* ioc_gone's update must be visible before reading ioc_count */
3424 * this also protects us from entering cfq_slab_kill() with
3425 * pending RCU callbacks
3427 if (elv_ioc_count_read(cfq_ioc_count))
3428 wait_for_completion(&all_gone);
3432 module_init(cfq_init);
3433 module_exit(cfq_exit);
3435 MODULE_AUTHOR("Jens Axboe");
3436 MODULE_LICENSE("GPL");
3437 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");