2 * CFQ, or complete fairness queueing, disk scheduler.
4 * Based on ideas from a previously unfinished io
5 * scheduler (round robin per-process disk scheduling) and Andrea Arcangeli.
7 * Copyright (C) 2003 Jens Axboe <axboe@kernel.dk>
9 #include <linux/module.h>
10 #include <linux/blkdev.h>
11 #include <linux/elevator.h>
12 #include <linux/jiffies.h>
13 #include <linux/rbtree.h>
14 #include <linux/ioprio.h>
15 #include <linux/blktrace_api.h>
16 #include "blk-cgroup.h"
21 /* max queue in one round of service */
22 static const int cfq_quantum = 4;
23 static const int cfq_fifo_expire[2] = { HZ / 4, HZ / 8 };
24 /* maximum backwards seek, in KiB */
25 static const int cfq_back_max = 16 * 1024;
26 /* penalty of a backwards seek */
27 static const int cfq_back_penalty = 2;
28 static const int cfq_slice_sync = HZ / 10;
29 static int cfq_slice_async = HZ / 25;
30 static const int cfq_slice_async_rq = 2;
31 static int cfq_slice_idle = HZ / 125;
32 static const int cfq_target_latency = HZ * 3/10; /* 300 ms */
33 static const int cfq_hist_divisor = 4;
36 * offset from end of service tree
38 #define CFQ_IDLE_DELAY (HZ / 5)
41 * below this threshold, we consider thinktime immediate
43 #define CFQ_MIN_TT (2)
46 * Allow merged cfqqs to perform this amount of seeky I/O before
47 * deciding to break the queues up again.
49 #define CFQQ_COOP_TOUT (HZ)
51 #define CFQ_SLICE_SCALE (5)
52 #define CFQ_HW_QUEUE_MIN (5)
53 #define CFQ_SERVICE_SHIFT 12
56 ((struct cfq_io_context *) (rq)->elevator_private)
57 #define RQ_CFQQ(rq) (struct cfq_queue *) ((rq)->elevator_private2)
59 static struct kmem_cache *cfq_pool;
60 static struct kmem_cache *cfq_ioc_pool;
62 static DEFINE_PER_CPU(unsigned long, cfq_ioc_count);
63 static struct completion *ioc_gone;
64 static DEFINE_SPINLOCK(ioc_gone_lock);
66 #define CFQ_PRIO_LISTS IOPRIO_BE_NR
67 #define cfq_class_idle(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
68 #define cfq_class_rt(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
70 #define sample_valid(samples) ((samples) > 80)
71 #define rb_entry_cfqg(node) rb_entry((node), struct cfq_group, rb_node)
74 * Most of our rbtree usage is for sorting with min extraction, so
75 * if we cache the leftmost node we don't have to walk down the tree
76 * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should
77 * move this into the elevator for the rq sorting as well.
84 struct rb_node *active;
85 unsigned total_weight;
87 #define CFQ_RB_ROOT (struct cfq_rb_root) { RB_ROOT, NULL, 0, 0, }
90 * Per process-grouping structure
95 /* various state flags, see below */
98 struct cfq_data *cfqd;
99 /* service_tree member */
100 struct rb_node rb_node;
101 /* service_tree key */
102 unsigned long rb_key;
103 /* prio tree member */
104 struct rb_node p_node;
105 /* prio tree root we belong to, if any */
106 struct rb_root *p_root;
107 /* sorted list of pending requests */
108 struct rb_root sort_list;
109 /* if fifo isn't expired, next request to serve */
110 struct request *next_rq;
111 /* requests queued in sort_list */
113 /* currently allocated requests */
115 /* fifo list of requests in sort_list */
116 struct list_head fifo;
118 /* time when queue got scheduled in to dispatch first request. */
119 unsigned long dispatch_start;
120 unsigned int allocated_slice;
121 /* time when first request from queue completed and slice started. */
122 unsigned long slice_start;
123 unsigned long slice_end;
125 unsigned int slice_dispatch;
127 /* pending metadata requests */
129 /* number of requests that are on the dispatch list or inside driver */
132 /* io prio of this group */
133 unsigned short ioprio, org_ioprio;
134 unsigned short ioprio_class, org_ioprio_class;
136 unsigned int seek_samples;
139 sector_t last_request_pos;
140 unsigned long seeky_start;
144 struct cfq_rb_root *service_tree;
145 struct cfq_queue *new_cfqq;
146 struct cfq_group *cfqg;
147 struct cfq_group *orig_cfqg;
148 /* Sectors dispatched in current dispatch round */
149 unsigned long nr_sectors;
153 * First index in the service_trees.
154 * IDLE is handled separately, so it has negative index
163 * Second index in the service_trees.
167 SYNC_NOIDLE_WORKLOAD = 1,
171 /* This is per cgroup per device grouping structure */
173 /* group service_tree member */
174 struct rb_node rb_node;
176 /* group service_tree key */
181 /* number of cfqq currently on this group */
184 /* Per group busy queus average. Useful for workload slice calc. */
185 unsigned int busy_queues_avg[2];
187 * rr lists of queues with requests, onle rr for each priority class.
188 * Counts are embedded in the cfq_rb_root
190 struct cfq_rb_root service_trees[2][3];
191 struct cfq_rb_root service_tree_idle;
193 unsigned long saved_workload_slice;
194 enum wl_type_t saved_workload;
195 enum wl_prio_t saved_serving_prio;
196 struct blkio_group blkg;
197 #ifdef CONFIG_CFQ_GROUP_IOSCHED
198 struct hlist_node cfqd_node;
204 * Per block device queue structure
207 struct request_queue *queue;
208 /* Root service tree for cfq_groups */
209 struct cfq_rb_root grp_service_tree;
210 struct cfq_group root_group;
211 /* Number of active cfq groups on group service tree */
215 * The priority currently being served
217 enum wl_prio_t serving_prio;
218 enum wl_type_t serving_type;
219 unsigned long workload_expires;
220 struct cfq_group *serving_group;
221 bool noidle_tree_requires_idle;
224 * Each priority tree is sorted by next_request position. These
225 * trees are used when determining if two or more queues are
226 * interleaving requests (see cfq_close_cooperator).
228 struct rb_root prio_trees[CFQ_PRIO_LISTS];
230 unsigned int busy_queues;
236 * queue-depth detection
242 * -1 => indeterminate, (cfq will behave as if NCQ is present, to allow better detection)
243 * 1 => NCQ is present (hw_tag_est_depth is the estimated max depth)
246 int hw_tag_est_depth;
247 unsigned int hw_tag_samples;
250 * idle window management
252 struct timer_list idle_slice_timer;
253 struct work_struct unplug_work;
255 struct cfq_queue *active_queue;
256 struct cfq_io_context *active_cic;
259 * async queue for each priority case
261 struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
262 struct cfq_queue *async_idle_cfqq;
264 sector_t last_position;
267 * tunables, see top of file
269 unsigned int cfq_quantum;
270 unsigned int cfq_fifo_expire[2];
271 unsigned int cfq_back_penalty;
272 unsigned int cfq_back_max;
273 unsigned int cfq_slice[2];
274 unsigned int cfq_slice_async_rq;
275 unsigned int cfq_slice_idle;
276 unsigned int cfq_latency;
277 unsigned int cfq_group_isolation;
279 struct list_head cic_list;
282 * Fallback dummy cfqq for extreme OOM conditions
284 struct cfq_queue oom_cfqq;
286 unsigned long last_delayed_sync;
288 /* List of cfq groups being managed on this device*/
289 struct hlist_head cfqg_list;
293 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd);
295 static struct cfq_rb_root *service_tree_for(struct cfq_group *cfqg,
298 struct cfq_data *cfqd)
303 if (prio == IDLE_WORKLOAD)
304 return &cfqg->service_tree_idle;
306 return &cfqg->service_trees[prio][type];
309 enum cfqq_state_flags {
310 CFQ_CFQQ_FLAG_on_rr = 0, /* on round-robin busy list */
311 CFQ_CFQQ_FLAG_wait_request, /* waiting for a request */
312 CFQ_CFQQ_FLAG_must_dispatch, /* must be allowed a dispatch */
313 CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
314 CFQ_CFQQ_FLAG_fifo_expire, /* FIFO checked in this slice */
315 CFQ_CFQQ_FLAG_idle_window, /* slice idling enabled */
316 CFQ_CFQQ_FLAG_prio_changed, /* task priority has changed */
317 CFQ_CFQQ_FLAG_slice_new, /* no requests dispatched in slice */
318 CFQ_CFQQ_FLAG_sync, /* synchronous queue */
319 CFQ_CFQQ_FLAG_coop, /* cfqq is shared */
320 CFQ_CFQQ_FLAG_deep, /* sync cfqq experienced large depth */
321 CFQ_CFQQ_FLAG_wait_busy, /* Waiting for next request */
324 #define CFQ_CFQQ_FNS(name) \
325 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \
327 (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name); \
329 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \
331 (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \
333 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \
335 return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \
339 CFQ_CFQQ_FNS(wait_request);
340 CFQ_CFQQ_FNS(must_dispatch);
341 CFQ_CFQQ_FNS(must_alloc_slice);
342 CFQ_CFQQ_FNS(fifo_expire);
343 CFQ_CFQQ_FNS(idle_window);
344 CFQ_CFQQ_FNS(prio_changed);
345 CFQ_CFQQ_FNS(slice_new);
349 CFQ_CFQQ_FNS(wait_busy);
352 #ifdef CONFIG_DEBUG_CFQ_IOSCHED
353 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
354 blk_add_trace_msg((cfqd)->queue, "cfq%d%c %s " fmt, (cfqq)->pid, \
355 cfq_cfqq_sync((cfqq)) ? 'S' : 'A', \
356 blkg_path(&(cfqq)->cfqg->blkg), ##args);
358 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...) \
359 blk_add_trace_msg((cfqd)->queue, "%s " fmt, \
360 blkg_path(&(cfqg)->blkg), ##args); \
363 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
364 blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args)
365 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...) do {} while (0);
367 #define cfq_log(cfqd, fmt, args...) \
368 blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
370 /* Traverses through cfq group service trees */
371 #define for_each_cfqg_st(cfqg, i, j, st) \
372 for (i = 0; i <= IDLE_WORKLOAD; i++) \
373 for (j = 0, st = i < IDLE_WORKLOAD ? &cfqg->service_trees[i][j]\
374 : &cfqg->service_tree_idle; \
375 (i < IDLE_WORKLOAD && j <= SYNC_WORKLOAD) || \
376 (i == IDLE_WORKLOAD && j == 0); \
377 j++, st = i < IDLE_WORKLOAD ? \
378 &cfqg->service_trees[i][j]: NULL) \
381 static inline enum wl_prio_t cfqq_prio(struct cfq_queue *cfqq)
383 if (cfq_class_idle(cfqq))
384 return IDLE_WORKLOAD;
385 if (cfq_class_rt(cfqq))
391 static enum wl_type_t cfqq_type(struct cfq_queue *cfqq)
393 if (!cfq_cfqq_sync(cfqq))
394 return ASYNC_WORKLOAD;
395 if (!cfq_cfqq_idle_window(cfqq))
396 return SYNC_NOIDLE_WORKLOAD;
397 return SYNC_WORKLOAD;
400 static inline int cfq_group_busy_queues_wl(enum wl_prio_t wl,
401 struct cfq_data *cfqd,
402 struct cfq_group *cfqg)
404 if (wl == IDLE_WORKLOAD)
405 return cfqg->service_tree_idle.count;
407 return cfqg->service_trees[wl][ASYNC_WORKLOAD].count
408 + cfqg->service_trees[wl][SYNC_NOIDLE_WORKLOAD].count
409 + cfqg->service_trees[wl][SYNC_WORKLOAD].count;
412 static inline int cfqg_busy_async_queues(struct cfq_data *cfqd,
413 struct cfq_group *cfqg)
415 return cfqg->service_trees[RT_WORKLOAD][ASYNC_WORKLOAD].count
416 + cfqg->service_trees[BE_WORKLOAD][ASYNC_WORKLOAD].count;
419 static void cfq_dispatch_insert(struct request_queue *, struct request *);
420 static struct cfq_queue *cfq_get_queue(struct cfq_data *, bool,
421 struct io_context *, gfp_t);
422 static struct cfq_io_context *cfq_cic_lookup(struct cfq_data *,
423 struct io_context *);
425 static inline int rq_in_driver(struct cfq_data *cfqd)
427 return cfqd->rq_in_driver[0] + cfqd->rq_in_driver[1];
430 static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_context *cic,
433 return cic->cfqq[is_sync];
436 static inline void cic_set_cfqq(struct cfq_io_context *cic,
437 struct cfq_queue *cfqq, bool is_sync)
439 cic->cfqq[is_sync] = cfqq;
443 * We regard a request as SYNC, if it's either a read or has the SYNC bit
444 * set (in which case it could also be direct WRITE).
446 static inline bool cfq_bio_sync(struct bio *bio)
448 return bio_data_dir(bio) == READ || bio_rw_flagged(bio, BIO_RW_SYNCIO);
452 * scheduler run of queue, if there are requests pending and no one in the
453 * driver that will restart queueing
455 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
457 if (cfqd->busy_queues) {
458 cfq_log(cfqd, "schedule dispatch");
459 kblockd_schedule_work(cfqd->queue, &cfqd->unplug_work);
463 static int cfq_queue_empty(struct request_queue *q)
465 struct cfq_data *cfqd = q->elevator->elevator_data;
467 return !cfqd->rq_queued;
471 * Scale schedule slice based on io priority. Use the sync time slice only
472 * if a queue is marked sync and has sync io queued. A sync queue with async
473 * io only, should not get full sync slice length.
475 static inline int cfq_prio_slice(struct cfq_data *cfqd, bool sync,
478 const int base_slice = cfqd->cfq_slice[sync];
480 WARN_ON(prio >= IOPRIO_BE_NR);
482 return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
486 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
488 return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
491 static inline u64 cfq_scale_slice(unsigned long delta, struct cfq_group *cfqg)
493 u64 d = delta << CFQ_SERVICE_SHIFT;
495 d = d * BLKIO_WEIGHT_DEFAULT;
496 do_div(d, cfqg->weight);
500 static inline u64 max_vdisktime(u64 min_vdisktime, u64 vdisktime)
502 s64 delta = (s64)(vdisktime - min_vdisktime);
504 min_vdisktime = vdisktime;
506 return min_vdisktime;
509 static inline u64 min_vdisktime(u64 min_vdisktime, u64 vdisktime)
511 s64 delta = (s64)(vdisktime - min_vdisktime);
513 min_vdisktime = vdisktime;
515 return min_vdisktime;
518 static void update_min_vdisktime(struct cfq_rb_root *st)
520 u64 vdisktime = st->min_vdisktime;
521 struct cfq_group *cfqg;
524 cfqg = rb_entry_cfqg(st->active);
525 vdisktime = cfqg->vdisktime;
529 cfqg = rb_entry_cfqg(st->left);
530 vdisktime = min_vdisktime(vdisktime, cfqg->vdisktime);
533 st->min_vdisktime = max_vdisktime(st->min_vdisktime, vdisktime);
537 * get averaged number of queues of RT/BE priority.
538 * average is updated, with a formula that gives more weight to higher numbers,
539 * to quickly follows sudden increases and decrease slowly
542 static inline unsigned cfq_group_get_avg_queues(struct cfq_data *cfqd,
543 struct cfq_group *cfqg, bool rt)
545 unsigned min_q, max_q;
546 unsigned mult = cfq_hist_divisor - 1;
547 unsigned round = cfq_hist_divisor / 2;
548 unsigned busy = cfq_group_busy_queues_wl(rt, cfqd, cfqg);
550 min_q = min(cfqg->busy_queues_avg[rt], busy);
551 max_q = max(cfqg->busy_queues_avg[rt], busy);
552 cfqg->busy_queues_avg[rt] = (mult * max_q + min_q + round) /
554 return cfqg->busy_queues_avg[rt];
557 static inline unsigned
558 cfq_group_slice(struct cfq_data *cfqd, struct cfq_group *cfqg)
560 struct cfq_rb_root *st = &cfqd->grp_service_tree;
562 return cfq_target_latency * cfqg->weight / st->total_weight;
566 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
568 unsigned slice = cfq_prio_to_slice(cfqd, cfqq);
569 if (cfqd->cfq_latency) {
571 * interested queues (we consider only the ones with the same
572 * priority class in the cfq group)
574 unsigned iq = cfq_group_get_avg_queues(cfqd, cfqq->cfqg,
576 unsigned sync_slice = cfqd->cfq_slice[1];
577 unsigned expect_latency = sync_slice * iq;
578 unsigned group_slice = cfq_group_slice(cfqd, cfqq->cfqg);
580 if (expect_latency > group_slice) {
581 unsigned base_low_slice = 2 * cfqd->cfq_slice_idle;
582 /* scale low_slice according to IO priority
583 * and sync vs async */
585 min(slice, base_low_slice * slice / sync_slice);
586 /* the adapted slice value is scaled to fit all iqs
587 * into the target latency */
588 slice = max(slice * group_slice / expect_latency,
592 cfqq->slice_start = jiffies;
593 cfqq->slice_end = jiffies + slice;
594 cfqq->allocated_slice = slice;
595 cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
599 * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
600 * isn't valid until the first request from the dispatch is activated
601 * and the slice time set.
603 static inline bool cfq_slice_used(struct cfq_queue *cfqq)
605 if (cfq_cfqq_slice_new(cfqq))
607 if (time_before(jiffies, cfqq->slice_end))
614 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
615 * We choose the request that is closest to the head right now. Distance
616 * behind the head is penalized and only allowed to a certain extent.
618 static struct request *
619 cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2, sector_t last)
621 sector_t s1, s2, d1 = 0, d2 = 0;
622 unsigned long back_max;
623 #define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */
624 #define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */
625 unsigned wrap = 0; /* bit mask: requests behind the disk head? */
627 if (rq1 == NULL || rq1 == rq2)
632 if (rq_is_sync(rq1) && !rq_is_sync(rq2))
634 else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
636 if (rq_is_meta(rq1) && !rq_is_meta(rq2))
638 else if (rq_is_meta(rq2) && !rq_is_meta(rq1))
641 s1 = blk_rq_pos(rq1);
642 s2 = blk_rq_pos(rq2);
645 * by definition, 1KiB is 2 sectors
647 back_max = cfqd->cfq_back_max * 2;
650 * Strict one way elevator _except_ in the case where we allow
651 * short backward seeks which are biased as twice the cost of a
652 * similar forward seek.
656 else if (s1 + back_max >= last)
657 d1 = (last - s1) * cfqd->cfq_back_penalty;
659 wrap |= CFQ_RQ1_WRAP;
663 else if (s2 + back_max >= last)
664 d2 = (last - s2) * cfqd->cfq_back_penalty;
666 wrap |= CFQ_RQ2_WRAP;
668 /* Found required data */
671 * By doing switch() on the bit mask "wrap" we avoid having to
672 * check two variables for all permutations: --> faster!
675 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
691 case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
694 * Since both rqs are wrapped,
695 * start with the one that's further behind head
696 * (--> only *one* back seek required),
697 * since back seek takes more time than forward.
707 * The below is leftmost cache rbtree addon
709 static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
711 /* Service tree is empty */
716 root->left = rb_first(&root->rb);
719 return rb_entry(root->left, struct cfq_queue, rb_node);
724 static struct cfq_group *cfq_rb_first_group(struct cfq_rb_root *root)
727 root->left = rb_first(&root->rb);
730 return rb_entry_cfqg(root->left);
735 static void rb_erase_init(struct rb_node *n, struct rb_root *root)
741 static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
745 rb_erase_init(n, &root->rb);
750 * would be nice to take fifo expire time into account as well
752 static struct request *
753 cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
754 struct request *last)
756 struct rb_node *rbnext = rb_next(&last->rb_node);
757 struct rb_node *rbprev = rb_prev(&last->rb_node);
758 struct request *next = NULL, *prev = NULL;
760 BUG_ON(RB_EMPTY_NODE(&last->rb_node));
763 prev = rb_entry_rq(rbprev);
766 next = rb_entry_rq(rbnext);
768 rbnext = rb_first(&cfqq->sort_list);
769 if (rbnext && rbnext != &last->rb_node)
770 next = rb_entry_rq(rbnext);
773 return cfq_choose_req(cfqd, next, prev, blk_rq_pos(last));
776 static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
777 struct cfq_queue *cfqq)
780 * just an approximation, should be ok.
782 return (cfqq->cfqg->nr_cfqq - 1) * (cfq_prio_slice(cfqd, 1, 0) -
783 cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
787 cfqg_key(struct cfq_rb_root *st, struct cfq_group *cfqg)
789 return cfqg->vdisktime - st->min_vdisktime;
793 __cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
795 struct rb_node **node = &st->rb.rb_node;
796 struct rb_node *parent = NULL;
797 struct cfq_group *__cfqg;
798 s64 key = cfqg_key(st, cfqg);
801 while (*node != NULL) {
803 __cfqg = rb_entry_cfqg(parent);
805 if (key < cfqg_key(st, __cfqg))
806 node = &parent->rb_left;
808 node = &parent->rb_right;
814 st->left = &cfqg->rb_node;
816 rb_link_node(&cfqg->rb_node, parent, node);
817 rb_insert_color(&cfqg->rb_node, &st->rb);
821 cfq_group_service_tree_add(struct cfq_data *cfqd, struct cfq_group *cfqg)
823 struct cfq_rb_root *st = &cfqd->grp_service_tree;
824 struct cfq_group *__cfqg;
832 * Currently put the group at the end. Later implement something
833 * so that groups get lesser vtime based on their weights, so that
834 * if group does not loose all if it was not continously backlogged.
836 n = rb_last(&st->rb);
838 __cfqg = rb_entry_cfqg(n);
839 cfqg->vdisktime = __cfqg->vdisktime + CFQ_IDLE_DELAY;
841 cfqg->vdisktime = st->min_vdisktime;
843 __cfq_group_service_tree_add(st, cfqg);
846 st->total_weight += cfqg->weight;
850 cfq_group_service_tree_del(struct cfq_data *cfqd, struct cfq_group *cfqg)
852 struct cfq_rb_root *st = &cfqd->grp_service_tree;
854 if (st->active == &cfqg->rb_node)
857 BUG_ON(cfqg->nr_cfqq < 1);
860 /* If there are other cfq queues under this group, don't delete it */
864 cfq_log_cfqg(cfqd, cfqg, "del_from_rr group");
867 st->total_weight -= cfqg->weight;
868 if (!RB_EMPTY_NODE(&cfqg->rb_node))
869 cfq_rb_erase(&cfqg->rb_node, st);
870 cfqg->saved_workload_slice = 0;
871 blkiocg_update_blkio_group_dequeue_stats(&cfqg->blkg, 1);
874 static inline unsigned int cfq_cfqq_slice_usage(struct cfq_queue *cfqq)
876 unsigned int slice_used;
879 * Queue got expired before even a single request completed or
880 * got expired immediately after first request completion.
882 if (!cfqq->slice_start || cfqq->slice_start == jiffies) {
884 * Also charge the seek time incurred to the group, otherwise
885 * if there are mutiple queues in the group, each can dispatch
886 * a single request on seeky media and cause lots of seek time
887 * and group will never know it.
889 slice_used = max_t(unsigned, (jiffies - cfqq->dispatch_start),
892 slice_used = jiffies - cfqq->slice_start;
893 if (slice_used > cfqq->allocated_slice)
894 slice_used = cfqq->allocated_slice;
897 cfq_log_cfqq(cfqq->cfqd, cfqq, "sl_used=%u sect=%lu", slice_used,
902 static void cfq_group_served(struct cfq_data *cfqd, struct cfq_group *cfqg,
903 struct cfq_queue *cfqq)
905 struct cfq_rb_root *st = &cfqd->grp_service_tree;
906 unsigned int used_sl, charge_sl;
907 int nr_sync = cfqg->nr_cfqq - cfqg_busy_async_queues(cfqd, cfqg)
908 - cfqg->service_tree_idle.count;
911 used_sl = charge_sl = cfq_cfqq_slice_usage(cfqq);
913 if (!cfq_cfqq_sync(cfqq) && !nr_sync)
914 charge_sl = cfqq->allocated_slice;
916 /* Can't update vdisktime while group is on service tree */
917 cfq_rb_erase(&cfqg->rb_node, st);
918 cfqg->vdisktime += cfq_scale_slice(charge_sl, cfqg);
919 __cfq_group_service_tree_add(st, cfqg);
921 /* This group is being expired. Save the context */
922 if (time_after(cfqd->workload_expires, jiffies)) {
923 cfqg->saved_workload_slice = cfqd->workload_expires
925 cfqg->saved_workload = cfqd->serving_type;
926 cfqg->saved_serving_prio = cfqd->serving_prio;
928 cfqg->saved_workload_slice = 0;
930 cfq_log_cfqg(cfqd, cfqg, "served: vt=%llu min_vt=%llu", cfqg->vdisktime,
932 blkiocg_update_blkio_group_stats(&cfqg->blkg, used_sl,
936 #ifdef CONFIG_CFQ_GROUP_IOSCHED
937 static inline struct cfq_group *cfqg_of_blkg(struct blkio_group *blkg)
940 return container_of(blkg, struct cfq_group, blkg);
945 cfq_update_blkio_group_weight(struct blkio_group *blkg, unsigned int weight)
947 cfqg_of_blkg(blkg)->weight = weight;
950 static struct cfq_group *
951 cfq_find_alloc_cfqg(struct cfq_data *cfqd, struct cgroup *cgroup, int create)
953 struct blkio_cgroup *blkcg = cgroup_to_blkio_cgroup(cgroup);
954 struct cfq_group *cfqg = NULL;
957 struct cfq_rb_root *st;
958 struct backing_dev_info *bdi = &cfqd->queue->backing_dev_info;
959 unsigned int major, minor;
961 /* Do we need to take this reference */
962 if (!blkiocg_css_tryget(blkcg))
965 cfqg = cfqg_of_blkg(blkiocg_lookup_group(blkcg, key));
969 cfqg = kzalloc_node(sizeof(*cfqg), GFP_ATOMIC, cfqd->queue->node);
973 cfqg->weight = blkcg->weight;
974 for_each_cfqg_st(cfqg, i, j, st)
976 RB_CLEAR_NODE(&cfqg->rb_node);
979 * Take the initial reference that will be released on destroy
980 * This can be thought of a joint reference by cgroup and
981 * elevator which will be dropped by either elevator exit
982 * or cgroup deletion path depending on who is exiting first.
984 atomic_set(&cfqg->ref, 1);
986 /* Add group onto cgroup list */
987 sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
988 blkiocg_add_blkio_group(blkcg, &cfqg->blkg, (void *)cfqd,
989 MKDEV(major, minor));
991 /* Add group on cfqd list */
992 hlist_add_head(&cfqg->cfqd_node, &cfqd->cfqg_list);
995 blkiocg_css_put(blkcg);
1000 * Search for the cfq group current task belongs to. If create = 1, then also
1001 * create the cfq group if it does not exist. request_queue lock must be held.
1003 static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd, int create)
1005 struct cgroup *cgroup;
1006 struct cfq_group *cfqg = NULL;
1009 cgroup = task_cgroup(current, blkio_subsys_id);
1010 cfqg = cfq_find_alloc_cfqg(cfqd, cgroup, create);
1011 if (!cfqg && create)
1012 cfqg = &cfqd->root_group;
1017 static void cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg)
1019 /* Currently, all async queues are mapped to root group */
1020 if (!cfq_cfqq_sync(cfqq))
1021 cfqg = &cfqq->cfqd->root_group;
1024 /* cfqq reference on cfqg */
1025 atomic_inc(&cfqq->cfqg->ref);
1028 static void cfq_put_cfqg(struct cfq_group *cfqg)
1030 struct cfq_rb_root *st;
1033 BUG_ON(atomic_read(&cfqg->ref) <= 0);
1034 if (!atomic_dec_and_test(&cfqg->ref))
1036 for_each_cfqg_st(cfqg, i, j, st)
1037 BUG_ON(!RB_EMPTY_ROOT(&st->rb) || st->active != NULL);
1041 static void cfq_destroy_cfqg(struct cfq_data *cfqd, struct cfq_group *cfqg)
1043 /* Something wrong if we are trying to remove same group twice */
1044 BUG_ON(hlist_unhashed(&cfqg->cfqd_node));
1046 hlist_del_init(&cfqg->cfqd_node);
1049 * Put the reference taken at the time of creation so that when all
1050 * queues are gone, group can be destroyed.
1055 static void cfq_release_cfq_groups(struct cfq_data *cfqd)
1057 struct hlist_node *pos, *n;
1058 struct cfq_group *cfqg;
1060 hlist_for_each_entry_safe(cfqg, pos, n, &cfqd->cfqg_list, cfqd_node) {
1062 * If cgroup removal path got to blk_group first and removed
1063 * it from cgroup list, then it will take care of destroying
1066 if (!blkiocg_del_blkio_group(&cfqg->blkg))
1067 cfq_destroy_cfqg(cfqd, cfqg);
1072 * Blk cgroup controller notification saying that blkio_group object is being
1073 * delinked as associated cgroup object is going away. That also means that
1074 * no new IO will come in this group. So get rid of this group as soon as
1075 * any pending IO in the group is finished.
1077 * This function is called under rcu_read_lock(). key is the rcu protected
1078 * pointer. That means "key" is a valid cfq_data pointer as long as we are rcu
1081 * "key" was fetched from blkio_group under blkio_cgroup->lock. That means
1082 * it should not be NULL as even if elevator was exiting, cgroup deltion
1083 * path got to it first.
1085 void cfq_unlink_blkio_group(void *key, struct blkio_group *blkg)
1087 unsigned long flags;
1088 struct cfq_data *cfqd = key;
1090 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1091 cfq_destroy_cfqg(cfqd, cfqg_of_blkg(blkg));
1092 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1095 #else /* GROUP_IOSCHED */
1096 static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd, int create)
1098 return &cfqd->root_group;
1101 cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) {
1105 static void cfq_release_cfq_groups(struct cfq_data *cfqd) {}
1106 static inline void cfq_put_cfqg(struct cfq_group *cfqg) {}
1108 #endif /* GROUP_IOSCHED */
1111 * The cfqd->service_trees holds all pending cfq_queue's that have
1112 * requests waiting to be processed. It is sorted in the order that
1113 * we will service the queues.
1115 static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1118 struct rb_node **p, *parent;
1119 struct cfq_queue *__cfqq;
1120 unsigned long rb_key;
1121 struct cfq_rb_root *service_tree;
1124 int group_changed = 0;
1126 #ifdef CONFIG_CFQ_GROUP_IOSCHED
1127 if (!cfqd->cfq_group_isolation
1128 && cfqq_type(cfqq) == SYNC_NOIDLE_WORKLOAD
1129 && cfqq->cfqg && cfqq->cfqg != &cfqd->root_group) {
1130 /* Move this cfq to root group */
1131 cfq_log_cfqq(cfqd, cfqq, "moving to root group");
1132 if (!RB_EMPTY_NODE(&cfqq->rb_node))
1133 cfq_group_service_tree_del(cfqd, cfqq->cfqg);
1134 cfqq->orig_cfqg = cfqq->cfqg;
1135 cfqq->cfqg = &cfqd->root_group;
1136 atomic_inc(&cfqd->root_group.ref);
1138 } else if (!cfqd->cfq_group_isolation
1139 && cfqq_type(cfqq) == SYNC_WORKLOAD && cfqq->orig_cfqg) {
1140 /* cfqq is sequential now needs to go to its original group */
1141 BUG_ON(cfqq->cfqg != &cfqd->root_group);
1142 if (!RB_EMPTY_NODE(&cfqq->rb_node))
1143 cfq_group_service_tree_del(cfqd, cfqq->cfqg);
1144 cfq_put_cfqg(cfqq->cfqg);
1145 cfqq->cfqg = cfqq->orig_cfqg;
1146 cfqq->orig_cfqg = NULL;
1148 cfq_log_cfqq(cfqd, cfqq, "moved to origin group");
1152 service_tree = service_tree_for(cfqq->cfqg, cfqq_prio(cfqq),
1153 cfqq_type(cfqq), cfqd);
1154 if (cfq_class_idle(cfqq)) {
1155 rb_key = CFQ_IDLE_DELAY;
1156 parent = rb_last(&service_tree->rb);
1157 if (parent && parent != &cfqq->rb_node) {
1158 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1159 rb_key += __cfqq->rb_key;
1162 } else if (!add_front) {
1164 * Get our rb key offset. Subtract any residual slice
1165 * value carried from last service. A negative resid
1166 * count indicates slice overrun, and this should position
1167 * the next service time further away in the tree.
1169 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
1170 rb_key -= cfqq->slice_resid;
1171 cfqq->slice_resid = 0;
1174 __cfqq = cfq_rb_first(service_tree);
1175 rb_key += __cfqq ? __cfqq->rb_key : jiffies;
1178 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1181 * same position, nothing more to do
1183 if (rb_key == cfqq->rb_key &&
1184 cfqq->service_tree == service_tree)
1187 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1188 cfqq->service_tree = NULL;
1193 cfqq->service_tree = service_tree;
1194 p = &service_tree->rb.rb_node;
1199 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1202 * sort by key, that represents service time.
1204 if (time_before(rb_key, __cfqq->rb_key))
1207 n = &(*p)->rb_right;
1215 service_tree->left = &cfqq->rb_node;
1217 cfqq->rb_key = rb_key;
1218 rb_link_node(&cfqq->rb_node, parent, p);
1219 rb_insert_color(&cfqq->rb_node, &service_tree->rb);
1220 service_tree->count++;
1221 if ((add_front || !new_cfqq) && !group_changed)
1223 cfq_group_service_tree_add(cfqd, cfqq->cfqg);
1226 static struct cfq_queue *
1227 cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
1228 sector_t sector, struct rb_node **ret_parent,
1229 struct rb_node ***rb_link)
1231 struct rb_node **p, *parent;
1232 struct cfq_queue *cfqq = NULL;
1240 cfqq = rb_entry(parent, struct cfq_queue, p_node);
1243 * Sort strictly based on sector. Smallest to the left,
1244 * largest to the right.
1246 if (sector > blk_rq_pos(cfqq->next_rq))
1247 n = &(*p)->rb_right;
1248 else if (sector < blk_rq_pos(cfqq->next_rq))
1256 *ret_parent = parent;
1262 static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1264 struct rb_node **p, *parent;
1265 struct cfq_queue *__cfqq;
1268 rb_erase(&cfqq->p_node, cfqq->p_root);
1269 cfqq->p_root = NULL;
1272 if (cfq_class_idle(cfqq))
1277 cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
1278 __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
1279 blk_rq_pos(cfqq->next_rq), &parent, &p);
1281 rb_link_node(&cfqq->p_node, parent, p);
1282 rb_insert_color(&cfqq->p_node, cfqq->p_root);
1284 cfqq->p_root = NULL;
1288 * Update cfqq's position in the service tree.
1290 static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1293 * Resorting requires the cfqq to be on the RR list already.
1295 if (cfq_cfqq_on_rr(cfqq)) {
1296 cfq_service_tree_add(cfqd, cfqq, 0);
1297 cfq_prio_tree_add(cfqd, cfqq);
1302 * add to busy list of queues for service, trying to be fair in ordering
1303 * the pending list according to last request service
1305 static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1307 cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
1308 BUG_ON(cfq_cfqq_on_rr(cfqq));
1309 cfq_mark_cfqq_on_rr(cfqq);
1310 cfqd->busy_queues++;
1312 cfq_resort_rr_list(cfqd, cfqq);
1316 * Called when the cfqq no longer has requests pending, remove it from
1319 static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1321 cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
1322 BUG_ON(!cfq_cfqq_on_rr(cfqq));
1323 cfq_clear_cfqq_on_rr(cfqq);
1325 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1326 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1327 cfqq->service_tree = NULL;
1330 rb_erase(&cfqq->p_node, cfqq->p_root);
1331 cfqq->p_root = NULL;
1334 cfq_group_service_tree_del(cfqd, cfqq->cfqg);
1335 BUG_ON(!cfqd->busy_queues);
1336 cfqd->busy_queues--;
1340 * rb tree support functions
1342 static void cfq_del_rq_rb(struct request *rq)
1344 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1345 const int sync = rq_is_sync(rq);
1347 BUG_ON(!cfqq->queued[sync]);
1348 cfqq->queued[sync]--;
1350 elv_rb_del(&cfqq->sort_list, rq);
1352 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) {
1354 * Queue will be deleted from service tree when we actually
1355 * expire it later. Right now just remove it from prio tree
1359 rb_erase(&cfqq->p_node, cfqq->p_root);
1360 cfqq->p_root = NULL;
1365 static void cfq_add_rq_rb(struct request *rq)
1367 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1368 struct cfq_data *cfqd = cfqq->cfqd;
1369 struct request *__alias, *prev;
1371 cfqq->queued[rq_is_sync(rq)]++;
1374 * looks a little odd, but the first insert might return an alias.
1375 * if that happens, put the alias on the dispatch list
1377 while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL)
1378 cfq_dispatch_insert(cfqd->queue, __alias);
1380 if (!cfq_cfqq_on_rr(cfqq))
1381 cfq_add_cfqq_rr(cfqd, cfqq);
1384 * check if this request is a better next-serve candidate
1386 prev = cfqq->next_rq;
1387 cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq, cfqd->last_position);
1390 * adjust priority tree position, if ->next_rq changes
1392 if (prev != cfqq->next_rq)
1393 cfq_prio_tree_add(cfqd, cfqq);
1395 BUG_ON(!cfqq->next_rq);
1398 static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
1400 elv_rb_del(&cfqq->sort_list, rq);
1401 cfqq->queued[rq_is_sync(rq)]--;
1405 static struct request *
1406 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
1408 struct task_struct *tsk = current;
1409 struct cfq_io_context *cic;
1410 struct cfq_queue *cfqq;
1412 cic = cfq_cic_lookup(cfqd, tsk->io_context);
1416 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
1418 sector_t sector = bio->bi_sector + bio_sectors(bio);
1420 return elv_rb_find(&cfqq->sort_list, sector);
1426 static void cfq_activate_request(struct request_queue *q, struct request *rq)
1428 struct cfq_data *cfqd = q->elevator->elevator_data;
1430 cfqd->rq_in_driver[rq_is_sync(rq)]++;
1431 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
1432 rq_in_driver(cfqd));
1434 cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
1437 static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
1439 struct cfq_data *cfqd = q->elevator->elevator_data;
1440 const int sync = rq_is_sync(rq);
1442 WARN_ON(!cfqd->rq_in_driver[sync]);
1443 cfqd->rq_in_driver[sync]--;
1444 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
1445 rq_in_driver(cfqd));
1448 static void cfq_remove_request(struct request *rq)
1450 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1452 if (cfqq->next_rq == rq)
1453 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
1455 list_del_init(&rq->queuelist);
1458 cfqq->cfqd->rq_queued--;
1459 if (rq_is_meta(rq)) {
1460 WARN_ON(!cfqq->meta_pending);
1461 cfqq->meta_pending--;
1465 static int cfq_merge(struct request_queue *q, struct request **req,
1468 struct cfq_data *cfqd = q->elevator->elevator_data;
1469 struct request *__rq;
1471 __rq = cfq_find_rq_fmerge(cfqd, bio);
1472 if (__rq && elv_rq_merge_ok(__rq, bio)) {
1474 return ELEVATOR_FRONT_MERGE;
1477 return ELEVATOR_NO_MERGE;
1480 static void cfq_merged_request(struct request_queue *q, struct request *req,
1483 if (type == ELEVATOR_FRONT_MERGE) {
1484 struct cfq_queue *cfqq = RQ_CFQQ(req);
1486 cfq_reposition_rq_rb(cfqq, req);
1491 cfq_merged_requests(struct request_queue *q, struct request *rq,
1492 struct request *next)
1494 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1496 * reposition in fifo if next is older than rq
1498 if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
1499 time_before(rq_fifo_time(next), rq_fifo_time(rq))) {
1500 list_move(&rq->queuelist, &next->queuelist);
1501 rq_set_fifo_time(rq, rq_fifo_time(next));
1504 if (cfqq->next_rq == next)
1506 cfq_remove_request(next);
1509 static int cfq_allow_merge(struct request_queue *q, struct request *rq,
1512 struct cfq_data *cfqd = q->elevator->elevator_data;
1513 struct cfq_io_context *cic;
1514 struct cfq_queue *cfqq;
1516 /* Deny merge if bio and rq don't belong to same cfq group */
1517 if ((RQ_CFQQ(rq))->cfqg != cfq_get_cfqg(cfqd, 0))
1520 * Disallow merge of a sync bio into an async request.
1522 if (cfq_bio_sync(bio) && !rq_is_sync(rq))
1526 * Lookup the cfqq that this bio will be queued with. Allow
1527 * merge only if rq is queued there.
1529 cic = cfq_cic_lookup(cfqd, current->io_context);
1533 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
1534 return cfqq == RQ_CFQQ(rq);
1537 static void __cfq_set_active_queue(struct cfq_data *cfqd,
1538 struct cfq_queue *cfqq)
1541 cfq_log_cfqq(cfqd, cfqq, "set_active");
1542 cfqq->slice_start = 0;
1543 cfqq->dispatch_start = jiffies;
1544 cfqq->allocated_slice = 0;
1545 cfqq->slice_end = 0;
1546 cfqq->slice_dispatch = 0;
1547 cfqq->nr_sectors = 0;
1549 cfq_clear_cfqq_wait_request(cfqq);
1550 cfq_clear_cfqq_must_dispatch(cfqq);
1551 cfq_clear_cfqq_must_alloc_slice(cfqq);
1552 cfq_clear_cfqq_fifo_expire(cfqq);
1553 cfq_mark_cfqq_slice_new(cfqq);
1555 del_timer(&cfqd->idle_slice_timer);
1558 cfqd->active_queue = cfqq;
1562 * current cfqq expired its slice (or was too idle), select new one
1565 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1568 cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
1570 if (cfq_cfqq_wait_request(cfqq))
1571 del_timer(&cfqd->idle_slice_timer);
1573 cfq_clear_cfqq_wait_request(cfqq);
1574 cfq_clear_cfqq_wait_busy(cfqq);
1577 * store what was left of this slice, if the queue idled/timed out
1579 if (timed_out && !cfq_cfqq_slice_new(cfqq)) {
1580 cfqq->slice_resid = cfqq->slice_end - jiffies;
1581 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
1584 cfq_group_served(cfqd, cfqq->cfqg, cfqq);
1586 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
1587 cfq_del_cfqq_rr(cfqd, cfqq);
1589 cfq_resort_rr_list(cfqd, cfqq);
1591 if (cfqq == cfqd->active_queue)
1592 cfqd->active_queue = NULL;
1594 if (&cfqq->cfqg->rb_node == cfqd->grp_service_tree.active)
1595 cfqd->grp_service_tree.active = NULL;
1597 if (cfqd->active_cic) {
1598 put_io_context(cfqd->active_cic->ioc);
1599 cfqd->active_cic = NULL;
1603 static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out)
1605 struct cfq_queue *cfqq = cfqd->active_queue;
1608 __cfq_slice_expired(cfqd, cfqq, timed_out);
1612 * Get next queue for service. Unless we have a queue preemption,
1613 * we'll simply select the first cfqq in the service tree.
1615 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
1617 struct cfq_rb_root *service_tree =
1618 service_tree_for(cfqd->serving_group, cfqd->serving_prio,
1619 cfqd->serving_type, cfqd);
1621 if (!cfqd->rq_queued)
1624 /* There is nothing to dispatch */
1627 if (RB_EMPTY_ROOT(&service_tree->rb))
1629 return cfq_rb_first(service_tree);
1632 static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd)
1634 struct cfq_group *cfqg;
1635 struct cfq_queue *cfqq;
1637 struct cfq_rb_root *st;
1639 if (!cfqd->rq_queued)
1642 cfqg = cfq_get_next_cfqg(cfqd);
1646 for_each_cfqg_st(cfqg, i, j, st)
1647 if ((cfqq = cfq_rb_first(st)) != NULL)
1653 * Get and set a new active queue for service.
1655 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
1656 struct cfq_queue *cfqq)
1659 cfqq = cfq_get_next_queue(cfqd);
1661 __cfq_set_active_queue(cfqd, cfqq);
1665 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
1668 if (blk_rq_pos(rq) >= cfqd->last_position)
1669 return blk_rq_pos(rq) - cfqd->last_position;
1671 return cfqd->last_position - blk_rq_pos(rq);
1674 #define CFQQ_SEEK_THR 8 * 1024
1675 #define CFQQ_SEEKY(cfqq) ((cfqq)->seek_mean > CFQQ_SEEK_THR)
1677 static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1680 sector_t sdist = cfqq->seek_mean;
1682 if (!sample_valid(cfqq->seek_samples))
1683 sdist = CFQQ_SEEK_THR;
1685 return cfq_dist_from_last(cfqd, rq) <= sdist;
1688 static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
1689 struct cfq_queue *cur_cfqq)
1691 struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
1692 struct rb_node *parent, *node;
1693 struct cfq_queue *__cfqq;
1694 sector_t sector = cfqd->last_position;
1696 if (RB_EMPTY_ROOT(root))
1700 * First, if we find a request starting at the end of the last
1701 * request, choose it.
1703 __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
1708 * If the exact sector wasn't found, the parent of the NULL leaf
1709 * will contain the closest sector.
1711 __cfqq = rb_entry(parent, struct cfq_queue, p_node);
1712 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1715 if (blk_rq_pos(__cfqq->next_rq) < sector)
1716 node = rb_next(&__cfqq->p_node);
1718 node = rb_prev(&__cfqq->p_node);
1722 __cfqq = rb_entry(node, struct cfq_queue, p_node);
1723 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1731 * cur_cfqq - passed in so that we don't decide that the current queue is
1732 * closely cooperating with itself.
1734 * So, basically we're assuming that that cur_cfqq has dispatched at least
1735 * one request, and that cfqd->last_position reflects a position on the disk
1736 * associated with the I/O issued by cur_cfqq. I'm not sure this is a valid
1739 static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
1740 struct cfq_queue *cur_cfqq)
1742 struct cfq_queue *cfqq;
1744 if (!cfq_cfqq_sync(cur_cfqq))
1746 if (CFQQ_SEEKY(cur_cfqq))
1750 * Don't search priority tree if it's the only queue in the group.
1752 if (cur_cfqq->cfqg->nr_cfqq == 1)
1756 * We should notice if some of the queues are cooperating, eg
1757 * working closely on the same area of the disk. In that case,
1758 * we can group them together and don't waste time idling.
1760 cfqq = cfqq_close(cfqd, cur_cfqq);
1764 /* If new queue belongs to different cfq_group, don't choose it */
1765 if (cur_cfqq->cfqg != cfqq->cfqg)
1769 * It only makes sense to merge sync queues.
1771 if (!cfq_cfqq_sync(cfqq))
1773 if (CFQQ_SEEKY(cfqq))
1777 * Do not merge queues of different priority classes
1779 if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq))
1786 * Determine whether we should enforce idle window for this queue.
1789 static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1791 enum wl_prio_t prio = cfqq_prio(cfqq);
1792 struct cfq_rb_root *service_tree = cfqq->service_tree;
1794 BUG_ON(!service_tree);
1795 BUG_ON(!service_tree->count);
1797 /* We never do for idle class queues. */
1798 if (prio == IDLE_WORKLOAD)
1801 /* We do for queues that were marked with idle window flag. */
1802 if (cfq_cfqq_idle_window(cfqq) &&
1803 !(blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag))
1807 * Otherwise, we do only if they are the last ones
1808 * in their service tree.
1810 return service_tree->count == 1;
1813 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
1815 struct cfq_queue *cfqq = cfqd->active_queue;
1816 struct cfq_io_context *cic;
1820 * SSD device without seek penalty, disable idling. But only do so
1821 * for devices that support queuing, otherwise we still have a problem
1822 * with sync vs async workloads.
1824 if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
1827 WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
1828 WARN_ON(cfq_cfqq_slice_new(cfqq));
1831 * idle is disabled, either manually or by past process history
1833 if (!cfqd->cfq_slice_idle || !cfq_should_idle(cfqd, cfqq))
1837 * still active requests from this queue, don't idle
1839 if (cfqq->dispatched)
1843 * task has exited, don't wait
1845 cic = cfqd->active_cic;
1846 if (!cic || !atomic_read(&cic->ioc->nr_tasks))
1850 * If our average think time is larger than the remaining time
1851 * slice, then don't idle. This avoids overrunning the allotted
1854 if (sample_valid(cic->ttime_samples) &&
1855 (cfqq->slice_end - jiffies < cic->ttime_mean))
1858 cfq_mark_cfqq_wait_request(cfqq);
1860 sl = cfqd->cfq_slice_idle;
1862 mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
1863 cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu", sl);
1867 * Move request from internal lists to the request queue dispatch list.
1869 static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
1871 struct cfq_data *cfqd = q->elevator->elevator_data;
1872 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1874 cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
1876 cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
1877 cfq_remove_request(rq);
1879 elv_dispatch_sort(q, rq);
1881 if (cfq_cfqq_sync(cfqq))
1882 cfqd->sync_flight++;
1883 cfqq->nr_sectors += blk_rq_sectors(rq);
1887 * return expired entry, or NULL to just start from scratch in rbtree
1889 static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
1891 struct request *rq = NULL;
1893 if (cfq_cfqq_fifo_expire(cfqq))
1896 cfq_mark_cfqq_fifo_expire(cfqq);
1898 if (list_empty(&cfqq->fifo))
1901 rq = rq_entry_fifo(cfqq->fifo.next);
1902 if (time_before(jiffies, rq_fifo_time(rq)))
1905 cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq);
1910 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1912 const int base_rq = cfqd->cfq_slice_async_rq;
1914 WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
1916 return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
1920 * Must be called with the queue_lock held.
1922 static int cfqq_process_refs(struct cfq_queue *cfqq)
1924 int process_refs, io_refs;
1926 io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE];
1927 process_refs = atomic_read(&cfqq->ref) - io_refs;
1928 BUG_ON(process_refs < 0);
1929 return process_refs;
1932 static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq)
1934 int process_refs, new_process_refs;
1935 struct cfq_queue *__cfqq;
1937 /* Avoid a circular list and skip interim queue merges */
1938 while ((__cfqq = new_cfqq->new_cfqq)) {
1944 process_refs = cfqq_process_refs(cfqq);
1946 * If the process for the cfqq has gone away, there is no
1947 * sense in merging the queues.
1949 if (process_refs == 0)
1953 * Merge in the direction of the lesser amount of work.
1955 new_process_refs = cfqq_process_refs(new_cfqq);
1956 if (new_process_refs >= process_refs) {
1957 cfqq->new_cfqq = new_cfqq;
1958 atomic_add(process_refs, &new_cfqq->ref);
1960 new_cfqq->new_cfqq = cfqq;
1961 atomic_add(new_process_refs, &cfqq->ref);
1965 static enum wl_type_t cfq_choose_wl(struct cfq_data *cfqd,
1966 struct cfq_group *cfqg, enum wl_prio_t prio,
1969 struct cfq_queue *queue;
1971 bool key_valid = false;
1972 unsigned long lowest_key = 0;
1973 enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD;
1977 * When priorities switched, we prefer starting
1978 * from SYNC_NOIDLE (first choice), or just SYNC
1981 if (service_tree_for(cfqg, prio, cur_best, cfqd)->count)
1983 cur_best = SYNC_WORKLOAD;
1984 if (service_tree_for(cfqg, prio, cur_best, cfqd)->count)
1987 return ASYNC_WORKLOAD;
1990 for (i = 0; i < 3; ++i) {
1991 /* otherwise, select the one with lowest rb_key */
1992 queue = cfq_rb_first(service_tree_for(cfqg, prio, i, cfqd));
1994 (!key_valid || time_before(queue->rb_key, lowest_key))) {
1995 lowest_key = queue->rb_key;
2004 static void choose_service_tree(struct cfq_data *cfqd, struct cfq_group *cfqg)
2006 enum wl_prio_t previous_prio = cfqd->serving_prio;
2010 struct cfq_rb_root *st;
2011 unsigned group_slice;
2014 cfqd->serving_prio = IDLE_WORKLOAD;
2015 cfqd->workload_expires = jiffies + 1;
2019 /* Choose next priority. RT > BE > IDLE */
2020 if (cfq_group_busy_queues_wl(RT_WORKLOAD, cfqd, cfqg))
2021 cfqd->serving_prio = RT_WORKLOAD;
2022 else if (cfq_group_busy_queues_wl(BE_WORKLOAD, cfqd, cfqg))
2023 cfqd->serving_prio = BE_WORKLOAD;
2025 cfqd->serving_prio = IDLE_WORKLOAD;
2026 cfqd->workload_expires = jiffies + 1;
2031 * For RT and BE, we have to choose also the type
2032 * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
2035 prio_changed = (cfqd->serving_prio != previous_prio);
2036 st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type,
2041 * If priority didn't change, check workload expiration,
2042 * and that we still have other queues ready
2044 if (!prio_changed && count &&
2045 !time_after(jiffies, cfqd->workload_expires))
2048 /* otherwise select new workload type */
2049 cfqd->serving_type =
2050 cfq_choose_wl(cfqd, cfqg, cfqd->serving_prio, prio_changed);
2051 st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type,
2056 * the workload slice is computed as a fraction of target latency
2057 * proportional to the number of queues in that workload, over
2058 * all the queues in the same priority class
2060 group_slice = cfq_group_slice(cfqd, cfqg);
2062 slice = group_slice * count /
2063 max_t(unsigned, cfqg->busy_queues_avg[cfqd->serving_prio],
2064 cfq_group_busy_queues_wl(cfqd->serving_prio, cfqd, cfqg));
2066 if (cfqd->serving_type == ASYNC_WORKLOAD) {
2070 * Async queues are currently system wide. Just taking
2071 * proportion of queues with-in same group will lead to higher
2072 * async ratio system wide as generally root group is going
2073 * to have higher weight. A more accurate thing would be to
2074 * calculate system wide asnc/sync ratio.
2076 tmp = cfq_target_latency * cfqg_busy_async_queues(cfqd, cfqg);
2077 tmp = tmp/cfqd->busy_queues;
2078 slice = min_t(unsigned, slice, tmp);
2080 /* async workload slice is scaled down according to
2081 * the sync/async slice ratio. */
2082 slice = slice * cfqd->cfq_slice[0] / cfqd->cfq_slice[1];
2084 /* sync workload slice is at least 2 * cfq_slice_idle */
2085 slice = max(slice, 2 * cfqd->cfq_slice_idle);
2087 slice = max_t(unsigned, slice, CFQ_MIN_TT);
2088 cfqd->workload_expires = jiffies + slice;
2089 cfqd->noidle_tree_requires_idle = false;
2092 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd)
2094 struct cfq_rb_root *st = &cfqd->grp_service_tree;
2095 struct cfq_group *cfqg;
2097 if (RB_EMPTY_ROOT(&st->rb))
2099 cfqg = cfq_rb_first_group(st);
2100 st->active = &cfqg->rb_node;
2101 update_min_vdisktime(st);
2105 static void cfq_choose_cfqg(struct cfq_data *cfqd)
2107 struct cfq_group *cfqg = cfq_get_next_cfqg(cfqd);
2109 cfqd->serving_group = cfqg;
2111 /* Restore the workload type data */
2112 if (cfqg->saved_workload_slice) {
2113 cfqd->workload_expires = jiffies + cfqg->saved_workload_slice;
2114 cfqd->serving_type = cfqg->saved_workload;
2115 cfqd->serving_prio = cfqg->saved_serving_prio;
2117 cfqd->workload_expires = jiffies - 1;
2119 choose_service_tree(cfqd, cfqg);
2123 * Select a queue for service. If we have a current active queue,
2124 * check whether to continue servicing it, or retrieve and set a new one.
2126 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
2128 struct cfq_queue *cfqq, *new_cfqq = NULL;
2130 cfqq = cfqd->active_queue;
2134 if (!cfqd->rq_queued)
2138 * We were waiting for group to get backlogged. Expire the queue
2140 if (cfq_cfqq_wait_busy(cfqq) && !RB_EMPTY_ROOT(&cfqq->sort_list))
2144 * The active queue has run out of time, expire it and select new.
2146 if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq)) {
2148 * If slice had not expired at the completion of last request
2149 * we might not have turned on wait_busy flag. Don't expire
2150 * the queue yet. Allow the group to get backlogged.
2152 * The very fact that we have used the slice, that means we
2153 * have been idling all along on this queue and it should be
2154 * ok to wait for this request to complete.
2156 if (cfqq->cfqg->nr_cfqq == 1 && RB_EMPTY_ROOT(&cfqq->sort_list)
2157 && cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2165 * The active queue has requests and isn't expired, allow it to
2168 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
2172 * If another queue has a request waiting within our mean seek
2173 * distance, let it run. The expire code will check for close
2174 * cooperators and put the close queue at the front of the service
2175 * tree. If possible, merge the expiring queue with the new cfqq.
2177 new_cfqq = cfq_close_cooperator(cfqd, cfqq);
2179 if (!cfqq->new_cfqq)
2180 cfq_setup_merge(cfqq, new_cfqq);
2185 * No requests pending. If the active queue still has requests in
2186 * flight or is idling for a new request, allow either of these
2187 * conditions to happen (or time out) before selecting a new queue.
2189 if (timer_pending(&cfqd->idle_slice_timer) ||
2190 (cfqq->dispatched && cfq_should_idle(cfqd, cfqq))) {
2196 cfq_slice_expired(cfqd, 0);
2199 * Current queue expired. Check if we have to switch to a new
2203 cfq_choose_cfqg(cfqd);
2205 cfqq = cfq_set_active_queue(cfqd, new_cfqq);
2210 static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
2214 while (cfqq->next_rq) {
2215 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
2219 BUG_ON(!list_empty(&cfqq->fifo));
2221 /* By default cfqq is not expired if it is empty. Do it explicitly */
2222 __cfq_slice_expired(cfqq->cfqd, cfqq, 0);
2227 * Drain our current requests. Used for barriers and when switching
2228 * io schedulers on-the-fly.
2230 static int cfq_forced_dispatch(struct cfq_data *cfqd)
2232 struct cfq_queue *cfqq;
2235 while ((cfqq = cfq_get_next_queue_forced(cfqd)) != NULL)
2236 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
2238 cfq_slice_expired(cfqd, 0);
2239 BUG_ON(cfqd->busy_queues);
2241 cfq_log(cfqd, "forced_dispatch=%d", dispatched);
2245 static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2247 unsigned int max_dispatch;
2250 * Drain async requests before we start sync IO
2252 if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_driver[BLK_RW_ASYNC])
2256 * If this is an async queue and we have sync IO in flight, let it wait
2258 if (cfqd->sync_flight && !cfq_cfqq_sync(cfqq))
2261 max_dispatch = cfqd->cfq_quantum;
2262 if (cfq_class_idle(cfqq))
2266 * Does this cfqq already have too much IO in flight?
2268 if (cfqq->dispatched >= max_dispatch) {
2270 * idle queue must always only have a single IO in flight
2272 if (cfq_class_idle(cfqq))
2276 * We have other queues, don't allow more IO from this one
2278 if (cfqd->busy_queues > 1)
2282 * Sole queue user, no limit
2288 * Async queues must wait a bit before being allowed dispatch.
2289 * We also ramp up the dispatch depth gradually for async IO,
2290 * based on the last sync IO we serviced
2292 if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
2293 unsigned long last_sync = jiffies - cfqd->last_delayed_sync;
2296 depth = last_sync / cfqd->cfq_slice[1];
2297 if (!depth && !cfqq->dispatched)
2299 if (depth < max_dispatch)
2300 max_dispatch = depth;
2304 * If we're below the current max, allow a dispatch
2306 return cfqq->dispatched < max_dispatch;
2310 * Dispatch a request from cfqq, moving them to the request queue
2313 static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2317 BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
2319 if (!cfq_may_dispatch(cfqd, cfqq))
2323 * follow expired path, else get first next available
2325 rq = cfq_check_fifo(cfqq);
2330 * insert request into driver dispatch list
2332 cfq_dispatch_insert(cfqd->queue, rq);
2334 if (!cfqd->active_cic) {
2335 struct cfq_io_context *cic = RQ_CIC(rq);
2337 atomic_long_inc(&cic->ioc->refcount);
2338 cfqd->active_cic = cic;
2345 * Find the cfqq that we need to service and move a request from that to the
2348 static int cfq_dispatch_requests(struct request_queue *q, int force)
2350 struct cfq_data *cfqd = q->elevator->elevator_data;
2351 struct cfq_queue *cfqq;
2353 if (!cfqd->busy_queues)
2356 if (unlikely(force))
2357 return cfq_forced_dispatch(cfqd);
2359 cfqq = cfq_select_queue(cfqd);
2364 * Dispatch a request from this cfqq, if it is allowed
2366 if (!cfq_dispatch_request(cfqd, cfqq))
2369 cfqq->slice_dispatch++;
2370 cfq_clear_cfqq_must_dispatch(cfqq);
2373 * expire an async queue immediately if it has used up its slice. idle
2374 * queue always expire after 1 dispatch round.
2376 if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
2377 cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
2378 cfq_class_idle(cfqq))) {
2379 cfqq->slice_end = jiffies + 1;
2380 cfq_slice_expired(cfqd, 0);
2383 cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
2388 * task holds one reference to the queue, dropped when task exits. each rq
2389 * in-flight on this queue also holds a reference, dropped when rq is freed.
2391 * Each cfq queue took a reference on the parent group. Drop it now.
2392 * queue lock must be held here.
2394 static void cfq_put_queue(struct cfq_queue *cfqq)
2396 struct cfq_data *cfqd = cfqq->cfqd;
2397 struct cfq_group *cfqg, *orig_cfqg;
2399 BUG_ON(atomic_read(&cfqq->ref) <= 0);
2401 if (!atomic_dec_and_test(&cfqq->ref))
2404 cfq_log_cfqq(cfqd, cfqq, "put_queue");
2405 BUG_ON(rb_first(&cfqq->sort_list));
2406 BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
2408 orig_cfqg = cfqq->orig_cfqg;
2410 if (unlikely(cfqd->active_queue == cfqq)) {
2411 __cfq_slice_expired(cfqd, cfqq, 0);
2412 cfq_schedule_dispatch(cfqd);
2415 BUG_ON(cfq_cfqq_on_rr(cfqq));
2416 kmem_cache_free(cfq_pool, cfqq);
2419 cfq_put_cfqg(orig_cfqg);
2423 * Must always be called with the rcu_read_lock() held
2426 __call_for_each_cic(struct io_context *ioc,
2427 void (*func)(struct io_context *, struct cfq_io_context *))
2429 struct cfq_io_context *cic;
2430 struct hlist_node *n;
2432 hlist_for_each_entry_rcu(cic, n, &ioc->cic_list, cic_list)
2437 * Call func for each cic attached to this ioc.
2440 call_for_each_cic(struct io_context *ioc,
2441 void (*func)(struct io_context *, struct cfq_io_context *))
2444 __call_for_each_cic(ioc, func);
2448 static void cfq_cic_free_rcu(struct rcu_head *head)
2450 struct cfq_io_context *cic;
2452 cic = container_of(head, struct cfq_io_context, rcu_head);
2454 kmem_cache_free(cfq_ioc_pool, cic);
2455 elv_ioc_count_dec(cfq_ioc_count);
2459 * CFQ scheduler is exiting, grab exit lock and check
2460 * the pending io context count. If it hits zero,
2461 * complete ioc_gone and set it back to NULL
2463 spin_lock(&ioc_gone_lock);
2464 if (ioc_gone && !elv_ioc_count_read(cfq_ioc_count)) {
2468 spin_unlock(&ioc_gone_lock);
2472 static void cfq_cic_free(struct cfq_io_context *cic)
2474 call_rcu(&cic->rcu_head, cfq_cic_free_rcu);
2477 static void cic_free_func(struct io_context *ioc, struct cfq_io_context *cic)
2479 unsigned long flags;
2481 BUG_ON(!cic->dead_key);
2483 spin_lock_irqsave(&ioc->lock, flags);
2484 radix_tree_delete(&ioc->radix_root, cic->dead_key);
2485 hlist_del_rcu(&cic->cic_list);
2486 spin_unlock_irqrestore(&ioc->lock, flags);
2492 * Must be called with rcu_read_lock() held or preemption otherwise disabled.
2493 * Only two callers of this - ->dtor() which is called with the rcu_read_lock(),
2494 * and ->trim() which is called with the task lock held
2496 static void cfq_free_io_context(struct io_context *ioc)
2499 * ioc->refcount is zero here, or we are called from elv_unregister(),
2500 * so no more cic's are allowed to be linked into this ioc. So it
2501 * should be ok to iterate over the known list, we will see all cic's
2502 * since no new ones are added.
2504 __call_for_each_cic(ioc, cic_free_func);
2507 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2509 struct cfq_queue *__cfqq, *next;
2511 if (unlikely(cfqq == cfqd->active_queue)) {
2512 __cfq_slice_expired(cfqd, cfqq, 0);
2513 cfq_schedule_dispatch(cfqd);
2517 * If this queue was scheduled to merge with another queue, be
2518 * sure to drop the reference taken on that queue (and others in
2519 * the merge chain). See cfq_setup_merge and cfq_merge_cfqqs.
2521 __cfqq = cfqq->new_cfqq;
2523 if (__cfqq == cfqq) {
2524 WARN(1, "cfqq->new_cfqq loop detected\n");
2527 next = __cfqq->new_cfqq;
2528 cfq_put_queue(__cfqq);
2532 cfq_put_queue(cfqq);
2535 static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
2536 struct cfq_io_context *cic)
2538 struct io_context *ioc = cic->ioc;
2540 list_del_init(&cic->queue_list);
2543 * Make sure key == NULL is seen for dead queues
2546 cic->dead_key = (unsigned long) cic->key;
2549 if (ioc->ioc_data == cic)
2550 rcu_assign_pointer(ioc->ioc_data, NULL);
2552 if (cic->cfqq[BLK_RW_ASYNC]) {
2553 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]);
2554 cic->cfqq[BLK_RW_ASYNC] = NULL;
2557 if (cic->cfqq[BLK_RW_SYNC]) {
2558 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]);
2559 cic->cfqq[BLK_RW_SYNC] = NULL;
2563 static void cfq_exit_single_io_context(struct io_context *ioc,
2564 struct cfq_io_context *cic)
2566 struct cfq_data *cfqd = cic->key;
2569 struct request_queue *q = cfqd->queue;
2570 unsigned long flags;
2572 spin_lock_irqsave(q->queue_lock, flags);
2575 * Ensure we get a fresh copy of the ->key to prevent
2576 * race between exiting task and queue
2578 smp_read_barrier_depends();
2580 __cfq_exit_single_io_context(cfqd, cic);
2582 spin_unlock_irqrestore(q->queue_lock, flags);
2587 * The process that ioc belongs to has exited, we need to clean up
2588 * and put the internal structures we have that belongs to that process.
2590 static void cfq_exit_io_context(struct io_context *ioc)
2592 call_for_each_cic(ioc, cfq_exit_single_io_context);
2595 static struct cfq_io_context *
2596 cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
2598 struct cfq_io_context *cic;
2600 cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO,
2603 cic->last_end_request = jiffies;
2604 INIT_LIST_HEAD(&cic->queue_list);
2605 INIT_HLIST_NODE(&cic->cic_list);
2606 cic->dtor = cfq_free_io_context;
2607 cic->exit = cfq_exit_io_context;
2608 elv_ioc_count_inc(cfq_ioc_count);
2614 static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc)
2616 struct task_struct *tsk = current;
2619 if (!cfq_cfqq_prio_changed(cfqq))
2622 ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio);
2623 switch (ioprio_class) {
2625 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
2626 case IOPRIO_CLASS_NONE:
2628 * no prio set, inherit CPU scheduling settings
2630 cfqq->ioprio = task_nice_ioprio(tsk);
2631 cfqq->ioprio_class = task_nice_ioclass(tsk);
2633 case IOPRIO_CLASS_RT:
2634 cfqq->ioprio = task_ioprio(ioc);
2635 cfqq->ioprio_class = IOPRIO_CLASS_RT;
2637 case IOPRIO_CLASS_BE:
2638 cfqq->ioprio = task_ioprio(ioc);
2639 cfqq->ioprio_class = IOPRIO_CLASS_BE;
2641 case IOPRIO_CLASS_IDLE:
2642 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
2644 cfq_clear_cfqq_idle_window(cfqq);
2649 * keep track of original prio settings in case we have to temporarily
2650 * elevate the priority of this queue
2652 cfqq->org_ioprio = cfqq->ioprio;
2653 cfqq->org_ioprio_class = cfqq->ioprio_class;
2654 cfq_clear_cfqq_prio_changed(cfqq);
2657 static void changed_ioprio(struct io_context *ioc, struct cfq_io_context *cic)
2659 struct cfq_data *cfqd = cic->key;
2660 struct cfq_queue *cfqq;
2661 unsigned long flags;
2663 if (unlikely(!cfqd))
2666 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2668 cfqq = cic->cfqq[BLK_RW_ASYNC];
2670 struct cfq_queue *new_cfqq;
2671 new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic->ioc,
2674 cic->cfqq[BLK_RW_ASYNC] = new_cfqq;
2675 cfq_put_queue(cfqq);
2679 cfqq = cic->cfqq[BLK_RW_SYNC];
2681 cfq_mark_cfqq_prio_changed(cfqq);
2683 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2686 static void cfq_ioc_set_ioprio(struct io_context *ioc)
2688 call_for_each_cic(ioc, changed_ioprio);
2689 ioc->ioprio_changed = 0;
2692 static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2693 pid_t pid, bool is_sync)
2695 RB_CLEAR_NODE(&cfqq->rb_node);
2696 RB_CLEAR_NODE(&cfqq->p_node);
2697 INIT_LIST_HEAD(&cfqq->fifo);
2699 atomic_set(&cfqq->ref, 0);
2702 cfq_mark_cfqq_prio_changed(cfqq);
2705 if (!cfq_class_idle(cfqq))
2706 cfq_mark_cfqq_idle_window(cfqq);
2707 cfq_mark_cfqq_sync(cfqq);
2712 #ifdef CONFIG_CFQ_GROUP_IOSCHED
2713 static void changed_cgroup(struct io_context *ioc, struct cfq_io_context *cic)
2715 struct cfq_queue *sync_cfqq = cic_to_cfqq(cic, 1);
2716 struct cfq_data *cfqd = cic->key;
2717 unsigned long flags;
2718 struct request_queue *q;
2720 if (unlikely(!cfqd))
2725 spin_lock_irqsave(q->queue_lock, flags);
2729 * Drop reference to sync queue. A new sync queue will be
2730 * assigned in new group upon arrival of a fresh request.
2732 cfq_log_cfqq(cfqd, sync_cfqq, "changed cgroup");
2733 cic_set_cfqq(cic, NULL, 1);
2734 cfq_put_queue(sync_cfqq);
2737 spin_unlock_irqrestore(q->queue_lock, flags);
2740 static void cfq_ioc_set_cgroup(struct io_context *ioc)
2742 call_for_each_cic(ioc, changed_cgroup);
2743 ioc->cgroup_changed = 0;
2745 #endif /* CONFIG_CFQ_GROUP_IOSCHED */
2747 static struct cfq_queue *
2748 cfq_find_alloc_queue(struct cfq_data *cfqd, bool is_sync,
2749 struct io_context *ioc, gfp_t gfp_mask)
2751 struct cfq_queue *cfqq, *new_cfqq = NULL;
2752 struct cfq_io_context *cic;
2753 struct cfq_group *cfqg;
2756 cfqg = cfq_get_cfqg(cfqd, 1);
2757 cic = cfq_cic_lookup(cfqd, ioc);
2758 /* cic always exists here */
2759 cfqq = cic_to_cfqq(cic, is_sync);
2762 * Always try a new alloc if we fell back to the OOM cfqq
2763 * originally, since it should just be a temporary situation.
2765 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
2770 } else if (gfp_mask & __GFP_WAIT) {
2771 spin_unlock_irq(cfqd->queue->queue_lock);
2772 new_cfqq = kmem_cache_alloc_node(cfq_pool,
2773 gfp_mask | __GFP_ZERO,
2775 spin_lock_irq(cfqd->queue->queue_lock);
2779 cfqq = kmem_cache_alloc_node(cfq_pool,
2780 gfp_mask | __GFP_ZERO,
2785 cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
2786 cfq_init_prio_data(cfqq, ioc);
2787 cfq_link_cfqq_cfqg(cfqq, cfqg);
2788 cfq_log_cfqq(cfqd, cfqq, "alloced");
2790 cfqq = &cfqd->oom_cfqq;
2794 kmem_cache_free(cfq_pool, new_cfqq);
2799 static struct cfq_queue **
2800 cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
2802 switch (ioprio_class) {
2803 case IOPRIO_CLASS_RT:
2804 return &cfqd->async_cfqq[0][ioprio];
2805 case IOPRIO_CLASS_BE:
2806 return &cfqd->async_cfqq[1][ioprio];
2807 case IOPRIO_CLASS_IDLE:
2808 return &cfqd->async_idle_cfqq;
2814 static struct cfq_queue *
2815 cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct io_context *ioc,
2818 const int ioprio = task_ioprio(ioc);
2819 const int ioprio_class = task_ioprio_class(ioc);
2820 struct cfq_queue **async_cfqq = NULL;
2821 struct cfq_queue *cfqq = NULL;
2824 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
2829 cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask);
2832 * pin the queue now that it's allocated, scheduler exit will prune it
2834 if (!is_sync && !(*async_cfqq)) {
2835 atomic_inc(&cfqq->ref);
2839 atomic_inc(&cfqq->ref);
2844 * We drop cfq io contexts lazily, so we may find a dead one.
2847 cfq_drop_dead_cic(struct cfq_data *cfqd, struct io_context *ioc,
2848 struct cfq_io_context *cic)
2850 unsigned long flags;
2852 WARN_ON(!list_empty(&cic->queue_list));
2854 spin_lock_irqsave(&ioc->lock, flags);
2856 BUG_ON(ioc->ioc_data == cic);
2858 radix_tree_delete(&ioc->radix_root, (unsigned long) cfqd);
2859 hlist_del_rcu(&cic->cic_list);
2860 spin_unlock_irqrestore(&ioc->lock, flags);
2865 static struct cfq_io_context *
2866 cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc)
2868 struct cfq_io_context *cic;
2869 unsigned long flags;
2878 * we maintain a last-hit cache, to avoid browsing over the tree
2880 cic = rcu_dereference(ioc->ioc_data);
2881 if (cic && cic->key == cfqd) {
2887 cic = radix_tree_lookup(&ioc->radix_root, (unsigned long) cfqd);
2891 /* ->key must be copied to avoid race with cfq_exit_queue() */
2894 cfq_drop_dead_cic(cfqd, ioc, cic);
2899 spin_lock_irqsave(&ioc->lock, flags);
2900 rcu_assign_pointer(ioc->ioc_data, cic);
2901 spin_unlock_irqrestore(&ioc->lock, flags);
2909 * Add cic into ioc, using cfqd as the search key. This enables us to lookup
2910 * the process specific cfq io context when entered from the block layer.
2911 * Also adds the cic to a per-cfqd list, used when this queue is removed.
2913 static int cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
2914 struct cfq_io_context *cic, gfp_t gfp_mask)
2916 unsigned long flags;
2919 ret = radix_tree_preload(gfp_mask);
2924 spin_lock_irqsave(&ioc->lock, flags);
2925 ret = radix_tree_insert(&ioc->radix_root,
2926 (unsigned long) cfqd, cic);
2928 hlist_add_head_rcu(&cic->cic_list, &ioc->cic_list);
2929 spin_unlock_irqrestore(&ioc->lock, flags);
2931 radix_tree_preload_end();
2934 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2935 list_add(&cic->queue_list, &cfqd->cic_list);
2936 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2941 printk(KERN_ERR "cfq: cic link failed!\n");
2947 * Setup general io context and cfq io context. There can be several cfq
2948 * io contexts per general io context, if this process is doing io to more
2949 * than one device managed by cfq.
2951 static struct cfq_io_context *
2952 cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
2954 struct io_context *ioc = NULL;
2955 struct cfq_io_context *cic;
2957 might_sleep_if(gfp_mask & __GFP_WAIT);
2959 ioc = get_io_context(gfp_mask, cfqd->queue->node);
2963 cic = cfq_cic_lookup(cfqd, ioc);
2967 cic = cfq_alloc_io_context(cfqd, gfp_mask);
2971 if (cfq_cic_link(cfqd, ioc, cic, gfp_mask))
2975 smp_read_barrier_depends();
2976 if (unlikely(ioc->ioprio_changed))
2977 cfq_ioc_set_ioprio(ioc);
2979 #ifdef CONFIG_CFQ_GROUP_IOSCHED
2980 if (unlikely(ioc->cgroup_changed))
2981 cfq_ioc_set_cgroup(ioc);
2987 put_io_context(ioc);
2992 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
2994 unsigned long elapsed = jiffies - cic->last_end_request;
2995 unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
2997 cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
2998 cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
2999 cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
3003 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3009 if (!cfqq->last_request_pos)
3011 else if (cfqq->last_request_pos < blk_rq_pos(rq))
3012 sdist = blk_rq_pos(rq) - cfqq->last_request_pos;
3014 sdist = cfqq->last_request_pos - blk_rq_pos(rq);
3017 * Don't allow the seek distance to get too large from the
3018 * odd fragment, pagein, etc
3020 if (cfqq->seek_samples <= 60) /* second&third seek */
3021 sdist = min(sdist, (cfqq->seek_mean * 4) + 2*1024*1024);
3023 sdist = min(sdist, (cfqq->seek_mean * 4) + 2*1024*64);
3025 cfqq->seek_samples = (7*cfqq->seek_samples + 256) / 8;
3026 cfqq->seek_total = (7*cfqq->seek_total + (u64)256*sdist) / 8;
3027 total = cfqq->seek_total + (cfqq->seek_samples/2);
3028 do_div(total, cfqq->seek_samples);
3029 cfqq->seek_mean = (sector_t)total;
3032 * If this cfqq is shared between multiple processes, check to
3033 * make sure that those processes are still issuing I/Os within
3034 * the mean seek distance. If not, it may be time to break the
3035 * queues apart again.
3037 if (cfq_cfqq_coop(cfqq)) {
3038 if (CFQQ_SEEKY(cfqq) && !cfqq->seeky_start)
3039 cfqq->seeky_start = jiffies;
3040 else if (!CFQQ_SEEKY(cfqq))
3041 cfqq->seeky_start = 0;
3046 * Disable idle window if the process thinks too long or seeks so much that
3050 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3051 struct cfq_io_context *cic)
3053 int old_idle, enable_idle;
3056 * Don't idle for async or idle io prio class
3058 if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
3061 enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
3063 if (cfqq->queued[0] + cfqq->queued[1] >= 4)
3064 cfq_mark_cfqq_deep(cfqq);
3066 if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle ||
3067 (!cfq_cfqq_deep(cfqq) && sample_valid(cfqq->seek_samples)
3068 && CFQQ_SEEKY(cfqq)))
3070 else if (sample_valid(cic->ttime_samples)) {
3071 if (cic->ttime_mean > cfqd->cfq_slice_idle)
3077 if (old_idle != enable_idle) {
3078 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
3080 cfq_mark_cfqq_idle_window(cfqq);
3082 cfq_clear_cfqq_idle_window(cfqq);
3087 * Check if new_cfqq should preempt the currently active queue. Return 0 for
3088 * no or if we aren't sure, a 1 will cause a preempt.
3091 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
3094 struct cfq_queue *cfqq;
3096 cfqq = cfqd->active_queue;
3100 if (cfq_class_idle(new_cfqq))
3103 if (cfq_class_idle(cfqq))
3107 * if the new request is sync, but the currently running queue is
3108 * not, let the sync request have priority.
3110 if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
3113 if (new_cfqq->cfqg != cfqq->cfqg)
3116 if (cfq_slice_used(cfqq))
3119 /* Allow preemption only if we are idling on sync-noidle tree */
3120 if (cfqd->serving_type == SYNC_NOIDLE_WORKLOAD &&
3121 cfqq_type(new_cfqq) == SYNC_NOIDLE_WORKLOAD &&
3122 new_cfqq->service_tree->count == 2 &&
3123 RB_EMPTY_ROOT(&cfqq->sort_list))
3127 * So both queues are sync. Let the new request get disk time if
3128 * it's a metadata request and the current queue is doing regular IO.
3130 if (rq_is_meta(rq) && !cfqq->meta_pending)
3134 * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
3136 if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
3139 if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
3143 * if this request is as-good as one we would expect from the
3144 * current cfqq, let it preempt
3146 if (cfq_rq_close(cfqd, cfqq, rq))
3153 * cfqq preempts the active queue. if we allowed preempt with no slice left,
3154 * let it have half of its nominal slice.
3156 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3158 cfq_log_cfqq(cfqd, cfqq, "preempt");
3159 cfq_slice_expired(cfqd, 1);
3162 * Put the new queue at the front of the of the current list,
3163 * so we know that it will be selected next.
3165 BUG_ON(!cfq_cfqq_on_rr(cfqq));
3167 cfq_service_tree_add(cfqd, cfqq, 1);
3169 cfqq->slice_end = 0;
3170 cfq_mark_cfqq_slice_new(cfqq);
3174 * Called when a new fs request (rq) is added (to cfqq). Check if there's
3175 * something we should do about it
3178 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3181 struct cfq_io_context *cic = RQ_CIC(rq);
3185 cfqq->meta_pending++;
3187 cfq_update_io_thinktime(cfqd, cic);
3188 cfq_update_io_seektime(cfqd, cfqq, rq);
3189 cfq_update_idle_window(cfqd, cfqq, cic);
3191 cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
3193 if (cfqq == cfqd->active_queue) {
3195 * Remember that we saw a request from this process, but
3196 * don't start queuing just yet. Otherwise we risk seeing lots
3197 * of tiny requests, because we disrupt the normal plugging
3198 * and merging. If the request is already larger than a single
3199 * page, let it rip immediately. For that case we assume that
3200 * merging is already done. Ditto for a busy system that
3201 * has other work pending, don't risk delaying until the
3202 * idle timer unplug to continue working.
3204 if (cfq_cfqq_wait_request(cfqq)) {
3205 if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
3206 cfqd->busy_queues > 1) {
3207 del_timer(&cfqd->idle_slice_timer);
3208 cfq_clear_cfqq_wait_request(cfqq);
3209 __blk_run_queue(cfqd->queue);
3211 cfq_mark_cfqq_must_dispatch(cfqq);
3213 } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
3215 * not the active queue - expire current slice if it is
3216 * idle and has expired it's mean thinktime or this new queue
3217 * has some old slice time left and is of higher priority or
3218 * this new queue is RT and the current one is BE
3220 cfq_preempt_queue(cfqd, cfqq);
3221 __blk_run_queue(cfqd->queue);
3225 static void cfq_insert_request(struct request_queue *q, struct request *rq)
3227 struct cfq_data *cfqd = q->elevator->elevator_data;
3228 struct cfq_queue *cfqq = RQ_CFQQ(rq);
3230 cfq_log_cfqq(cfqd, cfqq, "insert_request");
3231 cfq_init_prio_data(cfqq, RQ_CIC(rq)->ioc);
3233 rq_set_fifo_time(rq, jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)]);
3234 list_add_tail(&rq->queuelist, &cfqq->fifo);
3237 cfq_rq_enqueued(cfqd, cfqq, rq);
3241 * Update hw_tag based on peak queue depth over 50 samples under
3244 static void cfq_update_hw_tag(struct cfq_data *cfqd)
3246 struct cfq_queue *cfqq = cfqd->active_queue;
3248 if (rq_in_driver(cfqd) > cfqd->hw_tag_est_depth)
3249 cfqd->hw_tag_est_depth = rq_in_driver(cfqd);
3251 if (cfqd->hw_tag == 1)
3254 if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
3255 rq_in_driver(cfqd) <= CFQ_HW_QUEUE_MIN)
3259 * If active queue hasn't enough requests and can idle, cfq might not
3260 * dispatch sufficient requests to hardware. Don't zero hw_tag in this
3263 if (cfqq && cfq_cfqq_idle_window(cfqq) &&
3264 cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] <
3265 CFQ_HW_QUEUE_MIN && rq_in_driver(cfqd) < CFQ_HW_QUEUE_MIN)
3268 if (cfqd->hw_tag_samples++ < 50)
3271 if (cfqd->hw_tag_est_depth >= CFQ_HW_QUEUE_MIN)
3277 static bool cfq_should_wait_busy(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3279 struct cfq_io_context *cic = cfqd->active_cic;
3281 /* If there are other queues in the group, don't wait */
3282 if (cfqq->cfqg->nr_cfqq > 1)
3285 if (cfq_slice_used(cfqq))
3288 /* if slice left is less than think time, wait busy */
3289 if (cic && sample_valid(cic->ttime_samples)
3290 && (cfqq->slice_end - jiffies < cic->ttime_mean))
3294 * If think times is less than a jiffy than ttime_mean=0 and above
3295 * will not be true. It might happen that slice has not expired yet
3296 * but will expire soon (4-5 ns) during select_queue(). To cover the
3297 * case where think time is less than a jiffy, mark the queue wait
3298 * busy if only 1 jiffy is left in the slice.
3300 if (cfqq->slice_end - jiffies == 1)
3306 static void cfq_completed_request(struct request_queue *q, struct request *rq)
3308 struct cfq_queue *cfqq = RQ_CFQQ(rq);
3309 struct cfq_data *cfqd = cfqq->cfqd;
3310 const int sync = rq_is_sync(rq);
3314 cfq_log_cfqq(cfqd, cfqq, "complete rqnoidle %d", !!rq_noidle(rq));
3316 cfq_update_hw_tag(cfqd);
3318 WARN_ON(!cfqd->rq_in_driver[sync]);
3319 WARN_ON(!cfqq->dispatched);
3320 cfqd->rq_in_driver[sync]--;
3323 if (cfq_cfqq_sync(cfqq))
3324 cfqd->sync_flight--;
3327 RQ_CIC(rq)->last_end_request = now;
3328 if (!time_after(rq->start_time + cfqd->cfq_fifo_expire[1], now))
3329 cfqd->last_delayed_sync = now;
3333 * If this is the active queue, check if it needs to be expired,
3334 * or if we want to idle in case it has no pending requests.
3336 if (cfqd->active_queue == cfqq) {
3337 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
3339 if (cfq_cfqq_slice_new(cfqq)) {
3340 cfq_set_prio_slice(cfqd, cfqq);
3341 cfq_clear_cfqq_slice_new(cfqq);
3345 * Should we wait for next request to come in before we expire
3348 if (cfq_should_wait_busy(cfqd, cfqq)) {
3349 cfqq->slice_end = jiffies + cfqd->cfq_slice_idle;
3350 cfq_mark_cfqq_wait_busy(cfqq);
3354 * Idling is not enabled on:
3356 * - idle-priority queues
3358 * - queues with still some requests queued
3359 * - when there is a close cooperator
3361 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
3362 cfq_slice_expired(cfqd, 1);
3363 else if (sync && cfqq_empty &&
3364 !cfq_close_cooperator(cfqd, cfqq)) {
3365 cfqd->noidle_tree_requires_idle |= !rq_noidle(rq);
3367 * Idling is enabled for SYNC_WORKLOAD.
3368 * SYNC_NOIDLE_WORKLOAD idles at the end of the tree
3369 * only if we processed at least one !rq_noidle request
3371 if (cfqd->serving_type == SYNC_WORKLOAD
3372 || cfqd->noidle_tree_requires_idle
3373 || cfqq->cfqg->nr_cfqq == 1)
3374 cfq_arm_slice_timer(cfqd);
3378 if (!rq_in_driver(cfqd))
3379 cfq_schedule_dispatch(cfqd);
3383 * we temporarily boost lower priority queues if they are holding fs exclusive
3384 * resources. they are boosted to normal prio (CLASS_BE/4)
3386 static void cfq_prio_boost(struct cfq_queue *cfqq)
3388 if (has_fs_excl()) {
3390 * boost idle prio on transactions that would lock out other
3391 * users of the filesystem
3393 if (cfq_class_idle(cfqq))
3394 cfqq->ioprio_class = IOPRIO_CLASS_BE;
3395 if (cfqq->ioprio > IOPRIO_NORM)
3396 cfqq->ioprio = IOPRIO_NORM;
3399 * unboost the queue (if needed)
3401 cfqq->ioprio_class = cfqq->org_ioprio_class;
3402 cfqq->ioprio = cfqq->org_ioprio;
3406 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
3408 if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
3409 cfq_mark_cfqq_must_alloc_slice(cfqq);
3410 return ELV_MQUEUE_MUST;
3413 return ELV_MQUEUE_MAY;
3416 static int cfq_may_queue(struct request_queue *q, int rw)
3418 struct cfq_data *cfqd = q->elevator->elevator_data;
3419 struct task_struct *tsk = current;
3420 struct cfq_io_context *cic;
3421 struct cfq_queue *cfqq;
3424 * don't force setup of a queue from here, as a call to may_queue
3425 * does not necessarily imply that a request actually will be queued.
3426 * so just lookup a possibly existing queue, or return 'may queue'
3429 cic = cfq_cic_lookup(cfqd, tsk->io_context);
3431 return ELV_MQUEUE_MAY;
3433 cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
3435 cfq_init_prio_data(cfqq, cic->ioc);
3436 cfq_prio_boost(cfqq);
3438 return __cfq_may_queue(cfqq);
3441 return ELV_MQUEUE_MAY;
3445 * queue lock held here
3447 static void cfq_put_request(struct request *rq)
3449 struct cfq_queue *cfqq = RQ_CFQQ(rq);
3452 const int rw = rq_data_dir(rq);
3454 BUG_ON(!cfqq->allocated[rw]);
3455 cfqq->allocated[rw]--;
3457 put_io_context(RQ_CIC(rq)->ioc);
3459 rq->elevator_private = NULL;
3460 rq->elevator_private2 = NULL;
3462 cfq_put_queue(cfqq);
3466 static struct cfq_queue *
3467 cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_context *cic,
3468 struct cfq_queue *cfqq)
3470 cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq);
3471 cic_set_cfqq(cic, cfqq->new_cfqq, 1);
3472 cfq_mark_cfqq_coop(cfqq->new_cfqq);
3473 cfq_put_queue(cfqq);
3474 return cic_to_cfqq(cic, 1);
3477 static int should_split_cfqq(struct cfq_queue *cfqq)
3479 if (cfqq->seeky_start &&
3480 time_after(jiffies, cfqq->seeky_start + CFQQ_COOP_TOUT))
3486 * Returns NULL if a new cfqq should be allocated, or the old cfqq if this
3487 * was the last process referring to said cfqq.
3489 static struct cfq_queue *
3490 split_cfqq(struct cfq_io_context *cic, struct cfq_queue *cfqq)
3492 if (cfqq_process_refs(cfqq) == 1) {
3493 cfqq->seeky_start = 0;
3494 cfqq->pid = current->pid;
3495 cfq_clear_cfqq_coop(cfqq);
3499 cic_set_cfqq(cic, NULL, 1);
3500 cfq_put_queue(cfqq);
3504 * Allocate cfq data structures associated with this request.
3507 cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
3509 struct cfq_data *cfqd = q->elevator->elevator_data;
3510 struct cfq_io_context *cic;
3511 const int rw = rq_data_dir(rq);
3512 const bool is_sync = rq_is_sync(rq);
3513 struct cfq_queue *cfqq;
3514 unsigned long flags;
3516 might_sleep_if(gfp_mask & __GFP_WAIT);
3518 cic = cfq_get_io_context(cfqd, gfp_mask);
3520 spin_lock_irqsave(q->queue_lock, flags);
3526 cfqq = cic_to_cfqq(cic, is_sync);
3527 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
3528 cfqq = cfq_get_queue(cfqd, is_sync, cic->ioc, gfp_mask);
3529 cic_set_cfqq(cic, cfqq, is_sync);
3532 * If the queue was seeky for too long, break it apart.
3534 if (cfq_cfqq_coop(cfqq) && should_split_cfqq(cfqq)) {
3535 cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq");
3536 cfqq = split_cfqq(cic, cfqq);
3542 * Check to see if this queue is scheduled to merge with
3543 * another, closely cooperating queue. The merging of
3544 * queues happens here as it must be done in process context.
3545 * The reference on new_cfqq was taken in merge_cfqqs.
3548 cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq);
3551 cfqq->allocated[rw]++;
3552 atomic_inc(&cfqq->ref);
3554 spin_unlock_irqrestore(q->queue_lock, flags);
3556 rq->elevator_private = cic;
3557 rq->elevator_private2 = cfqq;
3562 put_io_context(cic->ioc);
3564 cfq_schedule_dispatch(cfqd);
3565 spin_unlock_irqrestore(q->queue_lock, flags);
3566 cfq_log(cfqd, "set_request fail");
3570 static void cfq_kick_queue(struct work_struct *work)
3572 struct cfq_data *cfqd =
3573 container_of(work, struct cfq_data, unplug_work);
3574 struct request_queue *q = cfqd->queue;
3576 spin_lock_irq(q->queue_lock);
3577 __blk_run_queue(cfqd->queue);
3578 spin_unlock_irq(q->queue_lock);
3582 * Timer running if the active_queue is currently idling inside its time slice
3584 static void cfq_idle_slice_timer(unsigned long data)
3586 struct cfq_data *cfqd = (struct cfq_data *) data;
3587 struct cfq_queue *cfqq;
3588 unsigned long flags;
3591 cfq_log(cfqd, "idle timer fired");
3593 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
3595 cfqq = cfqd->active_queue;
3600 * We saw a request before the queue expired, let it through
3602 if (cfq_cfqq_must_dispatch(cfqq))
3608 if (cfq_slice_used(cfqq))
3612 * only expire and reinvoke request handler, if there are
3613 * other queues with pending requests
3615 if (!cfqd->busy_queues)
3619 * not expired and it has a request pending, let it dispatch
3621 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3625 * Queue depth flag is reset only when the idle didn't succeed
3627 cfq_clear_cfqq_deep(cfqq);
3630 cfq_slice_expired(cfqd, timed_out);
3632 cfq_schedule_dispatch(cfqd);
3634 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
3637 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
3639 del_timer_sync(&cfqd->idle_slice_timer);
3640 cancel_work_sync(&cfqd->unplug_work);
3643 static void cfq_put_async_queues(struct cfq_data *cfqd)
3647 for (i = 0; i < IOPRIO_BE_NR; i++) {
3648 if (cfqd->async_cfqq[0][i])
3649 cfq_put_queue(cfqd->async_cfqq[0][i]);
3650 if (cfqd->async_cfqq[1][i])
3651 cfq_put_queue(cfqd->async_cfqq[1][i]);
3654 if (cfqd->async_idle_cfqq)
3655 cfq_put_queue(cfqd->async_idle_cfqq);
3658 static void cfq_cfqd_free(struct rcu_head *head)
3660 kfree(container_of(head, struct cfq_data, rcu));
3663 static void cfq_exit_queue(struct elevator_queue *e)
3665 struct cfq_data *cfqd = e->elevator_data;
3666 struct request_queue *q = cfqd->queue;
3668 cfq_shutdown_timer_wq(cfqd);
3670 spin_lock_irq(q->queue_lock);
3672 if (cfqd->active_queue)
3673 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
3675 while (!list_empty(&cfqd->cic_list)) {
3676 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
3677 struct cfq_io_context,
3680 __cfq_exit_single_io_context(cfqd, cic);
3683 cfq_put_async_queues(cfqd);
3684 cfq_release_cfq_groups(cfqd);
3685 blkiocg_del_blkio_group(&cfqd->root_group.blkg);
3687 spin_unlock_irq(q->queue_lock);
3689 cfq_shutdown_timer_wq(cfqd);
3691 /* Wait for cfqg->blkg->key accessors to exit their grace periods. */
3692 call_rcu(&cfqd->rcu, cfq_cfqd_free);
3695 static void *cfq_init_queue(struct request_queue *q)
3697 struct cfq_data *cfqd;
3699 struct cfq_group *cfqg;
3700 struct cfq_rb_root *st;
3702 cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
3706 /* Init root service tree */
3707 cfqd->grp_service_tree = CFQ_RB_ROOT;
3709 /* Init root group */
3710 cfqg = &cfqd->root_group;
3711 for_each_cfqg_st(cfqg, i, j, st)
3713 RB_CLEAR_NODE(&cfqg->rb_node);
3715 /* Give preference to root group over other groups */
3716 cfqg->weight = 2*BLKIO_WEIGHT_DEFAULT;
3718 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3720 * Take a reference to root group which we never drop. This is just
3721 * to make sure that cfq_put_cfqg() does not try to kfree root group
3723 atomic_set(&cfqg->ref, 1);
3724 blkiocg_add_blkio_group(&blkio_root_cgroup, &cfqg->blkg, (void *)cfqd,
3728 * Not strictly needed (since RB_ROOT just clears the node and we
3729 * zeroed cfqd on alloc), but better be safe in case someone decides
3730 * to add magic to the rb code
3732 for (i = 0; i < CFQ_PRIO_LISTS; i++)
3733 cfqd->prio_trees[i] = RB_ROOT;
3736 * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
3737 * Grab a permanent reference to it, so that the normal code flow
3738 * will not attempt to free it.
3740 cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
3741 atomic_inc(&cfqd->oom_cfqq.ref);
3742 cfq_link_cfqq_cfqg(&cfqd->oom_cfqq, &cfqd->root_group);
3744 INIT_LIST_HEAD(&cfqd->cic_list);
3748 init_timer(&cfqd->idle_slice_timer);
3749 cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
3750 cfqd->idle_slice_timer.data = (unsigned long) cfqd;
3752 INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
3754 cfqd->cfq_quantum = cfq_quantum;
3755 cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
3756 cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
3757 cfqd->cfq_back_max = cfq_back_max;
3758 cfqd->cfq_back_penalty = cfq_back_penalty;
3759 cfqd->cfq_slice[0] = cfq_slice_async;
3760 cfqd->cfq_slice[1] = cfq_slice_sync;
3761 cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
3762 cfqd->cfq_slice_idle = cfq_slice_idle;
3763 cfqd->cfq_latency = 1;
3764 cfqd->cfq_group_isolation = 0;
3767 * we optimistically start assuming sync ops weren't delayed in last
3768 * second, in order to have larger depth for async operations.
3770 cfqd->last_delayed_sync = jiffies - HZ;
3771 INIT_RCU_HEAD(&cfqd->rcu);
3775 static void cfq_slab_kill(void)
3778 * Caller already ensured that pending RCU callbacks are completed,
3779 * so we should have no busy allocations at this point.
3782 kmem_cache_destroy(cfq_pool);
3784 kmem_cache_destroy(cfq_ioc_pool);
3787 static int __init cfq_slab_setup(void)
3789 cfq_pool = KMEM_CACHE(cfq_queue, 0);
3793 cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0);
3804 * sysfs parts below -->
3807 cfq_var_show(unsigned int var, char *page)
3809 return sprintf(page, "%d\n", var);
3813 cfq_var_store(unsigned int *var, const char *page, size_t count)
3815 char *p = (char *) page;
3817 *var = simple_strtoul(p, &p, 10);
3821 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
3822 static ssize_t __FUNC(struct elevator_queue *e, char *page) \
3824 struct cfq_data *cfqd = e->elevator_data; \
3825 unsigned int __data = __VAR; \
3827 __data = jiffies_to_msecs(__data); \
3828 return cfq_var_show(__data, (page)); \
3830 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
3831 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
3832 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
3833 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
3834 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
3835 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
3836 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
3837 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
3838 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
3839 SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
3840 SHOW_FUNCTION(cfq_group_isolation_show, cfqd->cfq_group_isolation, 0);
3841 #undef SHOW_FUNCTION
3843 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
3844 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
3846 struct cfq_data *cfqd = e->elevator_data; \
3847 unsigned int __data; \
3848 int ret = cfq_var_store(&__data, (page), count); \
3849 if (__data < (MIN)) \
3851 else if (__data > (MAX)) \
3854 *(__PTR) = msecs_to_jiffies(__data); \
3856 *(__PTR) = __data; \
3859 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
3860 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
3862 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
3864 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
3865 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
3867 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
3868 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
3869 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
3870 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
3872 STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
3873 STORE_FUNCTION(cfq_group_isolation_store, &cfqd->cfq_group_isolation, 0, 1, 0);
3874 #undef STORE_FUNCTION
3876 #define CFQ_ATTR(name) \
3877 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
3879 static struct elv_fs_entry cfq_attrs[] = {
3881 CFQ_ATTR(fifo_expire_sync),
3882 CFQ_ATTR(fifo_expire_async),
3883 CFQ_ATTR(back_seek_max),
3884 CFQ_ATTR(back_seek_penalty),
3885 CFQ_ATTR(slice_sync),
3886 CFQ_ATTR(slice_async),
3887 CFQ_ATTR(slice_async_rq),
3888 CFQ_ATTR(slice_idle),
3889 CFQ_ATTR(low_latency),
3890 CFQ_ATTR(group_isolation),
3894 static struct elevator_type iosched_cfq = {
3896 .elevator_merge_fn = cfq_merge,
3897 .elevator_merged_fn = cfq_merged_request,
3898 .elevator_merge_req_fn = cfq_merged_requests,
3899 .elevator_allow_merge_fn = cfq_allow_merge,
3900 .elevator_dispatch_fn = cfq_dispatch_requests,
3901 .elevator_add_req_fn = cfq_insert_request,
3902 .elevator_activate_req_fn = cfq_activate_request,
3903 .elevator_deactivate_req_fn = cfq_deactivate_request,
3904 .elevator_queue_empty_fn = cfq_queue_empty,
3905 .elevator_completed_req_fn = cfq_completed_request,
3906 .elevator_former_req_fn = elv_rb_former_request,
3907 .elevator_latter_req_fn = elv_rb_latter_request,
3908 .elevator_set_req_fn = cfq_set_request,
3909 .elevator_put_req_fn = cfq_put_request,
3910 .elevator_may_queue_fn = cfq_may_queue,
3911 .elevator_init_fn = cfq_init_queue,
3912 .elevator_exit_fn = cfq_exit_queue,
3913 .trim = cfq_free_io_context,
3915 .elevator_attrs = cfq_attrs,
3916 .elevator_name = "cfq",
3917 .elevator_owner = THIS_MODULE,
3920 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3921 static struct blkio_policy_type blkio_policy_cfq = {
3923 .blkio_unlink_group_fn = cfq_unlink_blkio_group,
3924 .blkio_update_group_weight_fn = cfq_update_blkio_group_weight,
3928 static struct blkio_policy_type blkio_policy_cfq;
3931 static int __init cfq_init(void)
3934 * could be 0 on HZ < 1000 setups
3936 if (!cfq_slice_async)
3937 cfq_slice_async = 1;
3938 if (!cfq_slice_idle)
3941 if (cfq_slab_setup())
3944 elv_register(&iosched_cfq);
3945 blkio_policy_register(&blkio_policy_cfq);
3950 static void __exit cfq_exit(void)
3952 DECLARE_COMPLETION_ONSTACK(all_gone);
3953 blkio_policy_unregister(&blkio_policy_cfq);
3954 elv_unregister(&iosched_cfq);
3955 ioc_gone = &all_gone;
3956 /* ioc_gone's update must be visible before reading ioc_count */
3960 * this also protects us from entering cfq_slab_kill() with
3961 * pending RCU callbacks
3963 if (elv_ioc_count_read(cfq_ioc_count))
3964 wait_for_completion(&all_gone);
3968 module_init(cfq_init);
3969 module_exit(cfq_exit);
3971 MODULE_AUTHOR("Jens Axboe");
3972 MODULE_LICENSE("GPL");
3973 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");