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)
45 #define CFQ_SLICE_SCALE (5)
46 #define CFQ_HW_QUEUE_MIN (5)
47 #define CFQ_SERVICE_SHIFT 12
49 #define CFQQ_SEEK_THR (sector_t)(8 * 100)
50 #define CFQQ_SEEKY(cfqq) (hweight32(cfqq->seek_history) > 32/8)
53 ((struct cfq_io_context *) (rq)->elevator_private)
54 #define RQ_CFQQ(rq) (struct cfq_queue *) ((rq)->elevator_private2)
56 static struct kmem_cache *cfq_pool;
57 static struct kmem_cache *cfq_ioc_pool;
59 static DEFINE_PER_CPU(unsigned long, cfq_ioc_count);
60 static struct completion *ioc_gone;
61 static DEFINE_SPINLOCK(ioc_gone_lock);
63 #define CFQ_PRIO_LISTS IOPRIO_BE_NR
64 #define cfq_class_idle(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
65 #define cfq_class_rt(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
67 #define sample_valid(samples) ((samples) > 80)
68 #define rb_entry_cfqg(node) rb_entry((node), struct cfq_group, rb_node)
71 * Most of our rbtree usage is for sorting with min extraction, so
72 * if we cache the leftmost node we don't have to walk down the tree
73 * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should
74 * move this into the elevator for the rq sorting as well.
81 struct rb_node *active;
82 unsigned total_weight;
84 #define CFQ_RB_ROOT (struct cfq_rb_root) { RB_ROOT, NULL, 0, 0, }
87 * Per process-grouping structure
92 /* various state flags, see below */
95 struct cfq_data *cfqd;
96 /* service_tree member */
97 struct rb_node rb_node;
98 /* service_tree key */
100 /* prio tree member */
101 struct rb_node p_node;
102 /* prio tree root we belong to, if any */
103 struct rb_root *p_root;
104 /* sorted list of pending requests */
105 struct rb_root sort_list;
106 /* if fifo isn't expired, next request to serve */
107 struct request *next_rq;
108 /* requests queued in sort_list */
110 /* currently allocated requests */
112 /* fifo list of requests in sort_list */
113 struct list_head fifo;
115 /* time when queue got scheduled in to dispatch first request. */
116 unsigned long dispatch_start;
117 unsigned int allocated_slice;
118 unsigned int slice_dispatch;
119 /* time when first request from queue completed and slice started. */
120 unsigned long slice_start;
121 unsigned long slice_end;
124 /* pending metadata requests */
126 /* number of requests that are on the dispatch list or inside driver */
129 /* io prio of this group */
130 unsigned short ioprio, org_ioprio;
131 unsigned short ioprio_class, org_ioprio_class;
136 sector_t last_request_pos;
138 struct cfq_rb_root *service_tree;
139 struct cfq_queue *new_cfqq;
140 struct cfq_group *cfqg;
141 struct cfq_group *orig_cfqg;
142 /* Sectors dispatched in current dispatch round */
143 unsigned long nr_sectors;
147 * First index in the service_trees.
148 * IDLE is handled separately, so it has negative index
157 * Second index in the service_trees.
161 SYNC_NOIDLE_WORKLOAD = 1,
165 /* This is per cgroup per device grouping structure */
167 /* group service_tree member */
168 struct rb_node rb_node;
170 /* group service_tree key */
175 /* number of cfqq currently on this group */
178 /* Per group busy queus average. Useful for workload slice calc. */
179 unsigned int busy_queues_avg[2];
181 * rr lists of queues with requests, onle rr for each priority class.
182 * Counts are embedded in the cfq_rb_root
184 struct cfq_rb_root service_trees[2][3];
185 struct cfq_rb_root service_tree_idle;
187 unsigned long saved_workload_slice;
188 enum wl_type_t saved_workload;
189 enum wl_prio_t saved_serving_prio;
190 struct blkio_group blkg;
191 #ifdef CONFIG_CFQ_GROUP_IOSCHED
192 struct hlist_node cfqd_node;
198 * Per block device queue structure
201 struct request_queue *queue;
202 /* Root service tree for cfq_groups */
203 struct cfq_rb_root grp_service_tree;
204 struct cfq_group root_group;
207 * The priority currently being served
209 enum wl_prio_t serving_prio;
210 enum wl_type_t serving_type;
211 unsigned long workload_expires;
212 struct cfq_group *serving_group;
213 bool noidle_tree_requires_idle;
216 * Each priority tree is sorted by next_request position. These
217 * trees are used when determining if two or more queues are
218 * interleaving requests (see cfq_close_cooperator).
220 struct rb_root prio_trees[CFQ_PRIO_LISTS];
222 unsigned int busy_queues;
228 * queue-depth detection
234 * -1 => indeterminate, (cfq will behave as if NCQ is present, to allow better detection)
235 * 1 => NCQ is present (hw_tag_est_depth is the estimated max depth)
238 int hw_tag_est_depth;
239 unsigned int hw_tag_samples;
242 * idle window management
244 struct timer_list idle_slice_timer;
245 struct work_struct unplug_work;
247 struct cfq_queue *active_queue;
248 struct cfq_io_context *active_cic;
251 * async queue for each priority case
253 struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
254 struct cfq_queue *async_idle_cfqq;
256 sector_t last_position;
259 * tunables, see top of file
261 unsigned int cfq_quantum;
262 unsigned int cfq_fifo_expire[2];
263 unsigned int cfq_back_penalty;
264 unsigned int cfq_back_max;
265 unsigned int cfq_slice[2];
266 unsigned int cfq_slice_async_rq;
267 unsigned int cfq_slice_idle;
268 unsigned int cfq_latency;
269 unsigned int cfq_group_isolation;
271 struct list_head cic_list;
274 * Fallback dummy cfqq for extreme OOM conditions
276 struct cfq_queue oom_cfqq;
278 unsigned long last_delayed_sync;
280 /* List of cfq groups being managed on this device*/
281 struct hlist_head cfqg_list;
285 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd);
287 static struct cfq_rb_root *service_tree_for(struct cfq_group *cfqg,
294 if (prio == IDLE_WORKLOAD)
295 return &cfqg->service_tree_idle;
297 return &cfqg->service_trees[prio][type];
300 enum cfqq_state_flags {
301 CFQ_CFQQ_FLAG_on_rr = 0, /* on round-robin busy list */
302 CFQ_CFQQ_FLAG_wait_request, /* waiting for a request */
303 CFQ_CFQQ_FLAG_must_dispatch, /* must be allowed a dispatch */
304 CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
305 CFQ_CFQQ_FLAG_fifo_expire, /* FIFO checked in this slice */
306 CFQ_CFQQ_FLAG_idle_window, /* slice idling enabled */
307 CFQ_CFQQ_FLAG_prio_changed, /* task priority has changed */
308 CFQ_CFQQ_FLAG_slice_new, /* no requests dispatched in slice */
309 CFQ_CFQQ_FLAG_sync, /* synchronous queue */
310 CFQ_CFQQ_FLAG_coop, /* cfqq is shared */
311 CFQ_CFQQ_FLAG_split_coop, /* shared cfqq will be splitted */
312 CFQ_CFQQ_FLAG_deep, /* sync cfqq experienced large depth */
313 CFQ_CFQQ_FLAG_wait_busy, /* Waiting for next request */
316 #define CFQ_CFQQ_FNS(name) \
317 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \
319 (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name); \
321 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \
323 (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \
325 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \
327 return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \
331 CFQ_CFQQ_FNS(wait_request);
332 CFQ_CFQQ_FNS(must_dispatch);
333 CFQ_CFQQ_FNS(must_alloc_slice);
334 CFQ_CFQQ_FNS(fifo_expire);
335 CFQ_CFQQ_FNS(idle_window);
336 CFQ_CFQQ_FNS(prio_changed);
337 CFQ_CFQQ_FNS(slice_new);
340 CFQ_CFQQ_FNS(split_coop);
342 CFQ_CFQQ_FNS(wait_busy);
345 #ifdef CONFIG_DEBUG_CFQ_IOSCHED
346 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
347 blk_add_trace_msg((cfqd)->queue, "cfq%d%c %s " fmt, (cfqq)->pid, \
348 cfq_cfqq_sync((cfqq)) ? 'S' : 'A', \
349 blkg_path(&(cfqq)->cfqg->blkg), ##args);
351 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...) \
352 blk_add_trace_msg((cfqd)->queue, "%s " fmt, \
353 blkg_path(&(cfqg)->blkg), ##args); \
356 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
357 blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args)
358 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...) do {} while (0);
360 #define cfq_log(cfqd, fmt, args...) \
361 blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
363 /* Traverses through cfq group service trees */
364 #define for_each_cfqg_st(cfqg, i, j, st) \
365 for (i = 0; i <= IDLE_WORKLOAD; i++) \
366 for (j = 0, st = i < IDLE_WORKLOAD ? &cfqg->service_trees[i][j]\
367 : &cfqg->service_tree_idle; \
368 (i < IDLE_WORKLOAD && j <= SYNC_WORKLOAD) || \
369 (i == IDLE_WORKLOAD && j == 0); \
370 j++, st = i < IDLE_WORKLOAD ? \
371 &cfqg->service_trees[i][j]: NULL) \
374 static inline enum wl_prio_t cfqq_prio(struct cfq_queue *cfqq)
376 if (cfq_class_idle(cfqq))
377 return IDLE_WORKLOAD;
378 if (cfq_class_rt(cfqq))
384 static enum wl_type_t cfqq_type(struct cfq_queue *cfqq)
386 if (!cfq_cfqq_sync(cfqq))
387 return ASYNC_WORKLOAD;
388 if (!cfq_cfqq_idle_window(cfqq))
389 return SYNC_NOIDLE_WORKLOAD;
390 return SYNC_WORKLOAD;
393 static inline int cfq_group_busy_queues_wl(enum wl_prio_t wl,
394 struct cfq_data *cfqd,
395 struct cfq_group *cfqg)
397 if (wl == IDLE_WORKLOAD)
398 return cfqg->service_tree_idle.count;
400 return cfqg->service_trees[wl][ASYNC_WORKLOAD].count
401 + cfqg->service_trees[wl][SYNC_NOIDLE_WORKLOAD].count
402 + cfqg->service_trees[wl][SYNC_WORKLOAD].count;
405 static inline int cfqg_busy_async_queues(struct cfq_data *cfqd,
406 struct cfq_group *cfqg)
408 return cfqg->service_trees[RT_WORKLOAD][ASYNC_WORKLOAD].count
409 + cfqg->service_trees[BE_WORKLOAD][ASYNC_WORKLOAD].count;
412 static void cfq_dispatch_insert(struct request_queue *, struct request *);
413 static struct cfq_queue *cfq_get_queue(struct cfq_data *, bool,
414 struct io_context *, gfp_t);
415 static struct cfq_io_context *cfq_cic_lookup(struct cfq_data *,
416 struct io_context *);
418 static inline int rq_in_driver(struct cfq_data *cfqd)
420 return cfqd->rq_in_driver[0] + cfqd->rq_in_driver[1];
423 static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_context *cic,
426 return cic->cfqq[is_sync];
429 static inline void cic_set_cfqq(struct cfq_io_context *cic,
430 struct cfq_queue *cfqq, bool is_sync)
432 cic->cfqq[is_sync] = cfqq;
436 * We regard a request as SYNC, if it's either a read or has the SYNC bit
437 * set (in which case it could also be direct WRITE).
439 static inline bool cfq_bio_sync(struct bio *bio)
441 return bio_data_dir(bio) == READ || bio_rw_flagged(bio, BIO_RW_SYNCIO);
445 * scheduler run of queue, if there are requests pending and no one in the
446 * driver that will restart queueing
448 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
450 if (cfqd->busy_queues) {
451 cfq_log(cfqd, "schedule dispatch");
452 kblockd_schedule_work(cfqd->queue, &cfqd->unplug_work);
456 static int cfq_queue_empty(struct request_queue *q)
458 struct cfq_data *cfqd = q->elevator->elevator_data;
460 return !cfqd->rq_queued;
464 * Scale schedule slice based on io priority. Use the sync time slice only
465 * if a queue is marked sync and has sync io queued. A sync queue with async
466 * io only, should not get full sync slice length.
468 static inline int cfq_prio_slice(struct cfq_data *cfqd, bool sync,
471 const int base_slice = cfqd->cfq_slice[sync];
473 WARN_ON(prio >= IOPRIO_BE_NR);
475 return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
479 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
481 return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
484 static inline u64 cfq_scale_slice(unsigned long delta, struct cfq_group *cfqg)
486 u64 d = delta << CFQ_SERVICE_SHIFT;
488 d = d * BLKIO_WEIGHT_DEFAULT;
489 do_div(d, cfqg->weight);
493 static inline u64 max_vdisktime(u64 min_vdisktime, u64 vdisktime)
495 s64 delta = (s64)(vdisktime - min_vdisktime);
497 min_vdisktime = vdisktime;
499 return min_vdisktime;
502 static inline u64 min_vdisktime(u64 min_vdisktime, u64 vdisktime)
504 s64 delta = (s64)(vdisktime - min_vdisktime);
506 min_vdisktime = vdisktime;
508 return min_vdisktime;
511 static void update_min_vdisktime(struct cfq_rb_root *st)
513 u64 vdisktime = st->min_vdisktime;
514 struct cfq_group *cfqg;
517 cfqg = rb_entry_cfqg(st->active);
518 vdisktime = cfqg->vdisktime;
522 cfqg = rb_entry_cfqg(st->left);
523 vdisktime = min_vdisktime(vdisktime, cfqg->vdisktime);
526 st->min_vdisktime = max_vdisktime(st->min_vdisktime, vdisktime);
530 * get averaged number of queues of RT/BE priority.
531 * average is updated, with a formula that gives more weight to higher numbers,
532 * to quickly follows sudden increases and decrease slowly
535 static inline unsigned cfq_group_get_avg_queues(struct cfq_data *cfqd,
536 struct cfq_group *cfqg, bool rt)
538 unsigned min_q, max_q;
539 unsigned mult = cfq_hist_divisor - 1;
540 unsigned round = cfq_hist_divisor / 2;
541 unsigned busy = cfq_group_busy_queues_wl(rt, cfqd, cfqg);
543 min_q = min(cfqg->busy_queues_avg[rt], busy);
544 max_q = max(cfqg->busy_queues_avg[rt], busy);
545 cfqg->busy_queues_avg[rt] = (mult * max_q + min_q + round) /
547 return cfqg->busy_queues_avg[rt];
550 static inline unsigned
551 cfq_group_slice(struct cfq_data *cfqd, struct cfq_group *cfqg)
553 struct cfq_rb_root *st = &cfqd->grp_service_tree;
555 return cfq_target_latency * cfqg->weight / st->total_weight;
559 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
561 unsigned slice = cfq_prio_to_slice(cfqd, cfqq);
562 if (cfqd->cfq_latency) {
564 * interested queues (we consider only the ones with the same
565 * priority class in the cfq group)
567 unsigned iq = cfq_group_get_avg_queues(cfqd, cfqq->cfqg,
569 unsigned sync_slice = cfqd->cfq_slice[1];
570 unsigned expect_latency = sync_slice * iq;
571 unsigned group_slice = cfq_group_slice(cfqd, cfqq->cfqg);
573 if (expect_latency > group_slice) {
574 unsigned base_low_slice = 2 * cfqd->cfq_slice_idle;
575 /* scale low_slice according to IO priority
576 * and sync vs async */
578 min(slice, base_low_slice * slice / sync_slice);
579 /* the adapted slice value is scaled to fit all iqs
580 * into the target latency */
581 slice = max(slice * group_slice / expect_latency,
585 cfqq->slice_start = jiffies;
586 cfqq->slice_end = jiffies + slice;
587 cfqq->allocated_slice = slice;
588 cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
592 * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
593 * isn't valid until the first request from the dispatch is activated
594 * and the slice time set.
596 static inline bool cfq_slice_used(struct cfq_queue *cfqq)
598 if (cfq_cfqq_slice_new(cfqq))
600 if (time_before(jiffies, cfqq->slice_end))
607 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
608 * We choose the request that is closest to the head right now. Distance
609 * behind the head is penalized and only allowed to a certain extent.
611 static struct request *
612 cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2, sector_t last)
614 sector_t s1, s2, d1 = 0, d2 = 0;
615 unsigned long back_max;
616 #define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */
617 #define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */
618 unsigned wrap = 0; /* bit mask: requests behind the disk head? */
620 if (rq1 == NULL || rq1 == rq2)
625 if (rq_is_sync(rq1) && !rq_is_sync(rq2))
627 else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
629 if (rq_is_meta(rq1) && !rq_is_meta(rq2))
631 else if (rq_is_meta(rq2) && !rq_is_meta(rq1))
634 s1 = blk_rq_pos(rq1);
635 s2 = blk_rq_pos(rq2);
638 * by definition, 1KiB is 2 sectors
640 back_max = cfqd->cfq_back_max * 2;
643 * Strict one way elevator _except_ in the case where we allow
644 * short backward seeks which are biased as twice the cost of a
645 * similar forward seek.
649 else if (s1 + back_max >= last)
650 d1 = (last - s1) * cfqd->cfq_back_penalty;
652 wrap |= CFQ_RQ1_WRAP;
656 else if (s2 + back_max >= last)
657 d2 = (last - s2) * cfqd->cfq_back_penalty;
659 wrap |= CFQ_RQ2_WRAP;
661 /* Found required data */
664 * By doing switch() on the bit mask "wrap" we avoid having to
665 * check two variables for all permutations: --> faster!
668 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
684 case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
687 * Since both rqs are wrapped,
688 * start with the one that's further behind head
689 * (--> only *one* back seek required),
690 * since back seek takes more time than forward.
700 * The below is leftmost cache rbtree addon
702 static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
704 /* Service tree is empty */
709 root->left = rb_first(&root->rb);
712 return rb_entry(root->left, struct cfq_queue, rb_node);
717 static struct cfq_group *cfq_rb_first_group(struct cfq_rb_root *root)
720 root->left = rb_first(&root->rb);
723 return rb_entry_cfqg(root->left);
728 static void rb_erase_init(struct rb_node *n, struct rb_root *root)
734 static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
738 rb_erase_init(n, &root->rb);
743 * would be nice to take fifo expire time into account as well
745 static struct request *
746 cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
747 struct request *last)
749 struct rb_node *rbnext = rb_next(&last->rb_node);
750 struct rb_node *rbprev = rb_prev(&last->rb_node);
751 struct request *next = NULL, *prev = NULL;
753 BUG_ON(RB_EMPTY_NODE(&last->rb_node));
756 prev = rb_entry_rq(rbprev);
759 next = rb_entry_rq(rbnext);
761 rbnext = rb_first(&cfqq->sort_list);
762 if (rbnext && rbnext != &last->rb_node)
763 next = rb_entry_rq(rbnext);
766 return cfq_choose_req(cfqd, next, prev, blk_rq_pos(last));
769 static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
770 struct cfq_queue *cfqq)
773 * just an approximation, should be ok.
775 return (cfqq->cfqg->nr_cfqq - 1) * (cfq_prio_slice(cfqd, 1, 0) -
776 cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
780 cfqg_key(struct cfq_rb_root *st, struct cfq_group *cfqg)
782 return cfqg->vdisktime - st->min_vdisktime;
786 __cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
788 struct rb_node **node = &st->rb.rb_node;
789 struct rb_node *parent = NULL;
790 struct cfq_group *__cfqg;
791 s64 key = cfqg_key(st, cfqg);
794 while (*node != NULL) {
796 __cfqg = rb_entry_cfqg(parent);
798 if (key < cfqg_key(st, __cfqg))
799 node = &parent->rb_left;
801 node = &parent->rb_right;
807 st->left = &cfqg->rb_node;
809 rb_link_node(&cfqg->rb_node, parent, node);
810 rb_insert_color(&cfqg->rb_node, &st->rb);
814 cfq_group_service_tree_add(struct cfq_data *cfqd, struct cfq_group *cfqg)
816 struct cfq_rb_root *st = &cfqd->grp_service_tree;
817 struct cfq_group *__cfqg;
825 * Currently put the group at the end. Later implement something
826 * so that groups get lesser vtime based on their weights, so that
827 * if group does not loose all if it was not continously backlogged.
829 n = rb_last(&st->rb);
831 __cfqg = rb_entry_cfqg(n);
832 cfqg->vdisktime = __cfqg->vdisktime + CFQ_IDLE_DELAY;
834 cfqg->vdisktime = st->min_vdisktime;
836 __cfq_group_service_tree_add(st, cfqg);
838 st->total_weight += cfqg->weight;
842 cfq_group_service_tree_del(struct cfq_data *cfqd, struct cfq_group *cfqg)
844 struct cfq_rb_root *st = &cfqd->grp_service_tree;
846 if (st->active == &cfqg->rb_node)
849 BUG_ON(cfqg->nr_cfqq < 1);
852 /* If there are other cfq queues under this group, don't delete it */
856 cfq_log_cfqg(cfqd, cfqg, "del_from_rr group");
858 st->total_weight -= cfqg->weight;
859 if (!RB_EMPTY_NODE(&cfqg->rb_node))
860 cfq_rb_erase(&cfqg->rb_node, st);
861 cfqg->saved_workload_slice = 0;
862 blkiocg_update_blkio_group_dequeue_stats(&cfqg->blkg, 1);
865 static inline unsigned int cfq_cfqq_slice_usage(struct cfq_queue *cfqq)
867 unsigned int slice_used;
870 * Queue got expired before even a single request completed or
871 * got expired immediately after first request completion.
873 if (!cfqq->slice_start || cfqq->slice_start == jiffies) {
875 * Also charge the seek time incurred to the group, otherwise
876 * if there are mutiple queues in the group, each can dispatch
877 * a single request on seeky media and cause lots of seek time
878 * and group will never know it.
880 slice_used = max_t(unsigned, (jiffies - cfqq->dispatch_start),
883 slice_used = jiffies - cfqq->slice_start;
884 if (slice_used > cfqq->allocated_slice)
885 slice_used = cfqq->allocated_slice;
888 cfq_log_cfqq(cfqq->cfqd, cfqq, "sl_used=%u sect=%lu", slice_used,
893 static void cfq_group_served(struct cfq_data *cfqd, struct cfq_group *cfqg,
894 struct cfq_queue *cfqq)
896 struct cfq_rb_root *st = &cfqd->grp_service_tree;
897 unsigned int used_sl, charge_sl;
898 int nr_sync = cfqg->nr_cfqq - cfqg_busy_async_queues(cfqd, cfqg)
899 - cfqg->service_tree_idle.count;
902 used_sl = charge_sl = cfq_cfqq_slice_usage(cfqq);
904 if (!cfq_cfqq_sync(cfqq) && !nr_sync)
905 charge_sl = cfqq->allocated_slice;
907 /* Can't update vdisktime while group is on service tree */
908 cfq_rb_erase(&cfqg->rb_node, st);
909 cfqg->vdisktime += cfq_scale_slice(charge_sl, cfqg);
910 __cfq_group_service_tree_add(st, cfqg);
912 /* This group is being expired. Save the context */
913 if (time_after(cfqd->workload_expires, jiffies)) {
914 cfqg->saved_workload_slice = cfqd->workload_expires
916 cfqg->saved_workload = cfqd->serving_type;
917 cfqg->saved_serving_prio = cfqd->serving_prio;
919 cfqg->saved_workload_slice = 0;
921 cfq_log_cfqg(cfqd, cfqg, "served: vt=%llu min_vt=%llu", cfqg->vdisktime,
923 blkiocg_update_blkio_group_stats(&cfqg->blkg, used_sl,
927 #ifdef CONFIG_CFQ_GROUP_IOSCHED
928 static inline struct cfq_group *cfqg_of_blkg(struct blkio_group *blkg)
931 return container_of(blkg, struct cfq_group, blkg);
936 cfq_update_blkio_group_weight(struct blkio_group *blkg, unsigned int weight)
938 cfqg_of_blkg(blkg)->weight = weight;
941 static struct cfq_group *
942 cfq_find_alloc_cfqg(struct cfq_data *cfqd, struct cgroup *cgroup, int create)
944 struct blkio_cgroup *blkcg = cgroup_to_blkio_cgroup(cgroup);
945 struct cfq_group *cfqg = NULL;
948 struct cfq_rb_root *st;
949 struct backing_dev_info *bdi = &cfqd->queue->backing_dev_info;
950 unsigned int major, minor;
952 cfqg = cfqg_of_blkg(blkiocg_lookup_group(blkcg, key));
956 cfqg = kzalloc_node(sizeof(*cfqg), GFP_ATOMIC, cfqd->queue->node);
960 cfqg->weight = blkcg->weight;
961 for_each_cfqg_st(cfqg, i, j, st)
963 RB_CLEAR_NODE(&cfqg->rb_node);
966 * Take the initial reference that will be released on destroy
967 * This can be thought of a joint reference by cgroup and
968 * elevator which will be dropped by either elevator exit
969 * or cgroup deletion path depending on who is exiting first.
971 atomic_set(&cfqg->ref, 1);
973 /* Add group onto cgroup list */
974 sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
975 blkiocg_add_blkio_group(blkcg, &cfqg->blkg, (void *)cfqd,
976 MKDEV(major, minor));
978 /* Add group on cfqd list */
979 hlist_add_head(&cfqg->cfqd_node, &cfqd->cfqg_list);
986 * Search for the cfq group current task belongs to. If create = 1, then also
987 * create the cfq group if it does not exist. request_queue lock must be held.
989 static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd, int create)
991 struct cgroup *cgroup;
992 struct cfq_group *cfqg = NULL;
995 cgroup = task_cgroup(current, blkio_subsys_id);
996 cfqg = cfq_find_alloc_cfqg(cfqd, cgroup, create);
998 cfqg = &cfqd->root_group;
1003 static void cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg)
1005 /* Currently, all async queues are mapped to root group */
1006 if (!cfq_cfqq_sync(cfqq))
1007 cfqg = &cfqq->cfqd->root_group;
1010 /* cfqq reference on cfqg */
1011 atomic_inc(&cfqq->cfqg->ref);
1014 static void cfq_put_cfqg(struct cfq_group *cfqg)
1016 struct cfq_rb_root *st;
1019 BUG_ON(atomic_read(&cfqg->ref) <= 0);
1020 if (!atomic_dec_and_test(&cfqg->ref))
1022 for_each_cfqg_st(cfqg, i, j, st)
1023 BUG_ON(!RB_EMPTY_ROOT(&st->rb) || st->active != NULL);
1027 static void cfq_destroy_cfqg(struct cfq_data *cfqd, struct cfq_group *cfqg)
1029 /* Something wrong if we are trying to remove same group twice */
1030 BUG_ON(hlist_unhashed(&cfqg->cfqd_node));
1032 hlist_del_init(&cfqg->cfqd_node);
1035 * Put the reference taken at the time of creation so that when all
1036 * queues are gone, group can be destroyed.
1041 static void cfq_release_cfq_groups(struct cfq_data *cfqd)
1043 struct hlist_node *pos, *n;
1044 struct cfq_group *cfqg;
1046 hlist_for_each_entry_safe(cfqg, pos, n, &cfqd->cfqg_list, cfqd_node) {
1048 * If cgroup removal path got to blk_group first and removed
1049 * it from cgroup list, then it will take care of destroying
1052 if (!blkiocg_del_blkio_group(&cfqg->blkg))
1053 cfq_destroy_cfqg(cfqd, cfqg);
1058 * Blk cgroup controller notification saying that blkio_group object is being
1059 * delinked as associated cgroup object is going away. That also means that
1060 * no new IO will come in this group. So get rid of this group as soon as
1061 * any pending IO in the group is finished.
1063 * This function is called under rcu_read_lock(). key is the rcu protected
1064 * pointer. That means "key" is a valid cfq_data pointer as long as we are rcu
1067 * "key" was fetched from blkio_group under blkio_cgroup->lock. That means
1068 * it should not be NULL as even if elevator was exiting, cgroup deltion
1069 * path got to it first.
1071 void cfq_unlink_blkio_group(void *key, struct blkio_group *blkg)
1073 unsigned long flags;
1074 struct cfq_data *cfqd = key;
1076 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1077 cfq_destroy_cfqg(cfqd, cfqg_of_blkg(blkg));
1078 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1081 #else /* GROUP_IOSCHED */
1082 static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd, int create)
1084 return &cfqd->root_group;
1087 cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) {
1091 static void cfq_release_cfq_groups(struct cfq_data *cfqd) {}
1092 static inline void cfq_put_cfqg(struct cfq_group *cfqg) {}
1094 #endif /* GROUP_IOSCHED */
1097 * The cfqd->service_trees holds all pending cfq_queue's that have
1098 * requests waiting to be processed. It is sorted in the order that
1099 * we will service the queues.
1101 static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1104 struct rb_node **p, *parent;
1105 struct cfq_queue *__cfqq;
1106 unsigned long rb_key;
1107 struct cfq_rb_root *service_tree;
1110 int group_changed = 0;
1112 #ifdef CONFIG_CFQ_GROUP_IOSCHED
1113 if (!cfqd->cfq_group_isolation
1114 && cfqq_type(cfqq) == SYNC_NOIDLE_WORKLOAD
1115 && cfqq->cfqg && cfqq->cfqg != &cfqd->root_group) {
1116 /* Move this cfq to root group */
1117 cfq_log_cfqq(cfqd, cfqq, "moving to root group");
1118 if (!RB_EMPTY_NODE(&cfqq->rb_node))
1119 cfq_group_service_tree_del(cfqd, cfqq->cfqg);
1120 cfqq->orig_cfqg = cfqq->cfqg;
1121 cfqq->cfqg = &cfqd->root_group;
1122 atomic_inc(&cfqd->root_group.ref);
1124 } else if (!cfqd->cfq_group_isolation
1125 && cfqq_type(cfqq) == SYNC_WORKLOAD && cfqq->orig_cfqg) {
1126 /* cfqq is sequential now needs to go to its original group */
1127 BUG_ON(cfqq->cfqg != &cfqd->root_group);
1128 if (!RB_EMPTY_NODE(&cfqq->rb_node))
1129 cfq_group_service_tree_del(cfqd, cfqq->cfqg);
1130 cfq_put_cfqg(cfqq->cfqg);
1131 cfqq->cfqg = cfqq->orig_cfqg;
1132 cfqq->orig_cfqg = NULL;
1134 cfq_log_cfqq(cfqd, cfqq, "moved to origin group");
1138 service_tree = service_tree_for(cfqq->cfqg, cfqq_prio(cfqq),
1140 if (cfq_class_idle(cfqq)) {
1141 rb_key = CFQ_IDLE_DELAY;
1142 parent = rb_last(&service_tree->rb);
1143 if (parent && parent != &cfqq->rb_node) {
1144 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1145 rb_key += __cfqq->rb_key;
1148 } else if (!add_front) {
1150 * Get our rb key offset. Subtract any residual slice
1151 * value carried from last service. A negative resid
1152 * count indicates slice overrun, and this should position
1153 * the next service time further away in the tree.
1155 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
1156 rb_key -= cfqq->slice_resid;
1157 cfqq->slice_resid = 0;
1160 __cfqq = cfq_rb_first(service_tree);
1161 rb_key += __cfqq ? __cfqq->rb_key : jiffies;
1164 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1167 * same position, nothing more to do
1169 if (rb_key == cfqq->rb_key &&
1170 cfqq->service_tree == service_tree)
1173 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1174 cfqq->service_tree = NULL;
1179 cfqq->service_tree = service_tree;
1180 p = &service_tree->rb.rb_node;
1185 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1188 * sort by key, that represents service time.
1190 if (time_before(rb_key, __cfqq->rb_key))
1193 n = &(*p)->rb_right;
1201 service_tree->left = &cfqq->rb_node;
1203 cfqq->rb_key = rb_key;
1204 rb_link_node(&cfqq->rb_node, parent, p);
1205 rb_insert_color(&cfqq->rb_node, &service_tree->rb);
1206 service_tree->count++;
1207 if ((add_front || !new_cfqq) && !group_changed)
1209 cfq_group_service_tree_add(cfqd, cfqq->cfqg);
1212 static struct cfq_queue *
1213 cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
1214 sector_t sector, struct rb_node **ret_parent,
1215 struct rb_node ***rb_link)
1217 struct rb_node **p, *parent;
1218 struct cfq_queue *cfqq = NULL;
1226 cfqq = rb_entry(parent, struct cfq_queue, p_node);
1229 * Sort strictly based on sector. Smallest to the left,
1230 * largest to the right.
1232 if (sector > blk_rq_pos(cfqq->next_rq))
1233 n = &(*p)->rb_right;
1234 else if (sector < blk_rq_pos(cfqq->next_rq))
1242 *ret_parent = parent;
1248 static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1250 struct rb_node **p, *parent;
1251 struct cfq_queue *__cfqq;
1254 rb_erase(&cfqq->p_node, cfqq->p_root);
1255 cfqq->p_root = NULL;
1258 if (cfq_class_idle(cfqq))
1263 cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
1264 __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
1265 blk_rq_pos(cfqq->next_rq), &parent, &p);
1267 rb_link_node(&cfqq->p_node, parent, p);
1268 rb_insert_color(&cfqq->p_node, cfqq->p_root);
1270 cfqq->p_root = NULL;
1274 * Update cfqq's position in the service tree.
1276 static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1279 * Resorting requires the cfqq to be on the RR list already.
1281 if (cfq_cfqq_on_rr(cfqq)) {
1282 cfq_service_tree_add(cfqd, cfqq, 0);
1283 cfq_prio_tree_add(cfqd, cfqq);
1288 * add to busy list of queues for service, trying to be fair in ordering
1289 * the pending list according to last request service
1291 static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1293 cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
1294 BUG_ON(cfq_cfqq_on_rr(cfqq));
1295 cfq_mark_cfqq_on_rr(cfqq);
1296 cfqd->busy_queues++;
1298 cfq_resort_rr_list(cfqd, cfqq);
1302 * Called when the cfqq no longer has requests pending, remove it from
1305 static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1307 cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
1308 BUG_ON(!cfq_cfqq_on_rr(cfqq));
1309 cfq_clear_cfqq_on_rr(cfqq);
1311 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1312 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1313 cfqq->service_tree = NULL;
1316 rb_erase(&cfqq->p_node, cfqq->p_root);
1317 cfqq->p_root = NULL;
1320 cfq_group_service_tree_del(cfqd, cfqq->cfqg);
1321 BUG_ON(!cfqd->busy_queues);
1322 cfqd->busy_queues--;
1326 * rb tree support functions
1328 static void cfq_del_rq_rb(struct request *rq)
1330 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1331 const int sync = rq_is_sync(rq);
1333 BUG_ON(!cfqq->queued[sync]);
1334 cfqq->queued[sync]--;
1336 elv_rb_del(&cfqq->sort_list, rq);
1338 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) {
1340 * Queue will be deleted from service tree when we actually
1341 * expire it later. Right now just remove it from prio tree
1345 rb_erase(&cfqq->p_node, cfqq->p_root);
1346 cfqq->p_root = NULL;
1351 static void cfq_add_rq_rb(struct request *rq)
1353 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1354 struct cfq_data *cfqd = cfqq->cfqd;
1355 struct request *__alias, *prev;
1357 cfqq->queued[rq_is_sync(rq)]++;
1360 * looks a little odd, but the first insert might return an alias.
1361 * if that happens, put the alias on the dispatch list
1363 while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL)
1364 cfq_dispatch_insert(cfqd->queue, __alias);
1366 if (!cfq_cfqq_on_rr(cfqq))
1367 cfq_add_cfqq_rr(cfqd, cfqq);
1370 * check if this request is a better next-serve candidate
1372 prev = cfqq->next_rq;
1373 cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq, cfqd->last_position);
1376 * adjust priority tree position, if ->next_rq changes
1378 if (prev != cfqq->next_rq)
1379 cfq_prio_tree_add(cfqd, cfqq);
1381 BUG_ON(!cfqq->next_rq);
1384 static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
1386 elv_rb_del(&cfqq->sort_list, rq);
1387 cfqq->queued[rq_is_sync(rq)]--;
1391 static struct request *
1392 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
1394 struct task_struct *tsk = current;
1395 struct cfq_io_context *cic;
1396 struct cfq_queue *cfqq;
1398 cic = cfq_cic_lookup(cfqd, tsk->io_context);
1402 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
1404 sector_t sector = bio->bi_sector + bio_sectors(bio);
1406 return elv_rb_find(&cfqq->sort_list, sector);
1412 static void cfq_activate_request(struct request_queue *q, struct request *rq)
1414 struct cfq_data *cfqd = q->elevator->elevator_data;
1416 cfqd->rq_in_driver[rq_is_sync(rq)]++;
1417 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
1418 rq_in_driver(cfqd));
1420 cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
1423 static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
1425 struct cfq_data *cfqd = q->elevator->elevator_data;
1426 const int sync = rq_is_sync(rq);
1428 WARN_ON(!cfqd->rq_in_driver[sync]);
1429 cfqd->rq_in_driver[sync]--;
1430 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
1431 rq_in_driver(cfqd));
1434 static void cfq_remove_request(struct request *rq)
1436 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1438 if (cfqq->next_rq == rq)
1439 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
1441 list_del_init(&rq->queuelist);
1444 cfqq->cfqd->rq_queued--;
1445 if (rq_is_meta(rq)) {
1446 WARN_ON(!cfqq->meta_pending);
1447 cfqq->meta_pending--;
1451 static int cfq_merge(struct request_queue *q, struct request **req,
1454 struct cfq_data *cfqd = q->elevator->elevator_data;
1455 struct request *__rq;
1457 __rq = cfq_find_rq_fmerge(cfqd, bio);
1458 if (__rq && elv_rq_merge_ok(__rq, bio)) {
1460 return ELEVATOR_FRONT_MERGE;
1463 return ELEVATOR_NO_MERGE;
1466 static void cfq_merged_request(struct request_queue *q, struct request *req,
1469 if (type == ELEVATOR_FRONT_MERGE) {
1470 struct cfq_queue *cfqq = RQ_CFQQ(req);
1472 cfq_reposition_rq_rb(cfqq, req);
1477 cfq_merged_requests(struct request_queue *q, struct request *rq,
1478 struct request *next)
1480 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1482 * reposition in fifo if next is older than rq
1484 if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
1485 time_before(rq_fifo_time(next), rq_fifo_time(rq))) {
1486 list_move(&rq->queuelist, &next->queuelist);
1487 rq_set_fifo_time(rq, rq_fifo_time(next));
1490 if (cfqq->next_rq == next)
1492 cfq_remove_request(next);
1495 static int cfq_allow_merge(struct request_queue *q, struct request *rq,
1498 struct cfq_data *cfqd = q->elevator->elevator_data;
1499 struct cfq_io_context *cic;
1500 struct cfq_queue *cfqq;
1503 * Disallow merge of a sync bio into an async request.
1505 if (cfq_bio_sync(bio) && !rq_is_sync(rq))
1509 * Lookup the cfqq that this bio will be queued with. Allow
1510 * merge only if rq is queued there.
1512 cic = cfq_cic_lookup(cfqd, current->io_context);
1516 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
1517 return cfqq == RQ_CFQQ(rq);
1520 static void __cfq_set_active_queue(struct cfq_data *cfqd,
1521 struct cfq_queue *cfqq)
1524 cfq_log_cfqq(cfqd, cfqq, "set_active");
1525 cfqq->slice_start = 0;
1526 cfqq->dispatch_start = jiffies;
1527 cfqq->allocated_slice = 0;
1528 cfqq->slice_end = 0;
1529 cfqq->slice_dispatch = 0;
1530 cfqq->nr_sectors = 0;
1532 cfq_clear_cfqq_wait_request(cfqq);
1533 cfq_clear_cfqq_must_dispatch(cfqq);
1534 cfq_clear_cfqq_must_alloc_slice(cfqq);
1535 cfq_clear_cfqq_fifo_expire(cfqq);
1536 cfq_mark_cfqq_slice_new(cfqq);
1538 del_timer(&cfqd->idle_slice_timer);
1541 cfqd->active_queue = cfqq;
1545 * current cfqq expired its slice (or was too idle), select new one
1548 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1551 cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
1553 if (cfq_cfqq_wait_request(cfqq))
1554 del_timer(&cfqd->idle_slice_timer);
1556 cfq_clear_cfqq_wait_request(cfqq);
1557 cfq_clear_cfqq_wait_busy(cfqq);
1560 * If this cfqq is shared between multiple processes, check to
1561 * make sure that those processes are still issuing I/Os within
1562 * the mean seek distance. If not, it may be time to break the
1563 * queues apart again.
1565 if (cfq_cfqq_coop(cfqq) && CFQQ_SEEKY(cfqq))
1566 cfq_mark_cfqq_split_coop(cfqq);
1569 * store what was left of this slice, if the queue idled/timed out
1571 if (timed_out && !cfq_cfqq_slice_new(cfqq)) {
1572 cfqq->slice_resid = cfqq->slice_end - jiffies;
1573 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
1576 cfq_group_served(cfqd, cfqq->cfqg, cfqq);
1578 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
1579 cfq_del_cfqq_rr(cfqd, cfqq);
1581 cfq_resort_rr_list(cfqd, cfqq);
1583 if (cfqq == cfqd->active_queue)
1584 cfqd->active_queue = NULL;
1586 if (&cfqq->cfqg->rb_node == cfqd->grp_service_tree.active)
1587 cfqd->grp_service_tree.active = NULL;
1589 if (cfqd->active_cic) {
1590 put_io_context(cfqd->active_cic->ioc);
1591 cfqd->active_cic = NULL;
1595 static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out)
1597 struct cfq_queue *cfqq = cfqd->active_queue;
1600 __cfq_slice_expired(cfqd, cfqq, timed_out);
1604 * Get next queue for service. Unless we have a queue preemption,
1605 * we'll simply select the first cfqq in the service tree.
1607 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
1609 struct cfq_rb_root *service_tree =
1610 service_tree_for(cfqd->serving_group, cfqd->serving_prio,
1611 cfqd->serving_type);
1613 if (!cfqd->rq_queued)
1616 /* There is nothing to dispatch */
1619 if (RB_EMPTY_ROOT(&service_tree->rb))
1621 return cfq_rb_first(service_tree);
1624 static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd)
1626 struct cfq_group *cfqg;
1627 struct cfq_queue *cfqq;
1629 struct cfq_rb_root *st;
1631 if (!cfqd->rq_queued)
1634 cfqg = cfq_get_next_cfqg(cfqd);
1638 for_each_cfqg_st(cfqg, i, j, st)
1639 if ((cfqq = cfq_rb_first(st)) != NULL)
1645 * Get and set a new active queue for service.
1647 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
1648 struct cfq_queue *cfqq)
1651 cfqq = cfq_get_next_queue(cfqd);
1653 __cfq_set_active_queue(cfqd, cfqq);
1657 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
1660 if (blk_rq_pos(rq) >= cfqd->last_position)
1661 return blk_rq_pos(rq) - cfqd->last_position;
1663 return cfqd->last_position - blk_rq_pos(rq);
1666 static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1667 struct request *rq, bool for_preempt)
1669 return cfq_dist_from_last(cfqd, rq) <= CFQQ_SEEK_THR;
1672 static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
1673 struct cfq_queue *cur_cfqq)
1675 struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
1676 struct rb_node *parent, *node;
1677 struct cfq_queue *__cfqq;
1678 sector_t sector = cfqd->last_position;
1680 if (RB_EMPTY_ROOT(root))
1684 * First, if we find a request starting at the end of the last
1685 * request, choose it.
1687 __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
1692 * If the exact sector wasn't found, the parent of the NULL leaf
1693 * will contain the closest sector.
1695 __cfqq = rb_entry(parent, struct cfq_queue, p_node);
1696 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq, false))
1699 if (blk_rq_pos(__cfqq->next_rq) < sector)
1700 node = rb_next(&__cfqq->p_node);
1702 node = rb_prev(&__cfqq->p_node);
1706 __cfqq = rb_entry(node, struct cfq_queue, p_node);
1707 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq, false))
1715 * cur_cfqq - passed in so that we don't decide that the current queue is
1716 * closely cooperating with itself.
1718 * So, basically we're assuming that that cur_cfqq has dispatched at least
1719 * one request, and that cfqd->last_position reflects a position on the disk
1720 * associated with the I/O issued by cur_cfqq. I'm not sure this is a valid
1723 static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
1724 struct cfq_queue *cur_cfqq)
1726 struct cfq_queue *cfqq;
1728 if (!cfq_cfqq_sync(cur_cfqq))
1730 if (CFQQ_SEEKY(cur_cfqq))
1734 * Don't search priority tree if it's the only queue in the group.
1736 if (cur_cfqq->cfqg->nr_cfqq == 1)
1740 * We should notice if some of the queues are cooperating, eg
1741 * working closely on the same area of the disk. In that case,
1742 * we can group them together and don't waste time idling.
1744 cfqq = cfqq_close(cfqd, cur_cfqq);
1748 /* If new queue belongs to different cfq_group, don't choose it */
1749 if (cur_cfqq->cfqg != cfqq->cfqg)
1753 * It only makes sense to merge sync queues.
1755 if (!cfq_cfqq_sync(cfqq))
1757 if (CFQQ_SEEKY(cfqq))
1761 * Do not merge queues of different priority classes
1763 if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq))
1770 * Determine whether we should enforce idle window for this queue.
1773 static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1775 enum wl_prio_t prio = cfqq_prio(cfqq);
1776 struct cfq_rb_root *service_tree = cfqq->service_tree;
1778 BUG_ON(!service_tree);
1779 BUG_ON(!service_tree->count);
1781 /* We never do for idle class queues. */
1782 if (prio == IDLE_WORKLOAD)
1785 /* We do for queues that were marked with idle window flag. */
1786 if (cfq_cfqq_idle_window(cfqq) &&
1787 !(blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag))
1791 * Otherwise, we do only if they are the last ones
1792 * in their service tree.
1794 return service_tree->count == 1 && cfq_cfqq_sync(cfqq);
1797 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
1799 struct cfq_queue *cfqq = cfqd->active_queue;
1800 struct cfq_io_context *cic;
1804 * SSD device without seek penalty, disable idling. But only do so
1805 * for devices that support queuing, otherwise we still have a problem
1806 * with sync vs async workloads.
1808 if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
1811 WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
1812 WARN_ON(cfq_cfqq_slice_new(cfqq));
1815 * idle is disabled, either manually or by past process history
1817 if (!cfqd->cfq_slice_idle || !cfq_should_idle(cfqd, cfqq))
1821 * still active requests from this queue, don't idle
1823 if (cfqq->dispatched)
1827 * task has exited, don't wait
1829 cic = cfqd->active_cic;
1830 if (!cic || !atomic_read(&cic->ioc->nr_tasks))
1834 * If our average think time is larger than the remaining time
1835 * slice, then don't idle. This avoids overrunning the allotted
1838 if (sample_valid(cic->ttime_samples) &&
1839 (cfqq->slice_end - jiffies < cic->ttime_mean))
1842 cfq_mark_cfqq_wait_request(cfqq);
1844 sl = cfqd->cfq_slice_idle;
1846 mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
1847 cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu", sl);
1851 * Move request from internal lists to the request queue dispatch list.
1853 static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
1855 struct cfq_data *cfqd = q->elevator->elevator_data;
1856 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1858 cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
1860 cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
1861 cfq_remove_request(rq);
1863 elv_dispatch_sort(q, rq);
1865 if (cfq_cfqq_sync(cfqq))
1866 cfqd->sync_flight++;
1867 cfqq->nr_sectors += blk_rq_sectors(rq);
1871 * return expired entry, or NULL to just start from scratch in rbtree
1873 static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
1875 struct request *rq = NULL;
1877 if (cfq_cfqq_fifo_expire(cfqq))
1880 cfq_mark_cfqq_fifo_expire(cfqq);
1882 if (list_empty(&cfqq->fifo))
1885 rq = rq_entry_fifo(cfqq->fifo.next);
1886 if (time_before(jiffies, rq_fifo_time(rq)))
1889 cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq);
1894 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1896 const int base_rq = cfqd->cfq_slice_async_rq;
1898 WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
1900 return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
1904 * Must be called with the queue_lock held.
1906 static int cfqq_process_refs(struct cfq_queue *cfqq)
1908 int process_refs, io_refs;
1910 io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE];
1911 process_refs = atomic_read(&cfqq->ref) - io_refs;
1912 BUG_ON(process_refs < 0);
1913 return process_refs;
1916 static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq)
1918 int process_refs, new_process_refs;
1919 struct cfq_queue *__cfqq;
1921 /* Avoid a circular list and skip interim queue merges */
1922 while ((__cfqq = new_cfqq->new_cfqq)) {
1928 process_refs = cfqq_process_refs(cfqq);
1930 * If the process for the cfqq has gone away, there is no
1931 * sense in merging the queues.
1933 if (process_refs == 0)
1937 * Merge in the direction of the lesser amount of work.
1939 new_process_refs = cfqq_process_refs(new_cfqq);
1940 if (new_process_refs >= process_refs) {
1941 cfqq->new_cfqq = new_cfqq;
1942 atomic_add(process_refs, &new_cfqq->ref);
1944 new_cfqq->new_cfqq = cfqq;
1945 atomic_add(new_process_refs, &cfqq->ref);
1949 static enum wl_type_t cfq_choose_wl(struct cfq_data *cfqd,
1950 struct cfq_group *cfqg, enum wl_prio_t prio)
1952 struct cfq_queue *queue;
1954 bool key_valid = false;
1955 unsigned long lowest_key = 0;
1956 enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD;
1958 for (i = 0; i <= SYNC_WORKLOAD; ++i) {
1959 /* select the one with lowest rb_key */
1960 queue = cfq_rb_first(service_tree_for(cfqg, prio, i));
1962 (!key_valid || time_before(queue->rb_key, lowest_key))) {
1963 lowest_key = queue->rb_key;
1972 static void choose_service_tree(struct cfq_data *cfqd, struct cfq_group *cfqg)
1976 struct cfq_rb_root *st;
1977 unsigned group_slice;
1980 cfqd->serving_prio = IDLE_WORKLOAD;
1981 cfqd->workload_expires = jiffies + 1;
1985 /* Choose next priority. RT > BE > IDLE */
1986 if (cfq_group_busy_queues_wl(RT_WORKLOAD, cfqd, cfqg))
1987 cfqd->serving_prio = RT_WORKLOAD;
1988 else if (cfq_group_busy_queues_wl(BE_WORKLOAD, cfqd, cfqg))
1989 cfqd->serving_prio = BE_WORKLOAD;
1991 cfqd->serving_prio = IDLE_WORKLOAD;
1992 cfqd->workload_expires = jiffies + 1;
1997 * For RT and BE, we have to choose also the type
1998 * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
2001 st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
2005 * check workload expiration, and that we still have other queues ready
2007 if (count && !time_after(jiffies, cfqd->workload_expires))
2010 /* otherwise select new workload type */
2011 cfqd->serving_type =
2012 cfq_choose_wl(cfqd, cfqg, cfqd->serving_prio);
2013 st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
2017 * the workload slice is computed as a fraction of target latency
2018 * proportional to the number of queues in that workload, over
2019 * all the queues in the same priority class
2021 group_slice = cfq_group_slice(cfqd, cfqg);
2023 slice = group_slice * count /
2024 max_t(unsigned, cfqg->busy_queues_avg[cfqd->serving_prio],
2025 cfq_group_busy_queues_wl(cfqd->serving_prio, cfqd, cfqg));
2027 if (cfqd->serving_type == ASYNC_WORKLOAD) {
2031 * Async queues are currently system wide. Just taking
2032 * proportion of queues with-in same group will lead to higher
2033 * async ratio system wide as generally root group is going
2034 * to have higher weight. A more accurate thing would be to
2035 * calculate system wide asnc/sync ratio.
2037 tmp = cfq_target_latency * cfqg_busy_async_queues(cfqd, cfqg);
2038 tmp = tmp/cfqd->busy_queues;
2039 slice = min_t(unsigned, slice, tmp);
2041 /* async workload slice is scaled down according to
2042 * the sync/async slice ratio. */
2043 slice = slice * cfqd->cfq_slice[0] / cfqd->cfq_slice[1];
2045 /* sync workload slice is at least 2 * cfq_slice_idle */
2046 slice = max(slice, 2 * cfqd->cfq_slice_idle);
2048 slice = max_t(unsigned, slice, CFQ_MIN_TT);
2049 cfqd->workload_expires = jiffies + slice;
2050 cfqd->noidle_tree_requires_idle = false;
2053 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd)
2055 struct cfq_rb_root *st = &cfqd->grp_service_tree;
2056 struct cfq_group *cfqg;
2058 if (RB_EMPTY_ROOT(&st->rb))
2060 cfqg = cfq_rb_first_group(st);
2061 st->active = &cfqg->rb_node;
2062 update_min_vdisktime(st);
2066 static void cfq_choose_cfqg(struct cfq_data *cfqd)
2068 struct cfq_group *cfqg = cfq_get_next_cfqg(cfqd);
2070 cfqd->serving_group = cfqg;
2072 /* Restore the workload type data */
2073 if (cfqg->saved_workload_slice) {
2074 cfqd->workload_expires = jiffies + cfqg->saved_workload_slice;
2075 cfqd->serving_type = cfqg->saved_workload;
2076 cfqd->serving_prio = cfqg->saved_serving_prio;
2078 cfqd->workload_expires = jiffies - 1;
2080 choose_service_tree(cfqd, cfqg);
2084 * Select a queue for service. If we have a current active queue,
2085 * check whether to continue servicing it, or retrieve and set a new one.
2087 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
2089 struct cfq_queue *cfqq, *new_cfqq = NULL;
2091 cfqq = cfqd->active_queue;
2095 if (!cfqd->rq_queued)
2099 * We were waiting for group to get backlogged. Expire the queue
2101 if (cfq_cfqq_wait_busy(cfqq) && !RB_EMPTY_ROOT(&cfqq->sort_list))
2105 * The active queue has run out of time, expire it and select new.
2107 if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq)) {
2109 * If slice had not expired at the completion of last request
2110 * we might not have turned on wait_busy flag. Don't expire
2111 * the queue yet. Allow the group to get backlogged.
2113 * The very fact that we have used the slice, that means we
2114 * have been idling all along on this queue and it should be
2115 * ok to wait for this request to complete.
2117 if (cfqq->cfqg->nr_cfqq == 1 && RB_EMPTY_ROOT(&cfqq->sort_list)
2118 && cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2126 * The active queue has requests and isn't expired, allow it to
2129 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
2133 * If another queue has a request waiting within our mean seek
2134 * distance, let it run. The expire code will check for close
2135 * cooperators and put the close queue at the front of the service
2136 * tree. If possible, merge the expiring queue with the new cfqq.
2138 new_cfqq = cfq_close_cooperator(cfqd, cfqq);
2140 if (!cfqq->new_cfqq)
2141 cfq_setup_merge(cfqq, new_cfqq);
2146 * No requests pending. If the active queue still has requests in
2147 * flight or is idling for a new request, allow either of these
2148 * conditions to happen (or time out) before selecting a new queue.
2150 if (timer_pending(&cfqd->idle_slice_timer) ||
2151 (cfqq->dispatched && cfq_should_idle(cfqd, cfqq))) {
2157 cfq_slice_expired(cfqd, 0);
2160 * Current queue expired. Check if we have to switch to a new
2164 cfq_choose_cfqg(cfqd);
2166 cfqq = cfq_set_active_queue(cfqd, new_cfqq);
2171 static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
2175 while (cfqq->next_rq) {
2176 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
2180 BUG_ON(!list_empty(&cfqq->fifo));
2182 /* By default cfqq is not expired if it is empty. Do it explicitly */
2183 __cfq_slice_expired(cfqq->cfqd, cfqq, 0);
2188 * Drain our current requests. Used for barriers and when switching
2189 * io schedulers on-the-fly.
2191 static int cfq_forced_dispatch(struct cfq_data *cfqd)
2193 struct cfq_queue *cfqq;
2196 while ((cfqq = cfq_get_next_queue_forced(cfqd)) != NULL)
2197 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
2199 cfq_slice_expired(cfqd, 0);
2200 BUG_ON(cfqd->busy_queues);
2202 cfq_log(cfqd, "forced_dispatch=%d", dispatched);
2206 static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2208 unsigned int max_dispatch;
2211 * Drain async requests before we start sync IO
2213 if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_driver[BLK_RW_ASYNC])
2217 * If this is an async queue and we have sync IO in flight, let it wait
2219 if (cfqd->sync_flight && !cfq_cfqq_sync(cfqq))
2222 max_dispatch = cfqd->cfq_quantum;
2223 if (cfq_class_idle(cfqq))
2227 * Does this cfqq already have too much IO in flight?
2229 if (cfqq->dispatched >= max_dispatch) {
2231 * idle queue must always only have a single IO in flight
2233 if (cfq_class_idle(cfqq))
2237 * We have other queues, don't allow more IO from this one
2239 if (cfqd->busy_queues > 1)
2243 * Sole queue user, no limit
2249 * Async queues must wait a bit before being allowed dispatch.
2250 * We also ramp up the dispatch depth gradually for async IO,
2251 * based on the last sync IO we serviced
2253 if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
2254 unsigned long last_sync = jiffies - cfqd->last_delayed_sync;
2257 depth = last_sync / cfqd->cfq_slice[1];
2258 if (!depth && !cfqq->dispatched)
2260 if (depth < max_dispatch)
2261 max_dispatch = depth;
2265 * If we're below the current max, allow a dispatch
2267 return cfqq->dispatched < max_dispatch;
2271 * Dispatch a request from cfqq, moving them to the request queue
2274 static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2278 BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
2280 if (!cfq_may_dispatch(cfqd, cfqq))
2284 * follow expired path, else get first next available
2286 rq = cfq_check_fifo(cfqq);
2291 * insert request into driver dispatch list
2293 cfq_dispatch_insert(cfqd->queue, rq);
2295 if (!cfqd->active_cic) {
2296 struct cfq_io_context *cic = RQ_CIC(rq);
2298 atomic_long_inc(&cic->ioc->refcount);
2299 cfqd->active_cic = cic;
2306 * Find the cfqq that we need to service and move a request from that to the
2309 static int cfq_dispatch_requests(struct request_queue *q, int force)
2311 struct cfq_data *cfqd = q->elevator->elevator_data;
2312 struct cfq_queue *cfqq;
2314 if (!cfqd->busy_queues)
2317 if (unlikely(force))
2318 return cfq_forced_dispatch(cfqd);
2320 cfqq = cfq_select_queue(cfqd);
2325 * Dispatch a request from this cfqq, if it is allowed
2327 if (!cfq_dispatch_request(cfqd, cfqq))
2330 cfqq->slice_dispatch++;
2331 cfq_clear_cfqq_must_dispatch(cfqq);
2334 * expire an async queue immediately if it has used up its slice. idle
2335 * queue always expire after 1 dispatch round.
2337 if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
2338 cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
2339 cfq_class_idle(cfqq))) {
2340 cfqq->slice_end = jiffies + 1;
2341 cfq_slice_expired(cfqd, 0);
2344 cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
2349 * task holds one reference to the queue, dropped when task exits. each rq
2350 * in-flight on this queue also holds a reference, dropped when rq is freed.
2352 * Each cfq queue took a reference on the parent group. Drop it now.
2353 * queue lock must be held here.
2355 static void cfq_put_queue(struct cfq_queue *cfqq)
2357 struct cfq_data *cfqd = cfqq->cfqd;
2358 struct cfq_group *cfqg, *orig_cfqg;
2360 BUG_ON(atomic_read(&cfqq->ref) <= 0);
2362 if (!atomic_dec_and_test(&cfqq->ref))
2365 cfq_log_cfqq(cfqd, cfqq, "put_queue");
2366 BUG_ON(rb_first(&cfqq->sort_list));
2367 BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
2369 orig_cfqg = cfqq->orig_cfqg;
2371 if (unlikely(cfqd->active_queue == cfqq)) {
2372 __cfq_slice_expired(cfqd, cfqq, 0);
2373 cfq_schedule_dispatch(cfqd);
2376 BUG_ON(cfq_cfqq_on_rr(cfqq));
2377 kmem_cache_free(cfq_pool, cfqq);
2380 cfq_put_cfqg(orig_cfqg);
2384 * Must always be called with the rcu_read_lock() held
2387 __call_for_each_cic(struct io_context *ioc,
2388 void (*func)(struct io_context *, struct cfq_io_context *))
2390 struct cfq_io_context *cic;
2391 struct hlist_node *n;
2393 hlist_for_each_entry_rcu(cic, n, &ioc->cic_list, cic_list)
2398 * Call func for each cic attached to this ioc.
2401 call_for_each_cic(struct io_context *ioc,
2402 void (*func)(struct io_context *, struct cfq_io_context *))
2405 __call_for_each_cic(ioc, func);
2409 static void cfq_cic_free_rcu(struct rcu_head *head)
2411 struct cfq_io_context *cic;
2413 cic = container_of(head, struct cfq_io_context, rcu_head);
2415 kmem_cache_free(cfq_ioc_pool, cic);
2416 elv_ioc_count_dec(cfq_ioc_count);
2420 * CFQ scheduler is exiting, grab exit lock and check
2421 * the pending io context count. If it hits zero,
2422 * complete ioc_gone and set it back to NULL
2424 spin_lock(&ioc_gone_lock);
2425 if (ioc_gone && !elv_ioc_count_read(cfq_ioc_count)) {
2429 spin_unlock(&ioc_gone_lock);
2433 static void cfq_cic_free(struct cfq_io_context *cic)
2435 call_rcu(&cic->rcu_head, cfq_cic_free_rcu);
2438 static void cic_free_func(struct io_context *ioc, struct cfq_io_context *cic)
2440 unsigned long flags;
2442 BUG_ON(!cic->dead_key);
2444 spin_lock_irqsave(&ioc->lock, flags);
2445 radix_tree_delete(&ioc->radix_root, cic->dead_key);
2446 hlist_del_rcu(&cic->cic_list);
2447 spin_unlock_irqrestore(&ioc->lock, flags);
2453 * Must be called with rcu_read_lock() held or preemption otherwise disabled.
2454 * Only two callers of this - ->dtor() which is called with the rcu_read_lock(),
2455 * and ->trim() which is called with the task lock held
2457 static void cfq_free_io_context(struct io_context *ioc)
2460 * ioc->refcount is zero here, or we are called from elv_unregister(),
2461 * so no more cic's are allowed to be linked into this ioc. So it
2462 * should be ok to iterate over the known list, we will see all cic's
2463 * since no new ones are added.
2465 __call_for_each_cic(ioc, cic_free_func);
2468 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2470 struct cfq_queue *__cfqq, *next;
2472 if (unlikely(cfqq == cfqd->active_queue)) {
2473 __cfq_slice_expired(cfqd, cfqq, 0);
2474 cfq_schedule_dispatch(cfqd);
2478 * If this queue was scheduled to merge with another queue, be
2479 * sure to drop the reference taken on that queue (and others in
2480 * the merge chain). See cfq_setup_merge and cfq_merge_cfqqs.
2482 __cfqq = cfqq->new_cfqq;
2484 if (__cfqq == cfqq) {
2485 WARN(1, "cfqq->new_cfqq loop detected\n");
2488 next = __cfqq->new_cfqq;
2489 cfq_put_queue(__cfqq);
2493 cfq_put_queue(cfqq);
2496 static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
2497 struct cfq_io_context *cic)
2499 struct io_context *ioc = cic->ioc;
2501 list_del_init(&cic->queue_list);
2504 * Make sure key == NULL is seen for dead queues
2507 cic->dead_key = (unsigned long) cic->key;
2510 if (ioc->ioc_data == cic)
2511 rcu_assign_pointer(ioc->ioc_data, NULL);
2513 if (cic->cfqq[BLK_RW_ASYNC]) {
2514 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]);
2515 cic->cfqq[BLK_RW_ASYNC] = NULL;
2518 if (cic->cfqq[BLK_RW_SYNC]) {
2519 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]);
2520 cic->cfqq[BLK_RW_SYNC] = NULL;
2524 static void cfq_exit_single_io_context(struct io_context *ioc,
2525 struct cfq_io_context *cic)
2527 struct cfq_data *cfqd = cic->key;
2530 struct request_queue *q = cfqd->queue;
2531 unsigned long flags;
2533 spin_lock_irqsave(q->queue_lock, flags);
2536 * Ensure we get a fresh copy of the ->key to prevent
2537 * race between exiting task and queue
2539 smp_read_barrier_depends();
2541 __cfq_exit_single_io_context(cfqd, cic);
2543 spin_unlock_irqrestore(q->queue_lock, flags);
2548 * The process that ioc belongs to has exited, we need to clean up
2549 * and put the internal structures we have that belongs to that process.
2551 static void cfq_exit_io_context(struct io_context *ioc)
2553 call_for_each_cic(ioc, cfq_exit_single_io_context);
2556 static struct cfq_io_context *
2557 cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
2559 struct cfq_io_context *cic;
2561 cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO,
2564 cic->last_end_request = jiffies;
2565 INIT_LIST_HEAD(&cic->queue_list);
2566 INIT_HLIST_NODE(&cic->cic_list);
2567 cic->dtor = cfq_free_io_context;
2568 cic->exit = cfq_exit_io_context;
2569 elv_ioc_count_inc(cfq_ioc_count);
2575 static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc)
2577 struct task_struct *tsk = current;
2580 if (!cfq_cfqq_prio_changed(cfqq))
2583 ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio);
2584 switch (ioprio_class) {
2586 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
2587 case IOPRIO_CLASS_NONE:
2589 * no prio set, inherit CPU scheduling settings
2591 cfqq->ioprio = task_nice_ioprio(tsk);
2592 cfqq->ioprio_class = task_nice_ioclass(tsk);
2594 case IOPRIO_CLASS_RT:
2595 cfqq->ioprio = task_ioprio(ioc);
2596 cfqq->ioprio_class = IOPRIO_CLASS_RT;
2598 case IOPRIO_CLASS_BE:
2599 cfqq->ioprio = task_ioprio(ioc);
2600 cfqq->ioprio_class = IOPRIO_CLASS_BE;
2602 case IOPRIO_CLASS_IDLE:
2603 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
2605 cfq_clear_cfqq_idle_window(cfqq);
2610 * keep track of original prio settings in case we have to temporarily
2611 * elevate the priority of this queue
2613 cfqq->org_ioprio = cfqq->ioprio;
2614 cfqq->org_ioprio_class = cfqq->ioprio_class;
2615 cfq_clear_cfqq_prio_changed(cfqq);
2618 static void changed_ioprio(struct io_context *ioc, struct cfq_io_context *cic)
2620 struct cfq_data *cfqd = cic->key;
2621 struct cfq_queue *cfqq;
2622 unsigned long flags;
2624 if (unlikely(!cfqd))
2627 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2629 cfqq = cic->cfqq[BLK_RW_ASYNC];
2631 struct cfq_queue *new_cfqq;
2632 new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic->ioc,
2635 cic->cfqq[BLK_RW_ASYNC] = new_cfqq;
2636 cfq_put_queue(cfqq);
2640 cfqq = cic->cfqq[BLK_RW_SYNC];
2642 cfq_mark_cfqq_prio_changed(cfqq);
2644 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2647 static void cfq_ioc_set_ioprio(struct io_context *ioc)
2649 call_for_each_cic(ioc, changed_ioprio);
2650 ioc->ioprio_changed = 0;
2653 static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2654 pid_t pid, bool is_sync)
2656 RB_CLEAR_NODE(&cfqq->rb_node);
2657 RB_CLEAR_NODE(&cfqq->p_node);
2658 INIT_LIST_HEAD(&cfqq->fifo);
2660 atomic_set(&cfqq->ref, 0);
2663 cfq_mark_cfqq_prio_changed(cfqq);
2666 if (!cfq_class_idle(cfqq))
2667 cfq_mark_cfqq_idle_window(cfqq);
2668 cfq_mark_cfqq_sync(cfqq);
2673 #ifdef CONFIG_CFQ_GROUP_IOSCHED
2674 static void changed_cgroup(struct io_context *ioc, struct cfq_io_context *cic)
2676 struct cfq_queue *sync_cfqq = cic_to_cfqq(cic, 1);
2677 struct cfq_data *cfqd = cic->key;
2678 unsigned long flags;
2679 struct request_queue *q;
2681 if (unlikely(!cfqd))
2686 spin_lock_irqsave(q->queue_lock, flags);
2690 * Drop reference to sync queue. A new sync queue will be
2691 * assigned in new group upon arrival of a fresh request.
2693 cfq_log_cfqq(cfqd, sync_cfqq, "changed cgroup");
2694 cic_set_cfqq(cic, NULL, 1);
2695 cfq_put_queue(sync_cfqq);
2698 spin_unlock_irqrestore(q->queue_lock, flags);
2701 static void cfq_ioc_set_cgroup(struct io_context *ioc)
2703 call_for_each_cic(ioc, changed_cgroup);
2704 ioc->cgroup_changed = 0;
2706 #endif /* CONFIG_CFQ_GROUP_IOSCHED */
2708 static struct cfq_queue *
2709 cfq_find_alloc_queue(struct cfq_data *cfqd, bool is_sync,
2710 struct io_context *ioc, gfp_t gfp_mask)
2712 struct cfq_queue *cfqq, *new_cfqq = NULL;
2713 struct cfq_io_context *cic;
2714 struct cfq_group *cfqg;
2717 cfqg = cfq_get_cfqg(cfqd, 1);
2718 cic = cfq_cic_lookup(cfqd, ioc);
2719 /* cic always exists here */
2720 cfqq = cic_to_cfqq(cic, is_sync);
2723 * Always try a new alloc if we fell back to the OOM cfqq
2724 * originally, since it should just be a temporary situation.
2726 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
2731 } else if (gfp_mask & __GFP_WAIT) {
2732 spin_unlock_irq(cfqd->queue->queue_lock);
2733 new_cfqq = kmem_cache_alloc_node(cfq_pool,
2734 gfp_mask | __GFP_ZERO,
2736 spin_lock_irq(cfqd->queue->queue_lock);
2740 cfqq = kmem_cache_alloc_node(cfq_pool,
2741 gfp_mask | __GFP_ZERO,
2746 cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
2747 cfq_init_prio_data(cfqq, ioc);
2748 cfq_link_cfqq_cfqg(cfqq, cfqg);
2749 cfq_log_cfqq(cfqd, cfqq, "alloced");
2751 cfqq = &cfqd->oom_cfqq;
2755 kmem_cache_free(cfq_pool, new_cfqq);
2760 static struct cfq_queue **
2761 cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
2763 switch (ioprio_class) {
2764 case IOPRIO_CLASS_RT:
2765 return &cfqd->async_cfqq[0][ioprio];
2766 case IOPRIO_CLASS_BE:
2767 return &cfqd->async_cfqq[1][ioprio];
2768 case IOPRIO_CLASS_IDLE:
2769 return &cfqd->async_idle_cfqq;
2775 static struct cfq_queue *
2776 cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct io_context *ioc,
2779 const int ioprio = task_ioprio(ioc);
2780 const int ioprio_class = task_ioprio_class(ioc);
2781 struct cfq_queue **async_cfqq = NULL;
2782 struct cfq_queue *cfqq = NULL;
2785 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
2790 cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask);
2793 * pin the queue now that it's allocated, scheduler exit will prune it
2795 if (!is_sync && !(*async_cfqq)) {
2796 atomic_inc(&cfqq->ref);
2800 atomic_inc(&cfqq->ref);
2805 * We drop cfq io contexts lazily, so we may find a dead one.
2808 cfq_drop_dead_cic(struct cfq_data *cfqd, struct io_context *ioc,
2809 struct cfq_io_context *cic)
2811 unsigned long flags;
2813 WARN_ON(!list_empty(&cic->queue_list));
2815 spin_lock_irqsave(&ioc->lock, flags);
2817 BUG_ON(ioc->ioc_data == cic);
2819 radix_tree_delete(&ioc->radix_root, (unsigned long) cfqd);
2820 hlist_del_rcu(&cic->cic_list);
2821 spin_unlock_irqrestore(&ioc->lock, flags);
2826 static struct cfq_io_context *
2827 cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc)
2829 struct cfq_io_context *cic;
2830 unsigned long flags;
2839 * we maintain a last-hit cache, to avoid browsing over the tree
2841 cic = rcu_dereference(ioc->ioc_data);
2842 if (cic && cic->key == cfqd) {
2848 cic = radix_tree_lookup(&ioc->radix_root, (unsigned long) cfqd);
2852 /* ->key must be copied to avoid race with cfq_exit_queue() */
2855 cfq_drop_dead_cic(cfqd, ioc, cic);
2860 spin_lock_irqsave(&ioc->lock, flags);
2861 rcu_assign_pointer(ioc->ioc_data, cic);
2862 spin_unlock_irqrestore(&ioc->lock, flags);
2870 * Add cic into ioc, using cfqd as the search key. This enables us to lookup
2871 * the process specific cfq io context when entered from the block layer.
2872 * Also adds the cic to a per-cfqd list, used when this queue is removed.
2874 static int cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
2875 struct cfq_io_context *cic, gfp_t gfp_mask)
2877 unsigned long flags;
2880 ret = radix_tree_preload(gfp_mask);
2885 spin_lock_irqsave(&ioc->lock, flags);
2886 ret = radix_tree_insert(&ioc->radix_root,
2887 (unsigned long) cfqd, cic);
2889 hlist_add_head_rcu(&cic->cic_list, &ioc->cic_list);
2890 spin_unlock_irqrestore(&ioc->lock, flags);
2892 radix_tree_preload_end();
2895 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2896 list_add(&cic->queue_list, &cfqd->cic_list);
2897 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2902 printk(KERN_ERR "cfq: cic link failed!\n");
2908 * Setup general io context and cfq io context. There can be several cfq
2909 * io contexts per general io context, if this process is doing io to more
2910 * than one device managed by cfq.
2912 static struct cfq_io_context *
2913 cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
2915 struct io_context *ioc = NULL;
2916 struct cfq_io_context *cic;
2918 might_sleep_if(gfp_mask & __GFP_WAIT);
2920 ioc = get_io_context(gfp_mask, cfqd->queue->node);
2924 cic = cfq_cic_lookup(cfqd, ioc);
2928 cic = cfq_alloc_io_context(cfqd, gfp_mask);
2932 if (cfq_cic_link(cfqd, ioc, cic, gfp_mask))
2936 smp_read_barrier_depends();
2937 if (unlikely(ioc->ioprio_changed))
2938 cfq_ioc_set_ioprio(ioc);
2940 #ifdef CONFIG_CFQ_GROUP_IOSCHED
2941 if (unlikely(ioc->cgroup_changed))
2942 cfq_ioc_set_cgroup(ioc);
2948 put_io_context(ioc);
2953 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
2955 unsigned long elapsed = jiffies - cic->last_end_request;
2956 unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
2958 cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
2959 cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
2960 cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
2964 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2968 if (cfqq->last_request_pos) {
2969 if (cfqq->last_request_pos < blk_rq_pos(rq))
2970 sdist = blk_rq_pos(rq) - cfqq->last_request_pos;
2972 sdist = cfqq->last_request_pos - blk_rq_pos(rq);
2975 cfqq->seek_history <<= 1;
2976 cfqq->seek_history |= (sdist > CFQQ_SEEK_THR);
2980 * Disable idle window if the process thinks too long or seeks so much that
2984 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2985 struct cfq_io_context *cic)
2987 int old_idle, enable_idle;
2990 * Don't idle for async or idle io prio class
2992 if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
2995 enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
2997 if (cfqq->queued[0] + cfqq->queued[1] >= 4)
2998 cfq_mark_cfqq_deep(cfqq);
3000 if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle ||
3001 (!cfq_cfqq_deep(cfqq) && CFQQ_SEEKY(cfqq)))
3003 else if (sample_valid(cic->ttime_samples)) {
3004 if (cic->ttime_mean > cfqd->cfq_slice_idle)
3010 if (old_idle != enable_idle) {
3011 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
3013 cfq_mark_cfqq_idle_window(cfqq);
3015 cfq_clear_cfqq_idle_window(cfqq);
3020 * Check if new_cfqq should preempt the currently active queue. Return 0 for
3021 * no or if we aren't sure, a 1 will cause a preempt.
3024 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
3027 struct cfq_queue *cfqq;
3029 cfqq = cfqd->active_queue;
3033 if (cfq_class_idle(new_cfqq))
3036 if (cfq_class_idle(cfqq))
3040 * Don't allow a non-RT request to preempt an ongoing RT cfqq timeslice.
3042 if (cfq_class_rt(cfqq) && !cfq_class_rt(new_cfqq))
3046 * if the new request is sync, but the currently running queue is
3047 * not, let the sync request have priority.
3049 if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
3052 if (new_cfqq->cfqg != cfqq->cfqg)
3055 if (cfq_slice_used(cfqq))
3058 /* Allow preemption only if we are idling on sync-noidle tree */
3059 if (cfqd->serving_type == SYNC_NOIDLE_WORKLOAD &&
3060 cfqq_type(new_cfqq) == SYNC_NOIDLE_WORKLOAD &&
3061 new_cfqq->service_tree->count == 2 &&
3062 RB_EMPTY_ROOT(&cfqq->sort_list))
3066 * So both queues are sync. Let the new request get disk time if
3067 * it's a metadata request and the current queue is doing regular IO.
3069 if (rq_is_meta(rq) && !cfqq->meta_pending)
3073 * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
3075 if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
3078 if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
3082 * if this request is as-good as one we would expect from the
3083 * current cfqq, let it preempt
3085 if (cfq_rq_close(cfqd, cfqq, rq, true))
3092 * cfqq preempts the active queue. if we allowed preempt with no slice left,
3093 * let it have half of its nominal slice.
3095 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3097 cfq_log_cfqq(cfqd, cfqq, "preempt");
3098 cfq_slice_expired(cfqd, 1);
3101 * Put the new queue at the front of the of the current list,
3102 * so we know that it will be selected next.
3104 BUG_ON(!cfq_cfqq_on_rr(cfqq));
3106 cfq_service_tree_add(cfqd, cfqq, 1);
3108 cfqq->slice_end = 0;
3109 cfq_mark_cfqq_slice_new(cfqq);
3113 * Called when a new fs request (rq) is added (to cfqq). Check if there's
3114 * something we should do about it
3117 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3120 struct cfq_io_context *cic = RQ_CIC(rq);
3124 cfqq->meta_pending++;
3126 cfq_update_io_thinktime(cfqd, cic);
3127 cfq_update_io_seektime(cfqd, cfqq, rq);
3128 cfq_update_idle_window(cfqd, cfqq, cic);
3130 cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
3132 if (cfqq == cfqd->active_queue) {
3134 * Remember that we saw a request from this process, but
3135 * don't start queuing just yet. Otherwise we risk seeing lots
3136 * of tiny requests, because we disrupt the normal plugging
3137 * and merging. If the request is already larger than a single
3138 * page, let it rip immediately. For that case we assume that
3139 * merging is already done. Ditto for a busy system that
3140 * has other work pending, don't risk delaying until the
3141 * idle timer unplug to continue working.
3143 if (cfq_cfqq_wait_request(cfqq)) {
3144 if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
3145 cfqd->busy_queues > 1) {
3146 del_timer(&cfqd->idle_slice_timer);
3147 cfq_clear_cfqq_wait_request(cfqq);
3148 __blk_run_queue(cfqd->queue);
3150 cfq_mark_cfqq_must_dispatch(cfqq);
3152 } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
3154 * not the active queue - expire current slice if it is
3155 * idle and has expired it's mean thinktime or this new queue
3156 * has some old slice time left and is of higher priority or
3157 * this new queue is RT and the current one is BE
3159 cfq_preempt_queue(cfqd, cfqq);
3160 __blk_run_queue(cfqd->queue);
3164 static void cfq_insert_request(struct request_queue *q, struct request *rq)
3166 struct cfq_data *cfqd = q->elevator->elevator_data;
3167 struct cfq_queue *cfqq = RQ_CFQQ(rq);
3169 cfq_log_cfqq(cfqd, cfqq, "insert_request");
3170 cfq_init_prio_data(cfqq, RQ_CIC(rq)->ioc);
3172 rq_set_fifo_time(rq, jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)]);
3173 list_add_tail(&rq->queuelist, &cfqq->fifo);
3176 cfq_rq_enqueued(cfqd, cfqq, rq);
3180 * Update hw_tag based on peak queue depth over 50 samples under
3183 static void cfq_update_hw_tag(struct cfq_data *cfqd)
3185 struct cfq_queue *cfqq = cfqd->active_queue;
3187 if (rq_in_driver(cfqd) > cfqd->hw_tag_est_depth)
3188 cfqd->hw_tag_est_depth = rq_in_driver(cfqd);
3190 if (cfqd->hw_tag == 1)
3193 if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
3194 rq_in_driver(cfqd) <= CFQ_HW_QUEUE_MIN)
3198 * If active queue hasn't enough requests and can idle, cfq might not
3199 * dispatch sufficient requests to hardware. Don't zero hw_tag in this
3202 if (cfqq && cfq_cfqq_idle_window(cfqq) &&
3203 cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] <
3204 CFQ_HW_QUEUE_MIN && rq_in_driver(cfqd) < CFQ_HW_QUEUE_MIN)
3207 if (cfqd->hw_tag_samples++ < 50)
3210 if (cfqd->hw_tag_est_depth >= CFQ_HW_QUEUE_MIN)
3216 static bool cfq_should_wait_busy(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3218 struct cfq_io_context *cic = cfqd->active_cic;
3220 /* If there are other queues in the group, don't wait */
3221 if (cfqq->cfqg->nr_cfqq > 1)
3224 if (cfq_slice_used(cfqq))
3227 /* if slice left is less than think time, wait busy */
3228 if (cic && sample_valid(cic->ttime_samples)
3229 && (cfqq->slice_end - jiffies < cic->ttime_mean))
3233 * If think times is less than a jiffy than ttime_mean=0 and above
3234 * will not be true. It might happen that slice has not expired yet
3235 * but will expire soon (4-5 ns) during select_queue(). To cover the
3236 * case where think time is less than a jiffy, mark the queue wait
3237 * busy if only 1 jiffy is left in the slice.
3239 if (cfqq->slice_end - jiffies == 1)
3245 static void cfq_completed_request(struct request_queue *q, struct request *rq)
3247 struct cfq_queue *cfqq = RQ_CFQQ(rq);
3248 struct cfq_data *cfqd = cfqq->cfqd;
3249 const int sync = rq_is_sync(rq);
3253 cfq_log_cfqq(cfqd, cfqq, "complete rqnoidle %d", !!rq_noidle(rq));
3255 cfq_update_hw_tag(cfqd);
3257 WARN_ON(!cfqd->rq_in_driver[sync]);
3258 WARN_ON(!cfqq->dispatched);
3259 cfqd->rq_in_driver[sync]--;
3262 if (cfq_cfqq_sync(cfqq))
3263 cfqd->sync_flight--;
3266 RQ_CIC(rq)->last_end_request = now;
3267 if (!time_after(rq->start_time + cfqd->cfq_fifo_expire[1], now))
3268 cfqd->last_delayed_sync = now;
3272 * If this is the active queue, check if it needs to be expired,
3273 * or if we want to idle in case it has no pending requests.
3275 if (cfqd->active_queue == cfqq) {
3276 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
3278 if (cfq_cfqq_slice_new(cfqq)) {
3279 cfq_set_prio_slice(cfqd, cfqq);
3280 cfq_clear_cfqq_slice_new(cfqq);
3284 * Should we wait for next request to come in before we expire
3287 if (cfq_should_wait_busy(cfqd, cfqq)) {
3288 cfqq->slice_end = jiffies + cfqd->cfq_slice_idle;
3289 cfq_mark_cfqq_wait_busy(cfqq);
3293 * Idling is not enabled on:
3295 * - idle-priority queues
3297 * - queues with still some requests queued
3298 * - when there is a close cooperator
3300 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
3301 cfq_slice_expired(cfqd, 1);
3302 else if (sync && cfqq_empty &&
3303 !cfq_close_cooperator(cfqd, cfqq)) {
3304 cfqd->noidle_tree_requires_idle |= !rq_noidle(rq);
3306 * Idling is enabled for SYNC_WORKLOAD.
3307 * SYNC_NOIDLE_WORKLOAD idles at the end of the tree
3308 * only if we processed at least one !rq_noidle request
3310 if (cfqd->serving_type == SYNC_WORKLOAD
3311 || cfqd->noidle_tree_requires_idle
3312 || cfqq->cfqg->nr_cfqq == 1)
3313 cfq_arm_slice_timer(cfqd);
3317 if (!rq_in_driver(cfqd))
3318 cfq_schedule_dispatch(cfqd);
3322 * we temporarily boost lower priority queues if they are holding fs exclusive
3323 * resources. they are boosted to normal prio (CLASS_BE/4)
3325 static void cfq_prio_boost(struct cfq_queue *cfqq)
3327 if (has_fs_excl()) {
3329 * boost idle prio on transactions that would lock out other
3330 * users of the filesystem
3332 if (cfq_class_idle(cfqq))
3333 cfqq->ioprio_class = IOPRIO_CLASS_BE;
3334 if (cfqq->ioprio > IOPRIO_NORM)
3335 cfqq->ioprio = IOPRIO_NORM;
3338 * unboost the queue (if needed)
3340 cfqq->ioprio_class = cfqq->org_ioprio_class;
3341 cfqq->ioprio = cfqq->org_ioprio;
3345 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
3347 if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
3348 cfq_mark_cfqq_must_alloc_slice(cfqq);
3349 return ELV_MQUEUE_MUST;
3352 return ELV_MQUEUE_MAY;
3355 static int cfq_may_queue(struct request_queue *q, int rw)
3357 struct cfq_data *cfqd = q->elevator->elevator_data;
3358 struct task_struct *tsk = current;
3359 struct cfq_io_context *cic;
3360 struct cfq_queue *cfqq;
3363 * don't force setup of a queue from here, as a call to may_queue
3364 * does not necessarily imply that a request actually will be queued.
3365 * so just lookup a possibly existing queue, or return 'may queue'
3368 cic = cfq_cic_lookup(cfqd, tsk->io_context);
3370 return ELV_MQUEUE_MAY;
3372 cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
3374 cfq_init_prio_data(cfqq, cic->ioc);
3375 cfq_prio_boost(cfqq);
3377 return __cfq_may_queue(cfqq);
3380 return ELV_MQUEUE_MAY;
3384 * queue lock held here
3386 static void cfq_put_request(struct request *rq)
3388 struct cfq_queue *cfqq = RQ_CFQQ(rq);
3391 const int rw = rq_data_dir(rq);
3393 BUG_ON(!cfqq->allocated[rw]);
3394 cfqq->allocated[rw]--;
3396 put_io_context(RQ_CIC(rq)->ioc);
3398 rq->elevator_private = NULL;
3399 rq->elevator_private2 = NULL;
3401 cfq_put_queue(cfqq);
3405 static struct cfq_queue *
3406 cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_context *cic,
3407 struct cfq_queue *cfqq)
3409 cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq);
3410 cic_set_cfqq(cic, cfqq->new_cfqq, 1);
3411 cfq_mark_cfqq_coop(cfqq->new_cfqq);
3412 cfq_put_queue(cfqq);
3413 return cic_to_cfqq(cic, 1);
3417 * Returns NULL if a new cfqq should be allocated, or the old cfqq if this
3418 * was the last process referring to said cfqq.
3420 static struct cfq_queue *
3421 split_cfqq(struct cfq_io_context *cic, struct cfq_queue *cfqq)
3423 if (cfqq_process_refs(cfqq) == 1) {
3424 cfqq->pid = current->pid;
3425 cfq_clear_cfqq_coop(cfqq);
3426 cfq_clear_cfqq_split_coop(cfqq);
3430 cic_set_cfqq(cic, NULL, 1);
3431 cfq_put_queue(cfqq);
3435 * Allocate cfq data structures associated with this request.
3438 cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
3440 struct cfq_data *cfqd = q->elevator->elevator_data;
3441 struct cfq_io_context *cic;
3442 const int rw = rq_data_dir(rq);
3443 const bool is_sync = rq_is_sync(rq);
3444 struct cfq_queue *cfqq;
3445 unsigned long flags;
3447 might_sleep_if(gfp_mask & __GFP_WAIT);
3449 cic = cfq_get_io_context(cfqd, gfp_mask);
3451 spin_lock_irqsave(q->queue_lock, flags);
3457 cfqq = cic_to_cfqq(cic, is_sync);
3458 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
3459 cfqq = cfq_get_queue(cfqd, is_sync, cic->ioc, gfp_mask);
3460 cic_set_cfqq(cic, cfqq, is_sync);
3463 * If the queue was seeky for too long, break it apart.
3465 if (cfq_cfqq_coop(cfqq) && cfq_cfqq_split_coop(cfqq)) {
3466 cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq");
3467 cfqq = split_cfqq(cic, cfqq);
3473 * Check to see if this queue is scheduled to merge with
3474 * another, closely cooperating queue. The merging of
3475 * queues happens here as it must be done in process context.
3476 * The reference on new_cfqq was taken in merge_cfqqs.
3479 cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq);
3482 cfqq->allocated[rw]++;
3483 atomic_inc(&cfqq->ref);
3485 spin_unlock_irqrestore(q->queue_lock, flags);
3487 rq->elevator_private = cic;
3488 rq->elevator_private2 = cfqq;
3493 put_io_context(cic->ioc);
3495 cfq_schedule_dispatch(cfqd);
3496 spin_unlock_irqrestore(q->queue_lock, flags);
3497 cfq_log(cfqd, "set_request fail");
3501 static void cfq_kick_queue(struct work_struct *work)
3503 struct cfq_data *cfqd =
3504 container_of(work, struct cfq_data, unplug_work);
3505 struct request_queue *q = cfqd->queue;
3507 spin_lock_irq(q->queue_lock);
3508 __blk_run_queue(cfqd->queue);
3509 spin_unlock_irq(q->queue_lock);
3513 * Timer running if the active_queue is currently idling inside its time slice
3515 static void cfq_idle_slice_timer(unsigned long data)
3517 struct cfq_data *cfqd = (struct cfq_data *) data;
3518 struct cfq_queue *cfqq;
3519 unsigned long flags;
3522 cfq_log(cfqd, "idle timer fired");
3524 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
3526 cfqq = cfqd->active_queue;
3531 * We saw a request before the queue expired, let it through
3533 if (cfq_cfqq_must_dispatch(cfqq))
3539 if (cfq_slice_used(cfqq))
3543 * only expire and reinvoke request handler, if there are
3544 * other queues with pending requests
3546 if (!cfqd->busy_queues)
3550 * not expired and it has a request pending, let it dispatch
3552 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3556 * Queue depth flag is reset only when the idle didn't succeed
3558 cfq_clear_cfqq_deep(cfqq);
3561 cfq_slice_expired(cfqd, timed_out);
3563 cfq_schedule_dispatch(cfqd);
3565 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
3568 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
3570 del_timer_sync(&cfqd->idle_slice_timer);
3571 cancel_work_sync(&cfqd->unplug_work);
3574 static void cfq_put_async_queues(struct cfq_data *cfqd)
3578 for (i = 0; i < IOPRIO_BE_NR; i++) {
3579 if (cfqd->async_cfqq[0][i])
3580 cfq_put_queue(cfqd->async_cfqq[0][i]);
3581 if (cfqd->async_cfqq[1][i])
3582 cfq_put_queue(cfqd->async_cfqq[1][i]);
3585 if (cfqd->async_idle_cfqq)
3586 cfq_put_queue(cfqd->async_idle_cfqq);
3589 static void cfq_cfqd_free(struct rcu_head *head)
3591 kfree(container_of(head, struct cfq_data, rcu));
3594 static void cfq_exit_queue(struct elevator_queue *e)
3596 struct cfq_data *cfqd = e->elevator_data;
3597 struct request_queue *q = cfqd->queue;
3599 cfq_shutdown_timer_wq(cfqd);
3601 spin_lock_irq(q->queue_lock);
3603 if (cfqd->active_queue)
3604 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
3606 while (!list_empty(&cfqd->cic_list)) {
3607 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
3608 struct cfq_io_context,
3611 __cfq_exit_single_io_context(cfqd, cic);
3614 cfq_put_async_queues(cfqd);
3615 cfq_release_cfq_groups(cfqd);
3616 blkiocg_del_blkio_group(&cfqd->root_group.blkg);
3618 spin_unlock_irq(q->queue_lock);
3620 cfq_shutdown_timer_wq(cfqd);
3622 /* Wait for cfqg->blkg->key accessors to exit their grace periods. */
3623 call_rcu(&cfqd->rcu, cfq_cfqd_free);
3626 static void *cfq_init_queue(struct request_queue *q)
3628 struct cfq_data *cfqd;
3630 struct cfq_group *cfqg;
3631 struct cfq_rb_root *st;
3633 cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
3637 /* Init root service tree */
3638 cfqd->grp_service_tree = CFQ_RB_ROOT;
3640 /* Init root group */
3641 cfqg = &cfqd->root_group;
3642 for_each_cfqg_st(cfqg, i, j, st)
3644 RB_CLEAR_NODE(&cfqg->rb_node);
3646 /* Give preference to root group over other groups */
3647 cfqg->weight = 2*BLKIO_WEIGHT_DEFAULT;
3649 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3651 * Take a reference to root group which we never drop. This is just
3652 * to make sure that cfq_put_cfqg() does not try to kfree root group
3654 atomic_set(&cfqg->ref, 1);
3655 blkiocg_add_blkio_group(&blkio_root_cgroup, &cfqg->blkg, (void *)cfqd,
3659 * Not strictly needed (since RB_ROOT just clears the node and we
3660 * zeroed cfqd on alloc), but better be safe in case someone decides
3661 * to add magic to the rb code
3663 for (i = 0; i < CFQ_PRIO_LISTS; i++)
3664 cfqd->prio_trees[i] = RB_ROOT;
3667 * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
3668 * Grab a permanent reference to it, so that the normal code flow
3669 * will not attempt to free it.
3671 cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
3672 atomic_inc(&cfqd->oom_cfqq.ref);
3673 cfq_link_cfqq_cfqg(&cfqd->oom_cfqq, &cfqd->root_group);
3675 INIT_LIST_HEAD(&cfqd->cic_list);
3679 init_timer(&cfqd->idle_slice_timer);
3680 cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
3681 cfqd->idle_slice_timer.data = (unsigned long) cfqd;
3683 INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
3685 cfqd->cfq_quantum = cfq_quantum;
3686 cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
3687 cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
3688 cfqd->cfq_back_max = cfq_back_max;
3689 cfqd->cfq_back_penalty = cfq_back_penalty;
3690 cfqd->cfq_slice[0] = cfq_slice_async;
3691 cfqd->cfq_slice[1] = cfq_slice_sync;
3692 cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
3693 cfqd->cfq_slice_idle = cfq_slice_idle;
3694 cfqd->cfq_latency = 1;
3695 cfqd->cfq_group_isolation = 0;
3698 * we optimistically start assuming sync ops weren't delayed in last
3699 * second, in order to have larger depth for async operations.
3701 cfqd->last_delayed_sync = jiffies - HZ;
3702 INIT_RCU_HEAD(&cfqd->rcu);
3706 static void cfq_slab_kill(void)
3709 * Caller already ensured that pending RCU callbacks are completed,
3710 * so we should have no busy allocations at this point.
3713 kmem_cache_destroy(cfq_pool);
3715 kmem_cache_destroy(cfq_ioc_pool);
3718 static int __init cfq_slab_setup(void)
3720 cfq_pool = KMEM_CACHE(cfq_queue, 0);
3724 cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0);
3735 * sysfs parts below -->
3738 cfq_var_show(unsigned int var, char *page)
3740 return sprintf(page, "%d\n", var);
3744 cfq_var_store(unsigned int *var, const char *page, size_t count)
3746 char *p = (char *) page;
3748 *var = simple_strtoul(p, &p, 10);
3752 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
3753 static ssize_t __FUNC(struct elevator_queue *e, char *page) \
3755 struct cfq_data *cfqd = e->elevator_data; \
3756 unsigned int __data = __VAR; \
3758 __data = jiffies_to_msecs(__data); \
3759 return cfq_var_show(__data, (page)); \
3761 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
3762 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
3763 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
3764 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
3765 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
3766 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
3767 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
3768 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
3769 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
3770 SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
3771 SHOW_FUNCTION(cfq_group_isolation_show, cfqd->cfq_group_isolation, 0);
3772 #undef SHOW_FUNCTION
3774 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
3775 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
3777 struct cfq_data *cfqd = e->elevator_data; \
3778 unsigned int __data; \
3779 int ret = cfq_var_store(&__data, (page), count); \
3780 if (__data < (MIN)) \
3782 else if (__data > (MAX)) \
3785 *(__PTR) = msecs_to_jiffies(__data); \
3787 *(__PTR) = __data; \
3790 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
3791 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
3793 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
3795 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
3796 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
3798 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
3799 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
3800 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
3801 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
3803 STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
3804 STORE_FUNCTION(cfq_group_isolation_store, &cfqd->cfq_group_isolation, 0, 1, 0);
3805 #undef STORE_FUNCTION
3807 #define CFQ_ATTR(name) \
3808 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
3810 static struct elv_fs_entry cfq_attrs[] = {
3812 CFQ_ATTR(fifo_expire_sync),
3813 CFQ_ATTR(fifo_expire_async),
3814 CFQ_ATTR(back_seek_max),
3815 CFQ_ATTR(back_seek_penalty),
3816 CFQ_ATTR(slice_sync),
3817 CFQ_ATTR(slice_async),
3818 CFQ_ATTR(slice_async_rq),
3819 CFQ_ATTR(slice_idle),
3820 CFQ_ATTR(low_latency),
3821 CFQ_ATTR(group_isolation),
3825 static struct elevator_type iosched_cfq = {
3827 .elevator_merge_fn = cfq_merge,
3828 .elevator_merged_fn = cfq_merged_request,
3829 .elevator_merge_req_fn = cfq_merged_requests,
3830 .elevator_allow_merge_fn = cfq_allow_merge,
3831 .elevator_dispatch_fn = cfq_dispatch_requests,
3832 .elevator_add_req_fn = cfq_insert_request,
3833 .elevator_activate_req_fn = cfq_activate_request,
3834 .elevator_deactivate_req_fn = cfq_deactivate_request,
3835 .elevator_queue_empty_fn = cfq_queue_empty,
3836 .elevator_completed_req_fn = cfq_completed_request,
3837 .elevator_former_req_fn = elv_rb_former_request,
3838 .elevator_latter_req_fn = elv_rb_latter_request,
3839 .elevator_set_req_fn = cfq_set_request,
3840 .elevator_put_req_fn = cfq_put_request,
3841 .elevator_may_queue_fn = cfq_may_queue,
3842 .elevator_init_fn = cfq_init_queue,
3843 .elevator_exit_fn = cfq_exit_queue,
3844 .trim = cfq_free_io_context,
3846 .elevator_attrs = cfq_attrs,
3847 .elevator_name = "cfq",
3848 .elevator_owner = THIS_MODULE,
3851 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3852 static struct blkio_policy_type blkio_policy_cfq = {
3854 .blkio_unlink_group_fn = cfq_unlink_blkio_group,
3855 .blkio_update_group_weight_fn = cfq_update_blkio_group_weight,
3859 static struct blkio_policy_type blkio_policy_cfq;
3862 static int __init cfq_init(void)
3865 * could be 0 on HZ < 1000 setups
3867 if (!cfq_slice_async)
3868 cfq_slice_async = 1;
3869 if (!cfq_slice_idle)
3872 if (cfq_slab_setup())
3875 elv_register(&iosched_cfq);
3876 blkio_policy_register(&blkio_policy_cfq);
3881 static void __exit cfq_exit(void)
3883 DECLARE_COMPLETION_ONSTACK(all_gone);
3884 blkio_policy_unregister(&blkio_policy_cfq);
3885 elv_unregister(&iosched_cfq);
3886 ioc_gone = &all_gone;
3887 /* ioc_gone's update must be visible before reading ioc_count */
3891 * this also protects us from entering cfq_slab_kill() with
3892 * pending RCU callbacks
3894 if (elv_ioc_count_read(cfq_ioc_count))
3895 wait_for_completion(&all_gone);
3899 module_init(cfq_init);
3900 module_exit(cfq_exit);
3902 MODULE_AUTHOR("Jens Axboe");
3903 MODULE_LICENSE("GPL");
3904 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");