2 * CFQ, or complete fairness queueing, disk scheduler.
4 * Based on ideas from a previously unfinished io
5 * scheduler (round robin per-process disk scheduling) and Andrea Arcangeli.
7 * Copyright (C) 2003 Jens Axboe <axboe@kernel.dk>
9 #include <linux/module.h>
10 #include <linux/blkdev.h>
11 #include <linux/elevator.h>
12 #include <linux/jiffies.h>
13 #include <linux/rbtree.h>
14 #include <linux/ioprio.h>
15 #include <linux/blktrace_api.h>
16 #include "blk-cgroup.h"
21 /* max queue in one round of service */
22 static const int cfq_quantum = 4;
23 static const int cfq_fifo_expire[2] = { HZ / 4, HZ / 8 };
24 /* maximum backwards seek, in KiB */
25 static const int cfq_back_max = 16 * 1024;
26 /* penalty of a backwards seek */
27 static const int cfq_back_penalty = 2;
28 static const int cfq_slice_sync = HZ / 10;
29 static int cfq_slice_async = HZ / 25;
30 static const int cfq_slice_async_rq = 2;
31 static int cfq_slice_idle = HZ / 125;
32 static const int cfq_target_latency = HZ * 3/10; /* 300 ms */
33 static const int cfq_hist_divisor = 4;
36 * offset from end of service tree
38 #define CFQ_IDLE_DELAY (HZ / 5)
41 * below this threshold, we consider thinktime immediate
43 #define CFQ_MIN_TT (2)
46 * Allow merged cfqqs to perform this amount of seeky I/O before
47 * deciding to break the queues up again.
49 #define CFQQ_COOP_TOUT (HZ)
51 #define CFQ_SLICE_SCALE (5)
52 #define CFQ_HW_QUEUE_MIN (5)
53 #define CFQ_SERVICE_SHIFT 12
56 ((struct cfq_io_context *) (rq)->elevator_private)
57 #define RQ_CFQQ(rq) (struct cfq_queue *) ((rq)->elevator_private2)
59 static struct kmem_cache *cfq_pool;
60 static struct kmem_cache *cfq_ioc_pool;
62 static DEFINE_PER_CPU(unsigned long, cfq_ioc_count);
63 static struct completion *ioc_gone;
64 static DEFINE_SPINLOCK(ioc_gone_lock);
66 #define CFQ_PRIO_LISTS IOPRIO_BE_NR
67 #define cfq_class_idle(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
68 #define cfq_class_rt(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
70 #define sample_valid(samples) ((samples) > 80)
71 #define rb_entry_cfqg(node) rb_entry((node), struct cfq_group, rb_node)
74 * Most of our rbtree usage is for sorting with min extraction, so
75 * if we cache the leftmost node we don't have to walk down the tree
76 * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should
77 * move this into the elevator for the rq sorting as well.
84 struct rb_node *active;
85 unsigned total_weight;
87 #define CFQ_RB_ROOT (struct cfq_rb_root) { RB_ROOT, NULL, 0, 0, }
90 * Per process-grouping structure
95 /* various state flags, see below */
98 struct cfq_data *cfqd;
99 /* service_tree member */
100 struct rb_node rb_node;
101 /* service_tree key */
102 unsigned long rb_key;
103 /* prio tree member */
104 struct rb_node p_node;
105 /* prio tree root we belong to, if any */
106 struct rb_root *p_root;
107 /* sorted list of pending requests */
108 struct rb_root sort_list;
109 /* if fifo isn't expired, next request to serve */
110 struct request *next_rq;
111 /* requests queued in sort_list */
113 /* currently allocated requests */
115 /* fifo list of requests in sort_list */
116 struct list_head fifo;
118 /* time when queue got scheduled in to dispatch first request. */
119 unsigned long dispatch_start;
120 unsigned int allocated_slice;
121 /* time when first request from queue completed and slice started. */
122 unsigned long slice_start;
123 unsigned long slice_end;
125 unsigned int slice_dispatch;
127 /* pending metadata requests */
129 /* number of requests that are on the dispatch list or inside driver */
132 /* io prio of this group */
133 unsigned short ioprio, org_ioprio;
134 unsigned short ioprio_class, org_ioprio_class;
136 unsigned int seek_samples;
139 sector_t last_request_pos;
140 unsigned long seeky_start;
144 struct cfq_rb_root *service_tree;
145 struct cfq_queue *new_cfqq;
146 struct cfq_group *cfqg;
147 struct cfq_group *orig_cfqg;
148 /* Sectors dispatched in current dispatch round */
149 unsigned long nr_sectors;
153 * First index in the service_trees.
154 * IDLE is handled separately, so it has negative index
163 * Second index in the service_trees.
167 SYNC_NOIDLE_WORKLOAD = 1,
171 /* This is per cgroup per device grouping structure */
173 /* group service_tree member */
174 struct rb_node rb_node;
176 /* group service_tree key */
181 /* number of cfqq currently on this group */
184 /* Per group busy queus average. Useful for workload slice calc. */
185 unsigned int busy_queues_avg[2];
187 * rr lists of queues with requests, onle rr for each priority class.
188 * Counts are embedded in the cfq_rb_root
190 struct cfq_rb_root service_trees[2][3];
191 struct cfq_rb_root service_tree_idle;
193 unsigned long saved_workload_slice;
194 enum wl_type_t saved_workload;
195 enum wl_prio_t saved_serving_prio;
196 struct blkio_group blkg;
197 #ifdef CONFIG_CFQ_GROUP_IOSCHED
198 struct hlist_node cfqd_node;
204 * Per block device queue structure
207 struct request_queue *queue;
208 /* Root service tree for cfq_groups */
209 struct cfq_rb_root grp_service_tree;
210 struct cfq_group root_group;
213 * The priority currently being served
215 enum wl_prio_t serving_prio;
216 enum wl_type_t serving_type;
217 unsigned long workload_expires;
218 struct cfq_group *serving_group;
219 bool noidle_tree_requires_idle;
222 * Each priority tree is sorted by next_request position. These
223 * trees are used when determining if two or more queues are
224 * interleaving requests (see cfq_close_cooperator).
226 struct rb_root prio_trees[CFQ_PRIO_LISTS];
228 unsigned int busy_queues;
234 * queue-depth detection
240 * -1 => indeterminate, (cfq will behave as if NCQ is present, to allow better detection)
241 * 1 => NCQ is present (hw_tag_est_depth is the estimated max depth)
244 int hw_tag_est_depth;
245 unsigned int hw_tag_samples;
248 * idle window management
250 struct timer_list idle_slice_timer;
251 struct work_struct unplug_work;
253 struct cfq_queue *active_queue;
254 struct cfq_io_context *active_cic;
257 * async queue for each priority case
259 struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
260 struct cfq_queue *async_idle_cfqq;
262 sector_t last_position;
265 * tunables, see top of file
267 unsigned int cfq_quantum;
268 unsigned int cfq_fifo_expire[2];
269 unsigned int cfq_back_penalty;
270 unsigned int cfq_back_max;
271 unsigned int cfq_slice[2];
272 unsigned int cfq_slice_async_rq;
273 unsigned int cfq_slice_idle;
274 unsigned int cfq_latency;
275 unsigned int cfq_group_isolation;
277 struct list_head cic_list;
280 * Fallback dummy cfqq for extreme OOM conditions
282 struct cfq_queue oom_cfqq;
284 unsigned long last_delayed_sync;
286 /* List of cfq groups being managed on this device*/
287 struct hlist_head cfqg_list;
291 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd);
293 static struct cfq_rb_root *service_tree_for(struct cfq_group *cfqg,
300 if (prio == IDLE_WORKLOAD)
301 return &cfqg->service_tree_idle;
303 return &cfqg->service_trees[prio][type];
306 enum cfqq_state_flags {
307 CFQ_CFQQ_FLAG_on_rr = 0, /* on round-robin busy list */
308 CFQ_CFQQ_FLAG_wait_request, /* waiting for a request */
309 CFQ_CFQQ_FLAG_must_dispatch, /* must be allowed a dispatch */
310 CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
311 CFQ_CFQQ_FLAG_fifo_expire, /* FIFO checked in this slice */
312 CFQ_CFQQ_FLAG_idle_window, /* slice idling enabled */
313 CFQ_CFQQ_FLAG_prio_changed, /* task priority has changed */
314 CFQ_CFQQ_FLAG_slice_new, /* no requests dispatched in slice */
315 CFQ_CFQQ_FLAG_sync, /* synchronous queue */
316 CFQ_CFQQ_FLAG_coop, /* cfqq is shared */
317 CFQ_CFQQ_FLAG_deep, /* sync cfqq experienced large depth */
318 CFQ_CFQQ_FLAG_wait_busy, /* Waiting for next request */
321 #define CFQ_CFQQ_FNS(name) \
322 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \
324 (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name); \
326 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \
328 (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \
330 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \
332 return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \
336 CFQ_CFQQ_FNS(wait_request);
337 CFQ_CFQQ_FNS(must_dispatch);
338 CFQ_CFQQ_FNS(must_alloc_slice);
339 CFQ_CFQQ_FNS(fifo_expire);
340 CFQ_CFQQ_FNS(idle_window);
341 CFQ_CFQQ_FNS(prio_changed);
342 CFQ_CFQQ_FNS(slice_new);
346 CFQ_CFQQ_FNS(wait_busy);
349 #ifdef CONFIG_DEBUG_CFQ_IOSCHED
350 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
351 blk_add_trace_msg((cfqd)->queue, "cfq%d%c %s " fmt, (cfqq)->pid, \
352 cfq_cfqq_sync((cfqq)) ? 'S' : 'A', \
353 blkg_path(&(cfqq)->cfqg->blkg), ##args);
355 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...) \
356 blk_add_trace_msg((cfqd)->queue, "%s " fmt, \
357 blkg_path(&(cfqg)->blkg), ##args); \
360 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
361 blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args)
362 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...) do {} while (0);
364 #define cfq_log(cfqd, fmt, args...) \
365 blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
367 /* Traverses through cfq group service trees */
368 #define for_each_cfqg_st(cfqg, i, j, st) \
369 for (i = 0; i <= IDLE_WORKLOAD; i++) \
370 for (j = 0, st = i < IDLE_WORKLOAD ? &cfqg->service_trees[i][j]\
371 : &cfqg->service_tree_idle; \
372 (i < IDLE_WORKLOAD && j <= SYNC_WORKLOAD) || \
373 (i == IDLE_WORKLOAD && j == 0); \
374 j++, st = i < IDLE_WORKLOAD ? \
375 &cfqg->service_trees[i][j]: NULL) \
378 static inline enum wl_prio_t cfqq_prio(struct cfq_queue *cfqq)
380 if (cfq_class_idle(cfqq))
381 return IDLE_WORKLOAD;
382 if (cfq_class_rt(cfqq))
388 static enum wl_type_t cfqq_type(struct cfq_queue *cfqq)
390 if (!cfq_cfqq_sync(cfqq))
391 return ASYNC_WORKLOAD;
392 if (!cfq_cfqq_idle_window(cfqq))
393 return SYNC_NOIDLE_WORKLOAD;
394 return SYNC_WORKLOAD;
397 static inline int cfq_group_busy_queues_wl(enum wl_prio_t wl,
398 struct cfq_data *cfqd,
399 struct cfq_group *cfqg)
401 if (wl == IDLE_WORKLOAD)
402 return cfqg->service_tree_idle.count;
404 return cfqg->service_trees[wl][ASYNC_WORKLOAD].count
405 + cfqg->service_trees[wl][SYNC_NOIDLE_WORKLOAD].count
406 + cfqg->service_trees[wl][SYNC_WORKLOAD].count;
409 static inline int cfqg_busy_async_queues(struct cfq_data *cfqd,
410 struct cfq_group *cfqg)
412 return cfqg->service_trees[RT_WORKLOAD][ASYNC_WORKLOAD].count
413 + cfqg->service_trees[BE_WORKLOAD][ASYNC_WORKLOAD].count;
416 static void cfq_dispatch_insert(struct request_queue *, struct request *);
417 static struct cfq_queue *cfq_get_queue(struct cfq_data *, bool,
418 struct io_context *, gfp_t);
419 static struct cfq_io_context *cfq_cic_lookup(struct cfq_data *,
420 struct io_context *);
422 static inline int rq_in_driver(struct cfq_data *cfqd)
424 return cfqd->rq_in_driver[0] + cfqd->rq_in_driver[1];
427 static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_context *cic,
430 return cic->cfqq[is_sync];
433 static inline void cic_set_cfqq(struct cfq_io_context *cic,
434 struct cfq_queue *cfqq, bool is_sync)
436 cic->cfqq[is_sync] = cfqq;
440 * We regard a request as SYNC, if it's either a read or has the SYNC bit
441 * set (in which case it could also be direct WRITE).
443 static inline bool cfq_bio_sync(struct bio *bio)
445 return bio_data_dir(bio) == READ || bio_rw_flagged(bio, BIO_RW_SYNCIO);
449 * scheduler run of queue, if there are requests pending and no one in the
450 * driver that will restart queueing
452 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
454 if (cfqd->busy_queues) {
455 cfq_log(cfqd, "schedule dispatch");
456 kblockd_schedule_work(cfqd->queue, &cfqd->unplug_work);
460 static int cfq_queue_empty(struct request_queue *q)
462 struct cfq_data *cfqd = q->elevator->elevator_data;
464 return !cfqd->rq_queued;
468 * Scale schedule slice based on io priority. Use the sync time slice only
469 * if a queue is marked sync and has sync io queued. A sync queue with async
470 * io only, should not get full sync slice length.
472 static inline int cfq_prio_slice(struct cfq_data *cfqd, bool sync,
475 const int base_slice = cfqd->cfq_slice[sync];
477 WARN_ON(prio >= IOPRIO_BE_NR);
479 return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
483 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
485 return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
488 static inline u64 cfq_scale_slice(unsigned long delta, struct cfq_group *cfqg)
490 u64 d = delta << CFQ_SERVICE_SHIFT;
492 d = d * BLKIO_WEIGHT_DEFAULT;
493 do_div(d, cfqg->weight);
497 static inline u64 max_vdisktime(u64 min_vdisktime, u64 vdisktime)
499 s64 delta = (s64)(vdisktime - min_vdisktime);
501 min_vdisktime = vdisktime;
503 return min_vdisktime;
506 static inline u64 min_vdisktime(u64 min_vdisktime, u64 vdisktime)
508 s64 delta = (s64)(vdisktime - min_vdisktime);
510 min_vdisktime = vdisktime;
512 return min_vdisktime;
515 static void update_min_vdisktime(struct cfq_rb_root *st)
517 u64 vdisktime = st->min_vdisktime;
518 struct cfq_group *cfqg;
521 cfqg = rb_entry_cfqg(st->active);
522 vdisktime = cfqg->vdisktime;
526 cfqg = rb_entry_cfqg(st->left);
527 vdisktime = min_vdisktime(vdisktime, cfqg->vdisktime);
530 st->min_vdisktime = max_vdisktime(st->min_vdisktime, vdisktime);
534 * get averaged number of queues of RT/BE priority.
535 * average is updated, with a formula that gives more weight to higher numbers,
536 * to quickly follows sudden increases and decrease slowly
539 static inline unsigned cfq_group_get_avg_queues(struct cfq_data *cfqd,
540 struct cfq_group *cfqg, bool rt)
542 unsigned min_q, max_q;
543 unsigned mult = cfq_hist_divisor - 1;
544 unsigned round = cfq_hist_divisor / 2;
545 unsigned busy = cfq_group_busy_queues_wl(rt, cfqd, cfqg);
547 min_q = min(cfqg->busy_queues_avg[rt], busy);
548 max_q = max(cfqg->busy_queues_avg[rt], busy);
549 cfqg->busy_queues_avg[rt] = (mult * max_q + min_q + round) /
551 return cfqg->busy_queues_avg[rt];
554 static inline unsigned
555 cfq_group_slice(struct cfq_data *cfqd, struct cfq_group *cfqg)
557 struct cfq_rb_root *st = &cfqd->grp_service_tree;
559 return cfq_target_latency * cfqg->weight / st->total_weight;
563 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
565 unsigned slice = cfq_prio_to_slice(cfqd, cfqq);
566 if (cfqd->cfq_latency) {
568 * interested queues (we consider only the ones with the same
569 * priority class in the cfq group)
571 unsigned iq = cfq_group_get_avg_queues(cfqd, cfqq->cfqg,
573 unsigned sync_slice = cfqd->cfq_slice[1];
574 unsigned expect_latency = sync_slice * iq;
575 unsigned group_slice = cfq_group_slice(cfqd, cfqq->cfqg);
577 if (expect_latency > group_slice) {
578 unsigned base_low_slice = 2 * cfqd->cfq_slice_idle;
579 /* scale low_slice according to IO priority
580 * and sync vs async */
582 min(slice, base_low_slice * slice / sync_slice);
583 /* the adapted slice value is scaled to fit all iqs
584 * into the target latency */
585 slice = max(slice * group_slice / expect_latency,
589 cfqq->slice_start = jiffies;
590 cfqq->slice_end = jiffies + slice;
591 cfqq->allocated_slice = slice;
592 cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
596 * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
597 * isn't valid until the first request from the dispatch is activated
598 * and the slice time set.
600 static inline bool cfq_slice_used(struct cfq_queue *cfqq)
602 if (cfq_cfqq_slice_new(cfqq))
604 if (time_before(jiffies, cfqq->slice_end))
611 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
612 * We choose the request that is closest to the head right now. Distance
613 * behind the head is penalized and only allowed to a certain extent.
615 static struct request *
616 cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2, sector_t last)
618 sector_t s1, s2, d1 = 0, d2 = 0;
619 unsigned long back_max;
620 #define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */
621 #define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */
622 unsigned wrap = 0; /* bit mask: requests behind the disk head? */
624 if (rq1 == NULL || rq1 == rq2)
629 if (rq_is_sync(rq1) && !rq_is_sync(rq2))
631 else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
633 if (rq_is_meta(rq1) && !rq_is_meta(rq2))
635 else if (rq_is_meta(rq2) && !rq_is_meta(rq1))
638 s1 = blk_rq_pos(rq1);
639 s2 = blk_rq_pos(rq2);
642 * by definition, 1KiB is 2 sectors
644 back_max = cfqd->cfq_back_max * 2;
647 * Strict one way elevator _except_ in the case where we allow
648 * short backward seeks which are biased as twice the cost of a
649 * similar forward seek.
653 else if (s1 + back_max >= last)
654 d1 = (last - s1) * cfqd->cfq_back_penalty;
656 wrap |= CFQ_RQ1_WRAP;
660 else if (s2 + back_max >= last)
661 d2 = (last - s2) * cfqd->cfq_back_penalty;
663 wrap |= CFQ_RQ2_WRAP;
665 /* Found required data */
668 * By doing switch() on the bit mask "wrap" we avoid having to
669 * check two variables for all permutations: --> faster!
672 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
688 case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
691 * Since both rqs are wrapped,
692 * start with the one that's further behind head
693 * (--> only *one* back seek required),
694 * since back seek takes more time than forward.
704 * The below is leftmost cache rbtree addon
706 static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
708 /* Service tree is empty */
713 root->left = rb_first(&root->rb);
716 return rb_entry(root->left, struct cfq_queue, rb_node);
721 static struct cfq_group *cfq_rb_first_group(struct cfq_rb_root *root)
724 root->left = rb_first(&root->rb);
727 return rb_entry_cfqg(root->left);
732 static void rb_erase_init(struct rb_node *n, struct rb_root *root)
738 static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
742 rb_erase_init(n, &root->rb);
747 * would be nice to take fifo expire time into account as well
749 static struct request *
750 cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
751 struct request *last)
753 struct rb_node *rbnext = rb_next(&last->rb_node);
754 struct rb_node *rbprev = rb_prev(&last->rb_node);
755 struct request *next = NULL, *prev = NULL;
757 BUG_ON(RB_EMPTY_NODE(&last->rb_node));
760 prev = rb_entry_rq(rbprev);
763 next = rb_entry_rq(rbnext);
765 rbnext = rb_first(&cfqq->sort_list);
766 if (rbnext && rbnext != &last->rb_node)
767 next = rb_entry_rq(rbnext);
770 return cfq_choose_req(cfqd, next, prev, blk_rq_pos(last));
773 static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
774 struct cfq_queue *cfqq)
777 * just an approximation, should be ok.
779 return (cfqq->cfqg->nr_cfqq - 1) * (cfq_prio_slice(cfqd, 1, 0) -
780 cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
784 cfqg_key(struct cfq_rb_root *st, struct cfq_group *cfqg)
786 return cfqg->vdisktime - st->min_vdisktime;
790 __cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
792 struct rb_node **node = &st->rb.rb_node;
793 struct rb_node *parent = NULL;
794 struct cfq_group *__cfqg;
795 s64 key = cfqg_key(st, cfqg);
798 while (*node != NULL) {
800 __cfqg = rb_entry_cfqg(parent);
802 if (key < cfqg_key(st, __cfqg))
803 node = &parent->rb_left;
805 node = &parent->rb_right;
811 st->left = &cfqg->rb_node;
813 rb_link_node(&cfqg->rb_node, parent, node);
814 rb_insert_color(&cfqg->rb_node, &st->rb);
818 cfq_group_service_tree_add(struct cfq_data *cfqd, struct cfq_group *cfqg)
820 struct cfq_rb_root *st = &cfqd->grp_service_tree;
821 struct cfq_group *__cfqg;
829 * Currently put the group at the end. Later implement something
830 * so that groups get lesser vtime based on their weights, so that
831 * if group does not loose all if it was not continously backlogged.
833 n = rb_last(&st->rb);
835 __cfqg = rb_entry_cfqg(n);
836 cfqg->vdisktime = __cfqg->vdisktime + CFQ_IDLE_DELAY;
838 cfqg->vdisktime = st->min_vdisktime;
840 __cfq_group_service_tree_add(st, cfqg);
842 st->total_weight += cfqg->weight;
846 cfq_group_service_tree_del(struct cfq_data *cfqd, struct cfq_group *cfqg)
848 struct cfq_rb_root *st = &cfqd->grp_service_tree;
850 if (st->active == &cfqg->rb_node)
853 BUG_ON(cfqg->nr_cfqq < 1);
856 /* If there are other cfq queues under this group, don't delete it */
860 cfq_log_cfqg(cfqd, cfqg, "del_from_rr group");
862 st->total_weight -= cfqg->weight;
863 if (!RB_EMPTY_NODE(&cfqg->rb_node))
864 cfq_rb_erase(&cfqg->rb_node, st);
865 cfqg->saved_workload_slice = 0;
866 blkiocg_update_blkio_group_dequeue_stats(&cfqg->blkg, 1);
869 static inline unsigned int cfq_cfqq_slice_usage(struct cfq_queue *cfqq)
871 unsigned int slice_used;
874 * Queue got expired before even a single request completed or
875 * got expired immediately after first request completion.
877 if (!cfqq->slice_start || cfqq->slice_start == jiffies) {
879 * Also charge the seek time incurred to the group, otherwise
880 * if there are mutiple queues in the group, each can dispatch
881 * a single request on seeky media and cause lots of seek time
882 * and group will never know it.
884 slice_used = max_t(unsigned, (jiffies - cfqq->dispatch_start),
887 slice_used = jiffies - cfqq->slice_start;
888 if (slice_used > cfqq->allocated_slice)
889 slice_used = cfqq->allocated_slice;
892 cfq_log_cfqq(cfqq->cfqd, cfqq, "sl_used=%u sect=%lu", slice_used,
897 static void cfq_group_served(struct cfq_data *cfqd, struct cfq_group *cfqg,
898 struct cfq_queue *cfqq)
900 struct cfq_rb_root *st = &cfqd->grp_service_tree;
901 unsigned int used_sl, charge_sl;
902 int nr_sync = cfqg->nr_cfqq - cfqg_busy_async_queues(cfqd, cfqg)
903 - cfqg->service_tree_idle.count;
906 used_sl = charge_sl = cfq_cfqq_slice_usage(cfqq);
908 if (!cfq_cfqq_sync(cfqq) && !nr_sync)
909 charge_sl = cfqq->allocated_slice;
911 /* Can't update vdisktime while group is on service tree */
912 cfq_rb_erase(&cfqg->rb_node, st);
913 cfqg->vdisktime += cfq_scale_slice(charge_sl, cfqg);
914 __cfq_group_service_tree_add(st, cfqg);
916 /* This group is being expired. Save the context */
917 if (time_after(cfqd->workload_expires, jiffies)) {
918 cfqg->saved_workload_slice = cfqd->workload_expires
920 cfqg->saved_workload = cfqd->serving_type;
921 cfqg->saved_serving_prio = cfqd->serving_prio;
923 cfqg->saved_workload_slice = 0;
925 cfq_log_cfqg(cfqd, cfqg, "served: vt=%llu min_vt=%llu", cfqg->vdisktime,
927 blkiocg_update_blkio_group_stats(&cfqg->blkg, used_sl,
931 #ifdef CONFIG_CFQ_GROUP_IOSCHED
932 static inline struct cfq_group *cfqg_of_blkg(struct blkio_group *blkg)
935 return container_of(blkg, struct cfq_group, blkg);
940 cfq_update_blkio_group_weight(struct blkio_group *blkg, unsigned int weight)
942 cfqg_of_blkg(blkg)->weight = weight;
945 static struct cfq_group *
946 cfq_find_alloc_cfqg(struct cfq_data *cfqd, struct cgroup *cgroup, int create)
948 struct blkio_cgroup *blkcg = cgroup_to_blkio_cgroup(cgroup);
949 struct cfq_group *cfqg = NULL;
952 struct cfq_rb_root *st;
953 struct backing_dev_info *bdi = &cfqd->queue->backing_dev_info;
954 unsigned int major, minor;
956 /* Do we need to take this reference */
957 if (!blkiocg_css_tryget(blkcg))
960 cfqg = cfqg_of_blkg(blkiocg_lookup_group(blkcg, key));
964 cfqg = kzalloc_node(sizeof(*cfqg), GFP_ATOMIC, cfqd->queue->node);
968 cfqg->weight = blkcg->weight;
969 for_each_cfqg_st(cfqg, i, j, st)
971 RB_CLEAR_NODE(&cfqg->rb_node);
974 * Take the initial reference that will be released on destroy
975 * This can be thought of a joint reference by cgroup and
976 * elevator which will be dropped by either elevator exit
977 * or cgroup deletion path depending on who is exiting first.
979 atomic_set(&cfqg->ref, 1);
981 /* Add group onto cgroup list */
982 sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
983 blkiocg_add_blkio_group(blkcg, &cfqg->blkg, (void *)cfqd,
984 MKDEV(major, minor));
986 /* Add group on cfqd list */
987 hlist_add_head(&cfqg->cfqd_node, &cfqd->cfqg_list);
990 blkiocg_css_put(blkcg);
995 * Search for the cfq group current task belongs to. If create = 1, then also
996 * create the cfq group if it does not exist. request_queue lock must be held.
998 static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd, int create)
1000 struct cgroup *cgroup;
1001 struct cfq_group *cfqg = NULL;
1004 cgroup = task_cgroup(current, blkio_subsys_id);
1005 cfqg = cfq_find_alloc_cfqg(cfqd, cgroup, create);
1006 if (!cfqg && create)
1007 cfqg = &cfqd->root_group;
1012 static void cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg)
1014 /* Currently, all async queues are mapped to root group */
1015 if (!cfq_cfqq_sync(cfqq))
1016 cfqg = &cfqq->cfqd->root_group;
1019 /* cfqq reference on cfqg */
1020 atomic_inc(&cfqq->cfqg->ref);
1023 static void cfq_put_cfqg(struct cfq_group *cfqg)
1025 struct cfq_rb_root *st;
1028 BUG_ON(atomic_read(&cfqg->ref) <= 0);
1029 if (!atomic_dec_and_test(&cfqg->ref))
1031 for_each_cfqg_st(cfqg, i, j, st)
1032 BUG_ON(!RB_EMPTY_ROOT(&st->rb) || st->active != NULL);
1036 static void cfq_destroy_cfqg(struct cfq_data *cfqd, struct cfq_group *cfqg)
1038 /* Something wrong if we are trying to remove same group twice */
1039 BUG_ON(hlist_unhashed(&cfqg->cfqd_node));
1041 hlist_del_init(&cfqg->cfqd_node);
1044 * Put the reference taken at the time of creation so that when all
1045 * queues are gone, group can be destroyed.
1050 static void cfq_release_cfq_groups(struct cfq_data *cfqd)
1052 struct hlist_node *pos, *n;
1053 struct cfq_group *cfqg;
1055 hlist_for_each_entry_safe(cfqg, pos, n, &cfqd->cfqg_list, cfqd_node) {
1057 * If cgroup removal path got to blk_group first and removed
1058 * it from cgroup list, then it will take care of destroying
1061 if (!blkiocg_del_blkio_group(&cfqg->blkg))
1062 cfq_destroy_cfqg(cfqd, cfqg);
1067 * Blk cgroup controller notification saying that blkio_group object is being
1068 * delinked as associated cgroup object is going away. That also means that
1069 * no new IO will come in this group. So get rid of this group as soon as
1070 * any pending IO in the group is finished.
1072 * This function is called under rcu_read_lock(). key is the rcu protected
1073 * pointer. That means "key" is a valid cfq_data pointer as long as we are rcu
1076 * "key" was fetched from blkio_group under blkio_cgroup->lock. That means
1077 * it should not be NULL as even if elevator was exiting, cgroup deltion
1078 * path got to it first.
1080 void cfq_unlink_blkio_group(void *key, struct blkio_group *blkg)
1082 unsigned long flags;
1083 struct cfq_data *cfqd = key;
1085 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1086 cfq_destroy_cfqg(cfqd, cfqg_of_blkg(blkg));
1087 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1090 #else /* GROUP_IOSCHED */
1091 static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd, int create)
1093 return &cfqd->root_group;
1096 cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) {
1100 static void cfq_release_cfq_groups(struct cfq_data *cfqd) {}
1101 static inline void cfq_put_cfqg(struct cfq_group *cfqg) {}
1103 #endif /* GROUP_IOSCHED */
1106 * The cfqd->service_trees holds all pending cfq_queue's that have
1107 * requests waiting to be processed. It is sorted in the order that
1108 * we will service the queues.
1110 static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1113 struct rb_node **p, *parent;
1114 struct cfq_queue *__cfqq;
1115 unsigned long rb_key;
1116 struct cfq_rb_root *service_tree;
1119 int group_changed = 0;
1121 #ifdef CONFIG_CFQ_GROUP_IOSCHED
1122 if (!cfqd->cfq_group_isolation
1123 && cfqq_type(cfqq) == SYNC_NOIDLE_WORKLOAD
1124 && cfqq->cfqg && cfqq->cfqg != &cfqd->root_group) {
1125 /* Move this cfq to root group */
1126 cfq_log_cfqq(cfqd, cfqq, "moving to root group");
1127 if (!RB_EMPTY_NODE(&cfqq->rb_node))
1128 cfq_group_service_tree_del(cfqd, cfqq->cfqg);
1129 cfqq->orig_cfqg = cfqq->cfqg;
1130 cfqq->cfqg = &cfqd->root_group;
1131 atomic_inc(&cfqd->root_group.ref);
1133 } else if (!cfqd->cfq_group_isolation
1134 && cfqq_type(cfqq) == SYNC_WORKLOAD && cfqq->orig_cfqg) {
1135 /* cfqq is sequential now needs to go to its original group */
1136 BUG_ON(cfqq->cfqg != &cfqd->root_group);
1137 if (!RB_EMPTY_NODE(&cfqq->rb_node))
1138 cfq_group_service_tree_del(cfqd, cfqq->cfqg);
1139 cfq_put_cfqg(cfqq->cfqg);
1140 cfqq->cfqg = cfqq->orig_cfqg;
1141 cfqq->orig_cfqg = NULL;
1143 cfq_log_cfqq(cfqd, cfqq, "moved to origin group");
1147 service_tree = service_tree_for(cfqq->cfqg, cfqq_prio(cfqq),
1149 if (cfq_class_idle(cfqq)) {
1150 rb_key = CFQ_IDLE_DELAY;
1151 parent = rb_last(&service_tree->rb);
1152 if (parent && parent != &cfqq->rb_node) {
1153 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1154 rb_key += __cfqq->rb_key;
1157 } else if (!add_front) {
1159 * Get our rb key offset. Subtract any residual slice
1160 * value carried from last service. A negative resid
1161 * count indicates slice overrun, and this should position
1162 * the next service time further away in the tree.
1164 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
1165 rb_key -= cfqq->slice_resid;
1166 cfqq->slice_resid = 0;
1169 __cfqq = cfq_rb_first(service_tree);
1170 rb_key += __cfqq ? __cfqq->rb_key : jiffies;
1173 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1176 * same position, nothing more to do
1178 if (rb_key == cfqq->rb_key &&
1179 cfqq->service_tree == service_tree)
1182 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1183 cfqq->service_tree = NULL;
1188 cfqq->service_tree = service_tree;
1189 p = &service_tree->rb.rb_node;
1194 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1197 * sort by key, that represents service time.
1199 if (time_before(rb_key, __cfqq->rb_key))
1202 n = &(*p)->rb_right;
1210 service_tree->left = &cfqq->rb_node;
1212 cfqq->rb_key = rb_key;
1213 rb_link_node(&cfqq->rb_node, parent, p);
1214 rb_insert_color(&cfqq->rb_node, &service_tree->rb);
1215 service_tree->count++;
1216 if ((add_front || !new_cfqq) && !group_changed)
1218 cfq_group_service_tree_add(cfqd, cfqq->cfqg);
1221 static struct cfq_queue *
1222 cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
1223 sector_t sector, struct rb_node **ret_parent,
1224 struct rb_node ***rb_link)
1226 struct rb_node **p, *parent;
1227 struct cfq_queue *cfqq = NULL;
1235 cfqq = rb_entry(parent, struct cfq_queue, p_node);
1238 * Sort strictly based on sector. Smallest to the left,
1239 * largest to the right.
1241 if (sector > blk_rq_pos(cfqq->next_rq))
1242 n = &(*p)->rb_right;
1243 else if (sector < blk_rq_pos(cfqq->next_rq))
1251 *ret_parent = parent;
1257 static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1259 struct rb_node **p, *parent;
1260 struct cfq_queue *__cfqq;
1263 rb_erase(&cfqq->p_node, cfqq->p_root);
1264 cfqq->p_root = NULL;
1267 if (cfq_class_idle(cfqq))
1272 cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
1273 __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
1274 blk_rq_pos(cfqq->next_rq), &parent, &p);
1276 rb_link_node(&cfqq->p_node, parent, p);
1277 rb_insert_color(&cfqq->p_node, cfqq->p_root);
1279 cfqq->p_root = NULL;
1283 * Update cfqq's position in the service tree.
1285 static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1288 * Resorting requires the cfqq to be on the RR list already.
1290 if (cfq_cfqq_on_rr(cfqq)) {
1291 cfq_service_tree_add(cfqd, cfqq, 0);
1292 cfq_prio_tree_add(cfqd, cfqq);
1297 * add to busy list of queues for service, trying to be fair in ordering
1298 * the pending list according to last request service
1300 static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1302 cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
1303 BUG_ON(cfq_cfqq_on_rr(cfqq));
1304 cfq_mark_cfqq_on_rr(cfqq);
1305 cfqd->busy_queues++;
1307 cfq_resort_rr_list(cfqd, cfqq);
1311 * Called when the cfqq no longer has requests pending, remove it from
1314 static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1316 cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
1317 BUG_ON(!cfq_cfqq_on_rr(cfqq));
1318 cfq_clear_cfqq_on_rr(cfqq);
1320 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1321 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1322 cfqq->service_tree = NULL;
1325 rb_erase(&cfqq->p_node, cfqq->p_root);
1326 cfqq->p_root = NULL;
1329 cfq_group_service_tree_del(cfqd, cfqq->cfqg);
1330 BUG_ON(!cfqd->busy_queues);
1331 cfqd->busy_queues--;
1335 * rb tree support functions
1337 static void cfq_del_rq_rb(struct request *rq)
1339 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1340 const int sync = rq_is_sync(rq);
1342 BUG_ON(!cfqq->queued[sync]);
1343 cfqq->queued[sync]--;
1345 elv_rb_del(&cfqq->sort_list, rq);
1347 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) {
1349 * Queue will be deleted from service tree when we actually
1350 * expire it later. Right now just remove it from prio tree
1354 rb_erase(&cfqq->p_node, cfqq->p_root);
1355 cfqq->p_root = NULL;
1360 static void cfq_add_rq_rb(struct request *rq)
1362 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1363 struct cfq_data *cfqd = cfqq->cfqd;
1364 struct request *__alias, *prev;
1366 cfqq->queued[rq_is_sync(rq)]++;
1369 * looks a little odd, but the first insert might return an alias.
1370 * if that happens, put the alias on the dispatch list
1372 while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL)
1373 cfq_dispatch_insert(cfqd->queue, __alias);
1375 if (!cfq_cfqq_on_rr(cfqq))
1376 cfq_add_cfqq_rr(cfqd, cfqq);
1379 * check if this request is a better next-serve candidate
1381 prev = cfqq->next_rq;
1382 cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq, cfqd->last_position);
1385 * adjust priority tree position, if ->next_rq changes
1387 if (prev != cfqq->next_rq)
1388 cfq_prio_tree_add(cfqd, cfqq);
1390 BUG_ON(!cfqq->next_rq);
1393 static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
1395 elv_rb_del(&cfqq->sort_list, rq);
1396 cfqq->queued[rq_is_sync(rq)]--;
1400 static struct request *
1401 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
1403 struct task_struct *tsk = current;
1404 struct cfq_io_context *cic;
1405 struct cfq_queue *cfqq;
1407 cic = cfq_cic_lookup(cfqd, tsk->io_context);
1411 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
1413 sector_t sector = bio->bi_sector + bio_sectors(bio);
1415 return elv_rb_find(&cfqq->sort_list, sector);
1421 static void cfq_activate_request(struct request_queue *q, struct request *rq)
1423 struct cfq_data *cfqd = q->elevator->elevator_data;
1425 cfqd->rq_in_driver[rq_is_sync(rq)]++;
1426 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
1427 rq_in_driver(cfqd));
1429 cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
1432 static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
1434 struct cfq_data *cfqd = q->elevator->elevator_data;
1435 const int sync = rq_is_sync(rq);
1437 WARN_ON(!cfqd->rq_in_driver[sync]);
1438 cfqd->rq_in_driver[sync]--;
1439 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
1440 rq_in_driver(cfqd));
1443 static void cfq_remove_request(struct request *rq)
1445 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1447 if (cfqq->next_rq == rq)
1448 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
1450 list_del_init(&rq->queuelist);
1453 cfqq->cfqd->rq_queued--;
1454 if (rq_is_meta(rq)) {
1455 WARN_ON(!cfqq->meta_pending);
1456 cfqq->meta_pending--;
1460 static int cfq_merge(struct request_queue *q, struct request **req,
1463 struct cfq_data *cfqd = q->elevator->elevator_data;
1464 struct request *__rq;
1466 __rq = cfq_find_rq_fmerge(cfqd, bio);
1467 if (__rq && elv_rq_merge_ok(__rq, bio)) {
1469 return ELEVATOR_FRONT_MERGE;
1472 return ELEVATOR_NO_MERGE;
1475 static void cfq_merged_request(struct request_queue *q, struct request *req,
1478 if (type == ELEVATOR_FRONT_MERGE) {
1479 struct cfq_queue *cfqq = RQ_CFQQ(req);
1481 cfq_reposition_rq_rb(cfqq, req);
1486 cfq_merged_requests(struct request_queue *q, struct request *rq,
1487 struct request *next)
1489 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1491 * reposition in fifo if next is older than rq
1493 if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
1494 time_before(rq_fifo_time(next), rq_fifo_time(rq))) {
1495 list_move(&rq->queuelist, &next->queuelist);
1496 rq_set_fifo_time(rq, rq_fifo_time(next));
1499 if (cfqq->next_rq == next)
1501 cfq_remove_request(next);
1504 static int cfq_allow_merge(struct request_queue *q, struct request *rq,
1507 struct cfq_data *cfqd = q->elevator->elevator_data;
1508 struct cfq_io_context *cic;
1509 struct cfq_queue *cfqq;
1512 * Disallow merge of a sync bio into an async request.
1514 if (cfq_bio_sync(bio) && !rq_is_sync(rq))
1518 * Lookup the cfqq that this bio will be queued with. Allow
1519 * merge only if rq is queued there.
1521 cic = cfq_cic_lookup(cfqd, current->io_context);
1525 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
1526 return cfqq == RQ_CFQQ(rq);
1529 static void __cfq_set_active_queue(struct cfq_data *cfqd,
1530 struct cfq_queue *cfqq)
1533 cfq_log_cfqq(cfqd, cfqq, "set_active");
1534 cfqq->slice_start = 0;
1535 cfqq->dispatch_start = jiffies;
1536 cfqq->allocated_slice = 0;
1537 cfqq->slice_end = 0;
1538 cfqq->slice_dispatch = 0;
1539 cfqq->nr_sectors = 0;
1541 cfq_clear_cfqq_wait_request(cfqq);
1542 cfq_clear_cfqq_must_dispatch(cfqq);
1543 cfq_clear_cfqq_must_alloc_slice(cfqq);
1544 cfq_clear_cfqq_fifo_expire(cfqq);
1545 cfq_mark_cfqq_slice_new(cfqq);
1547 del_timer(&cfqd->idle_slice_timer);
1550 cfqd->active_queue = cfqq;
1554 * current cfqq expired its slice (or was too idle), select new one
1557 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1560 cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
1562 if (cfq_cfqq_wait_request(cfqq))
1563 del_timer(&cfqd->idle_slice_timer);
1565 cfq_clear_cfqq_wait_request(cfqq);
1566 cfq_clear_cfqq_wait_busy(cfqq);
1569 * store what was left of this slice, if the queue idled/timed out
1571 if (timed_out && !cfq_cfqq_slice_new(cfqq)) {
1572 cfqq->slice_resid = cfqq->slice_end - jiffies;
1573 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
1576 cfq_group_served(cfqd, cfqq->cfqg, cfqq);
1578 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
1579 cfq_del_cfqq_rr(cfqd, cfqq);
1581 cfq_resort_rr_list(cfqd, cfqq);
1583 if (cfqq == cfqd->active_queue)
1584 cfqd->active_queue = NULL;
1586 if (&cfqq->cfqg->rb_node == cfqd->grp_service_tree.active)
1587 cfqd->grp_service_tree.active = NULL;
1589 if (cfqd->active_cic) {
1590 put_io_context(cfqd->active_cic->ioc);
1591 cfqd->active_cic = NULL;
1595 static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out)
1597 struct cfq_queue *cfqq = cfqd->active_queue;
1600 __cfq_slice_expired(cfqd, cfqq, timed_out);
1604 * Get next queue for service. Unless we have a queue preemption,
1605 * we'll simply select the first cfqq in the service tree.
1607 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
1609 struct cfq_rb_root *service_tree =
1610 service_tree_for(cfqd->serving_group, cfqd->serving_prio,
1611 cfqd->serving_type);
1613 if (!cfqd->rq_queued)
1616 /* There is nothing to dispatch */
1619 if (RB_EMPTY_ROOT(&service_tree->rb))
1621 return cfq_rb_first(service_tree);
1624 static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd)
1626 struct cfq_group *cfqg;
1627 struct cfq_queue *cfqq;
1629 struct cfq_rb_root *st;
1631 if (!cfqd->rq_queued)
1634 cfqg = cfq_get_next_cfqg(cfqd);
1638 for_each_cfqg_st(cfqg, i, j, st)
1639 if ((cfqq = cfq_rb_first(st)) != NULL)
1645 * Get and set a new active queue for service.
1647 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
1648 struct cfq_queue *cfqq)
1651 cfqq = cfq_get_next_queue(cfqd);
1653 __cfq_set_active_queue(cfqd, cfqq);
1657 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
1660 if (blk_rq_pos(rq) >= cfqd->last_position)
1661 return blk_rq_pos(rq) - cfqd->last_position;
1663 return cfqd->last_position - blk_rq_pos(rq);
1666 #define CFQQ_SEEK_THR 8 * 1024
1667 #define CFQQ_SEEKY(cfqq) ((cfqq)->seek_mean > CFQQ_SEEK_THR)
1669 static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1672 sector_t sdist = cfqq->seek_mean;
1674 if (!sample_valid(cfqq->seek_samples))
1675 sdist = CFQQ_SEEK_THR;
1677 return cfq_dist_from_last(cfqd, rq) <= sdist;
1680 static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
1681 struct cfq_queue *cur_cfqq)
1683 struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
1684 struct rb_node *parent, *node;
1685 struct cfq_queue *__cfqq;
1686 sector_t sector = cfqd->last_position;
1688 if (RB_EMPTY_ROOT(root))
1692 * First, if we find a request starting at the end of the last
1693 * request, choose it.
1695 __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
1700 * If the exact sector wasn't found, the parent of the NULL leaf
1701 * will contain the closest sector.
1703 __cfqq = rb_entry(parent, struct cfq_queue, p_node);
1704 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1707 if (blk_rq_pos(__cfqq->next_rq) < sector)
1708 node = rb_next(&__cfqq->p_node);
1710 node = rb_prev(&__cfqq->p_node);
1714 __cfqq = rb_entry(node, struct cfq_queue, p_node);
1715 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1723 * cur_cfqq - passed in so that we don't decide that the current queue is
1724 * closely cooperating with itself.
1726 * So, basically we're assuming that that cur_cfqq has dispatched at least
1727 * one request, and that cfqd->last_position reflects a position on the disk
1728 * associated with the I/O issued by cur_cfqq. I'm not sure this is a valid
1731 static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
1732 struct cfq_queue *cur_cfqq)
1734 struct cfq_queue *cfqq;
1736 if (!cfq_cfqq_sync(cur_cfqq))
1738 if (CFQQ_SEEKY(cur_cfqq))
1742 * Don't search priority tree if it's the only queue in the group.
1744 if (cur_cfqq->cfqg->nr_cfqq == 1)
1748 * We should notice if some of the queues are cooperating, eg
1749 * working closely on the same area of the disk. In that case,
1750 * we can group them together and don't waste time idling.
1752 cfqq = cfqq_close(cfqd, cur_cfqq);
1756 /* If new queue belongs to different cfq_group, don't choose it */
1757 if (cur_cfqq->cfqg != cfqq->cfqg)
1761 * It only makes sense to merge sync queues.
1763 if (!cfq_cfqq_sync(cfqq))
1765 if (CFQQ_SEEKY(cfqq))
1769 * Do not merge queues of different priority classes
1771 if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq))
1778 * Determine whether we should enforce idle window for this queue.
1781 static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1783 enum wl_prio_t prio = cfqq_prio(cfqq);
1784 struct cfq_rb_root *service_tree = cfqq->service_tree;
1786 BUG_ON(!service_tree);
1787 BUG_ON(!service_tree->count);
1789 /* We never do for idle class queues. */
1790 if (prio == IDLE_WORKLOAD)
1793 /* We do for queues that were marked with idle window flag. */
1794 if (cfq_cfqq_idle_window(cfqq) &&
1795 !(blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag))
1799 * Otherwise, we do only if they are the last ones
1800 * in their service tree.
1802 return service_tree->count == 1;
1805 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
1807 struct cfq_queue *cfqq = cfqd->active_queue;
1808 struct cfq_io_context *cic;
1812 * SSD device without seek penalty, disable idling. But only do so
1813 * for devices that support queuing, otherwise we still have a problem
1814 * with sync vs async workloads.
1816 if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
1819 WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
1820 WARN_ON(cfq_cfqq_slice_new(cfqq));
1823 * idle is disabled, either manually or by past process history
1825 if (!cfqd->cfq_slice_idle || !cfq_should_idle(cfqd, cfqq))
1829 * still active requests from this queue, don't idle
1831 if (cfqq->dispatched)
1835 * task has exited, don't wait
1837 cic = cfqd->active_cic;
1838 if (!cic || !atomic_read(&cic->ioc->nr_tasks))
1842 * If our average think time is larger than the remaining time
1843 * slice, then don't idle. This avoids overrunning the allotted
1846 if (sample_valid(cic->ttime_samples) &&
1847 (cfqq->slice_end - jiffies < cic->ttime_mean))
1850 cfq_mark_cfqq_wait_request(cfqq);
1852 sl = cfqd->cfq_slice_idle;
1854 mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
1855 cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu", sl);
1859 * Move request from internal lists to the request queue dispatch list.
1861 static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
1863 struct cfq_data *cfqd = q->elevator->elevator_data;
1864 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1866 cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
1868 cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
1869 cfq_remove_request(rq);
1871 elv_dispatch_sort(q, rq);
1873 if (cfq_cfqq_sync(cfqq))
1874 cfqd->sync_flight++;
1875 cfqq->nr_sectors += blk_rq_sectors(rq);
1879 * return expired entry, or NULL to just start from scratch in rbtree
1881 static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
1883 struct request *rq = NULL;
1885 if (cfq_cfqq_fifo_expire(cfqq))
1888 cfq_mark_cfqq_fifo_expire(cfqq);
1890 if (list_empty(&cfqq->fifo))
1893 rq = rq_entry_fifo(cfqq->fifo.next);
1894 if (time_before(jiffies, rq_fifo_time(rq)))
1897 cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq);
1902 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1904 const int base_rq = cfqd->cfq_slice_async_rq;
1906 WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
1908 return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
1912 * Must be called with the queue_lock held.
1914 static int cfqq_process_refs(struct cfq_queue *cfqq)
1916 int process_refs, io_refs;
1918 io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE];
1919 process_refs = atomic_read(&cfqq->ref) - io_refs;
1920 BUG_ON(process_refs < 0);
1921 return process_refs;
1924 static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq)
1926 int process_refs, new_process_refs;
1927 struct cfq_queue *__cfqq;
1929 /* Avoid a circular list and skip interim queue merges */
1930 while ((__cfqq = new_cfqq->new_cfqq)) {
1936 process_refs = cfqq_process_refs(cfqq);
1938 * If the process for the cfqq has gone away, there is no
1939 * sense in merging the queues.
1941 if (process_refs == 0)
1945 * Merge in the direction of the lesser amount of work.
1947 new_process_refs = cfqq_process_refs(new_cfqq);
1948 if (new_process_refs >= process_refs) {
1949 cfqq->new_cfqq = new_cfqq;
1950 atomic_add(process_refs, &new_cfqq->ref);
1952 new_cfqq->new_cfqq = cfqq;
1953 atomic_add(new_process_refs, &cfqq->ref);
1957 static enum wl_type_t cfq_choose_wl(struct cfq_data *cfqd,
1958 struct cfq_group *cfqg, enum wl_prio_t prio)
1960 struct cfq_queue *queue;
1962 bool key_valid = false;
1963 unsigned long lowest_key = 0;
1964 enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD;
1966 for (i = 0; i <= SYNC_WORKLOAD; ++i) {
1967 /* select the one with lowest rb_key */
1968 queue = cfq_rb_first(service_tree_for(cfqg, prio, i));
1970 (!key_valid || time_before(queue->rb_key, lowest_key))) {
1971 lowest_key = queue->rb_key;
1980 static void choose_service_tree(struct cfq_data *cfqd, struct cfq_group *cfqg)
1984 struct cfq_rb_root *st;
1985 unsigned group_slice;
1988 cfqd->serving_prio = IDLE_WORKLOAD;
1989 cfqd->workload_expires = jiffies + 1;
1993 /* Choose next priority. RT > BE > IDLE */
1994 if (cfq_group_busy_queues_wl(RT_WORKLOAD, cfqd, cfqg))
1995 cfqd->serving_prio = RT_WORKLOAD;
1996 else if (cfq_group_busy_queues_wl(BE_WORKLOAD, cfqd, cfqg))
1997 cfqd->serving_prio = BE_WORKLOAD;
1999 cfqd->serving_prio = IDLE_WORKLOAD;
2000 cfqd->workload_expires = jiffies + 1;
2005 * For RT and BE, we have to choose also the type
2006 * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
2009 st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
2013 * check workload expiration, and that we still have other queues ready
2015 if (count && !time_after(jiffies, cfqd->workload_expires))
2018 /* otherwise select new workload type */
2019 cfqd->serving_type =
2020 cfq_choose_wl(cfqd, cfqg, cfqd->serving_prio);
2021 st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
2025 * the workload slice is computed as a fraction of target latency
2026 * proportional to the number of queues in that workload, over
2027 * all the queues in the same priority class
2029 group_slice = cfq_group_slice(cfqd, cfqg);
2031 slice = group_slice * count /
2032 max_t(unsigned, cfqg->busy_queues_avg[cfqd->serving_prio],
2033 cfq_group_busy_queues_wl(cfqd->serving_prio, cfqd, cfqg));
2035 if (cfqd->serving_type == ASYNC_WORKLOAD) {
2039 * Async queues are currently system wide. Just taking
2040 * proportion of queues with-in same group will lead to higher
2041 * async ratio system wide as generally root group is going
2042 * to have higher weight. A more accurate thing would be to
2043 * calculate system wide asnc/sync ratio.
2045 tmp = cfq_target_latency * cfqg_busy_async_queues(cfqd, cfqg);
2046 tmp = tmp/cfqd->busy_queues;
2047 slice = min_t(unsigned, slice, tmp);
2049 /* async workload slice is scaled down according to
2050 * the sync/async slice ratio. */
2051 slice = slice * cfqd->cfq_slice[0] / cfqd->cfq_slice[1];
2053 /* sync workload slice is at least 2 * cfq_slice_idle */
2054 slice = max(slice, 2 * cfqd->cfq_slice_idle);
2056 slice = max_t(unsigned, slice, CFQ_MIN_TT);
2057 cfqd->workload_expires = jiffies + slice;
2058 cfqd->noidle_tree_requires_idle = false;
2061 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd)
2063 struct cfq_rb_root *st = &cfqd->grp_service_tree;
2064 struct cfq_group *cfqg;
2066 if (RB_EMPTY_ROOT(&st->rb))
2068 cfqg = cfq_rb_first_group(st);
2069 st->active = &cfqg->rb_node;
2070 update_min_vdisktime(st);
2074 static void cfq_choose_cfqg(struct cfq_data *cfqd)
2076 struct cfq_group *cfqg = cfq_get_next_cfqg(cfqd);
2078 cfqd->serving_group = cfqg;
2080 /* Restore the workload type data */
2081 if (cfqg->saved_workload_slice) {
2082 cfqd->workload_expires = jiffies + cfqg->saved_workload_slice;
2083 cfqd->serving_type = cfqg->saved_workload;
2084 cfqd->serving_prio = cfqg->saved_serving_prio;
2086 cfqd->workload_expires = jiffies - 1;
2088 choose_service_tree(cfqd, cfqg);
2092 * Select a queue for service. If we have a current active queue,
2093 * check whether to continue servicing it, or retrieve and set a new one.
2095 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
2097 struct cfq_queue *cfqq, *new_cfqq = NULL;
2099 cfqq = cfqd->active_queue;
2103 if (!cfqd->rq_queued)
2107 * We were waiting for group to get backlogged. Expire the queue
2109 if (cfq_cfqq_wait_busy(cfqq) && !RB_EMPTY_ROOT(&cfqq->sort_list))
2113 * The active queue has run out of time, expire it and select new.
2115 if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq)) {
2117 * If slice had not expired at the completion of last request
2118 * we might not have turned on wait_busy flag. Don't expire
2119 * the queue yet. Allow the group to get backlogged.
2121 * The very fact that we have used the slice, that means we
2122 * have been idling all along on this queue and it should be
2123 * ok to wait for this request to complete.
2125 if (cfqq->cfqg->nr_cfqq == 1 && RB_EMPTY_ROOT(&cfqq->sort_list)
2126 && cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2134 * The active queue has requests and isn't expired, allow it to
2137 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
2141 * If another queue has a request waiting within our mean seek
2142 * distance, let it run. The expire code will check for close
2143 * cooperators and put the close queue at the front of the service
2144 * tree. If possible, merge the expiring queue with the new cfqq.
2146 new_cfqq = cfq_close_cooperator(cfqd, cfqq);
2148 if (!cfqq->new_cfqq)
2149 cfq_setup_merge(cfqq, new_cfqq);
2154 * No requests pending. If the active queue still has requests in
2155 * flight or is idling for a new request, allow either of these
2156 * conditions to happen (or time out) before selecting a new queue.
2158 if (timer_pending(&cfqd->idle_slice_timer) ||
2159 (cfqq->dispatched && cfq_should_idle(cfqd, cfqq))) {
2165 cfq_slice_expired(cfqd, 0);
2168 * Current queue expired. Check if we have to switch to a new
2172 cfq_choose_cfqg(cfqd);
2174 cfqq = cfq_set_active_queue(cfqd, new_cfqq);
2179 static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
2183 while (cfqq->next_rq) {
2184 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
2188 BUG_ON(!list_empty(&cfqq->fifo));
2190 /* By default cfqq is not expired if it is empty. Do it explicitly */
2191 __cfq_slice_expired(cfqq->cfqd, cfqq, 0);
2196 * Drain our current requests. Used for barriers and when switching
2197 * io schedulers on-the-fly.
2199 static int cfq_forced_dispatch(struct cfq_data *cfqd)
2201 struct cfq_queue *cfqq;
2204 while ((cfqq = cfq_get_next_queue_forced(cfqd)) != NULL)
2205 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
2207 cfq_slice_expired(cfqd, 0);
2208 BUG_ON(cfqd->busy_queues);
2210 cfq_log(cfqd, "forced_dispatch=%d", dispatched);
2214 static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2216 unsigned int max_dispatch;
2219 * Drain async requests before we start sync IO
2221 if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_driver[BLK_RW_ASYNC])
2225 * If this is an async queue and we have sync IO in flight, let it wait
2227 if (cfqd->sync_flight && !cfq_cfqq_sync(cfqq))
2230 max_dispatch = cfqd->cfq_quantum;
2231 if (cfq_class_idle(cfqq))
2235 * Does this cfqq already have too much IO in flight?
2237 if (cfqq->dispatched >= max_dispatch) {
2239 * idle queue must always only have a single IO in flight
2241 if (cfq_class_idle(cfqq))
2245 * We have other queues, don't allow more IO from this one
2247 if (cfqd->busy_queues > 1)
2251 * Sole queue user, no limit
2257 * Async queues must wait a bit before being allowed dispatch.
2258 * We also ramp up the dispatch depth gradually for async IO,
2259 * based on the last sync IO we serviced
2261 if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
2262 unsigned long last_sync = jiffies - cfqd->last_delayed_sync;
2265 depth = last_sync / cfqd->cfq_slice[1];
2266 if (!depth && !cfqq->dispatched)
2268 if (depth < max_dispatch)
2269 max_dispatch = depth;
2273 * If we're below the current max, allow a dispatch
2275 return cfqq->dispatched < max_dispatch;
2279 * Dispatch a request from cfqq, moving them to the request queue
2282 static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2286 BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
2288 if (!cfq_may_dispatch(cfqd, cfqq))
2292 * follow expired path, else get first next available
2294 rq = cfq_check_fifo(cfqq);
2299 * insert request into driver dispatch list
2301 cfq_dispatch_insert(cfqd->queue, rq);
2303 if (!cfqd->active_cic) {
2304 struct cfq_io_context *cic = RQ_CIC(rq);
2306 atomic_long_inc(&cic->ioc->refcount);
2307 cfqd->active_cic = cic;
2314 * Find the cfqq that we need to service and move a request from that to the
2317 static int cfq_dispatch_requests(struct request_queue *q, int force)
2319 struct cfq_data *cfqd = q->elevator->elevator_data;
2320 struct cfq_queue *cfqq;
2322 if (!cfqd->busy_queues)
2325 if (unlikely(force))
2326 return cfq_forced_dispatch(cfqd);
2328 cfqq = cfq_select_queue(cfqd);
2333 * Dispatch a request from this cfqq, if it is allowed
2335 if (!cfq_dispatch_request(cfqd, cfqq))
2338 cfqq->slice_dispatch++;
2339 cfq_clear_cfqq_must_dispatch(cfqq);
2342 * expire an async queue immediately if it has used up its slice. idle
2343 * queue always expire after 1 dispatch round.
2345 if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
2346 cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
2347 cfq_class_idle(cfqq))) {
2348 cfqq->slice_end = jiffies + 1;
2349 cfq_slice_expired(cfqd, 0);
2352 cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
2357 * task holds one reference to the queue, dropped when task exits. each rq
2358 * in-flight on this queue also holds a reference, dropped when rq is freed.
2360 * Each cfq queue took a reference on the parent group. Drop it now.
2361 * queue lock must be held here.
2363 static void cfq_put_queue(struct cfq_queue *cfqq)
2365 struct cfq_data *cfqd = cfqq->cfqd;
2366 struct cfq_group *cfqg, *orig_cfqg;
2368 BUG_ON(atomic_read(&cfqq->ref) <= 0);
2370 if (!atomic_dec_and_test(&cfqq->ref))
2373 cfq_log_cfqq(cfqd, cfqq, "put_queue");
2374 BUG_ON(rb_first(&cfqq->sort_list));
2375 BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
2377 orig_cfqg = cfqq->orig_cfqg;
2379 if (unlikely(cfqd->active_queue == cfqq)) {
2380 __cfq_slice_expired(cfqd, cfqq, 0);
2381 cfq_schedule_dispatch(cfqd);
2384 BUG_ON(cfq_cfqq_on_rr(cfqq));
2385 kmem_cache_free(cfq_pool, cfqq);
2388 cfq_put_cfqg(orig_cfqg);
2392 * Must always be called with the rcu_read_lock() held
2395 __call_for_each_cic(struct io_context *ioc,
2396 void (*func)(struct io_context *, struct cfq_io_context *))
2398 struct cfq_io_context *cic;
2399 struct hlist_node *n;
2401 hlist_for_each_entry_rcu(cic, n, &ioc->cic_list, cic_list)
2406 * Call func for each cic attached to this ioc.
2409 call_for_each_cic(struct io_context *ioc,
2410 void (*func)(struct io_context *, struct cfq_io_context *))
2413 __call_for_each_cic(ioc, func);
2417 static void cfq_cic_free_rcu(struct rcu_head *head)
2419 struct cfq_io_context *cic;
2421 cic = container_of(head, struct cfq_io_context, rcu_head);
2423 kmem_cache_free(cfq_ioc_pool, cic);
2424 elv_ioc_count_dec(cfq_ioc_count);
2428 * CFQ scheduler is exiting, grab exit lock and check
2429 * the pending io context count. If it hits zero,
2430 * complete ioc_gone and set it back to NULL
2432 spin_lock(&ioc_gone_lock);
2433 if (ioc_gone && !elv_ioc_count_read(cfq_ioc_count)) {
2437 spin_unlock(&ioc_gone_lock);
2441 static void cfq_cic_free(struct cfq_io_context *cic)
2443 call_rcu(&cic->rcu_head, cfq_cic_free_rcu);
2446 static void cic_free_func(struct io_context *ioc, struct cfq_io_context *cic)
2448 unsigned long flags;
2450 BUG_ON(!cic->dead_key);
2452 spin_lock_irqsave(&ioc->lock, flags);
2453 radix_tree_delete(&ioc->radix_root, cic->dead_key);
2454 hlist_del_rcu(&cic->cic_list);
2455 spin_unlock_irqrestore(&ioc->lock, flags);
2461 * Must be called with rcu_read_lock() held or preemption otherwise disabled.
2462 * Only two callers of this - ->dtor() which is called with the rcu_read_lock(),
2463 * and ->trim() which is called with the task lock held
2465 static void cfq_free_io_context(struct io_context *ioc)
2468 * ioc->refcount is zero here, or we are called from elv_unregister(),
2469 * so no more cic's are allowed to be linked into this ioc. So it
2470 * should be ok to iterate over the known list, we will see all cic's
2471 * since no new ones are added.
2473 __call_for_each_cic(ioc, cic_free_func);
2476 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2478 struct cfq_queue *__cfqq, *next;
2480 if (unlikely(cfqq == cfqd->active_queue)) {
2481 __cfq_slice_expired(cfqd, cfqq, 0);
2482 cfq_schedule_dispatch(cfqd);
2486 * If this queue was scheduled to merge with another queue, be
2487 * sure to drop the reference taken on that queue (and others in
2488 * the merge chain). See cfq_setup_merge and cfq_merge_cfqqs.
2490 __cfqq = cfqq->new_cfqq;
2492 if (__cfqq == cfqq) {
2493 WARN(1, "cfqq->new_cfqq loop detected\n");
2496 next = __cfqq->new_cfqq;
2497 cfq_put_queue(__cfqq);
2501 cfq_put_queue(cfqq);
2504 static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
2505 struct cfq_io_context *cic)
2507 struct io_context *ioc = cic->ioc;
2509 list_del_init(&cic->queue_list);
2512 * Make sure key == NULL is seen for dead queues
2515 cic->dead_key = (unsigned long) cic->key;
2518 if (ioc->ioc_data == cic)
2519 rcu_assign_pointer(ioc->ioc_data, NULL);
2521 if (cic->cfqq[BLK_RW_ASYNC]) {
2522 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]);
2523 cic->cfqq[BLK_RW_ASYNC] = NULL;
2526 if (cic->cfqq[BLK_RW_SYNC]) {
2527 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]);
2528 cic->cfqq[BLK_RW_SYNC] = NULL;
2532 static void cfq_exit_single_io_context(struct io_context *ioc,
2533 struct cfq_io_context *cic)
2535 struct cfq_data *cfqd = cic->key;
2538 struct request_queue *q = cfqd->queue;
2539 unsigned long flags;
2541 spin_lock_irqsave(q->queue_lock, flags);
2544 * Ensure we get a fresh copy of the ->key to prevent
2545 * race between exiting task and queue
2547 smp_read_barrier_depends();
2549 __cfq_exit_single_io_context(cfqd, cic);
2551 spin_unlock_irqrestore(q->queue_lock, flags);
2556 * The process that ioc belongs to has exited, we need to clean up
2557 * and put the internal structures we have that belongs to that process.
2559 static void cfq_exit_io_context(struct io_context *ioc)
2561 call_for_each_cic(ioc, cfq_exit_single_io_context);
2564 static struct cfq_io_context *
2565 cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
2567 struct cfq_io_context *cic;
2569 cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO,
2572 cic->last_end_request = jiffies;
2573 INIT_LIST_HEAD(&cic->queue_list);
2574 INIT_HLIST_NODE(&cic->cic_list);
2575 cic->dtor = cfq_free_io_context;
2576 cic->exit = cfq_exit_io_context;
2577 elv_ioc_count_inc(cfq_ioc_count);
2583 static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc)
2585 struct task_struct *tsk = current;
2588 if (!cfq_cfqq_prio_changed(cfqq))
2591 ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio);
2592 switch (ioprio_class) {
2594 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
2595 case IOPRIO_CLASS_NONE:
2597 * no prio set, inherit CPU scheduling settings
2599 cfqq->ioprio = task_nice_ioprio(tsk);
2600 cfqq->ioprio_class = task_nice_ioclass(tsk);
2602 case IOPRIO_CLASS_RT:
2603 cfqq->ioprio = task_ioprio(ioc);
2604 cfqq->ioprio_class = IOPRIO_CLASS_RT;
2606 case IOPRIO_CLASS_BE:
2607 cfqq->ioprio = task_ioprio(ioc);
2608 cfqq->ioprio_class = IOPRIO_CLASS_BE;
2610 case IOPRIO_CLASS_IDLE:
2611 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
2613 cfq_clear_cfqq_idle_window(cfqq);
2618 * keep track of original prio settings in case we have to temporarily
2619 * elevate the priority of this queue
2621 cfqq->org_ioprio = cfqq->ioprio;
2622 cfqq->org_ioprio_class = cfqq->ioprio_class;
2623 cfq_clear_cfqq_prio_changed(cfqq);
2626 static void changed_ioprio(struct io_context *ioc, struct cfq_io_context *cic)
2628 struct cfq_data *cfqd = cic->key;
2629 struct cfq_queue *cfqq;
2630 unsigned long flags;
2632 if (unlikely(!cfqd))
2635 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2637 cfqq = cic->cfqq[BLK_RW_ASYNC];
2639 struct cfq_queue *new_cfqq;
2640 new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic->ioc,
2643 cic->cfqq[BLK_RW_ASYNC] = new_cfqq;
2644 cfq_put_queue(cfqq);
2648 cfqq = cic->cfqq[BLK_RW_SYNC];
2650 cfq_mark_cfqq_prio_changed(cfqq);
2652 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2655 static void cfq_ioc_set_ioprio(struct io_context *ioc)
2657 call_for_each_cic(ioc, changed_ioprio);
2658 ioc->ioprio_changed = 0;
2661 static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2662 pid_t pid, bool is_sync)
2664 RB_CLEAR_NODE(&cfqq->rb_node);
2665 RB_CLEAR_NODE(&cfqq->p_node);
2666 INIT_LIST_HEAD(&cfqq->fifo);
2668 atomic_set(&cfqq->ref, 0);
2671 cfq_mark_cfqq_prio_changed(cfqq);
2674 if (!cfq_class_idle(cfqq))
2675 cfq_mark_cfqq_idle_window(cfqq);
2676 cfq_mark_cfqq_sync(cfqq);
2681 #ifdef CONFIG_CFQ_GROUP_IOSCHED
2682 static void changed_cgroup(struct io_context *ioc, struct cfq_io_context *cic)
2684 struct cfq_queue *sync_cfqq = cic_to_cfqq(cic, 1);
2685 struct cfq_data *cfqd = cic->key;
2686 unsigned long flags;
2687 struct request_queue *q;
2689 if (unlikely(!cfqd))
2694 spin_lock_irqsave(q->queue_lock, flags);
2698 * Drop reference to sync queue. A new sync queue will be
2699 * assigned in new group upon arrival of a fresh request.
2701 cfq_log_cfqq(cfqd, sync_cfqq, "changed cgroup");
2702 cic_set_cfqq(cic, NULL, 1);
2703 cfq_put_queue(sync_cfqq);
2706 spin_unlock_irqrestore(q->queue_lock, flags);
2709 static void cfq_ioc_set_cgroup(struct io_context *ioc)
2711 call_for_each_cic(ioc, changed_cgroup);
2712 ioc->cgroup_changed = 0;
2714 #endif /* CONFIG_CFQ_GROUP_IOSCHED */
2716 static struct cfq_queue *
2717 cfq_find_alloc_queue(struct cfq_data *cfqd, bool is_sync,
2718 struct io_context *ioc, gfp_t gfp_mask)
2720 struct cfq_queue *cfqq, *new_cfqq = NULL;
2721 struct cfq_io_context *cic;
2722 struct cfq_group *cfqg;
2725 cfqg = cfq_get_cfqg(cfqd, 1);
2726 cic = cfq_cic_lookup(cfqd, ioc);
2727 /* cic always exists here */
2728 cfqq = cic_to_cfqq(cic, is_sync);
2731 * Always try a new alloc if we fell back to the OOM cfqq
2732 * originally, since it should just be a temporary situation.
2734 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
2739 } else if (gfp_mask & __GFP_WAIT) {
2740 spin_unlock_irq(cfqd->queue->queue_lock);
2741 new_cfqq = kmem_cache_alloc_node(cfq_pool,
2742 gfp_mask | __GFP_ZERO,
2744 spin_lock_irq(cfqd->queue->queue_lock);
2748 cfqq = kmem_cache_alloc_node(cfq_pool,
2749 gfp_mask | __GFP_ZERO,
2754 cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
2755 cfq_init_prio_data(cfqq, ioc);
2756 cfq_link_cfqq_cfqg(cfqq, cfqg);
2757 cfq_log_cfqq(cfqd, cfqq, "alloced");
2759 cfqq = &cfqd->oom_cfqq;
2763 kmem_cache_free(cfq_pool, new_cfqq);
2768 static struct cfq_queue **
2769 cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
2771 switch (ioprio_class) {
2772 case IOPRIO_CLASS_RT:
2773 return &cfqd->async_cfqq[0][ioprio];
2774 case IOPRIO_CLASS_BE:
2775 return &cfqd->async_cfqq[1][ioprio];
2776 case IOPRIO_CLASS_IDLE:
2777 return &cfqd->async_idle_cfqq;
2783 static struct cfq_queue *
2784 cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct io_context *ioc,
2787 const int ioprio = task_ioprio(ioc);
2788 const int ioprio_class = task_ioprio_class(ioc);
2789 struct cfq_queue **async_cfqq = NULL;
2790 struct cfq_queue *cfqq = NULL;
2793 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
2798 cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask);
2801 * pin the queue now that it's allocated, scheduler exit will prune it
2803 if (!is_sync && !(*async_cfqq)) {
2804 atomic_inc(&cfqq->ref);
2808 atomic_inc(&cfqq->ref);
2813 * We drop cfq io contexts lazily, so we may find a dead one.
2816 cfq_drop_dead_cic(struct cfq_data *cfqd, struct io_context *ioc,
2817 struct cfq_io_context *cic)
2819 unsigned long flags;
2821 WARN_ON(!list_empty(&cic->queue_list));
2823 spin_lock_irqsave(&ioc->lock, flags);
2825 BUG_ON(ioc->ioc_data == cic);
2827 radix_tree_delete(&ioc->radix_root, (unsigned long) cfqd);
2828 hlist_del_rcu(&cic->cic_list);
2829 spin_unlock_irqrestore(&ioc->lock, flags);
2834 static struct cfq_io_context *
2835 cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc)
2837 struct cfq_io_context *cic;
2838 unsigned long flags;
2847 * we maintain a last-hit cache, to avoid browsing over the tree
2849 cic = rcu_dereference(ioc->ioc_data);
2850 if (cic && cic->key == cfqd) {
2856 cic = radix_tree_lookup(&ioc->radix_root, (unsigned long) cfqd);
2860 /* ->key must be copied to avoid race with cfq_exit_queue() */
2863 cfq_drop_dead_cic(cfqd, ioc, cic);
2868 spin_lock_irqsave(&ioc->lock, flags);
2869 rcu_assign_pointer(ioc->ioc_data, cic);
2870 spin_unlock_irqrestore(&ioc->lock, flags);
2878 * Add cic into ioc, using cfqd as the search key. This enables us to lookup
2879 * the process specific cfq io context when entered from the block layer.
2880 * Also adds the cic to a per-cfqd list, used when this queue is removed.
2882 static int cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
2883 struct cfq_io_context *cic, gfp_t gfp_mask)
2885 unsigned long flags;
2888 ret = radix_tree_preload(gfp_mask);
2893 spin_lock_irqsave(&ioc->lock, flags);
2894 ret = radix_tree_insert(&ioc->radix_root,
2895 (unsigned long) cfqd, cic);
2897 hlist_add_head_rcu(&cic->cic_list, &ioc->cic_list);
2898 spin_unlock_irqrestore(&ioc->lock, flags);
2900 radix_tree_preload_end();
2903 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2904 list_add(&cic->queue_list, &cfqd->cic_list);
2905 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2910 printk(KERN_ERR "cfq: cic link failed!\n");
2916 * Setup general io context and cfq io context. There can be several cfq
2917 * io contexts per general io context, if this process is doing io to more
2918 * than one device managed by cfq.
2920 static struct cfq_io_context *
2921 cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
2923 struct io_context *ioc = NULL;
2924 struct cfq_io_context *cic;
2926 might_sleep_if(gfp_mask & __GFP_WAIT);
2928 ioc = get_io_context(gfp_mask, cfqd->queue->node);
2932 cic = cfq_cic_lookup(cfqd, ioc);
2936 cic = cfq_alloc_io_context(cfqd, gfp_mask);
2940 if (cfq_cic_link(cfqd, ioc, cic, gfp_mask))
2944 smp_read_barrier_depends();
2945 if (unlikely(ioc->ioprio_changed))
2946 cfq_ioc_set_ioprio(ioc);
2948 #ifdef CONFIG_CFQ_GROUP_IOSCHED
2949 if (unlikely(ioc->cgroup_changed))
2950 cfq_ioc_set_cgroup(ioc);
2956 put_io_context(ioc);
2961 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
2963 unsigned long elapsed = jiffies - cic->last_end_request;
2964 unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
2966 cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
2967 cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
2968 cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
2972 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2978 if (!cfqq->last_request_pos)
2980 else if (cfqq->last_request_pos < blk_rq_pos(rq))
2981 sdist = blk_rq_pos(rq) - cfqq->last_request_pos;
2983 sdist = cfqq->last_request_pos - blk_rq_pos(rq);
2986 * Don't allow the seek distance to get too large from the
2987 * odd fragment, pagein, etc
2989 if (cfqq->seek_samples <= 60) /* second&third seek */
2990 sdist = min(sdist, (cfqq->seek_mean * 4) + 2*1024*1024);
2992 sdist = min(sdist, (cfqq->seek_mean * 4) + 2*1024*64);
2994 cfqq->seek_samples = (7*cfqq->seek_samples + 256) / 8;
2995 cfqq->seek_total = (7*cfqq->seek_total + (u64)256*sdist) / 8;
2996 total = cfqq->seek_total + (cfqq->seek_samples/2);
2997 do_div(total, cfqq->seek_samples);
2998 cfqq->seek_mean = (sector_t)total;
3001 * If this cfqq is shared between multiple processes, check to
3002 * make sure that those processes are still issuing I/Os within
3003 * the mean seek distance. If not, it may be time to break the
3004 * queues apart again.
3006 if (cfq_cfqq_coop(cfqq)) {
3007 if (CFQQ_SEEKY(cfqq) && !cfqq->seeky_start)
3008 cfqq->seeky_start = jiffies;
3009 else if (!CFQQ_SEEKY(cfqq))
3010 cfqq->seeky_start = 0;
3015 * Disable idle window if the process thinks too long or seeks so much that
3019 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3020 struct cfq_io_context *cic)
3022 int old_idle, enable_idle;
3025 * Don't idle for async or idle io prio class
3027 if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
3030 enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
3032 if (cfqq->queued[0] + cfqq->queued[1] >= 4)
3033 cfq_mark_cfqq_deep(cfqq);
3035 if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle ||
3036 (!cfq_cfqq_deep(cfqq) && sample_valid(cfqq->seek_samples)
3037 && CFQQ_SEEKY(cfqq)))
3039 else if (sample_valid(cic->ttime_samples)) {
3040 if (cic->ttime_mean > cfqd->cfq_slice_idle)
3046 if (old_idle != enable_idle) {
3047 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
3049 cfq_mark_cfqq_idle_window(cfqq);
3051 cfq_clear_cfqq_idle_window(cfqq);
3056 * Check if new_cfqq should preempt the currently active queue. Return 0 for
3057 * no or if we aren't sure, a 1 will cause a preempt.
3060 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
3063 struct cfq_queue *cfqq;
3065 cfqq = cfqd->active_queue;
3069 if (cfq_class_idle(new_cfqq))
3072 if (cfq_class_idle(cfqq))
3076 * if the new request is sync, but the currently running queue is
3077 * not, let the sync request have priority.
3079 if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
3082 if (new_cfqq->cfqg != cfqq->cfqg)
3085 if (cfq_slice_used(cfqq))
3088 /* Allow preemption only if we are idling on sync-noidle tree */
3089 if (cfqd->serving_type == SYNC_NOIDLE_WORKLOAD &&
3090 cfqq_type(new_cfqq) == SYNC_NOIDLE_WORKLOAD &&
3091 new_cfqq->service_tree->count == 2 &&
3092 RB_EMPTY_ROOT(&cfqq->sort_list))
3096 * So both queues are sync. Let the new request get disk time if
3097 * it's a metadata request and the current queue is doing regular IO.
3099 if (rq_is_meta(rq) && !cfqq->meta_pending)
3103 * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
3105 if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
3108 if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
3112 * if this request is as-good as one we would expect from the
3113 * current cfqq, let it preempt
3115 if (cfq_rq_close(cfqd, cfqq, rq))
3122 * cfqq preempts the active queue. if we allowed preempt with no slice left,
3123 * let it have half of its nominal slice.
3125 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3127 cfq_log_cfqq(cfqd, cfqq, "preempt");
3128 cfq_slice_expired(cfqd, 1);
3131 * Put the new queue at the front of the of the current list,
3132 * so we know that it will be selected next.
3134 BUG_ON(!cfq_cfqq_on_rr(cfqq));
3136 cfq_service_tree_add(cfqd, cfqq, 1);
3138 cfqq->slice_end = 0;
3139 cfq_mark_cfqq_slice_new(cfqq);
3143 * Called when a new fs request (rq) is added (to cfqq). Check if there's
3144 * something we should do about it
3147 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3150 struct cfq_io_context *cic = RQ_CIC(rq);
3154 cfqq->meta_pending++;
3156 cfq_update_io_thinktime(cfqd, cic);
3157 cfq_update_io_seektime(cfqd, cfqq, rq);
3158 cfq_update_idle_window(cfqd, cfqq, cic);
3160 cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
3162 if (cfqq == cfqd->active_queue) {
3164 * Remember that we saw a request from this process, but
3165 * don't start queuing just yet. Otherwise we risk seeing lots
3166 * of tiny requests, because we disrupt the normal plugging
3167 * and merging. If the request is already larger than a single
3168 * page, let it rip immediately. For that case we assume that
3169 * merging is already done. Ditto for a busy system that
3170 * has other work pending, don't risk delaying until the
3171 * idle timer unplug to continue working.
3173 if (cfq_cfqq_wait_request(cfqq)) {
3174 if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
3175 cfqd->busy_queues > 1) {
3176 del_timer(&cfqd->idle_slice_timer);
3177 cfq_clear_cfqq_wait_request(cfqq);
3178 __blk_run_queue(cfqd->queue);
3180 cfq_mark_cfqq_must_dispatch(cfqq);
3182 } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
3184 * not the active queue - expire current slice if it is
3185 * idle and has expired it's mean thinktime or this new queue
3186 * has some old slice time left and is of higher priority or
3187 * this new queue is RT and the current one is BE
3189 cfq_preempt_queue(cfqd, cfqq);
3190 __blk_run_queue(cfqd->queue);
3194 static void cfq_insert_request(struct request_queue *q, struct request *rq)
3196 struct cfq_data *cfqd = q->elevator->elevator_data;
3197 struct cfq_queue *cfqq = RQ_CFQQ(rq);
3199 cfq_log_cfqq(cfqd, cfqq, "insert_request");
3200 cfq_init_prio_data(cfqq, RQ_CIC(rq)->ioc);
3202 rq_set_fifo_time(rq, jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)]);
3203 list_add_tail(&rq->queuelist, &cfqq->fifo);
3206 cfq_rq_enqueued(cfqd, cfqq, rq);
3210 * Update hw_tag based on peak queue depth over 50 samples under
3213 static void cfq_update_hw_tag(struct cfq_data *cfqd)
3215 struct cfq_queue *cfqq = cfqd->active_queue;
3217 if (rq_in_driver(cfqd) > cfqd->hw_tag_est_depth)
3218 cfqd->hw_tag_est_depth = rq_in_driver(cfqd);
3220 if (cfqd->hw_tag == 1)
3223 if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
3224 rq_in_driver(cfqd) <= CFQ_HW_QUEUE_MIN)
3228 * If active queue hasn't enough requests and can idle, cfq might not
3229 * dispatch sufficient requests to hardware. Don't zero hw_tag in this
3232 if (cfqq && cfq_cfqq_idle_window(cfqq) &&
3233 cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] <
3234 CFQ_HW_QUEUE_MIN && rq_in_driver(cfqd) < CFQ_HW_QUEUE_MIN)
3237 if (cfqd->hw_tag_samples++ < 50)
3240 if (cfqd->hw_tag_est_depth >= CFQ_HW_QUEUE_MIN)
3246 static bool cfq_should_wait_busy(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3248 struct cfq_io_context *cic = cfqd->active_cic;
3250 /* If there are other queues in the group, don't wait */
3251 if (cfqq->cfqg->nr_cfqq > 1)
3254 if (cfq_slice_used(cfqq))
3257 /* if slice left is less than think time, wait busy */
3258 if (cic && sample_valid(cic->ttime_samples)
3259 && (cfqq->slice_end - jiffies < cic->ttime_mean))
3263 * If think times is less than a jiffy than ttime_mean=0 and above
3264 * will not be true. It might happen that slice has not expired yet
3265 * but will expire soon (4-5 ns) during select_queue(). To cover the
3266 * case where think time is less than a jiffy, mark the queue wait
3267 * busy if only 1 jiffy is left in the slice.
3269 if (cfqq->slice_end - jiffies == 1)
3275 static void cfq_completed_request(struct request_queue *q, struct request *rq)
3277 struct cfq_queue *cfqq = RQ_CFQQ(rq);
3278 struct cfq_data *cfqd = cfqq->cfqd;
3279 const int sync = rq_is_sync(rq);
3283 cfq_log_cfqq(cfqd, cfqq, "complete rqnoidle %d", !!rq_noidle(rq));
3285 cfq_update_hw_tag(cfqd);
3287 WARN_ON(!cfqd->rq_in_driver[sync]);
3288 WARN_ON(!cfqq->dispatched);
3289 cfqd->rq_in_driver[sync]--;
3292 if (cfq_cfqq_sync(cfqq))
3293 cfqd->sync_flight--;
3296 RQ_CIC(rq)->last_end_request = now;
3297 if (!time_after(rq->start_time + cfqd->cfq_fifo_expire[1], now))
3298 cfqd->last_delayed_sync = now;
3302 * If this is the active queue, check if it needs to be expired,
3303 * or if we want to idle in case it has no pending requests.
3305 if (cfqd->active_queue == cfqq) {
3306 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
3308 if (cfq_cfqq_slice_new(cfqq)) {
3309 cfq_set_prio_slice(cfqd, cfqq);
3310 cfq_clear_cfqq_slice_new(cfqq);
3314 * Should we wait for next request to come in before we expire
3317 if (cfq_should_wait_busy(cfqd, cfqq)) {
3318 cfqq->slice_end = jiffies + cfqd->cfq_slice_idle;
3319 cfq_mark_cfqq_wait_busy(cfqq);
3323 * Idling is not enabled on:
3325 * - idle-priority queues
3327 * - queues with still some requests queued
3328 * - when there is a close cooperator
3330 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
3331 cfq_slice_expired(cfqd, 1);
3332 else if (sync && cfqq_empty &&
3333 !cfq_close_cooperator(cfqd, cfqq)) {
3334 cfqd->noidle_tree_requires_idle |= !rq_noidle(rq);
3336 * Idling is enabled for SYNC_WORKLOAD.
3337 * SYNC_NOIDLE_WORKLOAD idles at the end of the tree
3338 * only if we processed at least one !rq_noidle request
3340 if (cfqd->serving_type == SYNC_WORKLOAD
3341 || cfqd->noidle_tree_requires_idle
3342 || cfqq->cfqg->nr_cfqq == 1)
3343 cfq_arm_slice_timer(cfqd);
3347 if (!rq_in_driver(cfqd))
3348 cfq_schedule_dispatch(cfqd);
3352 * we temporarily boost lower priority queues if they are holding fs exclusive
3353 * resources. they are boosted to normal prio (CLASS_BE/4)
3355 static void cfq_prio_boost(struct cfq_queue *cfqq)
3357 if (has_fs_excl()) {
3359 * boost idle prio on transactions that would lock out other
3360 * users of the filesystem
3362 if (cfq_class_idle(cfqq))
3363 cfqq->ioprio_class = IOPRIO_CLASS_BE;
3364 if (cfqq->ioprio > IOPRIO_NORM)
3365 cfqq->ioprio = IOPRIO_NORM;
3368 * unboost the queue (if needed)
3370 cfqq->ioprio_class = cfqq->org_ioprio_class;
3371 cfqq->ioprio = cfqq->org_ioprio;
3375 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
3377 if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
3378 cfq_mark_cfqq_must_alloc_slice(cfqq);
3379 return ELV_MQUEUE_MUST;
3382 return ELV_MQUEUE_MAY;
3385 static int cfq_may_queue(struct request_queue *q, int rw)
3387 struct cfq_data *cfqd = q->elevator->elevator_data;
3388 struct task_struct *tsk = current;
3389 struct cfq_io_context *cic;
3390 struct cfq_queue *cfqq;
3393 * don't force setup of a queue from here, as a call to may_queue
3394 * does not necessarily imply that a request actually will be queued.
3395 * so just lookup a possibly existing queue, or return 'may queue'
3398 cic = cfq_cic_lookup(cfqd, tsk->io_context);
3400 return ELV_MQUEUE_MAY;
3402 cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
3404 cfq_init_prio_data(cfqq, cic->ioc);
3405 cfq_prio_boost(cfqq);
3407 return __cfq_may_queue(cfqq);
3410 return ELV_MQUEUE_MAY;
3414 * queue lock held here
3416 static void cfq_put_request(struct request *rq)
3418 struct cfq_queue *cfqq = RQ_CFQQ(rq);
3421 const int rw = rq_data_dir(rq);
3423 BUG_ON(!cfqq->allocated[rw]);
3424 cfqq->allocated[rw]--;
3426 put_io_context(RQ_CIC(rq)->ioc);
3428 rq->elevator_private = NULL;
3429 rq->elevator_private2 = NULL;
3431 cfq_put_queue(cfqq);
3435 static struct cfq_queue *
3436 cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_context *cic,
3437 struct cfq_queue *cfqq)
3439 cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq);
3440 cic_set_cfqq(cic, cfqq->new_cfqq, 1);
3441 cfq_mark_cfqq_coop(cfqq->new_cfqq);
3442 cfq_put_queue(cfqq);
3443 return cic_to_cfqq(cic, 1);
3446 static int should_split_cfqq(struct cfq_queue *cfqq)
3448 if (cfqq->seeky_start &&
3449 time_after(jiffies, cfqq->seeky_start + CFQQ_COOP_TOUT))
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->seeky_start = 0;
3463 cfqq->pid = current->pid;
3464 cfq_clear_cfqq_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) && should_split_cfqq(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");