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 /* Sectors dispatched in current dispatch round */
148 unsigned long nr_sectors;
152 * First index in the service_trees.
153 * IDLE is handled separately, so it has negative index
162 * Second index in the service_trees.
166 SYNC_NOIDLE_WORKLOAD = 1,
170 /* This is per cgroup per device grouping structure */
172 /* group service_tree member */
173 struct rb_node rb_node;
175 /* group service_tree key */
180 /* number of cfqq currently on this group */
183 /* Per group busy queus average. Useful for workload slice calc. */
184 unsigned int busy_queues_avg[2];
186 * rr lists of queues with requests, onle rr for each priority class.
187 * Counts are embedded in the cfq_rb_root
189 struct cfq_rb_root service_trees[2][3];
190 struct cfq_rb_root service_tree_idle;
192 unsigned long saved_workload_slice;
193 enum wl_type_t saved_workload;
194 enum wl_prio_t saved_serving_prio;
195 struct blkio_group blkg;
196 #ifdef CONFIG_CFQ_GROUP_IOSCHED
197 struct hlist_node cfqd_node;
203 * Per block device queue structure
206 struct request_queue *queue;
207 /* Root service tree for cfq_groups */
208 struct cfq_rb_root grp_service_tree;
209 struct cfq_group root_group;
210 /* Number of active cfq groups on group service tree */
214 * The priority currently being served
216 enum wl_prio_t serving_prio;
217 enum wl_type_t serving_type;
218 unsigned long workload_expires;
219 struct cfq_group *serving_group;
220 bool noidle_tree_requires_idle;
223 * Each priority tree is sorted by next_request position. These
224 * trees are used when determining if two or more queues are
225 * interleaving requests (see cfq_close_cooperator).
227 struct rb_root prio_trees[CFQ_PRIO_LISTS];
229 unsigned int busy_queues;
235 * queue-depth detection
241 * -1 => indeterminate, (cfq will behave as if NCQ is present, to allow better detection)
242 * 1 => NCQ is present (hw_tag_est_depth is the estimated max depth)
245 int hw_tag_est_depth;
246 unsigned int hw_tag_samples;
249 * idle window management
251 struct timer_list idle_slice_timer;
252 struct work_struct unplug_work;
254 struct cfq_queue *active_queue;
255 struct cfq_io_context *active_cic;
258 * async queue for each priority case
260 struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
261 struct cfq_queue *async_idle_cfqq;
263 sector_t last_position;
266 * tunables, see top of file
268 unsigned int cfq_quantum;
269 unsigned int cfq_fifo_expire[2];
270 unsigned int cfq_back_penalty;
271 unsigned int cfq_back_max;
272 unsigned int cfq_slice[2];
273 unsigned int cfq_slice_async_rq;
274 unsigned int cfq_slice_idle;
275 unsigned int cfq_latency;
277 struct list_head cic_list;
280 * Fallback dummy cfqq for extreme OOM conditions
282 struct cfq_queue oom_cfqq;
284 unsigned long last_end_sync_rq;
286 /* List of cfq groups being managed on this device*/
287 struct hlist_head cfqg_list;
290 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd);
292 static struct cfq_rb_root *service_tree_for(struct cfq_group *cfqg,
295 struct cfq_data *cfqd)
300 if (prio == IDLE_WORKLOAD)
301 return &cfqg->service_tree_idle;
303 return &cfqg->service_trees[prio][type];
306 enum cfqq_state_flags {
307 CFQ_CFQQ_FLAG_on_rr = 0, /* on round-robin busy list */
308 CFQ_CFQQ_FLAG_wait_request, /* waiting for a request */
309 CFQ_CFQQ_FLAG_must_dispatch, /* must be allowed a dispatch */
310 CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
311 CFQ_CFQQ_FLAG_fifo_expire, /* FIFO checked in this slice */
312 CFQ_CFQQ_FLAG_idle_window, /* slice idling enabled */
313 CFQ_CFQQ_FLAG_prio_changed, /* task priority has changed */
314 CFQ_CFQQ_FLAG_slice_new, /* no requests dispatched in slice */
315 CFQ_CFQQ_FLAG_sync, /* synchronous queue */
316 CFQ_CFQQ_FLAG_coop, /* cfqq is shared */
317 CFQ_CFQQ_FLAG_deep, /* sync cfqq experienced large depth */
318 CFQ_CFQQ_FLAG_wait_busy, /* Waiting for next request */
319 CFQ_CFQQ_FLAG_wait_busy_done, /* Got new request. Expire the queue */
322 #define CFQ_CFQQ_FNS(name) \
323 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \
325 (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name); \
327 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \
329 (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \
331 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \
333 return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \
337 CFQ_CFQQ_FNS(wait_request);
338 CFQ_CFQQ_FNS(must_dispatch);
339 CFQ_CFQQ_FNS(must_alloc_slice);
340 CFQ_CFQQ_FNS(fifo_expire);
341 CFQ_CFQQ_FNS(idle_window);
342 CFQ_CFQQ_FNS(prio_changed);
343 CFQ_CFQQ_FNS(slice_new);
347 CFQ_CFQQ_FNS(wait_busy);
348 CFQ_CFQQ_FNS(wait_busy_done);
351 #ifdef CONFIG_DEBUG_CFQ_IOSCHED
352 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
353 blk_add_trace_msg((cfqd)->queue, "cfq%d%c %s " fmt, (cfqq)->pid, \
354 cfq_cfqq_sync((cfqq)) ? 'S' : 'A', \
355 blkg_path(&(cfqq)->cfqg->blkg), ##args);
357 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...) \
358 blk_add_trace_msg((cfqd)->queue, "%s " fmt, \
359 blkg_path(&(cfqg)->blkg), ##args); \
362 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
363 blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args)
364 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...) do {} while (0);
366 #define cfq_log(cfqd, fmt, args...) \
367 blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
369 /* Traverses through cfq group service trees */
370 #define for_each_cfqg_st(cfqg, i, j, st) \
371 for (i = 0; i <= IDLE_WORKLOAD; i++) \
372 for (j = 0, st = i < IDLE_WORKLOAD ? &cfqg->service_trees[i][j]\
373 : &cfqg->service_tree_idle; \
374 (i < IDLE_WORKLOAD && j <= SYNC_WORKLOAD) || \
375 (i == IDLE_WORKLOAD && j == 0); \
376 j++, st = i < IDLE_WORKLOAD ? \
377 &cfqg->service_trees[i][j]: NULL) \
380 static inline enum wl_prio_t cfqq_prio(struct cfq_queue *cfqq)
382 if (cfq_class_idle(cfqq))
383 return IDLE_WORKLOAD;
384 if (cfq_class_rt(cfqq))
390 static enum wl_type_t cfqq_type(struct cfq_queue *cfqq)
392 if (!cfq_cfqq_sync(cfqq))
393 return ASYNC_WORKLOAD;
394 if (!cfq_cfqq_idle_window(cfqq))
395 return SYNC_NOIDLE_WORKLOAD;
396 return SYNC_WORKLOAD;
399 static inline int cfq_group_busy_queues_wl(enum wl_prio_t wl,
400 struct cfq_data *cfqd,
401 struct cfq_group *cfqg)
403 if (wl == IDLE_WORKLOAD)
404 return cfqg->service_tree_idle.count;
406 return cfqg->service_trees[wl][ASYNC_WORKLOAD].count
407 + cfqg->service_trees[wl][SYNC_NOIDLE_WORKLOAD].count
408 + cfqg->service_trees[wl][SYNC_WORKLOAD].count;
411 static void cfq_dispatch_insert(struct request_queue *, struct request *);
412 static struct cfq_queue *cfq_get_queue(struct cfq_data *, bool,
413 struct io_context *, gfp_t);
414 static struct cfq_io_context *cfq_cic_lookup(struct cfq_data *,
415 struct io_context *);
417 static inline int rq_in_driver(struct cfq_data *cfqd)
419 return cfqd->rq_in_driver[0] + cfqd->rq_in_driver[1];
422 static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_context *cic,
425 return cic->cfqq[is_sync];
428 static inline void cic_set_cfqq(struct cfq_io_context *cic,
429 struct cfq_queue *cfqq, bool is_sync)
431 cic->cfqq[is_sync] = cfqq;
435 * We regard a request as SYNC, if it's either a read or has the SYNC bit
436 * set (in which case it could also be direct WRITE).
438 static inline bool cfq_bio_sync(struct bio *bio)
440 return bio_data_dir(bio) == READ || bio_rw_flagged(bio, BIO_RW_SYNCIO);
444 * scheduler run of queue, if there are requests pending and no one in the
445 * driver that will restart queueing
447 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
449 if (cfqd->busy_queues) {
450 cfq_log(cfqd, "schedule dispatch");
451 kblockd_schedule_work(cfqd->queue, &cfqd->unplug_work);
455 static int cfq_queue_empty(struct request_queue *q)
457 struct cfq_data *cfqd = q->elevator->elevator_data;
459 return !cfqd->rq_queued;
463 * Scale schedule slice based on io priority. Use the sync time slice only
464 * if a queue is marked sync and has sync io queued. A sync queue with async
465 * io only, should not get full sync slice length.
467 static inline int cfq_prio_slice(struct cfq_data *cfqd, bool sync,
470 const int base_slice = cfqd->cfq_slice[sync];
472 WARN_ON(prio >= IOPRIO_BE_NR);
474 return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
478 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
480 return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
483 static inline u64 cfq_scale_slice(unsigned long delta, struct cfq_group *cfqg)
485 u64 d = delta << CFQ_SERVICE_SHIFT;
487 d = d * BLKIO_WEIGHT_DEFAULT;
488 do_div(d, cfqg->weight);
492 static inline u64 max_vdisktime(u64 min_vdisktime, u64 vdisktime)
494 s64 delta = (s64)(vdisktime - min_vdisktime);
496 min_vdisktime = vdisktime;
498 return min_vdisktime;
501 static inline u64 min_vdisktime(u64 min_vdisktime, u64 vdisktime)
503 s64 delta = (s64)(vdisktime - min_vdisktime);
505 min_vdisktime = vdisktime;
507 return min_vdisktime;
510 static void update_min_vdisktime(struct cfq_rb_root *st)
512 u64 vdisktime = st->min_vdisktime;
513 struct cfq_group *cfqg;
516 cfqg = rb_entry_cfqg(st->active);
517 vdisktime = cfqg->vdisktime;
521 cfqg = rb_entry_cfqg(st->left);
522 vdisktime = min_vdisktime(vdisktime, cfqg->vdisktime);
525 st->min_vdisktime = max_vdisktime(st->min_vdisktime, vdisktime);
529 * get averaged number of queues of RT/BE priority.
530 * average is updated, with a formula that gives more weight to higher numbers,
531 * to quickly follows sudden increases and decrease slowly
534 static inline unsigned cfq_group_get_avg_queues(struct cfq_data *cfqd,
535 struct cfq_group *cfqg, bool rt)
537 unsigned min_q, max_q;
538 unsigned mult = cfq_hist_divisor - 1;
539 unsigned round = cfq_hist_divisor / 2;
540 unsigned busy = cfq_group_busy_queues_wl(rt, cfqd, cfqg);
542 min_q = min(cfqg->busy_queues_avg[rt], busy);
543 max_q = max(cfqg->busy_queues_avg[rt], busy);
544 cfqg->busy_queues_avg[rt] = (mult * max_q + min_q + round) /
546 return cfqg->busy_queues_avg[rt];
549 static inline unsigned
550 cfq_group_slice(struct cfq_data *cfqd, struct cfq_group *cfqg)
552 struct cfq_rb_root *st = &cfqd->grp_service_tree;
554 return cfq_target_latency * cfqg->weight / st->total_weight;
558 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
560 unsigned slice = cfq_prio_to_slice(cfqd, cfqq);
561 if (cfqd->cfq_latency) {
563 * interested queues (we consider only the ones with the same
564 * priority class in the cfq group)
566 unsigned iq = cfq_group_get_avg_queues(cfqd, cfqq->cfqg,
568 unsigned sync_slice = cfqd->cfq_slice[1];
569 unsigned expect_latency = sync_slice * iq;
570 unsigned group_slice = cfq_group_slice(cfqd, cfqq->cfqg);
572 if (expect_latency > group_slice) {
573 unsigned base_low_slice = 2 * cfqd->cfq_slice_idle;
574 /* scale low_slice according to IO priority
575 * and sync vs async */
577 min(slice, base_low_slice * slice / sync_slice);
578 /* the adapted slice value is scaled to fit all iqs
579 * into the target latency */
580 slice = max(slice * group_slice / expect_latency,
584 cfqq->slice_start = jiffies;
585 cfqq->slice_end = jiffies + slice;
586 cfqq->allocated_slice = slice;
587 cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
591 * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
592 * isn't valid until the first request from the dispatch is activated
593 * and the slice time set.
595 static inline bool cfq_slice_used(struct cfq_queue *cfqq)
597 if (cfq_cfqq_slice_new(cfqq))
599 if (time_before(jiffies, cfqq->slice_end))
606 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
607 * We choose the request that is closest to the head right now. Distance
608 * behind the head is penalized and only allowed to a certain extent.
610 static struct request *
611 cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2, sector_t last)
613 sector_t s1, s2, d1 = 0, d2 = 0;
614 unsigned long back_max;
615 #define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */
616 #define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */
617 unsigned wrap = 0; /* bit mask: requests behind the disk head? */
619 if (rq1 == NULL || rq1 == rq2)
624 if (rq_is_sync(rq1) && !rq_is_sync(rq2))
626 else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
628 if (rq_is_meta(rq1) && !rq_is_meta(rq2))
630 else if (rq_is_meta(rq2) && !rq_is_meta(rq1))
633 s1 = blk_rq_pos(rq1);
634 s2 = blk_rq_pos(rq2);
637 * by definition, 1KiB is 2 sectors
639 back_max = cfqd->cfq_back_max * 2;
642 * Strict one way elevator _except_ in the case where we allow
643 * short backward seeks which are biased as twice the cost of a
644 * similar forward seek.
648 else if (s1 + back_max >= last)
649 d1 = (last - s1) * cfqd->cfq_back_penalty;
651 wrap |= CFQ_RQ1_WRAP;
655 else if (s2 + back_max >= last)
656 d2 = (last - s2) * cfqd->cfq_back_penalty;
658 wrap |= CFQ_RQ2_WRAP;
660 /* Found required data */
663 * By doing switch() on the bit mask "wrap" we avoid having to
664 * check two variables for all permutations: --> faster!
667 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
683 case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
686 * Since both rqs are wrapped,
687 * start with the one that's further behind head
688 * (--> only *one* back seek required),
689 * since back seek takes more time than forward.
699 * The below is leftmost cache rbtree addon
701 static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
703 /* Service tree is empty */
708 root->left = rb_first(&root->rb);
711 return rb_entry(root->left, struct cfq_queue, rb_node);
716 static struct cfq_group *cfq_rb_first_group(struct cfq_rb_root *root)
719 root->left = rb_first(&root->rb);
722 return rb_entry_cfqg(root->left);
727 static void rb_erase_init(struct rb_node *n, struct rb_root *root)
733 static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
737 rb_erase_init(n, &root->rb);
742 * would be nice to take fifo expire time into account as well
744 static struct request *
745 cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
746 struct request *last)
748 struct rb_node *rbnext = rb_next(&last->rb_node);
749 struct rb_node *rbprev = rb_prev(&last->rb_node);
750 struct request *next = NULL, *prev = NULL;
752 BUG_ON(RB_EMPTY_NODE(&last->rb_node));
755 prev = rb_entry_rq(rbprev);
758 next = rb_entry_rq(rbnext);
760 rbnext = rb_first(&cfqq->sort_list);
761 if (rbnext && rbnext != &last->rb_node)
762 next = rb_entry_rq(rbnext);
765 return cfq_choose_req(cfqd, next, prev, blk_rq_pos(last));
768 static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
769 struct cfq_queue *cfqq)
772 * just an approximation, should be ok.
774 return (cfqq->cfqg->nr_cfqq - 1) * (cfq_prio_slice(cfqd, 1, 0) -
775 cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
779 cfqg_key(struct cfq_rb_root *st, struct cfq_group *cfqg)
781 return cfqg->vdisktime - st->min_vdisktime;
785 __cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
787 struct rb_node **node = &st->rb.rb_node;
788 struct rb_node *parent = NULL;
789 struct cfq_group *__cfqg;
790 s64 key = cfqg_key(st, cfqg);
793 while (*node != NULL) {
795 __cfqg = rb_entry_cfqg(parent);
797 if (key < cfqg_key(st, __cfqg))
798 node = &parent->rb_left;
800 node = &parent->rb_right;
806 st->left = &cfqg->rb_node;
808 rb_link_node(&cfqg->rb_node, parent, node);
809 rb_insert_color(&cfqg->rb_node, &st->rb);
813 cfq_group_service_tree_add(struct cfq_data *cfqd, struct cfq_group *cfqg)
815 struct cfq_rb_root *st = &cfqd->grp_service_tree;
816 struct cfq_group *__cfqg;
824 * Currently put the group at the end. Later implement something
825 * so that groups get lesser vtime based on their weights, so that
826 * if group does not loose all if it was not continously backlogged.
828 n = rb_last(&st->rb);
830 __cfqg = rb_entry_cfqg(n);
831 cfqg->vdisktime = __cfqg->vdisktime + CFQ_IDLE_DELAY;
833 cfqg->vdisktime = st->min_vdisktime;
835 __cfq_group_service_tree_add(st, cfqg);
838 st->total_weight += cfqg->weight;
842 cfq_group_service_tree_del(struct cfq_data *cfqd, struct cfq_group *cfqg)
844 struct cfq_rb_root *st = &cfqd->grp_service_tree;
846 if (st->active == &cfqg->rb_node)
849 BUG_ON(cfqg->nr_cfqq < 1);
852 /* If there are other cfq queues under this group, don't delete it */
856 cfq_log_cfqg(cfqd, cfqg, "del_from_rr group");
859 st->total_weight -= cfqg->weight;
860 if (!RB_EMPTY_NODE(&cfqg->rb_node))
861 cfq_rb_erase(&cfqg->rb_node, st);
862 cfqg->saved_workload_slice = 0;
863 blkiocg_update_blkio_group_dequeue_stats(&cfqg->blkg, 1);
866 static inline unsigned int cfq_cfqq_slice_usage(struct cfq_queue *cfqq)
868 unsigned int slice_used;
871 * Queue got expired before even a single request completed or
872 * got expired immediately after first request completion.
874 if (!cfqq->slice_start || cfqq->slice_start == jiffies) {
876 * Also charge the seek time incurred to the group, otherwise
877 * if there are mutiple queues in the group, each can dispatch
878 * a single request on seeky media and cause lots of seek time
879 * and group will never know it.
881 slice_used = max_t(unsigned, (jiffies - cfqq->dispatch_start),
884 slice_used = jiffies - cfqq->slice_start;
885 if (slice_used > cfqq->allocated_slice)
886 slice_used = cfqq->allocated_slice;
889 cfq_log_cfqq(cfqq->cfqd, cfqq, "sl_used=%u sect=%lu", slice_used,
894 static void cfq_group_served(struct cfq_data *cfqd, struct cfq_group *cfqg,
895 struct cfq_queue *cfqq)
897 struct cfq_rb_root *st = &cfqd->grp_service_tree;
898 unsigned int used_sl;
900 used_sl = cfq_cfqq_slice_usage(cfqq);
902 /* Can't update vdisktime while group is on service tree */
903 cfq_rb_erase(&cfqg->rb_node, st);
904 cfqg->vdisktime += cfq_scale_slice(used_sl, cfqg);
905 __cfq_group_service_tree_add(st, cfqg);
907 /* This group is being expired. Save the context */
908 if (time_after(cfqd->workload_expires, jiffies)) {
909 cfqg->saved_workload_slice = cfqd->workload_expires
911 cfqg->saved_workload = cfqd->serving_type;
912 cfqg->saved_serving_prio = cfqd->serving_prio;
914 cfqg->saved_workload_slice = 0;
916 cfq_log_cfqg(cfqd, cfqg, "served: vt=%llu min_vt=%llu", cfqg->vdisktime,
918 blkiocg_update_blkio_group_stats(&cfqg->blkg, used_sl,
922 #ifdef CONFIG_CFQ_GROUP_IOSCHED
923 static inline struct cfq_group *cfqg_of_blkg(struct blkio_group *blkg)
926 return container_of(blkg, struct cfq_group, blkg);
931 cfq_update_blkio_group_weight(struct blkio_group *blkg, unsigned int weight)
933 cfqg_of_blkg(blkg)->weight = weight;
936 static struct cfq_group *
937 cfq_find_alloc_cfqg(struct cfq_data *cfqd, struct cgroup *cgroup, int create)
939 struct blkio_cgroup *blkcg = cgroup_to_blkio_cgroup(cgroup);
940 struct cfq_group *cfqg = NULL;
943 struct cfq_rb_root *st;
944 struct backing_dev_info *bdi = &cfqd->queue->backing_dev_info;
945 unsigned int major, minor;
947 /* Do we need to take this reference */
948 if (!css_tryget(&blkcg->css))
951 cfqg = cfqg_of_blkg(blkiocg_lookup_group(blkcg, key));
955 cfqg = kzalloc_node(sizeof(*cfqg), GFP_ATOMIC, cfqd->queue->node);
959 cfqg->weight = blkcg->weight;
960 for_each_cfqg_st(cfqg, i, j, st)
962 RB_CLEAR_NODE(&cfqg->rb_node);
965 * Take the initial reference that will be released on destroy
966 * This can be thought of a joint reference by cgroup and
967 * elevator which will be dropped by either elevator exit
968 * or cgroup deletion path depending on who is exiting first.
970 atomic_set(&cfqg->ref, 1);
972 /* Add group onto cgroup list */
973 sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
974 blkiocg_add_blkio_group(blkcg, &cfqg->blkg, (void *)cfqd,
975 MKDEV(major, minor));
977 /* Add group on cfqd list */
978 hlist_add_head(&cfqg->cfqd_node, &cfqd->cfqg_list);
981 css_put(&blkcg->css);
986 * Search for the cfq group current task belongs to. If create = 1, then also
987 * create the cfq group if it does not exist. request_queue lock must be held.
989 static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd, int create)
991 struct cgroup *cgroup;
992 struct cfq_group *cfqg = NULL;
995 cgroup = task_cgroup(current, blkio_subsys_id);
996 cfqg = cfq_find_alloc_cfqg(cfqd, cgroup, create);
998 cfqg = &cfqd->root_group;
1003 static void cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg)
1005 /* Currently, all async queues are mapped to root group */
1006 if (!cfq_cfqq_sync(cfqq))
1007 cfqg = &cfqq->cfqd->root_group;
1010 /* cfqq reference on cfqg */
1011 atomic_inc(&cfqq->cfqg->ref);
1014 static void cfq_put_cfqg(struct cfq_group *cfqg)
1016 struct cfq_rb_root *st;
1019 BUG_ON(atomic_read(&cfqg->ref) <= 0);
1020 if (!atomic_dec_and_test(&cfqg->ref))
1022 for_each_cfqg_st(cfqg, i, j, st)
1023 BUG_ON(!RB_EMPTY_ROOT(&st->rb) || st->active != NULL);
1027 static void cfq_destroy_cfqg(struct cfq_data *cfqd, struct cfq_group *cfqg)
1029 /* Something wrong if we are trying to remove same group twice */
1030 BUG_ON(hlist_unhashed(&cfqg->cfqd_node));
1032 hlist_del_init(&cfqg->cfqd_node);
1035 * Put the reference taken at the time of creation so that when all
1036 * queues are gone, group can be destroyed.
1041 static void cfq_release_cfq_groups(struct cfq_data *cfqd)
1043 struct hlist_node *pos, *n;
1044 struct cfq_group *cfqg;
1046 hlist_for_each_entry_safe(cfqg, pos, n, &cfqd->cfqg_list, cfqd_node) {
1048 * If cgroup removal path got to blk_group first and removed
1049 * it from cgroup list, then it will take care of destroying
1052 if (!blkiocg_del_blkio_group(&cfqg->blkg))
1053 cfq_destroy_cfqg(cfqd, cfqg);
1058 * Blk cgroup controller notification saying that blkio_group object is being
1059 * delinked as associated cgroup object is going away. That also means that
1060 * no new IO will come in this group. So get rid of this group as soon as
1061 * any pending IO in the group is finished.
1063 * This function is called under rcu_read_lock(). key is the rcu protected
1064 * pointer. That means "key" is a valid cfq_data pointer as long as we are rcu
1067 * "key" was fetched from blkio_group under blkio_cgroup->lock. That means
1068 * it should not be NULL as even if elevator was exiting, cgroup deltion
1069 * path got to it first.
1071 void cfq_unlink_blkio_group(void *key, struct blkio_group *blkg)
1073 unsigned long flags;
1074 struct cfq_data *cfqd = key;
1076 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1077 cfq_destroy_cfqg(cfqd, cfqg_of_blkg(blkg));
1078 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1081 #else /* GROUP_IOSCHED */
1082 static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd, int create)
1084 return &cfqd->root_group;
1087 cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) {
1091 static void cfq_release_cfq_groups(struct cfq_data *cfqd) {}
1092 static inline void cfq_put_cfqg(struct cfq_group *cfqg) {}
1094 #endif /* GROUP_IOSCHED */
1097 * The cfqd->service_trees holds all pending cfq_queue's that have
1098 * requests waiting to be processed. It is sorted in the order that
1099 * we will service the queues.
1101 static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1104 struct rb_node **p, *parent;
1105 struct cfq_queue *__cfqq;
1106 unsigned long rb_key;
1107 struct cfq_rb_root *service_tree;
1111 service_tree = service_tree_for(cfqq->cfqg, cfqq_prio(cfqq),
1112 cfqq_type(cfqq), cfqd);
1113 if (cfq_class_idle(cfqq)) {
1114 rb_key = CFQ_IDLE_DELAY;
1115 parent = rb_last(&service_tree->rb);
1116 if (parent && parent != &cfqq->rb_node) {
1117 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1118 rb_key += __cfqq->rb_key;
1121 } else if (!add_front) {
1123 * Get our rb key offset. Subtract any residual slice
1124 * value carried from last service. A negative resid
1125 * count indicates slice overrun, and this should position
1126 * the next service time further away in the tree.
1128 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
1129 rb_key -= cfqq->slice_resid;
1130 cfqq->slice_resid = 0;
1133 __cfqq = cfq_rb_first(service_tree);
1134 rb_key += __cfqq ? __cfqq->rb_key : jiffies;
1137 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1140 * same position, nothing more to do
1142 if (rb_key == cfqq->rb_key &&
1143 cfqq->service_tree == service_tree)
1146 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1147 cfqq->service_tree = NULL;
1152 cfqq->service_tree = service_tree;
1153 p = &service_tree->rb.rb_node;
1158 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1161 * sort by key, that represents service time.
1163 if (time_before(rb_key, __cfqq->rb_key))
1166 n = &(*p)->rb_right;
1174 service_tree->left = &cfqq->rb_node;
1176 cfqq->rb_key = rb_key;
1177 rb_link_node(&cfqq->rb_node, parent, p);
1178 rb_insert_color(&cfqq->rb_node, &service_tree->rb);
1179 service_tree->count++;
1180 if (add_front || !new_cfqq)
1182 cfq_group_service_tree_add(cfqd, cfqq->cfqg);
1185 static struct cfq_queue *
1186 cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
1187 sector_t sector, struct rb_node **ret_parent,
1188 struct rb_node ***rb_link)
1190 struct rb_node **p, *parent;
1191 struct cfq_queue *cfqq = NULL;
1199 cfqq = rb_entry(parent, struct cfq_queue, p_node);
1202 * Sort strictly based on sector. Smallest to the left,
1203 * largest to the right.
1205 if (sector > blk_rq_pos(cfqq->next_rq))
1206 n = &(*p)->rb_right;
1207 else if (sector < blk_rq_pos(cfqq->next_rq))
1215 *ret_parent = parent;
1221 static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1223 struct rb_node **p, *parent;
1224 struct cfq_queue *__cfqq;
1227 rb_erase(&cfqq->p_node, cfqq->p_root);
1228 cfqq->p_root = NULL;
1231 if (cfq_class_idle(cfqq))
1236 cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
1237 __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
1238 blk_rq_pos(cfqq->next_rq), &parent, &p);
1240 rb_link_node(&cfqq->p_node, parent, p);
1241 rb_insert_color(&cfqq->p_node, cfqq->p_root);
1243 cfqq->p_root = NULL;
1247 * Update cfqq's position in the service tree.
1249 static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1252 * Resorting requires the cfqq to be on the RR list already.
1254 if (cfq_cfqq_on_rr(cfqq)) {
1255 cfq_service_tree_add(cfqd, cfqq, 0);
1256 cfq_prio_tree_add(cfqd, cfqq);
1261 * add to busy list of queues for service, trying to be fair in ordering
1262 * the pending list according to last request service
1264 static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1266 cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
1267 BUG_ON(cfq_cfqq_on_rr(cfqq));
1268 cfq_mark_cfqq_on_rr(cfqq);
1269 cfqd->busy_queues++;
1271 cfq_resort_rr_list(cfqd, cfqq);
1275 * Called when the cfqq no longer has requests pending, remove it from
1278 static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1280 cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
1281 BUG_ON(!cfq_cfqq_on_rr(cfqq));
1282 cfq_clear_cfqq_on_rr(cfqq);
1284 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1285 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1286 cfqq->service_tree = NULL;
1289 rb_erase(&cfqq->p_node, cfqq->p_root);
1290 cfqq->p_root = NULL;
1293 cfq_group_service_tree_del(cfqd, cfqq->cfqg);
1294 BUG_ON(!cfqd->busy_queues);
1295 cfqd->busy_queues--;
1299 * rb tree support functions
1301 static void cfq_del_rq_rb(struct request *rq)
1303 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1304 const int sync = rq_is_sync(rq);
1306 BUG_ON(!cfqq->queued[sync]);
1307 cfqq->queued[sync]--;
1309 elv_rb_del(&cfqq->sort_list, rq);
1311 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) {
1313 * Queue will be deleted from service tree when we actually
1314 * expire it later. Right now just remove it from prio tree
1318 rb_erase(&cfqq->p_node, cfqq->p_root);
1319 cfqq->p_root = NULL;
1324 static void cfq_add_rq_rb(struct request *rq)
1326 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1327 struct cfq_data *cfqd = cfqq->cfqd;
1328 struct request *__alias, *prev;
1330 cfqq->queued[rq_is_sync(rq)]++;
1333 * looks a little odd, but the first insert might return an alias.
1334 * if that happens, put the alias on the dispatch list
1336 while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL)
1337 cfq_dispatch_insert(cfqd->queue, __alias);
1339 if (!cfq_cfqq_on_rr(cfqq))
1340 cfq_add_cfqq_rr(cfqd, cfqq);
1343 * check if this request is a better next-serve candidate
1345 prev = cfqq->next_rq;
1346 cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq, cfqd->last_position);
1349 * adjust priority tree position, if ->next_rq changes
1351 if (prev != cfqq->next_rq)
1352 cfq_prio_tree_add(cfqd, cfqq);
1354 BUG_ON(!cfqq->next_rq);
1357 static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
1359 elv_rb_del(&cfqq->sort_list, rq);
1360 cfqq->queued[rq_is_sync(rq)]--;
1364 static struct request *
1365 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
1367 struct task_struct *tsk = current;
1368 struct cfq_io_context *cic;
1369 struct cfq_queue *cfqq;
1371 cic = cfq_cic_lookup(cfqd, tsk->io_context);
1375 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
1377 sector_t sector = bio->bi_sector + bio_sectors(bio);
1379 return elv_rb_find(&cfqq->sort_list, sector);
1385 static void cfq_activate_request(struct request_queue *q, struct request *rq)
1387 struct cfq_data *cfqd = q->elevator->elevator_data;
1389 cfqd->rq_in_driver[rq_is_sync(rq)]++;
1390 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
1391 rq_in_driver(cfqd));
1393 cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
1396 static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
1398 struct cfq_data *cfqd = q->elevator->elevator_data;
1399 const int sync = rq_is_sync(rq);
1401 WARN_ON(!cfqd->rq_in_driver[sync]);
1402 cfqd->rq_in_driver[sync]--;
1403 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
1404 rq_in_driver(cfqd));
1407 static void cfq_remove_request(struct request *rq)
1409 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1411 if (cfqq->next_rq == rq)
1412 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
1414 list_del_init(&rq->queuelist);
1417 cfqq->cfqd->rq_queued--;
1418 if (rq_is_meta(rq)) {
1419 WARN_ON(!cfqq->meta_pending);
1420 cfqq->meta_pending--;
1424 static int cfq_merge(struct request_queue *q, struct request **req,
1427 struct cfq_data *cfqd = q->elevator->elevator_data;
1428 struct request *__rq;
1430 __rq = cfq_find_rq_fmerge(cfqd, bio);
1431 if (__rq && elv_rq_merge_ok(__rq, bio)) {
1433 return ELEVATOR_FRONT_MERGE;
1436 return ELEVATOR_NO_MERGE;
1439 static void cfq_merged_request(struct request_queue *q, struct request *req,
1442 if (type == ELEVATOR_FRONT_MERGE) {
1443 struct cfq_queue *cfqq = RQ_CFQQ(req);
1445 cfq_reposition_rq_rb(cfqq, req);
1450 cfq_merged_requests(struct request_queue *q, struct request *rq,
1451 struct request *next)
1453 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1455 * reposition in fifo if next is older than rq
1457 if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
1458 time_before(rq_fifo_time(next), rq_fifo_time(rq))) {
1459 list_move(&rq->queuelist, &next->queuelist);
1460 rq_set_fifo_time(rq, rq_fifo_time(next));
1463 if (cfqq->next_rq == next)
1465 cfq_remove_request(next);
1468 static int cfq_allow_merge(struct request_queue *q, struct request *rq,
1471 struct cfq_data *cfqd = q->elevator->elevator_data;
1472 struct cfq_io_context *cic;
1473 struct cfq_queue *cfqq;
1475 /* Deny merge if bio and rq don't belong to same cfq group */
1476 if ((RQ_CFQQ(rq))->cfqg != cfq_get_cfqg(cfqd, 0))
1479 * Disallow merge of a sync bio into an async request.
1481 if (cfq_bio_sync(bio) && !rq_is_sync(rq))
1485 * Lookup the cfqq that this bio will be queued with. Allow
1486 * merge only if rq is queued there.
1488 cic = cfq_cic_lookup(cfqd, current->io_context);
1492 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
1493 return cfqq == RQ_CFQQ(rq);
1496 static void __cfq_set_active_queue(struct cfq_data *cfqd,
1497 struct cfq_queue *cfqq)
1500 cfq_log_cfqq(cfqd, cfqq, "set_active");
1501 cfqq->slice_start = 0;
1502 cfqq->dispatch_start = jiffies;
1503 cfqq->allocated_slice = 0;
1504 cfqq->slice_end = 0;
1505 cfqq->slice_dispatch = 0;
1506 cfqq->nr_sectors = 0;
1508 cfq_clear_cfqq_wait_request(cfqq);
1509 cfq_clear_cfqq_must_dispatch(cfqq);
1510 cfq_clear_cfqq_must_alloc_slice(cfqq);
1511 cfq_clear_cfqq_fifo_expire(cfqq);
1512 cfq_mark_cfqq_slice_new(cfqq);
1514 del_timer(&cfqd->idle_slice_timer);
1517 cfqd->active_queue = cfqq;
1521 * current cfqq expired its slice (or was too idle), select new one
1524 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1527 cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
1529 if (cfq_cfqq_wait_request(cfqq))
1530 del_timer(&cfqd->idle_slice_timer);
1532 cfq_clear_cfqq_wait_request(cfqq);
1533 cfq_clear_cfqq_wait_busy(cfqq);
1534 cfq_clear_cfqq_wait_busy_done(cfqq);
1537 * store what was left of this slice, if the queue idled/timed out
1539 if (timed_out && !cfq_cfqq_slice_new(cfqq)) {
1540 cfqq->slice_resid = cfqq->slice_end - jiffies;
1541 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
1544 cfq_group_served(cfqd, cfqq->cfqg, cfqq);
1546 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
1547 cfq_del_cfqq_rr(cfqd, cfqq);
1549 cfq_resort_rr_list(cfqd, cfqq);
1551 if (cfqq == cfqd->active_queue)
1552 cfqd->active_queue = NULL;
1554 if (&cfqq->cfqg->rb_node == cfqd->grp_service_tree.active)
1555 cfqd->grp_service_tree.active = NULL;
1557 if (cfqd->active_cic) {
1558 put_io_context(cfqd->active_cic->ioc);
1559 cfqd->active_cic = NULL;
1563 static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out)
1565 struct cfq_queue *cfqq = cfqd->active_queue;
1568 __cfq_slice_expired(cfqd, cfqq, timed_out);
1572 * Get next queue for service. Unless we have a queue preemption,
1573 * we'll simply select the first cfqq in the service tree.
1575 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
1577 struct cfq_rb_root *service_tree =
1578 service_tree_for(cfqd->serving_group, cfqd->serving_prio,
1579 cfqd->serving_type, cfqd);
1581 if (!cfqd->rq_queued)
1584 /* There is nothing to dispatch */
1587 if (RB_EMPTY_ROOT(&service_tree->rb))
1589 return cfq_rb_first(service_tree);
1592 static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd)
1594 struct cfq_group *cfqg;
1595 struct cfq_queue *cfqq;
1597 struct cfq_rb_root *st;
1599 if (!cfqd->rq_queued)
1602 cfqg = cfq_get_next_cfqg(cfqd);
1606 for_each_cfqg_st(cfqg, i, j, st)
1607 if ((cfqq = cfq_rb_first(st)) != NULL)
1613 * Get and set a new active queue for service.
1615 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
1616 struct cfq_queue *cfqq)
1619 cfqq = cfq_get_next_queue(cfqd);
1621 __cfq_set_active_queue(cfqd, cfqq);
1625 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
1628 if (blk_rq_pos(rq) >= cfqd->last_position)
1629 return blk_rq_pos(rq) - cfqd->last_position;
1631 return cfqd->last_position - blk_rq_pos(rq);
1634 #define CFQQ_SEEK_THR 8 * 1024
1635 #define CFQQ_SEEKY(cfqq) ((cfqq)->seek_mean > CFQQ_SEEK_THR)
1637 static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1640 sector_t sdist = cfqq->seek_mean;
1642 if (!sample_valid(cfqq->seek_samples))
1643 sdist = CFQQ_SEEK_THR;
1645 return cfq_dist_from_last(cfqd, rq) <= sdist;
1648 static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
1649 struct cfq_queue *cur_cfqq)
1651 struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
1652 struct rb_node *parent, *node;
1653 struct cfq_queue *__cfqq;
1654 sector_t sector = cfqd->last_position;
1656 if (RB_EMPTY_ROOT(root))
1660 * First, if we find a request starting at the end of the last
1661 * request, choose it.
1663 __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
1668 * If the exact sector wasn't found, the parent of the NULL leaf
1669 * will contain the closest sector.
1671 __cfqq = rb_entry(parent, struct cfq_queue, p_node);
1672 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1675 if (blk_rq_pos(__cfqq->next_rq) < sector)
1676 node = rb_next(&__cfqq->p_node);
1678 node = rb_prev(&__cfqq->p_node);
1682 __cfqq = rb_entry(node, struct cfq_queue, p_node);
1683 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1691 * cur_cfqq - passed in so that we don't decide that the current queue is
1692 * closely cooperating with itself.
1694 * So, basically we're assuming that that cur_cfqq has dispatched at least
1695 * one request, and that cfqd->last_position reflects a position on the disk
1696 * associated with the I/O issued by cur_cfqq. I'm not sure this is a valid
1699 static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
1700 struct cfq_queue *cur_cfqq)
1702 struct cfq_queue *cfqq;
1704 if (!cfq_cfqq_sync(cur_cfqq))
1706 if (CFQQ_SEEKY(cur_cfqq))
1710 * We should notice if some of the queues are cooperating, eg
1711 * working closely on the same area of the disk. In that case,
1712 * we can group them together and don't waste time idling.
1714 cfqq = cfqq_close(cfqd, cur_cfqq);
1718 /* If new queue belongs to different cfq_group, don't choose it */
1719 if (cur_cfqq->cfqg != cfqq->cfqg)
1723 * It only makes sense to merge sync queues.
1725 if (!cfq_cfqq_sync(cfqq))
1727 if (CFQQ_SEEKY(cfqq))
1731 * Do not merge queues of different priority classes
1733 if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq))
1740 * Determine whether we should enforce idle window for this queue.
1743 static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1745 enum wl_prio_t prio = cfqq_prio(cfqq);
1746 struct cfq_rb_root *service_tree = cfqq->service_tree;
1748 BUG_ON(!service_tree);
1749 BUG_ON(!service_tree->count);
1751 /* We never do for idle class queues. */
1752 if (prio == IDLE_WORKLOAD)
1755 /* We do for queues that were marked with idle window flag. */
1756 if (cfq_cfqq_idle_window(cfqq))
1760 * Otherwise, we do only if they are the last ones
1761 * in their service tree.
1763 return service_tree->count == 1;
1766 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
1768 struct cfq_queue *cfqq = cfqd->active_queue;
1769 struct cfq_io_context *cic;
1773 * SSD device without seek penalty, disable idling. But only do so
1774 * for devices that support queuing, otherwise we still have a problem
1775 * with sync vs async workloads.
1777 if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
1780 WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
1781 WARN_ON(cfq_cfqq_slice_new(cfqq));
1784 * idle is disabled, either manually or by past process history
1786 if (!cfqd->cfq_slice_idle || !cfq_should_idle(cfqd, cfqq))
1790 * still active requests from this queue, don't idle
1792 if (cfqq->dispatched)
1796 * task has exited, don't wait
1798 cic = cfqd->active_cic;
1799 if (!cic || !atomic_read(&cic->ioc->nr_tasks))
1803 * If our average think time is larger than the remaining time
1804 * slice, then don't idle. This avoids overrunning the allotted
1807 if (sample_valid(cic->ttime_samples) &&
1808 (cfqq->slice_end - jiffies < cic->ttime_mean))
1811 cfq_mark_cfqq_wait_request(cfqq);
1813 sl = cfqd->cfq_slice_idle;
1815 mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
1816 cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu", sl);
1820 * Move request from internal lists to the request queue dispatch list.
1822 static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
1824 struct cfq_data *cfqd = q->elevator->elevator_data;
1825 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1827 cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
1829 cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
1830 cfq_remove_request(rq);
1832 elv_dispatch_sort(q, rq);
1834 if (cfq_cfqq_sync(cfqq))
1835 cfqd->sync_flight++;
1836 cfqq->nr_sectors += blk_rq_sectors(rq);
1840 * return expired entry, or NULL to just start from scratch in rbtree
1842 static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
1844 struct request *rq = NULL;
1846 if (cfq_cfqq_fifo_expire(cfqq))
1849 cfq_mark_cfqq_fifo_expire(cfqq);
1851 if (list_empty(&cfqq->fifo))
1854 rq = rq_entry_fifo(cfqq->fifo.next);
1855 if (time_before(jiffies, rq_fifo_time(rq)))
1858 cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq);
1863 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1865 const int base_rq = cfqd->cfq_slice_async_rq;
1867 WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
1869 return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
1873 * Must be called with the queue_lock held.
1875 static int cfqq_process_refs(struct cfq_queue *cfqq)
1877 int process_refs, io_refs;
1879 io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE];
1880 process_refs = atomic_read(&cfqq->ref) - io_refs;
1881 BUG_ON(process_refs < 0);
1882 return process_refs;
1885 static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq)
1887 int process_refs, new_process_refs;
1888 struct cfq_queue *__cfqq;
1890 /* Avoid a circular list and skip interim queue merges */
1891 while ((__cfqq = new_cfqq->new_cfqq)) {
1897 process_refs = cfqq_process_refs(cfqq);
1899 * If the process for the cfqq has gone away, there is no
1900 * sense in merging the queues.
1902 if (process_refs == 0)
1906 * Merge in the direction of the lesser amount of work.
1908 new_process_refs = cfqq_process_refs(new_cfqq);
1909 if (new_process_refs >= process_refs) {
1910 cfqq->new_cfqq = new_cfqq;
1911 atomic_add(process_refs, &new_cfqq->ref);
1913 new_cfqq->new_cfqq = cfqq;
1914 atomic_add(new_process_refs, &cfqq->ref);
1918 static enum wl_type_t cfq_choose_wl(struct cfq_data *cfqd,
1919 struct cfq_group *cfqg, enum wl_prio_t prio,
1922 struct cfq_queue *queue;
1924 bool key_valid = false;
1925 unsigned long lowest_key = 0;
1926 enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD;
1930 * When priorities switched, we prefer starting
1931 * from SYNC_NOIDLE (first choice), or just SYNC
1934 if (service_tree_for(cfqg, prio, cur_best, cfqd)->count)
1936 cur_best = SYNC_WORKLOAD;
1937 if (service_tree_for(cfqg, prio, cur_best, cfqd)->count)
1940 return ASYNC_WORKLOAD;
1943 for (i = 0; i < 3; ++i) {
1944 /* otherwise, select the one with lowest rb_key */
1945 queue = cfq_rb_first(service_tree_for(cfqg, prio, i, cfqd));
1947 (!key_valid || time_before(queue->rb_key, lowest_key))) {
1948 lowest_key = queue->rb_key;
1957 static void choose_service_tree(struct cfq_data *cfqd, struct cfq_group *cfqg)
1959 enum wl_prio_t previous_prio = cfqd->serving_prio;
1963 struct cfq_rb_root *st;
1964 unsigned group_slice;
1967 cfqd->serving_prio = IDLE_WORKLOAD;
1968 cfqd->workload_expires = jiffies + 1;
1972 /* Choose next priority. RT > BE > IDLE */
1973 if (cfq_group_busy_queues_wl(RT_WORKLOAD, cfqd, cfqg))
1974 cfqd->serving_prio = RT_WORKLOAD;
1975 else if (cfq_group_busy_queues_wl(BE_WORKLOAD, cfqd, cfqg))
1976 cfqd->serving_prio = BE_WORKLOAD;
1978 cfqd->serving_prio = IDLE_WORKLOAD;
1979 cfqd->workload_expires = jiffies + 1;
1984 * For RT and BE, we have to choose also the type
1985 * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
1988 prio_changed = (cfqd->serving_prio != previous_prio);
1989 st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type,
1994 * If priority didn't change, check workload expiration,
1995 * and that we still have other queues ready
1997 if (!prio_changed && count &&
1998 !time_after(jiffies, cfqd->workload_expires))
2001 /* otherwise select new workload type */
2002 cfqd->serving_type =
2003 cfq_choose_wl(cfqd, cfqg, cfqd->serving_prio, prio_changed);
2004 st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type,
2009 * the workload slice is computed as a fraction of target latency
2010 * proportional to the number of queues in that workload, over
2011 * all the queues in the same priority class
2013 group_slice = cfq_group_slice(cfqd, cfqg);
2015 slice = group_slice * count /
2016 max_t(unsigned, cfqg->busy_queues_avg[cfqd->serving_prio],
2017 cfq_group_busy_queues_wl(cfqd->serving_prio, cfqd, cfqg));
2019 if (cfqd->serving_type == ASYNC_WORKLOAD)
2020 /* async workload slice is scaled down according to
2021 * the sync/async slice ratio. */
2022 slice = slice * cfqd->cfq_slice[0] / cfqd->cfq_slice[1];
2024 /* sync workload slice is at least 2 * cfq_slice_idle */
2025 slice = max(slice, 2 * cfqd->cfq_slice_idle);
2027 slice = max_t(unsigned, slice, CFQ_MIN_TT);
2028 cfqd->workload_expires = jiffies + slice;
2029 cfqd->noidle_tree_requires_idle = false;
2032 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd)
2034 struct cfq_rb_root *st = &cfqd->grp_service_tree;
2035 struct cfq_group *cfqg;
2037 if (RB_EMPTY_ROOT(&st->rb))
2039 cfqg = cfq_rb_first_group(st);
2040 st->active = &cfqg->rb_node;
2041 update_min_vdisktime(st);
2045 static void cfq_choose_cfqg(struct cfq_data *cfqd)
2047 struct cfq_group *cfqg = cfq_get_next_cfqg(cfqd);
2049 cfqd->serving_group = cfqg;
2051 /* Restore the workload type data */
2052 if (cfqg->saved_workload_slice) {
2053 cfqd->workload_expires = jiffies + cfqg->saved_workload_slice;
2054 cfqd->serving_type = cfqg->saved_workload;
2055 cfqd->serving_prio = cfqg->saved_serving_prio;
2057 choose_service_tree(cfqd, cfqg);
2061 * Select a queue for service. If we have a current active queue,
2062 * check whether to continue servicing it, or retrieve and set a new one.
2064 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
2066 struct cfq_queue *cfqq, *new_cfqq = NULL;
2068 cfqq = cfqd->active_queue;
2072 if (!cfqd->rq_queued)
2075 * The active queue has run out of time, expire it and select new.
2077 if ((cfq_slice_used(cfqq) || cfq_cfqq_wait_busy_done(cfqq))
2078 && !cfq_cfqq_must_dispatch(cfqq))
2082 * The active queue has requests and isn't expired, allow it to
2085 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
2089 * If another queue has a request waiting within our mean seek
2090 * distance, let it run. The expire code will check for close
2091 * cooperators and put the close queue at the front of the service
2092 * tree. If possible, merge the expiring queue with the new cfqq.
2094 new_cfqq = cfq_close_cooperator(cfqd, cfqq);
2096 if (!cfqq->new_cfqq)
2097 cfq_setup_merge(cfqq, new_cfqq);
2102 * No requests pending. If the active queue still has requests in
2103 * flight or is idling for a new request, allow either of these
2104 * conditions to happen (or time out) before selecting a new queue.
2106 if (timer_pending(&cfqd->idle_slice_timer) ||
2107 (cfqq->dispatched && cfq_should_idle(cfqd, cfqq))) {
2113 cfq_slice_expired(cfqd, 0);
2116 * Current queue expired. Check if we have to switch to a new
2120 cfq_choose_cfqg(cfqd);
2122 cfqq = cfq_set_active_queue(cfqd, new_cfqq);
2127 static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
2131 while (cfqq->next_rq) {
2132 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
2136 BUG_ON(!list_empty(&cfqq->fifo));
2138 /* By default cfqq is not expired if it is empty. Do it explicitly */
2139 __cfq_slice_expired(cfqq->cfqd, cfqq, 0);
2144 * Drain our current requests. Used for barriers and when switching
2145 * io schedulers on-the-fly.
2147 static int cfq_forced_dispatch(struct cfq_data *cfqd)
2149 struct cfq_queue *cfqq;
2152 while ((cfqq = cfq_get_next_queue_forced(cfqd)) != NULL)
2153 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
2155 cfq_slice_expired(cfqd, 0);
2156 BUG_ON(cfqd->busy_queues);
2158 cfq_log(cfqd, "forced_dispatch=%d", dispatched);
2162 static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2164 unsigned int max_dispatch;
2167 * Drain async requests before we start sync IO
2169 if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_driver[BLK_RW_ASYNC])
2173 * If this is an async queue and we have sync IO in flight, let it wait
2175 if (cfqd->sync_flight && !cfq_cfqq_sync(cfqq))
2178 max_dispatch = cfqd->cfq_quantum;
2179 if (cfq_class_idle(cfqq))
2183 * Does this cfqq already have too much IO in flight?
2185 if (cfqq->dispatched >= max_dispatch) {
2187 * idle queue must always only have a single IO in flight
2189 if (cfq_class_idle(cfqq))
2193 * We have other queues, don't allow more IO from this one
2195 if (cfqd->busy_queues > 1)
2199 * Sole queue user, no limit
2205 * Async queues must wait a bit before being allowed dispatch.
2206 * We also ramp up the dispatch depth gradually for async IO,
2207 * based on the last sync IO we serviced
2209 if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
2210 unsigned long last_sync = jiffies - cfqd->last_end_sync_rq;
2213 depth = last_sync / cfqd->cfq_slice[1];
2214 if (!depth && !cfqq->dispatched)
2216 if (depth < max_dispatch)
2217 max_dispatch = depth;
2221 * If we're below the current max, allow a dispatch
2223 return cfqq->dispatched < max_dispatch;
2227 * Dispatch a request from cfqq, moving them to the request queue
2230 static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2234 BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
2236 if (!cfq_may_dispatch(cfqd, cfqq))
2240 * follow expired path, else get first next available
2242 rq = cfq_check_fifo(cfqq);
2247 * insert request into driver dispatch list
2249 cfq_dispatch_insert(cfqd->queue, rq);
2251 if (!cfqd->active_cic) {
2252 struct cfq_io_context *cic = RQ_CIC(rq);
2254 atomic_long_inc(&cic->ioc->refcount);
2255 cfqd->active_cic = cic;
2262 * Find the cfqq that we need to service and move a request from that to the
2265 static int cfq_dispatch_requests(struct request_queue *q, int force)
2267 struct cfq_data *cfqd = q->elevator->elevator_data;
2268 struct cfq_queue *cfqq;
2270 if (!cfqd->busy_queues)
2273 if (unlikely(force))
2274 return cfq_forced_dispatch(cfqd);
2276 cfqq = cfq_select_queue(cfqd);
2281 * Dispatch a request from this cfqq, if it is allowed
2283 if (!cfq_dispatch_request(cfqd, cfqq))
2286 cfqq->slice_dispatch++;
2287 cfq_clear_cfqq_must_dispatch(cfqq);
2290 * expire an async queue immediately if it has used up its slice. idle
2291 * queue always expire after 1 dispatch round.
2293 if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
2294 cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
2295 cfq_class_idle(cfqq))) {
2296 cfqq->slice_end = jiffies + 1;
2297 cfq_slice_expired(cfqd, 0);
2300 cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
2305 * task holds one reference to the queue, dropped when task exits. each rq
2306 * in-flight on this queue also holds a reference, dropped when rq is freed.
2308 * Each cfq queue took a reference on the parent group. Drop it now.
2309 * queue lock must be held here.
2311 static void cfq_put_queue(struct cfq_queue *cfqq)
2313 struct cfq_data *cfqd = cfqq->cfqd;
2314 struct cfq_group *cfqg;
2316 BUG_ON(atomic_read(&cfqq->ref) <= 0);
2318 if (!atomic_dec_and_test(&cfqq->ref))
2321 cfq_log_cfqq(cfqd, cfqq, "put_queue");
2322 BUG_ON(rb_first(&cfqq->sort_list));
2323 BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
2326 if (unlikely(cfqd->active_queue == cfqq)) {
2327 __cfq_slice_expired(cfqd, cfqq, 0);
2328 cfq_schedule_dispatch(cfqd);
2331 BUG_ON(cfq_cfqq_on_rr(cfqq));
2332 kmem_cache_free(cfq_pool, cfqq);
2337 * Must always be called with the rcu_read_lock() held
2340 __call_for_each_cic(struct io_context *ioc,
2341 void (*func)(struct io_context *, struct cfq_io_context *))
2343 struct cfq_io_context *cic;
2344 struct hlist_node *n;
2346 hlist_for_each_entry_rcu(cic, n, &ioc->cic_list, cic_list)
2351 * Call func for each cic attached to this ioc.
2354 call_for_each_cic(struct io_context *ioc,
2355 void (*func)(struct io_context *, struct cfq_io_context *))
2358 __call_for_each_cic(ioc, func);
2362 static void cfq_cic_free_rcu(struct rcu_head *head)
2364 struct cfq_io_context *cic;
2366 cic = container_of(head, struct cfq_io_context, rcu_head);
2368 kmem_cache_free(cfq_ioc_pool, cic);
2369 elv_ioc_count_dec(cfq_ioc_count);
2373 * CFQ scheduler is exiting, grab exit lock and check
2374 * the pending io context count. If it hits zero,
2375 * complete ioc_gone and set it back to NULL
2377 spin_lock(&ioc_gone_lock);
2378 if (ioc_gone && !elv_ioc_count_read(cfq_ioc_count)) {
2382 spin_unlock(&ioc_gone_lock);
2386 static void cfq_cic_free(struct cfq_io_context *cic)
2388 call_rcu(&cic->rcu_head, cfq_cic_free_rcu);
2391 static void cic_free_func(struct io_context *ioc, struct cfq_io_context *cic)
2393 unsigned long flags;
2395 BUG_ON(!cic->dead_key);
2397 spin_lock_irqsave(&ioc->lock, flags);
2398 radix_tree_delete(&ioc->radix_root, cic->dead_key);
2399 hlist_del_rcu(&cic->cic_list);
2400 spin_unlock_irqrestore(&ioc->lock, flags);
2406 * Must be called with rcu_read_lock() held or preemption otherwise disabled.
2407 * Only two callers of this - ->dtor() which is called with the rcu_read_lock(),
2408 * and ->trim() which is called with the task lock held
2410 static void cfq_free_io_context(struct io_context *ioc)
2413 * ioc->refcount is zero here, or we are called from elv_unregister(),
2414 * so no more cic's are allowed to be linked into this ioc. So it
2415 * should be ok to iterate over the known list, we will see all cic's
2416 * since no new ones are added.
2418 __call_for_each_cic(ioc, cic_free_func);
2421 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2423 struct cfq_queue *__cfqq, *next;
2425 if (unlikely(cfqq == cfqd->active_queue)) {
2426 __cfq_slice_expired(cfqd, cfqq, 0);
2427 cfq_schedule_dispatch(cfqd);
2431 * If this queue was scheduled to merge with another queue, be
2432 * sure to drop the reference taken on that queue (and others in
2433 * the merge chain). See cfq_setup_merge and cfq_merge_cfqqs.
2435 __cfqq = cfqq->new_cfqq;
2437 if (__cfqq == cfqq) {
2438 WARN(1, "cfqq->new_cfqq loop detected\n");
2441 next = __cfqq->new_cfqq;
2442 cfq_put_queue(__cfqq);
2446 cfq_put_queue(cfqq);
2449 static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
2450 struct cfq_io_context *cic)
2452 struct io_context *ioc = cic->ioc;
2454 list_del_init(&cic->queue_list);
2457 * Make sure key == NULL is seen for dead queues
2460 cic->dead_key = (unsigned long) cic->key;
2463 if (ioc->ioc_data == cic)
2464 rcu_assign_pointer(ioc->ioc_data, NULL);
2466 if (cic->cfqq[BLK_RW_ASYNC]) {
2467 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]);
2468 cic->cfqq[BLK_RW_ASYNC] = NULL;
2471 if (cic->cfqq[BLK_RW_SYNC]) {
2472 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]);
2473 cic->cfqq[BLK_RW_SYNC] = NULL;
2477 static void cfq_exit_single_io_context(struct io_context *ioc,
2478 struct cfq_io_context *cic)
2480 struct cfq_data *cfqd = cic->key;
2483 struct request_queue *q = cfqd->queue;
2484 unsigned long flags;
2486 spin_lock_irqsave(q->queue_lock, flags);
2489 * Ensure we get a fresh copy of the ->key to prevent
2490 * race between exiting task and queue
2492 smp_read_barrier_depends();
2494 __cfq_exit_single_io_context(cfqd, cic);
2496 spin_unlock_irqrestore(q->queue_lock, flags);
2501 * The process that ioc belongs to has exited, we need to clean up
2502 * and put the internal structures we have that belongs to that process.
2504 static void cfq_exit_io_context(struct io_context *ioc)
2506 call_for_each_cic(ioc, cfq_exit_single_io_context);
2509 static struct cfq_io_context *
2510 cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
2512 struct cfq_io_context *cic;
2514 cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO,
2517 cic->last_end_request = jiffies;
2518 INIT_LIST_HEAD(&cic->queue_list);
2519 INIT_HLIST_NODE(&cic->cic_list);
2520 cic->dtor = cfq_free_io_context;
2521 cic->exit = cfq_exit_io_context;
2522 elv_ioc_count_inc(cfq_ioc_count);
2528 static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc)
2530 struct task_struct *tsk = current;
2533 if (!cfq_cfqq_prio_changed(cfqq))
2536 ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio);
2537 switch (ioprio_class) {
2539 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
2540 case IOPRIO_CLASS_NONE:
2542 * no prio set, inherit CPU scheduling settings
2544 cfqq->ioprio = task_nice_ioprio(tsk);
2545 cfqq->ioprio_class = task_nice_ioclass(tsk);
2547 case IOPRIO_CLASS_RT:
2548 cfqq->ioprio = task_ioprio(ioc);
2549 cfqq->ioprio_class = IOPRIO_CLASS_RT;
2551 case IOPRIO_CLASS_BE:
2552 cfqq->ioprio = task_ioprio(ioc);
2553 cfqq->ioprio_class = IOPRIO_CLASS_BE;
2555 case IOPRIO_CLASS_IDLE:
2556 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
2558 cfq_clear_cfqq_idle_window(cfqq);
2563 * keep track of original prio settings in case we have to temporarily
2564 * elevate the priority of this queue
2566 cfqq->org_ioprio = cfqq->ioprio;
2567 cfqq->org_ioprio_class = cfqq->ioprio_class;
2568 cfq_clear_cfqq_prio_changed(cfqq);
2571 static void changed_ioprio(struct io_context *ioc, struct cfq_io_context *cic)
2573 struct cfq_data *cfqd = cic->key;
2574 struct cfq_queue *cfqq;
2575 unsigned long flags;
2577 if (unlikely(!cfqd))
2580 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2582 cfqq = cic->cfqq[BLK_RW_ASYNC];
2584 struct cfq_queue *new_cfqq;
2585 new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic->ioc,
2588 cic->cfqq[BLK_RW_ASYNC] = new_cfqq;
2589 cfq_put_queue(cfqq);
2593 cfqq = cic->cfqq[BLK_RW_SYNC];
2595 cfq_mark_cfqq_prio_changed(cfqq);
2597 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2600 static void cfq_ioc_set_ioprio(struct io_context *ioc)
2602 call_for_each_cic(ioc, changed_ioprio);
2603 ioc->ioprio_changed = 0;
2606 static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2607 pid_t pid, bool is_sync)
2609 RB_CLEAR_NODE(&cfqq->rb_node);
2610 RB_CLEAR_NODE(&cfqq->p_node);
2611 INIT_LIST_HEAD(&cfqq->fifo);
2613 atomic_set(&cfqq->ref, 0);
2616 cfq_mark_cfqq_prio_changed(cfqq);
2619 if (!cfq_class_idle(cfqq))
2620 cfq_mark_cfqq_idle_window(cfqq);
2621 cfq_mark_cfqq_sync(cfqq);
2626 #ifdef CONFIG_CFQ_GROUP_IOSCHED
2627 static void changed_cgroup(struct io_context *ioc, struct cfq_io_context *cic)
2629 struct cfq_queue *sync_cfqq = cic_to_cfqq(cic, 1);
2630 struct cfq_data *cfqd = cic->key;
2631 unsigned long flags;
2632 struct request_queue *q;
2634 if (unlikely(!cfqd))
2639 spin_lock_irqsave(q->queue_lock, flags);
2643 * Drop reference to sync queue. A new sync queue will be
2644 * assigned in new group upon arrival of a fresh request.
2646 cfq_log_cfqq(cfqd, sync_cfqq, "changed cgroup");
2647 cic_set_cfqq(cic, NULL, 1);
2648 cfq_put_queue(sync_cfqq);
2651 spin_unlock_irqrestore(q->queue_lock, flags);
2654 static void cfq_ioc_set_cgroup(struct io_context *ioc)
2656 call_for_each_cic(ioc, changed_cgroup);
2657 ioc->cgroup_changed = 0;
2659 #endif /* CONFIG_CFQ_GROUP_IOSCHED */
2661 static struct cfq_queue *
2662 cfq_find_alloc_queue(struct cfq_data *cfqd, bool is_sync,
2663 struct io_context *ioc, gfp_t gfp_mask)
2665 struct cfq_queue *cfqq, *new_cfqq = NULL;
2666 struct cfq_io_context *cic;
2667 struct cfq_group *cfqg;
2670 cfqg = cfq_get_cfqg(cfqd, 1);
2671 cic = cfq_cic_lookup(cfqd, ioc);
2672 /* cic always exists here */
2673 cfqq = cic_to_cfqq(cic, is_sync);
2676 * Always try a new alloc if we fell back to the OOM cfqq
2677 * originally, since it should just be a temporary situation.
2679 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
2684 } else if (gfp_mask & __GFP_WAIT) {
2685 spin_unlock_irq(cfqd->queue->queue_lock);
2686 new_cfqq = kmem_cache_alloc_node(cfq_pool,
2687 gfp_mask | __GFP_ZERO,
2689 spin_lock_irq(cfqd->queue->queue_lock);
2693 cfqq = kmem_cache_alloc_node(cfq_pool,
2694 gfp_mask | __GFP_ZERO,
2699 cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
2700 cfq_init_prio_data(cfqq, ioc);
2701 cfq_link_cfqq_cfqg(cfqq, cfqg);
2702 cfq_log_cfqq(cfqd, cfqq, "alloced");
2704 cfqq = &cfqd->oom_cfqq;
2708 kmem_cache_free(cfq_pool, new_cfqq);
2713 static struct cfq_queue **
2714 cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
2716 switch (ioprio_class) {
2717 case IOPRIO_CLASS_RT:
2718 return &cfqd->async_cfqq[0][ioprio];
2719 case IOPRIO_CLASS_BE:
2720 return &cfqd->async_cfqq[1][ioprio];
2721 case IOPRIO_CLASS_IDLE:
2722 return &cfqd->async_idle_cfqq;
2728 static struct cfq_queue *
2729 cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct io_context *ioc,
2732 const int ioprio = task_ioprio(ioc);
2733 const int ioprio_class = task_ioprio_class(ioc);
2734 struct cfq_queue **async_cfqq = NULL;
2735 struct cfq_queue *cfqq = NULL;
2738 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
2743 cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask);
2746 * pin the queue now that it's allocated, scheduler exit will prune it
2748 if (!is_sync && !(*async_cfqq)) {
2749 atomic_inc(&cfqq->ref);
2753 atomic_inc(&cfqq->ref);
2758 * We drop cfq io contexts lazily, so we may find a dead one.
2761 cfq_drop_dead_cic(struct cfq_data *cfqd, struct io_context *ioc,
2762 struct cfq_io_context *cic)
2764 unsigned long flags;
2766 WARN_ON(!list_empty(&cic->queue_list));
2768 spin_lock_irqsave(&ioc->lock, flags);
2770 BUG_ON(ioc->ioc_data == cic);
2772 radix_tree_delete(&ioc->radix_root, (unsigned long) cfqd);
2773 hlist_del_rcu(&cic->cic_list);
2774 spin_unlock_irqrestore(&ioc->lock, flags);
2779 static struct cfq_io_context *
2780 cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc)
2782 struct cfq_io_context *cic;
2783 unsigned long flags;
2792 * we maintain a last-hit cache, to avoid browsing over the tree
2794 cic = rcu_dereference(ioc->ioc_data);
2795 if (cic && cic->key == cfqd) {
2801 cic = radix_tree_lookup(&ioc->radix_root, (unsigned long) cfqd);
2805 /* ->key must be copied to avoid race with cfq_exit_queue() */
2808 cfq_drop_dead_cic(cfqd, ioc, cic);
2813 spin_lock_irqsave(&ioc->lock, flags);
2814 rcu_assign_pointer(ioc->ioc_data, cic);
2815 spin_unlock_irqrestore(&ioc->lock, flags);
2823 * Add cic into ioc, using cfqd as the search key. This enables us to lookup
2824 * the process specific cfq io context when entered from the block layer.
2825 * Also adds the cic to a per-cfqd list, used when this queue is removed.
2827 static int cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
2828 struct cfq_io_context *cic, gfp_t gfp_mask)
2830 unsigned long flags;
2833 ret = radix_tree_preload(gfp_mask);
2838 spin_lock_irqsave(&ioc->lock, flags);
2839 ret = radix_tree_insert(&ioc->radix_root,
2840 (unsigned long) cfqd, cic);
2842 hlist_add_head_rcu(&cic->cic_list, &ioc->cic_list);
2843 spin_unlock_irqrestore(&ioc->lock, flags);
2845 radix_tree_preload_end();
2848 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2849 list_add(&cic->queue_list, &cfqd->cic_list);
2850 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2855 printk(KERN_ERR "cfq: cic link failed!\n");
2861 * Setup general io context and cfq io context. There can be several cfq
2862 * io contexts per general io context, if this process is doing io to more
2863 * than one device managed by cfq.
2865 static struct cfq_io_context *
2866 cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
2868 struct io_context *ioc = NULL;
2869 struct cfq_io_context *cic;
2871 might_sleep_if(gfp_mask & __GFP_WAIT);
2873 ioc = get_io_context(gfp_mask, cfqd->queue->node);
2877 cic = cfq_cic_lookup(cfqd, ioc);
2881 cic = cfq_alloc_io_context(cfqd, gfp_mask);
2885 if (cfq_cic_link(cfqd, ioc, cic, gfp_mask))
2889 smp_read_barrier_depends();
2890 if (unlikely(ioc->ioprio_changed))
2891 cfq_ioc_set_ioprio(ioc);
2893 #ifdef CONFIG_CFQ_GROUP_IOSCHED
2894 if (unlikely(ioc->cgroup_changed))
2895 cfq_ioc_set_cgroup(ioc);
2901 put_io_context(ioc);
2906 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
2908 unsigned long elapsed = jiffies - cic->last_end_request;
2909 unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
2911 cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
2912 cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
2913 cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
2917 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2923 if (!cfqq->last_request_pos)
2925 else if (cfqq->last_request_pos < blk_rq_pos(rq))
2926 sdist = blk_rq_pos(rq) - cfqq->last_request_pos;
2928 sdist = cfqq->last_request_pos - blk_rq_pos(rq);
2931 * Don't allow the seek distance to get too large from the
2932 * odd fragment, pagein, etc
2934 if (cfqq->seek_samples <= 60) /* second&third seek */
2935 sdist = min(sdist, (cfqq->seek_mean * 4) + 2*1024*1024);
2937 sdist = min(sdist, (cfqq->seek_mean * 4) + 2*1024*64);
2939 cfqq->seek_samples = (7*cfqq->seek_samples + 256) / 8;
2940 cfqq->seek_total = (7*cfqq->seek_total + (u64)256*sdist) / 8;
2941 total = cfqq->seek_total + (cfqq->seek_samples/2);
2942 do_div(total, cfqq->seek_samples);
2943 cfqq->seek_mean = (sector_t)total;
2946 * If this cfqq is shared between multiple processes, check to
2947 * make sure that those processes are still issuing I/Os within
2948 * the mean seek distance. If not, it may be time to break the
2949 * queues apart again.
2951 if (cfq_cfqq_coop(cfqq)) {
2952 if (CFQQ_SEEKY(cfqq) && !cfqq->seeky_start)
2953 cfqq->seeky_start = jiffies;
2954 else if (!CFQQ_SEEKY(cfqq))
2955 cfqq->seeky_start = 0;
2960 * Disable idle window if the process thinks too long or seeks so much that
2964 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2965 struct cfq_io_context *cic)
2967 int old_idle, enable_idle;
2970 * Don't idle for async or idle io prio class
2972 if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
2975 enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
2977 if (cfqq->queued[0] + cfqq->queued[1] >= 4)
2978 cfq_mark_cfqq_deep(cfqq);
2980 if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle ||
2981 (!cfq_cfqq_deep(cfqq) && sample_valid(cfqq->seek_samples)
2982 && CFQQ_SEEKY(cfqq)))
2984 else if (sample_valid(cic->ttime_samples)) {
2985 if (cic->ttime_mean > cfqd->cfq_slice_idle)
2991 if (old_idle != enable_idle) {
2992 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
2994 cfq_mark_cfqq_idle_window(cfqq);
2996 cfq_clear_cfqq_idle_window(cfqq);
3001 * Check if new_cfqq should preempt the currently active queue. Return 0 for
3002 * no or if we aren't sure, a 1 will cause a preempt.
3005 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
3008 struct cfq_queue *cfqq;
3010 cfqq = cfqd->active_queue;
3014 if (cfq_class_idle(new_cfqq))
3017 if (cfq_class_idle(cfqq))
3021 * if the new request is sync, but the currently running queue is
3022 * not, let the sync request have priority.
3024 if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
3027 if (new_cfqq->cfqg != cfqq->cfqg)
3030 if (cfq_slice_used(cfqq))
3033 /* Allow preemption only if we are idling on sync-noidle tree */
3034 if (cfqd->serving_type == SYNC_NOIDLE_WORKLOAD &&
3035 cfqq_type(new_cfqq) == SYNC_NOIDLE_WORKLOAD &&
3036 new_cfqq->service_tree->count == 2 &&
3037 RB_EMPTY_ROOT(&cfqq->sort_list))
3041 * So both queues are sync. Let the new request get disk time if
3042 * it's a metadata request and the current queue is doing regular IO.
3044 if (rq_is_meta(rq) && !cfqq->meta_pending)
3048 * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
3050 if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
3053 if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
3057 * if this request is as-good as one we would expect from the
3058 * current cfqq, let it preempt
3060 if (cfq_rq_close(cfqd, cfqq, rq))
3067 * cfqq preempts the active queue. if we allowed preempt with no slice left,
3068 * let it have half of its nominal slice.
3070 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3072 cfq_log_cfqq(cfqd, cfqq, "preempt");
3073 cfq_slice_expired(cfqd, 1);
3076 * Put the new queue at the front of the of the current list,
3077 * so we know that it will be selected next.
3079 BUG_ON(!cfq_cfqq_on_rr(cfqq));
3081 cfq_service_tree_add(cfqd, cfqq, 1);
3083 cfqq->slice_end = 0;
3084 cfq_mark_cfqq_slice_new(cfqq);
3088 * Called when a new fs request (rq) is added (to cfqq). Check if there's
3089 * something we should do about it
3092 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3095 struct cfq_io_context *cic = RQ_CIC(rq);
3099 cfqq->meta_pending++;
3101 cfq_update_io_thinktime(cfqd, cic);
3102 cfq_update_io_seektime(cfqd, cfqq, rq);
3103 cfq_update_idle_window(cfqd, cfqq, cic);
3105 cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
3107 if (cfqq == cfqd->active_queue) {
3108 if (cfq_cfqq_wait_busy(cfqq)) {
3109 cfq_clear_cfqq_wait_busy(cfqq);
3110 cfq_mark_cfqq_wait_busy_done(cfqq);
3113 * Remember that we saw a request from this process, but
3114 * don't start queuing just yet. Otherwise we risk seeing lots
3115 * of tiny requests, because we disrupt the normal plugging
3116 * and merging. If the request is already larger than a single
3117 * page, let it rip immediately. For that case we assume that
3118 * merging is already done. Ditto for a busy system that
3119 * has other work pending, don't risk delaying until the
3120 * idle timer unplug to continue working.
3122 if (cfq_cfqq_wait_request(cfqq)) {
3123 if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
3124 cfqd->busy_queues > 1) {
3125 del_timer(&cfqd->idle_slice_timer);
3126 __blk_run_queue(cfqd->queue);
3128 cfq_mark_cfqq_must_dispatch(cfqq);
3130 } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
3132 * not the active queue - expire current slice if it is
3133 * idle and has expired it's mean thinktime or this new queue
3134 * has some old slice time left and is of higher priority or
3135 * this new queue is RT and the current one is BE
3137 cfq_preempt_queue(cfqd, cfqq);
3138 __blk_run_queue(cfqd->queue);
3142 static void cfq_insert_request(struct request_queue *q, struct request *rq)
3144 struct cfq_data *cfqd = q->elevator->elevator_data;
3145 struct cfq_queue *cfqq = RQ_CFQQ(rq);
3147 cfq_log_cfqq(cfqd, cfqq, "insert_request");
3148 cfq_init_prio_data(cfqq, RQ_CIC(rq)->ioc);
3150 rq_set_fifo_time(rq, jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)]);
3151 list_add_tail(&rq->queuelist, &cfqq->fifo);
3154 cfq_rq_enqueued(cfqd, cfqq, rq);
3158 * Update hw_tag based on peak queue depth over 50 samples under
3161 static void cfq_update_hw_tag(struct cfq_data *cfqd)
3163 struct cfq_queue *cfqq = cfqd->active_queue;
3165 if (rq_in_driver(cfqd) > cfqd->hw_tag_est_depth)
3166 cfqd->hw_tag_est_depth = rq_in_driver(cfqd);
3168 if (cfqd->hw_tag == 1)
3171 if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
3172 rq_in_driver(cfqd) <= CFQ_HW_QUEUE_MIN)
3176 * If active queue hasn't enough requests and can idle, cfq might not
3177 * dispatch sufficient requests to hardware. Don't zero hw_tag in this
3180 if (cfqq && cfq_cfqq_idle_window(cfqq) &&
3181 cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] <
3182 CFQ_HW_QUEUE_MIN && rq_in_driver(cfqd) < CFQ_HW_QUEUE_MIN)
3185 if (cfqd->hw_tag_samples++ < 50)
3188 if (cfqd->hw_tag_est_depth >= CFQ_HW_QUEUE_MIN)
3194 static void cfq_completed_request(struct request_queue *q, struct request *rq)
3196 struct cfq_queue *cfqq = RQ_CFQQ(rq);
3197 struct cfq_data *cfqd = cfqq->cfqd;
3198 const int sync = rq_is_sync(rq);
3202 cfq_log_cfqq(cfqd, cfqq, "complete rqnoidle %d", !!rq_noidle(rq));
3204 cfq_update_hw_tag(cfqd);
3206 WARN_ON(!cfqd->rq_in_driver[sync]);
3207 WARN_ON(!cfqq->dispatched);
3208 cfqd->rq_in_driver[sync]--;
3211 if (cfq_cfqq_sync(cfqq))
3212 cfqd->sync_flight--;
3215 RQ_CIC(rq)->last_end_request = now;
3216 cfqd->last_end_sync_rq = now;
3220 * If this is the active queue, check if it needs to be expired,
3221 * or if we want to idle in case it has no pending requests.
3223 if (cfqd->active_queue == cfqq) {
3224 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
3226 if (cfq_cfqq_slice_new(cfqq)) {
3227 cfq_set_prio_slice(cfqd, cfqq);
3228 cfq_clear_cfqq_slice_new(cfqq);
3232 * If this queue consumed its slice and this is last queue
3233 * in the group, wait for next request before we expire
3236 if (cfq_slice_used(cfqq) && cfqq->cfqg->nr_cfqq == 1) {
3237 cfqq->slice_end = jiffies + cfqd->cfq_slice_idle;
3238 cfq_mark_cfqq_wait_busy(cfqq);
3242 * Idling is not enabled on:
3244 * - idle-priority queues
3246 * - queues with still some requests queued
3247 * - when there is a close cooperator
3249 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
3250 cfq_slice_expired(cfqd, 1);
3251 else if (sync && cfqq_empty &&
3252 !cfq_close_cooperator(cfqd, cfqq)) {
3253 cfqd->noidle_tree_requires_idle |= !rq_noidle(rq);
3255 * Idling is enabled for SYNC_WORKLOAD.
3256 * SYNC_NOIDLE_WORKLOAD idles at the end of the tree
3257 * only if we processed at least one !rq_noidle request
3259 if (cfqd->serving_type == SYNC_WORKLOAD
3260 || cfqd->noidle_tree_requires_idle)
3261 cfq_arm_slice_timer(cfqd);
3265 if (!rq_in_driver(cfqd))
3266 cfq_schedule_dispatch(cfqd);
3270 * we temporarily boost lower priority queues if they are holding fs exclusive
3271 * resources. they are boosted to normal prio (CLASS_BE/4)
3273 static void cfq_prio_boost(struct cfq_queue *cfqq)
3275 if (has_fs_excl()) {
3277 * boost idle prio on transactions that would lock out other
3278 * users of the filesystem
3280 if (cfq_class_idle(cfqq))
3281 cfqq->ioprio_class = IOPRIO_CLASS_BE;
3282 if (cfqq->ioprio > IOPRIO_NORM)
3283 cfqq->ioprio = IOPRIO_NORM;
3286 * unboost the queue (if needed)
3288 cfqq->ioprio_class = cfqq->org_ioprio_class;
3289 cfqq->ioprio = cfqq->org_ioprio;
3293 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
3295 if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
3296 cfq_mark_cfqq_must_alloc_slice(cfqq);
3297 return ELV_MQUEUE_MUST;
3300 return ELV_MQUEUE_MAY;
3303 static int cfq_may_queue(struct request_queue *q, int rw)
3305 struct cfq_data *cfqd = q->elevator->elevator_data;
3306 struct task_struct *tsk = current;
3307 struct cfq_io_context *cic;
3308 struct cfq_queue *cfqq;
3311 * don't force setup of a queue from here, as a call to may_queue
3312 * does not necessarily imply that a request actually will be queued.
3313 * so just lookup a possibly existing queue, or return 'may queue'
3316 cic = cfq_cic_lookup(cfqd, tsk->io_context);
3318 return ELV_MQUEUE_MAY;
3320 cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
3322 cfq_init_prio_data(cfqq, cic->ioc);
3323 cfq_prio_boost(cfqq);
3325 return __cfq_may_queue(cfqq);
3328 return ELV_MQUEUE_MAY;
3332 * queue lock held here
3334 static void cfq_put_request(struct request *rq)
3336 struct cfq_queue *cfqq = RQ_CFQQ(rq);
3339 const int rw = rq_data_dir(rq);
3341 BUG_ON(!cfqq->allocated[rw]);
3342 cfqq->allocated[rw]--;
3344 put_io_context(RQ_CIC(rq)->ioc);
3346 rq->elevator_private = NULL;
3347 rq->elevator_private2 = NULL;
3349 cfq_put_queue(cfqq);
3353 static struct cfq_queue *
3354 cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_context *cic,
3355 struct cfq_queue *cfqq)
3357 cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq);
3358 cic_set_cfqq(cic, cfqq->new_cfqq, 1);
3359 cfq_mark_cfqq_coop(cfqq->new_cfqq);
3360 cfq_put_queue(cfqq);
3361 return cic_to_cfqq(cic, 1);
3364 static int should_split_cfqq(struct cfq_queue *cfqq)
3366 if (cfqq->seeky_start &&
3367 time_after(jiffies, cfqq->seeky_start + CFQQ_COOP_TOUT))
3373 * Returns NULL if a new cfqq should be allocated, or the old cfqq if this
3374 * was the last process referring to said cfqq.
3376 static struct cfq_queue *
3377 split_cfqq(struct cfq_io_context *cic, struct cfq_queue *cfqq)
3379 if (cfqq_process_refs(cfqq) == 1) {
3380 cfqq->seeky_start = 0;
3381 cfqq->pid = current->pid;
3382 cfq_clear_cfqq_coop(cfqq);
3386 cic_set_cfqq(cic, NULL, 1);
3387 cfq_put_queue(cfqq);
3391 * Allocate cfq data structures associated with this request.
3394 cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
3396 struct cfq_data *cfqd = q->elevator->elevator_data;
3397 struct cfq_io_context *cic;
3398 const int rw = rq_data_dir(rq);
3399 const bool is_sync = rq_is_sync(rq);
3400 struct cfq_queue *cfqq;
3401 unsigned long flags;
3403 might_sleep_if(gfp_mask & __GFP_WAIT);
3405 cic = cfq_get_io_context(cfqd, gfp_mask);
3407 spin_lock_irqsave(q->queue_lock, flags);
3413 cfqq = cic_to_cfqq(cic, is_sync);
3414 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
3415 cfqq = cfq_get_queue(cfqd, is_sync, cic->ioc, gfp_mask);
3416 cic_set_cfqq(cic, cfqq, is_sync);
3419 * If the queue was seeky for too long, break it apart.
3421 if (cfq_cfqq_coop(cfqq) && should_split_cfqq(cfqq)) {
3422 cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq");
3423 cfqq = split_cfqq(cic, cfqq);
3429 * Check to see if this queue is scheduled to merge with
3430 * another, closely cooperating queue. The merging of
3431 * queues happens here as it must be done in process context.
3432 * The reference on new_cfqq was taken in merge_cfqqs.
3435 cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq);
3438 cfqq->allocated[rw]++;
3439 atomic_inc(&cfqq->ref);
3441 spin_unlock_irqrestore(q->queue_lock, flags);
3443 rq->elevator_private = cic;
3444 rq->elevator_private2 = cfqq;
3449 put_io_context(cic->ioc);
3451 cfq_schedule_dispatch(cfqd);
3452 spin_unlock_irqrestore(q->queue_lock, flags);
3453 cfq_log(cfqd, "set_request fail");
3457 static void cfq_kick_queue(struct work_struct *work)
3459 struct cfq_data *cfqd =
3460 container_of(work, struct cfq_data, unplug_work);
3461 struct request_queue *q = cfqd->queue;
3463 spin_lock_irq(q->queue_lock);
3464 __blk_run_queue(cfqd->queue);
3465 spin_unlock_irq(q->queue_lock);
3469 * Timer running if the active_queue is currently idling inside its time slice
3471 static void cfq_idle_slice_timer(unsigned long data)
3473 struct cfq_data *cfqd = (struct cfq_data *) data;
3474 struct cfq_queue *cfqq;
3475 unsigned long flags;
3478 cfq_log(cfqd, "idle timer fired");
3480 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
3482 cfqq = cfqd->active_queue;
3487 * We saw a request before the queue expired, let it through
3489 if (cfq_cfqq_must_dispatch(cfqq))
3495 if (cfq_slice_used(cfqq))
3499 * only expire and reinvoke request handler, if there are
3500 * other queues with pending requests
3502 if (!cfqd->busy_queues)
3506 * not expired and it has a request pending, let it dispatch
3508 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3512 * Queue depth flag is reset only when the idle didn't succeed
3514 cfq_clear_cfqq_deep(cfqq);
3517 cfq_slice_expired(cfqd, timed_out);
3519 cfq_schedule_dispatch(cfqd);
3521 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
3524 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
3526 del_timer_sync(&cfqd->idle_slice_timer);
3527 cancel_work_sync(&cfqd->unplug_work);
3530 static void cfq_put_async_queues(struct cfq_data *cfqd)
3534 for (i = 0; i < IOPRIO_BE_NR; i++) {
3535 if (cfqd->async_cfqq[0][i])
3536 cfq_put_queue(cfqd->async_cfqq[0][i]);
3537 if (cfqd->async_cfqq[1][i])
3538 cfq_put_queue(cfqd->async_cfqq[1][i]);
3541 if (cfqd->async_idle_cfqq)
3542 cfq_put_queue(cfqd->async_idle_cfqq);
3545 static void cfq_exit_queue(struct elevator_queue *e)
3547 struct cfq_data *cfqd = e->elevator_data;
3548 struct request_queue *q = cfqd->queue;
3550 cfq_shutdown_timer_wq(cfqd);
3552 spin_lock_irq(q->queue_lock);
3554 if (cfqd->active_queue)
3555 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
3557 while (!list_empty(&cfqd->cic_list)) {
3558 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
3559 struct cfq_io_context,
3562 __cfq_exit_single_io_context(cfqd, cic);
3565 cfq_put_async_queues(cfqd);
3566 cfq_release_cfq_groups(cfqd);
3567 blkiocg_del_blkio_group(&cfqd->root_group.blkg);
3569 spin_unlock_irq(q->queue_lock);
3571 cfq_shutdown_timer_wq(cfqd);
3573 /* Wait for cfqg->blkg->key accessors to exit their grace periods. */
3578 static void *cfq_init_queue(struct request_queue *q)
3580 struct cfq_data *cfqd;
3582 struct cfq_group *cfqg;
3583 struct cfq_rb_root *st;
3585 cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
3589 /* Init root service tree */
3590 cfqd->grp_service_tree = CFQ_RB_ROOT;
3592 /* Init root group */
3593 cfqg = &cfqd->root_group;
3594 for_each_cfqg_st(cfqg, i, j, st)
3596 RB_CLEAR_NODE(&cfqg->rb_node);
3598 /* Give preference to root group over other groups */
3599 cfqg->weight = 2*BLKIO_WEIGHT_DEFAULT;
3601 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3603 * Take a reference to root group which we never drop. This is just
3604 * to make sure that cfq_put_cfqg() does not try to kfree root group
3606 atomic_set(&cfqg->ref, 1);
3607 blkiocg_add_blkio_group(&blkio_root_cgroup, &cfqg->blkg, (void *)cfqd,
3611 * Not strictly needed (since RB_ROOT just clears the node and we
3612 * zeroed cfqd on alloc), but better be safe in case someone decides
3613 * to add magic to the rb code
3615 for (i = 0; i < CFQ_PRIO_LISTS; i++)
3616 cfqd->prio_trees[i] = RB_ROOT;
3619 * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
3620 * Grab a permanent reference to it, so that the normal code flow
3621 * will not attempt to free it.
3623 cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
3624 atomic_inc(&cfqd->oom_cfqq.ref);
3625 cfq_link_cfqq_cfqg(&cfqd->oom_cfqq, &cfqd->root_group);
3627 INIT_LIST_HEAD(&cfqd->cic_list);
3631 init_timer(&cfqd->idle_slice_timer);
3632 cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
3633 cfqd->idle_slice_timer.data = (unsigned long) cfqd;
3635 INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
3637 cfqd->cfq_quantum = cfq_quantum;
3638 cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
3639 cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
3640 cfqd->cfq_back_max = cfq_back_max;
3641 cfqd->cfq_back_penalty = cfq_back_penalty;
3642 cfqd->cfq_slice[0] = cfq_slice_async;
3643 cfqd->cfq_slice[1] = cfq_slice_sync;
3644 cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
3645 cfqd->cfq_slice_idle = cfq_slice_idle;
3646 cfqd->cfq_latency = 1;
3648 cfqd->last_end_sync_rq = jiffies;
3652 static void cfq_slab_kill(void)
3655 * Caller already ensured that pending RCU callbacks are completed,
3656 * so we should have no busy allocations at this point.
3659 kmem_cache_destroy(cfq_pool);
3661 kmem_cache_destroy(cfq_ioc_pool);
3664 static int __init cfq_slab_setup(void)
3666 cfq_pool = KMEM_CACHE(cfq_queue, 0);
3670 cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0);
3681 * sysfs parts below -->
3684 cfq_var_show(unsigned int var, char *page)
3686 return sprintf(page, "%d\n", var);
3690 cfq_var_store(unsigned int *var, const char *page, size_t count)
3692 char *p = (char *) page;
3694 *var = simple_strtoul(p, &p, 10);
3698 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
3699 static ssize_t __FUNC(struct elevator_queue *e, char *page) \
3701 struct cfq_data *cfqd = e->elevator_data; \
3702 unsigned int __data = __VAR; \
3704 __data = jiffies_to_msecs(__data); \
3705 return cfq_var_show(__data, (page)); \
3707 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
3708 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
3709 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
3710 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
3711 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
3712 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
3713 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
3714 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
3715 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
3716 SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
3717 #undef SHOW_FUNCTION
3719 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
3720 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
3722 struct cfq_data *cfqd = e->elevator_data; \
3723 unsigned int __data; \
3724 int ret = cfq_var_store(&__data, (page), count); \
3725 if (__data < (MIN)) \
3727 else if (__data > (MAX)) \
3730 *(__PTR) = msecs_to_jiffies(__data); \
3732 *(__PTR) = __data; \
3735 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
3736 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
3738 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
3740 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
3741 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
3743 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
3744 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
3745 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
3746 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
3748 STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
3749 #undef STORE_FUNCTION
3751 #define CFQ_ATTR(name) \
3752 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
3754 static struct elv_fs_entry cfq_attrs[] = {
3756 CFQ_ATTR(fifo_expire_sync),
3757 CFQ_ATTR(fifo_expire_async),
3758 CFQ_ATTR(back_seek_max),
3759 CFQ_ATTR(back_seek_penalty),
3760 CFQ_ATTR(slice_sync),
3761 CFQ_ATTR(slice_async),
3762 CFQ_ATTR(slice_async_rq),
3763 CFQ_ATTR(slice_idle),
3764 CFQ_ATTR(low_latency),
3768 static struct elevator_type iosched_cfq = {
3770 .elevator_merge_fn = cfq_merge,
3771 .elevator_merged_fn = cfq_merged_request,
3772 .elevator_merge_req_fn = cfq_merged_requests,
3773 .elevator_allow_merge_fn = cfq_allow_merge,
3774 .elevator_dispatch_fn = cfq_dispatch_requests,
3775 .elevator_add_req_fn = cfq_insert_request,
3776 .elevator_activate_req_fn = cfq_activate_request,
3777 .elevator_deactivate_req_fn = cfq_deactivate_request,
3778 .elevator_queue_empty_fn = cfq_queue_empty,
3779 .elevator_completed_req_fn = cfq_completed_request,
3780 .elevator_former_req_fn = elv_rb_former_request,
3781 .elevator_latter_req_fn = elv_rb_latter_request,
3782 .elevator_set_req_fn = cfq_set_request,
3783 .elevator_put_req_fn = cfq_put_request,
3784 .elevator_may_queue_fn = cfq_may_queue,
3785 .elevator_init_fn = cfq_init_queue,
3786 .elevator_exit_fn = cfq_exit_queue,
3787 .trim = cfq_free_io_context,
3789 .elevator_attrs = cfq_attrs,
3790 .elevator_name = "cfq",
3791 .elevator_owner = THIS_MODULE,
3794 static int __init cfq_init(void)
3797 * could be 0 on HZ < 1000 setups
3799 if (!cfq_slice_async)
3800 cfq_slice_async = 1;
3801 if (!cfq_slice_idle)
3804 if (cfq_slab_setup())
3807 elv_register(&iosched_cfq);
3812 static void __exit cfq_exit(void)
3814 DECLARE_COMPLETION_ONSTACK(all_gone);
3815 elv_unregister(&iosched_cfq);
3816 ioc_gone = &all_gone;
3817 /* ioc_gone's update must be visible before reading ioc_count */
3821 * this also protects us from entering cfq_slab_kill() with
3822 * pending RCU callbacks
3824 if (elv_ioc_count_read(cfq_ioc_count))
3825 wait_for_completion(&all_gone);
3829 module_init(cfq_init);
3830 module_exit(cfq_exit);
3832 MODULE_AUTHOR("Jens Axboe");
3833 MODULE_LICENSE("GPL");
3834 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");