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 = 8;
23 static const int cfq_fifo_expire[2] = { HZ / 4, HZ / 8 };
24 /* maximum backwards seek, in KiB */
25 static const int cfq_back_max = 16 * 1024;
26 /* penalty of a backwards seek */
27 static const int cfq_back_penalty = 2;
28 static const int cfq_slice_sync = HZ / 10;
29 static int cfq_slice_async = HZ / 25;
30 static const int cfq_slice_async_rq = 2;
31 static int cfq_slice_idle = HZ / 125;
32 static const int cfq_target_latency = HZ * 3/10; /* 300 ms */
33 static const int cfq_hist_divisor = 4;
36 * offset from end of service tree
38 #define CFQ_IDLE_DELAY (HZ / 5)
41 * below this threshold, we consider thinktime immediate
43 #define CFQ_MIN_TT (2)
45 #define CFQ_SLICE_SCALE (5)
46 #define CFQ_HW_QUEUE_MIN (5)
47 #define CFQ_SERVICE_SHIFT 12
49 #define CFQQ_SEEK_THR (sector_t)(8 * 100)
50 #define CFQQ_CLOSE_THR (sector_t)(8 * 1024)
51 #define CFQQ_SECT_THR_NONROT (sector_t)(2 * 32)
52 #define CFQQ_SEEKY(cfqq) (hweight32(cfqq->seek_history) > 32/8)
55 ((struct cfq_io_context *) (rq)->elevator_private)
56 #define RQ_CFQQ(rq) (struct cfq_queue *) ((rq)->elevator_private2)
58 static struct kmem_cache *cfq_pool;
59 static struct kmem_cache *cfq_ioc_pool;
61 static DEFINE_PER_CPU(unsigned long, cfq_ioc_count);
62 static struct completion *ioc_gone;
63 static DEFINE_SPINLOCK(ioc_gone_lock);
65 #define CFQ_PRIO_LISTS IOPRIO_BE_NR
66 #define cfq_class_idle(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
67 #define cfq_class_rt(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
69 #define sample_valid(samples) ((samples) > 80)
70 #define rb_entry_cfqg(node) rb_entry((node), struct cfq_group, rb_node)
73 * Most of our rbtree usage is for sorting with min extraction, so
74 * if we cache the leftmost node we don't have to walk down the tree
75 * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should
76 * move this into the elevator for the rq sorting as well.
82 unsigned total_weight;
84 struct rb_node *active;
86 #define CFQ_RB_ROOT (struct cfq_rb_root) { .rb = RB_ROOT, .left = NULL, \
87 .count = 0, .min_vdisktime = 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 unsigned int slice_dispatch;
122 /* time when first request from queue completed and slice started. */
123 unsigned long slice_start;
124 unsigned long slice_end;
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;
139 sector_t last_request_pos;
141 struct cfq_rb_root *service_tree;
142 struct cfq_queue *new_cfqq;
143 struct cfq_group *cfqg;
144 struct cfq_group *orig_cfqg;
145 /* Sectors dispatched in current dispatch round */
146 unsigned long nr_sectors;
150 * First index in the service_trees.
151 * IDLE is handled separately, so it has negative index
160 * Second index in the service_trees.
164 SYNC_NOIDLE_WORKLOAD = 1,
168 /* This is per cgroup per device grouping structure */
170 /* group service_tree member */
171 struct rb_node rb_node;
173 /* group service_tree key */
178 /* number of cfqq currently on this group */
181 /* Per group busy queus average. Useful for workload slice calc. */
182 unsigned int busy_queues_avg[2];
184 * rr lists of queues with requests, onle rr for each priority class.
185 * Counts are embedded in the cfq_rb_root
187 struct cfq_rb_root service_trees[2][3];
188 struct cfq_rb_root service_tree_idle;
190 unsigned long saved_workload_slice;
191 enum wl_type_t saved_workload;
192 enum wl_prio_t saved_serving_prio;
193 struct blkio_group blkg;
194 #ifdef CONFIG_CFQ_GROUP_IOSCHED
195 struct hlist_node cfqd_node;
201 * Per block device queue structure
204 struct request_queue *queue;
205 /* Root service tree for cfq_groups */
206 struct cfq_rb_root grp_service_tree;
207 struct cfq_group root_group;
210 * The priority currently being served
212 enum wl_prio_t serving_prio;
213 enum wl_type_t serving_type;
214 unsigned long workload_expires;
215 struct cfq_group *serving_group;
216 bool noidle_tree_requires_idle;
219 * Each priority tree is sorted by next_request position. These
220 * trees are used when determining if two or more queues are
221 * interleaving requests (see cfq_close_cooperator).
223 struct rb_root prio_trees[CFQ_PRIO_LISTS];
225 unsigned int busy_queues;
231 * queue-depth detection
237 * -1 => indeterminate, (cfq will behave as if NCQ is present, to allow better detection)
238 * 1 => NCQ is present (hw_tag_est_depth is the estimated max depth)
241 int hw_tag_est_depth;
242 unsigned int hw_tag_samples;
245 * idle window management
247 struct timer_list idle_slice_timer;
248 struct work_struct unplug_work;
250 struct cfq_queue *active_queue;
251 struct cfq_io_context *active_cic;
254 * async queue for each priority case
256 struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
257 struct cfq_queue *async_idle_cfqq;
259 sector_t last_position;
262 * tunables, see top of file
264 unsigned int cfq_quantum;
265 unsigned int cfq_fifo_expire[2];
266 unsigned int cfq_back_penalty;
267 unsigned int cfq_back_max;
268 unsigned int cfq_slice[2];
269 unsigned int cfq_slice_async_rq;
270 unsigned int cfq_slice_idle;
271 unsigned int cfq_latency;
272 unsigned int cfq_group_isolation;
274 struct list_head cic_list;
277 * Fallback dummy cfqq for extreme OOM conditions
279 struct cfq_queue oom_cfqq;
281 unsigned long last_delayed_sync;
283 /* List of cfq groups being managed on this device*/
284 struct hlist_head cfqg_list;
288 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd);
290 static struct cfq_rb_root *service_tree_for(struct cfq_group *cfqg,
297 if (prio == IDLE_WORKLOAD)
298 return &cfqg->service_tree_idle;
300 return &cfqg->service_trees[prio][type];
303 enum cfqq_state_flags {
304 CFQ_CFQQ_FLAG_on_rr = 0, /* on round-robin busy list */
305 CFQ_CFQQ_FLAG_wait_request, /* waiting for a request */
306 CFQ_CFQQ_FLAG_must_dispatch, /* must be allowed a dispatch */
307 CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
308 CFQ_CFQQ_FLAG_fifo_expire, /* FIFO checked in this slice */
309 CFQ_CFQQ_FLAG_idle_window, /* slice idling enabled */
310 CFQ_CFQQ_FLAG_prio_changed, /* task priority has changed */
311 CFQ_CFQQ_FLAG_slice_new, /* no requests dispatched in slice */
312 CFQ_CFQQ_FLAG_sync, /* synchronous queue */
313 CFQ_CFQQ_FLAG_coop, /* cfqq is shared */
314 CFQ_CFQQ_FLAG_split_coop, /* shared cfqq will be splitted */
315 CFQ_CFQQ_FLAG_deep, /* sync cfqq experienced large depth */
316 CFQ_CFQQ_FLAG_wait_busy, /* Waiting for next request */
319 #define CFQ_CFQQ_FNS(name) \
320 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \
322 (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name); \
324 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \
326 (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \
328 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \
330 return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \
334 CFQ_CFQQ_FNS(wait_request);
335 CFQ_CFQQ_FNS(must_dispatch);
336 CFQ_CFQQ_FNS(must_alloc_slice);
337 CFQ_CFQQ_FNS(fifo_expire);
338 CFQ_CFQQ_FNS(idle_window);
339 CFQ_CFQQ_FNS(prio_changed);
340 CFQ_CFQQ_FNS(slice_new);
343 CFQ_CFQQ_FNS(split_coop);
345 CFQ_CFQQ_FNS(wait_busy);
348 #ifdef CONFIG_DEBUG_CFQ_IOSCHED
349 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
350 blk_add_trace_msg((cfqd)->queue, "cfq%d%c %s " fmt, (cfqq)->pid, \
351 cfq_cfqq_sync((cfqq)) ? 'S' : 'A', \
352 blkg_path(&(cfqq)->cfqg->blkg), ##args);
354 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...) \
355 blk_add_trace_msg((cfqd)->queue, "%s " fmt, \
356 blkg_path(&(cfqg)->blkg), ##args); \
359 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
360 blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args)
361 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...) do {} while (0);
363 #define cfq_log(cfqd, fmt, args...) \
364 blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
366 /* Traverses through cfq group service trees */
367 #define for_each_cfqg_st(cfqg, i, j, st) \
368 for (i = 0; i <= IDLE_WORKLOAD; i++) \
369 for (j = 0, st = i < IDLE_WORKLOAD ? &cfqg->service_trees[i][j]\
370 : &cfqg->service_tree_idle; \
371 (i < IDLE_WORKLOAD && j <= SYNC_WORKLOAD) || \
372 (i == IDLE_WORKLOAD && j == 0); \
373 j++, st = i < IDLE_WORKLOAD ? \
374 &cfqg->service_trees[i][j]: NULL) \
377 static inline enum wl_prio_t cfqq_prio(struct cfq_queue *cfqq)
379 if (cfq_class_idle(cfqq))
380 return IDLE_WORKLOAD;
381 if (cfq_class_rt(cfqq))
387 static enum wl_type_t cfqq_type(struct cfq_queue *cfqq)
389 if (!cfq_cfqq_sync(cfqq))
390 return ASYNC_WORKLOAD;
391 if (!cfq_cfqq_idle_window(cfqq))
392 return SYNC_NOIDLE_WORKLOAD;
393 return SYNC_WORKLOAD;
396 static inline int cfq_group_busy_queues_wl(enum wl_prio_t wl,
397 struct cfq_data *cfqd,
398 struct cfq_group *cfqg)
400 if (wl == IDLE_WORKLOAD)
401 return cfqg->service_tree_idle.count;
403 return cfqg->service_trees[wl][ASYNC_WORKLOAD].count
404 + cfqg->service_trees[wl][SYNC_NOIDLE_WORKLOAD].count
405 + cfqg->service_trees[wl][SYNC_WORKLOAD].count;
408 static inline int cfqg_busy_async_queues(struct cfq_data *cfqd,
409 struct cfq_group *cfqg)
411 return cfqg->service_trees[RT_WORKLOAD][ASYNC_WORKLOAD].count
412 + cfqg->service_trees[BE_WORKLOAD][ASYNC_WORKLOAD].count;
415 static void cfq_dispatch_insert(struct request_queue *, struct request *);
416 static struct cfq_queue *cfq_get_queue(struct cfq_data *, bool,
417 struct io_context *, gfp_t);
418 static struct cfq_io_context *cfq_cic_lookup(struct cfq_data *,
419 struct io_context *);
421 static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_context *cic,
424 return cic->cfqq[is_sync];
427 static inline void cic_set_cfqq(struct cfq_io_context *cic,
428 struct cfq_queue *cfqq, bool is_sync)
430 cic->cfqq[is_sync] = cfqq;
434 * We regard a request as SYNC, if it's either a read or has the SYNC bit
435 * set (in which case it could also be direct WRITE).
437 static inline bool cfq_bio_sync(struct bio *bio)
439 return bio_data_dir(bio) == READ || bio_rw_flagged(bio, BIO_RW_SYNCIO);
443 * scheduler run of queue, if there are requests pending and no one in the
444 * driver that will restart queueing
446 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
448 if (cfqd->busy_queues) {
449 cfq_log(cfqd, "schedule dispatch");
450 kblockd_schedule_work(cfqd->queue, &cfqd->unplug_work);
454 static int cfq_queue_empty(struct request_queue *q)
456 struct cfq_data *cfqd = q->elevator->elevator_data;
458 return !cfqd->rq_queued;
462 * Scale schedule slice based on io priority. Use the sync time slice only
463 * if a queue is marked sync and has sync io queued. A sync queue with async
464 * io only, should not get full sync slice length.
466 static inline int cfq_prio_slice(struct cfq_data *cfqd, bool sync,
469 const int base_slice = cfqd->cfq_slice[sync];
471 WARN_ON(prio >= IOPRIO_BE_NR);
473 return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
477 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
479 return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
482 static inline u64 cfq_scale_slice(unsigned long delta, struct cfq_group *cfqg)
484 u64 d = delta << CFQ_SERVICE_SHIFT;
486 d = d * BLKIO_WEIGHT_DEFAULT;
487 do_div(d, cfqg->weight);
491 static inline u64 max_vdisktime(u64 min_vdisktime, u64 vdisktime)
493 s64 delta = (s64)(vdisktime - min_vdisktime);
495 min_vdisktime = vdisktime;
497 return min_vdisktime;
500 static inline u64 min_vdisktime(u64 min_vdisktime, u64 vdisktime)
502 s64 delta = (s64)(vdisktime - min_vdisktime);
504 min_vdisktime = vdisktime;
506 return min_vdisktime;
509 static void update_min_vdisktime(struct cfq_rb_root *st)
511 u64 vdisktime = st->min_vdisktime;
512 struct cfq_group *cfqg;
515 cfqg = rb_entry_cfqg(st->active);
516 vdisktime = cfqg->vdisktime;
520 cfqg = rb_entry_cfqg(st->left);
521 vdisktime = min_vdisktime(vdisktime, cfqg->vdisktime);
524 st->min_vdisktime = max_vdisktime(st->min_vdisktime, vdisktime);
528 * get averaged number of queues of RT/BE priority.
529 * average is updated, with a formula that gives more weight to higher numbers,
530 * to quickly follows sudden increases and decrease slowly
533 static inline unsigned cfq_group_get_avg_queues(struct cfq_data *cfqd,
534 struct cfq_group *cfqg, bool rt)
536 unsigned min_q, max_q;
537 unsigned mult = cfq_hist_divisor - 1;
538 unsigned round = cfq_hist_divisor / 2;
539 unsigned busy = cfq_group_busy_queues_wl(rt, cfqd, cfqg);
541 min_q = min(cfqg->busy_queues_avg[rt], busy);
542 max_q = max(cfqg->busy_queues_avg[rt], busy);
543 cfqg->busy_queues_avg[rt] = (mult * max_q + min_q + round) /
545 return cfqg->busy_queues_avg[rt];
548 static inline unsigned
549 cfq_group_slice(struct cfq_data *cfqd, struct cfq_group *cfqg)
551 struct cfq_rb_root *st = &cfqd->grp_service_tree;
553 return cfq_target_latency * cfqg->weight / st->total_weight;
557 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
559 unsigned slice = cfq_prio_to_slice(cfqd, cfqq);
560 if (cfqd->cfq_latency) {
562 * interested queues (we consider only the ones with the same
563 * priority class in the cfq group)
565 unsigned iq = cfq_group_get_avg_queues(cfqd, cfqq->cfqg,
567 unsigned sync_slice = cfqd->cfq_slice[1];
568 unsigned expect_latency = sync_slice * iq;
569 unsigned group_slice = cfq_group_slice(cfqd, cfqq->cfqg);
571 if (expect_latency > group_slice) {
572 unsigned base_low_slice = 2 * cfqd->cfq_slice_idle;
573 /* scale low_slice according to IO priority
574 * and sync vs async */
576 min(slice, base_low_slice * slice / sync_slice);
577 /* the adapted slice value is scaled to fit all iqs
578 * into the target latency */
579 slice = max(slice * group_slice / expect_latency,
583 cfqq->slice_start = jiffies;
584 cfqq->slice_end = jiffies + slice;
585 cfqq->allocated_slice = slice;
586 cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
590 * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
591 * isn't valid until the first request from the dispatch is activated
592 * and the slice time set.
594 static inline bool cfq_slice_used(struct cfq_queue *cfqq)
596 if (cfq_cfqq_slice_new(cfqq))
598 if (time_before(jiffies, cfqq->slice_end))
605 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
606 * We choose the request that is closest to the head right now. Distance
607 * behind the head is penalized and only allowed to a certain extent.
609 static struct request *
610 cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2, sector_t last)
612 sector_t s1, s2, d1 = 0, d2 = 0;
613 unsigned long back_max;
614 #define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */
615 #define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */
616 unsigned wrap = 0; /* bit mask: requests behind the disk head? */
618 if (rq1 == NULL || rq1 == rq2)
623 if (rq_is_sync(rq1) && !rq_is_sync(rq2))
625 else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
627 if (rq_is_meta(rq1) && !rq_is_meta(rq2))
629 else if (rq_is_meta(rq2) && !rq_is_meta(rq1))
632 s1 = blk_rq_pos(rq1);
633 s2 = blk_rq_pos(rq2);
636 * by definition, 1KiB is 2 sectors
638 back_max = cfqd->cfq_back_max * 2;
641 * Strict one way elevator _except_ in the case where we allow
642 * short backward seeks which are biased as twice the cost of a
643 * similar forward seek.
647 else if (s1 + back_max >= last)
648 d1 = (last - s1) * cfqd->cfq_back_penalty;
650 wrap |= CFQ_RQ1_WRAP;
654 else if (s2 + back_max >= last)
655 d2 = (last - s2) * cfqd->cfq_back_penalty;
657 wrap |= CFQ_RQ2_WRAP;
659 /* Found required data */
662 * By doing switch() on the bit mask "wrap" we avoid having to
663 * check two variables for all permutations: --> faster!
666 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
682 case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
685 * Since both rqs are wrapped,
686 * start with the one that's further behind head
687 * (--> only *one* back seek required),
688 * since back seek takes more time than forward.
698 * The below is leftmost cache rbtree addon
700 static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
702 /* Service tree is empty */
707 root->left = rb_first(&root->rb);
710 return rb_entry(root->left, struct cfq_queue, rb_node);
715 static struct cfq_group *cfq_rb_first_group(struct cfq_rb_root *root)
718 root->left = rb_first(&root->rb);
721 return rb_entry_cfqg(root->left);
726 static void rb_erase_init(struct rb_node *n, struct rb_root *root)
732 static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
736 rb_erase_init(n, &root->rb);
741 * would be nice to take fifo expire time into account as well
743 static struct request *
744 cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
745 struct request *last)
747 struct rb_node *rbnext = rb_next(&last->rb_node);
748 struct rb_node *rbprev = rb_prev(&last->rb_node);
749 struct request *next = NULL, *prev = NULL;
751 BUG_ON(RB_EMPTY_NODE(&last->rb_node));
754 prev = rb_entry_rq(rbprev);
757 next = rb_entry_rq(rbnext);
759 rbnext = rb_first(&cfqq->sort_list);
760 if (rbnext && rbnext != &last->rb_node)
761 next = rb_entry_rq(rbnext);
764 return cfq_choose_req(cfqd, next, prev, blk_rq_pos(last));
767 static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
768 struct cfq_queue *cfqq)
771 * just an approximation, should be ok.
773 return (cfqq->cfqg->nr_cfqq - 1) * (cfq_prio_slice(cfqd, 1, 0) -
774 cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
778 cfqg_key(struct cfq_rb_root *st, struct cfq_group *cfqg)
780 return cfqg->vdisktime - st->min_vdisktime;
784 __cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
786 struct rb_node **node = &st->rb.rb_node;
787 struct rb_node *parent = NULL;
788 struct cfq_group *__cfqg;
789 s64 key = cfqg_key(st, cfqg);
792 while (*node != NULL) {
794 __cfqg = rb_entry_cfqg(parent);
796 if (key < cfqg_key(st, __cfqg))
797 node = &parent->rb_left;
799 node = &parent->rb_right;
805 st->left = &cfqg->rb_node;
807 rb_link_node(&cfqg->rb_node, parent, node);
808 rb_insert_color(&cfqg->rb_node, &st->rb);
812 cfq_group_service_tree_add(struct cfq_data *cfqd, struct cfq_group *cfqg)
814 struct cfq_rb_root *st = &cfqd->grp_service_tree;
815 struct cfq_group *__cfqg;
823 * Currently put the group at the end. Later implement something
824 * so that groups get lesser vtime based on their weights, so that
825 * if group does not loose all if it was not continously backlogged.
827 n = rb_last(&st->rb);
829 __cfqg = rb_entry_cfqg(n);
830 cfqg->vdisktime = __cfqg->vdisktime + CFQ_IDLE_DELAY;
832 cfqg->vdisktime = st->min_vdisktime;
834 __cfq_group_service_tree_add(st, cfqg);
836 st->total_weight += cfqg->weight;
840 cfq_group_service_tree_del(struct cfq_data *cfqd, struct cfq_group *cfqg)
842 struct cfq_rb_root *st = &cfqd->grp_service_tree;
844 if (st->active == &cfqg->rb_node)
847 BUG_ON(cfqg->nr_cfqq < 1);
850 /* If there are other cfq queues under this group, don't delete it */
854 cfq_log_cfqg(cfqd, cfqg, "del_from_rr group");
856 st->total_weight -= cfqg->weight;
857 if (!RB_EMPTY_NODE(&cfqg->rb_node))
858 cfq_rb_erase(&cfqg->rb_node, st);
859 cfqg->saved_workload_slice = 0;
860 blkiocg_update_blkio_group_dequeue_stats(&cfqg->blkg, 1);
863 static inline unsigned int cfq_cfqq_slice_usage(struct cfq_queue *cfqq)
865 unsigned int slice_used;
868 * Queue got expired before even a single request completed or
869 * got expired immediately after first request completion.
871 if (!cfqq->slice_start || cfqq->slice_start == jiffies) {
873 * Also charge the seek time incurred to the group, otherwise
874 * if there are mutiple queues in the group, each can dispatch
875 * a single request on seeky media and cause lots of seek time
876 * and group will never know it.
878 slice_used = max_t(unsigned, (jiffies - cfqq->dispatch_start),
881 slice_used = jiffies - cfqq->slice_start;
882 if (slice_used > cfqq->allocated_slice)
883 slice_used = cfqq->allocated_slice;
886 cfq_log_cfqq(cfqq->cfqd, cfqq, "sl_used=%u sect=%lu", slice_used,
891 static void cfq_group_served(struct cfq_data *cfqd, struct cfq_group *cfqg,
892 struct cfq_queue *cfqq)
894 struct cfq_rb_root *st = &cfqd->grp_service_tree;
895 unsigned int used_sl, charge_sl;
896 int nr_sync = cfqg->nr_cfqq - cfqg_busy_async_queues(cfqd, cfqg)
897 - cfqg->service_tree_idle.count;
900 used_sl = charge_sl = cfq_cfqq_slice_usage(cfqq);
902 if (!cfq_cfqq_sync(cfqq) && !nr_sync)
903 charge_sl = cfqq->allocated_slice;
905 /* Can't update vdisktime while group is on service tree */
906 cfq_rb_erase(&cfqg->rb_node, st);
907 cfqg->vdisktime += cfq_scale_slice(charge_sl, cfqg);
908 __cfq_group_service_tree_add(st, cfqg);
910 /* This group is being expired. Save the context */
911 if (time_after(cfqd->workload_expires, jiffies)) {
912 cfqg->saved_workload_slice = cfqd->workload_expires
914 cfqg->saved_workload = cfqd->serving_type;
915 cfqg->saved_serving_prio = cfqd->serving_prio;
917 cfqg->saved_workload_slice = 0;
919 cfq_log_cfqg(cfqd, cfqg, "served: vt=%llu min_vt=%llu", cfqg->vdisktime,
921 blkiocg_update_blkio_group_stats(&cfqg->blkg, used_sl,
925 #ifdef CONFIG_CFQ_GROUP_IOSCHED
926 static inline struct cfq_group *cfqg_of_blkg(struct blkio_group *blkg)
929 return container_of(blkg, struct cfq_group, blkg);
934 cfq_update_blkio_group_weight(struct blkio_group *blkg, unsigned int weight)
936 cfqg_of_blkg(blkg)->weight = weight;
939 static struct cfq_group *
940 cfq_find_alloc_cfqg(struct cfq_data *cfqd, struct cgroup *cgroup, int create)
942 struct blkio_cgroup *blkcg = cgroup_to_blkio_cgroup(cgroup);
943 struct cfq_group *cfqg = NULL;
946 struct cfq_rb_root *st;
947 struct backing_dev_info *bdi = &cfqd->queue->backing_dev_info;
948 unsigned int major, minor;
950 cfqg = cfqg_of_blkg(blkiocg_lookup_group(blkcg, key));
951 if (cfqg && !cfqg->blkg.dev && bdi->dev && dev_name(bdi->dev)) {
952 sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
953 cfqg->blkg.dev = MKDEV(major, minor);
959 cfqg = kzalloc_node(sizeof(*cfqg), GFP_ATOMIC, cfqd->queue->node);
963 cfqg->weight = blkcg->weight;
964 for_each_cfqg_st(cfqg, i, j, st)
966 RB_CLEAR_NODE(&cfqg->rb_node);
969 * Take the initial reference that will be released on destroy
970 * This can be thought of a joint reference by cgroup and
971 * elevator which will be dropped by either elevator exit
972 * or cgroup deletion path depending on who is exiting first.
974 atomic_set(&cfqg->ref, 1);
976 /* Add group onto cgroup list */
977 sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
978 blkiocg_add_blkio_group(blkcg, &cfqg->blkg, (void *)cfqd,
979 MKDEV(major, minor));
981 /* Add group on cfqd list */
982 hlist_add_head(&cfqg->cfqd_node, &cfqd->cfqg_list);
989 * Search for the cfq group current task belongs to. If create = 1, then also
990 * create the cfq group if it does not exist. request_queue lock must be held.
992 static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd, int create)
994 struct cgroup *cgroup;
995 struct cfq_group *cfqg = NULL;
998 cgroup = task_cgroup(current, blkio_subsys_id);
999 cfqg = cfq_find_alloc_cfqg(cfqd, cgroup, create);
1000 if (!cfqg && create)
1001 cfqg = &cfqd->root_group;
1006 static void cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg)
1008 /* Currently, all async queues are mapped to root group */
1009 if (!cfq_cfqq_sync(cfqq))
1010 cfqg = &cfqq->cfqd->root_group;
1013 /* cfqq reference on cfqg */
1014 atomic_inc(&cfqq->cfqg->ref);
1017 static void cfq_put_cfqg(struct cfq_group *cfqg)
1019 struct cfq_rb_root *st;
1022 BUG_ON(atomic_read(&cfqg->ref) <= 0);
1023 if (!atomic_dec_and_test(&cfqg->ref))
1025 for_each_cfqg_st(cfqg, i, j, st)
1026 BUG_ON(!RB_EMPTY_ROOT(&st->rb) || st->active != NULL);
1030 static void cfq_destroy_cfqg(struct cfq_data *cfqd, struct cfq_group *cfqg)
1032 /* Something wrong if we are trying to remove same group twice */
1033 BUG_ON(hlist_unhashed(&cfqg->cfqd_node));
1035 hlist_del_init(&cfqg->cfqd_node);
1038 * Put the reference taken at the time of creation so that when all
1039 * queues are gone, group can be destroyed.
1044 static void cfq_release_cfq_groups(struct cfq_data *cfqd)
1046 struct hlist_node *pos, *n;
1047 struct cfq_group *cfqg;
1049 hlist_for_each_entry_safe(cfqg, pos, n, &cfqd->cfqg_list, cfqd_node) {
1051 * If cgroup removal path got to blk_group first and removed
1052 * it from cgroup list, then it will take care of destroying
1055 if (!blkiocg_del_blkio_group(&cfqg->blkg))
1056 cfq_destroy_cfqg(cfqd, cfqg);
1061 * Blk cgroup controller notification saying that blkio_group object is being
1062 * delinked as associated cgroup object is going away. That also means that
1063 * no new IO will come in this group. So get rid of this group as soon as
1064 * any pending IO in the group is finished.
1066 * This function is called under rcu_read_lock(). key is the rcu protected
1067 * pointer. That means "key" is a valid cfq_data pointer as long as we are rcu
1070 * "key" was fetched from blkio_group under blkio_cgroup->lock. That means
1071 * it should not be NULL as even if elevator was exiting, cgroup deltion
1072 * path got to it first.
1074 void cfq_unlink_blkio_group(void *key, struct blkio_group *blkg)
1076 unsigned long flags;
1077 struct cfq_data *cfqd = key;
1079 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1080 cfq_destroy_cfqg(cfqd, cfqg_of_blkg(blkg));
1081 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1084 #else /* GROUP_IOSCHED */
1085 static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd, int create)
1087 return &cfqd->root_group;
1090 cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) {
1094 static void cfq_release_cfq_groups(struct cfq_data *cfqd) {}
1095 static inline void cfq_put_cfqg(struct cfq_group *cfqg) {}
1097 #endif /* GROUP_IOSCHED */
1100 * The cfqd->service_trees holds all pending cfq_queue's that have
1101 * requests waiting to be processed. It is sorted in the order that
1102 * we will service the queues.
1104 static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1107 struct rb_node **p, *parent;
1108 struct cfq_queue *__cfqq;
1109 unsigned long rb_key;
1110 struct cfq_rb_root *service_tree;
1113 int group_changed = 0;
1115 #ifdef CONFIG_CFQ_GROUP_IOSCHED
1116 if (!cfqd->cfq_group_isolation
1117 && cfqq_type(cfqq) == SYNC_NOIDLE_WORKLOAD
1118 && cfqq->cfqg && cfqq->cfqg != &cfqd->root_group) {
1119 /* Move this cfq to root group */
1120 cfq_log_cfqq(cfqd, cfqq, "moving to root group");
1121 if (!RB_EMPTY_NODE(&cfqq->rb_node))
1122 cfq_group_service_tree_del(cfqd, cfqq->cfqg);
1123 cfqq->orig_cfqg = cfqq->cfqg;
1124 cfqq->cfqg = &cfqd->root_group;
1125 atomic_inc(&cfqd->root_group.ref);
1127 } else if (!cfqd->cfq_group_isolation
1128 && cfqq_type(cfqq) == SYNC_WORKLOAD && cfqq->orig_cfqg) {
1129 /* cfqq is sequential now needs to go to its original group */
1130 BUG_ON(cfqq->cfqg != &cfqd->root_group);
1131 if (!RB_EMPTY_NODE(&cfqq->rb_node))
1132 cfq_group_service_tree_del(cfqd, cfqq->cfqg);
1133 cfq_put_cfqg(cfqq->cfqg);
1134 cfqq->cfqg = cfqq->orig_cfqg;
1135 cfqq->orig_cfqg = NULL;
1137 cfq_log_cfqq(cfqd, cfqq, "moved to origin group");
1141 service_tree = service_tree_for(cfqq->cfqg, cfqq_prio(cfqq),
1143 if (cfq_class_idle(cfqq)) {
1144 rb_key = CFQ_IDLE_DELAY;
1145 parent = rb_last(&service_tree->rb);
1146 if (parent && parent != &cfqq->rb_node) {
1147 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1148 rb_key += __cfqq->rb_key;
1151 } else if (!add_front) {
1153 * Get our rb key offset. Subtract any residual slice
1154 * value carried from last service. A negative resid
1155 * count indicates slice overrun, and this should position
1156 * the next service time further away in the tree.
1158 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
1159 rb_key -= cfqq->slice_resid;
1160 cfqq->slice_resid = 0;
1163 __cfqq = cfq_rb_first(service_tree);
1164 rb_key += __cfqq ? __cfqq->rb_key : jiffies;
1167 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1170 * same position, nothing more to do
1172 if (rb_key == cfqq->rb_key &&
1173 cfqq->service_tree == service_tree)
1176 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1177 cfqq->service_tree = NULL;
1182 cfqq->service_tree = service_tree;
1183 p = &service_tree->rb.rb_node;
1188 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1191 * sort by key, that represents service time.
1193 if (time_before(rb_key, __cfqq->rb_key))
1196 n = &(*p)->rb_right;
1204 service_tree->left = &cfqq->rb_node;
1206 cfqq->rb_key = rb_key;
1207 rb_link_node(&cfqq->rb_node, parent, p);
1208 rb_insert_color(&cfqq->rb_node, &service_tree->rb);
1209 service_tree->count++;
1210 if ((add_front || !new_cfqq) && !group_changed)
1212 cfq_group_service_tree_add(cfqd, cfqq->cfqg);
1215 static struct cfq_queue *
1216 cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
1217 sector_t sector, struct rb_node **ret_parent,
1218 struct rb_node ***rb_link)
1220 struct rb_node **p, *parent;
1221 struct cfq_queue *cfqq = NULL;
1229 cfqq = rb_entry(parent, struct cfq_queue, p_node);
1232 * Sort strictly based on sector. Smallest to the left,
1233 * largest to the right.
1235 if (sector > blk_rq_pos(cfqq->next_rq))
1236 n = &(*p)->rb_right;
1237 else if (sector < blk_rq_pos(cfqq->next_rq))
1245 *ret_parent = parent;
1251 static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1253 struct rb_node **p, *parent;
1254 struct cfq_queue *__cfqq;
1257 rb_erase(&cfqq->p_node, cfqq->p_root);
1258 cfqq->p_root = NULL;
1261 if (cfq_class_idle(cfqq))
1266 cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
1267 __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
1268 blk_rq_pos(cfqq->next_rq), &parent, &p);
1270 rb_link_node(&cfqq->p_node, parent, p);
1271 rb_insert_color(&cfqq->p_node, cfqq->p_root);
1273 cfqq->p_root = NULL;
1277 * Update cfqq's position in the service tree.
1279 static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1282 * Resorting requires the cfqq to be on the RR list already.
1284 if (cfq_cfqq_on_rr(cfqq)) {
1285 cfq_service_tree_add(cfqd, cfqq, 0);
1286 cfq_prio_tree_add(cfqd, cfqq);
1291 * add to busy list of queues for service, trying to be fair in ordering
1292 * the pending list according to last request service
1294 static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1296 cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
1297 BUG_ON(cfq_cfqq_on_rr(cfqq));
1298 cfq_mark_cfqq_on_rr(cfqq);
1299 cfqd->busy_queues++;
1301 cfq_resort_rr_list(cfqd, cfqq);
1305 * Called when the cfqq no longer has requests pending, remove it from
1308 static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1310 cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
1311 BUG_ON(!cfq_cfqq_on_rr(cfqq));
1312 cfq_clear_cfqq_on_rr(cfqq);
1314 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1315 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1316 cfqq->service_tree = NULL;
1319 rb_erase(&cfqq->p_node, cfqq->p_root);
1320 cfqq->p_root = NULL;
1323 cfq_group_service_tree_del(cfqd, cfqq->cfqg);
1324 BUG_ON(!cfqd->busy_queues);
1325 cfqd->busy_queues--;
1329 * rb tree support functions
1331 static void cfq_del_rq_rb(struct request *rq)
1333 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1334 const int sync = rq_is_sync(rq);
1336 BUG_ON(!cfqq->queued[sync]);
1337 cfqq->queued[sync]--;
1339 elv_rb_del(&cfqq->sort_list, rq);
1341 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) {
1343 * Queue will be deleted from service tree when we actually
1344 * expire it later. Right now just remove it from prio tree
1348 rb_erase(&cfqq->p_node, cfqq->p_root);
1349 cfqq->p_root = NULL;
1354 static void cfq_add_rq_rb(struct request *rq)
1356 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1357 struct cfq_data *cfqd = cfqq->cfqd;
1358 struct request *__alias, *prev;
1360 cfqq->queued[rq_is_sync(rq)]++;
1363 * looks a little odd, but the first insert might return an alias.
1364 * if that happens, put the alias on the dispatch list
1366 while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL)
1367 cfq_dispatch_insert(cfqd->queue, __alias);
1369 if (!cfq_cfqq_on_rr(cfqq))
1370 cfq_add_cfqq_rr(cfqd, cfqq);
1373 * check if this request is a better next-serve candidate
1375 prev = cfqq->next_rq;
1376 cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq, cfqd->last_position);
1379 * adjust priority tree position, if ->next_rq changes
1381 if (prev != cfqq->next_rq)
1382 cfq_prio_tree_add(cfqd, cfqq);
1384 BUG_ON(!cfqq->next_rq);
1387 static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
1389 elv_rb_del(&cfqq->sort_list, rq);
1390 cfqq->queued[rq_is_sync(rq)]--;
1394 static struct request *
1395 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
1397 struct task_struct *tsk = current;
1398 struct cfq_io_context *cic;
1399 struct cfq_queue *cfqq;
1401 cic = cfq_cic_lookup(cfqd, tsk->io_context);
1405 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
1407 sector_t sector = bio->bi_sector + bio_sectors(bio);
1409 return elv_rb_find(&cfqq->sort_list, sector);
1415 static void cfq_activate_request(struct request_queue *q, struct request *rq)
1417 struct cfq_data *cfqd = q->elevator->elevator_data;
1419 cfqd->rq_in_driver++;
1420 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
1421 cfqd->rq_in_driver);
1423 cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
1426 static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
1428 struct cfq_data *cfqd = q->elevator->elevator_data;
1430 WARN_ON(!cfqd->rq_in_driver);
1431 cfqd->rq_in_driver--;
1432 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
1433 cfqd->rq_in_driver);
1436 static void cfq_remove_request(struct request *rq)
1438 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1440 if (cfqq->next_rq == rq)
1441 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
1443 list_del_init(&rq->queuelist);
1446 cfqq->cfqd->rq_queued--;
1447 if (rq_is_meta(rq)) {
1448 WARN_ON(!cfqq->meta_pending);
1449 cfqq->meta_pending--;
1453 static int cfq_merge(struct request_queue *q, struct request **req,
1456 struct cfq_data *cfqd = q->elevator->elevator_data;
1457 struct request *__rq;
1459 __rq = cfq_find_rq_fmerge(cfqd, bio);
1460 if (__rq && elv_rq_merge_ok(__rq, bio)) {
1462 return ELEVATOR_FRONT_MERGE;
1465 return ELEVATOR_NO_MERGE;
1468 static void cfq_merged_request(struct request_queue *q, struct request *req,
1471 if (type == ELEVATOR_FRONT_MERGE) {
1472 struct cfq_queue *cfqq = RQ_CFQQ(req);
1474 cfq_reposition_rq_rb(cfqq, req);
1479 cfq_merged_requests(struct request_queue *q, struct request *rq,
1480 struct request *next)
1482 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1484 * reposition in fifo if next is older than rq
1486 if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
1487 time_before(rq_fifo_time(next), rq_fifo_time(rq))) {
1488 list_move(&rq->queuelist, &next->queuelist);
1489 rq_set_fifo_time(rq, rq_fifo_time(next));
1492 if (cfqq->next_rq == next)
1494 cfq_remove_request(next);
1497 static int cfq_allow_merge(struct request_queue *q, struct request *rq,
1500 struct cfq_data *cfqd = q->elevator->elevator_data;
1501 struct cfq_io_context *cic;
1502 struct cfq_queue *cfqq;
1505 * Disallow merge of a sync bio into an async request.
1507 if (cfq_bio_sync(bio) && !rq_is_sync(rq))
1511 * Lookup the cfqq that this bio will be queued with. Allow
1512 * merge only if rq is queued there.
1514 cic = cfq_cic_lookup(cfqd, current->io_context);
1518 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
1519 return cfqq == RQ_CFQQ(rq);
1522 static void __cfq_set_active_queue(struct cfq_data *cfqd,
1523 struct cfq_queue *cfqq)
1526 cfq_log_cfqq(cfqd, cfqq, "set_active wl_prio:%d wl_type:%d",
1527 cfqd->serving_prio, cfqd->serving_type);
1528 cfqq->slice_start = 0;
1529 cfqq->dispatch_start = jiffies;
1530 cfqq->allocated_slice = 0;
1531 cfqq->slice_end = 0;
1532 cfqq->slice_dispatch = 0;
1533 cfqq->nr_sectors = 0;
1535 cfq_clear_cfqq_wait_request(cfqq);
1536 cfq_clear_cfqq_must_dispatch(cfqq);
1537 cfq_clear_cfqq_must_alloc_slice(cfqq);
1538 cfq_clear_cfqq_fifo_expire(cfqq);
1539 cfq_mark_cfqq_slice_new(cfqq);
1541 del_timer(&cfqd->idle_slice_timer);
1544 cfqd->active_queue = cfqq;
1548 * current cfqq expired its slice (or was too idle), select new one
1551 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1554 cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
1556 if (cfq_cfqq_wait_request(cfqq))
1557 del_timer(&cfqd->idle_slice_timer);
1559 cfq_clear_cfqq_wait_request(cfqq);
1560 cfq_clear_cfqq_wait_busy(cfqq);
1563 * If this cfqq is shared between multiple processes, check to
1564 * make sure that those processes are still issuing I/Os within
1565 * the mean seek distance. If not, it may be time to break the
1566 * queues apart again.
1568 if (cfq_cfqq_coop(cfqq) && CFQQ_SEEKY(cfqq))
1569 cfq_mark_cfqq_split_coop(cfqq);
1572 * store what was left of this slice, if the queue idled/timed out
1574 if (timed_out && !cfq_cfqq_slice_new(cfqq)) {
1575 cfqq->slice_resid = cfqq->slice_end - jiffies;
1576 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
1579 cfq_group_served(cfqd, cfqq->cfqg, cfqq);
1581 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
1582 cfq_del_cfqq_rr(cfqd, cfqq);
1584 cfq_resort_rr_list(cfqd, cfqq);
1586 if (cfqq == cfqd->active_queue)
1587 cfqd->active_queue = NULL;
1589 if (&cfqq->cfqg->rb_node == cfqd->grp_service_tree.active)
1590 cfqd->grp_service_tree.active = NULL;
1592 if (cfqd->active_cic) {
1593 put_io_context(cfqd->active_cic->ioc);
1594 cfqd->active_cic = NULL;
1598 static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out)
1600 struct cfq_queue *cfqq = cfqd->active_queue;
1603 __cfq_slice_expired(cfqd, cfqq, timed_out);
1607 * Get next queue for service. Unless we have a queue preemption,
1608 * we'll simply select the first cfqq in the service tree.
1610 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
1612 struct cfq_rb_root *service_tree =
1613 service_tree_for(cfqd->serving_group, cfqd->serving_prio,
1614 cfqd->serving_type);
1616 if (!cfqd->rq_queued)
1619 /* There is nothing to dispatch */
1622 if (RB_EMPTY_ROOT(&service_tree->rb))
1624 return cfq_rb_first(service_tree);
1627 static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd)
1629 struct cfq_group *cfqg;
1630 struct cfq_queue *cfqq;
1632 struct cfq_rb_root *st;
1634 if (!cfqd->rq_queued)
1637 cfqg = cfq_get_next_cfqg(cfqd);
1641 for_each_cfqg_st(cfqg, i, j, st)
1642 if ((cfqq = cfq_rb_first(st)) != NULL)
1648 * Get and set a new active queue for service.
1650 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
1651 struct cfq_queue *cfqq)
1654 cfqq = cfq_get_next_queue(cfqd);
1656 __cfq_set_active_queue(cfqd, cfqq);
1660 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
1663 if (blk_rq_pos(rq) >= cfqd->last_position)
1664 return blk_rq_pos(rq) - cfqd->last_position;
1666 return cfqd->last_position - blk_rq_pos(rq);
1669 static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1672 return cfq_dist_from_last(cfqd, rq) <= CFQQ_CLOSE_THR;
1675 static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
1676 struct cfq_queue *cur_cfqq)
1678 struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
1679 struct rb_node *parent, *node;
1680 struct cfq_queue *__cfqq;
1681 sector_t sector = cfqd->last_position;
1683 if (RB_EMPTY_ROOT(root))
1687 * First, if we find a request starting at the end of the last
1688 * request, choose it.
1690 __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
1695 * If the exact sector wasn't found, the parent of the NULL leaf
1696 * will contain the closest sector.
1698 __cfqq = rb_entry(parent, struct cfq_queue, p_node);
1699 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1702 if (blk_rq_pos(__cfqq->next_rq) < sector)
1703 node = rb_next(&__cfqq->p_node);
1705 node = rb_prev(&__cfqq->p_node);
1709 __cfqq = rb_entry(node, struct cfq_queue, p_node);
1710 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1718 * cur_cfqq - passed in so that we don't decide that the current queue is
1719 * closely cooperating with itself.
1721 * So, basically we're assuming that that cur_cfqq has dispatched at least
1722 * one request, and that cfqd->last_position reflects a position on the disk
1723 * associated with the I/O issued by cur_cfqq. I'm not sure this is a valid
1726 static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
1727 struct cfq_queue *cur_cfqq)
1729 struct cfq_queue *cfqq;
1731 if (cfq_class_idle(cur_cfqq))
1733 if (!cfq_cfqq_sync(cur_cfqq))
1735 if (CFQQ_SEEKY(cur_cfqq))
1739 * Don't search priority tree if it's the only queue in the group.
1741 if (cur_cfqq->cfqg->nr_cfqq == 1)
1745 * We should notice if some of the queues are cooperating, eg
1746 * working closely on the same area of the disk. In that case,
1747 * we can group them together and don't waste time idling.
1749 cfqq = cfqq_close(cfqd, cur_cfqq);
1753 /* If new queue belongs to different cfq_group, don't choose it */
1754 if (cur_cfqq->cfqg != cfqq->cfqg)
1758 * It only makes sense to merge sync queues.
1760 if (!cfq_cfqq_sync(cfqq))
1762 if (CFQQ_SEEKY(cfqq))
1766 * Do not merge queues of different priority classes
1768 if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq))
1775 * Determine whether we should enforce idle window for this queue.
1778 static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1780 enum wl_prio_t prio = cfqq_prio(cfqq);
1781 struct cfq_rb_root *service_tree = cfqq->service_tree;
1783 BUG_ON(!service_tree);
1784 BUG_ON(!service_tree->count);
1786 /* We never do for idle class queues. */
1787 if (prio == IDLE_WORKLOAD)
1790 /* We do for queues that were marked with idle window flag. */
1791 if (cfq_cfqq_idle_window(cfqq) &&
1792 !(blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag))
1796 * Otherwise, we do only if they are the last ones
1797 * in their service tree.
1799 if (service_tree->count == 1 && cfq_cfqq_sync(cfqq))
1801 cfq_log_cfqq(cfqd, cfqq, "Not idling. st->count:%d",
1802 service_tree->count);
1806 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
1808 struct cfq_queue *cfqq = cfqd->active_queue;
1809 struct cfq_io_context *cic;
1813 * SSD device without seek penalty, disable idling. But only do so
1814 * for devices that support queuing, otherwise we still have a problem
1815 * with sync vs async workloads.
1817 if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
1820 WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
1821 WARN_ON(cfq_cfqq_slice_new(cfqq));
1824 * idle is disabled, either manually or by past process history
1826 if (!cfqd->cfq_slice_idle || !cfq_should_idle(cfqd, cfqq))
1830 * still active requests from this queue, don't idle
1832 if (cfqq->dispatched)
1836 * task has exited, don't wait
1838 cic = cfqd->active_cic;
1839 if (!cic || !atomic_read(&cic->ioc->nr_tasks))
1843 * If our average think time is larger than the remaining time
1844 * slice, then don't idle. This avoids overrunning the allotted
1847 if (sample_valid(cic->ttime_samples) &&
1848 (cfqq->slice_end - jiffies < cic->ttime_mean)) {
1849 cfq_log_cfqq(cfqd, cfqq, "Not idling. think_time:%d",
1854 cfq_mark_cfqq_wait_request(cfqq);
1856 sl = cfqd->cfq_slice_idle;
1858 mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
1859 cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu", sl);
1863 * Move request from internal lists to the request queue dispatch list.
1865 static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
1867 struct cfq_data *cfqd = q->elevator->elevator_data;
1868 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1870 cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
1872 cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
1873 cfq_remove_request(rq);
1875 elv_dispatch_sort(q, rq);
1877 cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]++;
1878 cfqq->nr_sectors += blk_rq_sectors(rq);
1882 * return expired entry, or NULL to just start from scratch in rbtree
1884 static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
1886 struct request *rq = NULL;
1888 if (cfq_cfqq_fifo_expire(cfqq))
1891 cfq_mark_cfqq_fifo_expire(cfqq);
1893 if (list_empty(&cfqq->fifo))
1896 rq = rq_entry_fifo(cfqq->fifo.next);
1897 if (time_before(jiffies, rq_fifo_time(rq)))
1900 cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq);
1905 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1907 const int base_rq = cfqd->cfq_slice_async_rq;
1909 WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
1911 return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
1915 * Must be called with the queue_lock held.
1917 static int cfqq_process_refs(struct cfq_queue *cfqq)
1919 int process_refs, io_refs;
1921 io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE];
1922 process_refs = atomic_read(&cfqq->ref) - io_refs;
1923 BUG_ON(process_refs < 0);
1924 return process_refs;
1927 static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq)
1929 int process_refs, new_process_refs;
1930 struct cfq_queue *__cfqq;
1932 /* Avoid a circular list and skip interim queue merges */
1933 while ((__cfqq = new_cfqq->new_cfqq)) {
1939 process_refs = cfqq_process_refs(cfqq);
1941 * If the process for the cfqq has gone away, there is no
1942 * sense in merging the queues.
1944 if (process_refs == 0)
1948 * Merge in the direction of the lesser amount of work.
1950 new_process_refs = cfqq_process_refs(new_cfqq);
1951 if (new_process_refs >= process_refs) {
1952 cfqq->new_cfqq = new_cfqq;
1953 atomic_add(process_refs, &new_cfqq->ref);
1955 new_cfqq->new_cfqq = cfqq;
1956 atomic_add(new_process_refs, &cfqq->ref);
1960 static enum wl_type_t cfq_choose_wl(struct cfq_data *cfqd,
1961 struct cfq_group *cfqg, enum wl_prio_t prio)
1963 struct cfq_queue *queue;
1965 bool key_valid = false;
1966 unsigned long lowest_key = 0;
1967 enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD;
1969 for (i = 0; i <= SYNC_WORKLOAD; ++i) {
1970 /* select the one with lowest rb_key */
1971 queue = cfq_rb_first(service_tree_for(cfqg, prio, i));
1973 (!key_valid || time_before(queue->rb_key, lowest_key))) {
1974 lowest_key = queue->rb_key;
1983 static void choose_service_tree(struct cfq_data *cfqd, struct cfq_group *cfqg)
1987 struct cfq_rb_root *st;
1988 unsigned group_slice;
1991 cfqd->serving_prio = IDLE_WORKLOAD;
1992 cfqd->workload_expires = jiffies + 1;
1996 /* Choose next priority. RT > BE > IDLE */
1997 if (cfq_group_busy_queues_wl(RT_WORKLOAD, cfqd, cfqg))
1998 cfqd->serving_prio = RT_WORKLOAD;
1999 else if (cfq_group_busy_queues_wl(BE_WORKLOAD, cfqd, cfqg))
2000 cfqd->serving_prio = BE_WORKLOAD;
2002 cfqd->serving_prio = IDLE_WORKLOAD;
2003 cfqd->workload_expires = jiffies + 1;
2008 * For RT and BE, we have to choose also the type
2009 * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
2012 st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
2016 * check workload expiration, and that we still have other queues ready
2018 if (count && !time_after(jiffies, cfqd->workload_expires))
2021 /* otherwise select new workload type */
2022 cfqd->serving_type =
2023 cfq_choose_wl(cfqd, cfqg, cfqd->serving_prio);
2024 st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
2028 * the workload slice is computed as a fraction of target latency
2029 * proportional to the number of queues in that workload, over
2030 * all the queues in the same priority class
2032 group_slice = cfq_group_slice(cfqd, cfqg);
2034 slice = group_slice * count /
2035 max_t(unsigned, cfqg->busy_queues_avg[cfqd->serving_prio],
2036 cfq_group_busy_queues_wl(cfqd->serving_prio, cfqd, cfqg));
2038 if (cfqd->serving_type == ASYNC_WORKLOAD) {
2042 * Async queues are currently system wide. Just taking
2043 * proportion of queues with-in same group will lead to higher
2044 * async ratio system wide as generally root group is going
2045 * to have higher weight. A more accurate thing would be to
2046 * calculate system wide asnc/sync ratio.
2048 tmp = cfq_target_latency * cfqg_busy_async_queues(cfqd, cfqg);
2049 tmp = tmp/cfqd->busy_queues;
2050 slice = min_t(unsigned, slice, tmp);
2052 /* async workload slice is scaled down according to
2053 * the sync/async slice ratio. */
2054 slice = slice * cfqd->cfq_slice[0] / cfqd->cfq_slice[1];
2056 /* sync workload slice is at least 2 * cfq_slice_idle */
2057 slice = max(slice, 2 * cfqd->cfq_slice_idle);
2059 slice = max_t(unsigned, slice, CFQ_MIN_TT);
2060 cfq_log(cfqd, "workload slice:%d", slice);
2061 cfqd->workload_expires = jiffies + slice;
2062 cfqd->noidle_tree_requires_idle = false;
2065 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd)
2067 struct cfq_rb_root *st = &cfqd->grp_service_tree;
2068 struct cfq_group *cfqg;
2070 if (RB_EMPTY_ROOT(&st->rb))
2072 cfqg = cfq_rb_first_group(st);
2073 st->active = &cfqg->rb_node;
2074 update_min_vdisktime(st);
2078 static void cfq_choose_cfqg(struct cfq_data *cfqd)
2080 struct cfq_group *cfqg = cfq_get_next_cfqg(cfqd);
2082 cfqd->serving_group = cfqg;
2084 /* Restore the workload type data */
2085 if (cfqg->saved_workload_slice) {
2086 cfqd->workload_expires = jiffies + cfqg->saved_workload_slice;
2087 cfqd->serving_type = cfqg->saved_workload;
2088 cfqd->serving_prio = cfqg->saved_serving_prio;
2090 cfqd->workload_expires = jiffies - 1;
2092 choose_service_tree(cfqd, cfqg);
2096 * Select a queue for service. If we have a current active queue,
2097 * check whether to continue servicing it, or retrieve and set a new one.
2099 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
2101 struct cfq_queue *cfqq, *new_cfqq = NULL;
2103 cfqq = cfqd->active_queue;
2107 if (!cfqd->rq_queued)
2111 * We were waiting for group to get backlogged. Expire the queue
2113 if (cfq_cfqq_wait_busy(cfqq) && !RB_EMPTY_ROOT(&cfqq->sort_list))
2117 * The active queue has run out of time, expire it and select new.
2119 if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq)) {
2121 * If slice had not expired at the completion of last request
2122 * we might not have turned on wait_busy flag. Don't expire
2123 * the queue yet. Allow the group to get backlogged.
2125 * The very fact that we have used the slice, that means we
2126 * have been idling all along on this queue and it should be
2127 * ok to wait for this request to complete.
2129 if (cfqq->cfqg->nr_cfqq == 1 && RB_EMPTY_ROOT(&cfqq->sort_list)
2130 && cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2138 * The active queue has requests and isn't expired, allow it to
2141 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
2145 * If another queue has a request waiting within our mean seek
2146 * distance, let it run. The expire code will check for close
2147 * cooperators and put the close queue at the front of the service
2148 * tree. If possible, merge the expiring queue with the new cfqq.
2150 new_cfqq = cfq_close_cooperator(cfqd, cfqq);
2152 if (!cfqq->new_cfqq)
2153 cfq_setup_merge(cfqq, new_cfqq);
2158 * No requests pending. If the active queue still has requests in
2159 * flight or is idling for a new request, allow either of these
2160 * conditions to happen (or time out) before selecting a new queue.
2162 if (timer_pending(&cfqd->idle_slice_timer) ||
2163 (cfqq->dispatched && cfq_should_idle(cfqd, cfqq))) {
2169 cfq_slice_expired(cfqd, 0);
2172 * Current queue expired. Check if we have to switch to a new
2176 cfq_choose_cfqg(cfqd);
2178 cfqq = cfq_set_active_queue(cfqd, new_cfqq);
2183 static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
2187 while (cfqq->next_rq) {
2188 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
2192 BUG_ON(!list_empty(&cfqq->fifo));
2194 /* By default cfqq is not expired if it is empty. Do it explicitly */
2195 __cfq_slice_expired(cfqq->cfqd, cfqq, 0);
2200 * Drain our current requests. Used for barriers and when switching
2201 * io schedulers on-the-fly.
2203 static int cfq_forced_dispatch(struct cfq_data *cfqd)
2205 struct cfq_queue *cfqq;
2208 while ((cfqq = cfq_get_next_queue_forced(cfqd)) != NULL)
2209 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
2211 cfq_slice_expired(cfqd, 0);
2212 BUG_ON(cfqd->busy_queues);
2214 cfq_log(cfqd, "forced_dispatch=%d", dispatched);
2218 static inline bool cfq_slice_used_soon(struct cfq_data *cfqd,
2219 struct cfq_queue *cfqq)
2221 /* the queue hasn't finished any request, can't estimate */
2222 if (cfq_cfqq_slice_new(cfqq))
2224 if (time_after(jiffies + cfqd->cfq_slice_idle * cfqq->dispatched,
2231 static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2233 unsigned int max_dispatch;
2236 * Drain async requests before we start sync IO
2238 if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_flight[BLK_RW_ASYNC])
2242 * If this is an async queue and we have sync IO in flight, let it wait
2244 if (cfqd->rq_in_flight[BLK_RW_SYNC] && !cfq_cfqq_sync(cfqq))
2247 max_dispatch = max_t(unsigned int, cfqd->cfq_quantum / 2, 1);
2248 if (cfq_class_idle(cfqq))
2252 * Does this cfqq already have too much IO in flight?
2254 if (cfqq->dispatched >= max_dispatch) {
2256 * idle queue must always only have a single IO in flight
2258 if (cfq_class_idle(cfqq))
2262 * We have other queues, don't allow more IO from this one
2264 if (cfqd->busy_queues > 1 && cfq_slice_used_soon(cfqd, cfqq))
2268 * Sole queue user, no limit
2270 if (cfqd->busy_queues == 1)
2274 * Normally we start throttling cfqq when cfq_quantum/2
2275 * requests have been dispatched. But we can drive
2276 * deeper queue depths at the beginning of slice
2277 * subjected to upper limit of cfq_quantum.
2279 max_dispatch = cfqd->cfq_quantum;
2283 * Async queues must wait a bit before being allowed dispatch.
2284 * We also ramp up the dispatch depth gradually for async IO,
2285 * based on the last sync IO we serviced
2287 if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
2288 unsigned long last_sync = jiffies - cfqd->last_delayed_sync;
2291 depth = last_sync / cfqd->cfq_slice[1];
2292 if (!depth && !cfqq->dispatched)
2294 if (depth < max_dispatch)
2295 max_dispatch = depth;
2299 * If we're below the current max, allow a dispatch
2301 return cfqq->dispatched < max_dispatch;
2305 * Dispatch a request from cfqq, moving them to the request queue
2308 static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2312 BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
2314 if (!cfq_may_dispatch(cfqd, cfqq))
2318 * follow expired path, else get first next available
2320 rq = cfq_check_fifo(cfqq);
2325 * insert request into driver dispatch list
2327 cfq_dispatch_insert(cfqd->queue, rq);
2329 if (!cfqd->active_cic) {
2330 struct cfq_io_context *cic = RQ_CIC(rq);
2332 atomic_long_inc(&cic->ioc->refcount);
2333 cfqd->active_cic = cic;
2340 * Find the cfqq that we need to service and move a request from that to the
2343 static int cfq_dispatch_requests(struct request_queue *q, int force)
2345 struct cfq_data *cfqd = q->elevator->elevator_data;
2346 struct cfq_queue *cfqq;
2348 if (!cfqd->busy_queues)
2351 if (unlikely(force))
2352 return cfq_forced_dispatch(cfqd);
2354 cfqq = cfq_select_queue(cfqd);
2359 * Dispatch a request from this cfqq, if it is allowed
2361 if (!cfq_dispatch_request(cfqd, cfqq))
2364 cfqq->slice_dispatch++;
2365 cfq_clear_cfqq_must_dispatch(cfqq);
2368 * expire an async queue immediately if it has used up its slice. idle
2369 * queue always expire after 1 dispatch round.
2371 if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
2372 cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
2373 cfq_class_idle(cfqq))) {
2374 cfqq->slice_end = jiffies + 1;
2375 cfq_slice_expired(cfqd, 0);
2378 cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
2383 * task holds one reference to the queue, dropped when task exits. each rq
2384 * in-flight on this queue also holds a reference, dropped when rq is freed.
2386 * Each cfq queue took a reference on the parent group. Drop it now.
2387 * queue lock must be held here.
2389 static void cfq_put_queue(struct cfq_queue *cfqq)
2391 struct cfq_data *cfqd = cfqq->cfqd;
2392 struct cfq_group *cfqg, *orig_cfqg;
2394 BUG_ON(atomic_read(&cfqq->ref) <= 0);
2396 if (!atomic_dec_and_test(&cfqq->ref))
2399 cfq_log_cfqq(cfqd, cfqq, "put_queue");
2400 BUG_ON(rb_first(&cfqq->sort_list));
2401 BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
2403 orig_cfqg = cfqq->orig_cfqg;
2405 if (unlikely(cfqd->active_queue == cfqq)) {
2406 __cfq_slice_expired(cfqd, cfqq, 0);
2407 cfq_schedule_dispatch(cfqd);
2410 BUG_ON(cfq_cfqq_on_rr(cfqq));
2411 kmem_cache_free(cfq_pool, cfqq);
2414 cfq_put_cfqg(orig_cfqg);
2418 * Must always be called with the rcu_read_lock() held
2421 __call_for_each_cic(struct io_context *ioc,
2422 void (*func)(struct io_context *, struct cfq_io_context *))
2424 struct cfq_io_context *cic;
2425 struct hlist_node *n;
2427 hlist_for_each_entry_rcu(cic, n, &ioc->cic_list, cic_list)
2432 * Call func for each cic attached to this ioc.
2435 call_for_each_cic(struct io_context *ioc,
2436 void (*func)(struct io_context *, struct cfq_io_context *))
2439 __call_for_each_cic(ioc, func);
2443 static void cfq_cic_free_rcu(struct rcu_head *head)
2445 struct cfq_io_context *cic;
2447 cic = container_of(head, struct cfq_io_context, rcu_head);
2449 kmem_cache_free(cfq_ioc_pool, cic);
2450 elv_ioc_count_dec(cfq_ioc_count);
2454 * CFQ scheduler is exiting, grab exit lock and check
2455 * the pending io context count. If it hits zero,
2456 * complete ioc_gone and set it back to NULL
2458 spin_lock(&ioc_gone_lock);
2459 if (ioc_gone && !elv_ioc_count_read(cfq_ioc_count)) {
2463 spin_unlock(&ioc_gone_lock);
2467 static void cfq_cic_free(struct cfq_io_context *cic)
2469 call_rcu(&cic->rcu_head, cfq_cic_free_rcu);
2472 static void cic_free_func(struct io_context *ioc, struct cfq_io_context *cic)
2474 unsigned long flags;
2476 BUG_ON(!cic->dead_key);
2478 spin_lock_irqsave(&ioc->lock, flags);
2479 radix_tree_delete(&ioc->radix_root, cic->dead_key);
2480 hlist_del_rcu(&cic->cic_list);
2481 spin_unlock_irqrestore(&ioc->lock, flags);
2487 * Must be called with rcu_read_lock() held or preemption otherwise disabled.
2488 * Only two callers of this - ->dtor() which is called with the rcu_read_lock(),
2489 * and ->trim() which is called with the task lock held
2491 static void cfq_free_io_context(struct io_context *ioc)
2494 * ioc->refcount is zero here, or we are called from elv_unregister(),
2495 * so no more cic's are allowed to be linked into this ioc. So it
2496 * should be ok to iterate over the known list, we will see all cic's
2497 * since no new ones are added.
2499 __call_for_each_cic(ioc, cic_free_func);
2502 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2504 struct cfq_queue *__cfqq, *next;
2506 if (unlikely(cfqq == cfqd->active_queue)) {
2507 __cfq_slice_expired(cfqd, cfqq, 0);
2508 cfq_schedule_dispatch(cfqd);
2512 * If this queue was scheduled to merge with another queue, be
2513 * sure to drop the reference taken on that queue (and others in
2514 * the merge chain). See cfq_setup_merge and cfq_merge_cfqqs.
2516 __cfqq = cfqq->new_cfqq;
2518 if (__cfqq == cfqq) {
2519 WARN(1, "cfqq->new_cfqq loop detected\n");
2522 next = __cfqq->new_cfqq;
2523 cfq_put_queue(__cfqq);
2527 cfq_put_queue(cfqq);
2530 static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
2531 struct cfq_io_context *cic)
2533 struct io_context *ioc = cic->ioc;
2535 list_del_init(&cic->queue_list);
2538 * Make sure key == NULL is seen for dead queues
2541 cic->dead_key = (unsigned long) cic->key;
2544 if (ioc->ioc_data == cic)
2545 rcu_assign_pointer(ioc->ioc_data, NULL);
2547 if (cic->cfqq[BLK_RW_ASYNC]) {
2548 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]);
2549 cic->cfqq[BLK_RW_ASYNC] = NULL;
2552 if (cic->cfqq[BLK_RW_SYNC]) {
2553 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]);
2554 cic->cfqq[BLK_RW_SYNC] = NULL;
2558 static void cfq_exit_single_io_context(struct io_context *ioc,
2559 struct cfq_io_context *cic)
2561 struct cfq_data *cfqd = cic->key;
2564 struct request_queue *q = cfqd->queue;
2565 unsigned long flags;
2567 spin_lock_irqsave(q->queue_lock, flags);
2570 * Ensure we get a fresh copy of the ->key to prevent
2571 * race between exiting task and queue
2573 smp_read_barrier_depends();
2575 __cfq_exit_single_io_context(cfqd, cic);
2577 spin_unlock_irqrestore(q->queue_lock, flags);
2582 * The process that ioc belongs to has exited, we need to clean up
2583 * and put the internal structures we have that belongs to that process.
2585 static void cfq_exit_io_context(struct io_context *ioc)
2587 call_for_each_cic(ioc, cfq_exit_single_io_context);
2590 static struct cfq_io_context *
2591 cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
2593 struct cfq_io_context *cic;
2595 cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO,
2598 cic->last_end_request = jiffies;
2599 INIT_LIST_HEAD(&cic->queue_list);
2600 INIT_HLIST_NODE(&cic->cic_list);
2601 cic->dtor = cfq_free_io_context;
2602 cic->exit = cfq_exit_io_context;
2603 elv_ioc_count_inc(cfq_ioc_count);
2609 static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc)
2611 struct task_struct *tsk = current;
2614 if (!cfq_cfqq_prio_changed(cfqq))
2617 ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio);
2618 switch (ioprio_class) {
2620 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
2621 case IOPRIO_CLASS_NONE:
2623 * no prio set, inherit CPU scheduling settings
2625 cfqq->ioprio = task_nice_ioprio(tsk);
2626 cfqq->ioprio_class = task_nice_ioclass(tsk);
2628 case IOPRIO_CLASS_RT:
2629 cfqq->ioprio = task_ioprio(ioc);
2630 cfqq->ioprio_class = IOPRIO_CLASS_RT;
2632 case IOPRIO_CLASS_BE:
2633 cfqq->ioprio = task_ioprio(ioc);
2634 cfqq->ioprio_class = IOPRIO_CLASS_BE;
2636 case IOPRIO_CLASS_IDLE:
2637 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
2639 cfq_clear_cfqq_idle_window(cfqq);
2644 * keep track of original prio settings in case we have to temporarily
2645 * elevate the priority of this queue
2647 cfqq->org_ioprio = cfqq->ioprio;
2648 cfqq->org_ioprio_class = cfqq->ioprio_class;
2649 cfq_clear_cfqq_prio_changed(cfqq);
2652 static void changed_ioprio(struct io_context *ioc, struct cfq_io_context *cic)
2654 struct cfq_data *cfqd = cic->key;
2655 struct cfq_queue *cfqq;
2656 unsigned long flags;
2658 if (unlikely(!cfqd))
2661 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2663 cfqq = cic->cfqq[BLK_RW_ASYNC];
2665 struct cfq_queue *new_cfqq;
2666 new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic->ioc,
2669 cic->cfqq[BLK_RW_ASYNC] = new_cfqq;
2670 cfq_put_queue(cfqq);
2674 cfqq = cic->cfqq[BLK_RW_SYNC];
2676 cfq_mark_cfqq_prio_changed(cfqq);
2678 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2681 static void cfq_ioc_set_ioprio(struct io_context *ioc)
2683 call_for_each_cic(ioc, changed_ioprio);
2684 ioc->ioprio_changed = 0;
2687 static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2688 pid_t pid, bool is_sync)
2690 RB_CLEAR_NODE(&cfqq->rb_node);
2691 RB_CLEAR_NODE(&cfqq->p_node);
2692 INIT_LIST_HEAD(&cfqq->fifo);
2694 atomic_set(&cfqq->ref, 0);
2697 cfq_mark_cfqq_prio_changed(cfqq);
2700 if (!cfq_class_idle(cfqq))
2701 cfq_mark_cfqq_idle_window(cfqq);
2702 cfq_mark_cfqq_sync(cfqq);
2707 #ifdef CONFIG_CFQ_GROUP_IOSCHED
2708 static void changed_cgroup(struct io_context *ioc, struct cfq_io_context *cic)
2710 struct cfq_queue *sync_cfqq = cic_to_cfqq(cic, 1);
2711 struct cfq_data *cfqd = cic->key;
2712 unsigned long flags;
2713 struct request_queue *q;
2715 if (unlikely(!cfqd))
2720 spin_lock_irqsave(q->queue_lock, flags);
2724 * Drop reference to sync queue. A new sync queue will be
2725 * assigned in new group upon arrival of a fresh request.
2727 cfq_log_cfqq(cfqd, sync_cfqq, "changed cgroup");
2728 cic_set_cfqq(cic, NULL, 1);
2729 cfq_put_queue(sync_cfqq);
2732 spin_unlock_irqrestore(q->queue_lock, flags);
2735 static void cfq_ioc_set_cgroup(struct io_context *ioc)
2737 call_for_each_cic(ioc, changed_cgroup);
2738 ioc->cgroup_changed = 0;
2740 #endif /* CONFIG_CFQ_GROUP_IOSCHED */
2742 static struct cfq_queue *
2743 cfq_find_alloc_queue(struct cfq_data *cfqd, bool is_sync,
2744 struct io_context *ioc, gfp_t gfp_mask)
2746 struct cfq_queue *cfqq, *new_cfqq = NULL;
2747 struct cfq_io_context *cic;
2748 struct cfq_group *cfqg;
2751 cfqg = cfq_get_cfqg(cfqd, 1);
2752 cic = cfq_cic_lookup(cfqd, ioc);
2753 /* cic always exists here */
2754 cfqq = cic_to_cfqq(cic, is_sync);
2757 * Always try a new alloc if we fell back to the OOM cfqq
2758 * originally, since it should just be a temporary situation.
2760 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
2765 } else if (gfp_mask & __GFP_WAIT) {
2766 spin_unlock_irq(cfqd->queue->queue_lock);
2767 new_cfqq = kmem_cache_alloc_node(cfq_pool,
2768 gfp_mask | __GFP_ZERO,
2770 spin_lock_irq(cfqd->queue->queue_lock);
2774 cfqq = kmem_cache_alloc_node(cfq_pool,
2775 gfp_mask | __GFP_ZERO,
2780 cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
2781 cfq_init_prio_data(cfqq, ioc);
2782 cfq_link_cfqq_cfqg(cfqq, cfqg);
2783 cfq_log_cfqq(cfqd, cfqq, "alloced");
2785 cfqq = &cfqd->oom_cfqq;
2789 kmem_cache_free(cfq_pool, new_cfqq);
2794 static struct cfq_queue **
2795 cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
2797 switch (ioprio_class) {
2798 case IOPRIO_CLASS_RT:
2799 return &cfqd->async_cfqq[0][ioprio];
2800 case IOPRIO_CLASS_BE:
2801 return &cfqd->async_cfqq[1][ioprio];
2802 case IOPRIO_CLASS_IDLE:
2803 return &cfqd->async_idle_cfqq;
2809 static struct cfq_queue *
2810 cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct io_context *ioc,
2813 const int ioprio = task_ioprio(ioc);
2814 const int ioprio_class = task_ioprio_class(ioc);
2815 struct cfq_queue **async_cfqq = NULL;
2816 struct cfq_queue *cfqq = NULL;
2819 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
2824 cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask);
2827 * pin the queue now that it's allocated, scheduler exit will prune it
2829 if (!is_sync && !(*async_cfqq)) {
2830 atomic_inc(&cfqq->ref);
2834 atomic_inc(&cfqq->ref);
2839 * We drop cfq io contexts lazily, so we may find a dead one.
2842 cfq_drop_dead_cic(struct cfq_data *cfqd, struct io_context *ioc,
2843 struct cfq_io_context *cic)
2845 unsigned long flags;
2847 WARN_ON(!list_empty(&cic->queue_list));
2849 spin_lock_irqsave(&ioc->lock, flags);
2851 BUG_ON(ioc->ioc_data == cic);
2853 radix_tree_delete(&ioc->radix_root, (unsigned long) cfqd);
2854 hlist_del_rcu(&cic->cic_list);
2855 spin_unlock_irqrestore(&ioc->lock, flags);
2860 static struct cfq_io_context *
2861 cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc)
2863 struct cfq_io_context *cic;
2864 unsigned long flags;
2873 * we maintain a last-hit cache, to avoid browsing over the tree
2875 cic = rcu_dereference(ioc->ioc_data);
2876 if (cic && cic->key == cfqd) {
2882 cic = radix_tree_lookup(&ioc->radix_root, (unsigned long) cfqd);
2886 /* ->key must be copied to avoid race with cfq_exit_queue() */
2889 cfq_drop_dead_cic(cfqd, ioc, cic);
2894 spin_lock_irqsave(&ioc->lock, flags);
2895 rcu_assign_pointer(ioc->ioc_data, cic);
2896 spin_unlock_irqrestore(&ioc->lock, flags);
2904 * Add cic into ioc, using cfqd as the search key. This enables us to lookup
2905 * the process specific cfq io context when entered from the block layer.
2906 * Also adds the cic to a per-cfqd list, used when this queue is removed.
2908 static int cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
2909 struct cfq_io_context *cic, gfp_t gfp_mask)
2911 unsigned long flags;
2914 ret = radix_tree_preload(gfp_mask);
2919 spin_lock_irqsave(&ioc->lock, flags);
2920 ret = radix_tree_insert(&ioc->radix_root,
2921 (unsigned long) cfqd, cic);
2923 hlist_add_head_rcu(&cic->cic_list, &ioc->cic_list);
2924 spin_unlock_irqrestore(&ioc->lock, flags);
2926 radix_tree_preload_end();
2929 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2930 list_add(&cic->queue_list, &cfqd->cic_list);
2931 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2936 printk(KERN_ERR "cfq: cic link failed!\n");
2942 * Setup general io context and cfq io context. There can be several cfq
2943 * io contexts per general io context, if this process is doing io to more
2944 * than one device managed by cfq.
2946 static struct cfq_io_context *
2947 cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
2949 struct io_context *ioc = NULL;
2950 struct cfq_io_context *cic;
2952 might_sleep_if(gfp_mask & __GFP_WAIT);
2954 ioc = get_io_context(gfp_mask, cfqd->queue->node);
2958 cic = cfq_cic_lookup(cfqd, ioc);
2962 cic = cfq_alloc_io_context(cfqd, gfp_mask);
2966 if (cfq_cic_link(cfqd, ioc, cic, gfp_mask))
2970 smp_read_barrier_depends();
2971 if (unlikely(ioc->ioprio_changed))
2972 cfq_ioc_set_ioprio(ioc);
2974 #ifdef CONFIG_CFQ_GROUP_IOSCHED
2975 if (unlikely(ioc->cgroup_changed))
2976 cfq_ioc_set_cgroup(ioc);
2982 put_io_context(ioc);
2987 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
2989 unsigned long elapsed = jiffies - cic->last_end_request;
2990 unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
2992 cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
2993 cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
2994 cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
2998 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3002 sector_t n_sec = blk_rq_sectors(rq);
3003 if (cfqq->last_request_pos) {
3004 if (cfqq->last_request_pos < blk_rq_pos(rq))
3005 sdist = blk_rq_pos(rq) - cfqq->last_request_pos;
3007 sdist = cfqq->last_request_pos - blk_rq_pos(rq);
3010 cfqq->seek_history <<= 1;
3011 if (blk_queue_nonrot(cfqd->queue))
3012 cfqq->seek_history |= (n_sec < CFQQ_SECT_THR_NONROT);
3014 cfqq->seek_history |= (sdist > CFQQ_SEEK_THR);
3018 * Disable idle window if the process thinks too long or seeks so much that
3022 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3023 struct cfq_io_context *cic)
3025 int old_idle, enable_idle;
3028 * Don't idle for async or idle io prio class
3030 if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
3033 enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
3035 if (cfqq->queued[0] + cfqq->queued[1] >= 4)
3036 cfq_mark_cfqq_deep(cfqq);
3038 if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle ||
3039 (!cfq_cfqq_deep(cfqq) && CFQQ_SEEKY(cfqq)))
3041 else if (sample_valid(cic->ttime_samples)) {
3042 if (cic->ttime_mean > cfqd->cfq_slice_idle)
3048 if (old_idle != enable_idle) {
3049 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
3051 cfq_mark_cfqq_idle_window(cfqq);
3053 cfq_clear_cfqq_idle_window(cfqq);
3058 * Check if new_cfqq should preempt the currently active queue. Return 0 for
3059 * no or if we aren't sure, a 1 will cause a preempt.
3062 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
3065 struct cfq_queue *cfqq;
3067 cfqq = cfqd->active_queue;
3071 if (cfq_class_idle(new_cfqq))
3074 if (cfq_class_idle(cfqq))
3078 * Don't allow a non-RT request to preempt an ongoing RT cfqq timeslice.
3080 if (cfq_class_rt(cfqq) && !cfq_class_rt(new_cfqq))
3084 * if the new request is sync, but the currently running queue is
3085 * not, let the sync request have priority.
3087 if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
3090 if (new_cfqq->cfqg != cfqq->cfqg)
3093 if (cfq_slice_used(cfqq))
3096 /* Allow preemption only if we are idling on sync-noidle tree */
3097 if (cfqd->serving_type == SYNC_NOIDLE_WORKLOAD &&
3098 cfqq_type(new_cfqq) == SYNC_NOIDLE_WORKLOAD &&
3099 new_cfqq->service_tree->count == 2 &&
3100 RB_EMPTY_ROOT(&cfqq->sort_list))
3104 * So both queues are sync. Let the new request get disk time if
3105 * it's a metadata request and the current queue is doing regular IO.
3107 if (rq_is_meta(rq) && !cfqq->meta_pending)
3111 * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
3113 if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
3116 if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
3120 * if this request is as-good as one we would expect from the
3121 * current cfqq, let it preempt
3123 if (cfq_rq_close(cfqd, cfqq, rq))
3130 * cfqq preempts the active queue. if we allowed preempt with no slice left,
3131 * let it have half of its nominal slice.
3133 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3135 cfq_log_cfqq(cfqd, cfqq, "preempt");
3136 cfq_slice_expired(cfqd, 1);
3139 * Put the new queue at the front of the of the current list,
3140 * so we know that it will be selected next.
3142 BUG_ON(!cfq_cfqq_on_rr(cfqq));
3144 cfq_service_tree_add(cfqd, cfqq, 1);
3146 cfqq->slice_end = 0;
3147 cfq_mark_cfqq_slice_new(cfqq);
3151 * Called when a new fs request (rq) is added (to cfqq). Check if there's
3152 * something we should do about it
3155 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3158 struct cfq_io_context *cic = RQ_CIC(rq);
3162 cfqq->meta_pending++;
3164 cfq_update_io_thinktime(cfqd, cic);
3165 cfq_update_io_seektime(cfqd, cfqq, rq);
3166 cfq_update_idle_window(cfqd, cfqq, cic);
3168 cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
3170 if (cfqq == cfqd->active_queue) {
3172 * Remember that we saw a request from this process, but
3173 * don't start queuing just yet. Otherwise we risk seeing lots
3174 * of tiny requests, because we disrupt the normal plugging
3175 * and merging. If the request is already larger than a single
3176 * page, let it rip immediately. For that case we assume that
3177 * merging is already done. Ditto for a busy system that
3178 * has other work pending, don't risk delaying until the
3179 * idle timer unplug to continue working.
3181 if (cfq_cfqq_wait_request(cfqq)) {
3182 if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
3183 cfqd->busy_queues > 1) {
3184 del_timer(&cfqd->idle_slice_timer);
3185 cfq_clear_cfqq_wait_request(cfqq);
3186 __blk_run_queue(cfqd->queue);
3188 cfq_mark_cfqq_must_dispatch(cfqq);
3190 } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
3192 * not the active queue - expire current slice if it is
3193 * idle and has expired it's mean thinktime or this new queue
3194 * has some old slice time left and is of higher priority or
3195 * this new queue is RT and the current one is BE
3197 cfq_preempt_queue(cfqd, cfqq);
3198 __blk_run_queue(cfqd->queue);
3202 static void cfq_insert_request(struct request_queue *q, struct request *rq)
3204 struct cfq_data *cfqd = q->elevator->elevator_data;
3205 struct cfq_queue *cfqq = RQ_CFQQ(rq);
3207 cfq_log_cfqq(cfqd, cfqq, "insert_request");
3208 cfq_init_prio_data(cfqq, RQ_CIC(rq)->ioc);
3210 rq_set_fifo_time(rq, jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)]);
3211 list_add_tail(&rq->queuelist, &cfqq->fifo);
3214 cfq_rq_enqueued(cfqd, cfqq, rq);
3218 * Update hw_tag based on peak queue depth over 50 samples under
3221 static void cfq_update_hw_tag(struct cfq_data *cfqd)
3223 struct cfq_queue *cfqq = cfqd->active_queue;
3225 if (cfqd->rq_in_driver > cfqd->hw_tag_est_depth)
3226 cfqd->hw_tag_est_depth = cfqd->rq_in_driver;
3228 if (cfqd->hw_tag == 1)
3231 if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
3232 cfqd->rq_in_driver <= CFQ_HW_QUEUE_MIN)
3236 * If active queue hasn't enough requests and can idle, cfq might not
3237 * dispatch sufficient requests to hardware. Don't zero hw_tag in this
3240 if (cfqq && cfq_cfqq_idle_window(cfqq) &&
3241 cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] <
3242 CFQ_HW_QUEUE_MIN && cfqd->rq_in_driver < CFQ_HW_QUEUE_MIN)
3245 if (cfqd->hw_tag_samples++ < 50)
3248 if (cfqd->hw_tag_est_depth >= CFQ_HW_QUEUE_MIN)
3254 static bool cfq_should_wait_busy(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3256 struct cfq_io_context *cic = cfqd->active_cic;
3258 /* If there are other queues in the group, don't wait */
3259 if (cfqq->cfqg->nr_cfqq > 1)
3262 if (cfq_slice_used(cfqq))
3265 /* if slice left is less than think time, wait busy */
3266 if (cic && sample_valid(cic->ttime_samples)
3267 && (cfqq->slice_end - jiffies < cic->ttime_mean))
3271 * If think times is less than a jiffy than ttime_mean=0 and above
3272 * will not be true. It might happen that slice has not expired yet
3273 * but will expire soon (4-5 ns) during select_queue(). To cover the
3274 * case where think time is less than a jiffy, mark the queue wait
3275 * busy if only 1 jiffy is left in the slice.
3277 if (cfqq->slice_end - jiffies == 1)
3283 static void cfq_completed_request(struct request_queue *q, struct request *rq)
3285 struct cfq_queue *cfqq = RQ_CFQQ(rq);
3286 struct cfq_data *cfqd = cfqq->cfqd;
3287 const int sync = rq_is_sync(rq);
3291 cfq_log_cfqq(cfqd, cfqq, "complete rqnoidle %d", !!rq_noidle(rq));
3293 cfq_update_hw_tag(cfqd);
3295 WARN_ON(!cfqd->rq_in_driver);
3296 WARN_ON(!cfqq->dispatched);
3297 cfqd->rq_in_driver--;
3300 cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]--;
3303 RQ_CIC(rq)->last_end_request = now;
3304 if (!time_after(rq->start_time + cfqd->cfq_fifo_expire[1], now))
3305 cfqd->last_delayed_sync = now;
3309 * If this is the active queue, check if it needs to be expired,
3310 * or if we want to idle in case it has no pending requests.
3312 if (cfqd->active_queue == cfqq) {
3313 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
3315 if (cfq_cfqq_slice_new(cfqq)) {
3316 cfq_set_prio_slice(cfqd, cfqq);
3317 cfq_clear_cfqq_slice_new(cfqq);
3321 * Should we wait for next request to come in before we expire
3324 if (cfq_should_wait_busy(cfqd, cfqq)) {
3325 cfqq->slice_end = jiffies + cfqd->cfq_slice_idle;
3326 cfq_mark_cfqq_wait_busy(cfqq);
3327 cfq_log_cfqq(cfqd, cfqq, "will busy wait");
3331 * Idling is not enabled on:
3333 * - idle-priority queues
3335 * - queues with still some requests queued
3336 * - when there is a close cooperator
3338 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
3339 cfq_slice_expired(cfqd, 1);
3340 else if (sync && cfqq_empty &&
3341 !cfq_close_cooperator(cfqd, cfqq)) {
3342 cfqd->noidle_tree_requires_idle |= !rq_noidle(rq);
3344 * Idling is enabled for SYNC_WORKLOAD.
3345 * SYNC_NOIDLE_WORKLOAD idles at the end of the tree
3346 * only if we processed at least one !rq_noidle request
3348 if (cfqd->serving_type == SYNC_WORKLOAD
3349 || cfqd->noidle_tree_requires_idle
3350 || cfqq->cfqg->nr_cfqq == 1)
3351 cfq_arm_slice_timer(cfqd);
3355 if (!cfqd->rq_in_driver)
3356 cfq_schedule_dispatch(cfqd);
3360 * we temporarily boost lower priority queues if they are holding fs exclusive
3361 * resources. they are boosted to normal prio (CLASS_BE/4)
3363 static void cfq_prio_boost(struct cfq_queue *cfqq)
3365 if (has_fs_excl()) {
3367 * boost idle prio on transactions that would lock out other
3368 * users of the filesystem
3370 if (cfq_class_idle(cfqq))
3371 cfqq->ioprio_class = IOPRIO_CLASS_BE;
3372 if (cfqq->ioprio > IOPRIO_NORM)
3373 cfqq->ioprio = IOPRIO_NORM;
3376 * unboost the queue (if needed)
3378 cfqq->ioprio_class = cfqq->org_ioprio_class;
3379 cfqq->ioprio = cfqq->org_ioprio;
3383 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
3385 if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
3386 cfq_mark_cfqq_must_alloc_slice(cfqq);
3387 return ELV_MQUEUE_MUST;
3390 return ELV_MQUEUE_MAY;
3393 static int cfq_may_queue(struct request_queue *q, int rw)
3395 struct cfq_data *cfqd = q->elevator->elevator_data;
3396 struct task_struct *tsk = current;
3397 struct cfq_io_context *cic;
3398 struct cfq_queue *cfqq;
3401 * don't force setup of a queue from here, as a call to may_queue
3402 * does not necessarily imply that a request actually will be queued.
3403 * so just lookup a possibly existing queue, or return 'may queue'
3406 cic = cfq_cic_lookup(cfqd, tsk->io_context);
3408 return ELV_MQUEUE_MAY;
3410 cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
3412 cfq_init_prio_data(cfqq, cic->ioc);
3413 cfq_prio_boost(cfqq);
3415 return __cfq_may_queue(cfqq);
3418 return ELV_MQUEUE_MAY;
3422 * queue lock held here
3424 static void cfq_put_request(struct request *rq)
3426 struct cfq_queue *cfqq = RQ_CFQQ(rq);
3429 const int rw = rq_data_dir(rq);
3431 BUG_ON(!cfqq->allocated[rw]);
3432 cfqq->allocated[rw]--;
3434 put_io_context(RQ_CIC(rq)->ioc);
3436 rq->elevator_private = NULL;
3437 rq->elevator_private2 = NULL;
3439 cfq_put_queue(cfqq);
3443 static struct cfq_queue *
3444 cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_context *cic,
3445 struct cfq_queue *cfqq)
3447 cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq);
3448 cic_set_cfqq(cic, cfqq->new_cfqq, 1);
3449 cfq_mark_cfqq_coop(cfqq->new_cfqq);
3450 cfq_put_queue(cfqq);
3451 return cic_to_cfqq(cic, 1);
3455 * Returns NULL if a new cfqq should be allocated, or the old cfqq if this
3456 * was the last process referring to said cfqq.
3458 static struct cfq_queue *
3459 split_cfqq(struct cfq_io_context *cic, struct cfq_queue *cfqq)
3461 if (cfqq_process_refs(cfqq) == 1) {
3462 cfqq->pid = current->pid;
3463 cfq_clear_cfqq_coop(cfqq);
3464 cfq_clear_cfqq_split_coop(cfqq);
3468 cic_set_cfqq(cic, NULL, 1);
3469 cfq_put_queue(cfqq);
3473 * Allocate cfq data structures associated with this request.
3476 cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
3478 struct cfq_data *cfqd = q->elevator->elevator_data;
3479 struct cfq_io_context *cic;
3480 const int rw = rq_data_dir(rq);
3481 const bool is_sync = rq_is_sync(rq);
3482 struct cfq_queue *cfqq;
3483 unsigned long flags;
3485 might_sleep_if(gfp_mask & __GFP_WAIT);
3487 cic = cfq_get_io_context(cfqd, gfp_mask);
3489 spin_lock_irqsave(q->queue_lock, flags);
3495 cfqq = cic_to_cfqq(cic, is_sync);
3496 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
3497 cfqq = cfq_get_queue(cfqd, is_sync, cic->ioc, gfp_mask);
3498 cic_set_cfqq(cic, cfqq, is_sync);
3501 * If the queue was seeky for too long, break it apart.
3503 if (cfq_cfqq_coop(cfqq) && cfq_cfqq_split_coop(cfqq)) {
3504 cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq");
3505 cfqq = split_cfqq(cic, cfqq);
3511 * Check to see if this queue is scheduled to merge with
3512 * another, closely cooperating queue. The merging of
3513 * queues happens here as it must be done in process context.
3514 * The reference on new_cfqq was taken in merge_cfqqs.
3517 cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq);
3520 cfqq->allocated[rw]++;
3521 atomic_inc(&cfqq->ref);
3523 spin_unlock_irqrestore(q->queue_lock, flags);
3525 rq->elevator_private = cic;
3526 rq->elevator_private2 = cfqq;
3531 put_io_context(cic->ioc);
3533 cfq_schedule_dispatch(cfqd);
3534 spin_unlock_irqrestore(q->queue_lock, flags);
3535 cfq_log(cfqd, "set_request fail");
3539 static void cfq_kick_queue(struct work_struct *work)
3541 struct cfq_data *cfqd =
3542 container_of(work, struct cfq_data, unplug_work);
3543 struct request_queue *q = cfqd->queue;
3545 spin_lock_irq(q->queue_lock);
3546 __blk_run_queue(cfqd->queue);
3547 spin_unlock_irq(q->queue_lock);
3551 * Timer running if the active_queue is currently idling inside its time slice
3553 static void cfq_idle_slice_timer(unsigned long data)
3555 struct cfq_data *cfqd = (struct cfq_data *) data;
3556 struct cfq_queue *cfqq;
3557 unsigned long flags;
3560 cfq_log(cfqd, "idle timer fired");
3562 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
3564 cfqq = cfqd->active_queue;
3569 * We saw a request before the queue expired, let it through
3571 if (cfq_cfqq_must_dispatch(cfqq))
3577 if (cfq_slice_used(cfqq))
3581 * only expire and reinvoke request handler, if there are
3582 * other queues with pending requests
3584 if (!cfqd->busy_queues)
3588 * not expired and it has a request pending, let it dispatch
3590 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3594 * Queue depth flag is reset only when the idle didn't succeed
3596 cfq_clear_cfqq_deep(cfqq);
3599 cfq_slice_expired(cfqd, timed_out);
3601 cfq_schedule_dispatch(cfqd);
3603 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
3606 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
3608 del_timer_sync(&cfqd->idle_slice_timer);
3609 cancel_work_sync(&cfqd->unplug_work);
3612 static void cfq_put_async_queues(struct cfq_data *cfqd)
3616 for (i = 0; i < IOPRIO_BE_NR; i++) {
3617 if (cfqd->async_cfqq[0][i])
3618 cfq_put_queue(cfqd->async_cfqq[0][i]);
3619 if (cfqd->async_cfqq[1][i])
3620 cfq_put_queue(cfqd->async_cfqq[1][i]);
3623 if (cfqd->async_idle_cfqq)
3624 cfq_put_queue(cfqd->async_idle_cfqq);
3627 static void cfq_cfqd_free(struct rcu_head *head)
3629 kfree(container_of(head, struct cfq_data, rcu));
3632 static void cfq_exit_queue(struct elevator_queue *e)
3634 struct cfq_data *cfqd = e->elevator_data;
3635 struct request_queue *q = cfqd->queue;
3637 cfq_shutdown_timer_wq(cfqd);
3639 spin_lock_irq(q->queue_lock);
3641 if (cfqd->active_queue)
3642 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
3644 while (!list_empty(&cfqd->cic_list)) {
3645 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
3646 struct cfq_io_context,
3649 __cfq_exit_single_io_context(cfqd, cic);
3652 cfq_put_async_queues(cfqd);
3653 cfq_release_cfq_groups(cfqd);
3654 blkiocg_del_blkio_group(&cfqd->root_group.blkg);
3656 spin_unlock_irq(q->queue_lock);
3658 cfq_shutdown_timer_wq(cfqd);
3660 /* Wait for cfqg->blkg->key accessors to exit their grace periods. */
3661 call_rcu(&cfqd->rcu, cfq_cfqd_free);
3664 static void *cfq_init_queue(struct request_queue *q)
3666 struct cfq_data *cfqd;
3668 struct cfq_group *cfqg;
3669 struct cfq_rb_root *st;
3671 cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
3675 /* Init root service tree */
3676 cfqd->grp_service_tree = CFQ_RB_ROOT;
3678 /* Init root group */
3679 cfqg = &cfqd->root_group;
3680 for_each_cfqg_st(cfqg, i, j, st)
3682 RB_CLEAR_NODE(&cfqg->rb_node);
3684 /* Give preference to root group over other groups */
3685 cfqg->weight = 2*BLKIO_WEIGHT_DEFAULT;
3687 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3689 * Take a reference to root group which we never drop. This is just
3690 * to make sure that cfq_put_cfqg() does not try to kfree root group
3692 atomic_set(&cfqg->ref, 1);
3693 blkiocg_add_blkio_group(&blkio_root_cgroup, &cfqg->blkg, (void *)cfqd,
3697 * Not strictly needed (since RB_ROOT just clears the node and we
3698 * zeroed cfqd on alloc), but better be safe in case someone decides
3699 * to add magic to the rb code
3701 for (i = 0; i < CFQ_PRIO_LISTS; i++)
3702 cfqd->prio_trees[i] = RB_ROOT;
3705 * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
3706 * Grab a permanent reference to it, so that the normal code flow
3707 * will not attempt to free it.
3709 cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
3710 atomic_inc(&cfqd->oom_cfqq.ref);
3711 cfq_link_cfqq_cfqg(&cfqd->oom_cfqq, &cfqd->root_group);
3713 INIT_LIST_HEAD(&cfqd->cic_list);
3717 init_timer(&cfqd->idle_slice_timer);
3718 cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
3719 cfqd->idle_slice_timer.data = (unsigned long) cfqd;
3721 INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
3723 cfqd->cfq_quantum = cfq_quantum;
3724 cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
3725 cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
3726 cfqd->cfq_back_max = cfq_back_max;
3727 cfqd->cfq_back_penalty = cfq_back_penalty;
3728 cfqd->cfq_slice[0] = cfq_slice_async;
3729 cfqd->cfq_slice[1] = cfq_slice_sync;
3730 cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
3731 cfqd->cfq_slice_idle = cfq_slice_idle;
3732 cfqd->cfq_latency = 1;
3733 cfqd->cfq_group_isolation = 0;
3736 * we optimistically start assuming sync ops weren't delayed in last
3737 * second, in order to have larger depth for async operations.
3739 cfqd->last_delayed_sync = jiffies - HZ;
3740 INIT_RCU_HEAD(&cfqd->rcu);
3744 static void cfq_slab_kill(void)
3747 * Caller already ensured that pending RCU callbacks are completed,
3748 * so we should have no busy allocations at this point.
3751 kmem_cache_destroy(cfq_pool);
3753 kmem_cache_destroy(cfq_ioc_pool);
3756 static int __init cfq_slab_setup(void)
3758 cfq_pool = KMEM_CACHE(cfq_queue, 0);
3762 cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0);
3773 * sysfs parts below -->
3776 cfq_var_show(unsigned int var, char *page)
3778 return sprintf(page, "%d\n", var);
3782 cfq_var_store(unsigned int *var, const char *page, size_t count)
3784 char *p = (char *) page;
3786 *var = simple_strtoul(p, &p, 10);
3790 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
3791 static ssize_t __FUNC(struct elevator_queue *e, char *page) \
3793 struct cfq_data *cfqd = e->elevator_data; \
3794 unsigned int __data = __VAR; \
3796 __data = jiffies_to_msecs(__data); \
3797 return cfq_var_show(__data, (page)); \
3799 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
3800 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
3801 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
3802 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
3803 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
3804 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
3805 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
3806 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
3807 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
3808 SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
3809 SHOW_FUNCTION(cfq_group_isolation_show, cfqd->cfq_group_isolation, 0);
3810 #undef SHOW_FUNCTION
3812 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
3813 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
3815 struct cfq_data *cfqd = e->elevator_data; \
3816 unsigned int __data; \
3817 int ret = cfq_var_store(&__data, (page), count); \
3818 if (__data < (MIN)) \
3820 else if (__data > (MAX)) \
3823 *(__PTR) = msecs_to_jiffies(__data); \
3825 *(__PTR) = __data; \
3828 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
3829 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
3831 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
3833 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
3834 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
3836 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
3837 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
3838 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
3839 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
3841 STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
3842 STORE_FUNCTION(cfq_group_isolation_store, &cfqd->cfq_group_isolation, 0, 1, 0);
3843 #undef STORE_FUNCTION
3845 #define CFQ_ATTR(name) \
3846 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
3848 static struct elv_fs_entry cfq_attrs[] = {
3850 CFQ_ATTR(fifo_expire_sync),
3851 CFQ_ATTR(fifo_expire_async),
3852 CFQ_ATTR(back_seek_max),
3853 CFQ_ATTR(back_seek_penalty),
3854 CFQ_ATTR(slice_sync),
3855 CFQ_ATTR(slice_async),
3856 CFQ_ATTR(slice_async_rq),
3857 CFQ_ATTR(slice_idle),
3858 CFQ_ATTR(low_latency),
3859 CFQ_ATTR(group_isolation),
3863 static struct elevator_type iosched_cfq = {
3865 .elevator_merge_fn = cfq_merge,
3866 .elevator_merged_fn = cfq_merged_request,
3867 .elevator_merge_req_fn = cfq_merged_requests,
3868 .elevator_allow_merge_fn = cfq_allow_merge,
3869 .elevator_dispatch_fn = cfq_dispatch_requests,
3870 .elevator_add_req_fn = cfq_insert_request,
3871 .elevator_activate_req_fn = cfq_activate_request,
3872 .elevator_deactivate_req_fn = cfq_deactivate_request,
3873 .elevator_queue_empty_fn = cfq_queue_empty,
3874 .elevator_completed_req_fn = cfq_completed_request,
3875 .elevator_former_req_fn = elv_rb_former_request,
3876 .elevator_latter_req_fn = elv_rb_latter_request,
3877 .elevator_set_req_fn = cfq_set_request,
3878 .elevator_put_req_fn = cfq_put_request,
3879 .elevator_may_queue_fn = cfq_may_queue,
3880 .elevator_init_fn = cfq_init_queue,
3881 .elevator_exit_fn = cfq_exit_queue,
3882 .trim = cfq_free_io_context,
3884 .elevator_attrs = cfq_attrs,
3885 .elevator_name = "cfq",
3886 .elevator_owner = THIS_MODULE,
3889 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3890 static struct blkio_policy_type blkio_policy_cfq = {
3892 .blkio_unlink_group_fn = cfq_unlink_blkio_group,
3893 .blkio_update_group_weight_fn = cfq_update_blkio_group_weight,
3897 static struct blkio_policy_type blkio_policy_cfq;
3900 static int __init cfq_init(void)
3903 * could be 0 on HZ < 1000 setups
3905 if (!cfq_slice_async)
3906 cfq_slice_async = 1;
3907 if (!cfq_slice_idle)
3910 if (cfq_slab_setup())
3913 elv_register(&iosched_cfq);
3914 blkio_policy_register(&blkio_policy_cfq);
3919 static void __exit cfq_exit(void)
3921 DECLARE_COMPLETION_ONSTACK(all_gone);
3922 blkio_policy_unregister(&blkio_policy_cfq);
3923 elv_unregister(&iosched_cfq);
3924 ioc_gone = &all_gone;
3925 /* ioc_gone's update must be visible before reading ioc_count */
3929 * this also protects us from entering cfq_slab_kill() with
3930 * pending RCU callbacks
3932 if (elv_ioc_count_read(cfq_ioc_count))
3933 wait_for_completion(&all_gone);
3937 module_init(cfq_init);
3938 module_exit(cfq_exit);
3940 MODULE_AUTHOR("Jens Axboe");
3941 MODULE_LICENSE("GPL");
3942 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");