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.
89 struct cfq_ttime ttime;
91 #define CFQ_RB_ROOT (struct cfq_rb_root) { .rb = RB_ROOT, \
92 .ttime = {.last_end_request = jiffies,},}
95 * Per process-grouping structure
100 /* various state flags, see below */
102 /* parent cfq_data */
103 struct cfq_data *cfqd;
104 /* service_tree member */
105 struct rb_node rb_node;
106 /* service_tree key */
107 unsigned long rb_key;
108 /* prio tree member */
109 struct rb_node p_node;
110 /* prio tree root we belong to, if any */
111 struct rb_root *p_root;
112 /* sorted list of pending requests */
113 struct rb_root sort_list;
114 /* if fifo isn't expired, next request to serve */
115 struct request *next_rq;
116 /* requests queued in sort_list */
118 /* currently allocated requests */
120 /* fifo list of requests in sort_list */
121 struct list_head fifo;
123 /* time when queue got scheduled in to dispatch first request. */
124 unsigned long dispatch_start;
125 unsigned int allocated_slice;
126 unsigned int slice_dispatch;
127 /* time when first request from queue completed and slice started. */
128 unsigned long slice_start;
129 unsigned long slice_end;
132 /* pending priority requests */
134 /* number of requests that are on the dispatch list or inside driver */
137 /* io prio of this group */
138 unsigned short ioprio, org_ioprio;
139 unsigned short ioprio_class;
144 sector_t last_request_pos;
146 struct cfq_rb_root *service_tree;
147 struct cfq_queue *new_cfqq;
148 struct cfq_group *cfqg;
149 /* Number of sectors dispatched from queue in single dispatch round */
150 unsigned long nr_sectors;
154 * First index in the service_trees.
155 * IDLE is handled separately, so it has negative index
165 * Second index in the service_trees.
169 SYNC_NOIDLE_WORKLOAD = 1,
174 #ifdef CONFIG_CFQ_GROUP_IOSCHED
175 /* total bytes transferred */
176 struct blkg_rwstat service_bytes;
177 /* total IOs serviced, post merge */
178 struct blkg_rwstat serviced;
179 /* number of ios merged */
180 struct blkg_rwstat merged;
181 /* total time spent on device in ns, may not be accurate w/ queueing */
182 struct blkg_rwstat service_time;
183 /* total time spent waiting in scheduler queue in ns */
184 struct blkg_rwstat wait_time;
185 /* number of IOs queued up */
186 struct blkg_rwstat queued;
187 /* total sectors transferred */
188 struct blkg_stat sectors;
189 /* total disk time and nr sectors dispatched by this group */
190 struct blkg_stat time;
191 #ifdef CONFIG_DEBUG_BLK_CGROUP
192 /* time not charged to this cgroup */
193 struct blkg_stat unaccounted_time;
194 /* sum of number of ios queued across all samples */
195 struct blkg_stat avg_queue_size_sum;
196 /* count of samples taken for average */
197 struct blkg_stat avg_queue_size_samples;
198 /* how many times this group has been removed from service tree */
199 struct blkg_stat dequeue;
200 /* total time spent waiting for it to be assigned a timeslice. */
201 struct blkg_stat group_wait_time;
202 /* time spent idling for this blkcg_gq */
203 struct blkg_stat idle_time;
204 /* total time with empty current active q with other requests queued */
205 struct blkg_stat empty_time;
206 /* fields after this shouldn't be cleared on stat reset */
207 uint64_t start_group_wait_time;
208 uint64_t start_idle_time;
209 uint64_t start_empty_time;
211 #endif /* CONFIG_DEBUG_BLK_CGROUP */
212 #endif /* CONFIG_CFQ_GROUP_IOSCHED */
215 /* This is per cgroup per device grouping structure */
217 /* must be the first member */
218 struct blkg_policy_data pd;
220 /* group service_tree member */
221 struct rb_node rb_node;
223 /* group service_tree key */
227 * The number of active cfqgs and sum of their weights under this
228 * cfqg. This covers this cfqg's leaf_weight and all children's
229 * weights, but does not cover weights of further descendants.
231 * If a cfqg is on the service tree, it's active. An active cfqg
232 * also activates its parent and contributes to the children_weight
236 unsigned int children_weight;
239 * vfraction is the fraction of vdisktime that the tasks in this
240 * cfqg are entitled to. This is determined by compounding the
241 * ratios walking up from this cfqg to the root.
243 * It is in fixed point w/ CFQ_SERVICE_SHIFT and the sum of all
244 * vfractions on a service tree is approximately 1. The sum may
245 * deviate a bit due to rounding errors and fluctuations caused by
246 * cfqgs entering and leaving the service tree.
248 unsigned int vfraction;
251 * There are two weights - (internal) weight is the weight of this
252 * cfqg against the sibling cfqgs. leaf_weight is the wight of
253 * this cfqg against the child cfqgs. For the root cfqg, both
254 * weights are kept in sync for backward compatibility.
257 unsigned int new_weight;
258 unsigned int dev_weight;
260 unsigned int leaf_weight;
261 unsigned int new_leaf_weight;
262 unsigned int dev_leaf_weight;
264 /* number of cfqq currently on this group */
268 * Per group busy queues average. Useful for workload slice calc. We
269 * create the array for each prio class but at run time it is used
270 * only for RT and BE class and slot for IDLE class remains unused.
271 * This is primarily done to avoid confusion and a gcc warning.
273 unsigned int busy_queues_avg[CFQ_PRIO_NR];
275 * rr lists of queues with requests. We maintain service trees for
276 * RT and BE classes. These trees are subdivided in subclasses
277 * of SYNC, SYNC_NOIDLE and ASYNC based on workload type. For IDLE
278 * class there is no subclassification and all the cfq queues go on
279 * a single tree service_tree_idle.
280 * Counts are embedded in the cfq_rb_root
282 struct cfq_rb_root service_trees[2][3];
283 struct cfq_rb_root service_tree_idle;
285 unsigned long saved_wl_slice;
286 enum wl_type_t saved_wl_type;
287 enum wl_class_t saved_wl_class;
289 /* number of requests that are on the dispatch list or inside driver */
291 struct cfq_ttime ttime;
292 struct cfqg_stats stats;
296 struct io_cq icq; /* must be the first member */
297 struct cfq_queue *cfqq[2];
298 struct cfq_ttime ttime;
299 int ioprio; /* the current ioprio */
300 #ifdef CONFIG_CFQ_GROUP_IOSCHED
301 uint64_t blkcg_id; /* the current blkcg ID */
306 * Per block device queue structure
309 struct request_queue *queue;
310 /* Root service tree for cfq_groups */
311 struct cfq_rb_root grp_service_tree;
312 struct cfq_group *root_group;
315 * The priority currently being served
317 enum wl_class_t serving_wl_class;
318 enum wl_type_t serving_wl_type;
319 unsigned long workload_expires;
320 struct cfq_group *serving_group;
323 * Each priority tree is sorted by next_request position. These
324 * trees are used when determining if two or more queues are
325 * interleaving requests (see cfq_close_cooperator).
327 struct rb_root prio_trees[CFQ_PRIO_LISTS];
329 unsigned int busy_queues;
330 unsigned int busy_sync_queues;
336 * queue-depth detection
342 * -1 => indeterminate, (cfq will behave as if NCQ is present, to allow better detection)
343 * 1 => NCQ is present (hw_tag_est_depth is the estimated max depth)
346 int hw_tag_est_depth;
347 unsigned int hw_tag_samples;
350 * idle window management
352 struct timer_list idle_slice_timer;
353 struct work_struct unplug_work;
355 struct cfq_queue *active_queue;
356 struct cfq_io_cq *active_cic;
359 * async queue for each priority case
361 struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
362 struct cfq_queue *async_idle_cfqq;
364 sector_t last_position;
367 * tunables, see top of file
369 unsigned int cfq_quantum;
370 unsigned int cfq_fifo_expire[2];
371 unsigned int cfq_back_penalty;
372 unsigned int cfq_back_max;
373 unsigned int cfq_slice[2];
374 unsigned int cfq_slice_async_rq;
375 unsigned int cfq_slice_idle;
376 unsigned int cfq_group_idle;
377 unsigned int cfq_latency;
378 unsigned int cfq_target_latency;
381 * Fallback dummy cfqq for extreme OOM conditions
383 struct cfq_queue oom_cfqq;
385 unsigned long last_delayed_sync;
388 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd);
390 static struct cfq_rb_root *st_for(struct cfq_group *cfqg,
391 enum wl_class_t class,
397 if (class == IDLE_WORKLOAD)
398 return &cfqg->service_tree_idle;
400 return &cfqg->service_trees[class][type];
403 enum cfqq_state_flags {
404 CFQ_CFQQ_FLAG_on_rr = 0, /* on round-robin busy list */
405 CFQ_CFQQ_FLAG_wait_request, /* waiting for a request */
406 CFQ_CFQQ_FLAG_must_dispatch, /* must be allowed a dispatch */
407 CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
408 CFQ_CFQQ_FLAG_fifo_expire, /* FIFO checked in this slice */
409 CFQ_CFQQ_FLAG_idle_window, /* slice idling enabled */
410 CFQ_CFQQ_FLAG_prio_changed, /* task priority has changed */
411 CFQ_CFQQ_FLAG_slice_new, /* no requests dispatched in slice */
412 CFQ_CFQQ_FLAG_sync, /* synchronous queue */
413 CFQ_CFQQ_FLAG_coop, /* cfqq is shared */
414 CFQ_CFQQ_FLAG_split_coop, /* shared cfqq will be splitted */
415 CFQ_CFQQ_FLAG_deep, /* sync cfqq experienced large depth */
416 CFQ_CFQQ_FLAG_wait_busy, /* Waiting for next request */
419 #define CFQ_CFQQ_FNS(name) \
420 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \
422 (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name); \
424 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \
426 (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \
428 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \
430 return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \
434 CFQ_CFQQ_FNS(wait_request);
435 CFQ_CFQQ_FNS(must_dispatch);
436 CFQ_CFQQ_FNS(must_alloc_slice);
437 CFQ_CFQQ_FNS(fifo_expire);
438 CFQ_CFQQ_FNS(idle_window);
439 CFQ_CFQQ_FNS(prio_changed);
440 CFQ_CFQQ_FNS(slice_new);
443 CFQ_CFQQ_FNS(split_coop);
445 CFQ_CFQQ_FNS(wait_busy);
448 static inline struct cfq_group *pd_to_cfqg(struct blkg_policy_data *pd)
450 return pd ? container_of(pd, struct cfq_group, pd) : NULL;
453 static inline struct blkcg_gq *cfqg_to_blkg(struct cfq_group *cfqg)
455 return pd_to_blkg(&cfqg->pd);
458 #if defined(CONFIG_CFQ_GROUP_IOSCHED) && defined(CONFIG_DEBUG_BLK_CGROUP)
460 /* cfqg stats flags */
461 enum cfqg_stats_flags {
462 CFQG_stats_waiting = 0,
467 #define CFQG_FLAG_FNS(name) \
468 static inline void cfqg_stats_mark_##name(struct cfqg_stats *stats) \
470 stats->flags |= (1 << CFQG_stats_##name); \
472 static inline void cfqg_stats_clear_##name(struct cfqg_stats *stats) \
474 stats->flags &= ~(1 << CFQG_stats_##name); \
476 static inline int cfqg_stats_##name(struct cfqg_stats *stats) \
478 return (stats->flags & (1 << CFQG_stats_##name)) != 0; \
481 CFQG_FLAG_FNS(waiting)
482 CFQG_FLAG_FNS(idling)
486 /* This should be called with the queue_lock held. */
487 static void cfqg_stats_update_group_wait_time(struct cfqg_stats *stats)
489 unsigned long long now;
491 if (!cfqg_stats_waiting(stats))
495 if (time_after64(now, stats->start_group_wait_time))
496 blkg_stat_add(&stats->group_wait_time,
497 now - stats->start_group_wait_time);
498 cfqg_stats_clear_waiting(stats);
501 /* This should be called with the queue_lock held. */
502 static void cfqg_stats_set_start_group_wait_time(struct cfq_group *cfqg,
503 struct cfq_group *curr_cfqg)
505 struct cfqg_stats *stats = &cfqg->stats;
507 if (cfqg_stats_waiting(stats))
509 if (cfqg == curr_cfqg)
511 stats->start_group_wait_time = sched_clock();
512 cfqg_stats_mark_waiting(stats);
515 /* This should be called with the queue_lock held. */
516 static void cfqg_stats_end_empty_time(struct cfqg_stats *stats)
518 unsigned long long now;
520 if (!cfqg_stats_empty(stats))
524 if (time_after64(now, stats->start_empty_time))
525 blkg_stat_add(&stats->empty_time,
526 now - stats->start_empty_time);
527 cfqg_stats_clear_empty(stats);
530 static void cfqg_stats_update_dequeue(struct cfq_group *cfqg)
532 blkg_stat_add(&cfqg->stats.dequeue, 1);
535 static void cfqg_stats_set_start_empty_time(struct cfq_group *cfqg)
537 struct cfqg_stats *stats = &cfqg->stats;
539 if (blkg_rwstat_sum(&stats->queued))
543 * group is already marked empty. This can happen if cfqq got new
544 * request in parent group and moved to this group while being added
545 * to service tree. Just ignore the event and move on.
547 if (cfqg_stats_empty(stats))
550 stats->start_empty_time = sched_clock();
551 cfqg_stats_mark_empty(stats);
554 static void cfqg_stats_update_idle_time(struct cfq_group *cfqg)
556 struct cfqg_stats *stats = &cfqg->stats;
558 if (cfqg_stats_idling(stats)) {
559 unsigned long long now = sched_clock();
561 if (time_after64(now, stats->start_idle_time))
562 blkg_stat_add(&stats->idle_time,
563 now - stats->start_idle_time);
564 cfqg_stats_clear_idling(stats);
568 static void cfqg_stats_set_start_idle_time(struct cfq_group *cfqg)
570 struct cfqg_stats *stats = &cfqg->stats;
572 BUG_ON(cfqg_stats_idling(stats));
574 stats->start_idle_time = sched_clock();
575 cfqg_stats_mark_idling(stats);
578 static void cfqg_stats_update_avg_queue_size(struct cfq_group *cfqg)
580 struct cfqg_stats *stats = &cfqg->stats;
582 blkg_stat_add(&stats->avg_queue_size_sum,
583 blkg_rwstat_sum(&stats->queued));
584 blkg_stat_add(&stats->avg_queue_size_samples, 1);
585 cfqg_stats_update_group_wait_time(stats);
588 #else /* CONFIG_CFQ_GROUP_IOSCHED && CONFIG_DEBUG_BLK_CGROUP */
590 static inline void cfqg_stats_set_start_group_wait_time(struct cfq_group *cfqg, struct cfq_group *curr_cfqg) { }
591 static inline void cfqg_stats_end_empty_time(struct cfqg_stats *stats) { }
592 static inline void cfqg_stats_update_dequeue(struct cfq_group *cfqg) { }
593 static inline void cfqg_stats_set_start_empty_time(struct cfq_group *cfqg) { }
594 static inline void cfqg_stats_update_idle_time(struct cfq_group *cfqg) { }
595 static inline void cfqg_stats_set_start_idle_time(struct cfq_group *cfqg) { }
596 static inline void cfqg_stats_update_avg_queue_size(struct cfq_group *cfqg) { }
598 #endif /* CONFIG_CFQ_GROUP_IOSCHED && CONFIG_DEBUG_BLK_CGROUP */
600 #ifdef CONFIG_CFQ_GROUP_IOSCHED
602 static struct blkcg_policy blkcg_policy_cfq;
604 static inline struct cfq_group *blkg_to_cfqg(struct blkcg_gq *blkg)
606 return pd_to_cfqg(blkg_to_pd(blkg, &blkcg_policy_cfq));
609 static inline struct cfq_group *cfqg_parent(struct cfq_group *cfqg)
611 struct blkcg_gq *pblkg = cfqg_to_blkg(cfqg)->parent;
613 return pblkg ? blkg_to_cfqg(pblkg) : NULL;
616 static inline void cfqg_get(struct cfq_group *cfqg)
618 return blkg_get(cfqg_to_blkg(cfqg));
621 static inline void cfqg_put(struct cfq_group *cfqg)
623 return blkg_put(cfqg_to_blkg(cfqg));
626 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...) do { \
629 blkg_path(cfqg_to_blkg((cfqq)->cfqg), __pbuf, sizeof(__pbuf)); \
630 blk_add_trace_msg((cfqd)->queue, "cfq%d%c%c %s " fmt, (cfqq)->pid, \
631 cfq_cfqq_sync((cfqq)) ? 'S' : 'A', \
632 cfqq_type((cfqq)) == SYNC_NOIDLE_WORKLOAD ? 'N' : ' ',\
636 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...) do { \
639 blkg_path(cfqg_to_blkg(cfqg), __pbuf, sizeof(__pbuf)); \
640 blk_add_trace_msg((cfqd)->queue, "%s " fmt, __pbuf, ##args); \
643 static inline void cfqg_stats_update_io_add(struct cfq_group *cfqg,
644 struct cfq_group *curr_cfqg, int rw)
646 blkg_rwstat_add(&cfqg->stats.queued, rw, 1);
647 cfqg_stats_end_empty_time(&cfqg->stats);
648 cfqg_stats_set_start_group_wait_time(cfqg, curr_cfqg);
651 static inline void cfqg_stats_update_timeslice_used(struct cfq_group *cfqg,
652 unsigned long time, unsigned long unaccounted_time)
654 blkg_stat_add(&cfqg->stats.time, time);
655 #ifdef CONFIG_DEBUG_BLK_CGROUP
656 blkg_stat_add(&cfqg->stats.unaccounted_time, unaccounted_time);
660 static inline void cfqg_stats_update_io_remove(struct cfq_group *cfqg, int rw)
662 blkg_rwstat_add(&cfqg->stats.queued, rw, -1);
665 static inline void cfqg_stats_update_io_merged(struct cfq_group *cfqg, int rw)
667 blkg_rwstat_add(&cfqg->stats.merged, rw, 1);
670 static inline void cfqg_stats_update_dispatch(struct cfq_group *cfqg,
671 uint64_t bytes, int rw)
673 blkg_stat_add(&cfqg->stats.sectors, bytes >> 9);
674 blkg_rwstat_add(&cfqg->stats.serviced, rw, 1);
675 blkg_rwstat_add(&cfqg->stats.service_bytes, rw, bytes);
678 static inline void cfqg_stats_update_completion(struct cfq_group *cfqg,
679 uint64_t start_time, uint64_t io_start_time, int rw)
681 struct cfqg_stats *stats = &cfqg->stats;
682 unsigned long long now = sched_clock();
684 if (time_after64(now, io_start_time))
685 blkg_rwstat_add(&stats->service_time, rw, now - io_start_time);
686 if (time_after64(io_start_time, start_time))
687 blkg_rwstat_add(&stats->wait_time, rw,
688 io_start_time - start_time);
691 static void cfq_pd_reset_stats(struct blkcg_gq *blkg)
693 struct cfq_group *cfqg = blkg_to_cfqg(blkg);
694 struct cfqg_stats *stats = &cfqg->stats;
696 /* queued stats shouldn't be cleared */
697 blkg_rwstat_reset(&stats->service_bytes);
698 blkg_rwstat_reset(&stats->serviced);
699 blkg_rwstat_reset(&stats->merged);
700 blkg_rwstat_reset(&stats->service_time);
701 blkg_rwstat_reset(&stats->wait_time);
702 blkg_stat_reset(&stats->time);
703 #ifdef CONFIG_DEBUG_BLK_CGROUP
704 blkg_stat_reset(&stats->unaccounted_time);
705 blkg_stat_reset(&stats->avg_queue_size_sum);
706 blkg_stat_reset(&stats->avg_queue_size_samples);
707 blkg_stat_reset(&stats->dequeue);
708 blkg_stat_reset(&stats->group_wait_time);
709 blkg_stat_reset(&stats->idle_time);
710 blkg_stat_reset(&stats->empty_time);
714 #else /* CONFIG_CFQ_GROUP_IOSCHED */
716 static inline struct cfq_group *cfqg_parent(struct cfq_group *cfqg) { return NULL; }
717 static inline void cfqg_get(struct cfq_group *cfqg) { }
718 static inline void cfqg_put(struct cfq_group *cfqg) { }
720 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
721 blk_add_trace_msg((cfqd)->queue, "cfq%d%c%c " fmt, (cfqq)->pid, \
722 cfq_cfqq_sync((cfqq)) ? 'S' : 'A', \
723 cfqq_type((cfqq)) == SYNC_NOIDLE_WORKLOAD ? 'N' : ' ',\
725 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...) do {} while (0)
727 static inline void cfqg_stats_update_io_add(struct cfq_group *cfqg,
728 struct cfq_group *curr_cfqg, int rw) { }
729 static inline void cfqg_stats_update_timeslice_used(struct cfq_group *cfqg,
730 unsigned long time, unsigned long unaccounted_time) { }
731 static inline void cfqg_stats_update_io_remove(struct cfq_group *cfqg, int rw) { }
732 static inline void cfqg_stats_update_io_merged(struct cfq_group *cfqg, int rw) { }
733 static inline void cfqg_stats_update_dispatch(struct cfq_group *cfqg,
734 uint64_t bytes, int rw) { }
735 static inline void cfqg_stats_update_completion(struct cfq_group *cfqg,
736 uint64_t start_time, uint64_t io_start_time, int rw) { }
738 #endif /* CONFIG_CFQ_GROUP_IOSCHED */
740 #define cfq_log(cfqd, fmt, args...) \
741 blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
743 /* Traverses through cfq group service trees */
744 #define for_each_cfqg_st(cfqg, i, j, st) \
745 for (i = 0; i <= IDLE_WORKLOAD; i++) \
746 for (j = 0, st = i < IDLE_WORKLOAD ? &cfqg->service_trees[i][j]\
747 : &cfqg->service_tree_idle; \
748 (i < IDLE_WORKLOAD && j <= SYNC_WORKLOAD) || \
749 (i == IDLE_WORKLOAD && j == 0); \
750 j++, st = i < IDLE_WORKLOAD ? \
751 &cfqg->service_trees[i][j]: NULL) \
753 static inline bool cfq_io_thinktime_big(struct cfq_data *cfqd,
754 struct cfq_ttime *ttime, bool group_idle)
757 if (!sample_valid(ttime->ttime_samples))
760 slice = cfqd->cfq_group_idle;
762 slice = cfqd->cfq_slice_idle;
763 return ttime->ttime_mean > slice;
766 static inline bool iops_mode(struct cfq_data *cfqd)
769 * If we are not idling on queues and it is a NCQ drive, parallel
770 * execution of requests is on and measuring time is not possible
771 * in most of the cases until and unless we drive shallower queue
772 * depths and that becomes a performance bottleneck. In such cases
773 * switch to start providing fairness in terms of number of IOs.
775 if (!cfqd->cfq_slice_idle && cfqd->hw_tag)
781 static inline enum wl_class_t cfqq_class(struct cfq_queue *cfqq)
783 if (cfq_class_idle(cfqq))
784 return IDLE_WORKLOAD;
785 if (cfq_class_rt(cfqq))
791 static enum wl_type_t cfqq_type(struct cfq_queue *cfqq)
793 if (!cfq_cfqq_sync(cfqq))
794 return ASYNC_WORKLOAD;
795 if (!cfq_cfqq_idle_window(cfqq))
796 return SYNC_NOIDLE_WORKLOAD;
797 return SYNC_WORKLOAD;
800 static inline int cfq_group_busy_queues_wl(enum wl_class_t wl_class,
801 struct cfq_data *cfqd,
802 struct cfq_group *cfqg)
804 if (wl_class == IDLE_WORKLOAD)
805 return cfqg->service_tree_idle.count;
807 return cfqg->service_trees[wl_class][ASYNC_WORKLOAD].count +
808 cfqg->service_trees[wl_class][SYNC_NOIDLE_WORKLOAD].count +
809 cfqg->service_trees[wl_class][SYNC_WORKLOAD].count;
812 static inline int cfqg_busy_async_queues(struct cfq_data *cfqd,
813 struct cfq_group *cfqg)
815 return cfqg->service_trees[RT_WORKLOAD][ASYNC_WORKLOAD].count +
816 cfqg->service_trees[BE_WORKLOAD][ASYNC_WORKLOAD].count;
819 static void cfq_dispatch_insert(struct request_queue *, struct request *);
820 static struct cfq_queue *cfq_get_queue(struct cfq_data *cfqd, bool is_sync,
821 struct cfq_io_cq *cic, struct bio *bio,
824 static inline struct cfq_io_cq *icq_to_cic(struct io_cq *icq)
826 /* cic->icq is the first member, %NULL will convert to %NULL */
827 return container_of(icq, struct cfq_io_cq, icq);
830 static inline struct cfq_io_cq *cfq_cic_lookup(struct cfq_data *cfqd,
831 struct io_context *ioc)
834 return icq_to_cic(ioc_lookup_icq(ioc, cfqd->queue));
838 static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_cq *cic, bool is_sync)
840 return cic->cfqq[is_sync];
843 static inline void cic_set_cfqq(struct cfq_io_cq *cic, struct cfq_queue *cfqq,
846 cic->cfqq[is_sync] = cfqq;
849 static inline struct cfq_data *cic_to_cfqd(struct cfq_io_cq *cic)
851 return cic->icq.q->elevator->elevator_data;
855 * We regard a request as SYNC, if it's either a read or has the SYNC bit
856 * set (in which case it could also be direct WRITE).
858 static inline bool cfq_bio_sync(struct bio *bio)
860 return bio_data_dir(bio) == READ || (bio->bi_rw & REQ_SYNC);
864 * scheduler run of queue, if there are requests pending and no one in the
865 * driver that will restart queueing
867 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
869 if (cfqd->busy_queues) {
870 cfq_log(cfqd, "schedule dispatch");
871 kblockd_schedule_work(cfqd->queue, &cfqd->unplug_work);
876 * Scale schedule slice based on io priority. Use the sync time slice only
877 * if a queue is marked sync and has sync io queued. A sync queue with async
878 * io only, should not get full sync slice length.
880 static inline int cfq_prio_slice(struct cfq_data *cfqd, bool sync,
883 const int base_slice = cfqd->cfq_slice[sync];
885 WARN_ON(prio >= IOPRIO_BE_NR);
887 return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
891 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
893 return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
897 * cfqg_scale_charge - scale disk time charge according to cfqg weight
898 * @charge: disk time being charged
899 * @vfraction: vfraction of the cfqg, fixed point w/ CFQ_SERVICE_SHIFT
901 * Scale @charge according to @vfraction, which is in range (0, 1]. The
902 * scaling is inversely proportional.
904 * scaled = charge / vfraction
906 * The result is also in fixed point w/ CFQ_SERVICE_SHIFT.
908 static inline u64 cfqg_scale_charge(unsigned long charge,
909 unsigned int vfraction)
911 u64 c = charge << CFQ_SERVICE_SHIFT; /* make it fixed point */
913 /* charge / vfraction */
914 c <<= CFQ_SERVICE_SHIFT;
915 do_div(c, vfraction);
919 static inline u64 max_vdisktime(u64 min_vdisktime, u64 vdisktime)
921 s64 delta = (s64)(vdisktime - min_vdisktime);
923 min_vdisktime = vdisktime;
925 return min_vdisktime;
928 static inline u64 min_vdisktime(u64 min_vdisktime, u64 vdisktime)
930 s64 delta = (s64)(vdisktime - min_vdisktime);
932 min_vdisktime = vdisktime;
934 return min_vdisktime;
937 static void update_min_vdisktime(struct cfq_rb_root *st)
939 struct cfq_group *cfqg;
942 cfqg = rb_entry_cfqg(st->left);
943 st->min_vdisktime = max_vdisktime(st->min_vdisktime,
949 * get averaged number of queues of RT/BE priority.
950 * average is updated, with a formula that gives more weight to higher numbers,
951 * to quickly follows sudden increases and decrease slowly
954 static inline unsigned cfq_group_get_avg_queues(struct cfq_data *cfqd,
955 struct cfq_group *cfqg, bool rt)
957 unsigned min_q, max_q;
958 unsigned mult = cfq_hist_divisor - 1;
959 unsigned round = cfq_hist_divisor / 2;
960 unsigned busy = cfq_group_busy_queues_wl(rt, cfqd, cfqg);
962 min_q = min(cfqg->busy_queues_avg[rt], busy);
963 max_q = max(cfqg->busy_queues_avg[rt], busy);
964 cfqg->busy_queues_avg[rt] = (mult * max_q + min_q + round) /
966 return cfqg->busy_queues_avg[rt];
969 static inline unsigned
970 cfq_group_slice(struct cfq_data *cfqd, struct cfq_group *cfqg)
972 return cfqd->cfq_target_latency * cfqg->vfraction >> CFQ_SERVICE_SHIFT;
975 static inline unsigned
976 cfq_scaled_cfqq_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
978 unsigned slice = cfq_prio_to_slice(cfqd, cfqq);
979 if (cfqd->cfq_latency) {
981 * interested queues (we consider only the ones with the same
982 * priority class in the cfq group)
984 unsigned iq = cfq_group_get_avg_queues(cfqd, cfqq->cfqg,
986 unsigned sync_slice = cfqd->cfq_slice[1];
987 unsigned expect_latency = sync_slice * iq;
988 unsigned group_slice = cfq_group_slice(cfqd, cfqq->cfqg);
990 if (expect_latency > group_slice) {
991 unsigned base_low_slice = 2 * cfqd->cfq_slice_idle;
992 /* scale low_slice according to IO priority
993 * and sync vs async */
995 min(slice, base_low_slice * slice / sync_slice);
996 /* the adapted slice value is scaled to fit all iqs
997 * into the target latency */
998 slice = max(slice * group_slice / expect_latency,
1006 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1008 unsigned slice = cfq_scaled_cfqq_slice(cfqd, cfqq);
1010 cfqq->slice_start = jiffies;
1011 cfqq->slice_end = jiffies + slice;
1012 cfqq->allocated_slice = slice;
1013 cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
1017 * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
1018 * isn't valid until the first request from the dispatch is activated
1019 * and the slice time set.
1021 static inline bool cfq_slice_used(struct cfq_queue *cfqq)
1023 if (cfq_cfqq_slice_new(cfqq))
1025 if (time_before(jiffies, cfqq->slice_end))
1032 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
1033 * We choose the request that is closest to the head right now. Distance
1034 * behind the head is penalized and only allowed to a certain extent.
1036 static struct request *
1037 cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2, sector_t last)
1039 sector_t s1, s2, d1 = 0, d2 = 0;
1040 unsigned long back_max;
1041 #define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */
1042 #define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */
1043 unsigned wrap = 0; /* bit mask: requests behind the disk head? */
1045 if (rq1 == NULL || rq1 == rq2)
1050 if (rq_is_sync(rq1) != rq_is_sync(rq2))
1051 return rq_is_sync(rq1) ? rq1 : rq2;
1053 if ((rq1->cmd_flags ^ rq2->cmd_flags) & REQ_PRIO)
1054 return rq1->cmd_flags & REQ_PRIO ? rq1 : rq2;
1056 s1 = blk_rq_pos(rq1);
1057 s2 = blk_rq_pos(rq2);
1060 * by definition, 1KiB is 2 sectors
1062 back_max = cfqd->cfq_back_max * 2;
1065 * Strict one way elevator _except_ in the case where we allow
1066 * short backward seeks which are biased as twice the cost of a
1067 * similar forward seek.
1071 else if (s1 + back_max >= last)
1072 d1 = (last - s1) * cfqd->cfq_back_penalty;
1074 wrap |= CFQ_RQ1_WRAP;
1078 else if (s2 + back_max >= last)
1079 d2 = (last - s2) * cfqd->cfq_back_penalty;
1081 wrap |= CFQ_RQ2_WRAP;
1083 /* Found required data */
1086 * By doing switch() on the bit mask "wrap" we avoid having to
1087 * check two variables for all permutations: --> faster!
1090 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
1106 case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
1109 * Since both rqs are wrapped,
1110 * start with the one that's further behind head
1111 * (--> only *one* back seek required),
1112 * since back seek takes more time than forward.
1122 * The below is leftmost cache rbtree addon
1124 static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
1126 /* Service tree is empty */
1131 root->left = rb_first(&root->rb);
1134 return rb_entry(root->left, struct cfq_queue, rb_node);
1139 static struct cfq_group *cfq_rb_first_group(struct cfq_rb_root *root)
1142 root->left = rb_first(&root->rb);
1145 return rb_entry_cfqg(root->left);
1150 static void rb_erase_init(struct rb_node *n, struct rb_root *root)
1156 static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
1158 if (root->left == n)
1160 rb_erase_init(n, &root->rb);
1165 * would be nice to take fifo expire time into account as well
1167 static struct request *
1168 cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1169 struct request *last)
1171 struct rb_node *rbnext = rb_next(&last->rb_node);
1172 struct rb_node *rbprev = rb_prev(&last->rb_node);
1173 struct request *next = NULL, *prev = NULL;
1175 BUG_ON(RB_EMPTY_NODE(&last->rb_node));
1178 prev = rb_entry_rq(rbprev);
1181 next = rb_entry_rq(rbnext);
1183 rbnext = rb_first(&cfqq->sort_list);
1184 if (rbnext && rbnext != &last->rb_node)
1185 next = rb_entry_rq(rbnext);
1188 return cfq_choose_req(cfqd, next, prev, blk_rq_pos(last));
1191 static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
1192 struct cfq_queue *cfqq)
1195 * just an approximation, should be ok.
1197 return (cfqq->cfqg->nr_cfqq - 1) * (cfq_prio_slice(cfqd, 1, 0) -
1198 cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
1202 cfqg_key(struct cfq_rb_root *st, struct cfq_group *cfqg)
1204 return cfqg->vdisktime - st->min_vdisktime;
1208 __cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
1210 struct rb_node **node = &st->rb.rb_node;
1211 struct rb_node *parent = NULL;
1212 struct cfq_group *__cfqg;
1213 s64 key = cfqg_key(st, cfqg);
1216 while (*node != NULL) {
1218 __cfqg = rb_entry_cfqg(parent);
1220 if (key < cfqg_key(st, __cfqg))
1221 node = &parent->rb_left;
1223 node = &parent->rb_right;
1229 st->left = &cfqg->rb_node;
1231 rb_link_node(&cfqg->rb_node, parent, node);
1232 rb_insert_color(&cfqg->rb_node, &st->rb);
1236 cfq_update_group_weight(struct cfq_group *cfqg)
1238 BUG_ON(!RB_EMPTY_NODE(&cfqg->rb_node));
1240 if (cfqg->new_weight) {
1241 cfqg->weight = cfqg->new_weight;
1242 cfqg->new_weight = 0;
1245 if (cfqg->new_leaf_weight) {
1246 cfqg->leaf_weight = cfqg->new_leaf_weight;
1247 cfqg->new_leaf_weight = 0;
1252 cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
1254 unsigned int vfr = 1 << CFQ_SERVICE_SHIFT; /* start with 1 */
1255 struct cfq_group *pos = cfqg;
1256 struct cfq_group *parent;
1259 /* add to the service tree */
1260 BUG_ON(!RB_EMPTY_NODE(&cfqg->rb_node));
1262 cfq_update_group_weight(cfqg);
1263 __cfq_group_service_tree_add(st, cfqg);
1266 * Activate @cfqg and calculate the portion of vfraction @cfqg is
1267 * entitled to. vfraction is calculated by walking the tree
1268 * towards the root calculating the fraction it has at each level.
1269 * The compounded ratio is how much vfraction @cfqg owns.
1271 * Start with the proportion tasks in this cfqg has against active
1272 * children cfqgs - its leaf_weight against children_weight.
1274 propagate = !pos->nr_active++;
1275 pos->children_weight += pos->leaf_weight;
1276 vfr = vfr * pos->leaf_weight / pos->children_weight;
1279 * Compound ->weight walking up the tree. Both activation and
1280 * vfraction calculation are done in the same loop. Propagation
1281 * stops once an already activated node is met. vfraction
1282 * calculation should always continue to the root.
1284 while ((parent = cfqg_parent(pos))) {
1286 propagate = !parent->nr_active++;
1287 parent->children_weight += pos->weight;
1289 vfr = vfr * pos->weight / parent->children_weight;
1293 cfqg->vfraction = max_t(unsigned, vfr, 1);
1297 cfq_group_notify_queue_add(struct cfq_data *cfqd, struct cfq_group *cfqg)
1299 struct cfq_rb_root *st = &cfqd->grp_service_tree;
1300 struct cfq_group *__cfqg;
1304 if (!RB_EMPTY_NODE(&cfqg->rb_node))
1308 * Currently put the group at the end. Later implement something
1309 * so that groups get lesser vtime based on their weights, so that
1310 * if group does not loose all if it was not continuously backlogged.
1312 n = rb_last(&st->rb);
1314 __cfqg = rb_entry_cfqg(n);
1315 cfqg->vdisktime = __cfqg->vdisktime + CFQ_IDLE_DELAY;
1317 cfqg->vdisktime = st->min_vdisktime;
1318 cfq_group_service_tree_add(st, cfqg);
1322 cfq_group_service_tree_del(struct cfq_rb_root *st, struct cfq_group *cfqg)
1324 struct cfq_group *pos = cfqg;
1328 * Undo activation from cfq_group_service_tree_add(). Deactivate
1329 * @cfqg and propagate deactivation upwards.
1331 propagate = !--pos->nr_active;
1332 pos->children_weight -= pos->leaf_weight;
1335 struct cfq_group *parent = cfqg_parent(pos);
1337 /* @pos has 0 nr_active at this point */
1338 WARN_ON_ONCE(pos->children_weight);
1344 propagate = !--parent->nr_active;
1345 parent->children_weight -= pos->weight;
1349 /* remove from the service tree */
1350 if (!RB_EMPTY_NODE(&cfqg->rb_node))
1351 cfq_rb_erase(&cfqg->rb_node, st);
1355 cfq_group_notify_queue_del(struct cfq_data *cfqd, struct cfq_group *cfqg)
1357 struct cfq_rb_root *st = &cfqd->grp_service_tree;
1359 BUG_ON(cfqg->nr_cfqq < 1);
1362 /* If there are other cfq queues under this group, don't delete it */
1366 cfq_log_cfqg(cfqd, cfqg, "del_from_rr group");
1367 cfq_group_service_tree_del(st, cfqg);
1368 cfqg->saved_wl_slice = 0;
1369 cfqg_stats_update_dequeue(cfqg);
1372 static inline unsigned int cfq_cfqq_slice_usage(struct cfq_queue *cfqq,
1373 unsigned int *unaccounted_time)
1375 unsigned int slice_used;
1378 * Queue got expired before even a single request completed or
1379 * got expired immediately after first request completion.
1381 if (!cfqq->slice_start || cfqq->slice_start == jiffies) {
1383 * Also charge the seek time incurred to the group, otherwise
1384 * if there are mutiple queues in the group, each can dispatch
1385 * a single request on seeky media and cause lots of seek time
1386 * and group will never know it.
1388 slice_used = max_t(unsigned, (jiffies - cfqq->dispatch_start),
1391 slice_used = jiffies - cfqq->slice_start;
1392 if (slice_used > cfqq->allocated_slice) {
1393 *unaccounted_time = slice_used - cfqq->allocated_slice;
1394 slice_used = cfqq->allocated_slice;
1396 if (time_after(cfqq->slice_start, cfqq->dispatch_start))
1397 *unaccounted_time += cfqq->slice_start -
1398 cfqq->dispatch_start;
1404 static void cfq_group_served(struct cfq_data *cfqd, struct cfq_group *cfqg,
1405 struct cfq_queue *cfqq)
1407 struct cfq_rb_root *st = &cfqd->grp_service_tree;
1408 unsigned int used_sl, charge, unaccounted_sl = 0;
1409 int nr_sync = cfqg->nr_cfqq - cfqg_busy_async_queues(cfqd, cfqg)
1410 - cfqg->service_tree_idle.count;
1413 BUG_ON(nr_sync < 0);
1414 used_sl = charge = cfq_cfqq_slice_usage(cfqq, &unaccounted_sl);
1416 if (iops_mode(cfqd))
1417 charge = cfqq->slice_dispatch;
1418 else if (!cfq_cfqq_sync(cfqq) && !nr_sync)
1419 charge = cfqq->allocated_slice;
1422 * Can't update vdisktime while on service tree and cfqg->vfraction
1423 * is valid only while on it. Cache vfr, leave the service tree,
1424 * update vdisktime and go back on. The re-addition to the tree
1425 * will also update the weights as necessary.
1427 vfr = cfqg->vfraction;
1428 cfq_group_service_tree_del(st, cfqg);
1429 cfqg->vdisktime += cfqg_scale_charge(charge, vfr);
1430 cfq_group_service_tree_add(st, cfqg);
1432 /* This group is being expired. Save the context */
1433 if (time_after(cfqd->workload_expires, jiffies)) {
1434 cfqg->saved_wl_slice = cfqd->workload_expires
1436 cfqg->saved_wl_type = cfqd->serving_wl_type;
1437 cfqg->saved_wl_class = cfqd->serving_wl_class;
1439 cfqg->saved_wl_slice = 0;
1441 cfq_log_cfqg(cfqd, cfqg, "served: vt=%llu min_vt=%llu", cfqg->vdisktime,
1443 cfq_log_cfqq(cfqq->cfqd, cfqq,
1444 "sl_used=%u disp=%u charge=%u iops=%u sect=%lu",
1445 used_sl, cfqq->slice_dispatch, charge,
1446 iops_mode(cfqd), cfqq->nr_sectors);
1447 cfqg_stats_update_timeslice_used(cfqg, used_sl, unaccounted_sl);
1448 cfqg_stats_set_start_empty_time(cfqg);
1452 * cfq_init_cfqg_base - initialize base part of a cfq_group
1453 * @cfqg: cfq_group to initialize
1455 * Initialize the base part which is used whether %CONFIG_CFQ_GROUP_IOSCHED
1456 * is enabled or not.
1458 static void cfq_init_cfqg_base(struct cfq_group *cfqg)
1460 struct cfq_rb_root *st;
1463 for_each_cfqg_st(cfqg, i, j, st)
1465 RB_CLEAR_NODE(&cfqg->rb_node);
1467 cfqg->ttime.last_end_request = jiffies;
1470 #ifdef CONFIG_CFQ_GROUP_IOSCHED
1471 static void cfq_pd_init(struct blkcg_gq *blkg)
1473 struct cfq_group *cfqg = blkg_to_cfqg(blkg);
1475 cfq_init_cfqg_base(cfqg);
1476 cfqg->weight = blkg->blkcg->cfq_weight;
1477 cfqg->leaf_weight = blkg->blkcg->cfq_leaf_weight;
1481 * Search for the cfq group current task belongs to. request_queue lock must
1484 static struct cfq_group *cfq_lookup_create_cfqg(struct cfq_data *cfqd,
1485 struct blkcg *blkcg)
1487 struct request_queue *q = cfqd->queue;
1488 struct cfq_group *cfqg = NULL;
1490 /* avoid lookup for the common case where there's no blkcg */
1491 if (blkcg == &blkcg_root) {
1492 cfqg = cfqd->root_group;
1494 struct blkcg_gq *blkg;
1496 blkg = blkg_lookup_create(blkcg, q);
1498 cfqg = blkg_to_cfqg(blkg);
1504 static void cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg)
1506 /* Currently, all async queues are mapped to root group */
1507 if (!cfq_cfqq_sync(cfqq))
1508 cfqg = cfqq->cfqd->root_group;
1511 /* cfqq reference on cfqg */
1515 static u64 cfqg_prfill_weight_device(struct seq_file *sf,
1516 struct blkg_policy_data *pd, int off)
1518 struct cfq_group *cfqg = pd_to_cfqg(pd);
1520 if (!cfqg->dev_weight)
1522 return __blkg_prfill_u64(sf, pd, cfqg->dev_weight);
1525 static int cfqg_print_weight_device(struct cgroup *cgrp, struct cftype *cft,
1526 struct seq_file *sf)
1528 blkcg_print_blkgs(sf, cgroup_to_blkcg(cgrp),
1529 cfqg_prfill_weight_device, &blkcg_policy_cfq, 0,
1534 static u64 cfqg_prfill_leaf_weight_device(struct seq_file *sf,
1535 struct blkg_policy_data *pd, int off)
1537 struct cfq_group *cfqg = pd_to_cfqg(pd);
1539 if (!cfqg->dev_leaf_weight)
1541 return __blkg_prfill_u64(sf, pd, cfqg->dev_leaf_weight);
1544 static int cfqg_print_leaf_weight_device(struct cgroup *cgrp,
1546 struct seq_file *sf)
1548 blkcg_print_blkgs(sf, cgroup_to_blkcg(cgrp),
1549 cfqg_prfill_leaf_weight_device, &blkcg_policy_cfq, 0,
1554 static int cfq_print_weight(struct cgroup *cgrp, struct cftype *cft,
1555 struct seq_file *sf)
1557 seq_printf(sf, "%u\n", cgroup_to_blkcg(cgrp)->cfq_weight);
1561 static int cfq_print_leaf_weight(struct cgroup *cgrp, struct cftype *cft,
1562 struct seq_file *sf)
1564 seq_printf(sf, "%u\n",
1565 cgroup_to_blkcg(cgrp)->cfq_leaf_weight);
1569 static int __cfqg_set_weight_device(struct cgroup *cgrp, struct cftype *cft,
1570 const char *buf, bool is_leaf_weight)
1572 struct blkcg *blkcg = cgroup_to_blkcg(cgrp);
1573 struct blkg_conf_ctx ctx;
1574 struct cfq_group *cfqg;
1577 ret = blkg_conf_prep(blkcg, &blkcg_policy_cfq, buf, &ctx);
1582 cfqg = blkg_to_cfqg(ctx.blkg);
1583 if (!ctx.v || (ctx.v >= CFQ_WEIGHT_MIN && ctx.v <= CFQ_WEIGHT_MAX)) {
1584 if (!is_leaf_weight) {
1585 cfqg->dev_weight = ctx.v;
1586 cfqg->new_weight = ctx.v ?: blkcg->cfq_weight;
1588 cfqg->dev_leaf_weight = ctx.v;
1589 cfqg->new_leaf_weight = ctx.v ?: blkcg->cfq_leaf_weight;
1594 blkg_conf_finish(&ctx);
1598 static int cfqg_set_weight_device(struct cgroup *cgrp, struct cftype *cft,
1601 return __cfqg_set_weight_device(cgrp, cft, buf, false);
1604 static int cfqg_set_leaf_weight_device(struct cgroup *cgrp, struct cftype *cft,
1607 return __cfqg_set_weight_device(cgrp, cft, buf, true);
1610 static int __cfq_set_weight(struct cgroup *cgrp, struct cftype *cft, u64 val,
1611 bool is_leaf_weight)
1613 struct blkcg *blkcg = cgroup_to_blkcg(cgrp);
1614 struct blkcg_gq *blkg;
1615 struct hlist_node *n;
1617 if (val < CFQ_WEIGHT_MIN || val > CFQ_WEIGHT_MAX)
1620 spin_lock_irq(&blkcg->lock);
1622 if (!is_leaf_weight)
1623 blkcg->cfq_weight = val;
1625 blkcg->cfq_leaf_weight = val;
1627 hlist_for_each_entry(blkg, n, &blkcg->blkg_list, blkcg_node) {
1628 struct cfq_group *cfqg = blkg_to_cfqg(blkg);
1633 if (!is_leaf_weight) {
1634 if (!cfqg->dev_weight)
1635 cfqg->new_weight = blkcg->cfq_weight;
1637 if (!cfqg->dev_leaf_weight)
1638 cfqg->new_leaf_weight = blkcg->cfq_leaf_weight;
1642 spin_unlock_irq(&blkcg->lock);
1646 static int cfq_set_weight(struct cgroup *cgrp, struct cftype *cft, u64 val)
1648 return __cfq_set_weight(cgrp, cft, val, false);
1651 static int cfq_set_leaf_weight(struct cgroup *cgrp, struct cftype *cft, u64 val)
1653 return __cfq_set_weight(cgrp, cft, val, true);
1656 static int cfqg_print_stat(struct cgroup *cgrp, struct cftype *cft,
1657 struct seq_file *sf)
1659 struct blkcg *blkcg = cgroup_to_blkcg(cgrp);
1661 blkcg_print_blkgs(sf, blkcg, blkg_prfill_stat, &blkcg_policy_cfq,
1662 cft->private, false);
1666 static int cfqg_print_rwstat(struct cgroup *cgrp, struct cftype *cft,
1667 struct seq_file *sf)
1669 struct blkcg *blkcg = cgroup_to_blkcg(cgrp);
1671 blkcg_print_blkgs(sf, blkcg, blkg_prfill_rwstat, &blkcg_policy_cfq,
1672 cft->private, true);
1676 #ifdef CONFIG_DEBUG_BLK_CGROUP
1677 static u64 cfqg_prfill_avg_queue_size(struct seq_file *sf,
1678 struct blkg_policy_data *pd, int off)
1680 struct cfq_group *cfqg = pd_to_cfqg(pd);
1681 u64 samples = blkg_stat_read(&cfqg->stats.avg_queue_size_samples);
1685 v = blkg_stat_read(&cfqg->stats.avg_queue_size_sum);
1688 __blkg_prfill_u64(sf, pd, v);
1692 /* print avg_queue_size */
1693 static int cfqg_print_avg_queue_size(struct cgroup *cgrp, struct cftype *cft,
1694 struct seq_file *sf)
1696 struct blkcg *blkcg = cgroup_to_blkcg(cgrp);
1698 blkcg_print_blkgs(sf, blkcg, cfqg_prfill_avg_queue_size,
1699 &blkcg_policy_cfq, 0, false);
1702 #endif /* CONFIG_DEBUG_BLK_CGROUP */
1704 static struct cftype cfq_blkcg_files[] = {
1705 /* on root, weight is mapped to leaf_weight */
1707 .name = "weight_device",
1708 .flags = CFTYPE_ONLY_ON_ROOT,
1709 .read_seq_string = cfqg_print_leaf_weight_device,
1710 .write_string = cfqg_set_leaf_weight_device,
1711 .max_write_len = 256,
1715 .flags = CFTYPE_ONLY_ON_ROOT,
1716 .read_seq_string = cfq_print_leaf_weight,
1717 .write_u64 = cfq_set_leaf_weight,
1720 /* no such mapping necessary for !roots */
1722 .name = "weight_device",
1723 .flags = CFTYPE_NOT_ON_ROOT,
1724 .read_seq_string = cfqg_print_weight_device,
1725 .write_string = cfqg_set_weight_device,
1726 .max_write_len = 256,
1730 .flags = CFTYPE_NOT_ON_ROOT,
1731 .read_seq_string = cfq_print_weight,
1732 .write_u64 = cfq_set_weight,
1736 .name = "leaf_weight_device",
1737 .read_seq_string = cfqg_print_leaf_weight_device,
1738 .write_string = cfqg_set_leaf_weight_device,
1739 .max_write_len = 256,
1742 .name = "leaf_weight",
1743 .read_seq_string = cfq_print_leaf_weight,
1744 .write_u64 = cfq_set_leaf_weight,
1749 .private = offsetof(struct cfq_group, stats.time),
1750 .read_seq_string = cfqg_print_stat,
1754 .private = offsetof(struct cfq_group, stats.sectors),
1755 .read_seq_string = cfqg_print_stat,
1758 .name = "io_service_bytes",
1759 .private = offsetof(struct cfq_group, stats.service_bytes),
1760 .read_seq_string = cfqg_print_rwstat,
1763 .name = "io_serviced",
1764 .private = offsetof(struct cfq_group, stats.serviced),
1765 .read_seq_string = cfqg_print_rwstat,
1768 .name = "io_service_time",
1769 .private = offsetof(struct cfq_group, stats.service_time),
1770 .read_seq_string = cfqg_print_rwstat,
1773 .name = "io_wait_time",
1774 .private = offsetof(struct cfq_group, stats.wait_time),
1775 .read_seq_string = cfqg_print_rwstat,
1778 .name = "io_merged",
1779 .private = offsetof(struct cfq_group, stats.merged),
1780 .read_seq_string = cfqg_print_rwstat,
1783 .name = "io_queued",
1784 .private = offsetof(struct cfq_group, stats.queued),
1785 .read_seq_string = cfqg_print_rwstat,
1787 #ifdef CONFIG_DEBUG_BLK_CGROUP
1789 .name = "avg_queue_size",
1790 .read_seq_string = cfqg_print_avg_queue_size,
1793 .name = "group_wait_time",
1794 .private = offsetof(struct cfq_group, stats.group_wait_time),
1795 .read_seq_string = cfqg_print_stat,
1798 .name = "idle_time",
1799 .private = offsetof(struct cfq_group, stats.idle_time),
1800 .read_seq_string = cfqg_print_stat,
1803 .name = "empty_time",
1804 .private = offsetof(struct cfq_group, stats.empty_time),
1805 .read_seq_string = cfqg_print_stat,
1809 .private = offsetof(struct cfq_group, stats.dequeue),
1810 .read_seq_string = cfqg_print_stat,
1813 .name = "unaccounted_time",
1814 .private = offsetof(struct cfq_group, stats.unaccounted_time),
1815 .read_seq_string = cfqg_print_stat,
1817 #endif /* CONFIG_DEBUG_BLK_CGROUP */
1820 #else /* GROUP_IOSCHED */
1821 static struct cfq_group *cfq_lookup_create_cfqg(struct cfq_data *cfqd,
1822 struct blkcg *blkcg)
1824 return cfqd->root_group;
1828 cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) {
1832 #endif /* GROUP_IOSCHED */
1835 * The cfqd->service_trees holds all pending cfq_queue's that have
1836 * requests waiting to be processed. It is sorted in the order that
1837 * we will service the queues.
1839 static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1842 struct rb_node **p, *parent;
1843 struct cfq_queue *__cfqq;
1844 unsigned long rb_key;
1845 struct cfq_rb_root *st;
1849 st = st_for(cfqq->cfqg, cfqq_class(cfqq), cfqq_type(cfqq));
1850 if (cfq_class_idle(cfqq)) {
1851 rb_key = CFQ_IDLE_DELAY;
1852 parent = rb_last(&st->rb);
1853 if (parent && parent != &cfqq->rb_node) {
1854 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1855 rb_key += __cfqq->rb_key;
1858 } else if (!add_front) {
1860 * Get our rb key offset. Subtract any residual slice
1861 * value carried from last service. A negative resid
1862 * count indicates slice overrun, and this should position
1863 * the next service time further away in the tree.
1865 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
1866 rb_key -= cfqq->slice_resid;
1867 cfqq->slice_resid = 0;
1870 __cfqq = cfq_rb_first(st);
1871 rb_key += __cfqq ? __cfqq->rb_key : jiffies;
1874 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1877 * same position, nothing more to do
1879 if (rb_key == cfqq->rb_key && cfqq->service_tree == st)
1882 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1883 cfqq->service_tree = NULL;
1888 cfqq->service_tree = st;
1889 p = &st->rb.rb_node;
1892 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1895 * sort by key, that represents service time.
1897 if (time_before(rb_key, __cfqq->rb_key))
1898 p = &parent->rb_left;
1900 p = &parent->rb_right;
1906 st->left = &cfqq->rb_node;
1908 cfqq->rb_key = rb_key;
1909 rb_link_node(&cfqq->rb_node, parent, p);
1910 rb_insert_color(&cfqq->rb_node, &st->rb);
1912 if (add_front || !new_cfqq)
1914 cfq_group_notify_queue_add(cfqd, cfqq->cfqg);
1917 static struct cfq_queue *
1918 cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
1919 sector_t sector, struct rb_node **ret_parent,
1920 struct rb_node ***rb_link)
1922 struct rb_node **p, *parent;
1923 struct cfq_queue *cfqq = NULL;
1931 cfqq = rb_entry(parent, struct cfq_queue, p_node);
1934 * Sort strictly based on sector. Smallest to the left,
1935 * largest to the right.
1937 if (sector > blk_rq_pos(cfqq->next_rq))
1938 n = &(*p)->rb_right;
1939 else if (sector < blk_rq_pos(cfqq->next_rq))
1947 *ret_parent = parent;
1953 static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1955 struct rb_node **p, *parent;
1956 struct cfq_queue *__cfqq;
1959 rb_erase(&cfqq->p_node, cfqq->p_root);
1960 cfqq->p_root = NULL;
1963 if (cfq_class_idle(cfqq))
1968 cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
1969 __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
1970 blk_rq_pos(cfqq->next_rq), &parent, &p);
1972 rb_link_node(&cfqq->p_node, parent, p);
1973 rb_insert_color(&cfqq->p_node, cfqq->p_root);
1975 cfqq->p_root = NULL;
1979 * Update cfqq's position in the service tree.
1981 static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1984 * Resorting requires the cfqq to be on the RR list already.
1986 if (cfq_cfqq_on_rr(cfqq)) {
1987 cfq_service_tree_add(cfqd, cfqq, 0);
1988 cfq_prio_tree_add(cfqd, cfqq);
1993 * add to busy list of queues for service, trying to be fair in ordering
1994 * the pending list according to last request service
1996 static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1998 cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
1999 BUG_ON(cfq_cfqq_on_rr(cfqq));
2000 cfq_mark_cfqq_on_rr(cfqq);
2001 cfqd->busy_queues++;
2002 if (cfq_cfqq_sync(cfqq))
2003 cfqd->busy_sync_queues++;
2005 cfq_resort_rr_list(cfqd, cfqq);
2009 * Called when the cfqq no longer has requests pending, remove it from
2012 static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2014 cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
2015 BUG_ON(!cfq_cfqq_on_rr(cfqq));
2016 cfq_clear_cfqq_on_rr(cfqq);
2018 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
2019 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
2020 cfqq->service_tree = NULL;
2023 rb_erase(&cfqq->p_node, cfqq->p_root);
2024 cfqq->p_root = NULL;
2027 cfq_group_notify_queue_del(cfqd, cfqq->cfqg);
2028 BUG_ON(!cfqd->busy_queues);
2029 cfqd->busy_queues--;
2030 if (cfq_cfqq_sync(cfqq))
2031 cfqd->busy_sync_queues--;
2035 * rb tree support functions
2037 static void cfq_del_rq_rb(struct request *rq)
2039 struct cfq_queue *cfqq = RQ_CFQQ(rq);
2040 const int sync = rq_is_sync(rq);
2042 BUG_ON(!cfqq->queued[sync]);
2043 cfqq->queued[sync]--;
2045 elv_rb_del(&cfqq->sort_list, rq);
2047 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) {
2049 * Queue will be deleted from service tree when we actually
2050 * expire it later. Right now just remove it from prio tree
2054 rb_erase(&cfqq->p_node, cfqq->p_root);
2055 cfqq->p_root = NULL;
2060 static void cfq_add_rq_rb(struct request *rq)
2062 struct cfq_queue *cfqq = RQ_CFQQ(rq);
2063 struct cfq_data *cfqd = cfqq->cfqd;
2064 struct request *prev;
2066 cfqq->queued[rq_is_sync(rq)]++;
2068 elv_rb_add(&cfqq->sort_list, rq);
2070 if (!cfq_cfqq_on_rr(cfqq))
2071 cfq_add_cfqq_rr(cfqd, cfqq);
2074 * check if this request is a better next-serve candidate
2076 prev = cfqq->next_rq;
2077 cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq, cfqd->last_position);
2080 * adjust priority tree position, if ->next_rq changes
2082 if (prev != cfqq->next_rq)
2083 cfq_prio_tree_add(cfqd, cfqq);
2085 BUG_ON(!cfqq->next_rq);
2088 static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
2090 elv_rb_del(&cfqq->sort_list, rq);
2091 cfqq->queued[rq_is_sync(rq)]--;
2092 cfqg_stats_update_io_remove(RQ_CFQG(rq), rq->cmd_flags);
2094 cfqg_stats_update_io_add(RQ_CFQG(rq), cfqq->cfqd->serving_group,
2098 static struct request *
2099 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
2101 struct task_struct *tsk = current;
2102 struct cfq_io_cq *cic;
2103 struct cfq_queue *cfqq;
2105 cic = cfq_cic_lookup(cfqd, tsk->io_context);
2109 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
2111 sector_t sector = bio->bi_sector + bio_sectors(bio);
2113 return elv_rb_find(&cfqq->sort_list, sector);
2119 static void cfq_activate_request(struct request_queue *q, struct request *rq)
2121 struct cfq_data *cfqd = q->elevator->elevator_data;
2123 cfqd->rq_in_driver++;
2124 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
2125 cfqd->rq_in_driver);
2127 cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
2130 static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
2132 struct cfq_data *cfqd = q->elevator->elevator_data;
2134 WARN_ON(!cfqd->rq_in_driver);
2135 cfqd->rq_in_driver--;
2136 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
2137 cfqd->rq_in_driver);
2140 static void cfq_remove_request(struct request *rq)
2142 struct cfq_queue *cfqq = RQ_CFQQ(rq);
2144 if (cfqq->next_rq == rq)
2145 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
2147 list_del_init(&rq->queuelist);
2150 cfqq->cfqd->rq_queued--;
2151 cfqg_stats_update_io_remove(RQ_CFQG(rq), rq->cmd_flags);
2152 if (rq->cmd_flags & REQ_PRIO) {
2153 WARN_ON(!cfqq->prio_pending);
2154 cfqq->prio_pending--;
2158 static int cfq_merge(struct request_queue *q, struct request **req,
2161 struct cfq_data *cfqd = q->elevator->elevator_data;
2162 struct request *__rq;
2164 __rq = cfq_find_rq_fmerge(cfqd, bio);
2165 if (__rq && elv_rq_merge_ok(__rq, bio)) {
2167 return ELEVATOR_FRONT_MERGE;
2170 return ELEVATOR_NO_MERGE;
2173 static void cfq_merged_request(struct request_queue *q, struct request *req,
2176 if (type == ELEVATOR_FRONT_MERGE) {
2177 struct cfq_queue *cfqq = RQ_CFQQ(req);
2179 cfq_reposition_rq_rb(cfqq, req);
2183 static void cfq_bio_merged(struct request_queue *q, struct request *req,
2186 cfqg_stats_update_io_merged(RQ_CFQG(req), bio->bi_rw);
2190 cfq_merged_requests(struct request_queue *q, struct request *rq,
2191 struct request *next)
2193 struct cfq_queue *cfqq = RQ_CFQQ(rq);
2194 struct cfq_data *cfqd = q->elevator->elevator_data;
2197 * reposition in fifo if next is older than rq
2199 if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
2200 time_before(rq_fifo_time(next), rq_fifo_time(rq)) &&
2201 cfqq == RQ_CFQQ(next)) {
2202 list_move(&rq->queuelist, &next->queuelist);
2203 rq_set_fifo_time(rq, rq_fifo_time(next));
2206 if (cfqq->next_rq == next)
2208 cfq_remove_request(next);
2209 cfqg_stats_update_io_merged(RQ_CFQG(rq), next->cmd_flags);
2211 cfqq = RQ_CFQQ(next);
2213 * all requests of this queue are merged to other queues, delete it
2214 * from the service tree. If it's the active_queue,
2215 * cfq_dispatch_requests() will choose to expire it or do idle
2217 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list) &&
2218 cfqq != cfqd->active_queue)
2219 cfq_del_cfqq_rr(cfqd, cfqq);
2222 static int cfq_allow_merge(struct request_queue *q, struct request *rq,
2225 struct cfq_data *cfqd = q->elevator->elevator_data;
2226 struct cfq_io_cq *cic;
2227 struct cfq_queue *cfqq;
2230 * Disallow merge of a sync bio into an async request.
2232 if (cfq_bio_sync(bio) && !rq_is_sync(rq))
2236 * Lookup the cfqq that this bio will be queued with and allow
2237 * merge only if rq is queued there.
2239 cic = cfq_cic_lookup(cfqd, current->io_context);
2243 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
2244 return cfqq == RQ_CFQQ(rq);
2247 static inline void cfq_del_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2249 del_timer(&cfqd->idle_slice_timer);
2250 cfqg_stats_update_idle_time(cfqq->cfqg);
2253 static void __cfq_set_active_queue(struct cfq_data *cfqd,
2254 struct cfq_queue *cfqq)
2257 cfq_log_cfqq(cfqd, cfqq, "set_active wl_class:%d wl_type:%d",
2258 cfqd->serving_wl_class, cfqd->serving_wl_type);
2259 cfqg_stats_update_avg_queue_size(cfqq->cfqg);
2260 cfqq->slice_start = 0;
2261 cfqq->dispatch_start = jiffies;
2262 cfqq->allocated_slice = 0;
2263 cfqq->slice_end = 0;
2264 cfqq->slice_dispatch = 0;
2265 cfqq->nr_sectors = 0;
2267 cfq_clear_cfqq_wait_request(cfqq);
2268 cfq_clear_cfqq_must_dispatch(cfqq);
2269 cfq_clear_cfqq_must_alloc_slice(cfqq);
2270 cfq_clear_cfqq_fifo_expire(cfqq);
2271 cfq_mark_cfqq_slice_new(cfqq);
2273 cfq_del_timer(cfqd, cfqq);
2276 cfqd->active_queue = cfqq;
2280 * current cfqq expired its slice (or was too idle), select new one
2283 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2286 cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
2288 if (cfq_cfqq_wait_request(cfqq))
2289 cfq_del_timer(cfqd, cfqq);
2291 cfq_clear_cfqq_wait_request(cfqq);
2292 cfq_clear_cfqq_wait_busy(cfqq);
2295 * If this cfqq is shared between multiple processes, check to
2296 * make sure that those processes are still issuing I/Os within
2297 * the mean seek distance. If not, it may be time to break the
2298 * queues apart again.
2300 if (cfq_cfqq_coop(cfqq) && CFQQ_SEEKY(cfqq))
2301 cfq_mark_cfqq_split_coop(cfqq);
2304 * store what was left of this slice, if the queue idled/timed out
2307 if (cfq_cfqq_slice_new(cfqq))
2308 cfqq->slice_resid = cfq_scaled_cfqq_slice(cfqd, cfqq);
2310 cfqq->slice_resid = cfqq->slice_end - jiffies;
2311 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
2314 cfq_group_served(cfqd, cfqq->cfqg, cfqq);
2316 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
2317 cfq_del_cfqq_rr(cfqd, cfqq);
2319 cfq_resort_rr_list(cfqd, cfqq);
2321 if (cfqq == cfqd->active_queue)
2322 cfqd->active_queue = NULL;
2324 if (cfqd->active_cic) {
2325 put_io_context(cfqd->active_cic->icq.ioc);
2326 cfqd->active_cic = NULL;
2330 static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out)
2332 struct cfq_queue *cfqq = cfqd->active_queue;
2335 __cfq_slice_expired(cfqd, cfqq, timed_out);
2339 * Get next queue for service. Unless we have a queue preemption,
2340 * we'll simply select the first cfqq in the service tree.
2342 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
2344 struct cfq_rb_root *st = st_for(cfqd->serving_group,
2345 cfqd->serving_wl_class, cfqd->serving_wl_type);
2347 if (!cfqd->rq_queued)
2350 /* There is nothing to dispatch */
2353 if (RB_EMPTY_ROOT(&st->rb))
2355 return cfq_rb_first(st);
2358 static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd)
2360 struct cfq_group *cfqg;
2361 struct cfq_queue *cfqq;
2363 struct cfq_rb_root *st;
2365 if (!cfqd->rq_queued)
2368 cfqg = cfq_get_next_cfqg(cfqd);
2372 for_each_cfqg_st(cfqg, i, j, st)
2373 if ((cfqq = cfq_rb_first(st)) != NULL)
2379 * Get and set a new active queue for service.
2381 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
2382 struct cfq_queue *cfqq)
2385 cfqq = cfq_get_next_queue(cfqd);
2387 __cfq_set_active_queue(cfqd, cfqq);
2391 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
2394 if (blk_rq_pos(rq) >= cfqd->last_position)
2395 return blk_rq_pos(rq) - cfqd->last_position;
2397 return cfqd->last_position - blk_rq_pos(rq);
2400 static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2403 return cfq_dist_from_last(cfqd, rq) <= CFQQ_CLOSE_THR;
2406 static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
2407 struct cfq_queue *cur_cfqq)
2409 struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
2410 struct rb_node *parent, *node;
2411 struct cfq_queue *__cfqq;
2412 sector_t sector = cfqd->last_position;
2414 if (RB_EMPTY_ROOT(root))
2418 * First, if we find a request starting at the end of the last
2419 * request, choose it.
2421 __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
2426 * If the exact sector wasn't found, the parent of the NULL leaf
2427 * will contain the closest sector.
2429 __cfqq = rb_entry(parent, struct cfq_queue, p_node);
2430 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
2433 if (blk_rq_pos(__cfqq->next_rq) < sector)
2434 node = rb_next(&__cfqq->p_node);
2436 node = rb_prev(&__cfqq->p_node);
2440 __cfqq = rb_entry(node, struct cfq_queue, p_node);
2441 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
2449 * cur_cfqq - passed in so that we don't decide that the current queue is
2450 * closely cooperating with itself.
2452 * So, basically we're assuming that that cur_cfqq has dispatched at least
2453 * one request, and that cfqd->last_position reflects a position on the disk
2454 * associated with the I/O issued by cur_cfqq. I'm not sure this is a valid
2457 static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
2458 struct cfq_queue *cur_cfqq)
2460 struct cfq_queue *cfqq;
2462 if (cfq_class_idle(cur_cfqq))
2464 if (!cfq_cfqq_sync(cur_cfqq))
2466 if (CFQQ_SEEKY(cur_cfqq))
2470 * Don't search priority tree if it's the only queue in the group.
2472 if (cur_cfqq->cfqg->nr_cfqq == 1)
2476 * We should notice if some of the queues are cooperating, eg
2477 * working closely on the same area of the disk. In that case,
2478 * we can group them together and don't waste time idling.
2480 cfqq = cfqq_close(cfqd, cur_cfqq);
2484 /* If new queue belongs to different cfq_group, don't choose it */
2485 if (cur_cfqq->cfqg != cfqq->cfqg)
2489 * It only makes sense to merge sync queues.
2491 if (!cfq_cfqq_sync(cfqq))
2493 if (CFQQ_SEEKY(cfqq))
2497 * Do not merge queues of different priority classes
2499 if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq))
2506 * Determine whether we should enforce idle window for this queue.
2509 static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2511 enum wl_class_t wl_class = cfqq_class(cfqq);
2512 struct cfq_rb_root *st = cfqq->service_tree;
2517 if (!cfqd->cfq_slice_idle)
2520 /* We never do for idle class queues. */
2521 if (wl_class == IDLE_WORKLOAD)
2524 /* We do for queues that were marked with idle window flag. */
2525 if (cfq_cfqq_idle_window(cfqq) &&
2526 !(blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag))
2530 * Otherwise, we do only if they are the last ones
2531 * in their service tree.
2533 if (st->count == 1 && cfq_cfqq_sync(cfqq) &&
2534 !cfq_io_thinktime_big(cfqd, &st->ttime, false))
2536 cfq_log_cfqq(cfqd, cfqq, "Not idling. st->count:%d", st->count);
2540 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
2542 struct cfq_queue *cfqq = cfqd->active_queue;
2543 struct cfq_io_cq *cic;
2544 unsigned long sl, group_idle = 0;
2547 * SSD device without seek penalty, disable idling. But only do so
2548 * for devices that support queuing, otherwise we still have a problem
2549 * with sync vs async workloads.
2551 if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
2554 WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
2555 WARN_ON(cfq_cfqq_slice_new(cfqq));
2558 * idle is disabled, either manually or by past process history
2560 if (!cfq_should_idle(cfqd, cfqq)) {
2561 /* no queue idling. Check for group idling */
2562 if (cfqd->cfq_group_idle)
2563 group_idle = cfqd->cfq_group_idle;
2569 * still active requests from this queue, don't idle
2571 if (cfqq->dispatched)
2575 * task has exited, don't wait
2577 cic = cfqd->active_cic;
2578 if (!cic || !atomic_read(&cic->icq.ioc->active_ref))
2582 * If our average think time is larger than the remaining time
2583 * slice, then don't idle. This avoids overrunning the allotted
2586 if (sample_valid(cic->ttime.ttime_samples) &&
2587 (cfqq->slice_end - jiffies < cic->ttime.ttime_mean)) {
2588 cfq_log_cfqq(cfqd, cfqq, "Not idling. think_time:%lu",
2589 cic->ttime.ttime_mean);
2593 /* There are other queues in the group, don't do group idle */
2594 if (group_idle && cfqq->cfqg->nr_cfqq > 1)
2597 cfq_mark_cfqq_wait_request(cfqq);
2600 sl = cfqd->cfq_group_idle;
2602 sl = cfqd->cfq_slice_idle;
2604 mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
2605 cfqg_stats_set_start_idle_time(cfqq->cfqg);
2606 cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu group_idle: %d", sl,
2607 group_idle ? 1 : 0);
2611 * Move request from internal lists to the request queue dispatch list.
2613 static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
2615 struct cfq_data *cfqd = q->elevator->elevator_data;
2616 struct cfq_queue *cfqq = RQ_CFQQ(rq);
2618 cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
2620 cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
2621 cfq_remove_request(rq);
2623 (RQ_CFQG(rq))->dispatched++;
2624 elv_dispatch_sort(q, rq);
2626 cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]++;
2627 cfqq->nr_sectors += blk_rq_sectors(rq);
2628 cfqg_stats_update_dispatch(cfqq->cfqg, blk_rq_bytes(rq), rq->cmd_flags);
2632 * return expired entry, or NULL to just start from scratch in rbtree
2634 static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
2636 struct request *rq = NULL;
2638 if (cfq_cfqq_fifo_expire(cfqq))
2641 cfq_mark_cfqq_fifo_expire(cfqq);
2643 if (list_empty(&cfqq->fifo))
2646 rq = rq_entry_fifo(cfqq->fifo.next);
2647 if (time_before(jiffies, rq_fifo_time(rq)))
2650 cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq);
2655 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2657 const int base_rq = cfqd->cfq_slice_async_rq;
2659 WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
2661 return 2 * base_rq * (IOPRIO_BE_NR - cfqq->ioprio);
2665 * Must be called with the queue_lock held.
2667 static int cfqq_process_refs(struct cfq_queue *cfqq)
2669 int process_refs, io_refs;
2671 io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE];
2672 process_refs = cfqq->ref - io_refs;
2673 BUG_ON(process_refs < 0);
2674 return process_refs;
2677 static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq)
2679 int process_refs, new_process_refs;
2680 struct cfq_queue *__cfqq;
2683 * If there are no process references on the new_cfqq, then it is
2684 * unsafe to follow the ->new_cfqq chain as other cfqq's in the
2685 * chain may have dropped their last reference (not just their
2686 * last process reference).
2688 if (!cfqq_process_refs(new_cfqq))
2691 /* Avoid a circular list and skip interim queue merges */
2692 while ((__cfqq = new_cfqq->new_cfqq)) {
2698 process_refs = cfqq_process_refs(cfqq);
2699 new_process_refs = cfqq_process_refs(new_cfqq);
2701 * If the process for the cfqq has gone away, there is no
2702 * sense in merging the queues.
2704 if (process_refs == 0 || new_process_refs == 0)
2708 * Merge in the direction of the lesser amount of work.
2710 if (new_process_refs >= process_refs) {
2711 cfqq->new_cfqq = new_cfqq;
2712 new_cfqq->ref += process_refs;
2714 new_cfqq->new_cfqq = cfqq;
2715 cfqq->ref += new_process_refs;
2719 static enum wl_type_t cfq_choose_wl_type(struct cfq_data *cfqd,
2720 struct cfq_group *cfqg, enum wl_class_t wl_class)
2722 struct cfq_queue *queue;
2724 bool key_valid = false;
2725 unsigned long lowest_key = 0;
2726 enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD;
2728 for (i = 0; i <= SYNC_WORKLOAD; ++i) {
2729 /* select the one with lowest rb_key */
2730 queue = cfq_rb_first(st_for(cfqg, wl_class, i));
2732 (!key_valid || time_before(queue->rb_key, lowest_key))) {
2733 lowest_key = queue->rb_key;
2743 choose_wl_class_and_type(struct cfq_data *cfqd, struct cfq_group *cfqg)
2747 struct cfq_rb_root *st;
2748 unsigned group_slice;
2749 enum wl_class_t original_class = cfqd->serving_wl_class;
2751 /* Choose next priority. RT > BE > IDLE */
2752 if (cfq_group_busy_queues_wl(RT_WORKLOAD, cfqd, cfqg))
2753 cfqd->serving_wl_class = RT_WORKLOAD;
2754 else if (cfq_group_busy_queues_wl(BE_WORKLOAD, cfqd, cfqg))
2755 cfqd->serving_wl_class = BE_WORKLOAD;
2757 cfqd->serving_wl_class = IDLE_WORKLOAD;
2758 cfqd->workload_expires = jiffies + 1;
2762 if (original_class != cfqd->serving_wl_class)
2766 * For RT and BE, we have to choose also the type
2767 * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
2770 st = st_for(cfqg, cfqd->serving_wl_class, cfqd->serving_wl_type);
2774 * check workload expiration, and that we still have other queues ready
2776 if (count && !time_after(jiffies, cfqd->workload_expires))
2780 /* otherwise select new workload type */
2781 cfqd->serving_wl_type = cfq_choose_wl_type(cfqd, cfqg,
2782 cfqd->serving_wl_class);
2783 st = st_for(cfqg, cfqd->serving_wl_class, cfqd->serving_wl_type);
2787 * the workload slice is computed as a fraction of target latency
2788 * proportional to the number of queues in that workload, over
2789 * all the queues in the same priority class
2791 group_slice = cfq_group_slice(cfqd, cfqg);
2793 slice = group_slice * count /
2794 max_t(unsigned, cfqg->busy_queues_avg[cfqd->serving_wl_class],
2795 cfq_group_busy_queues_wl(cfqd->serving_wl_class, cfqd,
2798 if (cfqd->serving_wl_type == ASYNC_WORKLOAD) {
2802 * Async queues are currently system wide. Just taking
2803 * proportion of queues with-in same group will lead to higher
2804 * async ratio system wide as generally root group is going
2805 * to have higher weight. A more accurate thing would be to
2806 * calculate system wide asnc/sync ratio.
2808 tmp = cfqd->cfq_target_latency *
2809 cfqg_busy_async_queues(cfqd, cfqg);
2810 tmp = tmp/cfqd->busy_queues;
2811 slice = min_t(unsigned, slice, tmp);
2813 /* async workload slice is scaled down according to
2814 * the sync/async slice ratio. */
2815 slice = slice * cfqd->cfq_slice[0] / cfqd->cfq_slice[1];
2817 /* sync workload slice is at least 2 * cfq_slice_idle */
2818 slice = max(slice, 2 * cfqd->cfq_slice_idle);
2820 slice = max_t(unsigned, slice, CFQ_MIN_TT);
2821 cfq_log(cfqd, "workload slice:%d", slice);
2822 cfqd->workload_expires = jiffies + slice;
2825 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd)
2827 struct cfq_rb_root *st = &cfqd->grp_service_tree;
2828 struct cfq_group *cfqg;
2830 if (RB_EMPTY_ROOT(&st->rb))
2832 cfqg = cfq_rb_first_group(st);
2833 update_min_vdisktime(st);
2837 static void cfq_choose_cfqg(struct cfq_data *cfqd)
2839 struct cfq_group *cfqg = cfq_get_next_cfqg(cfqd);
2841 cfqd->serving_group = cfqg;
2843 /* Restore the workload type data */
2844 if (cfqg->saved_wl_slice) {
2845 cfqd->workload_expires = jiffies + cfqg->saved_wl_slice;
2846 cfqd->serving_wl_type = cfqg->saved_wl_type;
2847 cfqd->serving_wl_class = cfqg->saved_wl_class;
2849 cfqd->workload_expires = jiffies - 1;
2851 choose_wl_class_and_type(cfqd, cfqg);
2855 * Select a queue for service. If we have a current active queue,
2856 * check whether to continue servicing it, or retrieve and set a new one.
2858 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
2860 struct cfq_queue *cfqq, *new_cfqq = NULL;
2862 cfqq = cfqd->active_queue;
2866 if (!cfqd->rq_queued)
2870 * We were waiting for group to get backlogged. Expire the queue
2872 if (cfq_cfqq_wait_busy(cfqq) && !RB_EMPTY_ROOT(&cfqq->sort_list))
2876 * The active queue has run out of time, expire it and select new.
2878 if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq)) {
2880 * If slice had not expired at the completion of last request
2881 * we might not have turned on wait_busy flag. Don't expire
2882 * the queue yet. Allow the group to get backlogged.
2884 * The very fact that we have used the slice, that means we
2885 * have been idling all along on this queue and it should be
2886 * ok to wait for this request to complete.
2888 if (cfqq->cfqg->nr_cfqq == 1 && RB_EMPTY_ROOT(&cfqq->sort_list)
2889 && cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2893 goto check_group_idle;
2897 * The active queue has requests and isn't expired, allow it to
2900 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
2904 * If another queue has a request waiting within our mean seek
2905 * distance, let it run. The expire code will check for close
2906 * cooperators and put the close queue at the front of the service
2907 * tree. If possible, merge the expiring queue with the new cfqq.
2909 new_cfqq = cfq_close_cooperator(cfqd, cfqq);
2911 if (!cfqq->new_cfqq)
2912 cfq_setup_merge(cfqq, new_cfqq);
2917 * No requests pending. If the active queue still has requests in
2918 * flight or is idling for a new request, allow either of these
2919 * conditions to happen (or time out) before selecting a new queue.
2921 if (timer_pending(&cfqd->idle_slice_timer)) {
2927 * This is a deep seek queue, but the device is much faster than
2928 * the queue can deliver, don't idle
2930 if (CFQQ_SEEKY(cfqq) && cfq_cfqq_idle_window(cfqq) &&
2931 (cfq_cfqq_slice_new(cfqq) ||
2932 (cfqq->slice_end - jiffies > jiffies - cfqq->slice_start))) {
2933 cfq_clear_cfqq_deep(cfqq);
2934 cfq_clear_cfqq_idle_window(cfqq);
2937 if (cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2943 * If group idle is enabled and there are requests dispatched from
2944 * this group, wait for requests to complete.
2947 if (cfqd->cfq_group_idle && cfqq->cfqg->nr_cfqq == 1 &&
2948 cfqq->cfqg->dispatched &&
2949 !cfq_io_thinktime_big(cfqd, &cfqq->cfqg->ttime, true)) {
2955 cfq_slice_expired(cfqd, 0);
2958 * Current queue expired. Check if we have to switch to a new
2962 cfq_choose_cfqg(cfqd);
2964 cfqq = cfq_set_active_queue(cfqd, new_cfqq);
2969 static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
2973 while (cfqq->next_rq) {
2974 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
2978 BUG_ON(!list_empty(&cfqq->fifo));
2980 /* By default cfqq is not expired if it is empty. Do it explicitly */
2981 __cfq_slice_expired(cfqq->cfqd, cfqq, 0);
2986 * Drain our current requests. Used for barriers and when switching
2987 * io schedulers on-the-fly.
2989 static int cfq_forced_dispatch(struct cfq_data *cfqd)
2991 struct cfq_queue *cfqq;
2994 /* Expire the timeslice of the current active queue first */
2995 cfq_slice_expired(cfqd, 0);
2996 while ((cfqq = cfq_get_next_queue_forced(cfqd)) != NULL) {
2997 __cfq_set_active_queue(cfqd, cfqq);
2998 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
3001 BUG_ON(cfqd->busy_queues);
3003 cfq_log(cfqd, "forced_dispatch=%d", dispatched);
3007 static inline bool cfq_slice_used_soon(struct cfq_data *cfqd,
3008 struct cfq_queue *cfqq)
3010 /* the queue hasn't finished any request, can't estimate */
3011 if (cfq_cfqq_slice_new(cfqq))
3013 if (time_after(jiffies + cfqd->cfq_slice_idle * cfqq->dispatched,
3020 static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3022 unsigned int max_dispatch;
3025 * Drain async requests before we start sync IO
3027 if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_flight[BLK_RW_ASYNC])
3031 * If this is an async queue and we have sync IO in flight, let it wait
3033 if (cfqd->rq_in_flight[BLK_RW_SYNC] && !cfq_cfqq_sync(cfqq))
3036 max_dispatch = max_t(unsigned int, cfqd->cfq_quantum / 2, 1);
3037 if (cfq_class_idle(cfqq))
3041 * Does this cfqq already have too much IO in flight?
3043 if (cfqq->dispatched >= max_dispatch) {
3044 bool promote_sync = false;
3046 * idle queue must always only have a single IO in flight
3048 if (cfq_class_idle(cfqq))
3052 * If there is only one sync queue
3053 * we can ignore async queue here and give the sync
3054 * queue no dispatch limit. The reason is a sync queue can
3055 * preempt async queue, limiting the sync queue doesn't make
3056 * sense. This is useful for aiostress test.
3058 if (cfq_cfqq_sync(cfqq) && cfqd->busy_sync_queues == 1)
3059 promote_sync = true;
3062 * We have other queues, don't allow more IO from this one
3064 if (cfqd->busy_queues > 1 && cfq_slice_used_soon(cfqd, cfqq) &&
3069 * Sole queue user, no limit
3071 if (cfqd->busy_queues == 1 || promote_sync)
3075 * Normally we start throttling cfqq when cfq_quantum/2
3076 * requests have been dispatched. But we can drive
3077 * deeper queue depths at the beginning of slice
3078 * subjected to upper limit of cfq_quantum.
3080 max_dispatch = cfqd->cfq_quantum;
3084 * Async queues must wait a bit before being allowed dispatch.
3085 * We also ramp up the dispatch depth gradually for async IO,
3086 * based on the last sync IO we serviced
3088 if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
3089 unsigned long last_sync = jiffies - cfqd->last_delayed_sync;
3092 depth = last_sync / cfqd->cfq_slice[1];
3093 if (!depth && !cfqq->dispatched)
3095 if (depth < max_dispatch)
3096 max_dispatch = depth;
3100 * If we're below the current max, allow a dispatch
3102 return cfqq->dispatched < max_dispatch;
3106 * Dispatch a request from cfqq, moving them to the request queue
3109 static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3113 BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
3115 if (!cfq_may_dispatch(cfqd, cfqq))
3119 * follow expired path, else get first next available
3121 rq = cfq_check_fifo(cfqq);
3126 * insert request into driver dispatch list
3128 cfq_dispatch_insert(cfqd->queue, rq);
3130 if (!cfqd->active_cic) {
3131 struct cfq_io_cq *cic = RQ_CIC(rq);
3133 atomic_long_inc(&cic->icq.ioc->refcount);
3134 cfqd->active_cic = cic;
3141 * Find the cfqq that we need to service and move a request from that to the
3144 static int cfq_dispatch_requests(struct request_queue *q, int force)
3146 struct cfq_data *cfqd = q->elevator->elevator_data;
3147 struct cfq_queue *cfqq;
3149 if (!cfqd->busy_queues)
3152 if (unlikely(force))
3153 return cfq_forced_dispatch(cfqd);
3155 cfqq = cfq_select_queue(cfqd);
3160 * Dispatch a request from this cfqq, if it is allowed
3162 if (!cfq_dispatch_request(cfqd, cfqq))
3165 cfqq->slice_dispatch++;
3166 cfq_clear_cfqq_must_dispatch(cfqq);
3169 * expire an async queue immediately if it has used up its slice. idle
3170 * queue always expire after 1 dispatch round.
3172 if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
3173 cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
3174 cfq_class_idle(cfqq))) {
3175 cfqq->slice_end = jiffies + 1;
3176 cfq_slice_expired(cfqd, 0);
3179 cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
3184 * task holds one reference to the queue, dropped when task exits. each rq
3185 * in-flight on this queue also holds a reference, dropped when rq is freed.
3187 * Each cfq queue took a reference on the parent group. Drop it now.
3188 * queue lock must be held here.
3190 static void cfq_put_queue(struct cfq_queue *cfqq)
3192 struct cfq_data *cfqd = cfqq->cfqd;
3193 struct cfq_group *cfqg;
3195 BUG_ON(cfqq->ref <= 0);
3201 cfq_log_cfqq(cfqd, cfqq, "put_queue");
3202 BUG_ON(rb_first(&cfqq->sort_list));
3203 BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
3206 if (unlikely(cfqd->active_queue == cfqq)) {
3207 __cfq_slice_expired(cfqd, cfqq, 0);
3208 cfq_schedule_dispatch(cfqd);
3211 BUG_ON(cfq_cfqq_on_rr(cfqq));
3212 kmem_cache_free(cfq_pool, cfqq);
3216 static void cfq_put_cooperator(struct cfq_queue *cfqq)
3218 struct cfq_queue *__cfqq, *next;
3221 * If this queue was scheduled to merge with another queue, be
3222 * sure to drop the reference taken on that queue (and others in
3223 * the merge chain). See cfq_setup_merge and cfq_merge_cfqqs.
3225 __cfqq = cfqq->new_cfqq;
3227 if (__cfqq == cfqq) {
3228 WARN(1, "cfqq->new_cfqq loop detected\n");
3231 next = __cfqq->new_cfqq;
3232 cfq_put_queue(__cfqq);
3237 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3239 if (unlikely(cfqq == cfqd->active_queue)) {
3240 __cfq_slice_expired(cfqd, cfqq, 0);
3241 cfq_schedule_dispatch(cfqd);
3244 cfq_put_cooperator(cfqq);
3246 cfq_put_queue(cfqq);
3249 static void cfq_init_icq(struct io_cq *icq)
3251 struct cfq_io_cq *cic = icq_to_cic(icq);
3253 cic->ttime.last_end_request = jiffies;
3256 static void cfq_exit_icq(struct io_cq *icq)
3258 struct cfq_io_cq *cic = icq_to_cic(icq);
3259 struct cfq_data *cfqd = cic_to_cfqd(cic);
3261 if (cic->cfqq[BLK_RW_ASYNC]) {
3262 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]);
3263 cic->cfqq[BLK_RW_ASYNC] = NULL;
3266 if (cic->cfqq[BLK_RW_SYNC]) {
3267 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]);
3268 cic->cfqq[BLK_RW_SYNC] = NULL;
3272 static void cfq_init_prio_data(struct cfq_queue *cfqq, struct cfq_io_cq *cic)
3274 struct task_struct *tsk = current;
3277 if (!cfq_cfqq_prio_changed(cfqq))
3280 ioprio_class = IOPRIO_PRIO_CLASS(cic->ioprio);
3281 switch (ioprio_class) {
3283 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
3284 case IOPRIO_CLASS_NONE:
3286 * no prio set, inherit CPU scheduling settings
3288 cfqq->ioprio = task_nice_ioprio(tsk);
3289 cfqq->ioprio_class = task_nice_ioclass(tsk);
3291 case IOPRIO_CLASS_RT:
3292 cfqq->ioprio = IOPRIO_PRIO_DATA(cic->ioprio);
3293 cfqq->ioprio_class = IOPRIO_CLASS_RT;
3295 case IOPRIO_CLASS_BE:
3296 cfqq->ioprio = IOPRIO_PRIO_DATA(cic->ioprio);
3297 cfqq->ioprio_class = IOPRIO_CLASS_BE;
3299 case IOPRIO_CLASS_IDLE:
3300 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
3302 cfq_clear_cfqq_idle_window(cfqq);
3307 * keep track of original prio settings in case we have to temporarily
3308 * elevate the priority of this queue
3310 cfqq->org_ioprio = cfqq->ioprio;
3311 cfq_clear_cfqq_prio_changed(cfqq);
3314 static void check_ioprio_changed(struct cfq_io_cq *cic, struct bio *bio)
3316 int ioprio = cic->icq.ioc->ioprio;
3317 struct cfq_data *cfqd = cic_to_cfqd(cic);
3318 struct cfq_queue *cfqq;
3321 * Check whether ioprio has changed. The condition may trigger
3322 * spuriously on a newly created cic but there's no harm.
3324 if (unlikely(!cfqd) || likely(cic->ioprio == ioprio))
3327 cfqq = cic->cfqq[BLK_RW_ASYNC];
3329 struct cfq_queue *new_cfqq;
3330 new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic, bio,
3333 cic->cfqq[BLK_RW_ASYNC] = new_cfqq;
3334 cfq_put_queue(cfqq);
3338 cfqq = cic->cfqq[BLK_RW_SYNC];
3340 cfq_mark_cfqq_prio_changed(cfqq);
3342 cic->ioprio = ioprio;
3345 static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3346 pid_t pid, bool is_sync)
3348 RB_CLEAR_NODE(&cfqq->rb_node);
3349 RB_CLEAR_NODE(&cfqq->p_node);
3350 INIT_LIST_HEAD(&cfqq->fifo);
3355 cfq_mark_cfqq_prio_changed(cfqq);
3358 if (!cfq_class_idle(cfqq))
3359 cfq_mark_cfqq_idle_window(cfqq);
3360 cfq_mark_cfqq_sync(cfqq);
3365 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3366 static void check_blkcg_changed(struct cfq_io_cq *cic, struct bio *bio)
3368 struct cfq_data *cfqd = cic_to_cfqd(cic);
3369 struct cfq_queue *sync_cfqq;
3373 id = bio_blkcg(bio)->id;
3377 * Check whether blkcg has changed. The condition may trigger
3378 * spuriously on a newly created cic but there's no harm.
3380 if (unlikely(!cfqd) || likely(cic->blkcg_id == id))
3383 sync_cfqq = cic_to_cfqq(cic, 1);
3386 * Drop reference to sync queue. A new sync queue will be
3387 * assigned in new group upon arrival of a fresh request.
3389 cfq_log_cfqq(cfqd, sync_cfqq, "changed cgroup");
3390 cic_set_cfqq(cic, NULL, 1);
3391 cfq_put_queue(sync_cfqq);
3397 static inline void check_blkcg_changed(struct cfq_io_cq *cic, struct bio *bio) { }
3398 #endif /* CONFIG_CFQ_GROUP_IOSCHED */
3400 static struct cfq_queue *
3401 cfq_find_alloc_queue(struct cfq_data *cfqd, bool is_sync, struct cfq_io_cq *cic,
3402 struct bio *bio, gfp_t gfp_mask)
3404 struct blkcg *blkcg;
3405 struct cfq_queue *cfqq, *new_cfqq = NULL;
3406 struct cfq_group *cfqg;
3411 blkcg = bio_blkcg(bio);
3412 cfqg = cfq_lookup_create_cfqg(cfqd, blkcg);
3413 cfqq = cic_to_cfqq(cic, is_sync);
3416 * Always try a new alloc if we fell back to the OOM cfqq
3417 * originally, since it should just be a temporary situation.
3419 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
3424 } else if (gfp_mask & __GFP_WAIT) {
3426 spin_unlock_irq(cfqd->queue->queue_lock);
3427 new_cfqq = kmem_cache_alloc_node(cfq_pool,
3428 gfp_mask | __GFP_ZERO,
3430 spin_lock_irq(cfqd->queue->queue_lock);
3434 cfqq = kmem_cache_alloc_node(cfq_pool,
3435 gfp_mask | __GFP_ZERO,
3440 cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
3441 cfq_init_prio_data(cfqq, cic);
3442 cfq_link_cfqq_cfqg(cfqq, cfqg);
3443 cfq_log_cfqq(cfqd, cfqq, "alloced");
3445 cfqq = &cfqd->oom_cfqq;
3449 kmem_cache_free(cfq_pool, new_cfqq);
3455 static struct cfq_queue **
3456 cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
3458 switch (ioprio_class) {
3459 case IOPRIO_CLASS_RT:
3460 return &cfqd->async_cfqq[0][ioprio];
3461 case IOPRIO_CLASS_NONE:
3462 ioprio = IOPRIO_NORM;
3464 case IOPRIO_CLASS_BE:
3465 return &cfqd->async_cfqq[1][ioprio];
3466 case IOPRIO_CLASS_IDLE:
3467 return &cfqd->async_idle_cfqq;
3473 static struct cfq_queue *
3474 cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct cfq_io_cq *cic,
3475 struct bio *bio, gfp_t gfp_mask)
3477 const int ioprio_class = IOPRIO_PRIO_CLASS(cic->ioprio);
3478 const int ioprio = IOPRIO_PRIO_DATA(cic->ioprio);
3479 struct cfq_queue **async_cfqq = NULL;
3480 struct cfq_queue *cfqq = NULL;
3483 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
3488 cfqq = cfq_find_alloc_queue(cfqd, is_sync, cic, bio, gfp_mask);
3491 * pin the queue now that it's allocated, scheduler exit will prune it
3493 if (!is_sync && !(*async_cfqq)) {
3503 __cfq_update_io_thinktime(struct cfq_ttime *ttime, unsigned long slice_idle)
3505 unsigned long elapsed = jiffies - ttime->last_end_request;
3506 elapsed = min(elapsed, 2UL * slice_idle);
3508 ttime->ttime_samples = (7*ttime->ttime_samples + 256) / 8;
3509 ttime->ttime_total = (7*ttime->ttime_total + 256*elapsed) / 8;
3510 ttime->ttime_mean = (ttime->ttime_total + 128) / ttime->ttime_samples;
3514 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3515 struct cfq_io_cq *cic)
3517 if (cfq_cfqq_sync(cfqq)) {
3518 __cfq_update_io_thinktime(&cic->ttime, cfqd->cfq_slice_idle);
3519 __cfq_update_io_thinktime(&cfqq->service_tree->ttime,
3520 cfqd->cfq_slice_idle);
3522 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3523 __cfq_update_io_thinktime(&cfqq->cfqg->ttime, cfqd->cfq_group_idle);
3528 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3532 sector_t n_sec = blk_rq_sectors(rq);
3533 if (cfqq->last_request_pos) {
3534 if (cfqq->last_request_pos < blk_rq_pos(rq))
3535 sdist = blk_rq_pos(rq) - cfqq->last_request_pos;
3537 sdist = cfqq->last_request_pos - blk_rq_pos(rq);
3540 cfqq->seek_history <<= 1;
3541 if (blk_queue_nonrot(cfqd->queue))
3542 cfqq->seek_history |= (n_sec < CFQQ_SECT_THR_NONROT);
3544 cfqq->seek_history |= (sdist > CFQQ_SEEK_THR);
3548 * Disable idle window if the process thinks too long or seeks so much that
3552 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3553 struct cfq_io_cq *cic)
3555 int old_idle, enable_idle;
3558 * Don't idle for async or idle io prio class
3560 if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
3563 enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
3565 if (cfqq->queued[0] + cfqq->queued[1] >= 4)
3566 cfq_mark_cfqq_deep(cfqq);
3568 if (cfqq->next_rq && (cfqq->next_rq->cmd_flags & REQ_NOIDLE))
3570 else if (!atomic_read(&cic->icq.ioc->active_ref) ||
3571 !cfqd->cfq_slice_idle ||
3572 (!cfq_cfqq_deep(cfqq) && CFQQ_SEEKY(cfqq)))
3574 else if (sample_valid(cic->ttime.ttime_samples)) {
3575 if (cic->ttime.ttime_mean > cfqd->cfq_slice_idle)
3581 if (old_idle != enable_idle) {
3582 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
3584 cfq_mark_cfqq_idle_window(cfqq);
3586 cfq_clear_cfqq_idle_window(cfqq);
3591 * Check if new_cfqq should preempt the currently active queue. Return 0 for
3592 * no or if we aren't sure, a 1 will cause a preempt.
3595 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
3598 struct cfq_queue *cfqq;
3600 cfqq = cfqd->active_queue;
3604 if (cfq_class_idle(new_cfqq))
3607 if (cfq_class_idle(cfqq))
3611 * Don't allow a non-RT request to preempt an ongoing RT cfqq timeslice.
3613 if (cfq_class_rt(cfqq) && !cfq_class_rt(new_cfqq))
3617 * if the new request is sync, but the currently running queue is
3618 * not, let the sync request have priority.
3620 if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
3623 if (new_cfqq->cfqg != cfqq->cfqg)
3626 if (cfq_slice_used(cfqq))
3629 /* Allow preemption only if we are idling on sync-noidle tree */
3630 if (cfqd->serving_wl_type == SYNC_NOIDLE_WORKLOAD &&
3631 cfqq_type(new_cfqq) == SYNC_NOIDLE_WORKLOAD &&
3632 new_cfqq->service_tree->count == 2 &&
3633 RB_EMPTY_ROOT(&cfqq->sort_list))
3637 * So both queues are sync. Let the new request get disk time if
3638 * it's a metadata request and the current queue is doing regular IO.
3640 if ((rq->cmd_flags & REQ_PRIO) && !cfqq->prio_pending)
3644 * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
3646 if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
3649 /* An idle queue should not be idle now for some reason */
3650 if (RB_EMPTY_ROOT(&cfqq->sort_list) && !cfq_should_idle(cfqd, cfqq))
3653 if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
3657 * if this request is as-good as one we would expect from the
3658 * current cfqq, let it preempt
3660 if (cfq_rq_close(cfqd, cfqq, rq))
3667 * cfqq preempts the active queue. if we allowed preempt with no slice left,
3668 * let it have half of its nominal slice.
3670 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3672 enum wl_type_t old_type = cfqq_type(cfqd->active_queue);
3674 cfq_log_cfqq(cfqd, cfqq, "preempt");
3675 cfq_slice_expired(cfqd, 1);
3678 * workload type is changed, don't save slice, otherwise preempt
3681 if (old_type != cfqq_type(cfqq))
3682 cfqq->cfqg->saved_wl_slice = 0;
3685 * Put the new queue at the front of the of the current list,
3686 * so we know that it will be selected next.
3688 BUG_ON(!cfq_cfqq_on_rr(cfqq));
3690 cfq_service_tree_add(cfqd, cfqq, 1);
3692 cfqq->slice_end = 0;
3693 cfq_mark_cfqq_slice_new(cfqq);
3697 * Called when a new fs request (rq) is added (to cfqq). Check if there's
3698 * something we should do about it
3701 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3704 struct cfq_io_cq *cic = RQ_CIC(rq);
3707 if (rq->cmd_flags & REQ_PRIO)
3708 cfqq->prio_pending++;
3710 cfq_update_io_thinktime(cfqd, cfqq, cic);
3711 cfq_update_io_seektime(cfqd, cfqq, rq);
3712 cfq_update_idle_window(cfqd, cfqq, cic);
3714 cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
3716 if (cfqq == cfqd->active_queue) {
3718 * Remember that we saw a request from this process, but
3719 * don't start queuing just yet. Otherwise we risk seeing lots
3720 * of tiny requests, because we disrupt the normal plugging
3721 * and merging. If the request is already larger than a single
3722 * page, let it rip immediately. For that case we assume that
3723 * merging is already done. Ditto for a busy system that
3724 * has other work pending, don't risk delaying until the
3725 * idle timer unplug to continue working.
3727 if (cfq_cfqq_wait_request(cfqq)) {
3728 if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
3729 cfqd->busy_queues > 1) {
3730 cfq_del_timer(cfqd, cfqq);
3731 cfq_clear_cfqq_wait_request(cfqq);
3732 __blk_run_queue(cfqd->queue);
3734 cfqg_stats_update_idle_time(cfqq->cfqg);
3735 cfq_mark_cfqq_must_dispatch(cfqq);
3738 } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
3740 * not the active queue - expire current slice if it is
3741 * idle and has expired it's mean thinktime or this new queue
3742 * has some old slice time left and is of higher priority or
3743 * this new queue is RT and the current one is BE
3745 cfq_preempt_queue(cfqd, cfqq);
3746 __blk_run_queue(cfqd->queue);
3750 static void cfq_insert_request(struct request_queue *q, struct request *rq)
3752 struct cfq_data *cfqd = q->elevator->elevator_data;
3753 struct cfq_queue *cfqq = RQ_CFQQ(rq);
3755 cfq_log_cfqq(cfqd, cfqq, "insert_request");
3756 cfq_init_prio_data(cfqq, RQ_CIC(rq));
3758 rq_set_fifo_time(rq, jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)]);
3759 list_add_tail(&rq->queuelist, &cfqq->fifo);
3761 cfqg_stats_update_io_add(RQ_CFQG(rq), cfqd->serving_group,
3763 cfq_rq_enqueued(cfqd, cfqq, rq);
3767 * Update hw_tag based on peak queue depth over 50 samples under
3770 static void cfq_update_hw_tag(struct cfq_data *cfqd)
3772 struct cfq_queue *cfqq = cfqd->active_queue;
3774 if (cfqd->rq_in_driver > cfqd->hw_tag_est_depth)
3775 cfqd->hw_tag_est_depth = cfqd->rq_in_driver;
3777 if (cfqd->hw_tag == 1)
3780 if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
3781 cfqd->rq_in_driver <= CFQ_HW_QUEUE_MIN)
3785 * If active queue hasn't enough requests and can idle, cfq might not
3786 * dispatch sufficient requests to hardware. Don't zero hw_tag in this
3789 if (cfqq && cfq_cfqq_idle_window(cfqq) &&
3790 cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] <
3791 CFQ_HW_QUEUE_MIN && cfqd->rq_in_driver < CFQ_HW_QUEUE_MIN)
3794 if (cfqd->hw_tag_samples++ < 50)
3797 if (cfqd->hw_tag_est_depth >= CFQ_HW_QUEUE_MIN)
3803 static bool cfq_should_wait_busy(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3805 struct cfq_io_cq *cic = cfqd->active_cic;
3807 /* If the queue already has requests, don't wait */
3808 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3811 /* If there are other queues in the group, don't wait */
3812 if (cfqq->cfqg->nr_cfqq > 1)
3815 /* the only queue in the group, but think time is big */
3816 if (cfq_io_thinktime_big(cfqd, &cfqq->cfqg->ttime, true))
3819 if (cfq_slice_used(cfqq))
3822 /* if slice left is less than think time, wait busy */
3823 if (cic && sample_valid(cic->ttime.ttime_samples)
3824 && (cfqq->slice_end - jiffies < cic->ttime.ttime_mean))
3828 * If think times is less than a jiffy than ttime_mean=0 and above
3829 * will not be true. It might happen that slice has not expired yet
3830 * but will expire soon (4-5 ns) during select_queue(). To cover the
3831 * case where think time is less than a jiffy, mark the queue wait
3832 * busy if only 1 jiffy is left in the slice.
3834 if (cfqq->slice_end - jiffies == 1)
3840 static void cfq_completed_request(struct request_queue *q, struct request *rq)
3842 struct cfq_queue *cfqq = RQ_CFQQ(rq);
3843 struct cfq_data *cfqd = cfqq->cfqd;
3844 const int sync = rq_is_sync(rq);
3848 cfq_log_cfqq(cfqd, cfqq, "complete rqnoidle %d",
3849 !!(rq->cmd_flags & REQ_NOIDLE));
3851 cfq_update_hw_tag(cfqd);
3853 WARN_ON(!cfqd->rq_in_driver);
3854 WARN_ON(!cfqq->dispatched);
3855 cfqd->rq_in_driver--;
3857 (RQ_CFQG(rq))->dispatched--;
3858 cfqg_stats_update_completion(cfqq->cfqg, rq_start_time_ns(rq),
3859 rq_io_start_time_ns(rq), rq->cmd_flags);
3861 cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]--;
3864 struct cfq_rb_root *st;
3866 RQ_CIC(rq)->ttime.last_end_request = now;
3868 if (cfq_cfqq_on_rr(cfqq))
3869 st = cfqq->service_tree;
3871 st = st_for(cfqq->cfqg, cfqq_class(cfqq),
3874 st->ttime.last_end_request = now;
3875 if (!time_after(rq->start_time + cfqd->cfq_fifo_expire[1], now))
3876 cfqd->last_delayed_sync = now;
3879 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3880 cfqq->cfqg->ttime.last_end_request = now;
3884 * If this is the active queue, check if it needs to be expired,
3885 * or if we want to idle in case it has no pending requests.
3887 if (cfqd->active_queue == cfqq) {
3888 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
3890 if (cfq_cfqq_slice_new(cfqq)) {
3891 cfq_set_prio_slice(cfqd, cfqq);
3892 cfq_clear_cfqq_slice_new(cfqq);
3896 * Should we wait for next request to come in before we expire
3899 if (cfq_should_wait_busy(cfqd, cfqq)) {
3900 unsigned long extend_sl = cfqd->cfq_slice_idle;
3901 if (!cfqd->cfq_slice_idle)
3902 extend_sl = cfqd->cfq_group_idle;
3903 cfqq->slice_end = jiffies + extend_sl;
3904 cfq_mark_cfqq_wait_busy(cfqq);
3905 cfq_log_cfqq(cfqd, cfqq, "will busy wait");
3909 * Idling is not enabled on:
3911 * - idle-priority queues
3913 * - queues with still some requests queued
3914 * - when there is a close cooperator
3916 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
3917 cfq_slice_expired(cfqd, 1);
3918 else if (sync && cfqq_empty &&
3919 !cfq_close_cooperator(cfqd, cfqq)) {
3920 cfq_arm_slice_timer(cfqd);
3924 if (!cfqd->rq_in_driver)
3925 cfq_schedule_dispatch(cfqd);
3928 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
3930 if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
3931 cfq_mark_cfqq_must_alloc_slice(cfqq);
3932 return ELV_MQUEUE_MUST;
3935 return ELV_MQUEUE_MAY;
3938 static int cfq_may_queue(struct request_queue *q, int rw)
3940 struct cfq_data *cfqd = q->elevator->elevator_data;
3941 struct task_struct *tsk = current;
3942 struct cfq_io_cq *cic;
3943 struct cfq_queue *cfqq;
3946 * don't force setup of a queue from here, as a call to may_queue
3947 * does not necessarily imply that a request actually will be queued.
3948 * so just lookup a possibly existing queue, or return 'may queue'
3951 cic = cfq_cic_lookup(cfqd, tsk->io_context);
3953 return ELV_MQUEUE_MAY;
3955 cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
3957 cfq_init_prio_data(cfqq, cic);
3959 return __cfq_may_queue(cfqq);
3962 return ELV_MQUEUE_MAY;
3966 * queue lock held here
3968 static void cfq_put_request(struct request *rq)
3970 struct cfq_queue *cfqq = RQ_CFQQ(rq);
3973 const int rw = rq_data_dir(rq);
3975 BUG_ON(!cfqq->allocated[rw]);
3976 cfqq->allocated[rw]--;
3978 /* Put down rq reference on cfqg */
3979 cfqg_put(RQ_CFQG(rq));
3980 rq->elv.priv[0] = NULL;
3981 rq->elv.priv[1] = NULL;
3983 cfq_put_queue(cfqq);
3987 static struct cfq_queue *
3988 cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_cq *cic,
3989 struct cfq_queue *cfqq)
3991 cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq);
3992 cic_set_cfqq(cic, cfqq->new_cfqq, 1);
3993 cfq_mark_cfqq_coop(cfqq->new_cfqq);
3994 cfq_put_queue(cfqq);
3995 return cic_to_cfqq(cic, 1);
3999 * Returns NULL if a new cfqq should be allocated, or the old cfqq if this
4000 * was the last process referring to said cfqq.
4002 static struct cfq_queue *
4003 split_cfqq(struct cfq_io_cq *cic, struct cfq_queue *cfqq)
4005 if (cfqq_process_refs(cfqq) == 1) {
4006 cfqq->pid = current->pid;
4007 cfq_clear_cfqq_coop(cfqq);
4008 cfq_clear_cfqq_split_coop(cfqq);
4012 cic_set_cfqq(cic, NULL, 1);
4014 cfq_put_cooperator(cfqq);
4016 cfq_put_queue(cfqq);
4020 * Allocate cfq data structures associated with this request.
4023 cfq_set_request(struct request_queue *q, struct request *rq, struct bio *bio,
4026 struct cfq_data *cfqd = q->elevator->elevator_data;
4027 struct cfq_io_cq *cic = icq_to_cic(rq->elv.icq);
4028 const int rw = rq_data_dir(rq);
4029 const bool is_sync = rq_is_sync(rq);
4030 struct cfq_queue *cfqq;
4032 might_sleep_if(gfp_mask & __GFP_WAIT);
4034 spin_lock_irq(q->queue_lock);
4036 check_ioprio_changed(cic, bio);
4037 check_blkcg_changed(cic, bio);
4039 cfqq = cic_to_cfqq(cic, is_sync);
4040 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
4041 cfqq = cfq_get_queue(cfqd, is_sync, cic, bio, gfp_mask);
4042 cic_set_cfqq(cic, cfqq, is_sync);
4045 * If the queue was seeky for too long, break it apart.
4047 if (cfq_cfqq_coop(cfqq) && cfq_cfqq_split_coop(cfqq)) {
4048 cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq");
4049 cfqq = split_cfqq(cic, cfqq);
4055 * Check to see if this queue is scheduled to merge with
4056 * another, closely cooperating queue. The merging of
4057 * queues happens here as it must be done in process context.
4058 * The reference on new_cfqq was taken in merge_cfqqs.
4061 cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq);
4064 cfqq->allocated[rw]++;
4067 cfqg_get(cfqq->cfqg);
4068 rq->elv.priv[0] = cfqq;
4069 rq->elv.priv[1] = cfqq->cfqg;
4070 spin_unlock_irq(q->queue_lock);
4074 static void cfq_kick_queue(struct work_struct *work)
4076 struct cfq_data *cfqd =
4077 container_of(work, struct cfq_data, unplug_work);
4078 struct request_queue *q = cfqd->queue;
4080 spin_lock_irq(q->queue_lock);
4081 __blk_run_queue(cfqd->queue);
4082 spin_unlock_irq(q->queue_lock);
4086 * Timer running if the active_queue is currently idling inside its time slice
4088 static void cfq_idle_slice_timer(unsigned long data)
4090 struct cfq_data *cfqd = (struct cfq_data *) data;
4091 struct cfq_queue *cfqq;
4092 unsigned long flags;
4095 cfq_log(cfqd, "idle timer fired");
4097 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
4099 cfqq = cfqd->active_queue;
4104 * We saw a request before the queue expired, let it through
4106 if (cfq_cfqq_must_dispatch(cfqq))
4112 if (cfq_slice_used(cfqq))
4116 * only expire and reinvoke request handler, if there are
4117 * other queues with pending requests
4119 if (!cfqd->busy_queues)
4123 * not expired and it has a request pending, let it dispatch
4125 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
4129 * Queue depth flag is reset only when the idle didn't succeed
4131 cfq_clear_cfqq_deep(cfqq);
4134 cfq_slice_expired(cfqd, timed_out);
4136 cfq_schedule_dispatch(cfqd);
4138 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
4141 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
4143 del_timer_sync(&cfqd->idle_slice_timer);
4144 cancel_work_sync(&cfqd->unplug_work);
4147 static void cfq_put_async_queues(struct cfq_data *cfqd)
4151 for (i = 0; i < IOPRIO_BE_NR; i++) {
4152 if (cfqd->async_cfqq[0][i])
4153 cfq_put_queue(cfqd->async_cfqq[0][i]);
4154 if (cfqd->async_cfqq[1][i])
4155 cfq_put_queue(cfqd->async_cfqq[1][i]);
4158 if (cfqd->async_idle_cfqq)
4159 cfq_put_queue(cfqd->async_idle_cfqq);
4162 static void cfq_exit_queue(struct elevator_queue *e)
4164 struct cfq_data *cfqd = e->elevator_data;
4165 struct request_queue *q = cfqd->queue;
4167 cfq_shutdown_timer_wq(cfqd);
4169 spin_lock_irq(q->queue_lock);
4171 if (cfqd->active_queue)
4172 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
4174 cfq_put_async_queues(cfqd);
4176 spin_unlock_irq(q->queue_lock);
4178 cfq_shutdown_timer_wq(cfqd);
4180 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4181 blkcg_deactivate_policy(q, &blkcg_policy_cfq);
4183 kfree(cfqd->root_group);
4188 static int cfq_init_queue(struct request_queue *q)
4190 struct cfq_data *cfqd;
4191 struct blkcg_gq *blkg __maybe_unused;
4194 cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
4199 q->elevator->elevator_data = cfqd;
4201 /* Init root service tree */
4202 cfqd->grp_service_tree = CFQ_RB_ROOT;
4204 /* Init root group and prefer root group over other groups by default */
4205 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4206 ret = blkcg_activate_policy(q, &blkcg_policy_cfq);
4210 cfqd->root_group = blkg_to_cfqg(q->root_blkg);
4213 cfqd->root_group = kzalloc_node(sizeof(*cfqd->root_group),
4214 GFP_KERNEL, cfqd->queue->node);
4215 if (!cfqd->root_group)
4218 cfq_init_cfqg_base(cfqd->root_group);
4220 cfqd->root_group->weight = 2 * CFQ_WEIGHT_DEFAULT;
4221 cfqd->root_group->leaf_weight = 2 * CFQ_WEIGHT_DEFAULT;
4224 * Not strictly needed (since RB_ROOT just clears the node and we
4225 * zeroed cfqd on alloc), but better be safe in case someone decides
4226 * to add magic to the rb code
4228 for (i = 0; i < CFQ_PRIO_LISTS; i++)
4229 cfqd->prio_trees[i] = RB_ROOT;
4232 * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
4233 * Grab a permanent reference to it, so that the normal code flow
4234 * will not attempt to free it. oom_cfqq is linked to root_group
4235 * but shouldn't hold a reference as it'll never be unlinked. Lose
4236 * the reference from linking right away.
4238 cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
4239 cfqd->oom_cfqq.ref++;
4241 spin_lock_irq(q->queue_lock);
4242 cfq_link_cfqq_cfqg(&cfqd->oom_cfqq, cfqd->root_group);
4243 cfqg_put(cfqd->root_group);
4244 spin_unlock_irq(q->queue_lock);
4246 init_timer(&cfqd->idle_slice_timer);
4247 cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
4248 cfqd->idle_slice_timer.data = (unsigned long) cfqd;
4250 INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
4252 cfqd->cfq_quantum = cfq_quantum;
4253 cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
4254 cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
4255 cfqd->cfq_back_max = cfq_back_max;
4256 cfqd->cfq_back_penalty = cfq_back_penalty;
4257 cfqd->cfq_slice[0] = cfq_slice_async;
4258 cfqd->cfq_slice[1] = cfq_slice_sync;
4259 cfqd->cfq_target_latency = cfq_target_latency;
4260 cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
4261 cfqd->cfq_slice_idle = cfq_slice_idle;
4262 cfqd->cfq_group_idle = cfq_group_idle;
4263 cfqd->cfq_latency = 1;
4266 * we optimistically start assuming sync ops weren't delayed in last
4267 * second, in order to have larger depth for async operations.
4269 cfqd->last_delayed_sync = jiffies - HZ;
4278 * sysfs parts below -->
4281 cfq_var_show(unsigned int var, char *page)
4283 return sprintf(page, "%d\n", var);
4287 cfq_var_store(unsigned int *var, const char *page, size_t count)
4289 char *p = (char *) page;
4291 *var = simple_strtoul(p, &p, 10);
4295 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
4296 static ssize_t __FUNC(struct elevator_queue *e, char *page) \
4298 struct cfq_data *cfqd = e->elevator_data; \
4299 unsigned int __data = __VAR; \
4301 __data = jiffies_to_msecs(__data); \
4302 return cfq_var_show(__data, (page)); \
4304 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
4305 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
4306 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
4307 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
4308 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
4309 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
4310 SHOW_FUNCTION(cfq_group_idle_show, cfqd->cfq_group_idle, 1);
4311 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
4312 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
4313 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
4314 SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
4315 SHOW_FUNCTION(cfq_target_latency_show, cfqd->cfq_target_latency, 1);
4316 #undef SHOW_FUNCTION
4318 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
4319 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
4321 struct cfq_data *cfqd = e->elevator_data; \
4322 unsigned int __data; \
4323 int ret = cfq_var_store(&__data, (page), count); \
4324 if (__data < (MIN)) \
4326 else if (__data > (MAX)) \
4329 *(__PTR) = msecs_to_jiffies(__data); \
4331 *(__PTR) = __data; \
4334 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
4335 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
4337 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
4339 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
4340 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
4342 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
4343 STORE_FUNCTION(cfq_group_idle_store, &cfqd->cfq_group_idle, 0, UINT_MAX, 1);
4344 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
4345 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
4346 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
4348 STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
4349 STORE_FUNCTION(cfq_target_latency_store, &cfqd->cfq_target_latency, 1, UINT_MAX, 1);
4350 #undef STORE_FUNCTION
4352 #define CFQ_ATTR(name) \
4353 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
4355 static struct elv_fs_entry cfq_attrs[] = {
4357 CFQ_ATTR(fifo_expire_sync),
4358 CFQ_ATTR(fifo_expire_async),
4359 CFQ_ATTR(back_seek_max),
4360 CFQ_ATTR(back_seek_penalty),
4361 CFQ_ATTR(slice_sync),
4362 CFQ_ATTR(slice_async),
4363 CFQ_ATTR(slice_async_rq),
4364 CFQ_ATTR(slice_idle),
4365 CFQ_ATTR(group_idle),
4366 CFQ_ATTR(low_latency),
4367 CFQ_ATTR(target_latency),
4371 static struct elevator_type iosched_cfq = {
4373 .elevator_merge_fn = cfq_merge,
4374 .elevator_merged_fn = cfq_merged_request,
4375 .elevator_merge_req_fn = cfq_merged_requests,
4376 .elevator_allow_merge_fn = cfq_allow_merge,
4377 .elevator_bio_merged_fn = cfq_bio_merged,
4378 .elevator_dispatch_fn = cfq_dispatch_requests,
4379 .elevator_add_req_fn = cfq_insert_request,
4380 .elevator_activate_req_fn = cfq_activate_request,
4381 .elevator_deactivate_req_fn = cfq_deactivate_request,
4382 .elevator_completed_req_fn = cfq_completed_request,
4383 .elevator_former_req_fn = elv_rb_former_request,
4384 .elevator_latter_req_fn = elv_rb_latter_request,
4385 .elevator_init_icq_fn = cfq_init_icq,
4386 .elevator_exit_icq_fn = cfq_exit_icq,
4387 .elevator_set_req_fn = cfq_set_request,
4388 .elevator_put_req_fn = cfq_put_request,
4389 .elevator_may_queue_fn = cfq_may_queue,
4390 .elevator_init_fn = cfq_init_queue,
4391 .elevator_exit_fn = cfq_exit_queue,
4393 .icq_size = sizeof(struct cfq_io_cq),
4394 .icq_align = __alignof__(struct cfq_io_cq),
4395 .elevator_attrs = cfq_attrs,
4396 .elevator_name = "cfq",
4397 .elevator_owner = THIS_MODULE,
4400 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4401 static struct blkcg_policy blkcg_policy_cfq = {
4402 .pd_size = sizeof(struct cfq_group),
4403 .cftypes = cfq_blkcg_files,
4405 .pd_init_fn = cfq_pd_init,
4406 .pd_reset_stats_fn = cfq_pd_reset_stats,
4410 static int __init cfq_init(void)
4415 * could be 0 on HZ < 1000 setups
4417 if (!cfq_slice_async)
4418 cfq_slice_async = 1;
4419 if (!cfq_slice_idle)
4422 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4423 if (!cfq_group_idle)
4426 ret = blkcg_policy_register(&blkcg_policy_cfq);
4434 cfq_pool = KMEM_CACHE(cfq_queue, 0);
4438 ret = elv_register(&iosched_cfq);
4445 kmem_cache_destroy(cfq_pool);
4447 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4448 blkcg_policy_unregister(&blkcg_policy_cfq);
4453 static void __exit cfq_exit(void)
4455 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4456 blkcg_policy_unregister(&blkcg_policy_cfq);
4458 elv_unregister(&iosched_cfq);
4459 kmem_cache_destroy(cfq_pool);
4462 module_init(cfq_init);
4463 module_exit(cfq_exit);
4465 MODULE_AUTHOR("Jens Axboe");
4466 MODULE_LICENSE("GPL");
4467 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");