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)
70 /* blkio-related constants */
71 #define CFQ_WEIGHT_MIN 10
72 #define CFQ_WEIGHT_MAX 1000
73 #define CFQ_WEIGHT_DEFAULT 500
76 unsigned long last_end_request;
78 unsigned long ttime_total;
79 unsigned long ttime_samples;
80 unsigned long ttime_mean;
84 * Most of our rbtree usage is for sorting with min extraction, so
85 * if we cache the leftmost node we don't have to walk down the tree
86 * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should
87 * move this into the elevator for the rq sorting as well.
94 struct cfq_ttime ttime;
96 #define CFQ_RB_ROOT (struct cfq_rb_root) { .rb = RB_ROOT, \
97 .ttime = {.last_end_request = jiffies,},}
100 * Per process-grouping structure
103 /* reference count */
105 /* various state flags, see below */
107 /* parent cfq_data */
108 struct cfq_data *cfqd;
109 /* service_tree member */
110 struct rb_node rb_node;
111 /* service_tree key */
112 unsigned long rb_key;
113 /* prio tree member */
114 struct rb_node p_node;
115 /* prio tree root we belong to, if any */
116 struct rb_root *p_root;
117 /* sorted list of pending requests */
118 struct rb_root sort_list;
119 /* if fifo isn't expired, next request to serve */
120 struct request *next_rq;
121 /* requests queued in sort_list */
123 /* currently allocated requests */
125 /* fifo list of requests in sort_list */
126 struct list_head fifo;
128 /* time when queue got scheduled in to dispatch first request. */
129 unsigned long dispatch_start;
130 unsigned int allocated_slice;
131 unsigned int slice_dispatch;
132 /* time when first request from queue completed and slice started. */
133 unsigned long slice_start;
134 unsigned long slice_end;
137 /* pending priority requests */
139 /* number of requests that are on the dispatch list or inside driver */
142 /* io prio of this group */
143 unsigned short ioprio, org_ioprio;
144 unsigned short ioprio_class;
149 sector_t last_request_pos;
151 struct cfq_rb_root *service_tree;
152 struct cfq_queue *new_cfqq;
153 struct cfq_group *cfqg;
154 /* Number of sectors dispatched from queue in single dispatch round */
155 unsigned long nr_sectors;
159 * First index in the service_trees.
160 * IDLE is handled separately, so it has negative index
170 * Second index in the service_trees.
174 SYNC_NOIDLE_WORKLOAD = 1,
179 #ifdef CONFIG_CFQ_GROUP_IOSCHED
180 /* total bytes transferred */
181 struct blkg_rwstat service_bytes;
182 /* total IOs serviced, post merge */
183 struct blkg_rwstat serviced;
184 /* number of ios merged */
185 struct blkg_rwstat merged;
186 /* total time spent on device in ns, may not be accurate w/ queueing */
187 struct blkg_rwstat service_time;
188 /* total time spent waiting in scheduler queue in ns */
189 struct blkg_rwstat wait_time;
190 /* number of IOs queued up */
191 struct blkg_rwstat queued;
192 /* total sectors transferred */
193 struct blkg_stat sectors;
194 /* total disk time and nr sectors dispatched by this group */
195 struct blkg_stat time;
196 #ifdef CONFIG_DEBUG_BLK_CGROUP
197 /* time not charged to this cgroup */
198 struct blkg_stat unaccounted_time;
199 /* sum of number of ios queued across all samples */
200 struct blkg_stat avg_queue_size_sum;
201 /* count of samples taken for average */
202 struct blkg_stat avg_queue_size_samples;
203 /* how many times this group has been removed from service tree */
204 struct blkg_stat dequeue;
205 /* total time spent waiting for it to be assigned a timeslice. */
206 struct blkg_stat group_wait_time;
207 /* time spent idling for this blkcg_gq */
208 struct blkg_stat idle_time;
209 /* total time with empty current active q with other requests queued */
210 struct blkg_stat empty_time;
211 /* fields after this shouldn't be cleared on stat reset */
212 uint64_t start_group_wait_time;
213 uint64_t start_idle_time;
214 uint64_t start_empty_time;
216 #endif /* CONFIG_DEBUG_BLK_CGROUP */
217 #endif /* CONFIG_CFQ_GROUP_IOSCHED */
220 /* Per-cgroup data */
221 struct cfq_group_data {
222 /* must be the first member */
223 struct blkcg_policy_data pd;
226 unsigned int leaf_weight;
229 /* This is per cgroup per device grouping structure */
231 /* must be the first member */
232 struct blkg_policy_data pd;
234 /* group service_tree member */
235 struct rb_node rb_node;
237 /* group service_tree key */
241 * The number of active cfqgs and sum of their weights under this
242 * cfqg. This covers this cfqg's leaf_weight and all children's
243 * weights, but does not cover weights of further descendants.
245 * If a cfqg is on the service tree, it's active. An active cfqg
246 * also activates its parent and contributes to the children_weight
250 unsigned int children_weight;
253 * vfraction is the fraction of vdisktime that the tasks in this
254 * cfqg are entitled to. This is determined by compounding the
255 * ratios walking up from this cfqg to the root.
257 * It is in fixed point w/ CFQ_SERVICE_SHIFT and the sum of all
258 * vfractions on a service tree is approximately 1. The sum may
259 * deviate a bit due to rounding errors and fluctuations caused by
260 * cfqgs entering and leaving the service tree.
262 unsigned int vfraction;
265 * There are two weights - (internal) weight is the weight of this
266 * cfqg against the sibling cfqgs. leaf_weight is the wight of
267 * this cfqg against the child cfqgs. For the root cfqg, both
268 * weights are kept in sync for backward compatibility.
271 unsigned int new_weight;
272 unsigned int dev_weight;
274 unsigned int leaf_weight;
275 unsigned int new_leaf_weight;
276 unsigned int dev_leaf_weight;
278 /* number of cfqq currently on this group */
282 * Per group busy queues average. Useful for workload slice calc. We
283 * create the array for each prio class but at run time it is used
284 * only for RT and BE class and slot for IDLE class remains unused.
285 * This is primarily done to avoid confusion and a gcc warning.
287 unsigned int busy_queues_avg[CFQ_PRIO_NR];
289 * rr lists of queues with requests. We maintain service trees for
290 * RT and BE classes. These trees are subdivided in subclasses
291 * of SYNC, SYNC_NOIDLE and ASYNC based on workload type. For IDLE
292 * class there is no subclassification and all the cfq queues go on
293 * a single tree service_tree_idle.
294 * Counts are embedded in the cfq_rb_root
296 struct cfq_rb_root service_trees[2][3];
297 struct cfq_rb_root service_tree_idle;
299 unsigned long saved_wl_slice;
300 enum wl_type_t saved_wl_type;
301 enum wl_class_t saved_wl_class;
303 /* number of requests that are on the dispatch list or inside driver */
305 struct cfq_ttime ttime;
306 struct cfqg_stats stats; /* stats for this cfqg */
307 struct cfqg_stats dead_stats; /* stats pushed from dead children */
311 struct io_cq icq; /* must be the first member */
312 struct cfq_queue *cfqq[2];
313 struct cfq_ttime ttime;
314 int ioprio; /* the current ioprio */
315 #ifdef CONFIG_CFQ_GROUP_IOSCHED
316 uint64_t blkcg_serial_nr; /* the current blkcg serial */
321 * Per block device queue structure
324 struct request_queue *queue;
325 /* Root service tree for cfq_groups */
326 struct cfq_rb_root grp_service_tree;
327 struct cfq_group *root_group;
330 * The priority currently being served
332 enum wl_class_t serving_wl_class;
333 enum wl_type_t serving_wl_type;
334 unsigned long workload_expires;
335 struct cfq_group *serving_group;
338 * Each priority tree is sorted by next_request position. These
339 * trees are used when determining if two or more queues are
340 * interleaving requests (see cfq_close_cooperator).
342 struct rb_root prio_trees[CFQ_PRIO_LISTS];
344 unsigned int busy_queues;
345 unsigned int busy_sync_queues;
351 * queue-depth detection
357 * -1 => indeterminate, (cfq will behave as if NCQ is present, to allow better detection)
358 * 1 => NCQ is present (hw_tag_est_depth is the estimated max depth)
361 int hw_tag_est_depth;
362 unsigned int hw_tag_samples;
365 * idle window management
367 struct timer_list idle_slice_timer;
368 struct work_struct unplug_work;
370 struct cfq_queue *active_queue;
371 struct cfq_io_cq *active_cic;
374 * async queue for each priority case
376 struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
377 struct cfq_queue *async_idle_cfqq;
379 sector_t last_position;
382 * tunables, see top of file
384 unsigned int cfq_quantum;
385 unsigned int cfq_fifo_expire[2];
386 unsigned int cfq_back_penalty;
387 unsigned int cfq_back_max;
388 unsigned int cfq_slice[2];
389 unsigned int cfq_slice_async_rq;
390 unsigned int cfq_slice_idle;
391 unsigned int cfq_group_idle;
392 unsigned int cfq_latency;
393 unsigned int cfq_target_latency;
396 * Fallback dummy cfqq for extreme OOM conditions
398 struct cfq_queue oom_cfqq;
400 unsigned long last_delayed_sync;
403 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd);
405 static struct cfq_rb_root *st_for(struct cfq_group *cfqg,
406 enum wl_class_t class,
412 if (class == IDLE_WORKLOAD)
413 return &cfqg->service_tree_idle;
415 return &cfqg->service_trees[class][type];
418 enum cfqq_state_flags {
419 CFQ_CFQQ_FLAG_on_rr = 0, /* on round-robin busy list */
420 CFQ_CFQQ_FLAG_wait_request, /* waiting for a request */
421 CFQ_CFQQ_FLAG_must_dispatch, /* must be allowed a dispatch */
422 CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
423 CFQ_CFQQ_FLAG_fifo_expire, /* FIFO checked in this slice */
424 CFQ_CFQQ_FLAG_idle_window, /* slice idling enabled */
425 CFQ_CFQQ_FLAG_prio_changed, /* task priority has changed */
426 CFQ_CFQQ_FLAG_slice_new, /* no requests dispatched in slice */
427 CFQ_CFQQ_FLAG_sync, /* synchronous queue */
428 CFQ_CFQQ_FLAG_coop, /* cfqq is shared */
429 CFQ_CFQQ_FLAG_split_coop, /* shared cfqq will be splitted */
430 CFQ_CFQQ_FLAG_deep, /* sync cfqq experienced large depth */
431 CFQ_CFQQ_FLAG_wait_busy, /* Waiting for next request */
434 #define CFQ_CFQQ_FNS(name) \
435 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \
437 (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name); \
439 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \
441 (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \
443 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \
445 return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \
449 CFQ_CFQQ_FNS(wait_request);
450 CFQ_CFQQ_FNS(must_dispatch);
451 CFQ_CFQQ_FNS(must_alloc_slice);
452 CFQ_CFQQ_FNS(fifo_expire);
453 CFQ_CFQQ_FNS(idle_window);
454 CFQ_CFQQ_FNS(prio_changed);
455 CFQ_CFQQ_FNS(slice_new);
458 CFQ_CFQQ_FNS(split_coop);
460 CFQ_CFQQ_FNS(wait_busy);
463 static inline struct cfq_group *pd_to_cfqg(struct blkg_policy_data *pd)
465 return pd ? container_of(pd, struct cfq_group, pd) : NULL;
468 static struct cfq_group_data
469 *cpd_to_cfqgd(struct blkcg_policy_data *cpd)
471 return cpd ? container_of(cpd, struct cfq_group_data, pd) : NULL;
474 static inline struct blkcg_gq *cfqg_to_blkg(struct cfq_group *cfqg)
476 return pd_to_blkg(&cfqg->pd);
479 #if defined(CONFIG_CFQ_GROUP_IOSCHED) && defined(CONFIG_DEBUG_BLK_CGROUP)
481 /* cfqg stats flags */
482 enum cfqg_stats_flags {
483 CFQG_stats_waiting = 0,
488 #define CFQG_FLAG_FNS(name) \
489 static inline void cfqg_stats_mark_##name(struct cfqg_stats *stats) \
491 stats->flags |= (1 << CFQG_stats_##name); \
493 static inline void cfqg_stats_clear_##name(struct cfqg_stats *stats) \
495 stats->flags &= ~(1 << CFQG_stats_##name); \
497 static inline int cfqg_stats_##name(struct cfqg_stats *stats) \
499 return (stats->flags & (1 << CFQG_stats_##name)) != 0; \
502 CFQG_FLAG_FNS(waiting)
503 CFQG_FLAG_FNS(idling)
507 /* This should be called with the queue_lock held. */
508 static void cfqg_stats_update_group_wait_time(struct cfqg_stats *stats)
510 unsigned long long now;
512 if (!cfqg_stats_waiting(stats))
516 if (time_after64(now, stats->start_group_wait_time))
517 blkg_stat_add(&stats->group_wait_time,
518 now - stats->start_group_wait_time);
519 cfqg_stats_clear_waiting(stats);
522 /* This should be called with the queue_lock held. */
523 static void cfqg_stats_set_start_group_wait_time(struct cfq_group *cfqg,
524 struct cfq_group *curr_cfqg)
526 struct cfqg_stats *stats = &cfqg->stats;
528 if (cfqg_stats_waiting(stats))
530 if (cfqg == curr_cfqg)
532 stats->start_group_wait_time = sched_clock();
533 cfqg_stats_mark_waiting(stats);
536 /* This should be called with the queue_lock held. */
537 static void cfqg_stats_end_empty_time(struct cfqg_stats *stats)
539 unsigned long long now;
541 if (!cfqg_stats_empty(stats))
545 if (time_after64(now, stats->start_empty_time))
546 blkg_stat_add(&stats->empty_time,
547 now - stats->start_empty_time);
548 cfqg_stats_clear_empty(stats);
551 static void cfqg_stats_update_dequeue(struct cfq_group *cfqg)
553 blkg_stat_add(&cfqg->stats.dequeue, 1);
556 static void cfqg_stats_set_start_empty_time(struct cfq_group *cfqg)
558 struct cfqg_stats *stats = &cfqg->stats;
560 if (blkg_rwstat_total(&stats->queued))
564 * group is already marked empty. This can happen if cfqq got new
565 * request in parent group and moved to this group while being added
566 * to service tree. Just ignore the event and move on.
568 if (cfqg_stats_empty(stats))
571 stats->start_empty_time = sched_clock();
572 cfqg_stats_mark_empty(stats);
575 static void cfqg_stats_update_idle_time(struct cfq_group *cfqg)
577 struct cfqg_stats *stats = &cfqg->stats;
579 if (cfqg_stats_idling(stats)) {
580 unsigned long long now = sched_clock();
582 if (time_after64(now, stats->start_idle_time))
583 blkg_stat_add(&stats->idle_time,
584 now - stats->start_idle_time);
585 cfqg_stats_clear_idling(stats);
589 static void cfqg_stats_set_start_idle_time(struct cfq_group *cfqg)
591 struct cfqg_stats *stats = &cfqg->stats;
593 BUG_ON(cfqg_stats_idling(stats));
595 stats->start_idle_time = sched_clock();
596 cfqg_stats_mark_idling(stats);
599 static void cfqg_stats_update_avg_queue_size(struct cfq_group *cfqg)
601 struct cfqg_stats *stats = &cfqg->stats;
603 blkg_stat_add(&stats->avg_queue_size_sum,
604 blkg_rwstat_total(&stats->queued));
605 blkg_stat_add(&stats->avg_queue_size_samples, 1);
606 cfqg_stats_update_group_wait_time(stats);
609 #else /* CONFIG_CFQ_GROUP_IOSCHED && CONFIG_DEBUG_BLK_CGROUP */
611 static inline void cfqg_stats_set_start_group_wait_time(struct cfq_group *cfqg, struct cfq_group *curr_cfqg) { }
612 static inline void cfqg_stats_end_empty_time(struct cfqg_stats *stats) { }
613 static inline void cfqg_stats_update_dequeue(struct cfq_group *cfqg) { }
614 static inline void cfqg_stats_set_start_empty_time(struct cfq_group *cfqg) { }
615 static inline void cfqg_stats_update_idle_time(struct cfq_group *cfqg) { }
616 static inline void cfqg_stats_set_start_idle_time(struct cfq_group *cfqg) { }
617 static inline void cfqg_stats_update_avg_queue_size(struct cfq_group *cfqg) { }
619 #endif /* CONFIG_CFQ_GROUP_IOSCHED && CONFIG_DEBUG_BLK_CGROUP */
621 #ifdef CONFIG_CFQ_GROUP_IOSCHED
623 static struct blkcg_policy blkcg_policy_cfq;
625 static inline struct cfq_group *blkg_to_cfqg(struct blkcg_gq *blkg)
627 return pd_to_cfqg(blkg_to_pd(blkg, &blkcg_policy_cfq));
630 static struct cfq_group_data *blkcg_to_cfqgd(struct blkcg *blkcg)
632 return cpd_to_cfqgd(blkcg_to_cpd(blkcg, &blkcg_policy_cfq));
635 static inline struct cfq_group *cfqg_parent(struct cfq_group *cfqg)
637 struct blkcg_gq *pblkg = cfqg_to_blkg(cfqg)->parent;
639 return pblkg ? blkg_to_cfqg(pblkg) : NULL;
642 static inline void cfqg_get(struct cfq_group *cfqg)
644 return blkg_get(cfqg_to_blkg(cfqg));
647 static inline void cfqg_put(struct cfq_group *cfqg)
649 return blkg_put(cfqg_to_blkg(cfqg));
652 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...) do { \
655 blkg_path(cfqg_to_blkg((cfqq)->cfqg), __pbuf, sizeof(__pbuf)); \
656 blk_add_trace_msg((cfqd)->queue, "cfq%d%c%c %s " fmt, (cfqq)->pid, \
657 cfq_cfqq_sync((cfqq)) ? 'S' : 'A', \
658 cfqq_type((cfqq)) == SYNC_NOIDLE_WORKLOAD ? 'N' : ' ',\
662 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...) do { \
665 blkg_path(cfqg_to_blkg(cfqg), __pbuf, sizeof(__pbuf)); \
666 blk_add_trace_msg((cfqd)->queue, "%s " fmt, __pbuf, ##args); \
669 static inline void cfqg_stats_update_io_add(struct cfq_group *cfqg,
670 struct cfq_group *curr_cfqg, int rw)
672 blkg_rwstat_add(&cfqg->stats.queued, rw, 1);
673 cfqg_stats_end_empty_time(&cfqg->stats);
674 cfqg_stats_set_start_group_wait_time(cfqg, curr_cfqg);
677 static inline void cfqg_stats_update_timeslice_used(struct cfq_group *cfqg,
678 unsigned long time, unsigned long unaccounted_time)
680 blkg_stat_add(&cfqg->stats.time, time);
681 #ifdef CONFIG_DEBUG_BLK_CGROUP
682 blkg_stat_add(&cfqg->stats.unaccounted_time, unaccounted_time);
686 static inline void cfqg_stats_update_io_remove(struct cfq_group *cfqg, int rw)
688 blkg_rwstat_add(&cfqg->stats.queued, rw, -1);
691 static inline void cfqg_stats_update_io_merged(struct cfq_group *cfqg, int rw)
693 blkg_rwstat_add(&cfqg->stats.merged, rw, 1);
696 static inline void cfqg_stats_update_dispatch(struct cfq_group *cfqg,
697 uint64_t bytes, int rw)
699 blkg_stat_add(&cfqg->stats.sectors, bytes >> 9);
700 blkg_rwstat_add(&cfqg->stats.serviced, rw, 1);
701 blkg_rwstat_add(&cfqg->stats.service_bytes, rw, bytes);
704 static inline void cfqg_stats_update_completion(struct cfq_group *cfqg,
705 uint64_t start_time, uint64_t io_start_time, int rw)
707 struct cfqg_stats *stats = &cfqg->stats;
708 unsigned long long now = sched_clock();
710 if (time_after64(now, io_start_time))
711 blkg_rwstat_add(&stats->service_time, rw, now - io_start_time);
712 if (time_after64(io_start_time, start_time))
713 blkg_rwstat_add(&stats->wait_time, rw,
714 io_start_time - start_time);
718 static void cfqg_stats_reset(struct cfqg_stats *stats)
720 /* queued stats shouldn't be cleared */
721 blkg_rwstat_reset(&stats->service_bytes);
722 blkg_rwstat_reset(&stats->serviced);
723 blkg_rwstat_reset(&stats->merged);
724 blkg_rwstat_reset(&stats->service_time);
725 blkg_rwstat_reset(&stats->wait_time);
726 blkg_stat_reset(&stats->time);
727 #ifdef CONFIG_DEBUG_BLK_CGROUP
728 blkg_stat_reset(&stats->unaccounted_time);
729 blkg_stat_reset(&stats->avg_queue_size_sum);
730 blkg_stat_reset(&stats->avg_queue_size_samples);
731 blkg_stat_reset(&stats->dequeue);
732 blkg_stat_reset(&stats->group_wait_time);
733 blkg_stat_reset(&stats->idle_time);
734 blkg_stat_reset(&stats->empty_time);
739 static void cfqg_stats_merge(struct cfqg_stats *to, struct cfqg_stats *from)
741 /* queued stats shouldn't be cleared */
742 blkg_rwstat_merge(&to->service_bytes, &from->service_bytes);
743 blkg_rwstat_merge(&to->serviced, &from->serviced);
744 blkg_rwstat_merge(&to->merged, &from->merged);
745 blkg_rwstat_merge(&to->service_time, &from->service_time);
746 blkg_rwstat_merge(&to->wait_time, &from->wait_time);
747 blkg_stat_merge(&from->time, &from->time);
748 #ifdef CONFIG_DEBUG_BLK_CGROUP
749 blkg_stat_merge(&to->unaccounted_time, &from->unaccounted_time);
750 blkg_stat_merge(&to->avg_queue_size_sum, &from->avg_queue_size_sum);
751 blkg_stat_merge(&to->avg_queue_size_samples, &from->avg_queue_size_samples);
752 blkg_stat_merge(&to->dequeue, &from->dequeue);
753 blkg_stat_merge(&to->group_wait_time, &from->group_wait_time);
754 blkg_stat_merge(&to->idle_time, &from->idle_time);
755 blkg_stat_merge(&to->empty_time, &from->empty_time);
760 * Transfer @cfqg's stats to its parent's dead_stats so that the ancestors'
761 * recursive stats can still account for the amount used by this cfqg after
764 static void cfqg_stats_xfer_dead(struct cfq_group *cfqg)
766 struct cfq_group *parent = cfqg_parent(cfqg);
768 lockdep_assert_held(cfqg_to_blkg(cfqg)->q->queue_lock);
770 if (unlikely(!parent))
773 cfqg_stats_merge(&parent->dead_stats, &cfqg->stats);
774 cfqg_stats_merge(&parent->dead_stats, &cfqg->dead_stats);
775 cfqg_stats_reset(&cfqg->stats);
776 cfqg_stats_reset(&cfqg->dead_stats);
779 #else /* CONFIG_CFQ_GROUP_IOSCHED */
781 static inline struct cfq_group *cfqg_parent(struct cfq_group *cfqg) { return NULL; }
782 static inline void cfqg_get(struct cfq_group *cfqg) { }
783 static inline void cfqg_put(struct cfq_group *cfqg) { }
785 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
786 blk_add_trace_msg((cfqd)->queue, "cfq%d%c%c " fmt, (cfqq)->pid, \
787 cfq_cfqq_sync((cfqq)) ? 'S' : 'A', \
788 cfqq_type((cfqq)) == SYNC_NOIDLE_WORKLOAD ? 'N' : ' ',\
790 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...) do {} while (0)
792 static inline void cfqg_stats_update_io_add(struct cfq_group *cfqg,
793 struct cfq_group *curr_cfqg, int rw) { }
794 static inline void cfqg_stats_update_timeslice_used(struct cfq_group *cfqg,
795 unsigned long time, unsigned long unaccounted_time) { }
796 static inline void cfqg_stats_update_io_remove(struct cfq_group *cfqg, int rw) { }
797 static inline void cfqg_stats_update_io_merged(struct cfq_group *cfqg, int rw) { }
798 static inline void cfqg_stats_update_dispatch(struct cfq_group *cfqg,
799 uint64_t bytes, int rw) { }
800 static inline void cfqg_stats_update_completion(struct cfq_group *cfqg,
801 uint64_t start_time, uint64_t io_start_time, int rw) { }
803 #endif /* CONFIG_CFQ_GROUP_IOSCHED */
805 #define cfq_log(cfqd, fmt, args...) \
806 blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
808 /* Traverses through cfq group service trees */
809 #define for_each_cfqg_st(cfqg, i, j, st) \
810 for (i = 0; i <= IDLE_WORKLOAD; i++) \
811 for (j = 0, st = i < IDLE_WORKLOAD ? &cfqg->service_trees[i][j]\
812 : &cfqg->service_tree_idle; \
813 (i < IDLE_WORKLOAD && j <= SYNC_WORKLOAD) || \
814 (i == IDLE_WORKLOAD && j == 0); \
815 j++, st = i < IDLE_WORKLOAD ? \
816 &cfqg->service_trees[i][j]: NULL) \
818 static inline bool cfq_io_thinktime_big(struct cfq_data *cfqd,
819 struct cfq_ttime *ttime, bool group_idle)
822 if (!sample_valid(ttime->ttime_samples))
825 slice = cfqd->cfq_group_idle;
827 slice = cfqd->cfq_slice_idle;
828 return ttime->ttime_mean > slice;
831 static inline bool iops_mode(struct cfq_data *cfqd)
834 * If we are not idling on queues and it is a NCQ drive, parallel
835 * execution of requests is on and measuring time is not possible
836 * in most of the cases until and unless we drive shallower queue
837 * depths and that becomes a performance bottleneck. In such cases
838 * switch to start providing fairness in terms of number of IOs.
840 if (!cfqd->cfq_slice_idle && cfqd->hw_tag)
846 static inline enum wl_class_t cfqq_class(struct cfq_queue *cfqq)
848 if (cfq_class_idle(cfqq))
849 return IDLE_WORKLOAD;
850 if (cfq_class_rt(cfqq))
856 static enum wl_type_t cfqq_type(struct cfq_queue *cfqq)
858 if (!cfq_cfqq_sync(cfqq))
859 return ASYNC_WORKLOAD;
860 if (!cfq_cfqq_idle_window(cfqq))
861 return SYNC_NOIDLE_WORKLOAD;
862 return SYNC_WORKLOAD;
865 static inline int cfq_group_busy_queues_wl(enum wl_class_t wl_class,
866 struct cfq_data *cfqd,
867 struct cfq_group *cfqg)
869 if (wl_class == IDLE_WORKLOAD)
870 return cfqg->service_tree_idle.count;
872 return cfqg->service_trees[wl_class][ASYNC_WORKLOAD].count +
873 cfqg->service_trees[wl_class][SYNC_NOIDLE_WORKLOAD].count +
874 cfqg->service_trees[wl_class][SYNC_WORKLOAD].count;
877 static inline int cfqg_busy_async_queues(struct cfq_data *cfqd,
878 struct cfq_group *cfqg)
880 return cfqg->service_trees[RT_WORKLOAD][ASYNC_WORKLOAD].count +
881 cfqg->service_trees[BE_WORKLOAD][ASYNC_WORKLOAD].count;
884 static void cfq_dispatch_insert(struct request_queue *, struct request *);
885 static struct cfq_queue *cfq_get_queue(struct cfq_data *cfqd, bool is_sync,
886 struct cfq_io_cq *cic, struct bio *bio,
889 static inline struct cfq_io_cq *icq_to_cic(struct io_cq *icq)
891 /* cic->icq is the first member, %NULL will convert to %NULL */
892 return container_of(icq, struct cfq_io_cq, icq);
895 static inline struct cfq_io_cq *cfq_cic_lookup(struct cfq_data *cfqd,
896 struct io_context *ioc)
899 return icq_to_cic(ioc_lookup_icq(ioc, cfqd->queue));
903 static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_cq *cic, bool is_sync)
905 return cic->cfqq[is_sync];
908 static inline void cic_set_cfqq(struct cfq_io_cq *cic, struct cfq_queue *cfqq,
911 cic->cfqq[is_sync] = cfqq;
914 static inline struct cfq_data *cic_to_cfqd(struct cfq_io_cq *cic)
916 return cic->icq.q->elevator->elevator_data;
920 * We regard a request as SYNC, if it's either a read or has the SYNC bit
921 * set (in which case it could also be direct WRITE).
923 static inline bool cfq_bio_sync(struct bio *bio)
925 return bio_data_dir(bio) == READ || (bio->bi_rw & REQ_SYNC);
929 * scheduler run of queue, if there are requests pending and no one in the
930 * driver that will restart queueing
932 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
934 if (cfqd->busy_queues) {
935 cfq_log(cfqd, "schedule dispatch");
936 kblockd_schedule_work(&cfqd->unplug_work);
941 * Scale schedule slice based on io priority. Use the sync time slice only
942 * if a queue is marked sync and has sync io queued. A sync queue with async
943 * io only, should not get full sync slice length.
945 static inline int cfq_prio_slice(struct cfq_data *cfqd, bool sync,
948 const int base_slice = cfqd->cfq_slice[sync];
950 WARN_ON(prio >= IOPRIO_BE_NR);
952 return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
956 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
958 return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
962 * cfqg_scale_charge - scale disk time charge according to cfqg weight
963 * @charge: disk time being charged
964 * @vfraction: vfraction of the cfqg, fixed point w/ CFQ_SERVICE_SHIFT
966 * Scale @charge according to @vfraction, which is in range (0, 1]. The
967 * scaling is inversely proportional.
969 * scaled = charge / vfraction
971 * The result is also in fixed point w/ CFQ_SERVICE_SHIFT.
973 static inline u64 cfqg_scale_charge(unsigned long charge,
974 unsigned int vfraction)
976 u64 c = charge << CFQ_SERVICE_SHIFT; /* make it fixed point */
978 /* charge / vfraction */
979 c <<= CFQ_SERVICE_SHIFT;
980 do_div(c, vfraction);
984 static inline u64 max_vdisktime(u64 min_vdisktime, u64 vdisktime)
986 s64 delta = (s64)(vdisktime - min_vdisktime);
988 min_vdisktime = vdisktime;
990 return min_vdisktime;
993 static inline u64 min_vdisktime(u64 min_vdisktime, u64 vdisktime)
995 s64 delta = (s64)(vdisktime - min_vdisktime);
997 min_vdisktime = vdisktime;
999 return min_vdisktime;
1002 static void update_min_vdisktime(struct cfq_rb_root *st)
1004 struct cfq_group *cfqg;
1007 cfqg = rb_entry_cfqg(st->left);
1008 st->min_vdisktime = max_vdisktime(st->min_vdisktime,
1014 * get averaged number of queues of RT/BE priority.
1015 * average is updated, with a formula that gives more weight to higher numbers,
1016 * to quickly follows sudden increases and decrease slowly
1019 static inline unsigned cfq_group_get_avg_queues(struct cfq_data *cfqd,
1020 struct cfq_group *cfqg, bool rt)
1022 unsigned min_q, max_q;
1023 unsigned mult = cfq_hist_divisor - 1;
1024 unsigned round = cfq_hist_divisor / 2;
1025 unsigned busy = cfq_group_busy_queues_wl(rt, cfqd, cfqg);
1027 min_q = min(cfqg->busy_queues_avg[rt], busy);
1028 max_q = max(cfqg->busy_queues_avg[rt], busy);
1029 cfqg->busy_queues_avg[rt] = (mult * max_q + min_q + round) /
1031 return cfqg->busy_queues_avg[rt];
1034 static inline unsigned
1035 cfq_group_slice(struct cfq_data *cfqd, struct cfq_group *cfqg)
1037 return cfqd->cfq_target_latency * cfqg->vfraction >> CFQ_SERVICE_SHIFT;
1040 static inline unsigned
1041 cfq_scaled_cfqq_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1043 unsigned slice = cfq_prio_to_slice(cfqd, cfqq);
1044 if (cfqd->cfq_latency) {
1046 * interested queues (we consider only the ones with the same
1047 * priority class in the cfq group)
1049 unsigned iq = cfq_group_get_avg_queues(cfqd, cfqq->cfqg,
1050 cfq_class_rt(cfqq));
1051 unsigned sync_slice = cfqd->cfq_slice[1];
1052 unsigned expect_latency = sync_slice * iq;
1053 unsigned group_slice = cfq_group_slice(cfqd, cfqq->cfqg);
1055 if (expect_latency > group_slice) {
1056 unsigned base_low_slice = 2 * cfqd->cfq_slice_idle;
1057 /* scale low_slice according to IO priority
1058 * and sync vs async */
1059 unsigned low_slice =
1060 min(slice, base_low_slice * slice / sync_slice);
1061 /* the adapted slice value is scaled to fit all iqs
1062 * into the target latency */
1063 slice = max(slice * group_slice / expect_latency,
1071 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1073 unsigned slice = cfq_scaled_cfqq_slice(cfqd, cfqq);
1075 cfqq->slice_start = jiffies;
1076 cfqq->slice_end = jiffies + slice;
1077 cfqq->allocated_slice = slice;
1078 cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
1082 * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
1083 * isn't valid until the first request from the dispatch is activated
1084 * and the slice time set.
1086 static inline bool cfq_slice_used(struct cfq_queue *cfqq)
1088 if (cfq_cfqq_slice_new(cfqq))
1090 if (time_before(jiffies, cfqq->slice_end))
1097 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
1098 * We choose the request that is closest to the head right now. Distance
1099 * behind the head is penalized and only allowed to a certain extent.
1101 static struct request *
1102 cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2, sector_t last)
1104 sector_t s1, s2, d1 = 0, d2 = 0;
1105 unsigned long back_max;
1106 #define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */
1107 #define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */
1108 unsigned wrap = 0; /* bit mask: requests behind the disk head? */
1110 if (rq1 == NULL || rq1 == rq2)
1115 if (rq_is_sync(rq1) != rq_is_sync(rq2))
1116 return rq_is_sync(rq1) ? rq1 : rq2;
1118 if ((rq1->cmd_flags ^ rq2->cmd_flags) & REQ_PRIO)
1119 return rq1->cmd_flags & REQ_PRIO ? rq1 : rq2;
1121 s1 = blk_rq_pos(rq1);
1122 s2 = blk_rq_pos(rq2);
1125 * by definition, 1KiB is 2 sectors
1127 back_max = cfqd->cfq_back_max * 2;
1130 * Strict one way elevator _except_ in the case where we allow
1131 * short backward seeks which are biased as twice the cost of a
1132 * similar forward seek.
1136 else if (s1 + back_max >= last)
1137 d1 = (last - s1) * cfqd->cfq_back_penalty;
1139 wrap |= CFQ_RQ1_WRAP;
1143 else if (s2 + back_max >= last)
1144 d2 = (last - s2) * cfqd->cfq_back_penalty;
1146 wrap |= CFQ_RQ2_WRAP;
1148 /* Found required data */
1151 * By doing switch() on the bit mask "wrap" we avoid having to
1152 * check two variables for all permutations: --> faster!
1155 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
1171 case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
1174 * Since both rqs are wrapped,
1175 * start with the one that's further behind head
1176 * (--> only *one* back seek required),
1177 * since back seek takes more time than forward.
1187 * The below is leftmost cache rbtree addon
1189 static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
1191 /* Service tree is empty */
1196 root->left = rb_first(&root->rb);
1199 return rb_entry(root->left, struct cfq_queue, rb_node);
1204 static struct cfq_group *cfq_rb_first_group(struct cfq_rb_root *root)
1207 root->left = rb_first(&root->rb);
1210 return rb_entry_cfqg(root->left);
1215 static void rb_erase_init(struct rb_node *n, struct rb_root *root)
1221 static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
1223 if (root->left == n)
1225 rb_erase_init(n, &root->rb);
1230 * would be nice to take fifo expire time into account as well
1232 static struct request *
1233 cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1234 struct request *last)
1236 struct rb_node *rbnext = rb_next(&last->rb_node);
1237 struct rb_node *rbprev = rb_prev(&last->rb_node);
1238 struct request *next = NULL, *prev = NULL;
1240 BUG_ON(RB_EMPTY_NODE(&last->rb_node));
1243 prev = rb_entry_rq(rbprev);
1246 next = rb_entry_rq(rbnext);
1248 rbnext = rb_first(&cfqq->sort_list);
1249 if (rbnext && rbnext != &last->rb_node)
1250 next = rb_entry_rq(rbnext);
1253 return cfq_choose_req(cfqd, next, prev, blk_rq_pos(last));
1256 static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
1257 struct cfq_queue *cfqq)
1260 * just an approximation, should be ok.
1262 return (cfqq->cfqg->nr_cfqq - 1) * (cfq_prio_slice(cfqd, 1, 0) -
1263 cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
1267 cfqg_key(struct cfq_rb_root *st, struct cfq_group *cfqg)
1269 return cfqg->vdisktime - st->min_vdisktime;
1273 __cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
1275 struct rb_node **node = &st->rb.rb_node;
1276 struct rb_node *parent = NULL;
1277 struct cfq_group *__cfqg;
1278 s64 key = cfqg_key(st, cfqg);
1281 while (*node != NULL) {
1283 __cfqg = rb_entry_cfqg(parent);
1285 if (key < cfqg_key(st, __cfqg))
1286 node = &parent->rb_left;
1288 node = &parent->rb_right;
1294 st->left = &cfqg->rb_node;
1296 rb_link_node(&cfqg->rb_node, parent, node);
1297 rb_insert_color(&cfqg->rb_node, &st->rb);
1301 * This has to be called only on activation of cfqg
1304 cfq_update_group_weight(struct cfq_group *cfqg)
1306 if (cfqg->new_weight) {
1307 cfqg->weight = cfqg->new_weight;
1308 cfqg->new_weight = 0;
1313 cfq_update_group_leaf_weight(struct cfq_group *cfqg)
1315 BUG_ON(!RB_EMPTY_NODE(&cfqg->rb_node));
1317 if (cfqg->new_leaf_weight) {
1318 cfqg->leaf_weight = cfqg->new_leaf_weight;
1319 cfqg->new_leaf_weight = 0;
1324 cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
1326 unsigned int vfr = 1 << CFQ_SERVICE_SHIFT; /* start with 1 */
1327 struct cfq_group *pos = cfqg;
1328 struct cfq_group *parent;
1331 /* add to the service tree */
1332 BUG_ON(!RB_EMPTY_NODE(&cfqg->rb_node));
1335 * Update leaf_weight. We cannot update weight at this point
1336 * because cfqg might already have been activated and is
1337 * contributing its current weight to the parent's child_weight.
1339 cfq_update_group_leaf_weight(cfqg);
1340 __cfq_group_service_tree_add(st, cfqg);
1343 * Activate @cfqg and calculate the portion of vfraction @cfqg is
1344 * entitled to. vfraction is calculated by walking the tree
1345 * towards the root calculating the fraction it has at each level.
1346 * The compounded ratio is how much vfraction @cfqg owns.
1348 * Start with the proportion tasks in this cfqg has against active
1349 * children cfqgs - its leaf_weight against children_weight.
1351 propagate = !pos->nr_active++;
1352 pos->children_weight += pos->leaf_weight;
1353 vfr = vfr * pos->leaf_weight / pos->children_weight;
1356 * Compound ->weight walking up the tree. Both activation and
1357 * vfraction calculation are done in the same loop. Propagation
1358 * stops once an already activated node is met. vfraction
1359 * calculation should always continue to the root.
1361 while ((parent = cfqg_parent(pos))) {
1363 cfq_update_group_weight(pos);
1364 propagate = !parent->nr_active++;
1365 parent->children_weight += pos->weight;
1367 vfr = vfr * pos->weight / parent->children_weight;
1371 cfqg->vfraction = max_t(unsigned, vfr, 1);
1375 cfq_group_notify_queue_add(struct cfq_data *cfqd, struct cfq_group *cfqg)
1377 struct cfq_rb_root *st = &cfqd->grp_service_tree;
1378 struct cfq_group *__cfqg;
1382 if (!RB_EMPTY_NODE(&cfqg->rb_node))
1386 * Currently put the group at the end. Later implement something
1387 * so that groups get lesser vtime based on their weights, so that
1388 * if group does not loose all if it was not continuously backlogged.
1390 n = rb_last(&st->rb);
1392 __cfqg = rb_entry_cfqg(n);
1393 cfqg->vdisktime = __cfqg->vdisktime + CFQ_IDLE_DELAY;
1395 cfqg->vdisktime = st->min_vdisktime;
1396 cfq_group_service_tree_add(st, cfqg);
1400 cfq_group_service_tree_del(struct cfq_rb_root *st, struct cfq_group *cfqg)
1402 struct cfq_group *pos = cfqg;
1406 * Undo activation from cfq_group_service_tree_add(). Deactivate
1407 * @cfqg and propagate deactivation upwards.
1409 propagate = !--pos->nr_active;
1410 pos->children_weight -= pos->leaf_weight;
1413 struct cfq_group *parent = cfqg_parent(pos);
1415 /* @pos has 0 nr_active at this point */
1416 WARN_ON_ONCE(pos->children_weight);
1422 propagate = !--parent->nr_active;
1423 parent->children_weight -= pos->weight;
1427 /* remove from the service tree */
1428 if (!RB_EMPTY_NODE(&cfqg->rb_node))
1429 cfq_rb_erase(&cfqg->rb_node, st);
1433 cfq_group_notify_queue_del(struct cfq_data *cfqd, struct cfq_group *cfqg)
1435 struct cfq_rb_root *st = &cfqd->grp_service_tree;
1437 BUG_ON(cfqg->nr_cfqq < 1);
1440 /* If there are other cfq queues under this group, don't delete it */
1444 cfq_log_cfqg(cfqd, cfqg, "del_from_rr group");
1445 cfq_group_service_tree_del(st, cfqg);
1446 cfqg->saved_wl_slice = 0;
1447 cfqg_stats_update_dequeue(cfqg);
1450 static inline unsigned int cfq_cfqq_slice_usage(struct cfq_queue *cfqq,
1451 unsigned int *unaccounted_time)
1453 unsigned int slice_used;
1456 * Queue got expired before even a single request completed or
1457 * got expired immediately after first request completion.
1459 if (!cfqq->slice_start || cfqq->slice_start == jiffies) {
1461 * Also charge the seek time incurred to the group, otherwise
1462 * if there are mutiple queues in the group, each can dispatch
1463 * a single request on seeky media and cause lots of seek time
1464 * and group will never know it.
1466 slice_used = max_t(unsigned, (jiffies - cfqq->dispatch_start),
1469 slice_used = jiffies - cfqq->slice_start;
1470 if (slice_used > cfqq->allocated_slice) {
1471 *unaccounted_time = slice_used - cfqq->allocated_slice;
1472 slice_used = cfqq->allocated_slice;
1474 if (time_after(cfqq->slice_start, cfqq->dispatch_start))
1475 *unaccounted_time += cfqq->slice_start -
1476 cfqq->dispatch_start;
1482 static void cfq_group_served(struct cfq_data *cfqd, struct cfq_group *cfqg,
1483 struct cfq_queue *cfqq)
1485 struct cfq_rb_root *st = &cfqd->grp_service_tree;
1486 unsigned int used_sl, charge, unaccounted_sl = 0;
1487 int nr_sync = cfqg->nr_cfqq - cfqg_busy_async_queues(cfqd, cfqg)
1488 - cfqg->service_tree_idle.count;
1491 BUG_ON(nr_sync < 0);
1492 used_sl = charge = cfq_cfqq_slice_usage(cfqq, &unaccounted_sl);
1494 if (iops_mode(cfqd))
1495 charge = cfqq->slice_dispatch;
1496 else if (!cfq_cfqq_sync(cfqq) && !nr_sync)
1497 charge = cfqq->allocated_slice;
1500 * Can't update vdisktime while on service tree and cfqg->vfraction
1501 * is valid only while on it. Cache vfr, leave the service tree,
1502 * update vdisktime and go back on. The re-addition to the tree
1503 * will also update the weights as necessary.
1505 vfr = cfqg->vfraction;
1506 cfq_group_service_tree_del(st, cfqg);
1507 cfqg->vdisktime += cfqg_scale_charge(charge, vfr);
1508 cfq_group_service_tree_add(st, cfqg);
1510 /* This group is being expired. Save the context */
1511 if (time_after(cfqd->workload_expires, jiffies)) {
1512 cfqg->saved_wl_slice = cfqd->workload_expires
1514 cfqg->saved_wl_type = cfqd->serving_wl_type;
1515 cfqg->saved_wl_class = cfqd->serving_wl_class;
1517 cfqg->saved_wl_slice = 0;
1519 cfq_log_cfqg(cfqd, cfqg, "served: vt=%llu min_vt=%llu", cfqg->vdisktime,
1521 cfq_log_cfqq(cfqq->cfqd, cfqq,
1522 "sl_used=%u disp=%u charge=%u iops=%u sect=%lu",
1523 used_sl, cfqq->slice_dispatch, charge,
1524 iops_mode(cfqd), cfqq->nr_sectors);
1525 cfqg_stats_update_timeslice_used(cfqg, used_sl, unaccounted_sl);
1526 cfqg_stats_set_start_empty_time(cfqg);
1530 * cfq_init_cfqg_base - initialize base part of a cfq_group
1531 * @cfqg: cfq_group to initialize
1533 * Initialize the base part which is used whether %CONFIG_CFQ_GROUP_IOSCHED
1534 * is enabled or not.
1536 static void cfq_init_cfqg_base(struct cfq_group *cfqg)
1538 struct cfq_rb_root *st;
1541 for_each_cfqg_st(cfqg, i, j, st)
1543 RB_CLEAR_NODE(&cfqg->rb_node);
1545 cfqg->ttime.last_end_request = jiffies;
1548 #ifdef CONFIG_CFQ_GROUP_IOSCHED
1549 static void cfqg_stats_init(struct cfqg_stats *stats)
1551 blkg_rwstat_init(&stats->service_bytes);
1552 blkg_rwstat_init(&stats->serviced);
1553 blkg_rwstat_init(&stats->merged);
1554 blkg_rwstat_init(&stats->service_time);
1555 blkg_rwstat_init(&stats->wait_time);
1556 blkg_rwstat_init(&stats->queued);
1558 blkg_stat_init(&stats->sectors);
1559 blkg_stat_init(&stats->time);
1561 #ifdef CONFIG_DEBUG_BLK_CGROUP
1562 blkg_stat_init(&stats->unaccounted_time);
1563 blkg_stat_init(&stats->avg_queue_size_sum);
1564 blkg_stat_init(&stats->avg_queue_size_samples);
1565 blkg_stat_init(&stats->dequeue);
1566 blkg_stat_init(&stats->group_wait_time);
1567 blkg_stat_init(&stats->idle_time);
1568 blkg_stat_init(&stats->empty_time);
1572 static void cfq_cpd_init(const struct blkcg *blkcg)
1574 struct cfq_group_data *cgd =
1575 cpd_to_cfqgd(blkcg->pd[blkcg_policy_cfq.plid]);
1577 if (blkcg == &blkcg_root) {
1578 cgd->weight = 2 * CFQ_WEIGHT_DEFAULT;
1579 cgd->leaf_weight = 2 * CFQ_WEIGHT_DEFAULT;
1581 cgd->weight = CFQ_WEIGHT_DEFAULT;
1582 cgd->leaf_weight = CFQ_WEIGHT_DEFAULT;
1586 static void cfq_pd_init(struct blkcg_gq *blkg)
1588 struct cfq_group *cfqg = blkg_to_cfqg(blkg);
1589 struct cfq_group_data *cgd = blkcg_to_cfqgd(blkg->blkcg);
1591 cfq_init_cfqg_base(cfqg);
1592 cfqg->weight = cgd->weight;
1593 cfqg->leaf_weight = cgd->leaf_weight;
1594 cfqg_stats_init(&cfqg->stats);
1595 cfqg_stats_init(&cfqg->dead_stats);
1598 static void cfq_pd_offline(struct blkcg_gq *blkg)
1601 * @blkg is going offline and will be ignored by
1602 * blkg_[rw]stat_recursive_sum(). Transfer stats to the parent so
1603 * that they don't get lost. If IOs complete after this point, the
1604 * stats for them will be lost. Oh well...
1606 cfqg_stats_xfer_dead(blkg_to_cfqg(blkg));
1609 /* offset delta from cfqg->stats to cfqg->dead_stats */
1610 static const int dead_stats_off_delta = offsetof(struct cfq_group, dead_stats) -
1611 offsetof(struct cfq_group, stats);
1613 /* to be used by recursive prfill, sums live and dead stats recursively */
1614 static u64 cfqg_stat_pd_recursive_sum(struct blkg_policy_data *pd, int off)
1618 sum += blkg_stat_recursive_sum(pd, off);
1619 sum += blkg_stat_recursive_sum(pd, off + dead_stats_off_delta);
1623 /* to be used by recursive prfill, sums live and dead rwstats recursively */
1624 static struct blkg_rwstat cfqg_rwstat_pd_recursive_sum(struct blkg_policy_data *pd,
1627 struct blkg_rwstat a, b;
1629 a = blkg_rwstat_recursive_sum(pd, off);
1630 b = blkg_rwstat_recursive_sum(pd, off + dead_stats_off_delta);
1631 blkg_rwstat_merge(&a, &b);
1635 static void cfq_pd_reset_stats(struct blkcg_gq *blkg)
1637 struct cfq_group *cfqg = blkg_to_cfqg(blkg);
1639 cfqg_stats_reset(&cfqg->stats);
1640 cfqg_stats_reset(&cfqg->dead_stats);
1644 * Search for the cfq group current task belongs to. request_queue lock must
1647 static struct cfq_group *cfq_lookup_create_cfqg(struct cfq_data *cfqd,
1648 struct blkcg *blkcg)
1650 struct request_queue *q = cfqd->queue;
1651 struct cfq_group *cfqg = NULL;
1653 /* avoid lookup for the common case where there's no blkcg */
1654 if (blkcg == &blkcg_root) {
1655 cfqg = cfqd->root_group;
1657 struct blkcg_gq *blkg;
1659 blkg = blkg_lookup_create(blkcg, q);
1661 cfqg = blkg_to_cfqg(blkg);
1667 static void cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg)
1669 /* Currently, all async queues are mapped to root group */
1670 if (!cfq_cfqq_sync(cfqq))
1671 cfqg = cfqq->cfqd->root_group;
1674 /* cfqq reference on cfqg */
1678 static u64 cfqg_prfill_weight_device(struct seq_file *sf,
1679 struct blkg_policy_data *pd, int off)
1681 struct cfq_group *cfqg = pd_to_cfqg(pd);
1683 if (!cfqg->dev_weight)
1685 return __blkg_prfill_u64(sf, pd, cfqg->dev_weight);
1688 static int cfqg_print_weight_device(struct seq_file *sf, void *v)
1690 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1691 cfqg_prfill_weight_device, &blkcg_policy_cfq,
1696 static u64 cfqg_prfill_leaf_weight_device(struct seq_file *sf,
1697 struct blkg_policy_data *pd, int off)
1699 struct cfq_group *cfqg = pd_to_cfqg(pd);
1701 if (!cfqg->dev_leaf_weight)
1703 return __blkg_prfill_u64(sf, pd, cfqg->dev_leaf_weight);
1706 static int cfqg_print_leaf_weight_device(struct seq_file *sf, void *v)
1708 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1709 cfqg_prfill_leaf_weight_device, &blkcg_policy_cfq,
1714 static int cfq_print_weight(struct seq_file *sf, void *v)
1716 struct blkcg *blkcg = css_to_blkcg(seq_css(sf));
1718 seq_printf(sf, "%u\n", blkcg_to_cfqgd(blkcg)->weight);
1722 static int cfq_print_leaf_weight(struct seq_file *sf, void *v)
1724 struct blkcg *blkcg = css_to_blkcg(seq_css(sf));
1726 seq_printf(sf, "%u\n", blkcg_to_cfqgd(blkcg)->leaf_weight);
1730 static ssize_t __cfqg_set_weight_device(struct kernfs_open_file *of,
1731 char *buf, size_t nbytes, loff_t off,
1732 bool is_leaf_weight)
1734 struct blkcg *blkcg = css_to_blkcg(of_css(of));
1735 struct blkg_conf_ctx ctx;
1736 struct cfq_group *cfqg;
1737 struct cfq_group_data *cfqgd;
1740 ret = blkg_conf_prep(blkcg, &blkcg_policy_cfq, buf, &ctx);
1745 cfqg = blkg_to_cfqg(ctx.blkg);
1746 cfqgd = blkcg_to_cfqgd(blkcg);
1747 if (!ctx.v || (ctx.v >= CFQ_WEIGHT_MIN && ctx.v <= CFQ_WEIGHT_MAX)) {
1748 if (!is_leaf_weight) {
1749 cfqg->dev_weight = ctx.v;
1750 cfqg->new_weight = ctx.v ?: cfqgd->weight;
1752 cfqg->dev_leaf_weight = ctx.v;
1753 cfqg->new_leaf_weight = ctx.v ?: cfqgd->leaf_weight;
1758 blkg_conf_finish(&ctx);
1759 return ret ?: nbytes;
1762 static ssize_t cfqg_set_weight_device(struct kernfs_open_file *of,
1763 char *buf, size_t nbytes, loff_t off)
1765 return __cfqg_set_weight_device(of, buf, nbytes, off, false);
1768 static ssize_t cfqg_set_leaf_weight_device(struct kernfs_open_file *of,
1769 char *buf, size_t nbytes, loff_t off)
1771 return __cfqg_set_weight_device(of, buf, nbytes, off, true);
1774 static int __cfq_set_weight(struct cgroup_subsys_state *css, struct cftype *cft,
1775 u64 val, bool is_leaf_weight)
1777 struct blkcg *blkcg = css_to_blkcg(css);
1778 struct blkcg_gq *blkg;
1779 struct cfq_group_data *cfqgd;
1781 if (val < CFQ_WEIGHT_MIN || val > CFQ_WEIGHT_MAX)
1784 spin_lock_irq(&blkcg->lock);
1785 cfqgd = blkcg_to_cfqgd(blkcg);
1787 if (!is_leaf_weight)
1788 cfqgd->weight = val;
1790 cfqgd->leaf_weight = val;
1792 hlist_for_each_entry(blkg, &blkcg->blkg_list, blkcg_node) {
1793 struct cfq_group *cfqg = blkg_to_cfqg(blkg);
1798 if (!is_leaf_weight) {
1799 if (!cfqg->dev_weight)
1800 cfqg->new_weight = cfqgd->weight;
1802 if (!cfqg->dev_leaf_weight)
1803 cfqg->new_leaf_weight = cfqgd->leaf_weight;
1807 spin_unlock_irq(&blkcg->lock);
1811 static int cfq_set_weight(struct cgroup_subsys_state *css, struct cftype *cft,
1814 return __cfq_set_weight(css, cft, val, false);
1817 static int cfq_set_leaf_weight(struct cgroup_subsys_state *css,
1818 struct cftype *cft, u64 val)
1820 return __cfq_set_weight(css, cft, val, true);
1823 static int cfqg_print_stat(struct seq_file *sf, void *v)
1825 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), blkg_prfill_stat,
1826 &blkcg_policy_cfq, seq_cft(sf)->private, false);
1830 static int cfqg_print_rwstat(struct seq_file *sf, void *v)
1832 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), blkg_prfill_rwstat,
1833 &blkcg_policy_cfq, seq_cft(sf)->private, true);
1837 static u64 cfqg_prfill_stat_recursive(struct seq_file *sf,
1838 struct blkg_policy_data *pd, int off)
1840 u64 sum = cfqg_stat_pd_recursive_sum(pd, off);
1842 return __blkg_prfill_u64(sf, pd, sum);
1845 static u64 cfqg_prfill_rwstat_recursive(struct seq_file *sf,
1846 struct blkg_policy_data *pd, int off)
1848 struct blkg_rwstat sum = cfqg_rwstat_pd_recursive_sum(pd, off);
1850 return __blkg_prfill_rwstat(sf, pd, &sum);
1853 static int cfqg_print_stat_recursive(struct seq_file *sf, void *v)
1855 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1856 cfqg_prfill_stat_recursive, &blkcg_policy_cfq,
1857 seq_cft(sf)->private, false);
1861 static int cfqg_print_rwstat_recursive(struct seq_file *sf, void *v)
1863 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1864 cfqg_prfill_rwstat_recursive, &blkcg_policy_cfq,
1865 seq_cft(sf)->private, true);
1869 #ifdef CONFIG_DEBUG_BLK_CGROUP
1870 static u64 cfqg_prfill_avg_queue_size(struct seq_file *sf,
1871 struct blkg_policy_data *pd, int off)
1873 struct cfq_group *cfqg = pd_to_cfqg(pd);
1874 u64 samples = blkg_stat_read(&cfqg->stats.avg_queue_size_samples);
1878 v = blkg_stat_read(&cfqg->stats.avg_queue_size_sum);
1879 v = div64_u64(v, samples);
1881 __blkg_prfill_u64(sf, pd, v);
1885 /* print avg_queue_size */
1886 static int cfqg_print_avg_queue_size(struct seq_file *sf, void *v)
1888 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1889 cfqg_prfill_avg_queue_size, &blkcg_policy_cfq,
1893 #endif /* CONFIG_DEBUG_BLK_CGROUP */
1895 static struct cftype cfq_blkcg_files[] = {
1896 /* on root, weight is mapped to leaf_weight */
1898 .name = "weight_device",
1899 .flags = CFTYPE_ONLY_ON_ROOT,
1900 .seq_show = cfqg_print_leaf_weight_device,
1901 .write = cfqg_set_leaf_weight_device,
1905 .flags = CFTYPE_ONLY_ON_ROOT,
1906 .seq_show = cfq_print_leaf_weight,
1907 .write_u64 = cfq_set_leaf_weight,
1910 /* no such mapping necessary for !roots */
1912 .name = "weight_device",
1913 .flags = CFTYPE_NOT_ON_ROOT,
1914 .seq_show = cfqg_print_weight_device,
1915 .write = cfqg_set_weight_device,
1919 .flags = CFTYPE_NOT_ON_ROOT,
1920 .seq_show = cfq_print_weight,
1921 .write_u64 = cfq_set_weight,
1925 .name = "leaf_weight_device",
1926 .seq_show = cfqg_print_leaf_weight_device,
1927 .write = cfqg_set_leaf_weight_device,
1930 .name = "leaf_weight",
1931 .seq_show = cfq_print_leaf_weight,
1932 .write_u64 = cfq_set_leaf_weight,
1935 /* statistics, covers only the tasks in the cfqg */
1938 .private = offsetof(struct cfq_group, stats.time),
1939 .seq_show = cfqg_print_stat,
1943 .private = offsetof(struct cfq_group, stats.sectors),
1944 .seq_show = cfqg_print_stat,
1947 .name = "io_service_bytes",
1948 .private = offsetof(struct cfq_group, stats.service_bytes),
1949 .seq_show = cfqg_print_rwstat,
1952 .name = "io_serviced",
1953 .private = offsetof(struct cfq_group, stats.serviced),
1954 .seq_show = cfqg_print_rwstat,
1957 .name = "io_service_time",
1958 .private = offsetof(struct cfq_group, stats.service_time),
1959 .seq_show = cfqg_print_rwstat,
1962 .name = "io_wait_time",
1963 .private = offsetof(struct cfq_group, stats.wait_time),
1964 .seq_show = cfqg_print_rwstat,
1967 .name = "io_merged",
1968 .private = offsetof(struct cfq_group, stats.merged),
1969 .seq_show = cfqg_print_rwstat,
1972 .name = "io_queued",
1973 .private = offsetof(struct cfq_group, stats.queued),
1974 .seq_show = cfqg_print_rwstat,
1977 /* the same statictics which cover the cfqg and its descendants */
1979 .name = "time_recursive",
1980 .private = offsetof(struct cfq_group, stats.time),
1981 .seq_show = cfqg_print_stat_recursive,
1984 .name = "sectors_recursive",
1985 .private = offsetof(struct cfq_group, stats.sectors),
1986 .seq_show = cfqg_print_stat_recursive,
1989 .name = "io_service_bytes_recursive",
1990 .private = offsetof(struct cfq_group, stats.service_bytes),
1991 .seq_show = cfqg_print_rwstat_recursive,
1994 .name = "io_serviced_recursive",
1995 .private = offsetof(struct cfq_group, stats.serviced),
1996 .seq_show = cfqg_print_rwstat_recursive,
1999 .name = "io_service_time_recursive",
2000 .private = offsetof(struct cfq_group, stats.service_time),
2001 .seq_show = cfqg_print_rwstat_recursive,
2004 .name = "io_wait_time_recursive",
2005 .private = offsetof(struct cfq_group, stats.wait_time),
2006 .seq_show = cfqg_print_rwstat_recursive,
2009 .name = "io_merged_recursive",
2010 .private = offsetof(struct cfq_group, stats.merged),
2011 .seq_show = cfqg_print_rwstat_recursive,
2014 .name = "io_queued_recursive",
2015 .private = offsetof(struct cfq_group, stats.queued),
2016 .seq_show = cfqg_print_rwstat_recursive,
2018 #ifdef CONFIG_DEBUG_BLK_CGROUP
2020 .name = "avg_queue_size",
2021 .seq_show = cfqg_print_avg_queue_size,
2024 .name = "group_wait_time",
2025 .private = offsetof(struct cfq_group, stats.group_wait_time),
2026 .seq_show = cfqg_print_stat,
2029 .name = "idle_time",
2030 .private = offsetof(struct cfq_group, stats.idle_time),
2031 .seq_show = cfqg_print_stat,
2034 .name = "empty_time",
2035 .private = offsetof(struct cfq_group, stats.empty_time),
2036 .seq_show = cfqg_print_stat,
2040 .private = offsetof(struct cfq_group, stats.dequeue),
2041 .seq_show = cfqg_print_stat,
2044 .name = "unaccounted_time",
2045 .private = offsetof(struct cfq_group, stats.unaccounted_time),
2046 .seq_show = cfqg_print_stat,
2048 #endif /* CONFIG_DEBUG_BLK_CGROUP */
2051 #else /* GROUP_IOSCHED */
2052 static struct cfq_group *cfq_lookup_create_cfqg(struct cfq_data *cfqd,
2053 struct blkcg *blkcg)
2055 return cfqd->root_group;
2059 cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) {
2063 #endif /* GROUP_IOSCHED */
2066 * The cfqd->service_trees holds all pending cfq_queue's that have
2067 * requests waiting to be processed. It is sorted in the order that
2068 * we will service the queues.
2070 static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2073 struct rb_node **p, *parent;
2074 struct cfq_queue *__cfqq;
2075 unsigned long rb_key;
2076 struct cfq_rb_root *st;
2080 st = st_for(cfqq->cfqg, cfqq_class(cfqq), cfqq_type(cfqq));
2081 if (cfq_class_idle(cfqq)) {
2082 rb_key = CFQ_IDLE_DELAY;
2083 parent = rb_last(&st->rb);
2084 if (parent && parent != &cfqq->rb_node) {
2085 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
2086 rb_key += __cfqq->rb_key;
2089 } else if (!add_front) {
2091 * Get our rb key offset. Subtract any residual slice
2092 * value carried from last service. A negative resid
2093 * count indicates slice overrun, and this should position
2094 * the next service time further away in the tree.
2096 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
2097 rb_key -= cfqq->slice_resid;
2098 cfqq->slice_resid = 0;
2101 __cfqq = cfq_rb_first(st);
2102 rb_key += __cfqq ? __cfqq->rb_key : jiffies;
2105 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
2108 * same position, nothing more to do
2110 if (rb_key == cfqq->rb_key && cfqq->service_tree == st)
2113 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
2114 cfqq->service_tree = NULL;
2119 cfqq->service_tree = st;
2120 p = &st->rb.rb_node;
2123 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
2126 * sort by key, that represents service time.
2128 if (time_before(rb_key, __cfqq->rb_key))
2129 p = &parent->rb_left;
2131 p = &parent->rb_right;
2137 st->left = &cfqq->rb_node;
2139 cfqq->rb_key = rb_key;
2140 rb_link_node(&cfqq->rb_node, parent, p);
2141 rb_insert_color(&cfqq->rb_node, &st->rb);
2143 if (add_front || !new_cfqq)
2145 cfq_group_notify_queue_add(cfqd, cfqq->cfqg);
2148 static struct cfq_queue *
2149 cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
2150 sector_t sector, struct rb_node **ret_parent,
2151 struct rb_node ***rb_link)
2153 struct rb_node **p, *parent;
2154 struct cfq_queue *cfqq = NULL;
2162 cfqq = rb_entry(parent, struct cfq_queue, p_node);
2165 * Sort strictly based on sector. Smallest to the left,
2166 * largest to the right.
2168 if (sector > blk_rq_pos(cfqq->next_rq))
2169 n = &(*p)->rb_right;
2170 else if (sector < blk_rq_pos(cfqq->next_rq))
2178 *ret_parent = parent;
2184 static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2186 struct rb_node **p, *parent;
2187 struct cfq_queue *__cfqq;
2190 rb_erase(&cfqq->p_node, cfqq->p_root);
2191 cfqq->p_root = NULL;
2194 if (cfq_class_idle(cfqq))
2199 cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
2200 __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
2201 blk_rq_pos(cfqq->next_rq), &parent, &p);
2203 rb_link_node(&cfqq->p_node, parent, p);
2204 rb_insert_color(&cfqq->p_node, cfqq->p_root);
2206 cfqq->p_root = NULL;
2210 * Update cfqq's position in the service tree.
2212 static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2215 * Resorting requires the cfqq to be on the RR list already.
2217 if (cfq_cfqq_on_rr(cfqq)) {
2218 cfq_service_tree_add(cfqd, cfqq, 0);
2219 cfq_prio_tree_add(cfqd, cfqq);
2224 * add to busy list of queues for service, trying to be fair in ordering
2225 * the pending list according to last request service
2227 static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2229 cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
2230 BUG_ON(cfq_cfqq_on_rr(cfqq));
2231 cfq_mark_cfqq_on_rr(cfqq);
2232 cfqd->busy_queues++;
2233 if (cfq_cfqq_sync(cfqq))
2234 cfqd->busy_sync_queues++;
2236 cfq_resort_rr_list(cfqd, cfqq);
2240 * Called when the cfqq no longer has requests pending, remove it from
2243 static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2245 cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
2246 BUG_ON(!cfq_cfqq_on_rr(cfqq));
2247 cfq_clear_cfqq_on_rr(cfqq);
2249 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
2250 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
2251 cfqq->service_tree = NULL;
2254 rb_erase(&cfqq->p_node, cfqq->p_root);
2255 cfqq->p_root = NULL;
2258 cfq_group_notify_queue_del(cfqd, cfqq->cfqg);
2259 BUG_ON(!cfqd->busy_queues);
2260 cfqd->busy_queues--;
2261 if (cfq_cfqq_sync(cfqq))
2262 cfqd->busy_sync_queues--;
2266 * rb tree support functions
2268 static void cfq_del_rq_rb(struct request *rq)
2270 struct cfq_queue *cfqq = RQ_CFQQ(rq);
2271 const int sync = rq_is_sync(rq);
2273 BUG_ON(!cfqq->queued[sync]);
2274 cfqq->queued[sync]--;
2276 elv_rb_del(&cfqq->sort_list, rq);
2278 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) {
2280 * Queue will be deleted from service tree when we actually
2281 * expire it later. Right now just remove it from prio tree
2285 rb_erase(&cfqq->p_node, cfqq->p_root);
2286 cfqq->p_root = NULL;
2291 static void cfq_add_rq_rb(struct request *rq)
2293 struct cfq_queue *cfqq = RQ_CFQQ(rq);
2294 struct cfq_data *cfqd = cfqq->cfqd;
2295 struct request *prev;
2297 cfqq->queued[rq_is_sync(rq)]++;
2299 elv_rb_add(&cfqq->sort_list, rq);
2301 if (!cfq_cfqq_on_rr(cfqq))
2302 cfq_add_cfqq_rr(cfqd, cfqq);
2305 * check if this request is a better next-serve candidate
2307 prev = cfqq->next_rq;
2308 cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq, cfqd->last_position);
2311 * adjust priority tree position, if ->next_rq changes
2313 if (prev != cfqq->next_rq)
2314 cfq_prio_tree_add(cfqd, cfqq);
2316 BUG_ON(!cfqq->next_rq);
2319 static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
2321 elv_rb_del(&cfqq->sort_list, rq);
2322 cfqq->queued[rq_is_sync(rq)]--;
2323 cfqg_stats_update_io_remove(RQ_CFQG(rq), rq->cmd_flags);
2325 cfqg_stats_update_io_add(RQ_CFQG(rq), cfqq->cfqd->serving_group,
2329 static struct request *
2330 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
2332 struct task_struct *tsk = current;
2333 struct cfq_io_cq *cic;
2334 struct cfq_queue *cfqq;
2336 cic = cfq_cic_lookup(cfqd, tsk->io_context);
2340 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
2342 return elv_rb_find(&cfqq->sort_list, bio_end_sector(bio));
2347 static void cfq_activate_request(struct request_queue *q, struct request *rq)
2349 struct cfq_data *cfqd = q->elevator->elevator_data;
2351 cfqd->rq_in_driver++;
2352 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
2353 cfqd->rq_in_driver);
2355 cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
2358 static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
2360 struct cfq_data *cfqd = q->elevator->elevator_data;
2362 WARN_ON(!cfqd->rq_in_driver);
2363 cfqd->rq_in_driver--;
2364 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
2365 cfqd->rq_in_driver);
2368 static void cfq_remove_request(struct request *rq)
2370 struct cfq_queue *cfqq = RQ_CFQQ(rq);
2372 if (cfqq->next_rq == rq)
2373 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
2375 list_del_init(&rq->queuelist);
2378 cfqq->cfqd->rq_queued--;
2379 cfqg_stats_update_io_remove(RQ_CFQG(rq), rq->cmd_flags);
2380 if (rq->cmd_flags & REQ_PRIO) {
2381 WARN_ON(!cfqq->prio_pending);
2382 cfqq->prio_pending--;
2386 static int cfq_merge(struct request_queue *q, struct request **req,
2389 struct cfq_data *cfqd = q->elevator->elevator_data;
2390 struct request *__rq;
2392 __rq = cfq_find_rq_fmerge(cfqd, bio);
2393 if (__rq && elv_rq_merge_ok(__rq, bio)) {
2395 return ELEVATOR_FRONT_MERGE;
2398 return ELEVATOR_NO_MERGE;
2401 static void cfq_merged_request(struct request_queue *q, struct request *req,
2404 if (type == ELEVATOR_FRONT_MERGE) {
2405 struct cfq_queue *cfqq = RQ_CFQQ(req);
2407 cfq_reposition_rq_rb(cfqq, req);
2411 static void cfq_bio_merged(struct request_queue *q, struct request *req,
2414 cfqg_stats_update_io_merged(RQ_CFQG(req), bio->bi_rw);
2418 cfq_merged_requests(struct request_queue *q, struct request *rq,
2419 struct request *next)
2421 struct cfq_queue *cfqq = RQ_CFQQ(rq);
2422 struct cfq_data *cfqd = q->elevator->elevator_data;
2425 * reposition in fifo if next is older than rq
2427 if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
2428 time_before(next->fifo_time, rq->fifo_time) &&
2429 cfqq == RQ_CFQQ(next)) {
2430 list_move(&rq->queuelist, &next->queuelist);
2431 rq->fifo_time = next->fifo_time;
2434 if (cfqq->next_rq == next)
2436 cfq_remove_request(next);
2437 cfqg_stats_update_io_merged(RQ_CFQG(rq), next->cmd_flags);
2439 cfqq = RQ_CFQQ(next);
2441 * all requests of this queue are merged to other queues, delete it
2442 * from the service tree. If it's the active_queue,
2443 * cfq_dispatch_requests() will choose to expire it or do idle
2445 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list) &&
2446 cfqq != cfqd->active_queue)
2447 cfq_del_cfqq_rr(cfqd, cfqq);
2450 static int cfq_allow_merge(struct request_queue *q, struct request *rq,
2453 struct cfq_data *cfqd = q->elevator->elevator_data;
2454 struct cfq_io_cq *cic;
2455 struct cfq_queue *cfqq;
2458 * Disallow merge of a sync bio into an async request.
2460 if (cfq_bio_sync(bio) && !rq_is_sync(rq))
2464 * Lookup the cfqq that this bio will be queued with and allow
2465 * merge only if rq is queued there.
2467 cic = cfq_cic_lookup(cfqd, current->io_context);
2471 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
2472 return cfqq == RQ_CFQQ(rq);
2475 static inline void cfq_del_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2477 del_timer(&cfqd->idle_slice_timer);
2478 cfqg_stats_update_idle_time(cfqq->cfqg);
2481 static void __cfq_set_active_queue(struct cfq_data *cfqd,
2482 struct cfq_queue *cfqq)
2485 cfq_log_cfqq(cfqd, cfqq, "set_active wl_class:%d wl_type:%d",
2486 cfqd->serving_wl_class, cfqd->serving_wl_type);
2487 cfqg_stats_update_avg_queue_size(cfqq->cfqg);
2488 cfqq->slice_start = 0;
2489 cfqq->dispatch_start = jiffies;
2490 cfqq->allocated_slice = 0;
2491 cfqq->slice_end = 0;
2492 cfqq->slice_dispatch = 0;
2493 cfqq->nr_sectors = 0;
2495 cfq_clear_cfqq_wait_request(cfqq);
2496 cfq_clear_cfqq_must_dispatch(cfqq);
2497 cfq_clear_cfqq_must_alloc_slice(cfqq);
2498 cfq_clear_cfqq_fifo_expire(cfqq);
2499 cfq_mark_cfqq_slice_new(cfqq);
2501 cfq_del_timer(cfqd, cfqq);
2504 cfqd->active_queue = cfqq;
2508 * current cfqq expired its slice (or was too idle), select new one
2511 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2514 cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
2516 if (cfq_cfqq_wait_request(cfqq))
2517 cfq_del_timer(cfqd, cfqq);
2519 cfq_clear_cfqq_wait_request(cfqq);
2520 cfq_clear_cfqq_wait_busy(cfqq);
2523 * If this cfqq is shared between multiple processes, check to
2524 * make sure that those processes are still issuing I/Os within
2525 * the mean seek distance. If not, it may be time to break the
2526 * queues apart again.
2528 if (cfq_cfqq_coop(cfqq) && CFQQ_SEEKY(cfqq))
2529 cfq_mark_cfqq_split_coop(cfqq);
2532 * store what was left of this slice, if the queue idled/timed out
2535 if (cfq_cfqq_slice_new(cfqq))
2536 cfqq->slice_resid = cfq_scaled_cfqq_slice(cfqd, cfqq);
2538 cfqq->slice_resid = cfqq->slice_end - jiffies;
2539 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
2542 cfq_group_served(cfqd, cfqq->cfqg, cfqq);
2544 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
2545 cfq_del_cfqq_rr(cfqd, cfqq);
2547 cfq_resort_rr_list(cfqd, cfqq);
2549 if (cfqq == cfqd->active_queue)
2550 cfqd->active_queue = NULL;
2552 if (cfqd->active_cic) {
2553 put_io_context(cfqd->active_cic->icq.ioc);
2554 cfqd->active_cic = NULL;
2558 static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out)
2560 struct cfq_queue *cfqq = cfqd->active_queue;
2563 __cfq_slice_expired(cfqd, cfqq, timed_out);
2567 * Get next queue for service. Unless we have a queue preemption,
2568 * we'll simply select the first cfqq in the service tree.
2570 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
2572 struct cfq_rb_root *st = st_for(cfqd->serving_group,
2573 cfqd->serving_wl_class, cfqd->serving_wl_type);
2575 if (!cfqd->rq_queued)
2578 /* There is nothing to dispatch */
2581 if (RB_EMPTY_ROOT(&st->rb))
2583 return cfq_rb_first(st);
2586 static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd)
2588 struct cfq_group *cfqg;
2589 struct cfq_queue *cfqq;
2591 struct cfq_rb_root *st;
2593 if (!cfqd->rq_queued)
2596 cfqg = cfq_get_next_cfqg(cfqd);
2600 for_each_cfqg_st(cfqg, i, j, st)
2601 if ((cfqq = cfq_rb_first(st)) != NULL)
2607 * Get and set a new active queue for service.
2609 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
2610 struct cfq_queue *cfqq)
2613 cfqq = cfq_get_next_queue(cfqd);
2615 __cfq_set_active_queue(cfqd, cfqq);
2619 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
2622 if (blk_rq_pos(rq) >= cfqd->last_position)
2623 return blk_rq_pos(rq) - cfqd->last_position;
2625 return cfqd->last_position - blk_rq_pos(rq);
2628 static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2631 return cfq_dist_from_last(cfqd, rq) <= CFQQ_CLOSE_THR;
2634 static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
2635 struct cfq_queue *cur_cfqq)
2637 struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
2638 struct rb_node *parent, *node;
2639 struct cfq_queue *__cfqq;
2640 sector_t sector = cfqd->last_position;
2642 if (RB_EMPTY_ROOT(root))
2646 * First, if we find a request starting at the end of the last
2647 * request, choose it.
2649 __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
2654 * If the exact sector wasn't found, the parent of the NULL leaf
2655 * will contain the closest sector.
2657 __cfqq = rb_entry(parent, struct cfq_queue, p_node);
2658 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
2661 if (blk_rq_pos(__cfqq->next_rq) < sector)
2662 node = rb_next(&__cfqq->p_node);
2664 node = rb_prev(&__cfqq->p_node);
2668 __cfqq = rb_entry(node, struct cfq_queue, p_node);
2669 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
2677 * cur_cfqq - passed in so that we don't decide that the current queue is
2678 * closely cooperating with itself.
2680 * So, basically we're assuming that that cur_cfqq has dispatched at least
2681 * one request, and that cfqd->last_position reflects a position on the disk
2682 * associated with the I/O issued by cur_cfqq. I'm not sure this is a valid
2685 static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
2686 struct cfq_queue *cur_cfqq)
2688 struct cfq_queue *cfqq;
2690 if (cfq_class_idle(cur_cfqq))
2692 if (!cfq_cfqq_sync(cur_cfqq))
2694 if (CFQQ_SEEKY(cur_cfqq))
2698 * Don't search priority tree if it's the only queue in the group.
2700 if (cur_cfqq->cfqg->nr_cfqq == 1)
2704 * We should notice if some of the queues are cooperating, eg
2705 * working closely on the same area of the disk. In that case,
2706 * we can group them together and don't waste time idling.
2708 cfqq = cfqq_close(cfqd, cur_cfqq);
2712 /* If new queue belongs to different cfq_group, don't choose it */
2713 if (cur_cfqq->cfqg != cfqq->cfqg)
2717 * It only makes sense to merge sync queues.
2719 if (!cfq_cfqq_sync(cfqq))
2721 if (CFQQ_SEEKY(cfqq))
2725 * Do not merge queues of different priority classes
2727 if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq))
2734 * Determine whether we should enforce idle window for this queue.
2737 static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2739 enum wl_class_t wl_class = cfqq_class(cfqq);
2740 struct cfq_rb_root *st = cfqq->service_tree;
2745 if (!cfqd->cfq_slice_idle)
2748 /* We never do for idle class queues. */
2749 if (wl_class == IDLE_WORKLOAD)
2752 /* We do for queues that were marked with idle window flag. */
2753 if (cfq_cfqq_idle_window(cfqq) &&
2754 !(blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag))
2758 * Otherwise, we do only if they are the last ones
2759 * in their service tree.
2761 if (st->count == 1 && cfq_cfqq_sync(cfqq) &&
2762 !cfq_io_thinktime_big(cfqd, &st->ttime, false))
2764 cfq_log_cfqq(cfqd, cfqq, "Not idling. st->count:%d", st->count);
2768 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
2770 struct cfq_queue *cfqq = cfqd->active_queue;
2771 struct cfq_io_cq *cic;
2772 unsigned long sl, group_idle = 0;
2775 * SSD device without seek penalty, disable idling. But only do so
2776 * for devices that support queuing, otherwise we still have a problem
2777 * with sync vs async workloads.
2779 if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
2782 WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
2783 WARN_ON(cfq_cfqq_slice_new(cfqq));
2786 * idle is disabled, either manually or by past process history
2788 if (!cfq_should_idle(cfqd, cfqq)) {
2789 /* no queue idling. Check for group idling */
2790 if (cfqd->cfq_group_idle)
2791 group_idle = cfqd->cfq_group_idle;
2797 * still active requests from this queue, don't idle
2799 if (cfqq->dispatched)
2803 * task has exited, don't wait
2805 cic = cfqd->active_cic;
2806 if (!cic || !atomic_read(&cic->icq.ioc->active_ref))
2810 * If our average think time is larger than the remaining time
2811 * slice, then don't idle. This avoids overrunning the allotted
2814 if (sample_valid(cic->ttime.ttime_samples) &&
2815 (cfqq->slice_end - jiffies < cic->ttime.ttime_mean)) {
2816 cfq_log_cfqq(cfqd, cfqq, "Not idling. think_time:%lu",
2817 cic->ttime.ttime_mean);
2821 /* There are other queues in the group, don't do group idle */
2822 if (group_idle && cfqq->cfqg->nr_cfqq > 1)
2825 cfq_mark_cfqq_wait_request(cfqq);
2828 sl = cfqd->cfq_group_idle;
2830 sl = cfqd->cfq_slice_idle;
2832 mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
2833 cfqg_stats_set_start_idle_time(cfqq->cfqg);
2834 cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu group_idle: %d", sl,
2835 group_idle ? 1 : 0);
2839 * Move request from internal lists to the request queue dispatch list.
2841 static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
2843 struct cfq_data *cfqd = q->elevator->elevator_data;
2844 struct cfq_queue *cfqq = RQ_CFQQ(rq);
2846 cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
2848 cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
2849 cfq_remove_request(rq);
2851 (RQ_CFQG(rq))->dispatched++;
2852 elv_dispatch_sort(q, rq);
2854 cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]++;
2855 cfqq->nr_sectors += blk_rq_sectors(rq);
2856 cfqg_stats_update_dispatch(cfqq->cfqg, blk_rq_bytes(rq), rq->cmd_flags);
2860 * return expired entry, or NULL to just start from scratch in rbtree
2862 static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
2864 struct request *rq = NULL;
2866 if (cfq_cfqq_fifo_expire(cfqq))
2869 cfq_mark_cfqq_fifo_expire(cfqq);
2871 if (list_empty(&cfqq->fifo))
2874 rq = rq_entry_fifo(cfqq->fifo.next);
2875 if (time_before(jiffies, rq->fifo_time))
2878 cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq);
2883 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2885 const int base_rq = cfqd->cfq_slice_async_rq;
2887 WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
2889 return 2 * base_rq * (IOPRIO_BE_NR - cfqq->ioprio);
2893 * Must be called with the queue_lock held.
2895 static int cfqq_process_refs(struct cfq_queue *cfqq)
2897 int process_refs, io_refs;
2899 io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE];
2900 process_refs = cfqq->ref - io_refs;
2901 BUG_ON(process_refs < 0);
2902 return process_refs;
2905 static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq)
2907 int process_refs, new_process_refs;
2908 struct cfq_queue *__cfqq;
2911 * If there are no process references on the new_cfqq, then it is
2912 * unsafe to follow the ->new_cfqq chain as other cfqq's in the
2913 * chain may have dropped their last reference (not just their
2914 * last process reference).
2916 if (!cfqq_process_refs(new_cfqq))
2919 /* Avoid a circular list and skip interim queue merges */
2920 while ((__cfqq = new_cfqq->new_cfqq)) {
2926 process_refs = cfqq_process_refs(cfqq);
2927 new_process_refs = cfqq_process_refs(new_cfqq);
2929 * If the process for the cfqq has gone away, there is no
2930 * sense in merging the queues.
2932 if (process_refs == 0 || new_process_refs == 0)
2936 * Merge in the direction of the lesser amount of work.
2938 if (new_process_refs >= process_refs) {
2939 cfqq->new_cfqq = new_cfqq;
2940 new_cfqq->ref += process_refs;
2942 new_cfqq->new_cfqq = cfqq;
2943 cfqq->ref += new_process_refs;
2947 static enum wl_type_t cfq_choose_wl_type(struct cfq_data *cfqd,
2948 struct cfq_group *cfqg, enum wl_class_t wl_class)
2950 struct cfq_queue *queue;
2952 bool key_valid = false;
2953 unsigned long lowest_key = 0;
2954 enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD;
2956 for (i = 0; i <= SYNC_WORKLOAD; ++i) {
2957 /* select the one with lowest rb_key */
2958 queue = cfq_rb_first(st_for(cfqg, wl_class, i));
2960 (!key_valid || time_before(queue->rb_key, lowest_key))) {
2961 lowest_key = queue->rb_key;
2971 choose_wl_class_and_type(struct cfq_data *cfqd, struct cfq_group *cfqg)
2975 struct cfq_rb_root *st;
2976 unsigned group_slice;
2977 enum wl_class_t original_class = cfqd->serving_wl_class;
2979 /* Choose next priority. RT > BE > IDLE */
2980 if (cfq_group_busy_queues_wl(RT_WORKLOAD, cfqd, cfqg))
2981 cfqd->serving_wl_class = RT_WORKLOAD;
2982 else if (cfq_group_busy_queues_wl(BE_WORKLOAD, cfqd, cfqg))
2983 cfqd->serving_wl_class = BE_WORKLOAD;
2985 cfqd->serving_wl_class = IDLE_WORKLOAD;
2986 cfqd->workload_expires = jiffies + 1;
2990 if (original_class != cfqd->serving_wl_class)
2994 * For RT and BE, we have to choose also the type
2995 * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
2998 st = st_for(cfqg, cfqd->serving_wl_class, cfqd->serving_wl_type);
3002 * check workload expiration, and that we still have other queues ready
3004 if (count && !time_after(jiffies, cfqd->workload_expires))
3008 /* otherwise select new workload type */
3009 cfqd->serving_wl_type = cfq_choose_wl_type(cfqd, cfqg,
3010 cfqd->serving_wl_class);
3011 st = st_for(cfqg, cfqd->serving_wl_class, cfqd->serving_wl_type);
3015 * the workload slice is computed as a fraction of target latency
3016 * proportional to the number of queues in that workload, over
3017 * all the queues in the same priority class
3019 group_slice = cfq_group_slice(cfqd, cfqg);
3021 slice = group_slice * count /
3022 max_t(unsigned, cfqg->busy_queues_avg[cfqd->serving_wl_class],
3023 cfq_group_busy_queues_wl(cfqd->serving_wl_class, cfqd,
3026 if (cfqd->serving_wl_type == ASYNC_WORKLOAD) {
3030 * Async queues are currently system wide. Just taking
3031 * proportion of queues with-in same group will lead to higher
3032 * async ratio system wide as generally root group is going
3033 * to have higher weight. A more accurate thing would be to
3034 * calculate system wide asnc/sync ratio.
3036 tmp = cfqd->cfq_target_latency *
3037 cfqg_busy_async_queues(cfqd, cfqg);
3038 tmp = tmp/cfqd->busy_queues;
3039 slice = min_t(unsigned, slice, tmp);
3041 /* async workload slice is scaled down according to
3042 * the sync/async slice ratio. */
3043 slice = slice * cfqd->cfq_slice[0] / cfqd->cfq_slice[1];
3045 /* sync workload slice is at least 2 * cfq_slice_idle */
3046 slice = max(slice, 2 * cfqd->cfq_slice_idle);
3048 slice = max_t(unsigned, slice, CFQ_MIN_TT);
3049 cfq_log(cfqd, "workload slice:%d", slice);
3050 cfqd->workload_expires = jiffies + slice;
3053 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd)
3055 struct cfq_rb_root *st = &cfqd->grp_service_tree;
3056 struct cfq_group *cfqg;
3058 if (RB_EMPTY_ROOT(&st->rb))
3060 cfqg = cfq_rb_first_group(st);
3061 update_min_vdisktime(st);
3065 static void cfq_choose_cfqg(struct cfq_data *cfqd)
3067 struct cfq_group *cfqg = cfq_get_next_cfqg(cfqd);
3069 cfqd->serving_group = cfqg;
3071 /* Restore the workload type data */
3072 if (cfqg->saved_wl_slice) {
3073 cfqd->workload_expires = jiffies + cfqg->saved_wl_slice;
3074 cfqd->serving_wl_type = cfqg->saved_wl_type;
3075 cfqd->serving_wl_class = cfqg->saved_wl_class;
3077 cfqd->workload_expires = jiffies - 1;
3079 choose_wl_class_and_type(cfqd, cfqg);
3083 * Select a queue for service. If we have a current active queue,
3084 * check whether to continue servicing it, or retrieve and set a new one.
3086 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
3088 struct cfq_queue *cfqq, *new_cfqq = NULL;
3090 cfqq = cfqd->active_queue;
3094 if (!cfqd->rq_queued)
3098 * We were waiting for group to get backlogged. Expire the queue
3100 if (cfq_cfqq_wait_busy(cfqq) && !RB_EMPTY_ROOT(&cfqq->sort_list))
3104 * The active queue has run out of time, expire it and select new.
3106 if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq)) {
3108 * If slice had not expired at the completion of last request
3109 * we might not have turned on wait_busy flag. Don't expire
3110 * the queue yet. Allow the group to get backlogged.
3112 * The very fact that we have used the slice, that means we
3113 * have been idling all along on this queue and it should be
3114 * ok to wait for this request to complete.
3116 if (cfqq->cfqg->nr_cfqq == 1 && RB_EMPTY_ROOT(&cfqq->sort_list)
3117 && cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
3121 goto check_group_idle;
3125 * The active queue has requests and isn't expired, allow it to
3128 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3132 * If another queue has a request waiting within our mean seek
3133 * distance, let it run. The expire code will check for close
3134 * cooperators and put the close queue at the front of the service
3135 * tree. If possible, merge the expiring queue with the new cfqq.
3137 new_cfqq = cfq_close_cooperator(cfqd, cfqq);
3139 if (!cfqq->new_cfqq)
3140 cfq_setup_merge(cfqq, new_cfqq);
3145 * No requests pending. If the active queue still has requests in
3146 * flight or is idling for a new request, allow either of these
3147 * conditions to happen (or time out) before selecting a new queue.
3149 if (timer_pending(&cfqd->idle_slice_timer)) {
3155 * This is a deep seek queue, but the device is much faster than
3156 * the queue can deliver, don't idle
3158 if (CFQQ_SEEKY(cfqq) && cfq_cfqq_idle_window(cfqq) &&
3159 (cfq_cfqq_slice_new(cfqq) ||
3160 (cfqq->slice_end - jiffies > jiffies - cfqq->slice_start))) {
3161 cfq_clear_cfqq_deep(cfqq);
3162 cfq_clear_cfqq_idle_window(cfqq);
3165 if (cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
3171 * If group idle is enabled and there are requests dispatched from
3172 * this group, wait for requests to complete.
3175 if (cfqd->cfq_group_idle && cfqq->cfqg->nr_cfqq == 1 &&
3176 cfqq->cfqg->dispatched &&
3177 !cfq_io_thinktime_big(cfqd, &cfqq->cfqg->ttime, true)) {
3183 cfq_slice_expired(cfqd, 0);
3186 * Current queue expired. Check if we have to switch to a new
3190 cfq_choose_cfqg(cfqd);
3192 cfqq = cfq_set_active_queue(cfqd, new_cfqq);
3197 static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
3201 while (cfqq->next_rq) {
3202 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
3206 BUG_ON(!list_empty(&cfqq->fifo));
3208 /* By default cfqq is not expired if it is empty. Do it explicitly */
3209 __cfq_slice_expired(cfqq->cfqd, cfqq, 0);
3214 * Drain our current requests. Used for barriers and when switching
3215 * io schedulers on-the-fly.
3217 static int cfq_forced_dispatch(struct cfq_data *cfqd)
3219 struct cfq_queue *cfqq;
3222 /* Expire the timeslice of the current active queue first */
3223 cfq_slice_expired(cfqd, 0);
3224 while ((cfqq = cfq_get_next_queue_forced(cfqd)) != NULL) {
3225 __cfq_set_active_queue(cfqd, cfqq);
3226 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
3229 BUG_ON(cfqd->busy_queues);
3231 cfq_log(cfqd, "forced_dispatch=%d", dispatched);
3235 static inline bool cfq_slice_used_soon(struct cfq_data *cfqd,
3236 struct cfq_queue *cfqq)
3238 /* the queue hasn't finished any request, can't estimate */
3239 if (cfq_cfqq_slice_new(cfqq))
3241 if (time_after(jiffies + cfqd->cfq_slice_idle * cfqq->dispatched,
3248 static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3250 unsigned int max_dispatch;
3253 * Drain async requests before we start sync IO
3255 if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_flight[BLK_RW_ASYNC])
3259 * If this is an async queue and we have sync IO in flight, let it wait
3261 if (cfqd->rq_in_flight[BLK_RW_SYNC] && !cfq_cfqq_sync(cfqq))
3264 max_dispatch = max_t(unsigned int, cfqd->cfq_quantum / 2, 1);
3265 if (cfq_class_idle(cfqq))
3269 * Does this cfqq already have too much IO in flight?
3271 if (cfqq->dispatched >= max_dispatch) {
3272 bool promote_sync = false;
3274 * idle queue must always only have a single IO in flight
3276 if (cfq_class_idle(cfqq))
3280 * If there is only one sync queue
3281 * we can ignore async queue here and give the sync
3282 * queue no dispatch limit. The reason is a sync queue can
3283 * preempt async queue, limiting the sync queue doesn't make
3284 * sense. This is useful for aiostress test.
3286 if (cfq_cfqq_sync(cfqq) && cfqd->busy_sync_queues == 1)
3287 promote_sync = true;
3290 * We have other queues, don't allow more IO from this one
3292 if (cfqd->busy_queues > 1 && cfq_slice_used_soon(cfqd, cfqq) &&
3297 * Sole queue user, no limit
3299 if (cfqd->busy_queues == 1 || promote_sync)
3303 * Normally we start throttling cfqq when cfq_quantum/2
3304 * requests have been dispatched. But we can drive
3305 * deeper queue depths at the beginning of slice
3306 * subjected to upper limit of cfq_quantum.
3308 max_dispatch = cfqd->cfq_quantum;
3312 * Async queues must wait a bit before being allowed dispatch.
3313 * We also ramp up the dispatch depth gradually for async IO,
3314 * based on the last sync IO we serviced
3316 if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
3317 unsigned long last_sync = jiffies - cfqd->last_delayed_sync;
3320 depth = last_sync / cfqd->cfq_slice[1];
3321 if (!depth && !cfqq->dispatched)
3323 if (depth < max_dispatch)
3324 max_dispatch = depth;
3328 * If we're below the current max, allow a dispatch
3330 return cfqq->dispatched < max_dispatch;
3334 * Dispatch a request from cfqq, moving them to the request queue
3337 static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3341 BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
3343 if (!cfq_may_dispatch(cfqd, cfqq))
3347 * follow expired path, else get first next available
3349 rq = cfq_check_fifo(cfqq);
3354 * insert request into driver dispatch list
3356 cfq_dispatch_insert(cfqd->queue, rq);
3358 if (!cfqd->active_cic) {
3359 struct cfq_io_cq *cic = RQ_CIC(rq);
3361 atomic_long_inc(&cic->icq.ioc->refcount);
3362 cfqd->active_cic = cic;
3369 * Find the cfqq that we need to service and move a request from that to the
3372 static int cfq_dispatch_requests(struct request_queue *q, int force)
3374 struct cfq_data *cfqd = q->elevator->elevator_data;
3375 struct cfq_queue *cfqq;
3377 if (!cfqd->busy_queues)
3380 if (unlikely(force))
3381 return cfq_forced_dispatch(cfqd);
3383 cfqq = cfq_select_queue(cfqd);
3388 * Dispatch a request from this cfqq, if it is allowed
3390 if (!cfq_dispatch_request(cfqd, cfqq))
3393 cfqq->slice_dispatch++;
3394 cfq_clear_cfqq_must_dispatch(cfqq);
3397 * expire an async queue immediately if it has used up its slice. idle
3398 * queue always expire after 1 dispatch round.
3400 if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
3401 cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
3402 cfq_class_idle(cfqq))) {
3403 cfqq->slice_end = jiffies + 1;
3404 cfq_slice_expired(cfqd, 0);
3407 cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
3412 * task holds one reference to the queue, dropped when task exits. each rq
3413 * in-flight on this queue also holds a reference, dropped when rq is freed.
3415 * Each cfq queue took a reference on the parent group. Drop it now.
3416 * queue lock must be held here.
3418 static void cfq_put_queue(struct cfq_queue *cfqq)
3420 struct cfq_data *cfqd = cfqq->cfqd;
3421 struct cfq_group *cfqg;
3423 BUG_ON(cfqq->ref <= 0);
3429 cfq_log_cfqq(cfqd, cfqq, "put_queue");
3430 BUG_ON(rb_first(&cfqq->sort_list));
3431 BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
3434 if (unlikely(cfqd->active_queue == cfqq)) {
3435 __cfq_slice_expired(cfqd, cfqq, 0);
3436 cfq_schedule_dispatch(cfqd);
3439 BUG_ON(cfq_cfqq_on_rr(cfqq));
3440 kmem_cache_free(cfq_pool, cfqq);
3444 static void cfq_put_cooperator(struct cfq_queue *cfqq)
3446 struct cfq_queue *__cfqq, *next;
3449 * If this queue was scheduled to merge with another queue, be
3450 * sure to drop the reference taken on that queue (and others in
3451 * the merge chain). See cfq_setup_merge and cfq_merge_cfqqs.
3453 __cfqq = cfqq->new_cfqq;
3455 if (__cfqq == cfqq) {
3456 WARN(1, "cfqq->new_cfqq loop detected\n");
3459 next = __cfqq->new_cfqq;
3460 cfq_put_queue(__cfqq);
3465 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3467 if (unlikely(cfqq == cfqd->active_queue)) {
3468 __cfq_slice_expired(cfqd, cfqq, 0);
3469 cfq_schedule_dispatch(cfqd);
3472 cfq_put_cooperator(cfqq);
3474 cfq_put_queue(cfqq);
3477 static void cfq_init_icq(struct io_cq *icq)
3479 struct cfq_io_cq *cic = icq_to_cic(icq);
3481 cic->ttime.last_end_request = jiffies;
3484 static void cfq_exit_icq(struct io_cq *icq)
3486 struct cfq_io_cq *cic = icq_to_cic(icq);
3487 struct cfq_data *cfqd = cic_to_cfqd(cic);
3489 if (cic->cfqq[BLK_RW_ASYNC]) {
3490 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]);
3491 cic->cfqq[BLK_RW_ASYNC] = NULL;
3494 if (cic->cfqq[BLK_RW_SYNC]) {
3495 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]);
3496 cic->cfqq[BLK_RW_SYNC] = NULL;
3500 static void cfq_init_prio_data(struct cfq_queue *cfqq, struct cfq_io_cq *cic)
3502 struct task_struct *tsk = current;
3505 if (!cfq_cfqq_prio_changed(cfqq))
3508 ioprio_class = IOPRIO_PRIO_CLASS(cic->ioprio);
3509 switch (ioprio_class) {
3511 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
3512 case IOPRIO_CLASS_NONE:
3514 * no prio set, inherit CPU scheduling settings
3516 cfqq->ioprio = task_nice_ioprio(tsk);
3517 cfqq->ioprio_class = task_nice_ioclass(tsk);
3519 case IOPRIO_CLASS_RT:
3520 cfqq->ioprio = IOPRIO_PRIO_DATA(cic->ioprio);
3521 cfqq->ioprio_class = IOPRIO_CLASS_RT;
3523 case IOPRIO_CLASS_BE:
3524 cfqq->ioprio = IOPRIO_PRIO_DATA(cic->ioprio);
3525 cfqq->ioprio_class = IOPRIO_CLASS_BE;
3527 case IOPRIO_CLASS_IDLE:
3528 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
3530 cfq_clear_cfqq_idle_window(cfqq);
3535 * keep track of original prio settings in case we have to temporarily
3536 * elevate the priority of this queue
3538 cfqq->org_ioprio = cfqq->ioprio;
3539 cfq_clear_cfqq_prio_changed(cfqq);
3542 static void check_ioprio_changed(struct cfq_io_cq *cic, struct bio *bio)
3544 int ioprio = cic->icq.ioc->ioprio;
3545 struct cfq_data *cfqd = cic_to_cfqd(cic);
3546 struct cfq_queue *cfqq;
3549 * Check whether ioprio has changed. The condition may trigger
3550 * spuriously on a newly created cic but there's no harm.
3552 if (unlikely(!cfqd) || likely(cic->ioprio == ioprio))
3555 cfqq = cic->cfqq[BLK_RW_ASYNC];
3557 struct cfq_queue *new_cfqq;
3558 new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic, bio,
3561 cic->cfqq[BLK_RW_ASYNC] = new_cfqq;
3562 cfq_put_queue(cfqq);
3566 cfqq = cic->cfqq[BLK_RW_SYNC];
3568 cfq_mark_cfqq_prio_changed(cfqq);
3570 cic->ioprio = ioprio;
3573 static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3574 pid_t pid, bool is_sync)
3576 RB_CLEAR_NODE(&cfqq->rb_node);
3577 RB_CLEAR_NODE(&cfqq->p_node);
3578 INIT_LIST_HEAD(&cfqq->fifo);
3583 cfq_mark_cfqq_prio_changed(cfqq);
3586 if (!cfq_class_idle(cfqq))
3587 cfq_mark_cfqq_idle_window(cfqq);
3588 cfq_mark_cfqq_sync(cfqq);
3593 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3594 static void check_blkcg_changed(struct cfq_io_cq *cic, struct bio *bio)
3596 struct cfq_data *cfqd = cic_to_cfqd(cic);
3597 struct cfq_queue *sync_cfqq;
3601 serial_nr = bio_blkcg(bio)->css.serial_nr;
3605 * Check whether blkcg has changed. The condition may trigger
3606 * spuriously on a newly created cic but there's no harm.
3608 if (unlikely(!cfqd) || likely(cic->blkcg_serial_nr == serial_nr))
3611 sync_cfqq = cic_to_cfqq(cic, 1);
3614 * Drop reference to sync queue. A new sync queue will be
3615 * assigned in new group upon arrival of a fresh request.
3617 cfq_log_cfqq(cfqd, sync_cfqq, "changed cgroup");
3618 cic_set_cfqq(cic, NULL, 1);
3619 cfq_put_queue(sync_cfqq);
3622 cic->blkcg_serial_nr = serial_nr;
3625 static inline void check_blkcg_changed(struct cfq_io_cq *cic, struct bio *bio) { }
3626 #endif /* CONFIG_CFQ_GROUP_IOSCHED */
3628 static struct cfq_queue *
3629 cfq_find_alloc_queue(struct cfq_data *cfqd, bool is_sync, struct cfq_io_cq *cic,
3630 struct bio *bio, gfp_t gfp_mask)
3632 struct blkcg *blkcg;
3633 struct cfq_queue *cfqq, *new_cfqq = NULL;
3634 struct cfq_group *cfqg;
3639 blkcg = bio_blkcg(bio);
3640 cfqg = cfq_lookup_create_cfqg(cfqd, blkcg);
3642 cfqq = &cfqd->oom_cfqq;
3646 cfqq = cic_to_cfqq(cic, is_sync);
3649 * Always try a new alloc if we fell back to the OOM cfqq
3650 * originally, since it should just be a temporary situation.
3652 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
3657 } else if (gfp_mask & __GFP_WAIT) {
3659 spin_unlock_irq(cfqd->queue->queue_lock);
3660 new_cfqq = kmem_cache_alloc_node(cfq_pool,
3661 gfp_mask | __GFP_ZERO,
3663 spin_lock_irq(cfqd->queue->queue_lock);
3667 return &cfqd->oom_cfqq;
3669 cfqq = kmem_cache_alloc_node(cfq_pool,
3670 gfp_mask | __GFP_ZERO,
3675 cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
3676 cfq_init_prio_data(cfqq, cic);
3677 cfq_link_cfqq_cfqg(cfqq, cfqg);
3678 cfq_log_cfqq(cfqd, cfqq, "alloced");
3680 cfqq = &cfqd->oom_cfqq;
3684 kmem_cache_free(cfq_pool, new_cfqq);
3690 static struct cfq_queue **
3691 cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
3693 switch (ioprio_class) {
3694 case IOPRIO_CLASS_RT:
3695 return &cfqd->async_cfqq[0][ioprio];
3696 case IOPRIO_CLASS_NONE:
3697 ioprio = IOPRIO_NORM;
3699 case IOPRIO_CLASS_BE:
3700 return &cfqd->async_cfqq[1][ioprio];
3701 case IOPRIO_CLASS_IDLE:
3702 return &cfqd->async_idle_cfqq;
3708 static struct cfq_queue *
3709 cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct cfq_io_cq *cic,
3710 struct bio *bio, gfp_t gfp_mask)
3712 int ioprio_class = IOPRIO_PRIO_CLASS(cic->ioprio);
3713 int ioprio = IOPRIO_PRIO_DATA(cic->ioprio);
3714 struct cfq_queue **async_cfqq = NULL;
3715 struct cfq_queue *cfqq = NULL;
3718 if (!ioprio_valid(cic->ioprio)) {
3719 struct task_struct *tsk = current;
3720 ioprio = task_nice_ioprio(tsk);
3721 ioprio_class = task_nice_ioclass(tsk);
3723 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
3728 cfqq = cfq_find_alloc_queue(cfqd, is_sync, cic, bio, gfp_mask);
3731 * pin the queue now that it's allocated, scheduler exit will prune it
3733 if (!is_sync && !(*async_cfqq)) {
3743 __cfq_update_io_thinktime(struct cfq_ttime *ttime, unsigned long slice_idle)
3745 unsigned long elapsed = jiffies - ttime->last_end_request;
3746 elapsed = min(elapsed, 2UL * slice_idle);
3748 ttime->ttime_samples = (7*ttime->ttime_samples + 256) / 8;
3749 ttime->ttime_total = (7*ttime->ttime_total + 256*elapsed) / 8;
3750 ttime->ttime_mean = (ttime->ttime_total + 128) / ttime->ttime_samples;
3754 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3755 struct cfq_io_cq *cic)
3757 if (cfq_cfqq_sync(cfqq)) {
3758 __cfq_update_io_thinktime(&cic->ttime, cfqd->cfq_slice_idle);
3759 __cfq_update_io_thinktime(&cfqq->service_tree->ttime,
3760 cfqd->cfq_slice_idle);
3762 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3763 __cfq_update_io_thinktime(&cfqq->cfqg->ttime, cfqd->cfq_group_idle);
3768 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3772 sector_t n_sec = blk_rq_sectors(rq);
3773 if (cfqq->last_request_pos) {
3774 if (cfqq->last_request_pos < blk_rq_pos(rq))
3775 sdist = blk_rq_pos(rq) - cfqq->last_request_pos;
3777 sdist = cfqq->last_request_pos - blk_rq_pos(rq);
3780 cfqq->seek_history <<= 1;
3781 if (blk_queue_nonrot(cfqd->queue))
3782 cfqq->seek_history |= (n_sec < CFQQ_SECT_THR_NONROT);
3784 cfqq->seek_history |= (sdist > CFQQ_SEEK_THR);
3788 * Disable idle window if the process thinks too long or seeks so much that
3792 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3793 struct cfq_io_cq *cic)
3795 int old_idle, enable_idle;
3798 * Don't idle for async or idle io prio class
3800 if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
3803 enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
3805 if (cfqq->queued[0] + cfqq->queued[1] >= 4)
3806 cfq_mark_cfqq_deep(cfqq);
3808 if (cfqq->next_rq && (cfqq->next_rq->cmd_flags & REQ_NOIDLE))
3810 else if (!atomic_read(&cic->icq.ioc->active_ref) ||
3811 !cfqd->cfq_slice_idle ||
3812 (!cfq_cfqq_deep(cfqq) && CFQQ_SEEKY(cfqq)))
3814 else if (sample_valid(cic->ttime.ttime_samples)) {
3815 if (cic->ttime.ttime_mean > cfqd->cfq_slice_idle)
3821 if (old_idle != enable_idle) {
3822 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
3824 cfq_mark_cfqq_idle_window(cfqq);
3826 cfq_clear_cfqq_idle_window(cfqq);
3831 * Check if new_cfqq should preempt the currently active queue. Return 0 for
3832 * no or if we aren't sure, a 1 will cause a preempt.
3835 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
3838 struct cfq_queue *cfqq;
3840 cfqq = cfqd->active_queue;
3844 if (cfq_class_idle(new_cfqq))
3847 if (cfq_class_idle(cfqq))
3851 * Don't allow a non-RT request to preempt an ongoing RT cfqq timeslice.
3853 if (cfq_class_rt(cfqq) && !cfq_class_rt(new_cfqq))
3857 * if the new request is sync, but the currently running queue is
3858 * not, let the sync request have priority.
3860 if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
3863 if (new_cfqq->cfqg != cfqq->cfqg)
3866 if (cfq_slice_used(cfqq))
3869 /* Allow preemption only if we are idling on sync-noidle tree */
3870 if (cfqd->serving_wl_type == SYNC_NOIDLE_WORKLOAD &&
3871 cfqq_type(new_cfqq) == SYNC_NOIDLE_WORKLOAD &&
3872 new_cfqq->service_tree->count == 2 &&
3873 RB_EMPTY_ROOT(&cfqq->sort_list))
3877 * So both queues are sync. Let the new request get disk time if
3878 * it's a metadata request and the current queue is doing regular IO.
3880 if ((rq->cmd_flags & REQ_PRIO) && !cfqq->prio_pending)
3884 * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
3886 if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
3889 /* An idle queue should not be idle now for some reason */
3890 if (RB_EMPTY_ROOT(&cfqq->sort_list) && !cfq_should_idle(cfqd, cfqq))
3893 if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
3897 * if this request is as-good as one we would expect from the
3898 * current cfqq, let it preempt
3900 if (cfq_rq_close(cfqd, cfqq, rq))
3907 * cfqq preempts the active queue. if we allowed preempt with no slice left,
3908 * let it have half of its nominal slice.
3910 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3912 enum wl_type_t old_type = cfqq_type(cfqd->active_queue);
3914 cfq_log_cfqq(cfqd, cfqq, "preempt");
3915 cfq_slice_expired(cfqd, 1);
3918 * workload type is changed, don't save slice, otherwise preempt
3921 if (old_type != cfqq_type(cfqq))
3922 cfqq->cfqg->saved_wl_slice = 0;
3925 * Put the new queue at the front of the of the current list,
3926 * so we know that it will be selected next.
3928 BUG_ON(!cfq_cfqq_on_rr(cfqq));
3930 cfq_service_tree_add(cfqd, cfqq, 1);
3932 cfqq->slice_end = 0;
3933 cfq_mark_cfqq_slice_new(cfqq);
3937 * Called when a new fs request (rq) is added (to cfqq). Check if there's
3938 * something we should do about it
3941 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3944 struct cfq_io_cq *cic = RQ_CIC(rq);
3947 if (rq->cmd_flags & REQ_PRIO)
3948 cfqq->prio_pending++;
3950 cfq_update_io_thinktime(cfqd, cfqq, cic);
3951 cfq_update_io_seektime(cfqd, cfqq, rq);
3952 cfq_update_idle_window(cfqd, cfqq, cic);
3954 cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
3956 if (cfqq == cfqd->active_queue) {
3958 * Remember that we saw a request from this process, but
3959 * don't start queuing just yet. Otherwise we risk seeing lots
3960 * of tiny requests, because we disrupt the normal plugging
3961 * and merging. If the request is already larger than a single
3962 * page, let it rip immediately. For that case we assume that
3963 * merging is already done. Ditto for a busy system that
3964 * has other work pending, don't risk delaying until the
3965 * idle timer unplug to continue working.
3967 if (cfq_cfqq_wait_request(cfqq)) {
3968 if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
3969 cfqd->busy_queues > 1) {
3970 cfq_del_timer(cfqd, cfqq);
3971 cfq_clear_cfqq_wait_request(cfqq);
3972 __blk_run_queue(cfqd->queue);
3974 cfqg_stats_update_idle_time(cfqq->cfqg);
3975 cfq_mark_cfqq_must_dispatch(cfqq);
3978 } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
3980 * not the active queue - expire current slice if it is
3981 * idle and has expired it's mean thinktime or this new queue
3982 * has some old slice time left and is of higher priority or
3983 * this new queue is RT and the current one is BE
3985 cfq_preempt_queue(cfqd, cfqq);
3986 __blk_run_queue(cfqd->queue);
3990 static void cfq_insert_request(struct request_queue *q, struct request *rq)
3992 struct cfq_data *cfqd = q->elevator->elevator_data;
3993 struct cfq_queue *cfqq = RQ_CFQQ(rq);
3995 cfq_log_cfqq(cfqd, cfqq, "insert_request");
3996 cfq_init_prio_data(cfqq, RQ_CIC(rq));
3998 rq->fifo_time = jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)];
3999 list_add_tail(&rq->queuelist, &cfqq->fifo);
4001 cfqg_stats_update_io_add(RQ_CFQG(rq), cfqd->serving_group,
4003 cfq_rq_enqueued(cfqd, cfqq, rq);
4007 * Update hw_tag based on peak queue depth over 50 samples under
4010 static void cfq_update_hw_tag(struct cfq_data *cfqd)
4012 struct cfq_queue *cfqq = cfqd->active_queue;
4014 if (cfqd->rq_in_driver > cfqd->hw_tag_est_depth)
4015 cfqd->hw_tag_est_depth = cfqd->rq_in_driver;
4017 if (cfqd->hw_tag == 1)
4020 if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
4021 cfqd->rq_in_driver <= CFQ_HW_QUEUE_MIN)
4025 * If active queue hasn't enough requests and can idle, cfq might not
4026 * dispatch sufficient requests to hardware. Don't zero hw_tag in this
4029 if (cfqq && cfq_cfqq_idle_window(cfqq) &&
4030 cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] <
4031 CFQ_HW_QUEUE_MIN && cfqd->rq_in_driver < CFQ_HW_QUEUE_MIN)
4034 if (cfqd->hw_tag_samples++ < 50)
4037 if (cfqd->hw_tag_est_depth >= CFQ_HW_QUEUE_MIN)
4043 static bool cfq_should_wait_busy(struct cfq_data *cfqd, struct cfq_queue *cfqq)
4045 struct cfq_io_cq *cic = cfqd->active_cic;
4047 /* If the queue already has requests, don't wait */
4048 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
4051 /* If there are other queues in the group, don't wait */
4052 if (cfqq->cfqg->nr_cfqq > 1)
4055 /* the only queue in the group, but think time is big */
4056 if (cfq_io_thinktime_big(cfqd, &cfqq->cfqg->ttime, true))
4059 if (cfq_slice_used(cfqq))
4062 /* if slice left is less than think time, wait busy */
4063 if (cic && sample_valid(cic->ttime.ttime_samples)
4064 && (cfqq->slice_end - jiffies < cic->ttime.ttime_mean))
4068 * If think times is less than a jiffy than ttime_mean=0 and above
4069 * will not be true. It might happen that slice has not expired yet
4070 * but will expire soon (4-5 ns) during select_queue(). To cover the
4071 * case where think time is less than a jiffy, mark the queue wait
4072 * busy if only 1 jiffy is left in the slice.
4074 if (cfqq->slice_end - jiffies == 1)
4080 static void cfq_completed_request(struct request_queue *q, struct request *rq)
4082 struct cfq_queue *cfqq = RQ_CFQQ(rq);
4083 struct cfq_data *cfqd = cfqq->cfqd;
4084 const int sync = rq_is_sync(rq);
4088 cfq_log_cfqq(cfqd, cfqq, "complete rqnoidle %d",
4089 !!(rq->cmd_flags & REQ_NOIDLE));
4091 cfq_update_hw_tag(cfqd);
4093 WARN_ON(!cfqd->rq_in_driver);
4094 WARN_ON(!cfqq->dispatched);
4095 cfqd->rq_in_driver--;
4097 (RQ_CFQG(rq))->dispatched--;
4098 cfqg_stats_update_completion(cfqq->cfqg, rq_start_time_ns(rq),
4099 rq_io_start_time_ns(rq), rq->cmd_flags);
4101 cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]--;
4104 struct cfq_rb_root *st;
4106 RQ_CIC(rq)->ttime.last_end_request = now;
4108 if (cfq_cfqq_on_rr(cfqq))
4109 st = cfqq->service_tree;
4111 st = st_for(cfqq->cfqg, cfqq_class(cfqq),
4114 st->ttime.last_end_request = now;
4115 if (!time_after(rq->start_time + cfqd->cfq_fifo_expire[1], now))
4116 cfqd->last_delayed_sync = now;
4119 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4120 cfqq->cfqg->ttime.last_end_request = now;
4124 * If this is the active queue, check if it needs to be expired,
4125 * or if we want to idle in case it has no pending requests.
4127 if (cfqd->active_queue == cfqq) {
4128 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
4130 if (cfq_cfqq_slice_new(cfqq)) {
4131 cfq_set_prio_slice(cfqd, cfqq);
4132 cfq_clear_cfqq_slice_new(cfqq);
4136 * Should we wait for next request to come in before we expire
4139 if (cfq_should_wait_busy(cfqd, cfqq)) {
4140 unsigned long extend_sl = cfqd->cfq_slice_idle;
4141 if (!cfqd->cfq_slice_idle)
4142 extend_sl = cfqd->cfq_group_idle;
4143 cfqq->slice_end = jiffies + extend_sl;
4144 cfq_mark_cfqq_wait_busy(cfqq);
4145 cfq_log_cfqq(cfqd, cfqq, "will busy wait");
4149 * Idling is not enabled on:
4151 * - idle-priority queues
4153 * - queues with still some requests queued
4154 * - when there is a close cooperator
4156 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
4157 cfq_slice_expired(cfqd, 1);
4158 else if (sync && cfqq_empty &&
4159 !cfq_close_cooperator(cfqd, cfqq)) {
4160 cfq_arm_slice_timer(cfqd);
4164 if (!cfqd->rq_in_driver)
4165 cfq_schedule_dispatch(cfqd);
4168 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
4170 if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
4171 cfq_mark_cfqq_must_alloc_slice(cfqq);
4172 return ELV_MQUEUE_MUST;
4175 return ELV_MQUEUE_MAY;
4178 static int cfq_may_queue(struct request_queue *q, int rw)
4180 struct cfq_data *cfqd = q->elevator->elevator_data;
4181 struct task_struct *tsk = current;
4182 struct cfq_io_cq *cic;
4183 struct cfq_queue *cfqq;
4186 * don't force setup of a queue from here, as a call to may_queue
4187 * does not necessarily imply that a request actually will be queued.
4188 * so just lookup a possibly existing queue, or return 'may queue'
4191 cic = cfq_cic_lookup(cfqd, tsk->io_context);
4193 return ELV_MQUEUE_MAY;
4195 cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
4197 cfq_init_prio_data(cfqq, cic);
4199 return __cfq_may_queue(cfqq);
4202 return ELV_MQUEUE_MAY;
4206 * queue lock held here
4208 static void cfq_put_request(struct request *rq)
4210 struct cfq_queue *cfqq = RQ_CFQQ(rq);
4213 const int rw = rq_data_dir(rq);
4215 BUG_ON(!cfqq->allocated[rw]);
4216 cfqq->allocated[rw]--;
4218 /* Put down rq reference on cfqg */
4219 cfqg_put(RQ_CFQG(rq));
4220 rq->elv.priv[0] = NULL;
4221 rq->elv.priv[1] = NULL;
4223 cfq_put_queue(cfqq);
4227 static struct cfq_queue *
4228 cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_cq *cic,
4229 struct cfq_queue *cfqq)
4231 cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq);
4232 cic_set_cfqq(cic, cfqq->new_cfqq, 1);
4233 cfq_mark_cfqq_coop(cfqq->new_cfqq);
4234 cfq_put_queue(cfqq);
4235 return cic_to_cfqq(cic, 1);
4239 * Returns NULL if a new cfqq should be allocated, or the old cfqq if this
4240 * was the last process referring to said cfqq.
4242 static struct cfq_queue *
4243 split_cfqq(struct cfq_io_cq *cic, struct cfq_queue *cfqq)
4245 if (cfqq_process_refs(cfqq) == 1) {
4246 cfqq->pid = current->pid;
4247 cfq_clear_cfqq_coop(cfqq);
4248 cfq_clear_cfqq_split_coop(cfqq);
4252 cic_set_cfqq(cic, NULL, 1);
4254 cfq_put_cooperator(cfqq);
4256 cfq_put_queue(cfqq);
4260 * Allocate cfq data structures associated with this request.
4263 cfq_set_request(struct request_queue *q, struct request *rq, struct bio *bio,
4266 struct cfq_data *cfqd = q->elevator->elevator_data;
4267 struct cfq_io_cq *cic = icq_to_cic(rq->elv.icq);
4268 const int rw = rq_data_dir(rq);
4269 const bool is_sync = rq_is_sync(rq);
4270 struct cfq_queue *cfqq;
4272 might_sleep_if(gfp_mask & __GFP_WAIT);
4274 spin_lock_irq(q->queue_lock);
4276 check_ioprio_changed(cic, bio);
4277 check_blkcg_changed(cic, bio);
4279 cfqq = cic_to_cfqq(cic, is_sync);
4280 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
4281 cfqq = cfq_get_queue(cfqd, is_sync, cic, bio, gfp_mask);
4282 cic_set_cfqq(cic, cfqq, is_sync);
4285 * If the queue was seeky for too long, break it apart.
4287 if (cfq_cfqq_coop(cfqq) && cfq_cfqq_split_coop(cfqq)) {
4288 cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq");
4289 cfqq = split_cfqq(cic, cfqq);
4295 * Check to see if this queue is scheduled to merge with
4296 * another, closely cooperating queue. The merging of
4297 * queues happens here as it must be done in process context.
4298 * The reference on new_cfqq was taken in merge_cfqqs.
4301 cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq);
4304 cfqq->allocated[rw]++;
4307 cfqg_get(cfqq->cfqg);
4308 rq->elv.priv[0] = cfqq;
4309 rq->elv.priv[1] = cfqq->cfqg;
4310 spin_unlock_irq(q->queue_lock);
4314 static void cfq_kick_queue(struct work_struct *work)
4316 struct cfq_data *cfqd =
4317 container_of(work, struct cfq_data, unplug_work);
4318 struct request_queue *q = cfqd->queue;
4320 spin_lock_irq(q->queue_lock);
4321 __blk_run_queue(cfqd->queue);
4322 spin_unlock_irq(q->queue_lock);
4326 * Timer running if the active_queue is currently idling inside its time slice
4328 static void cfq_idle_slice_timer(unsigned long data)
4330 struct cfq_data *cfqd = (struct cfq_data *) data;
4331 struct cfq_queue *cfqq;
4332 unsigned long flags;
4335 cfq_log(cfqd, "idle timer fired");
4337 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
4339 cfqq = cfqd->active_queue;
4344 * We saw a request before the queue expired, let it through
4346 if (cfq_cfqq_must_dispatch(cfqq))
4352 if (cfq_slice_used(cfqq))
4356 * only expire and reinvoke request handler, if there are
4357 * other queues with pending requests
4359 if (!cfqd->busy_queues)
4363 * not expired and it has a request pending, let it dispatch
4365 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
4369 * Queue depth flag is reset only when the idle didn't succeed
4371 cfq_clear_cfqq_deep(cfqq);
4374 cfq_slice_expired(cfqd, timed_out);
4376 cfq_schedule_dispatch(cfqd);
4378 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
4381 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
4383 del_timer_sync(&cfqd->idle_slice_timer);
4384 cancel_work_sync(&cfqd->unplug_work);
4387 static void cfq_put_async_queues(struct cfq_data *cfqd)
4391 for (i = 0; i < IOPRIO_BE_NR; i++) {
4392 if (cfqd->async_cfqq[0][i])
4393 cfq_put_queue(cfqd->async_cfqq[0][i]);
4394 if (cfqd->async_cfqq[1][i])
4395 cfq_put_queue(cfqd->async_cfqq[1][i]);
4398 if (cfqd->async_idle_cfqq)
4399 cfq_put_queue(cfqd->async_idle_cfqq);
4402 static void cfq_exit_queue(struct elevator_queue *e)
4404 struct cfq_data *cfqd = e->elevator_data;
4405 struct request_queue *q = cfqd->queue;
4407 cfq_shutdown_timer_wq(cfqd);
4409 spin_lock_irq(q->queue_lock);
4411 if (cfqd->active_queue)
4412 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
4414 cfq_put_async_queues(cfqd);
4416 spin_unlock_irq(q->queue_lock);
4418 cfq_shutdown_timer_wq(cfqd);
4420 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4421 blkcg_deactivate_policy(q, &blkcg_policy_cfq);
4423 kfree(cfqd->root_group);
4428 static int cfq_init_queue(struct request_queue *q, struct elevator_type *e)
4430 struct cfq_data *cfqd;
4431 struct blkcg_gq *blkg __maybe_unused;
4433 struct elevator_queue *eq;
4435 eq = elevator_alloc(q, e);
4439 cfqd = kzalloc_node(sizeof(*cfqd), GFP_KERNEL, q->node);
4441 kobject_put(&eq->kobj);
4444 eq->elevator_data = cfqd;
4447 spin_lock_irq(q->queue_lock);
4449 spin_unlock_irq(q->queue_lock);
4451 /* Init root service tree */
4452 cfqd->grp_service_tree = CFQ_RB_ROOT;
4454 /* Init root group and prefer root group over other groups by default */
4455 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4456 ret = blkcg_activate_policy(q, &blkcg_policy_cfq);
4460 cfqd->root_group = blkg_to_cfqg(q->root_blkg);
4463 cfqd->root_group = kzalloc_node(sizeof(*cfqd->root_group),
4464 GFP_KERNEL, cfqd->queue->node);
4465 if (!cfqd->root_group)
4468 cfq_init_cfqg_base(cfqd->root_group);
4470 cfqd->root_group->weight = 2 * CFQ_WEIGHT_DEFAULT;
4471 cfqd->root_group->leaf_weight = 2 * CFQ_WEIGHT_DEFAULT;
4474 * Not strictly needed (since RB_ROOT just clears the node and we
4475 * zeroed cfqd on alloc), but better be safe in case someone decides
4476 * to add magic to the rb code
4478 for (i = 0; i < CFQ_PRIO_LISTS; i++)
4479 cfqd->prio_trees[i] = RB_ROOT;
4482 * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
4483 * Grab a permanent reference to it, so that the normal code flow
4484 * will not attempt to free it. oom_cfqq is linked to root_group
4485 * but shouldn't hold a reference as it'll never be unlinked. Lose
4486 * the reference from linking right away.
4488 cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
4489 cfqd->oom_cfqq.ref++;
4491 spin_lock_irq(q->queue_lock);
4492 cfq_link_cfqq_cfqg(&cfqd->oom_cfqq, cfqd->root_group);
4493 cfqg_put(cfqd->root_group);
4494 spin_unlock_irq(q->queue_lock);
4496 init_timer(&cfqd->idle_slice_timer);
4497 cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
4498 cfqd->idle_slice_timer.data = (unsigned long) cfqd;
4500 INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
4502 cfqd->cfq_quantum = cfq_quantum;
4503 cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
4504 cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
4505 cfqd->cfq_back_max = cfq_back_max;
4506 cfqd->cfq_back_penalty = cfq_back_penalty;
4507 cfqd->cfq_slice[0] = cfq_slice_async;
4508 cfqd->cfq_slice[1] = cfq_slice_sync;
4509 cfqd->cfq_target_latency = cfq_target_latency;
4510 cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
4511 cfqd->cfq_slice_idle = blk_queue_nonrot(q) ? 0 : cfq_slice_idle;
4512 cfqd->cfq_group_idle = cfq_group_idle;
4513 cfqd->cfq_latency = 1;
4516 * we optimistically start assuming sync ops weren't delayed in last
4517 * second, in order to have larger depth for async operations.
4519 cfqd->last_delayed_sync = jiffies - HZ;
4524 kobject_put(&eq->kobj);
4529 * sysfs parts below -->
4532 cfq_var_show(unsigned int var, char *page)
4534 return sprintf(page, "%u\n", var);
4538 cfq_var_store(unsigned int *var, const char *page, size_t count)
4540 char *p = (char *) page;
4542 *var = simple_strtoul(p, &p, 10);
4546 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
4547 static ssize_t __FUNC(struct elevator_queue *e, char *page) \
4549 struct cfq_data *cfqd = e->elevator_data; \
4550 unsigned int __data = __VAR; \
4552 __data = jiffies_to_msecs(__data); \
4553 return cfq_var_show(__data, (page)); \
4555 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
4556 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
4557 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
4558 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
4559 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
4560 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
4561 SHOW_FUNCTION(cfq_group_idle_show, cfqd->cfq_group_idle, 1);
4562 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
4563 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
4564 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
4565 SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
4566 SHOW_FUNCTION(cfq_target_latency_show, cfqd->cfq_target_latency, 1);
4567 #undef SHOW_FUNCTION
4569 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
4570 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
4572 struct cfq_data *cfqd = e->elevator_data; \
4573 unsigned int __data; \
4574 int ret = cfq_var_store(&__data, (page), count); \
4575 if (__data < (MIN)) \
4577 else if (__data > (MAX)) \
4580 *(__PTR) = msecs_to_jiffies(__data); \
4582 *(__PTR) = __data; \
4585 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
4586 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
4588 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
4590 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
4591 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
4593 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
4594 STORE_FUNCTION(cfq_group_idle_store, &cfqd->cfq_group_idle, 0, UINT_MAX, 1);
4595 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
4596 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
4597 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
4599 STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
4600 STORE_FUNCTION(cfq_target_latency_store, &cfqd->cfq_target_latency, 1, UINT_MAX, 1);
4601 #undef STORE_FUNCTION
4603 #define CFQ_ATTR(name) \
4604 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
4606 static struct elv_fs_entry cfq_attrs[] = {
4608 CFQ_ATTR(fifo_expire_sync),
4609 CFQ_ATTR(fifo_expire_async),
4610 CFQ_ATTR(back_seek_max),
4611 CFQ_ATTR(back_seek_penalty),
4612 CFQ_ATTR(slice_sync),
4613 CFQ_ATTR(slice_async),
4614 CFQ_ATTR(slice_async_rq),
4615 CFQ_ATTR(slice_idle),
4616 CFQ_ATTR(group_idle),
4617 CFQ_ATTR(low_latency),
4618 CFQ_ATTR(target_latency),
4622 static struct elevator_type iosched_cfq = {
4624 .elevator_merge_fn = cfq_merge,
4625 .elevator_merged_fn = cfq_merged_request,
4626 .elevator_merge_req_fn = cfq_merged_requests,
4627 .elevator_allow_merge_fn = cfq_allow_merge,
4628 .elevator_bio_merged_fn = cfq_bio_merged,
4629 .elevator_dispatch_fn = cfq_dispatch_requests,
4630 .elevator_add_req_fn = cfq_insert_request,
4631 .elevator_activate_req_fn = cfq_activate_request,
4632 .elevator_deactivate_req_fn = cfq_deactivate_request,
4633 .elevator_completed_req_fn = cfq_completed_request,
4634 .elevator_former_req_fn = elv_rb_former_request,
4635 .elevator_latter_req_fn = elv_rb_latter_request,
4636 .elevator_init_icq_fn = cfq_init_icq,
4637 .elevator_exit_icq_fn = cfq_exit_icq,
4638 .elevator_set_req_fn = cfq_set_request,
4639 .elevator_put_req_fn = cfq_put_request,
4640 .elevator_may_queue_fn = cfq_may_queue,
4641 .elevator_init_fn = cfq_init_queue,
4642 .elevator_exit_fn = cfq_exit_queue,
4644 .icq_size = sizeof(struct cfq_io_cq),
4645 .icq_align = __alignof__(struct cfq_io_cq),
4646 .elevator_attrs = cfq_attrs,
4647 .elevator_name = "cfq",
4648 .elevator_owner = THIS_MODULE,
4651 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4652 static struct blkcg_policy blkcg_policy_cfq = {
4653 .pd_size = sizeof(struct cfq_group),
4654 .cpd_size = sizeof(struct cfq_group_data),
4655 .cftypes = cfq_blkcg_files,
4657 .cpd_init_fn = cfq_cpd_init,
4658 .pd_init_fn = cfq_pd_init,
4659 .pd_offline_fn = cfq_pd_offline,
4660 .pd_reset_stats_fn = cfq_pd_reset_stats,
4664 static int __init cfq_init(void)
4669 * could be 0 on HZ < 1000 setups
4671 if (!cfq_slice_async)
4672 cfq_slice_async = 1;
4673 if (!cfq_slice_idle)
4676 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4677 if (!cfq_group_idle)
4680 ret = blkcg_policy_register(&blkcg_policy_cfq);
4688 cfq_pool = KMEM_CACHE(cfq_queue, 0);
4692 ret = elv_register(&iosched_cfq);
4699 kmem_cache_destroy(cfq_pool);
4701 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4702 blkcg_policy_unregister(&blkcg_policy_cfq);
4707 static void __exit cfq_exit(void)
4709 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4710 blkcg_policy_unregister(&blkcg_policy_cfq);
4712 elv_unregister(&iosched_cfq);
4713 kmem_cache_destroy(cfq_pool);
4716 module_init(cfq_init);
4717 module_exit(cfq_exit);
4719 MODULE_AUTHOR("Jens Axboe");
4720 MODULE_LICENSE("GPL");
4721 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");