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/slab.h>
11 #include <linux/blkdev.h>
12 #include <linux/elevator.h>
13 #include <linux/jiffies.h>
14 #include <linux/rbtree.h>
15 #include <linux/ioprio.h>
16 #include <linux/blktrace_api.h>
18 #include "blk-cgroup.h"
23 /* max queue in one round of service */
24 static const int cfq_quantum = 8;
25 static const int cfq_fifo_expire[2] = { HZ / 4, HZ / 8 };
26 /* maximum backwards seek, in KiB */
27 static const int cfq_back_max = 16 * 1024;
28 /* penalty of a backwards seek */
29 static const int cfq_back_penalty = 2;
30 static const int cfq_slice_sync = HZ / 10;
31 static int cfq_slice_async = HZ / 25;
32 static const int cfq_slice_async_rq = 2;
33 static int cfq_slice_idle = HZ / 125;
34 static int cfq_group_idle = HZ / 125;
35 static const int cfq_target_latency = HZ * 3/10; /* 300 ms */
36 static const int cfq_hist_divisor = 4;
39 * offset from end of service tree
41 #define CFQ_IDLE_DELAY (HZ / 5)
44 * below this threshold, we consider thinktime immediate
46 #define CFQ_MIN_TT (2)
48 #define CFQ_SLICE_SCALE (5)
49 #define CFQ_HW_QUEUE_MIN (5)
50 #define CFQ_SERVICE_SHIFT 12
52 #define CFQQ_SEEK_THR (sector_t)(8 * 100)
53 #define CFQQ_CLOSE_THR (sector_t)(8 * 1024)
54 #define CFQQ_SECT_THR_NONROT (sector_t)(2 * 32)
55 #define CFQQ_SEEKY(cfqq) (hweight32(cfqq->seek_history) > 32/8)
57 #define RQ_CIC(rq) icq_to_cic((rq)->elv.icq)
58 #define RQ_CFQQ(rq) (struct cfq_queue *) ((rq)->elv.priv[0])
59 #define RQ_CFQG(rq) (struct cfq_group *) ((rq)->elv.priv[1])
61 static struct kmem_cache *cfq_pool;
63 #define CFQ_PRIO_LISTS IOPRIO_BE_NR
64 #define cfq_class_idle(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
65 #define cfq_class_rt(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
67 #define sample_valid(samples) ((samples) > 80)
68 #define rb_entry_cfqg(node) rb_entry((node), struct cfq_group, rb_node)
71 unsigned long last_end_request;
73 unsigned long ttime_total;
74 unsigned long ttime_samples;
75 unsigned long ttime_mean;
79 * Most of our rbtree usage is for sorting with min extraction, so
80 * if we cache the leftmost node we don't have to walk down the tree
81 * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should
82 * move this into the elevator for the rq sorting as well.
88 unsigned total_weight;
90 struct cfq_ttime ttime;
92 #define CFQ_RB_ROOT (struct cfq_rb_root) { .rb = RB_ROOT, \
93 .ttime = {.last_end_request = jiffies,},}
96 * Per process-grouping structure
101 /* various state flags, see below */
103 /* parent cfq_data */
104 struct cfq_data *cfqd;
105 /* service_tree member */
106 struct rb_node rb_node;
107 /* service_tree key */
108 unsigned long rb_key;
109 /* prio tree member */
110 struct rb_node p_node;
111 /* prio tree root we belong to, if any */
112 struct rb_root *p_root;
113 /* sorted list of pending requests */
114 struct rb_root sort_list;
115 /* if fifo isn't expired, next request to serve */
116 struct request *next_rq;
117 /* requests queued in sort_list */
119 /* currently allocated requests */
121 /* fifo list of requests in sort_list */
122 struct list_head fifo;
124 /* time when queue got scheduled in to dispatch first request. */
125 unsigned long dispatch_start;
126 unsigned int allocated_slice;
127 unsigned int slice_dispatch;
128 /* time when first request from queue completed and slice started. */
129 unsigned long slice_start;
130 unsigned long slice_end;
133 /* pending priority requests */
135 /* number of requests that are on the dispatch list or inside driver */
138 /* io prio of this group */
139 unsigned short ioprio, org_ioprio;
140 unsigned short ioprio_class;
145 sector_t last_request_pos;
147 struct cfq_rb_root *service_tree;
148 struct cfq_queue *new_cfqq;
149 struct cfq_group *cfqg;
150 /* Number of sectors dispatched from queue in single dispatch round */
151 unsigned long nr_sectors;
155 * First index in the service_trees.
156 * IDLE is handled separately, so it has negative index
166 * Second index in the service_trees.
170 SYNC_NOIDLE_WORKLOAD = 1,
175 #ifdef CONFIG_CFQ_GROUP_IOSCHED
176 /* total bytes transferred */
177 struct blkg_rwstat service_bytes;
178 /* total IOs serviced, post merge */
179 struct blkg_rwstat serviced;
180 /* number of ios merged */
181 struct blkg_rwstat merged;
182 /* total time spent on device in ns, may not be accurate w/ queueing */
183 struct blkg_rwstat service_time;
184 /* total time spent waiting in scheduler queue in ns */
185 struct blkg_rwstat wait_time;
186 /* number of IOs queued up */
187 struct blkg_rwstat queued;
188 /* total sectors transferred */
189 struct blkg_stat sectors;
190 /* total disk time and nr sectors dispatched by this group */
191 struct blkg_stat time;
192 #ifdef CONFIG_DEBUG_BLK_CGROUP
193 /* time not charged to this cgroup */
194 struct blkg_stat unaccounted_time;
195 /* sum of number of ios queued across all samples */
196 struct blkg_stat avg_queue_size_sum;
197 /* count of samples taken for average */
198 struct blkg_stat avg_queue_size_samples;
199 /* how many times this group has been removed from service tree */
200 struct blkg_stat dequeue;
201 /* total time spent waiting for it to be assigned a timeslice. */
202 struct blkg_stat group_wait_time;
203 /* time spent idling for this blkcg_gq */
204 struct blkg_stat idle_time;
205 /* total time with empty current active q with other requests queued */
206 struct blkg_stat empty_time;
207 /* fields after this shouldn't be cleared on stat reset */
208 uint64_t start_group_wait_time;
209 uint64_t start_idle_time;
210 uint64_t start_empty_time;
212 #endif /* CONFIG_DEBUG_BLK_CGROUP */
213 #endif /* CONFIG_CFQ_GROUP_IOSCHED */
216 /* This is per cgroup per device grouping structure */
218 /* must be the first member */
219 struct blkg_policy_data pd;
221 /* group service_tree member */
222 struct rb_node rb_node;
224 /* group service_tree key */
228 * There are two weights - (internal) weight is the weight of this
229 * cfqg against the sibling cfqgs. leaf_weight is the wight of
230 * this cfqg against the child cfqgs. For the root cfqg, both
231 * weights are kept in sync for backward compatibility.
234 unsigned int new_weight;
235 unsigned int dev_weight;
237 unsigned int leaf_weight;
238 unsigned int new_leaf_weight;
239 unsigned int dev_leaf_weight;
241 /* number of cfqq currently on this group */
245 * Per group busy queues average. Useful for workload slice calc. We
246 * create the array for each prio class but at run time it is used
247 * only for RT and BE class and slot for IDLE class remains unused.
248 * This is primarily done to avoid confusion and a gcc warning.
250 unsigned int busy_queues_avg[CFQ_PRIO_NR];
252 * rr lists of queues with requests. We maintain service trees for
253 * RT and BE classes. These trees are subdivided in subclasses
254 * of SYNC, SYNC_NOIDLE and ASYNC based on workload type. For IDLE
255 * class there is no subclassification and all the cfq queues go on
256 * a single tree service_tree_idle.
257 * Counts are embedded in the cfq_rb_root
259 struct cfq_rb_root service_trees[2][3];
260 struct cfq_rb_root service_tree_idle;
262 unsigned long saved_wl_slice;
263 enum wl_type_t saved_wl_type;
264 enum wl_class_t saved_wl_class;
266 /* number of requests that are on the dispatch list or inside driver */
268 struct cfq_ttime ttime;
269 struct cfqg_stats stats;
273 struct io_cq icq; /* must be the first member */
274 struct cfq_queue *cfqq[2];
275 struct cfq_ttime ttime;
276 int ioprio; /* the current ioprio */
277 #ifdef CONFIG_CFQ_GROUP_IOSCHED
278 uint64_t blkcg_id; /* the current blkcg ID */
283 * Per block device queue structure
286 struct request_queue *queue;
287 /* Root service tree for cfq_groups */
288 struct cfq_rb_root grp_service_tree;
289 struct cfq_group *root_group;
292 * The priority currently being served
294 enum wl_class_t serving_wl_class;
295 enum wl_type_t serving_wl_type;
296 unsigned long workload_expires;
297 struct cfq_group *serving_group;
300 * Each priority tree is sorted by next_request position. These
301 * trees are used when determining if two or more queues are
302 * interleaving requests (see cfq_close_cooperator).
304 struct rb_root prio_trees[CFQ_PRIO_LISTS];
306 unsigned int busy_queues;
307 unsigned int busy_sync_queues;
313 * queue-depth detection
319 * -1 => indeterminate, (cfq will behave as if NCQ is present, to allow better detection)
320 * 1 => NCQ is present (hw_tag_est_depth is the estimated max depth)
323 int hw_tag_est_depth;
324 unsigned int hw_tag_samples;
327 * idle window management
329 struct timer_list idle_slice_timer;
330 struct work_struct unplug_work;
332 struct cfq_queue *active_queue;
333 struct cfq_io_cq *active_cic;
336 * async queue for each priority case
338 struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
339 struct cfq_queue *async_idle_cfqq;
341 sector_t last_position;
344 * tunables, see top of file
346 unsigned int cfq_quantum;
347 unsigned int cfq_fifo_expire[2];
348 unsigned int cfq_back_penalty;
349 unsigned int cfq_back_max;
350 unsigned int cfq_slice[2];
351 unsigned int cfq_slice_async_rq;
352 unsigned int cfq_slice_idle;
353 unsigned int cfq_group_idle;
354 unsigned int cfq_latency;
355 unsigned int cfq_target_latency;
358 * Fallback dummy cfqq for extreme OOM conditions
360 struct cfq_queue oom_cfqq;
362 unsigned long last_delayed_sync;
365 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd);
367 static struct cfq_rb_root *st_for(struct cfq_group *cfqg,
368 enum wl_class_t class,
374 if (class == IDLE_WORKLOAD)
375 return &cfqg->service_tree_idle;
377 return &cfqg->service_trees[class][type];
380 enum cfqq_state_flags {
381 CFQ_CFQQ_FLAG_on_rr = 0, /* on round-robin busy list */
382 CFQ_CFQQ_FLAG_wait_request, /* waiting for a request */
383 CFQ_CFQQ_FLAG_must_dispatch, /* must be allowed a dispatch */
384 CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
385 CFQ_CFQQ_FLAG_fifo_expire, /* FIFO checked in this slice */
386 CFQ_CFQQ_FLAG_idle_window, /* slice idling enabled */
387 CFQ_CFQQ_FLAG_prio_changed, /* task priority has changed */
388 CFQ_CFQQ_FLAG_slice_new, /* no requests dispatched in slice */
389 CFQ_CFQQ_FLAG_sync, /* synchronous queue */
390 CFQ_CFQQ_FLAG_coop, /* cfqq is shared */
391 CFQ_CFQQ_FLAG_split_coop, /* shared cfqq will be splitted */
392 CFQ_CFQQ_FLAG_deep, /* sync cfqq experienced large depth */
393 CFQ_CFQQ_FLAG_wait_busy, /* Waiting for next request */
396 #define CFQ_CFQQ_FNS(name) \
397 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \
399 (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name); \
401 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \
403 (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \
405 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \
407 return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \
411 CFQ_CFQQ_FNS(wait_request);
412 CFQ_CFQQ_FNS(must_dispatch);
413 CFQ_CFQQ_FNS(must_alloc_slice);
414 CFQ_CFQQ_FNS(fifo_expire);
415 CFQ_CFQQ_FNS(idle_window);
416 CFQ_CFQQ_FNS(prio_changed);
417 CFQ_CFQQ_FNS(slice_new);
420 CFQ_CFQQ_FNS(split_coop);
422 CFQ_CFQQ_FNS(wait_busy);
425 static inline struct cfq_group *pd_to_cfqg(struct blkg_policy_data *pd)
427 return pd ? container_of(pd, struct cfq_group, pd) : NULL;
430 static inline struct blkcg_gq *cfqg_to_blkg(struct cfq_group *cfqg)
432 return pd_to_blkg(&cfqg->pd);
435 #if defined(CONFIG_CFQ_GROUP_IOSCHED) && defined(CONFIG_DEBUG_BLK_CGROUP)
437 /* cfqg stats flags */
438 enum cfqg_stats_flags {
439 CFQG_stats_waiting = 0,
444 #define CFQG_FLAG_FNS(name) \
445 static inline void cfqg_stats_mark_##name(struct cfqg_stats *stats) \
447 stats->flags |= (1 << CFQG_stats_##name); \
449 static inline void cfqg_stats_clear_##name(struct cfqg_stats *stats) \
451 stats->flags &= ~(1 << CFQG_stats_##name); \
453 static inline int cfqg_stats_##name(struct cfqg_stats *stats) \
455 return (stats->flags & (1 << CFQG_stats_##name)) != 0; \
458 CFQG_FLAG_FNS(waiting)
459 CFQG_FLAG_FNS(idling)
463 /* This should be called with the queue_lock held. */
464 static void cfqg_stats_update_group_wait_time(struct cfqg_stats *stats)
466 unsigned long long now;
468 if (!cfqg_stats_waiting(stats))
472 if (time_after64(now, stats->start_group_wait_time))
473 blkg_stat_add(&stats->group_wait_time,
474 now - stats->start_group_wait_time);
475 cfqg_stats_clear_waiting(stats);
478 /* This should be called with the queue_lock held. */
479 static void cfqg_stats_set_start_group_wait_time(struct cfq_group *cfqg,
480 struct cfq_group *curr_cfqg)
482 struct cfqg_stats *stats = &cfqg->stats;
484 if (cfqg_stats_waiting(stats))
486 if (cfqg == curr_cfqg)
488 stats->start_group_wait_time = sched_clock();
489 cfqg_stats_mark_waiting(stats);
492 /* This should be called with the queue_lock held. */
493 static void cfqg_stats_end_empty_time(struct cfqg_stats *stats)
495 unsigned long long now;
497 if (!cfqg_stats_empty(stats))
501 if (time_after64(now, stats->start_empty_time))
502 blkg_stat_add(&stats->empty_time,
503 now - stats->start_empty_time);
504 cfqg_stats_clear_empty(stats);
507 static void cfqg_stats_update_dequeue(struct cfq_group *cfqg)
509 blkg_stat_add(&cfqg->stats.dequeue, 1);
512 static void cfqg_stats_set_start_empty_time(struct cfq_group *cfqg)
514 struct cfqg_stats *stats = &cfqg->stats;
516 if (blkg_rwstat_sum(&stats->queued))
520 * group is already marked empty. This can happen if cfqq got new
521 * request in parent group and moved to this group while being added
522 * to service tree. Just ignore the event and move on.
524 if (cfqg_stats_empty(stats))
527 stats->start_empty_time = sched_clock();
528 cfqg_stats_mark_empty(stats);
531 static void cfqg_stats_update_idle_time(struct cfq_group *cfqg)
533 struct cfqg_stats *stats = &cfqg->stats;
535 if (cfqg_stats_idling(stats)) {
536 unsigned long long now = sched_clock();
538 if (time_after64(now, stats->start_idle_time))
539 blkg_stat_add(&stats->idle_time,
540 now - stats->start_idle_time);
541 cfqg_stats_clear_idling(stats);
545 static void cfqg_stats_set_start_idle_time(struct cfq_group *cfqg)
547 struct cfqg_stats *stats = &cfqg->stats;
549 BUG_ON(cfqg_stats_idling(stats));
551 stats->start_idle_time = sched_clock();
552 cfqg_stats_mark_idling(stats);
555 static void cfqg_stats_update_avg_queue_size(struct cfq_group *cfqg)
557 struct cfqg_stats *stats = &cfqg->stats;
559 blkg_stat_add(&stats->avg_queue_size_sum,
560 blkg_rwstat_sum(&stats->queued));
561 blkg_stat_add(&stats->avg_queue_size_samples, 1);
562 cfqg_stats_update_group_wait_time(stats);
565 #else /* CONFIG_CFQ_GROUP_IOSCHED && CONFIG_DEBUG_BLK_CGROUP */
567 static inline void cfqg_stats_set_start_group_wait_time(struct cfq_group *cfqg, struct cfq_group *curr_cfqg) { }
568 static inline void cfqg_stats_end_empty_time(struct cfqg_stats *stats) { }
569 static inline void cfqg_stats_update_dequeue(struct cfq_group *cfqg) { }
570 static inline void cfqg_stats_set_start_empty_time(struct cfq_group *cfqg) { }
571 static inline void cfqg_stats_update_idle_time(struct cfq_group *cfqg) { }
572 static inline void cfqg_stats_set_start_idle_time(struct cfq_group *cfqg) { }
573 static inline void cfqg_stats_update_avg_queue_size(struct cfq_group *cfqg) { }
575 #endif /* CONFIG_CFQ_GROUP_IOSCHED && CONFIG_DEBUG_BLK_CGROUP */
577 #ifdef CONFIG_CFQ_GROUP_IOSCHED
579 static struct blkcg_policy blkcg_policy_cfq;
581 static inline struct cfq_group *blkg_to_cfqg(struct blkcg_gq *blkg)
583 return pd_to_cfqg(blkg_to_pd(blkg, &blkcg_policy_cfq));
586 static inline void cfqg_get(struct cfq_group *cfqg)
588 return blkg_get(cfqg_to_blkg(cfqg));
591 static inline void cfqg_put(struct cfq_group *cfqg)
593 return blkg_put(cfqg_to_blkg(cfqg));
596 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...) do { \
599 blkg_path(cfqg_to_blkg((cfqq)->cfqg), __pbuf, sizeof(__pbuf)); \
600 blk_add_trace_msg((cfqd)->queue, "cfq%d%c%c %s " fmt, (cfqq)->pid, \
601 cfq_cfqq_sync((cfqq)) ? 'S' : 'A', \
602 cfqq_type((cfqq)) == SYNC_NOIDLE_WORKLOAD ? 'N' : ' ',\
606 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...) do { \
609 blkg_path(cfqg_to_blkg(cfqg), __pbuf, sizeof(__pbuf)); \
610 blk_add_trace_msg((cfqd)->queue, "%s " fmt, __pbuf, ##args); \
613 static inline void cfqg_stats_update_io_add(struct cfq_group *cfqg,
614 struct cfq_group *curr_cfqg, int rw)
616 blkg_rwstat_add(&cfqg->stats.queued, rw, 1);
617 cfqg_stats_end_empty_time(&cfqg->stats);
618 cfqg_stats_set_start_group_wait_time(cfqg, curr_cfqg);
621 static inline void cfqg_stats_update_timeslice_used(struct cfq_group *cfqg,
622 unsigned long time, unsigned long unaccounted_time)
624 blkg_stat_add(&cfqg->stats.time, time);
625 #ifdef CONFIG_DEBUG_BLK_CGROUP
626 blkg_stat_add(&cfqg->stats.unaccounted_time, unaccounted_time);
630 static inline void cfqg_stats_update_io_remove(struct cfq_group *cfqg, int rw)
632 blkg_rwstat_add(&cfqg->stats.queued, rw, -1);
635 static inline void cfqg_stats_update_io_merged(struct cfq_group *cfqg, int rw)
637 blkg_rwstat_add(&cfqg->stats.merged, rw, 1);
640 static inline void cfqg_stats_update_dispatch(struct cfq_group *cfqg,
641 uint64_t bytes, int rw)
643 blkg_stat_add(&cfqg->stats.sectors, bytes >> 9);
644 blkg_rwstat_add(&cfqg->stats.serviced, rw, 1);
645 blkg_rwstat_add(&cfqg->stats.service_bytes, rw, bytes);
648 static inline void cfqg_stats_update_completion(struct cfq_group *cfqg,
649 uint64_t start_time, uint64_t io_start_time, int rw)
651 struct cfqg_stats *stats = &cfqg->stats;
652 unsigned long long now = sched_clock();
654 if (time_after64(now, io_start_time))
655 blkg_rwstat_add(&stats->service_time, rw, now - io_start_time);
656 if (time_after64(io_start_time, start_time))
657 blkg_rwstat_add(&stats->wait_time, rw,
658 io_start_time - start_time);
661 static void cfq_pd_reset_stats(struct blkcg_gq *blkg)
663 struct cfq_group *cfqg = blkg_to_cfqg(blkg);
664 struct cfqg_stats *stats = &cfqg->stats;
666 /* queued stats shouldn't be cleared */
667 blkg_rwstat_reset(&stats->service_bytes);
668 blkg_rwstat_reset(&stats->serviced);
669 blkg_rwstat_reset(&stats->merged);
670 blkg_rwstat_reset(&stats->service_time);
671 blkg_rwstat_reset(&stats->wait_time);
672 blkg_stat_reset(&stats->time);
673 #ifdef CONFIG_DEBUG_BLK_CGROUP
674 blkg_stat_reset(&stats->unaccounted_time);
675 blkg_stat_reset(&stats->avg_queue_size_sum);
676 blkg_stat_reset(&stats->avg_queue_size_samples);
677 blkg_stat_reset(&stats->dequeue);
678 blkg_stat_reset(&stats->group_wait_time);
679 blkg_stat_reset(&stats->idle_time);
680 blkg_stat_reset(&stats->empty_time);
684 #else /* CONFIG_CFQ_GROUP_IOSCHED */
686 static inline void cfqg_get(struct cfq_group *cfqg) { }
687 static inline void cfqg_put(struct cfq_group *cfqg) { }
689 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
690 blk_add_trace_msg((cfqd)->queue, "cfq%d%c%c " fmt, (cfqq)->pid, \
691 cfq_cfqq_sync((cfqq)) ? 'S' : 'A', \
692 cfqq_type((cfqq)) == SYNC_NOIDLE_WORKLOAD ? 'N' : ' ',\
694 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...) do {} while (0)
696 static inline void cfqg_stats_update_io_add(struct cfq_group *cfqg,
697 struct cfq_group *curr_cfqg, int rw) { }
698 static inline void cfqg_stats_update_timeslice_used(struct cfq_group *cfqg,
699 unsigned long time, unsigned long unaccounted_time) { }
700 static inline void cfqg_stats_update_io_remove(struct cfq_group *cfqg, int rw) { }
701 static inline void cfqg_stats_update_io_merged(struct cfq_group *cfqg, int rw) { }
702 static inline void cfqg_stats_update_dispatch(struct cfq_group *cfqg,
703 uint64_t bytes, int rw) { }
704 static inline void cfqg_stats_update_completion(struct cfq_group *cfqg,
705 uint64_t start_time, uint64_t io_start_time, int rw) { }
707 #endif /* CONFIG_CFQ_GROUP_IOSCHED */
709 #define cfq_log(cfqd, fmt, args...) \
710 blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
712 /* Traverses through cfq group service trees */
713 #define for_each_cfqg_st(cfqg, i, j, st) \
714 for (i = 0; i <= IDLE_WORKLOAD; i++) \
715 for (j = 0, st = i < IDLE_WORKLOAD ? &cfqg->service_trees[i][j]\
716 : &cfqg->service_tree_idle; \
717 (i < IDLE_WORKLOAD && j <= SYNC_WORKLOAD) || \
718 (i == IDLE_WORKLOAD && j == 0); \
719 j++, st = i < IDLE_WORKLOAD ? \
720 &cfqg->service_trees[i][j]: NULL) \
722 static inline bool cfq_io_thinktime_big(struct cfq_data *cfqd,
723 struct cfq_ttime *ttime, bool group_idle)
726 if (!sample_valid(ttime->ttime_samples))
729 slice = cfqd->cfq_group_idle;
731 slice = cfqd->cfq_slice_idle;
732 return ttime->ttime_mean > slice;
735 static inline bool iops_mode(struct cfq_data *cfqd)
738 * If we are not idling on queues and it is a NCQ drive, parallel
739 * execution of requests is on and measuring time is not possible
740 * in most of the cases until and unless we drive shallower queue
741 * depths and that becomes a performance bottleneck. In such cases
742 * switch to start providing fairness in terms of number of IOs.
744 if (!cfqd->cfq_slice_idle && cfqd->hw_tag)
750 static inline enum wl_class_t cfqq_class(struct cfq_queue *cfqq)
752 if (cfq_class_idle(cfqq))
753 return IDLE_WORKLOAD;
754 if (cfq_class_rt(cfqq))
760 static enum wl_type_t cfqq_type(struct cfq_queue *cfqq)
762 if (!cfq_cfqq_sync(cfqq))
763 return ASYNC_WORKLOAD;
764 if (!cfq_cfqq_idle_window(cfqq))
765 return SYNC_NOIDLE_WORKLOAD;
766 return SYNC_WORKLOAD;
769 static inline int cfq_group_busy_queues_wl(enum wl_class_t wl_class,
770 struct cfq_data *cfqd,
771 struct cfq_group *cfqg)
773 if (wl_class == IDLE_WORKLOAD)
774 return cfqg->service_tree_idle.count;
776 return cfqg->service_trees[wl_class][ASYNC_WORKLOAD].count +
777 cfqg->service_trees[wl_class][SYNC_NOIDLE_WORKLOAD].count +
778 cfqg->service_trees[wl_class][SYNC_WORKLOAD].count;
781 static inline int cfqg_busy_async_queues(struct cfq_data *cfqd,
782 struct cfq_group *cfqg)
784 return cfqg->service_trees[RT_WORKLOAD][ASYNC_WORKLOAD].count +
785 cfqg->service_trees[BE_WORKLOAD][ASYNC_WORKLOAD].count;
788 static void cfq_dispatch_insert(struct request_queue *, struct request *);
789 static struct cfq_queue *cfq_get_queue(struct cfq_data *cfqd, bool is_sync,
790 struct cfq_io_cq *cic, struct bio *bio,
793 static inline struct cfq_io_cq *icq_to_cic(struct io_cq *icq)
795 /* cic->icq is the first member, %NULL will convert to %NULL */
796 return container_of(icq, struct cfq_io_cq, icq);
799 static inline struct cfq_io_cq *cfq_cic_lookup(struct cfq_data *cfqd,
800 struct io_context *ioc)
803 return icq_to_cic(ioc_lookup_icq(ioc, cfqd->queue));
807 static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_cq *cic, bool is_sync)
809 return cic->cfqq[is_sync];
812 static inline void cic_set_cfqq(struct cfq_io_cq *cic, struct cfq_queue *cfqq,
815 cic->cfqq[is_sync] = cfqq;
818 static inline struct cfq_data *cic_to_cfqd(struct cfq_io_cq *cic)
820 return cic->icq.q->elevator->elevator_data;
824 * We regard a request as SYNC, if it's either a read or has the SYNC bit
825 * set (in which case it could also be direct WRITE).
827 static inline bool cfq_bio_sync(struct bio *bio)
829 return bio_data_dir(bio) == READ || (bio->bi_rw & REQ_SYNC);
833 * scheduler run of queue, if there are requests pending and no one in the
834 * driver that will restart queueing
836 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
838 if (cfqd->busy_queues) {
839 cfq_log(cfqd, "schedule dispatch");
840 kblockd_schedule_work(cfqd->queue, &cfqd->unplug_work);
845 * Scale schedule slice based on io priority. Use the sync time slice only
846 * if a queue is marked sync and has sync io queued. A sync queue with async
847 * io only, should not get full sync slice length.
849 static inline int cfq_prio_slice(struct cfq_data *cfqd, bool sync,
852 const int base_slice = cfqd->cfq_slice[sync];
854 WARN_ON(prio >= IOPRIO_BE_NR);
856 return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
860 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
862 return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
865 static inline u64 cfq_scale_slice(unsigned long delta, struct cfq_group *cfqg)
867 u64 d = delta << CFQ_SERVICE_SHIFT;
869 d = d * CFQ_WEIGHT_DEFAULT;
870 do_div(d, cfqg->weight);
874 static inline u64 max_vdisktime(u64 min_vdisktime, u64 vdisktime)
876 s64 delta = (s64)(vdisktime - min_vdisktime);
878 min_vdisktime = vdisktime;
880 return min_vdisktime;
883 static inline u64 min_vdisktime(u64 min_vdisktime, u64 vdisktime)
885 s64 delta = (s64)(vdisktime - min_vdisktime);
887 min_vdisktime = vdisktime;
889 return min_vdisktime;
892 static void update_min_vdisktime(struct cfq_rb_root *st)
894 struct cfq_group *cfqg;
897 cfqg = rb_entry_cfqg(st->left);
898 st->min_vdisktime = max_vdisktime(st->min_vdisktime,
904 * get averaged number of queues of RT/BE priority.
905 * average is updated, with a formula that gives more weight to higher numbers,
906 * to quickly follows sudden increases and decrease slowly
909 static inline unsigned cfq_group_get_avg_queues(struct cfq_data *cfqd,
910 struct cfq_group *cfqg, bool rt)
912 unsigned min_q, max_q;
913 unsigned mult = cfq_hist_divisor - 1;
914 unsigned round = cfq_hist_divisor / 2;
915 unsigned busy = cfq_group_busy_queues_wl(rt, cfqd, cfqg);
917 min_q = min(cfqg->busy_queues_avg[rt], busy);
918 max_q = max(cfqg->busy_queues_avg[rt], busy);
919 cfqg->busy_queues_avg[rt] = (mult * max_q + min_q + round) /
921 return cfqg->busy_queues_avg[rt];
924 static inline unsigned
925 cfq_group_slice(struct cfq_data *cfqd, struct cfq_group *cfqg)
927 struct cfq_rb_root *st = &cfqd->grp_service_tree;
929 return cfqd->cfq_target_latency * cfqg->weight / st->total_weight;
932 static inline unsigned
933 cfq_scaled_cfqq_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
935 unsigned slice = cfq_prio_to_slice(cfqd, cfqq);
936 if (cfqd->cfq_latency) {
938 * interested queues (we consider only the ones with the same
939 * priority class in the cfq group)
941 unsigned iq = cfq_group_get_avg_queues(cfqd, cfqq->cfqg,
943 unsigned sync_slice = cfqd->cfq_slice[1];
944 unsigned expect_latency = sync_slice * iq;
945 unsigned group_slice = cfq_group_slice(cfqd, cfqq->cfqg);
947 if (expect_latency > group_slice) {
948 unsigned base_low_slice = 2 * cfqd->cfq_slice_idle;
949 /* scale low_slice according to IO priority
950 * and sync vs async */
952 min(slice, base_low_slice * slice / sync_slice);
953 /* the adapted slice value is scaled to fit all iqs
954 * into the target latency */
955 slice = max(slice * group_slice / expect_latency,
963 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
965 unsigned slice = cfq_scaled_cfqq_slice(cfqd, cfqq);
967 cfqq->slice_start = jiffies;
968 cfqq->slice_end = jiffies + slice;
969 cfqq->allocated_slice = slice;
970 cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
974 * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
975 * isn't valid until the first request from the dispatch is activated
976 * and the slice time set.
978 static inline bool cfq_slice_used(struct cfq_queue *cfqq)
980 if (cfq_cfqq_slice_new(cfqq))
982 if (time_before(jiffies, cfqq->slice_end))
989 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
990 * We choose the request that is closest to the head right now. Distance
991 * behind the head is penalized and only allowed to a certain extent.
993 static struct request *
994 cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2, sector_t last)
996 sector_t s1, s2, d1 = 0, d2 = 0;
997 unsigned long back_max;
998 #define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */
999 #define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */
1000 unsigned wrap = 0; /* bit mask: requests behind the disk head? */
1002 if (rq1 == NULL || rq1 == rq2)
1007 if (rq_is_sync(rq1) != rq_is_sync(rq2))
1008 return rq_is_sync(rq1) ? rq1 : rq2;
1010 if ((rq1->cmd_flags ^ rq2->cmd_flags) & REQ_PRIO)
1011 return rq1->cmd_flags & REQ_PRIO ? rq1 : rq2;
1013 s1 = blk_rq_pos(rq1);
1014 s2 = blk_rq_pos(rq2);
1017 * by definition, 1KiB is 2 sectors
1019 back_max = cfqd->cfq_back_max * 2;
1022 * Strict one way elevator _except_ in the case where we allow
1023 * short backward seeks which are biased as twice the cost of a
1024 * similar forward seek.
1028 else if (s1 + back_max >= last)
1029 d1 = (last - s1) * cfqd->cfq_back_penalty;
1031 wrap |= CFQ_RQ1_WRAP;
1035 else if (s2 + back_max >= last)
1036 d2 = (last - s2) * cfqd->cfq_back_penalty;
1038 wrap |= CFQ_RQ2_WRAP;
1040 /* Found required data */
1043 * By doing switch() on the bit mask "wrap" we avoid having to
1044 * check two variables for all permutations: --> faster!
1047 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
1063 case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
1066 * Since both rqs are wrapped,
1067 * start with the one that's further behind head
1068 * (--> only *one* back seek required),
1069 * since back seek takes more time than forward.
1079 * The below is leftmost cache rbtree addon
1081 static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
1083 /* Service tree is empty */
1088 root->left = rb_first(&root->rb);
1091 return rb_entry(root->left, struct cfq_queue, rb_node);
1096 static struct cfq_group *cfq_rb_first_group(struct cfq_rb_root *root)
1099 root->left = rb_first(&root->rb);
1102 return rb_entry_cfqg(root->left);
1107 static void rb_erase_init(struct rb_node *n, struct rb_root *root)
1113 static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
1115 if (root->left == n)
1117 rb_erase_init(n, &root->rb);
1122 * would be nice to take fifo expire time into account as well
1124 static struct request *
1125 cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1126 struct request *last)
1128 struct rb_node *rbnext = rb_next(&last->rb_node);
1129 struct rb_node *rbprev = rb_prev(&last->rb_node);
1130 struct request *next = NULL, *prev = NULL;
1132 BUG_ON(RB_EMPTY_NODE(&last->rb_node));
1135 prev = rb_entry_rq(rbprev);
1138 next = rb_entry_rq(rbnext);
1140 rbnext = rb_first(&cfqq->sort_list);
1141 if (rbnext && rbnext != &last->rb_node)
1142 next = rb_entry_rq(rbnext);
1145 return cfq_choose_req(cfqd, next, prev, blk_rq_pos(last));
1148 static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
1149 struct cfq_queue *cfqq)
1152 * just an approximation, should be ok.
1154 return (cfqq->cfqg->nr_cfqq - 1) * (cfq_prio_slice(cfqd, 1, 0) -
1155 cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
1159 cfqg_key(struct cfq_rb_root *st, struct cfq_group *cfqg)
1161 return cfqg->vdisktime - st->min_vdisktime;
1165 __cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
1167 struct rb_node **node = &st->rb.rb_node;
1168 struct rb_node *parent = NULL;
1169 struct cfq_group *__cfqg;
1170 s64 key = cfqg_key(st, cfqg);
1173 while (*node != NULL) {
1175 __cfqg = rb_entry_cfqg(parent);
1177 if (key < cfqg_key(st, __cfqg))
1178 node = &parent->rb_left;
1180 node = &parent->rb_right;
1186 st->left = &cfqg->rb_node;
1188 rb_link_node(&cfqg->rb_node, parent, node);
1189 rb_insert_color(&cfqg->rb_node, &st->rb);
1193 cfq_update_group_weight(struct cfq_group *cfqg)
1195 BUG_ON(!RB_EMPTY_NODE(&cfqg->rb_node));
1197 if (cfqg->new_weight) {
1198 cfqg->weight = cfqg->new_weight;
1199 cfqg->new_weight = 0;
1202 if (cfqg->new_leaf_weight) {
1203 cfqg->leaf_weight = cfqg->new_leaf_weight;
1204 cfqg->new_leaf_weight = 0;
1209 cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
1211 BUG_ON(!RB_EMPTY_NODE(&cfqg->rb_node));
1213 cfq_update_group_weight(cfqg);
1214 __cfq_group_service_tree_add(st, cfqg);
1215 st->total_weight += cfqg->weight;
1219 cfq_group_notify_queue_add(struct cfq_data *cfqd, struct cfq_group *cfqg)
1221 struct cfq_rb_root *st = &cfqd->grp_service_tree;
1222 struct cfq_group *__cfqg;
1226 if (!RB_EMPTY_NODE(&cfqg->rb_node))
1230 * Currently put the group at the end. Later implement something
1231 * so that groups get lesser vtime based on their weights, so that
1232 * if group does not loose all if it was not continuously backlogged.
1234 n = rb_last(&st->rb);
1236 __cfqg = rb_entry_cfqg(n);
1237 cfqg->vdisktime = __cfqg->vdisktime + CFQ_IDLE_DELAY;
1239 cfqg->vdisktime = st->min_vdisktime;
1240 cfq_group_service_tree_add(st, cfqg);
1244 cfq_group_service_tree_del(struct cfq_rb_root *st, struct cfq_group *cfqg)
1246 st->total_weight -= cfqg->weight;
1247 if (!RB_EMPTY_NODE(&cfqg->rb_node))
1248 cfq_rb_erase(&cfqg->rb_node, st);
1252 cfq_group_notify_queue_del(struct cfq_data *cfqd, struct cfq_group *cfqg)
1254 struct cfq_rb_root *st = &cfqd->grp_service_tree;
1256 BUG_ON(cfqg->nr_cfqq < 1);
1259 /* If there are other cfq queues under this group, don't delete it */
1263 cfq_log_cfqg(cfqd, cfqg, "del_from_rr group");
1264 cfq_group_service_tree_del(st, cfqg);
1265 cfqg->saved_wl_slice = 0;
1266 cfqg_stats_update_dequeue(cfqg);
1269 static inline unsigned int cfq_cfqq_slice_usage(struct cfq_queue *cfqq,
1270 unsigned int *unaccounted_time)
1272 unsigned int slice_used;
1275 * Queue got expired before even a single request completed or
1276 * got expired immediately after first request completion.
1278 if (!cfqq->slice_start || cfqq->slice_start == jiffies) {
1280 * Also charge the seek time incurred to the group, otherwise
1281 * if there are mutiple queues in the group, each can dispatch
1282 * a single request on seeky media and cause lots of seek time
1283 * and group will never know it.
1285 slice_used = max_t(unsigned, (jiffies - cfqq->dispatch_start),
1288 slice_used = jiffies - cfqq->slice_start;
1289 if (slice_used > cfqq->allocated_slice) {
1290 *unaccounted_time = slice_used - cfqq->allocated_slice;
1291 slice_used = cfqq->allocated_slice;
1293 if (time_after(cfqq->slice_start, cfqq->dispatch_start))
1294 *unaccounted_time += cfqq->slice_start -
1295 cfqq->dispatch_start;
1301 static void cfq_group_served(struct cfq_data *cfqd, struct cfq_group *cfqg,
1302 struct cfq_queue *cfqq)
1304 struct cfq_rb_root *st = &cfqd->grp_service_tree;
1305 unsigned int used_sl, charge, unaccounted_sl = 0;
1306 int nr_sync = cfqg->nr_cfqq - cfqg_busy_async_queues(cfqd, cfqg)
1307 - cfqg->service_tree_idle.count;
1309 BUG_ON(nr_sync < 0);
1310 used_sl = charge = cfq_cfqq_slice_usage(cfqq, &unaccounted_sl);
1312 if (iops_mode(cfqd))
1313 charge = cfqq->slice_dispatch;
1314 else if (!cfq_cfqq_sync(cfqq) && !nr_sync)
1315 charge = cfqq->allocated_slice;
1317 /* Can't update vdisktime while group is on service tree */
1318 cfq_group_service_tree_del(st, cfqg);
1319 cfqg->vdisktime += cfq_scale_slice(charge, cfqg);
1320 /* If a new weight was requested, update now, off tree */
1321 cfq_group_service_tree_add(st, cfqg);
1323 /* This group is being expired. Save the context */
1324 if (time_after(cfqd->workload_expires, jiffies)) {
1325 cfqg->saved_wl_slice = cfqd->workload_expires
1327 cfqg->saved_wl_type = cfqd->serving_wl_type;
1328 cfqg->saved_wl_class = cfqd->serving_wl_class;
1330 cfqg->saved_wl_slice = 0;
1332 cfq_log_cfqg(cfqd, cfqg, "served: vt=%llu min_vt=%llu", cfqg->vdisktime,
1334 cfq_log_cfqq(cfqq->cfqd, cfqq,
1335 "sl_used=%u disp=%u charge=%u iops=%u sect=%lu",
1336 used_sl, cfqq->slice_dispatch, charge,
1337 iops_mode(cfqd), cfqq->nr_sectors);
1338 cfqg_stats_update_timeslice_used(cfqg, used_sl, unaccounted_sl);
1339 cfqg_stats_set_start_empty_time(cfqg);
1343 * cfq_init_cfqg_base - initialize base part of a cfq_group
1344 * @cfqg: cfq_group to initialize
1346 * Initialize the base part which is used whether %CONFIG_CFQ_GROUP_IOSCHED
1347 * is enabled or not.
1349 static void cfq_init_cfqg_base(struct cfq_group *cfqg)
1351 struct cfq_rb_root *st;
1354 for_each_cfqg_st(cfqg, i, j, st)
1356 RB_CLEAR_NODE(&cfqg->rb_node);
1358 cfqg->ttime.last_end_request = jiffies;
1361 #ifdef CONFIG_CFQ_GROUP_IOSCHED
1362 static void cfq_pd_init(struct blkcg_gq *blkg)
1364 struct cfq_group *cfqg = blkg_to_cfqg(blkg);
1366 cfq_init_cfqg_base(cfqg);
1367 cfqg->weight = blkg->blkcg->cfq_weight;
1368 cfqg->leaf_weight = blkg->blkcg->cfq_leaf_weight;
1372 * Search for the cfq group current task belongs to. request_queue lock must
1375 static struct cfq_group *cfq_lookup_create_cfqg(struct cfq_data *cfqd,
1376 struct blkcg *blkcg)
1378 struct request_queue *q = cfqd->queue;
1379 struct cfq_group *cfqg = NULL;
1381 /* avoid lookup for the common case where there's no blkcg */
1382 if (blkcg == &blkcg_root) {
1383 cfqg = cfqd->root_group;
1385 struct blkcg_gq *blkg;
1387 blkg = blkg_lookup_create(blkcg, q);
1389 cfqg = blkg_to_cfqg(blkg);
1395 static void cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg)
1397 /* Currently, all async queues are mapped to root group */
1398 if (!cfq_cfqq_sync(cfqq))
1399 cfqg = cfqq->cfqd->root_group;
1402 /* cfqq reference on cfqg */
1406 static u64 cfqg_prfill_weight_device(struct seq_file *sf,
1407 struct blkg_policy_data *pd, int off)
1409 struct cfq_group *cfqg = pd_to_cfqg(pd);
1411 if (!cfqg->dev_weight)
1413 return __blkg_prfill_u64(sf, pd, cfqg->dev_weight);
1416 static int cfqg_print_weight_device(struct cgroup *cgrp, struct cftype *cft,
1417 struct seq_file *sf)
1419 blkcg_print_blkgs(sf, cgroup_to_blkcg(cgrp),
1420 cfqg_prfill_weight_device, &blkcg_policy_cfq, 0,
1425 static u64 cfqg_prfill_leaf_weight_device(struct seq_file *sf,
1426 struct blkg_policy_data *pd, int off)
1428 struct cfq_group *cfqg = pd_to_cfqg(pd);
1430 if (!cfqg->dev_leaf_weight)
1432 return __blkg_prfill_u64(sf, pd, cfqg->dev_leaf_weight);
1435 static int cfqg_print_leaf_weight_device(struct cgroup *cgrp,
1437 struct seq_file *sf)
1439 blkcg_print_blkgs(sf, cgroup_to_blkcg(cgrp),
1440 cfqg_prfill_leaf_weight_device, &blkcg_policy_cfq, 0,
1445 static int cfq_print_weight(struct cgroup *cgrp, struct cftype *cft,
1446 struct seq_file *sf)
1448 seq_printf(sf, "%u\n", cgroup_to_blkcg(cgrp)->cfq_weight);
1452 static int cfq_print_leaf_weight(struct cgroup *cgrp, struct cftype *cft,
1453 struct seq_file *sf)
1455 seq_printf(sf, "%u\n",
1456 cgroup_to_blkcg(cgrp)->cfq_leaf_weight);
1460 static int __cfqg_set_weight_device(struct cgroup *cgrp, struct cftype *cft,
1461 const char *buf, bool is_leaf_weight)
1463 struct blkcg *blkcg = cgroup_to_blkcg(cgrp);
1464 struct blkg_conf_ctx ctx;
1465 struct cfq_group *cfqg;
1468 ret = blkg_conf_prep(blkcg, &blkcg_policy_cfq, buf, &ctx);
1473 cfqg = blkg_to_cfqg(ctx.blkg);
1474 if (!ctx.v || (ctx.v >= CFQ_WEIGHT_MIN && ctx.v <= CFQ_WEIGHT_MAX)) {
1475 if (!is_leaf_weight) {
1476 cfqg->dev_weight = ctx.v;
1477 cfqg->new_weight = ctx.v ?: blkcg->cfq_weight;
1479 cfqg->dev_leaf_weight = ctx.v;
1480 cfqg->new_leaf_weight = ctx.v ?: blkcg->cfq_leaf_weight;
1485 blkg_conf_finish(&ctx);
1489 static int cfqg_set_weight_device(struct cgroup *cgrp, struct cftype *cft,
1492 return __cfqg_set_weight_device(cgrp, cft, buf, false);
1495 static int cfqg_set_leaf_weight_device(struct cgroup *cgrp, struct cftype *cft,
1498 return __cfqg_set_weight_device(cgrp, cft, buf, true);
1501 static int __cfq_set_weight(struct cgroup *cgrp, struct cftype *cft, u64 val,
1502 bool is_leaf_weight)
1504 struct blkcg *blkcg = cgroup_to_blkcg(cgrp);
1505 struct blkcg_gq *blkg;
1506 struct hlist_node *n;
1508 if (val < CFQ_WEIGHT_MIN || val > CFQ_WEIGHT_MAX)
1511 spin_lock_irq(&blkcg->lock);
1513 if (!is_leaf_weight)
1514 blkcg->cfq_weight = val;
1516 blkcg->cfq_leaf_weight = val;
1518 hlist_for_each_entry(blkg, n, &blkcg->blkg_list, blkcg_node) {
1519 struct cfq_group *cfqg = blkg_to_cfqg(blkg);
1524 if (!is_leaf_weight) {
1525 if (!cfqg->dev_weight)
1526 cfqg->new_weight = blkcg->cfq_weight;
1528 if (!cfqg->dev_leaf_weight)
1529 cfqg->new_leaf_weight = blkcg->cfq_leaf_weight;
1533 spin_unlock_irq(&blkcg->lock);
1537 static int cfq_set_weight(struct cgroup *cgrp, struct cftype *cft, u64 val)
1539 return __cfq_set_weight(cgrp, cft, val, false);
1542 static int cfq_set_leaf_weight(struct cgroup *cgrp, struct cftype *cft, u64 val)
1544 return __cfq_set_weight(cgrp, cft, val, true);
1547 static int cfqg_print_stat(struct cgroup *cgrp, struct cftype *cft,
1548 struct seq_file *sf)
1550 struct blkcg *blkcg = cgroup_to_blkcg(cgrp);
1552 blkcg_print_blkgs(sf, blkcg, blkg_prfill_stat, &blkcg_policy_cfq,
1553 cft->private, false);
1557 static int cfqg_print_rwstat(struct cgroup *cgrp, struct cftype *cft,
1558 struct seq_file *sf)
1560 struct blkcg *blkcg = cgroup_to_blkcg(cgrp);
1562 blkcg_print_blkgs(sf, blkcg, blkg_prfill_rwstat, &blkcg_policy_cfq,
1563 cft->private, true);
1567 #ifdef CONFIG_DEBUG_BLK_CGROUP
1568 static u64 cfqg_prfill_avg_queue_size(struct seq_file *sf,
1569 struct blkg_policy_data *pd, int off)
1571 struct cfq_group *cfqg = pd_to_cfqg(pd);
1572 u64 samples = blkg_stat_read(&cfqg->stats.avg_queue_size_samples);
1576 v = blkg_stat_read(&cfqg->stats.avg_queue_size_sum);
1579 __blkg_prfill_u64(sf, pd, v);
1583 /* print avg_queue_size */
1584 static int cfqg_print_avg_queue_size(struct cgroup *cgrp, struct cftype *cft,
1585 struct seq_file *sf)
1587 struct blkcg *blkcg = cgroup_to_blkcg(cgrp);
1589 blkcg_print_blkgs(sf, blkcg, cfqg_prfill_avg_queue_size,
1590 &blkcg_policy_cfq, 0, false);
1593 #endif /* CONFIG_DEBUG_BLK_CGROUP */
1595 static struct cftype cfq_blkcg_files[] = {
1597 .name = "weight_device",
1598 .read_seq_string = cfqg_print_weight_device,
1599 .write_string = cfqg_set_weight_device,
1600 .max_write_len = 256,
1604 .read_seq_string = cfq_print_weight,
1605 .write_u64 = cfq_set_weight,
1608 /* on root, leaf_weight is mapped to weight */
1610 .name = "leaf_weight_device",
1611 .flags = CFTYPE_ONLY_ON_ROOT,
1612 .read_seq_string = cfqg_print_weight_device,
1613 .write_string = cfqg_set_weight_device,
1614 .max_write_len = 256,
1617 .name = "leaf_weight",
1618 .flags = CFTYPE_ONLY_ON_ROOT,
1619 .read_seq_string = cfq_print_weight,
1620 .write_u64 = cfq_set_weight,
1623 /* no such mapping necessary for !roots */
1625 .name = "leaf_weight_device",
1626 .flags = CFTYPE_NOT_ON_ROOT,
1627 .read_seq_string = cfqg_print_leaf_weight_device,
1628 .write_string = cfqg_set_leaf_weight_device,
1629 .max_write_len = 256,
1632 .name = "leaf_weight",
1633 .flags = CFTYPE_NOT_ON_ROOT,
1634 .read_seq_string = cfq_print_leaf_weight,
1635 .write_u64 = cfq_set_leaf_weight,
1640 .private = offsetof(struct cfq_group, stats.time),
1641 .read_seq_string = cfqg_print_stat,
1645 .private = offsetof(struct cfq_group, stats.sectors),
1646 .read_seq_string = cfqg_print_stat,
1649 .name = "io_service_bytes",
1650 .private = offsetof(struct cfq_group, stats.service_bytes),
1651 .read_seq_string = cfqg_print_rwstat,
1654 .name = "io_serviced",
1655 .private = offsetof(struct cfq_group, stats.serviced),
1656 .read_seq_string = cfqg_print_rwstat,
1659 .name = "io_service_time",
1660 .private = offsetof(struct cfq_group, stats.service_time),
1661 .read_seq_string = cfqg_print_rwstat,
1664 .name = "io_wait_time",
1665 .private = offsetof(struct cfq_group, stats.wait_time),
1666 .read_seq_string = cfqg_print_rwstat,
1669 .name = "io_merged",
1670 .private = offsetof(struct cfq_group, stats.merged),
1671 .read_seq_string = cfqg_print_rwstat,
1674 .name = "io_queued",
1675 .private = offsetof(struct cfq_group, stats.queued),
1676 .read_seq_string = cfqg_print_rwstat,
1678 #ifdef CONFIG_DEBUG_BLK_CGROUP
1680 .name = "avg_queue_size",
1681 .read_seq_string = cfqg_print_avg_queue_size,
1684 .name = "group_wait_time",
1685 .private = offsetof(struct cfq_group, stats.group_wait_time),
1686 .read_seq_string = cfqg_print_stat,
1689 .name = "idle_time",
1690 .private = offsetof(struct cfq_group, stats.idle_time),
1691 .read_seq_string = cfqg_print_stat,
1694 .name = "empty_time",
1695 .private = offsetof(struct cfq_group, stats.empty_time),
1696 .read_seq_string = cfqg_print_stat,
1700 .private = offsetof(struct cfq_group, stats.dequeue),
1701 .read_seq_string = cfqg_print_stat,
1704 .name = "unaccounted_time",
1705 .private = offsetof(struct cfq_group, stats.unaccounted_time),
1706 .read_seq_string = cfqg_print_stat,
1708 #endif /* CONFIG_DEBUG_BLK_CGROUP */
1711 #else /* GROUP_IOSCHED */
1712 static struct cfq_group *cfq_lookup_create_cfqg(struct cfq_data *cfqd,
1713 struct blkcg *blkcg)
1715 return cfqd->root_group;
1719 cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) {
1723 #endif /* GROUP_IOSCHED */
1726 * The cfqd->service_trees holds all pending cfq_queue's that have
1727 * requests waiting to be processed. It is sorted in the order that
1728 * we will service the queues.
1730 static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1733 struct rb_node **p, *parent;
1734 struct cfq_queue *__cfqq;
1735 unsigned long rb_key;
1736 struct cfq_rb_root *st;
1740 st = st_for(cfqq->cfqg, cfqq_class(cfqq), cfqq_type(cfqq));
1741 if (cfq_class_idle(cfqq)) {
1742 rb_key = CFQ_IDLE_DELAY;
1743 parent = rb_last(&st->rb);
1744 if (parent && parent != &cfqq->rb_node) {
1745 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1746 rb_key += __cfqq->rb_key;
1749 } else if (!add_front) {
1751 * Get our rb key offset. Subtract any residual slice
1752 * value carried from last service. A negative resid
1753 * count indicates slice overrun, and this should position
1754 * the next service time further away in the tree.
1756 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
1757 rb_key -= cfqq->slice_resid;
1758 cfqq->slice_resid = 0;
1761 __cfqq = cfq_rb_first(st);
1762 rb_key += __cfqq ? __cfqq->rb_key : jiffies;
1765 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1768 * same position, nothing more to do
1770 if (rb_key == cfqq->rb_key && cfqq->service_tree == st)
1773 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1774 cfqq->service_tree = NULL;
1779 cfqq->service_tree = st;
1780 p = &st->rb.rb_node;
1783 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1786 * sort by key, that represents service time.
1788 if (time_before(rb_key, __cfqq->rb_key))
1789 p = &parent->rb_left;
1791 p = &parent->rb_right;
1797 st->left = &cfqq->rb_node;
1799 cfqq->rb_key = rb_key;
1800 rb_link_node(&cfqq->rb_node, parent, p);
1801 rb_insert_color(&cfqq->rb_node, &st->rb);
1803 if (add_front || !new_cfqq)
1805 cfq_group_notify_queue_add(cfqd, cfqq->cfqg);
1808 static struct cfq_queue *
1809 cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
1810 sector_t sector, struct rb_node **ret_parent,
1811 struct rb_node ***rb_link)
1813 struct rb_node **p, *parent;
1814 struct cfq_queue *cfqq = NULL;
1822 cfqq = rb_entry(parent, struct cfq_queue, p_node);
1825 * Sort strictly based on sector. Smallest to the left,
1826 * largest to the right.
1828 if (sector > blk_rq_pos(cfqq->next_rq))
1829 n = &(*p)->rb_right;
1830 else if (sector < blk_rq_pos(cfqq->next_rq))
1838 *ret_parent = parent;
1844 static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1846 struct rb_node **p, *parent;
1847 struct cfq_queue *__cfqq;
1850 rb_erase(&cfqq->p_node, cfqq->p_root);
1851 cfqq->p_root = NULL;
1854 if (cfq_class_idle(cfqq))
1859 cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
1860 __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
1861 blk_rq_pos(cfqq->next_rq), &parent, &p);
1863 rb_link_node(&cfqq->p_node, parent, p);
1864 rb_insert_color(&cfqq->p_node, cfqq->p_root);
1866 cfqq->p_root = NULL;
1870 * Update cfqq's position in the service tree.
1872 static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1875 * Resorting requires the cfqq to be on the RR list already.
1877 if (cfq_cfqq_on_rr(cfqq)) {
1878 cfq_service_tree_add(cfqd, cfqq, 0);
1879 cfq_prio_tree_add(cfqd, cfqq);
1884 * add to busy list of queues for service, trying to be fair in ordering
1885 * the pending list according to last request service
1887 static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1889 cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
1890 BUG_ON(cfq_cfqq_on_rr(cfqq));
1891 cfq_mark_cfqq_on_rr(cfqq);
1892 cfqd->busy_queues++;
1893 if (cfq_cfqq_sync(cfqq))
1894 cfqd->busy_sync_queues++;
1896 cfq_resort_rr_list(cfqd, cfqq);
1900 * Called when the cfqq no longer has requests pending, remove it from
1903 static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1905 cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
1906 BUG_ON(!cfq_cfqq_on_rr(cfqq));
1907 cfq_clear_cfqq_on_rr(cfqq);
1909 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1910 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1911 cfqq->service_tree = NULL;
1914 rb_erase(&cfqq->p_node, cfqq->p_root);
1915 cfqq->p_root = NULL;
1918 cfq_group_notify_queue_del(cfqd, cfqq->cfqg);
1919 BUG_ON(!cfqd->busy_queues);
1920 cfqd->busy_queues--;
1921 if (cfq_cfqq_sync(cfqq))
1922 cfqd->busy_sync_queues--;
1926 * rb tree support functions
1928 static void cfq_del_rq_rb(struct request *rq)
1930 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1931 const int sync = rq_is_sync(rq);
1933 BUG_ON(!cfqq->queued[sync]);
1934 cfqq->queued[sync]--;
1936 elv_rb_del(&cfqq->sort_list, rq);
1938 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) {
1940 * Queue will be deleted from service tree when we actually
1941 * expire it later. Right now just remove it from prio tree
1945 rb_erase(&cfqq->p_node, cfqq->p_root);
1946 cfqq->p_root = NULL;
1951 static void cfq_add_rq_rb(struct request *rq)
1953 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1954 struct cfq_data *cfqd = cfqq->cfqd;
1955 struct request *prev;
1957 cfqq->queued[rq_is_sync(rq)]++;
1959 elv_rb_add(&cfqq->sort_list, rq);
1961 if (!cfq_cfqq_on_rr(cfqq))
1962 cfq_add_cfqq_rr(cfqd, cfqq);
1965 * check if this request is a better next-serve candidate
1967 prev = cfqq->next_rq;
1968 cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq, cfqd->last_position);
1971 * adjust priority tree position, if ->next_rq changes
1973 if (prev != cfqq->next_rq)
1974 cfq_prio_tree_add(cfqd, cfqq);
1976 BUG_ON(!cfqq->next_rq);
1979 static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
1981 elv_rb_del(&cfqq->sort_list, rq);
1982 cfqq->queued[rq_is_sync(rq)]--;
1983 cfqg_stats_update_io_remove(RQ_CFQG(rq), rq->cmd_flags);
1985 cfqg_stats_update_io_add(RQ_CFQG(rq), cfqq->cfqd->serving_group,
1989 static struct request *
1990 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
1992 struct task_struct *tsk = current;
1993 struct cfq_io_cq *cic;
1994 struct cfq_queue *cfqq;
1996 cic = cfq_cic_lookup(cfqd, tsk->io_context);
2000 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
2002 sector_t sector = bio->bi_sector + bio_sectors(bio);
2004 return elv_rb_find(&cfqq->sort_list, sector);
2010 static void cfq_activate_request(struct request_queue *q, struct request *rq)
2012 struct cfq_data *cfqd = q->elevator->elevator_data;
2014 cfqd->rq_in_driver++;
2015 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
2016 cfqd->rq_in_driver);
2018 cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
2021 static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
2023 struct cfq_data *cfqd = q->elevator->elevator_data;
2025 WARN_ON(!cfqd->rq_in_driver);
2026 cfqd->rq_in_driver--;
2027 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
2028 cfqd->rq_in_driver);
2031 static void cfq_remove_request(struct request *rq)
2033 struct cfq_queue *cfqq = RQ_CFQQ(rq);
2035 if (cfqq->next_rq == rq)
2036 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
2038 list_del_init(&rq->queuelist);
2041 cfqq->cfqd->rq_queued--;
2042 cfqg_stats_update_io_remove(RQ_CFQG(rq), rq->cmd_flags);
2043 if (rq->cmd_flags & REQ_PRIO) {
2044 WARN_ON(!cfqq->prio_pending);
2045 cfqq->prio_pending--;
2049 static int cfq_merge(struct request_queue *q, struct request **req,
2052 struct cfq_data *cfqd = q->elevator->elevator_data;
2053 struct request *__rq;
2055 __rq = cfq_find_rq_fmerge(cfqd, bio);
2056 if (__rq && elv_rq_merge_ok(__rq, bio)) {
2058 return ELEVATOR_FRONT_MERGE;
2061 return ELEVATOR_NO_MERGE;
2064 static void cfq_merged_request(struct request_queue *q, struct request *req,
2067 if (type == ELEVATOR_FRONT_MERGE) {
2068 struct cfq_queue *cfqq = RQ_CFQQ(req);
2070 cfq_reposition_rq_rb(cfqq, req);
2074 static void cfq_bio_merged(struct request_queue *q, struct request *req,
2077 cfqg_stats_update_io_merged(RQ_CFQG(req), bio->bi_rw);
2081 cfq_merged_requests(struct request_queue *q, struct request *rq,
2082 struct request *next)
2084 struct cfq_queue *cfqq = RQ_CFQQ(rq);
2085 struct cfq_data *cfqd = q->elevator->elevator_data;
2088 * reposition in fifo if next is older than rq
2090 if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
2091 time_before(rq_fifo_time(next), rq_fifo_time(rq)) &&
2092 cfqq == RQ_CFQQ(next)) {
2093 list_move(&rq->queuelist, &next->queuelist);
2094 rq_set_fifo_time(rq, rq_fifo_time(next));
2097 if (cfqq->next_rq == next)
2099 cfq_remove_request(next);
2100 cfqg_stats_update_io_merged(RQ_CFQG(rq), next->cmd_flags);
2102 cfqq = RQ_CFQQ(next);
2104 * all requests of this queue are merged to other queues, delete it
2105 * from the service tree. If it's the active_queue,
2106 * cfq_dispatch_requests() will choose to expire it or do idle
2108 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list) &&
2109 cfqq != cfqd->active_queue)
2110 cfq_del_cfqq_rr(cfqd, cfqq);
2113 static int cfq_allow_merge(struct request_queue *q, struct request *rq,
2116 struct cfq_data *cfqd = q->elevator->elevator_data;
2117 struct cfq_io_cq *cic;
2118 struct cfq_queue *cfqq;
2121 * Disallow merge of a sync bio into an async request.
2123 if (cfq_bio_sync(bio) && !rq_is_sync(rq))
2127 * Lookup the cfqq that this bio will be queued with and allow
2128 * merge only if rq is queued there.
2130 cic = cfq_cic_lookup(cfqd, current->io_context);
2134 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
2135 return cfqq == RQ_CFQQ(rq);
2138 static inline void cfq_del_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2140 del_timer(&cfqd->idle_slice_timer);
2141 cfqg_stats_update_idle_time(cfqq->cfqg);
2144 static void __cfq_set_active_queue(struct cfq_data *cfqd,
2145 struct cfq_queue *cfqq)
2148 cfq_log_cfqq(cfqd, cfqq, "set_active wl_class:%d wl_type:%d",
2149 cfqd->serving_wl_class, cfqd->serving_wl_type);
2150 cfqg_stats_update_avg_queue_size(cfqq->cfqg);
2151 cfqq->slice_start = 0;
2152 cfqq->dispatch_start = jiffies;
2153 cfqq->allocated_slice = 0;
2154 cfqq->slice_end = 0;
2155 cfqq->slice_dispatch = 0;
2156 cfqq->nr_sectors = 0;
2158 cfq_clear_cfqq_wait_request(cfqq);
2159 cfq_clear_cfqq_must_dispatch(cfqq);
2160 cfq_clear_cfqq_must_alloc_slice(cfqq);
2161 cfq_clear_cfqq_fifo_expire(cfqq);
2162 cfq_mark_cfqq_slice_new(cfqq);
2164 cfq_del_timer(cfqd, cfqq);
2167 cfqd->active_queue = cfqq;
2171 * current cfqq expired its slice (or was too idle), select new one
2174 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2177 cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
2179 if (cfq_cfqq_wait_request(cfqq))
2180 cfq_del_timer(cfqd, cfqq);
2182 cfq_clear_cfqq_wait_request(cfqq);
2183 cfq_clear_cfqq_wait_busy(cfqq);
2186 * If this cfqq is shared between multiple processes, check to
2187 * make sure that those processes are still issuing I/Os within
2188 * the mean seek distance. If not, it may be time to break the
2189 * queues apart again.
2191 if (cfq_cfqq_coop(cfqq) && CFQQ_SEEKY(cfqq))
2192 cfq_mark_cfqq_split_coop(cfqq);
2195 * store what was left of this slice, if the queue idled/timed out
2198 if (cfq_cfqq_slice_new(cfqq))
2199 cfqq->slice_resid = cfq_scaled_cfqq_slice(cfqd, cfqq);
2201 cfqq->slice_resid = cfqq->slice_end - jiffies;
2202 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
2205 cfq_group_served(cfqd, cfqq->cfqg, cfqq);
2207 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
2208 cfq_del_cfqq_rr(cfqd, cfqq);
2210 cfq_resort_rr_list(cfqd, cfqq);
2212 if (cfqq == cfqd->active_queue)
2213 cfqd->active_queue = NULL;
2215 if (cfqd->active_cic) {
2216 put_io_context(cfqd->active_cic->icq.ioc);
2217 cfqd->active_cic = NULL;
2221 static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out)
2223 struct cfq_queue *cfqq = cfqd->active_queue;
2226 __cfq_slice_expired(cfqd, cfqq, timed_out);
2230 * Get next queue for service. Unless we have a queue preemption,
2231 * we'll simply select the first cfqq in the service tree.
2233 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
2235 struct cfq_rb_root *st = st_for(cfqd->serving_group,
2236 cfqd->serving_wl_class, cfqd->serving_wl_type);
2238 if (!cfqd->rq_queued)
2241 /* There is nothing to dispatch */
2244 if (RB_EMPTY_ROOT(&st->rb))
2246 return cfq_rb_first(st);
2249 static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd)
2251 struct cfq_group *cfqg;
2252 struct cfq_queue *cfqq;
2254 struct cfq_rb_root *st;
2256 if (!cfqd->rq_queued)
2259 cfqg = cfq_get_next_cfqg(cfqd);
2263 for_each_cfqg_st(cfqg, i, j, st)
2264 if ((cfqq = cfq_rb_first(st)) != NULL)
2270 * Get and set a new active queue for service.
2272 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
2273 struct cfq_queue *cfqq)
2276 cfqq = cfq_get_next_queue(cfqd);
2278 __cfq_set_active_queue(cfqd, cfqq);
2282 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
2285 if (blk_rq_pos(rq) >= cfqd->last_position)
2286 return blk_rq_pos(rq) - cfqd->last_position;
2288 return cfqd->last_position - blk_rq_pos(rq);
2291 static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2294 return cfq_dist_from_last(cfqd, rq) <= CFQQ_CLOSE_THR;
2297 static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
2298 struct cfq_queue *cur_cfqq)
2300 struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
2301 struct rb_node *parent, *node;
2302 struct cfq_queue *__cfqq;
2303 sector_t sector = cfqd->last_position;
2305 if (RB_EMPTY_ROOT(root))
2309 * First, if we find a request starting at the end of the last
2310 * request, choose it.
2312 __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
2317 * If the exact sector wasn't found, the parent of the NULL leaf
2318 * will contain the closest sector.
2320 __cfqq = rb_entry(parent, struct cfq_queue, p_node);
2321 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
2324 if (blk_rq_pos(__cfqq->next_rq) < sector)
2325 node = rb_next(&__cfqq->p_node);
2327 node = rb_prev(&__cfqq->p_node);
2331 __cfqq = rb_entry(node, struct cfq_queue, p_node);
2332 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
2340 * cur_cfqq - passed in so that we don't decide that the current queue is
2341 * closely cooperating with itself.
2343 * So, basically we're assuming that that cur_cfqq has dispatched at least
2344 * one request, and that cfqd->last_position reflects a position on the disk
2345 * associated with the I/O issued by cur_cfqq. I'm not sure this is a valid
2348 static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
2349 struct cfq_queue *cur_cfqq)
2351 struct cfq_queue *cfqq;
2353 if (cfq_class_idle(cur_cfqq))
2355 if (!cfq_cfqq_sync(cur_cfqq))
2357 if (CFQQ_SEEKY(cur_cfqq))
2361 * Don't search priority tree if it's the only queue in the group.
2363 if (cur_cfqq->cfqg->nr_cfqq == 1)
2367 * We should notice if some of the queues are cooperating, eg
2368 * working closely on the same area of the disk. In that case,
2369 * we can group them together and don't waste time idling.
2371 cfqq = cfqq_close(cfqd, cur_cfqq);
2375 /* If new queue belongs to different cfq_group, don't choose it */
2376 if (cur_cfqq->cfqg != cfqq->cfqg)
2380 * It only makes sense to merge sync queues.
2382 if (!cfq_cfqq_sync(cfqq))
2384 if (CFQQ_SEEKY(cfqq))
2388 * Do not merge queues of different priority classes
2390 if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq))
2397 * Determine whether we should enforce idle window for this queue.
2400 static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2402 enum wl_class_t wl_class = cfqq_class(cfqq);
2403 struct cfq_rb_root *st = cfqq->service_tree;
2408 if (!cfqd->cfq_slice_idle)
2411 /* We never do for idle class queues. */
2412 if (wl_class == IDLE_WORKLOAD)
2415 /* We do for queues that were marked with idle window flag. */
2416 if (cfq_cfqq_idle_window(cfqq) &&
2417 !(blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag))
2421 * Otherwise, we do only if they are the last ones
2422 * in their service tree.
2424 if (st->count == 1 && cfq_cfqq_sync(cfqq) &&
2425 !cfq_io_thinktime_big(cfqd, &st->ttime, false))
2427 cfq_log_cfqq(cfqd, cfqq, "Not idling. st->count:%d", st->count);
2431 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
2433 struct cfq_queue *cfqq = cfqd->active_queue;
2434 struct cfq_io_cq *cic;
2435 unsigned long sl, group_idle = 0;
2438 * SSD device without seek penalty, disable idling. But only do so
2439 * for devices that support queuing, otherwise we still have a problem
2440 * with sync vs async workloads.
2442 if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
2445 WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
2446 WARN_ON(cfq_cfqq_slice_new(cfqq));
2449 * idle is disabled, either manually or by past process history
2451 if (!cfq_should_idle(cfqd, cfqq)) {
2452 /* no queue idling. Check for group idling */
2453 if (cfqd->cfq_group_idle)
2454 group_idle = cfqd->cfq_group_idle;
2460 * still active requests from this queue, don't idle
2462 if (cfqq->dispatched)
2466 * task has exited, don't wait
2468 cic = cfqd->active_cic;
2469 if (!cic || !atomic_read(&cic->icq.ioc->active_ref))
2473 * If our average think time is larger than the remaining time
2474 * slice, then don't idle. This avoids overrunning the allotted
2477 if (sample_valid(cic->ttime.ttime_samples) &&
2478 (cfqq->slice_end - jiffies < cic->ttime.ttime_mean)) {
2479 cfq_log_cfqq(cfqd, cfqq, "Not idling. think_time:%lu",
2480 cic->ttime.ttime_mean);
2484 /* There are other queues in the group, don't do group idle */
2485 if (group_idle && cfqq->cfqg->nr_cfqq > 1)
2488 cfq_mark_cfqq_wait_request(cfqq);
2491 sl = cfqd->cfq_group_idle;
2493 sl = cfqd->cfq_slice_idle;
2495 mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
2496 cfqg_stats_set_start_idle_time(cfqq->cfqg);
2497 cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu group_idle: %d", sl,
2498 group_idle ? 1 : 0);
2502 * Move request from internal lists to the request queue dispatch list.
2504 static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
2506 struct cfq_data *cfqd = q->elevator->elevator_data;
2507 struct cfq_queue *cfqq = RQ_CFQQ(rq);
2509 cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
2511 cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
2512 cfq_remove_request(rq);
2514 (RQ_CFQG(rq))->dispatched++;
2515 elv_dispatch_sort(q, rq);
2517 cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]++;
2518 cfqq->nr_sectors += blk_rq_sectors(rq);
2519 cfqg_stats_update_dispatch(cfqq->cfqg, blk_rq_bytes(rq), rq->cmd_flags);
2523 * return expired entry, or NULL to just start from scratch in rbtree
2525 static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
2527 struct request *rq = NULL;
2529 if (cfq_cfqq_fifo_expire(cfqq))
2532 cfq_mark_cfqq_fifo_expire(cfqq);
2534 if (list_empty(&cfqq->fifo))
2537 rq = rq_entry_fifo(cfqq->fifo.next);
2538 if (time_before(jiffies, rq_fifo_time(rq)))
2541 cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq);
2546 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2548 const int base_rq = cfqd->cfq_slice_async_rq;
2550 WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
2552 return 2 * base_rq * (IOPRIO_BE_NR - cfqq->ioprio);
2556 * Must be called with the queue_lock held.
2558 static int cfqq_process_refs(struct cfq_queue *cfqq)
2560 int process_refs, io_refs;
2562 io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE];
2563 process_refs = cfqq->ref - io_refs;
2564 BUG_ON(process_refs < 0);
2565 return process_refs;
2568 static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq)
2570 int process_refs, new_process_refs;
2571 struct cfq_queue *__cfqq;
2574 * If there are no process references on the new_cfqq, then it is
2575 * unsafe to follow the ->new_cfqq chain as other cfqq's in the
2576 * chain may have dropped their last reference (not just their
2577 * last process reference).
2579 if (!cfqq_process_refs(new_cfqq))
2582 /* Avoid a circular list and skip interim queue merges */
2583 while ((__cfqq = new_cfqq->new_cfqq)) {
2589 process_refs = cfqq_process_refs(cfqq);
2590 new_process_refs = cfqq_process_refs(new_cfqq);
2592 * If the process for the cfqq has gone away, there is no
2593 * sense in merging the queues.
2595 if (process_refs == 0 || new_process_refs == 0)
2599 * Merge in the direction of the lesser amount of work.
2601 if (new_process_refs >= process_refs) {
2602 cfqq->new_cfqq = new_cfqq;
2603 new_cfqq->ref += process_refs;
2605 new_cfqq->new_cfqq = cfqq;
2606 cfqq->ref += new_process_refs;
2610 static enum wl_type_t cfq_choose_wl_type(struct cfq_data *cfqd,
2611 struct cfq_group *cfqg, enum wl_class_t wl_class)
2613 struct cfq_queue *queue;
2615 bool key_valid = false;
2616 unsigned long lowest_key = 0;
2617 enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD;
2619 for (i = 0; i <= SYNC_WORKLOAD; ++i) {
2620 /* select the one with lowest rb_key */
2621 queue = cfq_rb_first(st_for(cfqg, wl_class, i));
2623 (!key_valid || time_before(queue->rb_key, lowest_key))) {
2624 lowest_key = queue->rb_key;
2634 choose_wl_class_and_type(struct cfq_data *cfqd, struct cfq_group *cfqg)
2638 struct cfq_rb_root *st;
2639 unsigned group_slice;
2640 enum wl_class_t original_class = cfqd->serving_wl_class;
2642 /* Choose next priority. RT > BE > IDLE */
2643 if (cfq_group_busy_queues_wl(RT_WORKLOAD, cfqd, cfqg))
2644 cfqd->serving_wl_class = RT_WORKLOAD;
2645 else if (cfq_group_busy_queues_wl(BE_WORKLOAD, cfqd, cfqg))
2646 cfqd->serving_wl_class = BE_WORKLOAD;
2648 cfqd->serving_wl_class = IDLE_WORKLOAD;
2649 cfqd->workload_expires = jiffies + 1;
2653 if (original_class != cfqd->serving_wl_class)
2657 * For RT and BE, we have to choose also the type
2658 * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
2661 st = st_for(cfqg, cfqd->serving_wl_class, cfqd->serving_wl_type);
2665 * check workload expiration, and that we still have other queues ready
2667 if (count && !time_after(jiffies, cfqd->workload_expires))
2671 /* otherwise select new workload type */
2672 cfqd->serving_wl_type = cfq_choose_wl_type(cfqd, cfqg,
2673 cfqd->serving_wl_class);
2674 st = st_for(cfqg, cfqd->serving_wl_class, cfqd->serving_wl_type);
2678 * the workload slice is computed as a fraction of target latency
2679 * proportional to the number of queues in that workload, over
2680 * all the queues in the same priority class
2682 group_slice = cfq_group_slice(cfqd, cfqg);
2684 slice = group_slice * count /
2685 max_t(unsigned, cfqg->busy_queues_avg[cfqd->serving_wl_class],
2686 cfq_group_busy_queues_wl(cfqd->serving_wl_class, cfqd,
2689 if (cfqd->serving_wl_type == ASYNC_WORKLOAD) {
2693 * Async queues are currently system wide. Just taking
2694 * proportion of queues with-in same group will lead to higher
2695 * async ratio system wide as generally root group is going
2696 * to have higher weight. A more accurate thing would be to
2697 * calculate system wide asnc/sync ratio.
2699 tmp = cfqd->cfq_target_latency *
2700 cfqg_busy_async_queues(cfqd, cfqg);
2701 tmp = tmp/cfqd->busy_queues;
2702 slice = min_t(unsigned, slice, tmp);
2704 /* async workload slice is scaled down according to
2705 * the sync/async slice ratio. */
2706 slice = slice * cfqd->cfq_slice[0] / cfqd->cfq_slice[1];
2708 /* sync workload slice is at least 2 * cfq_slice_idle */
2709 slice = max(slice, 2 * cfqd->cfq_slice_idle);
2711 slice = max_t(unsigned, slice, CFQ_MIN_TT);
2712 cfq_log(cfqd, "workload slice:%d", slice);
2713 cfqd->workload_expires = jiffies + slice;
2716 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd)
2718 struct cfq_rb_root *st = &cfqd->grp_service_tree;
2719 struct cfq_group *cfqg;
2721 if (RB_EMPTY_ROOT(&st->rb))
2723 cfqg = cfq_rb_first_group(st);
2724 update_min_vdisktime(st);
2728 static void cfq_choose_cfqg(struct cfq_data *cfqd)
2730 struct cfq_group *cfqg = cfq_get_next_cfqg(cfqd);
2732 cfqd->serving_group = cfqg;
2734 /* Restore the workload type data */
2735 if (cfqg->saved_wl_slice) {
2736 cfqd->workload_expires = jiffies + cfqg->saved_wl_slice;
2737 cfqd->serving_wl_type = cfqg->saved_wl_type;
2738 cfqd->serving_wl_class = cfqg->saved_wl_class;
2740 cfqd->workload_expires = jiffies - 1;
2742 choose_wl_class_and_type(cfqd, cfqg);
2746 * Select a queue for service. If we have a current active queue,
2747 * check whether to continue servicing it, or retrieve and set a new one.
2749 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
2751 struct cfq_queue *cfqq, *new_cfqq = NULL;
2753 cfqq = cfqd->active_queue;
2757 if (!cfqd->rq_queued)
2761 * We were waiting for group to get backlogged. Expire the queue
2763 if (cfq_cfqq_wait_busy(cfqq) && !RB_EMPTY_ROOT(&cfqq->sort_list))
2767 * The active queue has run out of time, expire it and select new.
2769 if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq)) {
2771 * If slice had not expired at the completion of last request
2772 * we might not have turned on wait_busy flag. Don't expire
2773 * the queue yet. Allow the group to get backlogged.
2775 * The very fact that we have used the slice, that means we
2776 * have been idling all along on this queue and it should be
2777 * ok to wait for this request to complete.
2779 if (cfqq->cfqg->nr_cfqq == 1 && RB_EMPTY_ROOT(&cfqq->sort_list)
2780 && cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2784 goto check_group_idle;
2788 * The active queue has requests and isn't expired, allow it to
2791 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
2795 * If another queue has a request waiting within our mean seek
2796 * distance, let it run. The expire code will check for close
2797 * cooperators and put the close queue at the front of the service
2798 * tree. If possible, merge the expiring queue with the new cfqq.
2800 new_cfqq = cfq_close_cooperator(cfqd, cfqq);
2802 if (!cfqq->new_cfqq)
2803 cfq_setup_merge(cfqq, new_cfqq);
2808 * No requests pending. If the active queue still has requests in
2809 * flight or is idling for a new request, allow either of these
2810 * conditions to happen (or time out) before selecting a new queue.
2812 if (timer_pending(&cfqd->idle_slice_timer)) {
2818 * This is a deep seek queue, but the device is much faster than
2819 * the queue can deliver, don't idle
2821 if (CFQQ_SEEKY(cfqq) && cfq_cfqq_idle_window(cfqq) &&
2822 (cfq_cfqq_slice_new(cfqq) ||
2823 (cfqq->slice_end - jiffies > jiffies - cfqq->slice_start))) {
2824 cfq_clear_cfqq_deep(cfqq);
2825 cfq_clear_cfqq_idle_window(cfqq);
2828 if (cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2834 * If group idle is enabled and there are requests dispatched from
2835 * this group, wait for requests to complete.
2838 if (cfqd->cfq_group_idle && cfqq->cfqg->nr_cfqq == 1 &&
2839 cfqq->cfqg->dispatched &&
2840 !cfq_io_thinktime_big(cfqd, &cfqq->cfqg->ttime, true)) {
2846 cfq_slice_expired(cfqd, 0);
2849 * Current queue expired. Check if we have to switch to a new
2853 cfq_choose_cfqg(cfqd);
2855 cfqq = cfq_set_active_queue(cfqd, new_cfqq);
2860 static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
2864 while (cfqq->next_rq) {
2865 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
2869 BUG_ON(!list_empty(&cfqq->fifo));
2871 /* By default cfqq is not expired if it is empty. Do it explicitly */
2872 __cfq_slice_expired(cfqq->cfqd, cfqq, 0);
2877 * Drain our current requests. Used for barriers and when switching
2878 * io schedulers on-the-fly.
2880 static int cfq_forced_dispatch(struct cfq_data *cfqd)
2882 struct cfq_queue *cfqq;
2885 /* Expire the timeslice of the current active queue first */
2886 cfq_slice_expired(cfqd, 0);
2887 while ((cfqq = cfq_get_next_queue_forced(cfqd)) != NULL) {
2888 __cfq_set_active_queue(cfqd, cfqq);
2889 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
2892 BUG_ON(cfqd->busy_queues);
2894 cfq_log(cfqd, "forced_dispatch=%d", dispatched);
2898 static inline bool cfq_slice_used_soon(struct cfq_data *cfqd,
2899 struct cfq_queue *cfqq)
2901 /* the queue hasn't finished any request, can't estimate */
2902 if (cfq_cfqq_slice_new(cfqq))
2904 if (time_after(jiffies + cfqd->cfq_slice_idle * cfqq->dispatched,
2911 static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2913 unsigned int max_dispatch;
2916 * Drain async requests before we start sync IO
2918 if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_flight[BLK_RW_ASYNC])
2922 * If this is an async queue and we have sync IO in flight, let it wait
2924 if (cfqd->rq_in_flight[BLK_RW_SYNC] && !cfq_cfqq_sync(cfqq))
2927 max_dispatch = max_t(unsigned int, cfqd->cfq_quantum / 2, 1);
2928 if (cfq_class_idle(cfqq))
2932 * Does this cfqq already have too much IO in flight?
2934 if (cfqq->dispatched >= max_dispatch) {
2935 bool promote_sync = false;
2937 * idle queue must always only have a single IO in flight
2939 if (cfq_class_idle(cfqq))
2943 * If there is only one sync queue
2944 * we can ignore async queue here and give the sync
2945 * queue no dispatch limit. The reason is a sync queue can
2946 * preempt async queue, limiting the sync queue doesn't make
2947 * sense. This is useful for aiostress test.
2949 if (cfq_cfqq_sync(cfqq) && cfqd->busy_sync_queues == 1)
2950 promote_sync = true;
2953 * We have other queues, don't allow more IO from this one
2955 if (cfqd->busy_queues > 1 && cfq_slice_used_soon(cfqd, cfqq) &&
2960 * Sole queue user, no limit
2962 if (cfqd->busy_queues == 1 || promote_sync)
2966 * Normally we start throttling cfqq when cfq_quantum/2
2967 * requests have been dispatched. But we can drive
2968 * deeper queue depths at the beginning of slice
2969 * subjected to upper limit of cfq_quantum.
2971 max_dispatch = cfqd->cfq_quantum;
2975 * Async queues must wait a bit before being allowed dispatch.
2976 * We also ramp up the dispatch depth gradually for async IO,
2977 * based on the last sync IO we serviced
2979 if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
2980 unsigned long last_sync = jiffies - cfqd->last_delayed_sync;
2983 depth = last_sync / cfqd->cfq_slice[1];
2984 if (!depth && !cfqq->dispatched)
2986 if (depth < max_dispatch)
2987 max_dispatch = depth;
2991 * If we're below the current max, allow a dispatch
2993 return cfqq->dispatched < max_dispatch;
2997 * Dispatch a request from cfqq, moving them to the request queue
3000 static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3004 BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
3006 if (!cfq_may_dispatch(cfqd, cfqq))
3010 * follow expired path, else get first next available
3012 rq = cfq_check_fifo(cfqq);
3017 * insert request into driver dispatch list
3019 cfq_dispatch_insert(cfqd->queue, rq);
3021 if (!cfqd->active_cic) {
3022 struct cfq_io_cq *cic = RQ_CIC(rq);
3024 atomic_long_inc(&cic->icq.ioc->refcount);
3025 cfqd->active_cic = cic;
3032 * Find the cfqq that we need to service and move a request from that to the
3035 static int cfq_dispatch_requests(struct request_queue *q, int force)
3037 struct cfq_data *cfqd = q->elevator->elevator_data;
3038 struct cfq_queue *cfqq;
3040 if (!cfqd->busy_queues)
3043 if (unlikely(force))
3044 return cfq_forced_dispatch(cfqd);
3046 cfqq = cfq_select_queue(cfqd);
3051 * Dispatch a request from this cfqq, if it is allowed
3053 if (!cfq_dispatch_request(cfqd, cfqq))
3056 cfqq->slice_dispatch++;
3057 cfq_clear_cfqq_must_dispatch(cfqq);
3060 * expire an async queue immediately if it has used up its slice. idle
3061 * queue always expire after 1 dispatch round.
3063 if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
3064 cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
3065 cfq_class_idle(cfqq))) {
3066 cfqq->slice_end = jiffies + 1;
3067 cfq_slice_expired(cfqd, 0);
3070 cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
3075 * task holds one reference to the queue, dropped when task exits. each rq
3076 * in-flight on this queue also holds a reference, dropped when rq is freed.
3078 * Each cfq queue took a reference on the parent group. Drop it now.
3079 * queue lock must be held here.
3081 static void cfq_put_queue(struct cfq_queue *cfqq)
3083 struct cfq_data *cfqd = cfqq->cfqd;
3084 struct cfq_group *cfqg;
3086 BUG_ON(cfqq->ref <= 0);
3092 cfq_log_cfqq(cfqd, cfqq, "put_queue");
3093 BUG_ON(rb_first(&cfqq->sort_list));
3094 BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
3097 if (unlikely(cfqd->active_queue == cfqq)) {
3098 __cfq_slice_expired(cfqd, cfqq, 0);
3099 cfq_schedule_dispatch(cfqd);
3102 BUG_ON(cfq_cfqq_on_rr(cfqq));
3103 kmem_cache_free(cfq_pool, cfqq);
3107 static void cfq_put_cooperator(struct cfq_queue *cfqq)
3109 struct cfq_queue *__cfqq, *next;
3112 * If this queue was scheduled to merge with another queue, be
3113 * sure to drop the reference taken on that queue (and others in
3114 * the merge chain). See cfq_setup_merge and cfq_merge_cfqqs.
3116 __cfqq = cfqq->new_cfqq;
3118 if (__cfqq == cfqq) {
3119 WARN(1, "cfqq->new_cfqq loop detected\n");
3122 next = __cfqq->new_cfqq;
3123 cfq_put_queue(__cfqq);
3128 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3130 if (unlikely(cfqq == cfqd->active_queue)) {
3131 __cfq_slice_expired(cfqd, cfqq, 0);
3132 cfq_schedule_dispatch(cfqd);
3135 cfq_put_cooperator(cfqq);
3137 cfq_put_queue(cfqq);
3140 static void cfq_init_icq(struct io_cq *icq)
3142 struct cfq_io_cq *cic = icq_to_cic(icq);
3144 cic->ttime.last_end_request = jiffies;
3147 static void cfq_exit_icq(struct io_cq *icq)
3149 struct cfq_io_cq *cic = icq_to_cic(icq);
3150 struct cfq_data *cfqd = cic_to_cfqd(cic);
3152 if (cic->cfqq[BLK_RW_ASYNC]) {
3153 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]);
3154 cic->cfqq[BLK_RW_ASYNC] = NULL;
3157 if (cic->cfqq[BLK_RW_SYNC]) {
3158 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]);
3159 cic->cfqq[BLK_RW_SYNC] = NULL;
3163 static void cfq_init_prio_data(struct cfq_queue *cfqq, struct cfq_io_cq *cic)
3165 struct task_struct *tsk = current;
3168 if (!cfq_cfqq_prio_changed(cfqq))
3171 ioprio_class = IOPRIO_PRIO_CLASS(cic->ioprio);
3172 switch (ioprio_class) {
3174 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
3175 case IOPRIO_CLASS_NONE:
3177 * no prio set, inherit CPU scheduling settings
3179 cfqq->ioprio = task_nice_ioprio(tsk);
3180 cfqq->ioprio_class = task_nice_ioclass(tsk);
3182 case IOPRIO_CLASS_RT:
3183 cfqq->ioprio = IOPRIO_PRIO_DATA(cic->ioprio);
3184 cfqq->ioprio_class = IOPRIO_CLASS_RT;
3186 case IOPRIO_CLASS_BE:
3187 cfqq->ioprio = IOPRIO_PRIO_DATA(cic->ioprio);
3188 cfqq->ioprio_class = IOPRIO_CLASS_BE;
3190 case IOPRIO_CLASS_IDLE:
3191 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
3193 cfq_clear_cfqq_idle_window(cfqq);
3198 * keep track of original prio settings in case we have to temporarily
3199 * elevate the priority of this queue
3201 cfqq->org_ioprio = cfqq->ioprio;
3202 cfq_clear_cfqq_prio_changed(cfqq);
3205 static void check_ioprio_changed(struct cfq_io_cq *cic, struct bio *bio)
3207 int ioprio = cic->icq.ioc->ioprio;
3208 struct cfq_data *cfqd = cic_to_cfqd(cic);
3209 struct cfq_queue *cfqq;
3212 * Check whether ioprio has changed. The condition may trigger
3213 * spuriously on a newly created cic but there's no harm.
3215 if (unlikely(!cfqd) || likely(cic->ioprio == ioprio))
3218 cfqq = cic->cfqq[BLK_RW_ASYNC];
3220 struct cfq_queue *new_cfqq;
3221 new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic, bio,
3224 cic->cfqq[BLK_RW_ASYNC] = new_cfqq;
3225 cfq_put_queue(cfqq);
3229 cfqq = cic->cfqq[BLK_RW_SYNC];
3231 cfq_mark_cfqq_prio_changed(cfqq);
3233 cic->ioprio = ioprio;
3236 static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3237 pid_t pid, bool is_sync)
3239 RB_CLEAR_NODE(&cfqq->rb_node);
3240 RB_CLEAR_NODE(&cfqq->p_node);
3241 INIT_LIST_HEAD(&cfqq->fifo);
3246 cfq_mark_cfqq_prio_changed(cfqq);
3249 if (!cfq_class_idle(cfqq))
3250 cfq_mark_cfqq_idle_window(cfqq);
3251 cfq_mark_cfqq_sync(cfqq);
3256 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3257 static void check_blkcg_changed(struct cfq_io_cq *cic, struct bio *bio)
3259 struct cfq_data *cfqd = cic_to_cfqd(cic);
3260 struct cfq_queue *sync_cfqq;
3264 id = bio_blkcg(bio)->id;
3268 * Check whether blkcg has changed. The condition may trigger
3269 * spuriously on a newly created cic but there's no harm.
3271 if (unlikely(!cfqd) || likely(cic->blkcg_id == id))
3274 sync_cfqq = cic_to_cfqq(cic, 1);
3277 * Drop reference to sync queue. A new sync queue will be
3278 * assigned in new group upon arrival of a fresh request.
3280 cfq_log_cfqq(cfqd, sync_cfqq, "changed cgroup");
3281 cic_set_cfqq(cic, NULL, 1);
3282 cfq_put_queue(sync_cfqq);
3288 static inline void check_blkcg_changed(struct cfq_io_cq *cic, struct bio *bio) { }
3289 #endif /* CONFIG_CFQ_GROUP_IOSCHED */
3291 static struct cfq_queue *
3292 cfq_find_alloc_queue(struct cfq_data *cfqd, bool is_sync, struct cfq_io_cq *cic,
3293 struct bio *bio, gfp_t gfp_mask)
3295 struct blkcg *blkcg;
3296 struct cfq_queue *cfqq, *new_cfqq = NULL;
3297 struct cfq_group *cfqg;
3302 blkcg = bio_blkcg(bio);
3303 cfqg = cfq_lookup_create_cfqg(cfqd, blkcg);
3304 cfqq = cic_to_cfqq(cic, is_sync);
3307 * Always try a new alloc if we fell back to the OOM cfqq
3308 * originally, since it should just be a temporary situation.
3310 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
3315 } else if (gfp_mask & __GFP_WAIT) {
3317 spin_unlock_irq(cfqd->queue->queue_lock);
3318 new_cfqq = kmem_cache_alloc_node(cfq_pool,
3319 gfp_mask | __GFP_ZERO,
3321 spin_lock_irq(cfqd->queue->queue_lock);
3325 cfqq = kmem_cache_alloc_node(cfq_pool,
3326 gfp_mask | __GFP_ZERO,
3331 cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
3332 cfq_init_prio_data(cfqq, cic);
3333 cfq_link_cfqq_cfqg(cfqq, cfqg);
3334 cfq_log_cfqq(cfqd, cfqq, "alloced");
3336 cfqq = &cfqd->oom_cfqq;
3340 kmem_cache_free(cfq_pool, new_cfqq);
3346 static struct cfq_queue **
3347 cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
3349 switch (ioprio_class) {
3350 case IOPRIO_CLASS_RT:
3351 return &cfqd->async_cfqq[0][ioprio];
3352 case IOPRIO_CLASS_NONE:
3353 ioprio = IOPRIO_NORM;
3355 case IOPRIO_CLASS_BE:
3356 return &cfqd->async_cfqq[1][ioprio];
3357 case IOPRIO_CLASS_IDLE:
3358 return &cfqd->async_idle_cfqq;
3364 static struct cfq_queue *
3365 cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct cfq_io_cq *cic,
3366 struct bio *bio, gfp_t gfp_mask)
3368 const int ioprio_class = IOPRIO_PRIO_CLASS(cic->ioprio);
3369 const int ioprio = IOPRIO_PRIO_DATA(cic->ioprio);
3370 struct cfq_queue **async_cfqq = NULL;
3371 struct cfq_queue *cfqq = NULL;
3374 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
3379 cfqq = cfq_find_alloc_queue(cfqd, is_sync, cic, bio, gfp_mask);
3382 * pin the queue now that it's allocated, scheduler exit will prune it
3384 if (!is_sync && !(*async_cfqq)) {
3394 __cfq_update_io_thinktime(struct cfq_ttime *ttime, unsigned long slice_idle)
3396 unsigned long elapsed = jiffies - ttime->last_end_request;
3397 elapsed = min(elapsed, 2UL * slice_idle);
3399 ttime->ttime_samples = (7*ttime->ttime_samples + 256) / 8;
3400 ttime->ttime_total = (7*ttime->ttime_total + 256*elapsed) / 8;
3401 ttime->ttime_mean = (ttime->ttime_total + 128) / ttime->ttime_samples;
3405 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3406 struct cfq_io_cq *cic)
3408 if (cfq_cfqq_sync(cfqq)) {
3409 __cfq_update_io_thinktime(&cic->ttime, cfqd->cfq_slice_idle);
3410 __cfq_update_io_thinktime(&cfqq->service_tree->ttime,
3411 cfqd->cfq_slice_idle);
3413 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3414 __cfq_update_io_thinktime(&cfqq->cfqg->ttime, cfqd->cfq_group_idle);
3419 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3423 sector_t n_sec = blk_rq_sectors(rq);
3424 if (cfqq->last_request_pos) {
3425 if (cfqq->last_request_pos < blk_rq_pos(rq))
3426 sdist = blk_rq_pos(rq) - cfqq->last_request_pos;
3428 sdist = cfqq->last_request_pos - blk_rq_pos(rq);
3431 cfqq->seek_history <<= 1;
3432 if (blk_queue_nonrot(cfqd->queue))
3433 cfqq->seek_history |= (n_sec < CFQQ_SECT_THR_NONROT);
3435 cfqq->seek_history |= (sdist > CFQQ_SEEK_THR);
3439 * Disable idle window if the process thinks too long or seeks so much that
3443 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3444 struct cfq_io_cq *cic)
3446 int old_idle, enable_idle;
3449 * Don't idle for async or idle io prio class
3451 if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
3454 enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
3456 if (cfqq->queued[0] + cfqq->queued[1] >= 4)
3457 cfq_mark_cfqq_deep(cfqq);
3459 if (cfqq->next_rq && (cfqq->next_rq->cmd_flags & REQ_NOIDLE))
3461 else if (!atomic_read(&cic->icq.ioc->active_ref) ||
3462 !cfqd->cfq_slice_idle ||
3463 (!cfq_cfqq_deep(cfqq) && CFQQ_SEEKY(cfqq)))
3465 else if (sample_valid(cic->ttime.ttime_samples)) {
3466 if (cic->ttime.ttime_mean > cfqd->cfq_slice_idle)
3472 if (old_idle != enable_idle) {
3473 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
3475 cfq_mark_cfqq_idle_window(cfqq);
3477 cfq_clear_cfqq_idle_window(cfqq);
3482 * Check if new_cfqq should preempt the currently active queue. Return 0 for
3483 * no or if we aren't sure, a 1 will cause a preempt.
3486 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
3489 struct cfq_queue *cfqq;
3491 cfqq = cfqd->active_queue;
3495 if (cfq_class_idle(new_cfqq))
3498 if (cfq_class_idle(cfqq))
3502 * Don't allow a non-RT request to preempt an ongoing RT cfqq timeslice.
3504 if (cfq_class_rt(cfqq) && !cfq_class_rt(new_cfqq))
3508 * if the new request is sync, but the currently running queue is
3509 * not, let the sync request have priority.
3511 if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
3514 if (new_cfqq->cfqg != cfqq->cfqg)
3517 if (cfq_slice_used(cfqq))
3520 /* Allow preemption only if we are idling on sync-noidle tree */
3521 if (cfqd->serving_wl_type == SYNC_NOIDLE_WORKLOAD &&
3522 cfqq_type(new_cfqq) == SYNC_NOIDLE_WORKLOAD &&
3523 new_cfqq->service_tree->count == 2 &&
3524 RB_EMPTY_ROOT(&cfqq->sort_list))
3528 * So both queues are sync. Let the new request get disk time if
3529 * it's a metadata request and the current queue is doing regular IO.
3531 if ((rq->cmd_flags & REQ_PRIO) && !cfqq->prio_pending)
3535 * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
3537 if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
3540 /* An idle queue should not be idle now for some reason */
3541 if (RB_EMPTY_ROOT(&cfqq->sort_list) && !cfq_should_idle(cfqd, cfqq))
3544 if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
3548 * if this request is as-good as one we would expect from the
3549 * current cfqq, let it preempt
3551 if (cfq_rq_close(cfqd, cfqq, rq))
3558 * cfqq preempts the active queue. if we allowed preempt with no slice left,
3559 * let it have half of its nominal slice.
3561 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3563 enum wl_type_t old_type = cfqq_type(cfqd->active_queue);
3565 cfq_log_cfqq(cfqd, cfqq, "preempt");
3566 cfq_slice_expired(cfqd, 1);
3569 * workload type is changed, don't save slice, otherwise preempt
3572 if (old_type != cfqq_type(cfqq))
3573 cfqq->cfqg->saved_wl_slice = 0;
3576 * Put the new queue at the front of the of the current list,
3577 * so we know that it will be selected next.
3579 BUG_ON(!cfq_cfqq_on_rr(cfqq));
3581 cfq_service_tree_add(cfqd, cfqq, 1);
3583 cfqq->slice_end = 0;
3584 cfq_mark_cfqq_slice_new(cfqq);
3588 * Called when a new fs request (rq) is added (to cfqq). Check if there's
3589 * something we should do about it
3592 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3595 struct cfq_io_cq *cic = RQ_CIC(rq);
3598 if (rq->cmd_flags & REQ_PRIO)
3599 cfqq->prio_pending++;
3601 cfq_update_io_thinktime(cfqd, cfqq, cic);
3602 cfq_update_io_seektime(cfqd, cfqq, rq);
3603 cfq_update_idle_window(cfqd, cfqq, cic);
3605 cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
3607 if (cfqq == cfqd->active_queue) {
3609 * Remember that we saw a request from this process, but
3610 * don't start queuing just yet. Otherwise we risk seeing lots
3611 * of tiny requests, because we disrupt the normal plugging
3612 * and merging. If the request is already larger than a single
3613 * page, let it rip immediately. For that case we assume that
3614 * merging is already done. Ditto for a busy system that
3615 * has other work pending, don't risk delaying until the
3616 * idle timer unplug to continue working.
3618 if (cfq_cfqq_wait_request(cfqq)) {
3619 if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
3620 cfqd->busy_queues > 1) {
3621 cfq_del_timer(cfqd, cfqq);
3622 cfq_clear_cfqq_wait_request(cfqq);
3623 __blk_run_queue(cfqd->queue);
3625 cfqg_stats_update_idle_time(cfqq->cfqg);
3626 cfq_mark_cfqq_must_dispatch(cfqq);
3629 } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
3631 * not the active queue - expire current slice if it is
3632 * idle and has expired it's mean thinktime or this new queue
3633 * has some old slice time left and is of higher priority or
3634 * this new queue is RT and the current one is BE
3636 cfq_preempt_queue(cfqd, cfqq);
3637 __blk_run_queue(cfqd->queue);
3641 static void cfq_insert_request(struct request_queue *q, struct request *rq)
3643 struct cfq_data *cfqd = q->elevator->elevator_data;
3644 struct cfq_queue *cfqq = RQ_CFQQ(rq);
3646 cfq_log_cfqq(cfqd, cfqq, "insert_request");
3647 cfq_init_prio_data(cfqq, RQ_CIC(rq));
3649 rq_set_fifo_time(rq, jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)]);
3650 list_add_tail(&rq->queuelist, &cfqq->fifo);
3652 cfqg_stats_update_io_add(RQ_CFQG(rq), cfqd->serving_group,
3654 cfq_rq_enqueued(cfqd, cfqq, rq);
3658 * Update hw_tag based on peak queue depth over 50 samples under
3661 static void cfq_update_hw_tag(struct cfq_data *cfqd)
3663 struct cfq_queue *cfqq = cfqd->active_queue;
3665 if (cfqd->rq_in_driver > cfqd->hw_tag_est_depth)
3666 cfqd->hw_tag_est_depth = cfqd->rq_in_driver;
3668 if (cfqd->hw_tag == 1)
3671 if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
3672 cfqd->rq_in_driver <= CFQ_HW_QUEUE_MIN)
3676 * If active queue hasn't enough requests and can idle, cfq might not
3677 * dispatch sufficient requests to hardware. Don't zero hw_tag in this
3680 if (cfqq && cfq_cfqq_idle_window(cfqq) &&
3681 cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] <
3682 CFQ_HW_QUEUE_MIN && cfqd->rq_in_driver < CFQ_HW_QUEUE_MIN)
3685 if (cfqd->hw_tag_samples++ < 50)
3688 if (cfqd->hw_tag_est_depth >= CFQ_HW_QUEUE_MIN)
3694 static bool cfq_should_wait_busy(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3696 struct cfq_io_cq *cic = cfqd->active_cic;
3698 /* If the queue already has requests, don't wait */
3699 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3702 /* If there are other queues in the group, don't wait */
3703 if (cfqq->cfqg->nr_cfqq > 1)
3706 /* the only queue in the group, but think time is big */
3707 if (cfq_io_thinktime_big(cfqd, &cfqq->cfqg->ttime, true))
3710 if (cfq_slice_used(cfqq))
3713 /* if slice left is less than think time, wait busy */
3714 if (cic && sample_valid(cic->ttime.ttime_samples)
3715 && (cfqq->slice_end - jiffies < cic->ttime.ttime_mean))
3719 * If think times is less than a jiffy than ttime_mean=0 and above
3720 * will not be true. It might happen that slice has not expired yet
3721 * but will expire soon (4-5 ns) during select_queue(). To cover the
3722 * case where think time is less than a jiffy, mark the queue wait
3723 * busy if only 1 jiffy is left in the slice.
3725 if (cfqq->slice_end - jiffies == 1)
3731 static void cfq_completed_request(struct request_queue *q, struct request *rq)
3733 struct cfq_queue *cfqq = RQ_CFQQ(rq);
3734 struct cfq_data *cfqd = cfqq->cfqd;
3735 const int sync = rq_is_sync(rq);
3739 cfq_log_cfqq(cfqd, cfqq, "complete rqnoidle %d",
3740 !!(rq->cmd_flags & REQ_NOIDLE));
3742 cfq_update_hw_tag(cfqd);
3744 WARN_ON(!cfqd->rq_in_driver);
3745 WARN_ON(!cfqq->dispatched);
3746 cfqd->rq_in_driver--;
3748 (RQ_CFQG(rq))->dispatched--;
3749 cfqg_stats_update_completion(cfqq->cfqg, rq_start_time_ns(rq),
3750 rq_io_start_time_ns(rq), rq->cmd_flags);
3752 cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]--;
3755 struct cfq_rb_root *st;
3757 RQ_CIC(rq)->ttime.last_end_request = now;
3759 if (cfq_cfqq_on_rr(cfqq))
3760 st = cfqq->service_tree;
3762 st = st_for(cfqq->cfqg, cfqq_class(cfqq),
3765 st->ttime.last_end_request = now;
3766 if (!time_after(rq->start_time + cfqd->cfq_fifo_expire[1], now))
3767 cfqd->last_delayed_sync = now;
3770 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3771 cfqq->cfqg->ttime.last_end_request = now;
3775 * If this is the active queue, check if it needs to be expired,
3776 * or if we want to idle in case it has no pending requests.
3778 if (cfqd->active_queue == cfqq) {
3779 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
3781 if (cfq_cfqq_slice_new(cfqq)) {
3782 cfq_set_prio_slice(cfqd, cfqq);
3783 cfq_clear_cfqq_slice_new(cfqq);
3787 * Should we wait for next request to come in before we expire
3790 if (cfq_should_wait_busy(cfqd, cfqq)) {
3791 unsigned long extend_sl = cfqd->cfq_slice_idle;
3792 if (!cfqd->cfq_slice_idle)
3793 extend_sl = cfqd->cfq_group_idle;
3794 cfqq->slice_end = jiffies + extend_sl;
3795 cfq_mark_cfqq_wait_busy(cfqq);
3796 cfq_log_cfqq(cfqd, cfqq, "will busy wait");
3800 * Idling is not enabled on:
3802 * - idle-priority queues
3804 * - queues with still some requests queued
3805 * - when there is a close cooperator
3807 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
3808 cfq_slice_expired(cfqd, 1);
3809 else if (sync && cfqq_empty &&
3810 !cfq_close_cooperator(cfqd, cfqq)) {
3811 cfq_arm_slice_timer(cfqd);
3815 if (!cfqd->rq_in_driver)
3816 cfq_schedule_dispatch(cfqd);
3819 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
3821 if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
3822 cfq_mark_cfqq_must_alloc_slice(cfqq);
3823 return ELV_MQUEUE_MUST;
3826 return ELV_MQUEUE_MAY;
3829 static int cfq_may_queue(struct request_queue *q, int rw)
3831 struct cfq_data *cfqd = q->elevator->elevator_data;
3832 struct task_struct *tsk = current;
3833 struct cfq_io_cq *cic;
3834 struct cfq_queue *cfqq;
3837 * don't force setup of a queue from here, as a call to may_queue
3838 * does not necessarily imply that a request actually will be queued.
3839 * so just lookup a possibly existing queue, or return 'may queue'
3842 cic = cfq_cic_lookup(cfqd, tsk->io_context);
3844 return ELV_MQUEUE_MAY;
3846 cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
3848 cfq_init_prio_data(cfqq, cic);
3850 return __cfq_may_queue(cfqq);
3853 return ELV_MQUEUE_MAY;
3857 * queue lock held here
3859 static void cfq_put_request(struct request *rq)
3861 struct cfq_queue *cfqq = RQ_CFQQ(rq);
3864 const int rw = rq_data_dir(rq);
3866 BUG_ON(!cfqq->allocated[rw]);
3867 cfqq->allocated[rw]--;
3869 /* Put down rq reference on cfqg */
3870 cfqg_put(RQ_CFQG(rq));
3871 rq->elv.priv[0] = NULL;
3872 rq->elv.priv[1] = NULL;
3874 cfq_put_queue(cfqq);
3878 static struct cfq_queue *
3879 cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_cq *cic,
3880 struct cfq_queue *cfqq)
3882 cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq);
3883 cic_set_cfqq(cic, cfqq->new_cfqq, 1);
3884 cfq_mark_cfqq_coop(cfqq->new_cfqq);
3885 cfq_put_queue(cfqq);
3886 return cic_to_cfqq(cic, 1);
3890 * Returns NULL if a new cfqq should be allocated, or the old cfqq if this
3891 * was the last process referring to said cfqq.
3893 static struct cfq_queue *
3894 split_cfqq(struct cfq_io_cq *cic, struct cfq_queue *cfqq)
3896 if (cfqq_process_refs(cfqq) == 1) {
3897 cfqq->pid = current->pid;
3898 cfq_clear_cfqq_coop(cfqq);
3899 cfq_clear_cfqq_split_coop(cfqq);
3903 cic_set_cfqq(cic, NULL, 1);
3905 cfq_put_cooperator(cfqq);
3907 cfq_put_queue(cfqq);
3911 * Allocate cfq data structures associated with this request.
3914 cfq_set_request(struct request_queue *q, struct request *rq, struct bio *bio,
3917 struct cfq_data *cfqd = q->elevator->elevator_data;
3918 struct cfq_io_cq *cic = icq_to_cic(rq->elv.icq);
3919 const int rw = rq_data_dir(rq);
3920 const bool is_sync = rq_is_sync(rq);
3921 struct cfq_queue *cfqq;
3923 might_sleep_if(gfp_mask & __GFP_WAIT);
3925 spin_lock_irq(q->queue_lock);
3927 check_ioprio_changed(cic, bio);
3928 check_blkcg_changed(cic, bio);
3930 cfqq = cic_to_cfqq(cic, is_sync);
3931 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
3932 cfqq = cfq_get_queue(cfqd, is_sync, cic, bio, gfp_mask);
3933 cic_set_cfqq(cic, cfqq, is_sync);
3936 * If the queue was seeky for too long, break it apart.
3938 if (cfq_cfqq_coop(cfqq) && cfq_cfqq_split_coop(cfqq)) {
3939 cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq");
3940 cfqq = split_cfqq(cic, cfqq);
3946 * Check to see if this queue is scheduled to merge with
3947 * another, closely cooperating queue. The merging of
3948 * queues happens here as it must be done in process context.
3949 * The reference on new_cfqq was taken in merge_cfqqs.
3952 cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq);
3955 cfqq->allocated[rw]++;
3958 cfqg_get(cfqq->cfqg);
3959 rq->elv.priv[0] = cfqq;
3960 rq->elv.priv[1] = cfqq->cfqg;
3961 spin_unlock_irq(q->queue_lock);
3965 static void cfq_kick_queue(struct work_struct *work)
3967 struct cfq_data *cfqd =
3968 container_of(work, struct cfq_data, unplug_work);
3969 struct request_queue *q = cfqd->queue;
3971 spin_lock_irq(q->queue_lock);
3972 __blk_run_queue(cfqd->queue);
3973 spin_unlock_irq(q->queue_lock);
3977 * Timer running if the active_queue is currently idling inside its time slice
3979 static void cfq_idle_slice_timer(unsigned long data)
3981 struct cfq_data *cfqd = (struct cfq_data *) data;
3982 struct cfq_queue *cfqq;
3983 unsigned long flags;
3986 cfq_log(cfqd, "idle timer fired");
3988 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
3990 cfqq = cfqd->active_queue;
3995 * We saw a request before the queue expired, let it through
3997 if (cfq_cfqq_must_dispatch(cfqq))
4003 if (cfq_slice_used(cfqq))
4007 * only expire and reinvoke request handler, if there are
4008 * other queues with pending requests
4010 if (!cfqd->busy_queues)
4014 * not expired and it has a request pending, let it dispatch
4016 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
4020 * Queue depth flag is reset only when the idle didn't succeed
4022 cfq_clear_cfqq_deep(cfqq);
4025 cfq_slice_expired(cfqd, timed_out);
4027 cfq_schedule_dispatch(cfqd);
4029 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
4032 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
4034 del_timer_sync(&cfqd->idle_slice_timer);
4035 cancel_work_sync(&cfqd->unplug_work);
4038 static void cfq_put_async_queues(struct cfq_data *cfqd)
4042 for (i = 0; i < IOPRIO_BE_NR; i++) {
4043 if (cfqd->async_cfqq[0][i])
4044 cfq_put_queue(cfqd->async_cfqq[0][i]);
4045 if (cfqd->async_cfqq[1][i])
4046 cfq_put_queue(cfqd->async_cfqq[1][i]);
4049 if (cfqd->async_idle_cfqq)
4050 cfq_put_queue(cfqd->async_idle_cfqq);
4053 static void cfq_exit_queue(struct elevator_queue *e)
4055 struct cfq_data *cfqd = e->elevator_data;
4056 struct request_queue *q = cfqd->queue;
4058 cfq_shutdown_timer_wq(cfqd);
4060 spin_lock_irq(q->queue_lock);
4062 if (cfqd->active_queue)
4063 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
4065 cfq_put_async_queues(cfqd);
4067 spin_unlock_irq(q->queue_lock);
4069 cfq_shutdown_timer_wq(cfqd);
4071 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4072 blkcg_deactivate_policy(q, &blkcg_policy_cfq);
4074 kfree(cfqd->root_group);
4079 static int cfq_init_queue(struct request_queue *q)
4081 struct cfq_data *cfqd;
4082 struct blkcg_gq *blkg __maybe_unused;
4085 cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
4090 q->elevator->elevator_data = cfqd;
4092 /* Init root service tree */
4093 cfqd->grp_service_tree = CFQ_RB_ROOT;
4095 /* Init root group and prefer root group over other groups by default */
4096 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4097 ret = blkcg_activate_policy(q, &blkcg_policy_cfq);
4101 cfqd->root_group = blkg_to_cfqg(q->root_blkg);
4104 cfqd->root_group = kzalloc_node(sizeof(*cfqd->root_group),
4105 GFP_KERNEL, cfqd->queue->node);
4106 if (!cfqd->root_group)
4109 cfq_init_cfqg_base(cfqd->root_group);
4111 cfqd->root_group->weight = 2 * CFQ_WEIGHT_DEFAULT;
4112 cfqd->root_group->leaf_weight = 2 * CFQ_WEIGHT_DEFAULT;
4115 * Not strictly needed (since RB_ROOT just clears the node and we
4116 * zeroed cfqd on alloc), but better be safe in case someone decides
4117 * to add magic to the rb code
4119 for (i = 0; i < CFQ_PRIO_LISTS; i++)
4120 cfqd->prio_trees[i] = RB_ROOT;
4123 * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
4124 * Grab a permanent reference to it, so that the normal code flow
4125 * will not attempt to free it. oom_cfqq is linked to root_group
4126 * but shouldn't hold a reference as it'll never be unlinked. Lose
4127 * the reference from linking right away.
4129 cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
4130 cfqd->oom_cfqq.ref++;
4132 spin_lock_irq(q->queue_lock);
4133 cfq_link_cfqq_cfqg(&cfqd->oom_cfqq, cfqd->root_group);
4134 cfqg_put(cfqd->root_group);
4135 spin_unlock_irq(q->queue_lock);
4137 init_timer(&cfqd->idle_slice_timer);
4138 cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
4139 cfqd->idle_slice_timer.data = (unsigned long) cfqd;
4141 INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
4143 cfqd->cfq_quantum = cfq_quantum;
4144 cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
4145 cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
4146 cfqd->cfq_back_max = cfq_back_max;
4147 cfqd->cfq_back_penalty = cfq_back_penalty;
4148 cfqd->cfq_slice[0] = cfq_slice_async;
4149 cfqd->cfq_slice[1] = cfq_slice_sync;
4150 cfqd->cfq_target_latency = cfq_target_latency;
4151 cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
4152 cfqd->cfq_slice_idle = cfq_slice_idle;
4153 cfqd->cfq_group_idle = cfq_group_idle;
4154 cfqd->cfq_latency = 1;
4157 * we optimistically start assuming sync ops weren't delayed in last
4158 * second, in order to have larger depth for async operations.
4160 cfqd->last_delayed_sync = jiffies - HZ;
4169 * sysfs parts below -->
4172 cfq_var_show(unsigned int var, char *page)
4174 return sprintf(page, "%d\n", var);
4178 cfq_var_store(unsigned int *var, const char *page, size_t count)
4180 char *p = (char *) page;
4182 *var = simple_strtoul(p, &p, 10);
4186 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
4187 static ssize_t __FUNC(struct elevator_queue *e, char *page) \
4189 struct cfq_data *cfqd = e->elevator_data; \
4190 unsigned int __data = __VAR; \
4192 __data = jiffies_to_msecs(__data); \
4193 return cfq_var_show(__data, (page)); \
4195 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
4196 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
4197 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
4198 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
4199 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
4200 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
4201 SHOW_FUNCTION(cfq_group_idle_show, cfqd->cfq_group_idle, 1);
4202 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
4203 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
4204 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
4205 SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
4206 SHOW_FUNCTION(cfq_target_latency_show, cfqd->cfq_target_latency, 1);
4207 #undef SHOW_FUNCTION
4209 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
4210 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
4212 struct cfq_data *cfqd = e->elevator_data; \
4213 unsigned int __data; \
4214 int ret = cfq_var_store(&__data, (page), count); \
4215 if (__data < (MIN)) \
4217 else if (__data > (MAX)) \
4220 *(__PTR) = msecs_to_jiffies(__data); \
4222 *(__PTR) = __data; \
4225 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
4226 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
4228 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
4230 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
4231 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
4233 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
4234 STORE_FUNCTION(cfq_group_idle_store, &cfqd->cfq_group_idle, 0, UINT_MAX, 1);
4235 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
4236 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
4237 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
4239 STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
4240 STORE_FUNCTION(cfq_target_latency_store, &cfqd->cfq_target_latency, 1, UINT_MAX, 1);
4241 #undef STORE_FUNCTION
4243 #define CFQ_ATTR(name) \
4244 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
4246 static struct elv_fs_entry cfq_attrs[] = {
4248 CFQ_ATTR(fifo_expire_sync),
4249 CFQ_ATTR(fifo_expire_async),
4250 CFQ_ATTR(back_seek_max),
4251 CFQ_ATTR(back_seek_penalty),
4252 CFQ_ATTR(slice_sync),
4253 CFQ_ATTR(slice_async),
4254 CFQ_ATTR(slice_async_rq),
4255 CFQ_ATTR(slice_idle),
4256 CFQ_ATTR(group_idle),
4257 CFQ_ATTR(low_latency),
4258 CFQ_ATTR(target_latency),
4262 static struct elevator_type iosched_cfq = {
4264 .elevator_merge_fn = cfq_merge,
4265 .elevator_merged_fn = cfq_merged_request,
4266 .elevator_merge_req_fn = cfq_merged_requests,
4267 .elevator_allow_merge_fn = cfq_allow_merge,
4268 .elevator_bio_merged_fn = cfq_bio_merged,
4269 .elevator_dispatch_fn = cfq_dispatch_requests,
4270 .elevator_add_req_fn = cfq_insert_request,
4271 .elevator_activate_req_fn = cfq_activate_request,
4272 .elevator_deactivate_req_fn = cfq_deactivate_request,
4273 .elevator_completed_req_fn = cfq_completed_request,
4274 .elevator_former_req_fn = elv_rb_former_request,
4275 .elevator_latter_req_fn = elv_rb_latter_request,
4276 .elevator_init_icq_fn = cfq_init_icq,
4277 .elevator_exit_icq_fn = cfq_exit_icq,
4278 .elevator_set_req_fn = cfq_set_request,
4279 .elevator_put_req_fn = cfq_put_request,
4280 .elevator_may_queue_fn = cfq_may_queue,
4281 .elevator_init_fn = cfq_init_queue,
4282 .elevator_exit_fn = cfq_exit_queue,
4284 .icq_size = sizeof(struct cfq_io_cq),
4285 .icq_align = __alignof__(struct cfq_io_cq),
4286 .elevator_attrs = cfq_attrs,
4287 .elevator_name = "cfq",
4288 .elevator_owner = THIS_MODULE,
4291 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4292 static struct blkcg_policy blkcg_policy_cfq = {
4293 .pd_size = sizeof(struct cfq_group),
4294 .cftypes = cfq_blkcg_files,
4296 .pd_init_fn = cfq_pd_init,
4297 .pd_reset_stats_fn = cfq_pd_reset_stats,
4301 static int __init cfq_init(void)
4306 * could be 0 on HZ < 1000 setups
4308 if (!cfq_slice_async)
4309 cfq_slice_async = 1;
4310 if (!cfq_slice_idle)
4313 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4314 if (!cfq_group_idle)
4317 ret = blkcg_policy_register(&blkcg_policy_cfq);
4325 cfq_pool = KMEM_CACHE(cfq_queue, 0);
4329 ret = elv_register(&iosched_cfq);
4336 kmem_cache_destroy(cfq_pool);
4338 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4339 blkcg_policy_unregister(&blkcg_policy_cfq);
4344 static void __exit cfq_exit(void)
4346 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4347 blkcg_policy_unregister(&blkcg_policy_cfq);
4349 elv_unregister(&iosched_cfq);
4350 kmem_cache_destroy(cfq_pool);
4353 module_init(cfq_init);
4354 module_exit(cfq_exit);
4356 MODULE_AUTHOR("Jens Axboe");
4357 MODULE_LICENSE("GPL");
4358 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");