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/ktime.h>
14 #include <linux/rbtree.h>
15 #include <linux/ioprio.h>
16 #include <linux/blktrace_api.h>
17 #include <linux/blk-cgroup.h>
24 /* max queue in one round of service */
25 static const int cfq_quantum = 8;
26 static const u64 cfq_fifo_expire[2] = { NSEC_PER_SEC / 4, NSEC_PER_SEC / 8 };
27 /* maximum backwards seek, in KiB */
28 static const int cfq_back_max = 16 * 1024;
29 /* penalty of a backwards seek */
30 static const int cfq_back_penalty = 2;
31 static const u64 cfq_slice_sync = NSEC_PER_SEC / 10;
32 static u64 cfq_slice_async = NSEC_PER_SEC / 25;
33 static const int cfq_slice_async_rq = 2;
34 static u64 cfq_slice_idle = NSEC_PER_SEC / 125;
35 static u64 cfq_group_idle = NSEC_PER_SEC / 125;
36 static const u64 cfq_target_latency = (u64)NSEC_PER_SEC * 3/10; /* 300 ms */
37 static const int cfq_hist_divisor = 4;
40 * offset from end of service tree
42 #define CFQ_IDLE_DELAY (NSEC_PER_SEC / 5)
45 * below this threshold, we consider thinktime immediate
47 #define CFQ_MIN_TT (2 * NSEC_PER_SEC / HZ)
49 #define CFQ_SLICE_SCALE (5)
50 #define CFQ_HW_QUEUE_MIN (5)
51 #define CFQ_SERVICE_SHIFT 12
53 #define CFQQ_SEEK_THR (sector_t)(8 * 100)
54 #define CFQQ_CLOSE_THR (sector_t)(8 * 1024)
55 #define CFQQ_SECT_THR_NONROT (sector_t)(2 * 32)
56 #define CFQQ_SEEKY(cfqq) (hweight32(cfqq->seek_history) > 32/8)
58 #define RQ_CIC(rq) icq_to_cic((rq)->elv.icq)
59 #define RQ_CFQQ(rq) (struct cfq_queue *) ((rq)->elv.priv[0])
60 #define RQ_CFQG(rq) (struct cfq_group *) ((rq)->elv.priv[1])
62 static struct kmem_cache *cfq_pool;
64 #define CFQ_PRIO_LISTS IOPRIO_BE_NR
65 #define cfq_class_idle(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
66 #define cfq_class_rt(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
68 #define sample_valid(samples) ((samples) > 80)
69 #define rb_entry_cfqg(node) rb_entry((node), struct cfq_group, rb_node)
71 /* blkio-related constants */
72 #define CFQ_WEIGHT_LEGACY_MIN 10
73 #define CFQ_WEIGHT_LEGACY_DFL 500
74 #define CFQ_WEIGHT_LEGACY_MAX 1000
81 unsigned long ttime_samples;
85 * Most of our rbtree usage is for sorting with min extraction, so
86 * if we cache the leftmost node we don't have to walk down the tree
87 * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should
88 * move this into the elevator for the rq sorting as well.
95 struct cfq_ttime ttime;
97 #define CFQ_RB_ROOT (struct cfq_rb_root) { .rb = RB_ROOT, \
98 .ttime = {.last_end_request = ktime_get_ns(),},}
101 * Per process-grouping structure
104 /* reference count */
106 /* various state flags, see below */
108 /* parent cfq_data */
109 struct cfq_data *cfqd;
110 /* service_tree member */
111 struct rb_node rb_node;
112 /* service_tree key */
114 /* prio tree member */
115 struct rb_node p_node;
116 /* prio tree root we belong to, if any */
117 struct rb_root *p_root;
118 /* sorted list of pending requests */
119 struct rb_root sort_list;
120 /* if fifo isn't expired, next request to serve */
121 struct request *next_rq;
122 /* requests queued in sort_list */
124 /* currently allocated requests */
126 /* fifo list of requests in sort_list */
127 struct list_head fifo;
129 /* time when queue got scheduled in to dispatch first request. */
133 /* time when first request from queue completed and slice started. */
138 /* pending priority requests */
140 /* number of requests that are on the dispatch list or inside driver */
143 /* io prio of this group */
144 unsigned short ioprio, org_ioprio;
145 unsigned short ioprio_class, org_ioprio_class;
150 sector_t last_request_pos;
152 struct cfq_rb_root *service_tree;
153 struct cfq_queue *new_cfqq;
154 struct cfq_group *cfqg;
155 /* Number of sectors dispatched from queue in single dispatch round */
156 unsigned long nr_sectors;
160 * First index in the service_trees.
161 * IDLE is handled separately, so it has negative index
171 * Second index in the service_trees.
175 SYNC_NOIDLE_WORKLOAD = 1,
180 #ifdef CONFIG_CFQ_GROUP_IOSCHED
181 /* number of ios merged */
182 struct blkg_rwstat merged;
183 /* total time spent on device in ns, may not be accurate w/ queueing */
184 struct blkg_rwstat service_time;
185 /* total time spent waiting in scheduler queue in ns */
186 struct blkg_rwstat wait_time;
187 /* number of IOs queued up */
188 struct blkg_rwstat queued;
189 /* total disk time and nr sectors dispatched by this group */
190 struct blkg_stat time;
191 #ifdef CONFIG_DEBUG_BLK_CGROUP
192 /* time not charged to this cgroup */
193 struct blkg_stat unaccounted_time;
194 /* sum of number of ios queued across all samples */
195 struct blkg_stat avg_queue_size_sum;
196 /* count of samples taken for average */
197 struct blkg_stat avg_queue_size_samples;
198 /* how many times this group has been removed from service tree */
199 struct blkg_stat dequeue;
200 /* total time spent waiting for it to be assigned a timeslice. */
201 struct blkg_stat group_wait_time;
202 /* time spent idling for this blkcg_gq */
203 struct blkg_stat idle_time;
204 /* total time with empty current active q with other requests queued */
205 struct blkg_stat empty_time;
206 /* fields after this shouldn't be cleared on stat reset */
207 uint64_t start_group_wait_time;
208 uint64_t start_idle_time;
209 uint64_t start_empty_time;
211 #endif /* CONFIG_DEBUG_BLK_CGROUP */
212 #endif /* CONFIG_CFQ_GROUP_IOSCHED */
215 /* Per-cgroup data */
216 struct cfq_group_data {
217 /* must be the first member */
218 struct blkcg_policy_data cpd;
221 unsigned int leaf_weight;
224 /* This is per cgroup per device grouping structure */
226 /* must be the first member */
227 struct blkg_policy_data pd;
229 /* group service_tree member */
230 struct rb_node rb_node;
232 /* group service_tree key */
236 * The number of active cfqgs and sum of their weights under this
237 * cfqg. This covers this cfqg's leaf_weight and all children's
238 * weights, but does not cover weights of further descendants.
240 * If a cfqg is on the service tree, it's active. An active cfqg
241 * also activates its parent and contributes to the children_weight
245 unsigned int children_weight;
248 * vfraction is the fraction of vdisktime that the tasks in this
249 * cfqg are entitled to. This is determined by compounding the
250 * ratios walking up from this cfqg to the root.
252 * It is in fixed point w/ CFQ_SERVICE_SHIFT and the sum of all
253 * vfractions on a service tree is approximately 1. The sum may
254 * deviate a bit due to rounding errors and fluctuations caused by
255 * cfqgs entering and leaving the service tree.
257 unsigned int vfraction;
260 * There are two weights - (internal) weight is the weight of this
261 * cfqg against the sibling cfqgs. leaf_weight is the wight of
262 * this cfqg against the child cfqgs. For the root cfqg, both
263 * weights are kept in sync for backward compatibility.
266 unsigned int new_weight;
267 unsigned int dev_weight;
269 unsigned int leaf_weight;
270 unsigned int new_leaf_weight;
271 unsigned int dev_leaf_weight;
273 /* number of cfqq currently on this group */
277 * Per group busy queues average. Useful for workload slice calc. We
278 * create the array for each prio class but at run time it is used
279 * only for RT and BE class and slot for IDLE class remains unused.
280 * This is primarily done to avoid confusion and a gcc warning.
282 unsigned int busy_queues_avg[CFQ_PRIO_NR];
284 * rr lists of queues with requests. We maintain service trees for
285 * RT and BE classes. These trees are subdivided in subclasses
286 * of SYNC, SYNC_NOIDLE and ASYNC based on workload type. For IDLE
287 * class there is no subclassification and all the cfq queues go on
288 * a single tree service_tree_idle.
289 * Counts are embedded in the cfq_rb_root
291 struct cfq_rb_root service_trees[2][3];
292 struct cfq_rb_root service_tree_idle;
295 enum wl_type_t saved_wl_type;
296 enum wl_class_t saved_wl_class;
298 /* number of requests that are on the dispatch list or inside driver */
300 struct cfq_ttime ttime;
301 struct cfqg_stats stats; /* stats for this cfqg */
303 /* async queue for each priority case */
304 struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
305 struct cfq_queue *async_idle_cfqq;
310 struct io_cq icq; /* must be the first member */
311 struct cfq_queue *cfqq[2];
312 struct cfq_ttime ttime;
313 int ioprio; /* the current ioprio */
314 #ifdef CONFIG_CFQ_GROUP_IOSCHED
315 uint64_t blkcg_serial_nr; /* the current blkcg serial */
320 * Per block device queue structure
323 struct request_queue *queue;
324 /* Root service tree for cfq_groups */
325 struct cfq_rb_root grp_service_tree;
326 struct cfq_group *root_group;
329 * The priority currently being served
331 enum wl_class_t serving_wl_class;
332 enum wl_type_t serving_wl_type;
333 u64 workload_expires;
334 struct cfq_group *serving_group;
337 * Each priority tree is sorted by next_request position. These
338 * trees are used when determining if two or more queues are
339 * interleaving requests (see cfq_close_cooperator).
341 struct rb_root prio_trees[CFQ_PRIO_LISTS];
343 unsigned int busy_queues;
344 unsigned int busy_sync_queues;
350 * queue-depth detection
356 * -1 => indeterminate, (cfq will behave as if NCQ is present, to allow better detection)
357 * 1 => NCQ is present (hw_tag_est_depth is the estimated max depth)
360 int hw_tag_est_depth;
361 unsigned int hw_tag_samples;
364 * idle window management
366 struct hrtimer idle_slice_timer;
367 struct work_struct unplug_work;
369 struct cfq_queue *active_queue;
370 struct cfq_io_cq *active_cic;
372 sector_t last_position;
375 * tunables, see top of file
377 unsigned int cfq_quantum;
378 unsigned int cfq_back_penalty;
379 unsigned int cfq_back_max;
380 unsigned int cfq_slice_async_rq;
381 unsigned int cfq_latency;
382 u64 cfq_fifo_expire[2];
386 u64 cfq_target_latency;
389 * Fallback dummy cfqq for extreme OOM conditions
391 struct cfq_queue oom_cfqq;
393 u64 last_delayed_sync;
396 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd);
397 static void cfq_put_queue(struct cfq_queue *cfqq);
399 static struct cfq_rb_root *st_for(struct cfq_group *cfqg,
400 enum wl_class_t class,
406 if (class == IDLE_WORKLOAD)
407 return &cfqg->service_tree_idle;
409 return &cfqg->service_trees[class][type];
412 enum cfqq_state_flags {
413 CFQ_CFQQ_FLAG_on_rr = 0, /* on round-robin busy list */
414 CFQ_CFQQ_FLAG_wait_request, /* waiting for a request */
415 CFQ_CFQQ_FLAG_must_dispatch, /* must be allowed a dispatch */
416 CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
417 CFQ_CFQQ_FLAG_fifo_expire, /* FIFO checked in this slice */
418 CFQ_CFQQ_FLAG_idle_window, /* slice idling enabled */
419 CFQ_CFQQ_FLAG_prio_changed, /* task priority has changed */
420 CFQ_CFQQ_FLAG_slice_new, /* no requests dispatched in slice */
421 CFQ_CFQQ_FLAG_sync, /* synchronous queue */
422 CFQ_CFQQ_FLAG_coop, /* cfqq is shared */
423 CFQ_CFQQ_FLAG_split_coop, /* shared cfqq will be splitted */
424 CFQ_CFQQ_FLAG_deep, /* sync cfqq experienced large depth */
425 CFQ_CFQQ_FLAG_wait_busy, /* Waiting for next request */
428 #define CFQ_CFQQ_FNS(name) \
429 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \
431 (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name); \
433 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \
435 (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \
437 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \
439 return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \
443 CFQ_CFQQ_FNS(wait_request);
444 CFQ_CFQQ_FNS(must_dispatch);
445 CFQ_CFQQ_FNS(must_alloc_slice);
446 CFQ_CFQQ_FNS(fifo_expire);
447 CFQ_CFQQ_FNS(idle_window);
448 CFQ_CFQQ_FNS(prio_changed);
449 CFQ_CFQQ_FNS(slice_new);
452 CFQ_CFQQ_FNS(split_coop);
454 CFQ_CFQQ_FNS(wait_busy);
457 #if defined(CONFIG_CFQ_GROUP_IOSCHED) && defined(CONFIG_DEBUG_BLK_CGROUP)
459 /* cfqg stats flags */
460 enum cfqg_stats_flags {
461 CFQG_stats_waiting = 0,
466 #define CFQG_FLAG_FNS(name) \
467 static inline void cfqg_stats_mark_##name(struct cfqg_stats *stats) \
469 stats->flags |= (1 << CFQG_stats_##name); \
471 static inline void cfqg_stats_clear_##name(struct cfqg_stats *stats) \
473 stats->flags &= ~(1 << CFQG_stats_##name); \
475 static inline int cfqg_stats_##name(struct cfqg_stats *stats) \
477 return (stats->flags & (1 << CFQG_stats_##name)) != 0; \
480 CFQG_FLAG_FNS(waiting)
481 CFQG_FLAG_FNS(idling)
485 /* This should be called with the queue_lock held. */
486 static void cfqg_stats_update_group_wait_time(struct cfqg_stats *stats)
488 unsigned long long now;
490 if (!cfqg_stats_waiting(stats))
494 if (time_after64(now, stats->start_group_wait_time))
495 blkg_stat_add(&stats->group_wait_time,
496 now - stats->start_group_wait_time);
497 cfqg_stats_clear_waiting(stats);
500 /* This should be called with the queue_lock held. */
501 static void cfqg_stats_set_start_group_wait_time(struct cfq_group *cfqg,
502 struct cfq_group *curr_cfqg)
504 struct cfqg_stats *stats = &cfqg->stats;
506 if (cfqg_stats_waiting(stats))
508 if (cfqg == curr_cfqg)
510 stats->start_group_wait_time = sched_clock();
511 cfqg_stats_mark_waiting(stats);
514 /* This should be called with the queue_lock held. */
515 static void cfqg_stats_end_empty_time(struct cfqg_stats *stats)
517 unsigned long long now;
519 if (!cfqg_stats_empty(stats))
523 if (time_after64(now, stats->start_empty_time))
524 blkg_stat_add(&stats->empty_time,
525 now - stats->start_empty_time);
526 cfqg_stats_clear_empty(stats);
529 static void cfqg_stats_update_dequeue(struct cfq_group *cfqg)
531 blkg_stat_add(&cfqg->stats.dequeue, 1);
534 static void cfqg_stats_set_start_empty_time(struct cfq_group *cfqg)
536 struct cfqg_stats *stats = &cfqg->stats;
538 if (blkg_rwstat_total(&stats->queued))
542 * group is already marked empty. This can happen if cfqq got new
543 * request in parent group and moved to this group while being added
544 * to service tree. Just ignore the event and move on.
546 if (cfqg_stats_empty(stats))
549 stats->start_empty_time = sched_clock();
550 cfqg_stats_mark_empty(stats);
553 static void cfqg_stats_update_idle_time(struct cfq_group *cfqg)
555 struct cfqg_stats *stats = &cfqg->stats;
557 if (cfqg_stats_idling(stats)) {
558 unsigned long long now = sched_clock();
560 if (time_after64(now, stats->start_idle_time))
561 blkg_stat_add(&stats->idle_time,
562 now - stats->start_idle_time);
563 cfqg_stats_clear_idling(stats);
567 static void cfqg_stats_set_start_idle_time(struct cfq_group *cfqg)
569 struct cfqg_stats *stats = &cfqg->stats;
571 BUG_ON(cfqg_stats_idling(stats));
573 stats->start_idle_time = sched_clock();
574 cfqg_stats_mark_idling(stats);
577 static void cfqg_stats_update_avg_queue_size(struct cfq_group *cfqg)
579 struct cfqg_stats *stats = &cfqg->stats;
581 blkg_stat_add(&stats->avg_queue_size_sum,
582 blkg_rwstat_total(&stats->queued));
583 blkg_stat_add(&stats->avg_queue_size_samples, 1);
584 cfqg_stats_update_group_wait_time(stats);
587 #else /* CONFIG_CFQ_GROUP_IOSCHED && CONFIG_DEBUG_BLK_CGROUP */
589 static inline void cfqg_stats_set_start_group_wait_time(struct cfq_group *cfqg, struct cfq_group *curr_cfqg) { }
590 static inline void cfqg_stats_end_empty_time(struct cfqg_stats *stats) { }
591 static inline void cfqg_stats_update_dequeue(struct cfq_group *cfqg) { }
592 static inline void cfqg_stats_set_start_empty_time(struct cfq_group *cfqg) { }
593 static inline void cfqg_stats_update_idle_time(struct cfq_group *cfqg) { }
594 static inline void cfqg_stats_set_start_idle_time(struct cfq_group *cfqg) { }
595 static inline void cfqg_stats_update_avg_queue_size(struct cfq_group *cfqg) { }
597 #endif /* CONFIG_CFQ_GROUP_IOSCHED && CONFIG_DEBUG_BLK_CGROUP */
599 #ifdef CONFIG_CFQ_GROUP_IOSCHED
601 static inline struct cfq_group *pd_to_cfqg(struct blkg_policy_data *pd)
603 return pd ? container_of(pd, struct cfq_group, pd) : NULL;
606 static struct cfq_group_data
607 *cpd_to_cfqgd(struct blkcg_policy_data *cpd)
609 return cpd ? container_of(cpd, struct cfq_group_data, cpd) : NULL;
612 static inline struct blkcg_gq *cfqg_to_blkg(struct cfq_group *cfqg)
614 return pd_to_blkg(&cfqg->pd);
617 static struct blkcg_policy blkcg_policy_cfq;
619 static inline struct cfq_group *blkg_to_cfqg(struct blkcg_gq *blkg)
621 return pd_to_cfqg(blkg_to_pd(blkg, &blkcg_policy_cfq));
624 static struct cfq_group_data *blkcg_to_cfqgd(struct blkcg *blkcg)
626 return cpd_to_cfqgd(blkcg_to_cpd(blkcg, &blkcg_policy_cfq));
629 static inline struct cfq_group *cfqg_parent(struct cfq_group *cfqg)
631 struct blkcg_gq *pblkg = cfqg_to_blkg(cfqg)->parent;
633 return pblkg ? blkg_to_cfqg(pblkg) : NULL;
636 static inline bool cfqg_is_descendant(struct cfq_group *cfqg,
637 struct cfq_group *ancestor)
639 return cgroup_is_descendant(cfqg_to_blkg(cfqg)->blkcg->css.cgroup,
640 cfqg_to_blkg(ancestor)->blkcg->css.cgroup);
643 static inline void cfqg_get(struct cfq_group *cfqg)
645 return blkg_get(cfqg_to_blkg(cfqg));
648 static inline void cfqg_put(struct cfq_group *cfqg)
650 return blkg_put(cfqg_to_blkg(cfqg));
653 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...) do { \
656 blkg_path(cfqg_to_blkg((cfqq)->cfqg), __pbuf, sizeof(__pbuf)); \
657 blk_add_trace_msg((cfqd)->queue, "cfq%d%c%c %s " fmt, (cfqq)->pid, \
658 cfq_cfqq_sync((cfqq)) ? 'S' : 'A', \
659 cfqq_type((cfqq)) == SYNC_NOIDLE_WORKLOAD ? 'N' : ' ',\
663 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...) do { \
666 blkg_path(cfqg_to_blkg(cfqg), __pbuf, sizeof(__pbuf)); \
667 blk_add_trace_msg((cfqd)->queue, "%s " fmt, __pbuf, ##args); \
670 static inline void cfqg_stats_update_io_add(struct cfq_group *cfqg,
671 struct cfq_group *curr_cfqg,
674 blkg_rwstat_add(&cfqg->stats.queued, op, 1);
675 cfqg_stats_end_empty_time(&cfqg->stats);
676 cfqg_stats_set_start_group_wait_time(cfqg, curr_cfqg);
679 static inline void cfqg_stats_update_timeslice_used(struct cfq_group *cfqg,
680 uint64_t time, unsigned long unaccounted_time)
682 blkg_stat_add(&cfqg->stats.time, time);
683 #ifdef CONFIG_DEBUG_BLK_CGROUP
684 blkg_stat_add(&cfqg->stats.unaccounted_time, unaccounted_time);
688 static inline void cfqg_stats_update_io_remove(struct cfq_group *cfqg,
691 blkg_rwstat_add(&cfqg->stats.queued, op, -1);
694 static inline void cfqg_stats_update_io_merged(struct cfq_group *cfqg,
697 blkg_rwstat_add(&cfqg->stats.merged, op, 1);
700 static inline void cfqg_stats_update_completion(struct cfq_group *cfqg,
701 uint64_t start_time, uint64_t io_start_time,
704 struct cfqg_stats *stats = &cfqg->stats;
705 unsigned long long now = sched_clock();
707 if (time_after64(now, io_start_time))
708 blkg_rwstat_add(&stats->service_time, op, now - io_start_time);
709 if (time_after64(io_start_time, start_time))
710 blkg_rwstat_add(&stats->wait_time, op,
711 io_start_time - start_time);
715 static void cfqg_stats_reset(struct cfqg_stats *stats)
717 /* queued stats shouldn't be cleared */
718 blkg_rwstat_reset(&stats->merged);
719 blkg_rwstat_reset(&stats->service_time);
720 blkg_rwstat_reset(&stats->wait_time);
721 blkg_stat_reset(&stats->time);
722 #ifdef CONFIG_DEBUG_BLK_CGROUP
723 blkg_stat_reset(&stats->unaccounted_time);
724 blkg_stat_reset(&stats->avg_queue_size_sum);
725 blkg_stat_reset(&stats->avg_queue_size_samples);
726 blkg_stat_reset(&stats->dequeue);
727 blkg_stat_reset(&stats->group_wait_time);
728 blkg_stat_reset(&stats->idle_time);
729 blkg_stat_reset(&stats->empty_time);
734 static void cfqg_stats_add_aux(struct cfqg_stats *to, struct cfqg_stats *from)
736 /* queued stats shouldn't be cleared */
737 blkg_rwstat_add_aux(&to->merged, &from->merged);
738 blkg_rwstat_add_aux(&to->service_time, &from->service_time);
739 blkg_rwstat_add_aux(&to->wait_time, &from->wait_time);
740 blkg_stat_add_aux(&from->time, &from->time);
741 #ifdef CONFIG_DEBUG_BLK_CGROUP
742 blkg_stat_add_aux(&to->unaccounted_time, &from->unaccounted_time);
743 blkg_stat_add_aux(&to->avg_queue_size_sum, &from->avg_queue_size_sum);
744 blkg_stat_add_aux(&to->avg_queue_size_samples, &from->avg_queue_size_samples);
745 blkg_stat_add_aux(&to->dequeue, &from->dequeue);
746 blkg_stat_add_aux(&to->group_wait_time, &from->group_wait_time);
747 blkg_stat_add_aux(&to->idle_time, &from->idle_time);
748 blkg_stat_add_aux(&to->empty_time, &from->empty_time);
753 * Transfer @cfqg's stats to its parent's aux counts so that the ancestors'
754 * recursive stats can still account for the amount used by this cfqg after
757 static void cfqg_stats_xfer_dead(struct cfq_group *cfqg)
759 struct cfq_group *parent = cfqg_parent(cfqg);
761 lockdep_assert_held(cfqg_to_blkg(cfqg)->q->queue_lock);
763 if (unlikely(!parent))
766 cfqg_stats_add_aux(&parent->stats, &cfqg->stats);
767 cfqg_stats_reset(&cfqg->stats);
770 #else /* CONFIG_CFQ_GROUP_IOSCHED */
772 static inline struct cfq_group *cfqg_parent(struct cfq_group *cfqg) { return NULL; }
773 static inline bool cfqg_is_descendant(struct cfq_group *cfqg,
774 struct cfq_group *ancestor)
778 static inline void cfqg_get(struct cfq_group *cfqg) { }
779 static inline void cfqg_put(struct cfq_group *cfqg) { }
781 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
782 blk_add_trace_msg((cfqd)->queue, "cfq%d%c%c " fmt, (cfqq)->pid, \
783 cfq_cfqq_sync((cfqq)) ? 'S' : 'A', \
784 cfqq_type((cfqq)) == SYNC_NOIDLE_WORKLOAD ? 'N' : ' ',\
786 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...) do {} while (0)
788 static inline void cfqg_stats_update_io_add(struct cfq_group *cfqg,
789 struct cfq_group *curr_cfqg, unsigned int op) { }
790 static inline void cfqg_stats_update_timeslice_used(struct cfq_group *cfqg,
791 uint64_t time, unsigned long unaccounted_time) { }
792 static inline void cfqg_stats_update_io_remove(struct cfq_group *cfqg,
794 static inline void cfqg_stats_update_io_merged(struct cfq_group *cfqg,
796 static inline void cfqg_stats_update_completion(struct cfq_group *cfqg,
797 uint64_t start_time, uint64_t io_start_time,
800 #endif /* CONFIG_CFQ_GROUP_IOSCHED */
802 #define cfq_log(cfqd, fmt, args...) \
803 blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
805 /* Traverses through cfq group service trees */
806 #define for_each_cfqg_st(cfqg, i, j, st) \
807 for (i = 0; i <= IDLE_WORKLOAD; i++) \
808 for (j = 0, st = i < IDLE_WORKLOAD ? &cfqg->service_trees[i][j]\
809 : &cfqg->service_tree_idle; \
810 (i < IDLE_WORKLOAD && j <= SYNC_WORKLOAD) || \
811 (i == IDLE_WORKLOAD && j == 0); \
812 j++, st = i < IDLE_WORKLOAD ? \
813 &cfqg->service_trees[i][j]: NULL) \
815 static inline bool cfq_io_thinktime_big(struct cfq_data *cfqd,
816 struct cfq_ttime *ttime, bool group_idle)
819 if (!sample_valid(ttime->ttime_samples))
822 slice = cfqd->cfq_group_idle;
824 slice = cfqd->cfq_slice_idle;
825 return ttime->ttime_mean > slice;
828 static inline bool iops_mode(struct cfq_data *cfqd)
831 * If we are not idling on queues and it is a NCQ drive, parallel
832 * execution of requests is on and measuring time is not possible
833 * in most of the cases until and unless we drive shallower queue
834 * depths and that becomes a performance bottleneck. In such cases
835 * switch to start providing fairness in terms of number of IOs.
837 if (!cfqd->cfq_slice_idle && cfqd->hw_tag)
843 static inline enum wl_class_t cfqq_class(struct cfq_queue *cfqq)
845 if (cfq_class_idle(cfqq))
846 return IDLE_WORKLOAD;
847 if (cfq_class_rt(cfqq))
853 static enum wl_type_t cfqq_type(struct cfq_queue *cfqq)
855 if (!cfq_cfqq_sync(cfqq))
856 return ASYNC_WORKLOAD;
857 if (!cfq_cfqq_idle_window(cfqq))
858 return SYNC_NOIDLE_WORKLOAD;
859 return SYNC_WORKLOAD;
862 static inline int cfq_group_busy_queues_wl(enum wl_class_t wl_class,
863 struct cfq_data *cfqd,
864 struct cfq_group *cfqg)
866 if (wl_class == IDLE_WORKLOAD)
867 return cfqg->service_tree_idle.count;
869 return cfqg->service_trees[wl_class][ASYNC_WORKLOAD].count +
870 cfqg->service_trees[wl_class][SYNC_NOIDLE_WORKLOAD].count +
871 cfqg->service_trees[wl_class][SYNC_WORKLOAD].count;
874 static inline int cfqg_busy_async_queues(struct cfq_data *cfqd,
875 struct cfq_group *cfqg)
877 return cfqg->service_trees[RT_WORKLOAD][ASYNC_WORKLOAD].count +
878 cfqg->service_trees[BE_WORKLOAD][ASYNC_WORKLOAD].count;
881 static void cfq_dispatch_insert(struct request_queue *, struct request *);
882 static struct cfq_queue *cfq_get_queue(struct cfq_data *cfqd, bool is_sync,
883 struct cfq_io_cq *cic, struct bio *bio);
885 static inline struct cfq_io_cq *icq_to_cic(struct io_cq *icq)
887 /* cic->icq is the first member, %NULL will convert to %NULL */
888 return container_of(icq, struct cfq_io_cq, icq);
891 static inline struct cfq_io_cq *cfq_cic_lookup(struct cfq_data *cfqd,
892 struct io_context *ioc)
895 return icq_to_cic(ioc_lookup_icq(ioc, cfqd->queue));
899 static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_cq *cic, bool is_sync)
901 return cic->cfqq[is_sync];
904 static inline void cic_set_cfqq(struct cfq_io_cq *cic, struct cfq_queue *cfqq,
907 cic->cfqq[is_sync] = cfqq;
910 static inline struct cfq_data *cic_to_cfqd(struct cfq_io_cq *cic)
912 return cic->icq.q->elevator->elevator_data;
916 * scheduler run of queue, if there are requests pending and no one in the
917 * driver that will restart queueing
919 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
921 if (cfqd->busy_queues) {
922 cfq_log(cfqd, "schedule dispatch");
923 kblockd_schedule_work(&cfqd->unplug_work);
928 * Scale schedule slice based on io priority. Use the sync time slice only
929 * if a queue is marked sync and has sync io queued. A sync queue with async
930 * io only, should not get full sync slice length.
932 static inline u64 cfq_prio_slice(struct cfq_data *cfqd, bool sync,
935 u64 base_slice = cfqd->cfq_slice[sync];
936 u64 slice = div_u64(base_slice, CFQ_SLICE_SCALE);
938 WARN_ON(prio >= IOPRIO_BE_NR);
940 return base_slice + (slice * (4 - prio));
944 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
946 return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
950 * cfqg_scale_charge - scale disk time charge according to cfqg weight
951 * @charge: disk time being charged
952 * @vfraction: vfraction of the cfqg, fixed point w/ CFQ_SERVICE_SHIFT
954 * Scale @charge according to @vfraction, which is in range (0, 1]. The
955 * scaling is inversely proportional.
957 * scaled = charge / vfraction
959 * The result is also in fixed point w/ CFQ_SERVICE_SHIFT.
961 static inline u64 cfqg_scale_charge(u64 charge,
962 unsigned int vfraction)
964 u64 c = charge << CFQ_SERVICE_SHIFT; /* make it fixed point */
966 /* charge / vfraction */
967 c <<= CFQ_SERVICE_SHIFT;
968 return div_u64(c, vfraction);
971 static inline u64 max_vdisktime(u64 min_vdisktime, u64 vdisktime)
973 s64 delta = (s64)(vdisktime - min_vdisktime);
975 min_vdisktime = vdisktime;
977 return min_vdisktime;
980 static inline u64 min_vdisktime(u64 min_vdisktime, u64 vdisktime)
982 s64 delta = (s64)(vdisktime - min_vdisktime);
984 min_vdisktime = vdisktime;
986 return min_vdisktime;
989 static void update_min_vdisktime(struct cfq_rb_root *st)
991 struct cfq_group *cfqg;
994 cfqg = rb_entry_cfqg(st->left);
995 st->min_vdisktime = max_vdisktime(st->min_vdisktime,
1001 * get averaged number of queues of RT/BE priority.
1002 * average is updated, with a formula that gives more weight to higher numbers,
1003 * to quickly follows sudden increases and decrease slowly
1006 static inline unsigned cfq_group_get_avg_queues(struct cfq_data *cfqd,
1007 struct cfq_group *cfqg, bool rt)
1009 unsigned min_q, max_q;
1010 unsigned mult = cfq_hist_divisor - 1;
1011 unsigned round = cfq_hist_divisor / 2;
1012 unsigned busy = cfq_group_busy_queues_wl(rt, cfqd, cfqg);
1014 min_q = min(cfqg->busy_queues_avg[rt], busy);
1015 max_q = max(cfqg->busy_queues_avg[rt], busy);
1016 cfqg->busy_queues_avg[rt] = (mult * max_q + min_q + round) /
1018 return cfqg->busy_queues_avg[rt];
1022 cfq_group_slice(struct cfq_data *cfqd, struct cfq_group *cfqg)
1024 return cfqd->cfq_target_latency * cfqg->vfraction >> CFQ_SERVICE_SHIFT;
1028 cfq_scaled_cfqq_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1030 u64 slice = cfq_prio_to_slice(cfqd, cfqq);
1031 if (cfqd->cfq_latency) {
1033 * interested queues (we consider only the ones with the same
1034 * priority class in the cfq group)
1036 unsigned iq = cfq_group_get_avg_queues(cfqd, cfqq->cfqg,
1037 cfq_class_rt(cfqq));
1038 u64 sync_slice = cfqd->cfq_slice[1];
1039 u64 expect_latency = sync_slice * iq;
1040 u64 group_slice = cfq_group_slice(cfqd, cfqq->cfqg);
1042 if (expect_latency > group_slice) {
1043 u64 base_low_slice = 2 * cfqd->cfq_slice_idle;
1046 /* scale low_slice according to IO priority
1047 * and sync vs async */
1048 low_slice = div64_u64(base_low_slice*slice, sync_slice);
1049 low_slice = min(slice, low_slice);
1050 /* the adapted slice value is scaled to fit all iqs
1051 * into the target latency */
1052 slice = div64_u64(slice*group_slice, expect_latency);
1053 slice = max(slice, low_slice);
1060 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1062 u64 slice = cfq_scaled_cfqq_slice(cfqd, cfqq);
1063 u64 now = ktime_get_ns();
1065 cfqq->slice_start = now;
1066 cfqq->slice_end = now + slice;
1067 cfqq->allocated_slice = slice;
1068 cfq_log_cfqq(cfqd, cfqq, "set_slice=%llu", cfqq->slice_end - now);
1072 * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
1073 * isn't valid until the first request from the dispatch is activated
1074 * and the slice time set.
1076 static inline bool cfq_slice_used(struct cfq_queue *cfqq)
1078 if (cfq_cfqq_slice_new(cfqq))
1080 if (ktime_get_ns() < cfqq->slice_end)
1087 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
1088 * We choose the request that is closest to the head right now. Distance
1089 * behind the head is penalized and only allowed to a certain extent.
1091 static struct request *
1092 cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2, sector_t last)
1094 sector_t s1, s2, d1 = 0, d2 = 0;
1095 unsigned long back_max;
1096 #define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */
1097 #define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */
1098 unsigned wrap = 0; /* bit mask: requests behind the disk head? */
1100 if (rq1 == NULL || rq1 == rq2)
1105 if (rq_is_sync(rq1) != rq_is_sync(rq2))
1106 return rq_is_sync(rq1) ? rq1 : rq2;
1108 if ((rq1->cmd_flags ^ rq2->cmd_flags) & REQ_PRIO)
1109 return rq1->cmd_flags & REQ_PRIO ? rq1 : rq2;
1111 s1 = blk_rq_pos(rq1);
1112 s2 = blk_rq_pos(rq2);
1115 * by definition, 1KiB is 2 sectors
1117 back_max = cfqd->cfq_back_max * 2;
1120 * Strict one way elevator _except_ in the case where we allow
1121 * short backward seeks which are biased as twice the cost of a
1122 * similar forward seek.
1126 else if (s1 + back_max >= last)
1127 d1 = (last - s1) * cfqd->cfq_back_penalty;
1129 wrap |= CFQ_RQ1_WRAP;
1133 else if (s2 + back_max >= last)
1134 d2 = (last - s2) * cfqd->cfq_back_penalty;
1136 wrap |= CFQ_RQ2_WRAP;
1138 /* Found required data */
1141 * By doing switch() on the bit mask "wrap" we avoid having to
1142 * check two variables for all permutations: --> faster!
1145 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
1161 case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
1164 * Since both rqs are wrapped,
1165 * start with the one that's further behind head
1166 * (--> only *one* back seek required),
1167 * since back seek takes more time than forward.
1177 * The below is leftmost cache rbtree addon
1179 static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
1181 /* Service tree is empty */
1186 root->left = rb_first(&root->rb);
1189 return rb_entry(root->left, struct cfq_queue, rb_node);
1194 static struct cfq_group *cfq_rb_first_group(struct cfq_rb_root *root)
1197 root->left = rb_first(&root->rb);
1200 return rb_entry_cfqg(root->left);
1205 static void rb_erase_init(struct rb_node *n, struct rb_root *root)
1211 static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
1213 if (root->left == n)
1215 rb_erase_init(n, &root->rb);
1220 * would be nice to take fifo expire time into account as well
1222 static struct request *
1223 cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1224 struct request *last)
1226 struct rb_node *rbnext = rb_next(&last->rb_node);
1227 struct rb_node *rbprev = rb_prev(&last->rb_node);
1228 struct request *next = NULL, *prev = NULL;
1230 BUG_ON(RB_EMPTY_NODE(&last->rb_node));
1233 prev = rb_entry_rq(rbprev);
1236 next = rb_entry_rq(rbnext);
1238 rbnext = rb_first(&cfqq->sort_list);
1239 if (rbnext && rbnext != &last->rb_node)
1240 next = rb_entry_rq(rbnext);
1243 return cfq_choose_req(cfqd, next, prev, blk_rq_pos(last));
1246 static u64 cfq_slice_offset(struct cfq_data *cfqd,
1247 struct cfq_queue *cfqq)
1250 * just an approximation, should be ok.
1252 return (cfqq->cfqg->nr_cfqq - 1) * (cfq_prio_slice(cfqd, 1, 0) -
1253 cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
1257 cfqg_key(struct cfq_rb_root *st, struct cfq_group *cfqg)
1259 return cfqg->vdisktime - st->min_vdisktime;
1263 __cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
1265 struct rb_node **node = &st->rb.rb_node;
1266 struct rb_node *parent = NULL;
1267 struct cfq_group *__cfqg;
1268 s64 key = cfqg_key(st, cfqg);
1271 while (*node != NULL) {
1273 __cfqg = rb_entry_cfqg(parent);
1275 if (key < cfqg_key(st, __cfqg))
1276 node = &parent->rb_left;
1278 node = &parent->rb_right;
1284 st->left = &cfqg->rb_node;
1286 rb_link_node(&cfqg->rb_node, parent, node);
1287 rb_insert_color(&cfqg->rb_node, &st->rb);
1291 * This has to be called only on activation of cfqg
1294 cfq_update_group_weight(struct cfq_group *cfqg)
1296 if (cfqg->new_weight) {
1297 cfqg->weight = cfqg->new_weight;
1298 cfqg->new_weight = 0;
1303 cfq_update_group_leaf_weight(struct cfq_group *cfqg)
1305 BUG_ON(!RB_EMPTY_NODE(&cfqg->rb_node));
1307 if (cfqg->new_leaf_weight) {
1308 cfqg->leaf_weight = cfqg->new_leaf_weight;
1309 cfqg->new_leaf_weight = 0;
1314 cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
1316 unsigned int vfr = 1 << CFQ_SERVICE_SHIFT; /* start with 1 */
1317 struct cfq_group *pos = cfqg;
1318 struct cfq_group *parent;
1321 /* add to the service tree */
1322 BUG_ON(!RB_EMPTY_NODE(&cfqg->rb_node));
1325 * Update leaf_weight. We cannot update weight at this point
1326 * because cfqg might already have been activated and is
1327 * contributing its current weight to the parent's child_weight.
1329 cfq_update_group_leaf_weight(cfqg);
1330 __cfq_group_service_tree_add(st, cfqg);
1333 * Activate @cfqg and calculate the portion of vfraction @cfqg is
1334 * entitled to. vfraction is calculated by walking the tree
1335 * towards the root calculating the fraction it has at each level.
1336 * The compounded ratio is how much vfraction @cfqg owns.
1338 * Start with the proportion tasks in this cfqg has against active
1339 * children cfqgs - its leaf_weight against children_weight.
1341 propagate = !pos->nr_active++;
1342 pos->children_weight += pos->leaf_weight;
1343 vfr = vfr * pos->leaf_weight / pos->children_weight;
1346 * Compound ->weight walking up the tree. Both activation and
1347 * vfraction calculation are done in the same loop. Propagation
1348 * stops once an already activated node is met. vfraction
1349 * calculation should always continue to the root.
1351 while ((parent = cfqg_parent(pos))) {
1353 cfq_update_group_weight(pos);
1354 propagate = !parent->nr_active++;
1355 parent->children_weight += pos->weight;
1357 vfr = vfr * pos->weight / parent->children_weight;
1361 cfqg->vfraction = max_t(unsigned, vfr, 1);
1365 cfq_group_notify_queue_add(struct cfq_data *cfqd, struct cfq_group *cfqg)
1367 struct cfq_rb_root *st = &cfqd->grp_service_tree;
1368 struct cfq_group *__cfqg;
1372 if (!RB_EMPTY_NODE(&cfqg->rb_node))
1376 * Currently put the group at the end. Later implement something
1377 * so that groups get lesser vtime based on their weights, so that
1378 * if group does not loose all if it was not continuously backlogged.
1380 n = rb_last(&st->rb);
1382 __cfqg = rb_entry_cfqg(n);
1383 cfqg->vdisktime = __cfqg->vdisktime + CFQ_IDLE_DELAY;
1385 cfqg->vdisktime = st->min_vdisktime;
1386 cfq_group_service_tree_add(st, cfqg);
1390 cfq_group_service_tree_del(struct cfq_rb_root *st, struct cfq_group *cfqg)
1392 struct cfq_group *pos = cfqg;
1396 * Undo activation from cfq_group_service_tree_add(). Deactivate
1397 * @cfqg and propagate deactivation upwards.
1399 propagate = !--pos->nr_active;
1400 pos->children_weight -= pos->leaf_weight;
1403 struct cfq_group *parent = cfqg_parent(pos);
1405 /* @pos has 0 nr_active at this point */
1406 WARN_ON_ONCE(pos->children_weight);
1412 propagate = !--parent->nr_active;
1413 parent->children_weight -= pos->weight;
1417 /* remove from the service tree */
1418 if (!RB_EMPTY_NODE(&cfqg->rb_node))
1419 cfq_rb_erase(&cfqg->rb_node, st);
1423 cfq_group_notify_queue_del(struct cfq_data *cfqd, struct cfq_group *cfqg)
1425 struct cfq_rb_root *st = &cfqd->grp_service_tree;
1427 BUG_ON(cfqg->nr_cfqq < 1);
1430 /* If there are other cfq queues under this group, don't delete it */
1434 cfq_log_cfqg(cfqd, cfqg, "del_from_rr group");
1435 cfq_group_service_tree_del(st, cfqg);
1436 cfqg->saved_wl_slice = 0;
1437 cfqg_stats_update_dequeue(cfqg);
1440 static inline u64 cfq_cfqq_slice_usage(struct cfq_queue *cfqq,
1441 u64 *unaccounted_time)
1444 u64 now = ktime_get_ns();
1447 * Queue got expired before even a single request completed or
1448 * got expired immediately after first request completion.
1450 if (!cfqq->slice_start || cfqq->slice_start == now) {
1452 * Also charge the seek time incurred to the group, otherwise
1453 * if there are mutiple queues in the group, each can dispatch
1454 * a single request on seeky media and cause lots of seek time
1455 * and group will never know it.
1457 slice_used = max_t(u64, (now - cfqq->dispatch_start),
1458 jiffies_to_nsecs(1));
1460 slice_used = now - cfqq->slice_start;
1461 if (slice_used > cfqq->allocated_slice) {
1462 *unaccounted_time = slice_used - cfqq->allocated_slice;
1463 slice_used = cfqq->allocated_slice;
1465 if (cfqq->slice_start > cfqq->dispatch_start)
1466 *unaccounted_time += cfqq->slice_start -
1467 cfqq->dispatch_start;
1473 static void cfq_group_served(struct cfq_data *cfqd, struct cfq_group *cfqg,
1474 struct cfq_queue *cfqq)
1476 struct cfq_rb_root *st = &cfqd->grp_service_tree;
1477 u64 used_sl, charge, unaccounted_sl = 0;
1478 int nr_sync = cfqg->nr_cfqq - cfqg_busy_async_queues(cfqd, cfqg)
1479 - cfqg->service_tree_idle.count;
1481 u64 now = ktime_get_ns();
1483 BUG_ON(nr_sync < 0);
1484 used_sl = charge = cfq_cfqq_slice_usage(cfqq, &unaccounted_sl);
1486 if (iops_mode(cfqd))
1487 charge = cfqq->slice_dispatch;
1488 else if (!cfq_cfqq_sync(cfqq) && !nr_sync)
1489 charge = cfqq->allocated_slice;
1492 * Can't update vdisktime while on service tree and cfqg->vfraction
1493 * is valid only while on it. Cache vfr, leave the service tree,
1494 * update vdisktime and go back on. The re-addition to the tree
1495 * will also update the weights as necessary.
1497 vfr = cfqg->vfraction;
1498 cfq_group_service_tree_del(st, cfqg);
1499 cfqg->vdisktime += cfqg_scale_charge(charge, vfr);
1500 cfq_group_service_tree_add(st, cfqg);
1502 /* This group is being expired. Save the context */
1503 if (cfqd->workload_expires > now) {
1504 cfqg->saved_wl_slice = cfqd->workload_expires - now;
1505 cfqg->saved_wl_type = cfqd->serving_wl_type;
1506 cfqg->saved_wl_class = cfqd->serving_wl_class;
1508 cfqg->saved_wl_slice = 0;
1510 cfq_log_cfqg(cfqd, cfqg, "served: vt=%llu min_vt=%llu", cfqg->vdisktime,
1512 cfq_log_cfqq(cfqq->cfqd, cfqq,
1513 "sl_used=%llu disp=%llu charge=%llu iops=%u sect=%lu",
1514 used_sl, cfqq->slice_dispatch, charge,
1515 iops_mode(cfqd), cfqq->nr_sectors);
1516 cfqg_stats_update_timeslice_used(cfqg, used_sl, unaccounted_sl);
1517 cfqg_stats_set_start_empty_time(cfqg);
1521 * cfq_init_cfqg_base - initialize base part of a cfq_group
1522 * @cfqg: cfq_group to initialize
1524 * Initialize the base part which is used whether %CONFIG_CFQ_GROUP_IOSCHED
1525 * is enabled or not.
1527 static void cfq_init_cfqg_base(struct cfq_group *cfqg)
1529 struct cfq_rb_root *st;
1532 for_each_cfqg_st(cfqg, i, j, st)
1534 RB_CLEAR_NODE(&cfqg->rb_node);
1536 cfqg->ttime.last_end_request = ktime_get_ns();
1539 #ifdef CONFIG_CFQ_GROUP_IOSCHED
1540 static int __cfq_set_weight(struct cgroup_subsys_state *css, u64 val,
1541 bool on_dfl, bool reset_dev, bool is_leaf_weight);
1543 static void cfqg_stats_exit(struct cfqg_stats *stats)
1545 blkg_rwstat_exit(&stats->merged);
1546 blkg_rwstat_exit(&stats->service_time);
1547 blkg_rwstat_exit(&stats->wait_time);
1548 blkg_rwstat_exit(&stats->queued);
1549 blkg_stat_exit(&stats->time);
1550 #ifdef CONFIG_DEBUG_BLK_CGROUP
1551 blkg_stat_exit(&stats->unaccounted_time);
1552 blkg_stat_exit(&stats->avg_queue_size_sum);
1553 blkg_stat_exit(&stats->avg_queue_size_samples);
1554 blkg_stat_exit(&stats->dequeue);
1555 blkg_stat_exit(&stats->group_wait_time);
1556 blkg_stat_exit(&stats->idle_time);
1557 blkg_stat_exit(&stats->empty_time);
1561 static int cfqg_stats_init(struct cfqg_stats *stats, gfp_t gfp)
1563 if (blkg_rwstat_init(&stats->merged, gfp) ||
1564 blkg_rwstat_init(&stats->service_time, gfp) ||
1565 blkg_rwstat_init(&stats->wait_time, gfp) ||
1566 blkg_rwstat_init(&stats->queued, gfp) ||
1567 blkg_stat_init(&stats->time, gfp))
1570 #ifdef CONFIG_DEBUG_BLK_CGROUP
1571 if (blkg_stat_init(&stats->unaccounted_time, gfp) ||
1572 blkg_stat_init(&stats->avg_queue_size_sum, gfp) ||
1573 blkg_stat_init(&stats->avg_queue_size_samples, gfp) ||
1574 blkg_stat_init(&stats->dequeue, gfp) ||
1575 blkg_stat_init(&stats->group_wait_time, gfp) ||
1576 blkg_stat_init(&stats->idle_time, gfp) ||
1577 blkg_stat_init(&stats->empty_time, gfp))
1582 cfqg_stats_exit(stats);
1586 static struct blkcg_policy_data *cfq_cpd_alloc(gfp_t gfp)
1588 struct cfq_group_data *cgd;
1590 cgd = kzalloc(sizeof(*cgd), gfp);
1596 static void cfq_cpd_init(struct blkcg_policy_data *cpd)
1598 struct cfq_group_data *cgd = cpd_to_cfqgd(cpd);
1599 unsigned int weight = cgroup_subsys_on_dfl(io_cgrp_subsys) ?
1600 CGROUP_WEIGHT_DFL : CFQ_WEIGHT_LEGACY_DFL;
1602 if (cpd_to_blkcg(cpd) == &blkcg_root)
1605 cgd->weight = weight;
1606 cgd->leaf_weight = weight;
1609 static void cfq_cpd_free(struct blkcg_policy_data *cpd)
1611 kfree(cpd_to_cfqgd(cpd));
1614 static void cfq_cpd_bind(struct blkcg_policy_data *cpd)
1616 struct blkcg *blkcg = cpd_to_blkcg(cpd);
1617 bool on_dfl = cgroup_subsys_on_dfl(io_cgrp_subsys);
1618 unsigned int weight = on_dfl ? CGROUP_WEIGHT_DFL : CFQ_WEIGHT_LEGACY_DFL;
1620 if (blkcg == &blkcg_root)
1623 WARN_ON_ONCE(__cfq_set_weight(&blkcg->css, weight, on_dfl, true, false));
1624 WARN_ON_ONCE(__cfq_set_weight(&blkcg->css, weight, on_dfl, true, true));
1627 static struct blkg_policy_data *cfq_pd_alloc(gfp_t gfp, int node)
1629 struct cfq_group *cfqg;
1631 cfqg = kzalloc_node(sizeof(*cfqg), gfp, node);
1635 cfq_init_cfqg_base(cfqg);
1636 if (cfqg_stats_init(&cfqg->stats, gfp)) {
1644 static void cfq_pd_init(struct blkg_policy_data *pd)
1646 struct cfq_group *cfqg = pd_to_cfqg(pd);
1647 struct cfq_group_data *cgd = blkcg_to_cfqgd(pd->blkg->blkcg);
1649 cfqg->weight = cgd->weight;
1650 cfqg->leaf_weight = cgd->leaf_weight;
1653 static void cfq_pd_offline(struct blkg_policy_data *pd)
1655 struct cfq_group *cfqg = pd_to_cfqg(pd);
1658 for (i = 0; i < IOPRIO_BE_NR; i++) {
1659 if (cfqg->async_cfqq[0][i])
1660 cfq_put_queue(cfqg->async_cfqq[0][i]);
1661 if (cfqg->async_cfqq[1][i])
1662 cfq_put_queue(cfqg->async_cfqq[1][i]);
1665 if (cfqg->async_idle_cfqq)
1666 cfq_put_queue(cfqg->async_idle_cfqq);
1669 * @blkg is going offline and will be ignored by
1670 * blkg_[rw]stat_recursive_sum(). Transfer stats to the parent so
1671 * that they don't get lost. If IOs complete after this point, the
1672 * stats for them will be lost. Oh well...
1674 cfqg_stats_xfer_dead(cfqg);
1677 static void cfq_pd_free(struct blkg_policy_data *pd)
1679 struct cfq_group *cfqg = pd_to_cfqg(pd);
1681 cfqg_stats_exit(&cfqg->stats);
1685 static void cfq_pd_reset_stats(struct blkg_policy_data *pd)
1687 struct cfq_group *cfqg = pd_to_cfqg(pd);
1689 cfqg_stats_reset(&cfqg->stats);
1692 static struct cfq_group *cfq_lookup_cfqg(struct cfq_data *cfqd,
1693 struct blkcg *blkcg)
1695 struct blkcg_gq *blkg;
1697 blkg = blkg_lookup(blkcg, cfqd->queue);
1699 return blkg_to_cfqg(blkg);
1703 static void cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg)
1706 /* cfqq reference on cfqg */
1710 static u64 cfqg_prfill_weight_device(struct seq_file *sf,
1711 struct blkg_policy_data *pd, int off)
1713 struct cfq_group *cfqg = pd_to_cfqg(pd);
1715 if (!cfqg->dev_weight)
1717 return __blkg_prfill_u64(sf, pd, cfqg->dev_weight);
1720 static int cfqg_print_weight_device(struct seq_file *sf, void *v)
1722 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1723 cfqg_prfill_weight_device, &blkcg_policy_cfq,
1728 static u64 cfqg_prfill_leaf_weight_device(struct seq_file *sf,
1729 struct blkg_policy_data *pd, int off)
1731 struct cfq_group *cfqg = pd_to_cfqg(pd);
1733 if (!cfqg->dev_leaf_weight)
1735 return __blkg_prfill_u64(sf, pd, cfqg->dev_leaf_weight);
1738 static int cfqg_print_leaf_weight_device(struct seq_file *sf, void *v)
1740 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1741 cfqg_prfill_leaf_weight_device, &blkcg_policy_cfq,
1746 static int cfq_print_weight(struct seq_file *sf, void *v)
1748 struct blkcg *blkcg = css_to_blkcg(seq_css(sf));
1749 struct cfq_group_data *cgd = blkcg_to_cfqgd(blkcg);
1750 unsigned int val = 0;
1755 seq_printf(sf, "%u\n", val);
1759 static int cfq_print_leaf_weight(struct seq_file *sf, void *v)
1761 struct blkcg *blkcg = css_to_blkcg(seq_css(sf));
1762 struct cfq_group_data *cgd = blkcg_to_cfqgd(blkcg);
1763 unsigned int val = 0;
1766 val = cgd->leaf_weight;
1768 seq_printf(sf, "%u\n", val);
1772 static ssize_t __cfqg_set_weight_device(struct kernfs_open_file *of,
1773 char *buf, size_t nbytes, loff_t off,
1774 bool on_dfl, bool is_leaf_weight)
1776 unsigned int min = on_dfl ? CGROUP_WEIGHT_MIN : CFQ_WEIGHT_LEGACY_MIN;
1777 unsigned int max = on_dfl ? CGROUP_WEIGHT_MAX : CFQ_WEIGHT_LEGACY_MAX;
1778 struct blkcg *blkcg = css_to_blkcg(of_css(of));
1779 struct blkg_conf_ctx ctx;
1780 struct cfq_group *cfqg;
1781 struct cfq_group_data *cfqgd;
1785 ret = blkg_conf_prep(blkcg, &blkcg_policy_cfq, buf, &ctx);
1789 if (sscanf(ctx.body, "%llu", &v) == 1) {
1790 /* require "default" on dfl */
1794 } else if (!strcmp(strim(ctx.body), "default")) {
1801 cfqg = blkg_to_cfqg(ctx.blkg);
1802 cfqgd = blkcg_to_cfqgd(blkcg);
1805 if (!v || (v >= min && v <= max)) {
1806 if (!is_leaf_weight) {
1807 cfqg->dev_weight = v;
1808 cfqg->new_weight = v ?: cfqgd->weight;
1810 cfqg->dev_leaf_weight = v;
1811 cfqg->new_leaf_weight = v ?: cfqgd->leaf_weight;
1816 blkg_conf_finish(&ctx);
1817 return ret ?: nbytes;
1820 static ssize_t cfqg_set_weight_device(struct kernfs_open_file *of,
1821 char *buf, size_t nbytes, loff_t off)
1823 return __cfqg_set_weight_device(of, buf, nbytes, off, false, false);
1826 static ssize_t cfqg_set_leaf_weight_device(struct kernfs_open_file *of,
1827 char *buf, size_t nbytes, loff_t off)
1829 return __cfqg_set_weight_device(of, buf, nbytes, off, false, true);
1832 static int __cfq_set_weight(struct cgroup_subsys_state *css, u64 val,
1833 bool on_dfl, bool reset_dev, bool is_leaf_weight)
1835 unsigned int min = on_dfl ? CGROUP_WEIGHT_MIN : CFQ_WEIGHT_LEGACY_MIN;
1836 unsigned int max = on_dfl ? CGROUP_WEIGHT_MAX : CFQ_WEIGHT_LEGACY_MAX;
1837 struct blkcg *blkcg = css_to_blkcg(css);
1838 struct blkcg_gq *blkg;
1839 struct cfq_group_data *cfqgd;
1842 if (val < min || val > max)
1845 spin_lock_irq(&blkcg->lock);
1846 cfqgd = blkcg_to_cfqgd(blkcg);
1852 if (!is_leaf_weight)
1853 cfqgd->weight = val;
1855 cfqgd->leaf_weight = val;
1857 hlist_for_each_entry(blkg, &blkcg->blkg_list, blkcg_node) {
1858 struct cfq_group *cfqg = blkg_to_cfqg(blkg);
1863 if (!is_leaf_weight) {
1865 cfqg->dev_weight = 0;
1866 if (!cfqg->dev_weight)
1867 cfqg->new_weight = cfqgd->weight;
1870 cfqg->dev_leaf_weight = 0;
1871 if (!cfqg->dev_leaf_weight)
1872 cfqg->new_leaf_weight = cfqgd->leaf_weight;
1877 spin_unlock_irq(&blkcg->lock);
1881 static int cfq_set_weight(struct cgroup_subsys_state *css, struct cftype *cft,
1884 return __cfq_set_weight(css, val, false, false, false);
1887 static int cfq_set_leaf_weight(struct cgroup_subsys_state *css,
1888 struct cftype *cft, u64 val)
1890 return __cfq_set_weight(css, val, false, false, true);
1893 static int cfqg_print_stat(struct seq_file *sf, void *v)
1895 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), blkg_prfill_stat,
1896 &blkcg_policy_cfq, seq_cft(sf)->private, false);
1900 static int cfqg_print_rwstat(struct seq_file *sf, void *v)
1902 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), blkg_prfill_rwstat,
1903 &blkcg_policy_cfq, seq_cft(sf)->private, true);
1907 static u64 cfqg_prfill_stat_recursive(struct seq_file *sf,
1908 struct blkg_policy_data *pd, int off)
1910 u64 sum = blkg_stat_recursive_sum(pd_to_blkg(pd),
1911 &blkcg_policy_cfq, off);
1912 return __blkg_prfill_u64(sf, pd, sum);
1915 static u64 cfqg_prfill_rwstat_recursive(struct seq_file *sf,
1916 struct blkg_policy_data *pd, int off)
1918 struct blkg_rwstat sum = blkg_rwstat_recursive_sum(pd_to_blkg(pd),
1919 &blkcg_policy_cfq, off);
1920 return __blkg_prfill_rwstat(sf, pd, &sum);
1923 static int cfqg_print_stat_recursive(struct seq_file *sf, void *v)
1925 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1926 cfqg_prfill_stat_recursive, &blkcg_policy_cfq,
1927 seq_cft(sf)->private, false);
1931 static int cfqg_print_rwstat_recursive(struct seq_file *sf, void *v)
1933 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1934 cfqg_prfill_rwstat_recursive, &blkcg_policy_cfq,
1935 seq_cft(sf)->private, true);
1939 static u64 cfqg_prfill_sectors(struct seq_file *sf, struct blkg_policy_data *pd,
1942 u64 sum = blkg_rwstat_total(&pd->blkg->stat_bytes);
1944 return __blkg_prfill_u64(sf, pd, sum >> 9);
1947 static int cfqg_print_stat_sectors(struct seq_file *sf, void *v)
1949 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1950 cfqg_prfill_sectors, &blkcg_policy_cfq, 0, false);
1954 static u64 cfqg_prfill_sectors_recursive(struct seq_file *sf,
1955 struct blkg_policy_data *pd, int off)
1957 struct blkg_rwstat tmp = blkg_rwstat_recursive_sum(pd->blkg, NULL,
1958 offsetof(struct blkcg_gq, stat_bytes));
1959 u64 sum = atomic64_read(&tmp.aux_cnt[BLKG_RWSTAT_READ]) +
1960 atomic64_read(&tmp.aux_cnt[BLKG_RWSTAT_WRITE]);
1962 return __blkg_prfill_u64(sf, pd, sum >> 9);
1965 static int cfqg_print_stat_sectors_recursive(struct seq_file *sf, void *v)
1967 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1968 cfqg_prfill_sectors_recursive, &blkcg_policy_cfq, 0,
1973 #ifdef CONFIG_DEBUG_BLK_CGROUP
1974 static u64 cfqg_prfill_avg_queue_size(struct seq_file *sf,
1975 struct blkg_policy_data *pd, int off)
1977 struct cfq_group *cfqg = pd_to_cfqg(pd);
1978 u64 samples = blkg_stat_read(&cfqg->stats.avg_queue_size_samples);
1982 v = blkg_stat_read(&cfqg->stats.avg_queue_size_sum);
1983 v = div64_u64(v, samples);
1985 __blkg_prfill_u64(sf, pd, v);
1989 /* print avg_queue_size */
1990 static int cfqg_print_avg_queue_size(struct seq_file *sf, void *v)
1992 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1993 cfqg_prfill_avg_queue_size, &blkcg_policy_cfq,
1997 #endif /* CONFIG_DEBUG_BLK_CGROUP */
1999 static struct cftype cfq_blkcg_legacy_files[] = {
2000 /* on root, weight is mapped to leaf_weight */
2002 .name = "weight_device",
2003 .flags = CFTYPE_ONLY_ON_ROOT,
2004 .seq_show = cfqg_print_leaf_weight_device,
2005 .write = cfqg_set_leaf_weight_device,
2009 .flags = CFTYPE_ONLY_ON_ROOT,
2010 .seq_show = cfq_print_leaf_weight,
2011 .write_u64 = cfq_set_leaf_weight,
2014 /* no such mapping necessary for !roots */
2016 .name = "weight_device",
2017 .flags = CFTYPE_NOT_ON_ROOT,
2018 .seq_show = cfqg_print_weight_device,
2019 .write = cfqg_set_weight_device,
2023 .flags = CFTYPE_NOT_ON_ROOT,
2024 .seq_show = cfq_print_weight,
2025 .write_u64 = cfq_set_weight,
2029 .name = "leaf_weight_device",
2030 .seq_show = cfqg_print_leaf_weight_device,
2031 .write = cfqg_set_leaf_weight_device,
2034 .name = "leaf_weight",
2035 .seq_show = cfq_print_leaf_weight,
2036 .write_u64 = cfq_set_leaf_weight,
2039 /* statistics, covers only the tasks in the cfqg */
2042 .private = offsetof(struct cfq_group, stats.time),
2043 .seq_show = cfqg_print_stat,
2047 .seq_show = cfqg_print_stat_sectors,
2050 .name = "io_service_bytes",
2051 .private = (unsigned long)&blkcg_policy_cfq,
2052 .seq_show = blkg_print_stat_bytes,
2055 .name = "io_serviced",
2056 .private = (unsigned long)&blkcg_policy_cfq,
2057 .seq_show = blkg_print_stat_ios,
2060 .name = "io_service_time",
2061 .private = offsetof(struct cfq_group, stats.service_time),
2062 .seq_show = cfqg_print_rwstat,
2065 .name = "io_wait_time",
2066 .private = offsetof(struct cfq_group, stats.wait_time),
2067 .seq_show = cfqg_print_rwstat,
2070 .name = "io_merged",
2071 .private = offsetof(struct cfq_group, stats.merged),
2072 .seq_show = cfqg_print_rwstat,
2075 .name = "io_queued",
2076 .private = offsetof(struct cfq_group, stats.queued),
2077 .seq_show = cfqg_print_rwstat,
2080 /* the same statictics which cover the cfqg and its descendants */
2082 .name = "time_recursive",
2083 .private = offsetof(struct cfq_group, stats.time),
2084 .seq_show = cfqg_print_stat_recursive,
2087 .name = "sectors_recursive",
2088 .seq_show = cfqg_print_stat_sectors_recursive,
2091 .name = "io_service_bytes_recursive",
2092 .private = (unsigned long)&blkcg_policy_cfq,
2093 .seq_show = blkg_print_stat_bytes_recursive,
2096 .name = "io_serviced_recursive",
2097 .private = (unsigned long)&blkcg_policy_cfq,
2098 .seq_show = blkg_print_stat_ios_recursive,
2101 .name = "io_service_time_recursive",
2102 .private = offsetof(struct cfq_group, stats.service_time),
2103 .seq_show = cfqg_print_rwstat_recursive,
2106 .name = "io_wait_time_recursive",
2107 .private = offsetof(struct cfq_group, stats.wait_time),
2108 .seq_show = cfqg_print_rwstat_recursive,
2111 .name = "io_merged_recursive",
2112 .private = offsetof(struct cfq_group, stats.merged),
2113 .seq_show = cfqg_print_rwstat_recursive,
2116 .name = "io_queued_recursive",
2117 .private = offsetof(struct cfq_group, stats.queued),
2118 .seq_show = cfqg_print_rwstat_recursive,
2120 #ifdef CONFIG_DEBUG_BLK_CGROUP
2122 .name = "avg_queue_size",
2123 .seq_show = cfqg_print_avg_queue_size,
2126 .name = "group_wait_time",
2127 .private = offsetof(struct cfq_group, stats.group_wait_time),
2128 .seq_show = cfqg_print_stat,
2131 .name = "idle_time",
2132 .private = offsetof(struct cfq_group, stats.idle_time),
2133 .seq_show = cfqg_print_stat,
2136 .name = "empty_time",
2137 .private = offsetof(struct cfq_group, stats.empty_time),
2138 .seq_show = cfqg_print_stat,
2142 .private = offsetof(struct cfq_group, stats.dequeue),
2143 .seq_show = cfqg_print_stat,
2146 .name = "unaccounted_time",
2147 .private = offsetof(struct cfq_group, stats.unaccounted_time),
2148 .seq_show = cfqg_print_stat,
2150 #endif /* CONFIG_DEBUG_BLK_CGROUP */
2154 static int cfq_print_weight_on_dfl(struct seq_file *sf, void *v)
2156 struct blkcg *blkcg = css_to_blkcg(seq_css(sf));
2157 struct cfq_group_data *cgd = blkcg_to_cfqgd(blkcg);
2159 seq_printf(sf, "default %u\n", cgd->weight);
2160 blkcg_print_blkgs(sf, blkcg, cfqg_prfill_weight_device,
2161 &blkcg_policy_cfq, 0, false);
2165 static ssize_t cfq_set_weight_on_dfl(struct kernfs_open_file *of,
2166 char *buf, size_t nbytes, loff_t off)
2174 /* "WEIGHT" or "default WEIGHT" sets the default weight */
2175 v = simple_strtoull(buf, &endp, 0);
2176 if (*endp == '\0' || sscanf(buf, "default %llu", &v) == 1) {
2177 ret = __cfq_set_weight(of_css(of), v, true, false, false);
2178 return ret ?: nbytes;
2181 /* "MAJ:MIN WEIGHT" */
2182 return __cfqg_set_weight_device(of, buf, nbytes, off, true, false);
2185 static struct cftype cfq_blkcg_files[] = {
2188 .flags = CFTYPE_NOT_ON_ROOT,
2189 .seq_show = cfq_print_weight_on_dfl,
2190 .write = cfq_set_weight_on_dfl,
2195 #else /* GROUP_IOSCHED */
2196 static struct cfq_group *cfq_lookup_cfqg(struct cfq_data *cfqd,
2197 struct blkcg *blkcg)
2199 return cfqd->root_group;
2203 cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) {
2207 #endif /* GROUP_IOSCHED */
2210 * The cfqd->service_trees holds all pending cfq_queue's that have
2211 * requests waiting to be processed. It is sorted in the order that
2212 * we will service the queues.
2214 static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2217 struct rb_node **p, *parent;
2218 struct cfq_queue *__cfqq;
2220 struct cfq_rb_root *st;
2223 u64 now = ktime_get_ns();
2225 st = st_for(cfqq->cfqg, cfqq_class(cfqq), cfqq_type(cfqq));
2226 if (cfq_class_idle(cfqq)) {
2227 rb_key = CFQ_IDLE_DELAY;
2228 parent = rb_last(&st->rb);
2229 if (parent && parent != &cfqq->rb_node) {
2230 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
2231 rb_key += __cfqq->rb_key;
2234 } else if (!add_front) {
2236 * Get our rb key offset. Subtract any residual slice
2237 * value carried from last service. A negative resid
2238 * count indicates slice overrun, and this should position
2239 * the next service time further away in the tree.
2241 rb_key = cfq_slice_offset(cfqd, cfqq) + now;
2242 rb_key -= cfqq->slice_resid;
2243 cfqq->slice_resid = 0;
2245 rb_key = -NSEC_PER_SEC;
2246 __cfqq = cfq_rb_first(st);
2247 rb_key += __cfqq ? __cfqq->rb_key : now;
2250 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
2253 * same position, nothing more to do
2255 if (rb_key == cfqq->rb_key && cfqq->service_tree == st)
2258 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
2259 cfqq->service_tree = NULL;
2264 cfqq->service_tree = st;
2265 p = &st->rb.rb_node;
2268 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
2271 * sort by key, that represents service time.
2273 if (rb_key < __cfqq->rb_key)
2274 p = &parent->rb_left;
2276 p = &parent->rb_right;
2282 st->left = &cfqq->rb_node;
2284 cfqq->rb_key = rb_key;
2285 rb_link_node(&cfqq->rb_node, parent, p);
2286 rb_insert_color(&cfqq->rb_node, &st->rb);
2288 if (add_front || !new_cfqq)
2290 cfq_group_notify_queue_add(cfqd, cfqq->cfqg);
2293 static struct cfq_queue *
2294 cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
2295 sector_t sector, struct rb_node **ret_parent,
2296 struct rb_node ***rb_link)
2298 struct rb_node **p, *parent;
2299 struct cfq_queue *cfqq = NULL;
2307 cfqq = rb_entry(parent, struct cfq_queue, p_node);
2310 * Sort strictly based on sector. Smallest to the left,
2311 * largest to the right.
2313 if (sector > blk_rq_pos(cfqq->next_rq))
2314 n = &(*p)->rb_right;
2315 else if (sector < blk_rq_pos(cfqq->next_rq))
2323 *ret_parent = parent;
2329 static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2331 struct rb_node **p, *parent;
2332 struct cfq_queue *__cfqq;
2335 rb_erase(&cfqq->p_node, cfqq->p_root);
2336 cfqq->p_root = NULL;
2339 if (cfq_class_idle(cfqq))
2344 cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
2345 __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
2346 blk_rq_pos(cfqq->next_rq), &parent, &p);
2348 rb_link_node(&cfqq->p_node, parent, p);
2349 rb_insert_color(&cfqq->p_node, cfqq->p_root);
2351 cfqq->p_root = NULL;
2355 * Update cfqq's position in the service tree.
2357 static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2360 * Resorting requires the cfqq to be on the RR list already.
2362 if (cfq_cfqq_on_rr(cfqq)) {
2363 cfq_service_tree_add(cfqd, cfqq, 0);
2364 cfq_prio_tree_add(cfqd, cfqq);
2369 * add to busy list of queues for service, trying to be fair in ordering
2370 * the pending list according to last request service
2372 static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2374 cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
2375 BUG_ON(cfq_cfqq_on_rr(cfqq));
2376 cfq_mark_cfqq_on_rr(cfqq);
2377 cfqd->busy_queues++;
2378 if (cfq_cfqq_sync(cfqq))
2379 cfqd->busy_sync_queues++;
2381 cfq_resort_rr_list(cfqd, cfqq);
2385 * Called when the cfqq no longer has requests pending, remove it from
2388 static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2390 cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
2391 BUG_ON(!cfq_cfqq_on_rr(cfqq));
2392 cfq_clear_cfqq_on_rr(cfqq);
2394 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
2395 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
2396 cfqq->service_tree = NULL;
2399 rb_erase(&cfqq->p_node, cfqq->p_root);
2400 cfqq->p_root = NULL;
2403 cfq_group_notify_queue_del(cfqd, cfqq->cfqg);
2404 BUG_ON(!cfqd->busy_queues);
2405 cfqd->busy_queues--;
2406 if (cfq_cfqq_sync(cfqq))
2407 cfqd->busy_sync_queues--;
2411 * rb tree support functions
2413 static void cfq_del_rq_rb(struct request *rq)
2415 struct cfq_queue *cfqq = RQ_CFQQ(rq);
2416 const int sync = rq_is_sync(rq);
2418 BUG_ON(!cfqq->queued[sync]);
2419 cfqq->queued[sync]--;
2421 elv_rb_del(&cfqq->sort_list, rq);
2423 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) {
2425 * Queue will be deleted from service tree when we actually
2426 * expire it later. Right now just remove it from prio tree
2430 rb_erase(&cfqq->p_node, cfqq->p_root);
2431 cfqq->p_root = NULL;
2436 static void cfq_add_rq_rb(struct request *rq)
2438 struct cfq_queue *cfqq = RQ_CFQQ(rq);
2439 struct cfq_data *cfqd = cfqq->cfqd;
2440 struct request *prev;
2442 cfqq->queued[rq_is_sync(rq)]++;
2444 elv_rb_add(&cfqq->sort_list, rq);
2446 if (!cfq_cfqq_on_rr(cfqq))
2447 cfq_add_cfqq_rr(cfqd, cfqq);
2450 * check if this request is a better next-serve candidate
2452 prev = cfqq->next_rq;
2453 cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq, cfqd->last_position);
2456 * adjust priority tree position, if ->next_rq changes
2458 if (prev != cfqq->next_rq)
2459 cfq_prio_tree_add(cfqd, cfqq);
2461 BUG_ON(!cfqq->next_rq);
2464 static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
2466 elv_rb_del(&cfqq->sort_list, rq);
2467 cfqq->queued[rq_is_sync(rq)]--;
2468 cfqg_stats_update_io_remove(RQ_CFQG(rq), rq->cmd_flags);
2470 cfqg_stats_update_io_add(RQ_CFQG(rq), cfqq->cfqd->serving_group,
2474 static struct request *
2475 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
2477 struct task_struct *tsk = current;
2478 struct cfq_io_cq *cic;
2479 struct cfq_queue *cfqq;
2481 cic = cfq_cic_lookup(cfqd, tsk->io_context);
2485 cfqq = cic_to_cfqq(cic, op_is_sync(bio->bi_opf));
2487 return elv_rb_find(&cfqq->sort_list, bio_end_sector(bio));
2492 static void cfq_activate_request(struct request_queue *q, struct request *rq)
2494 struct cfq_data *cfqd = q->elevator->elevator_data;
2496 cfqd->rq_in_driver++;
2497 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
2498 cfqd->rq_in_driver);
2500 cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
2503 static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
2505 struct cfq_data *cfqd = q->elevator->elevator_data;
2507 WARN_ON(!cfqd->rq_in_driver);
2508 cfqd->rq_in_driver--;
2509 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
2510 cfqd->rq_in_driver);
2513 static void cfq_remove_request(struct request *rq)
2515 struct cfq_queue *cfqq = RQ_CFQQ(rq);
2517 if (cfqq->next_rq == rq)
2518 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
2520 list_del_init(&rq->queuelist);
2523 cfqq->cfqd->rq_queued--;
2524 cfqg_stats_update_io_remove(RQ_CFQG(rq), rq->cmd_flags);
2525 if (rq->cmd_flags & REQ_PRIO) {
2526 WARN_ON(!cfqq->prio_pending);
2527 cfqq->prio_pending--;
2531 static enum elv_merge cfq_merge(struct request_queue *q, struct request **req,
2534 struct cfq_data *cfqd = q->elevator->elevator_data;
2535 struct request *__rq;
2537 __rq = cfq_find_rq_fmerge(cfqd, bio);
2538 if (__rq && elv_bio_merge_ok(__rq, bio)) {
2540 return ELEVATOR_FRONT_MERGE;
2543 return ELEVATOR_NO_MERGE;
2546 static void cfq_merged_request(struct request_queue *q, struct request *req,
2547 enum elv_merge type)
2549 if (type == ELEVATOR_FRONT_MERGE) {
2550 struct cfq_queue *cfqq = RQ_CFQQ(req);
2552 cfq_reposition_rq_rb(cfqq, req);
2556 static void cfq_bio_merged(struct request_queue *q, struct request *req,
2559 cfqg_stats_update_io_merged(RQ_CFQG(req), bio->bi_opf);
2563 cfq_merged_requests(struct request_queue *q, struct request *rq,
2564 struct request *next)
2566 struct cfq_queue *cfqq = RQ_CFQQ(rq);
2567 struct cfq_data *cfqd = q->elevator->elevator_data;
2570 * reposition in fifo if next is older than rq
2572 if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
2573 next->fifo_time < rq->fifo_time &&
2574 cfqq == RQ_CFQQ(next)) {
2575 list_move(&rq->queuelist, &next->queuelist);
2576 rq->fifo_time = next->fifo_time;
2579 if (cfqq->next_rq == next)
2581 cfq_remove_request(next);
2582 cfqg_stats_update_io_merged(RQ_CFQG(rq), next->cmd_flags);
2584 cfqq = RQ_CFQQ(next);
2586 * all requests of this queue are merged to other queues, delete it
2587 * from the service tree. If it's the active_queue,
2588 * cfq_dispatch_requests() will choose to expire it or do idle
2590 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list) &&
2591 cfqq != cfqd->active_queue)
2592 cfq_del_cfqq_rr(cfqd, cfqq);
2595 static int cfq_allow_bio_merge(struct request_queue *q, struct request *rq,
2598 struct cfq_data *cfqd = q->elevator->elevator_data;
2599 bool is_sync = op_is_sync(bio->bi_opf);
2600 struct cfq_io_cq *cic;
2601 struct cfq_queue *cfqq;
2604 * Disallow merge of a sync bio into an async request.
2606 if (is_sync && !rq_is_sync(rq))
2610 * Lookup the cfqq that this bio will be queued with and allow
2611 * merge only if rq is queued there.
2613 cic = cfq_cic_lookup(cfqd, current->io_context);
2617 cfqq = cic_to_cfqq(cic, is_sync);
2618 return cfqq == RQ_CFQQ(rq);
2621 static int cfq_allow_rq_merge(struct request_queue *q, struct request *rq,
2622 struct request *next)
2624 return RQ_CFQQ(rq) == RQ_CFQQ(next);
2627 static inline void cfq_del_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2629 hrtimer_try_to_cancel(&cfqd->idle_slice_timer);
2630 cfqg_stats_update_idle_time(cfqq->cfqg);
2633 static void __cfq_set_active_queue(struct cfq_data *cfqd,
2634 struct cfq_queue *cfqq)
2637 cfq_log_cfqq(cfqd, cfqq, "set_active wl_class:%d wl_type:%d",
2638 cfqd->serving_wl_class, cfqd->serving_wl_type);
2639 cfqg_stats_update_avg_queue_size(cfqq->cfqg);
2640 cfqq->slice_start = 0;
2641 cfqq->dispatch_start = ktime_get_ns();
2642 cfqq->allocated_slice = 0;
2643 cfqq->slice_end = 0;
2644 cfqq->slice_dispatch = 0;
2645 cfqq->nr_sectors = 0;
2647 cfq_clear_cfqq_wait_request(cfqq);
2648 cfq_clear_cfqq_must_dispatch(cfqq);
2649 cfq_clear_cfqq_must_alloc_slice(cfqq);
2650 cfq_clear_cfqq_fifo_expire(cfqq);
2651 cfq_mark_cfqq_slice_new(cfqq);
2653 cfq_del_timer(cfqd, cfqq);
2656 cfqd->active_queue = cfqq;
2660 * current cfqq expired its slice (or was too idle), select new one
2663 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2666 cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
2668 if (cfq_cfqq_wait_request(cfqq))
2669 cfq_del_timer(cfqd, cfqq);
2671 cfq_clear_cfqq_wait_request(cfqq);
2672 cfq_clear_cfqq_wait_busy(cfqq);
2675 * If this cfqq is shared between multiple processes, check to
2676 * make sure that those processes are still issuing I/Os within
2677 * the mean seek distance. If not, it may be time to break the
2678 * queues apart again.
2680 if (cfq_cfqq_coop(cfqq) && CFQQ_SEEKY(cfqq))
2681 cfq_mark_cfqq_split_coop(cfqq);
2684 * store what was left of this slice, if the queue idled/timed out
2687 if (cfq_cfqq_slice_new(cfqq))
2688 cfqq->slice_resid = cfq_scaled_cfqq_slice(cfqd, cfqq);
2690 cfqq->slice_resid = cfqq->slice_end - ktime_get_ns();
2691 cfq_log_cfqq(cfqd, cfqq, "resid=%lld", cfqq->slice_resid);
2694 cfq_group_served(cfqd, cfqq->cfqg, cfqq);
2696 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
2697 cfq_del_cfqq_rr(cfqd, cfqq);
2699 cfq_resort_rr_list(cfqd, cfqq);
2701 if (cfqq == cfqd->active_queue)
2702 cfqd->active_queue = NULL;
2704 if (cfqd->active_cic) {
2705 put_io_context(cfqd->active_cic->icq.ioc);
2706 cfqd->active_cic = NULL;
2710 static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out)
2712 struct cfq_queue *cfqq = cfqd->active_queue;
2715 __cfq_slice_expired(cfqd, cfqq, timed_out);
2719 * Get next queue for service. Unless we have a queue preemption,
2720 * we'll simply select the first cfqq in the service tree.
2722 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
2724 struct cfq_rb_root *st = st_for(cfqd->serving_group,
2725 cfqd->serving_wl_class, cfqd->serving_wl_type);
2727 if (!cfqd->rq_queued)
2730 /* There is nothing to dispatch */
2733 if (RB_EMPTY_ROOT(&st->rb))
2735 return cfq_rb_first(st);
2738 static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd)
2740 struct cfq_group *cfqg;
2741 struct cfq_queue *cfqq;
2743 struct cfq_rb_root *st;
2745 if (!cfqd->rq_queued)
2748 cfqg = cfq_get_next_cfqg(cfqd);
2752 for_each_cfqg_st(cfqg, i, j, st) {
2753 cfqq = cfq_rb_first(st);
2761 * Get and set a new active queue for service.
2763 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
2764 struct cfq_queue *cfqq)
2767 cfqq = cfq_get_next_queue(cfqd);
2769 __cfq_set_active_queue(cfqd, cfqq);
2773 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
2776 if (blk_rq_pos(rq) >= cfqd->last_position)
2777 return blk_rq_pos(rq) - cfqd->last_position;
2779 return cfqd->last_position - blk_rq_pos(rq);
2782 static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2785 return cfq_dist_from_last(cfqd, rq) <= CFQQ_CLOSE_THR;
2788 static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
2789 struct cfq_queue *cur_cfqq)
2791 struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
2792 struct rb_node *parent, *node;
2793 struct cfq_queue *__cfqq;
2794 sector_t sector = cfqd->last_position;
2796 if (RB_EMPTY_ROOT(root))
2800 * First, if we find a request starting at the end of the last
2801 * request, choose it.
2803 __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
2808 * If the exact sector wasn't found, the parent of the NULL leaf
2809 * will contain the closest sector.
2811 __cfqq = rb_entry(parent, struct cfq_queue, p_node);
2812 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
2815 if (blk_rq_pos(__cfqq->next_rq) < sector)
2816 node = rb_next(&__cfqq->p_node);
2818 node = rb_prev(&__cfqq->p_node);
2822 __cfqq = rb_entry(node, struct cfq_queue, p_node);
2823 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
2831 * cur_cfqq - passed in so that we don't decide that the current queue is
2832 * closely cooperating with itself.
2834 * So, basically we're assuming that that cur_cfqq has dispatched at least
2835 * one request, and that cfqd->last_position reflects a position on the disk
2836 * associated with the I/O issued by cur_cfqq. I'm not sure this is a valid
2839 static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
2840 struct cfq_queue *cur_cfqq)
2842 struct cfq_queue *cfqq;
2844 if (cfq_class_idle(cur_cfqq))
2846 if (!cfq_cfqq_sync(cur_cfqq))
2848 if (CFQQ_SEEKY(cur_cfqq))
2852 * Don't search priority tree if it's the only queue in the group.
2854 if (cur_cfqq->cfqg->nr_cfqq == 1)
2858 * We should notice if some of the queues are cooperating, eg
2859 * working closely on the same area of the disk. In that case,
2860 * we can group them together and don't waste time idling.
2862 cfqq = cfqq_close(cfqd, cur_cfqq);
2866 /* If new queue belongs to different cfq_group, don't choose it */
2867 if (cur_cfqq->cfqg != cfqq->cfqg)
2871 * It only makes sense to merge sync queues.
2873 if (!cfq_cfqq_sync(cfqq))
2875 if (CFQQ_SEEKY(cfqq))
2879 * Do not merge queues of different priority classes
2881 if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq))
2888 * Determine whether we should enforce idle window for this queue.
2891 static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2893 enum wl_class_t wl_class = cfqq_class(cfqq);
2894 struct cfq_rb_root *st = cfqq->service_tree;
2899 if (!cfqd->cfq_slice_idle)
2902 /* We never do for idle class queues. */
2903 if (wl_class == IDLE_WORKLOAD)
2906 /* We do for queues that were marked with idle window flag. */
2907 if (cfq_cfqq_idle_window(cfqq) &&
2908 !(blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag))
2912 * Otherwise, we do only if they are the last ones
2913 * in their service tree.
2915 if (st->count == 1 && cfq_cfqq_sync(cfqq) &&
2916 !cfq_io_thinktime_big(cfqd, &st->ttime, false))
2918 cfq_log_cfqq(cfqd, cfqq, "Not idling. st->count:%d", st->count);
2922 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
2924 struct cfq_queue *cfqq = cfqd->active_queue;
2925 struct cfq_rb_root *st = cfqq->service_tree;
2926 struct cfq_io_cq *cic;
2927 u64 sl, group_idle = 0;
2928 u64 now = ktime_get_ns();
2931 * SSD device without seek penalty, disable idling. But only do so
2932 * for devices that support queuing, otherwise we still have a problem
2933 * with sync vs async workloads.
2935 if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
2938 WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
2939 WARN_ON(cfq_cfqq_slice_new(cfqq));
2942 * idle is disabled, either manually or by past process history
2944 if (!cfq_should_idle(cfqd, cfqq)) {
2945 /* no queue idling. Check for group idling */
2946 if (cfqd->cfq_group_idle)
2947 group_idle = cfqd->cfq_group_idle;
2953 * still active requests from this queue, don't idle
2955 if (cfqq->dispatched)
2959 * task has exited, don't wait
2961 cic = cfqd->active_cic;
2962 if (!cic || !atomic_read(&cic->icq.ioc->active_ref))
2966 * If our average think time is larger than the remaining time
2967 * slice, then don't idle. This avoids overrunning the allotted
2970 if (sample_valid(cic->ttime.ttime_samples) &&
2971 (cfqq->slice_end - now < cic->ttime.ttime_mean)) {
2972 cfq_log_cfqq(cfqd, cfqq, "Not idling. think_time:%llu",
2973 cic->ttime.ttime_mean);
2978 * There are other queues in the group or this is the only group and
2979 * it has too big thinktime, don't do group idle.
2982 (cfqq->cfqg->nr_cfqq > 1 ||
2983 cfq_io_thinktime_big(cfqd, &st->ttime, true)))
2986 cfq_mark_cfqq_wait_request(cfqq);
2989 sl = cfqd->cfq_group_idle;
2991 sl = cfqd->cfq_slice_idle;
2993 hrtimer_start(&cfqd->idle_slice_timer, ns_to_ktime(sl),
2995 cfqg_stats_set_start_idle_time(cfqq->cfqg);
2996 cfq_log_cfqq(cfqd, cfqq, "arm_idle: %llu group_idle: %d", sl,
2997 group_idle ? 1 : 0);
3001 * Move request from internal lists to the request queue dispatch list.
3003 static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
3005 struct cfq_data *cfqd = q->elevator->elevator_data;
3006 struct cfq_queue *cfqq = RQ_CFQQ(rq);
3008 cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
3010 cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
3011 cfq_remove_request(rq);
3013 (RQ_CFQG(rq))->dispatched++;
3014 elv_dispatch_sort(q, rq);
3016 cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]++;
3017 cfqq->nr_sectors += blk_rq_sectors(rq);
3021 * return expired entry, or NULL to just start from scratch in rbtree
3023 static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
3025 struct request *rq = NULL;
3027 if (cfq_cfqq_fifo_expire(cfqq))
3030 cfq_mark_cfqq_fifo_expire(cfqq);
3032 if (list_empty(&cfqq->fifo))
3035 rq = rq_entry_fifo(cfqq->fifo.next);
3036 if (ktime_get_ns() < rq->fifo_time)
3043 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3045 const int base_rq = cfqd->cfq_slice_async_rq;
3047 WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
3049 return 2 * base_rq * (IOPRIO_BE_NR - cfqq->ioprio);
3053 * Must be called with the queue_lock held.
3055 static int cfqq_process_refs(struct cfq_queue *cfqq)
3057 int process_refs, io_refs;
3059 io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE];
3060 process_refs = cfqq->ref - io_refs;
3061 BUG_ON(process_refs < 0);
3062 return process_refs;
3065 static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq)
3067 int process_refs, new_process_refs;
3068 struct cfq_queue *__cfqq;
3071 * If there are no process references on the new_cfqq, then it is
3072 * unsafe to follow the ->new_cfqq chain as other cfqq's in the
3073 * chain may have dropped their last reference (not just their
3074 * last process reference).
3076 if (!cfqq_process_refs(new_cfqq))
3079 /* Avoid a circular list and skip interim queue merges */
3080 while ((__cfqq = new_cfqq->new_cfqq)) {
3086 process_refs = cfqq_process_refs(cfqq);
3087 new_process_refs = cfqq_process_refs(new_cfqq);
3089 * If the process for the cfqq has gone away, there is no
3090 * sense in merging the queues.
3092 if (process_refs == 0 || new_process_refs == 0)
3096 * Merge in the direction of the lesser amount of work.
3098 if (new_process_refs >= process_refs) {
3099 cfqq->new_cfqq = new_cfqq;
3100 new_cfqq->ref += process_refs;
3102 new_cfqq->new_cfqq = cfqq;
3103 cfqq->ref += new_process_refs;
3107 static enum wl_type_t cfq_choose_wl_type(struct cfq_data *cfqd,
3108 struct cfq_group *cfqg, enum wl_class_t wl_class)
3110 struct cfq_queue *queue;
3112 bool key_valid = false;
3114 enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD;
3116 for (i = 0; i <= SYNC_WORKLOAD; ++i) {
3117 /* select the one with lowest rb_key */
3118 queue = cfq_rb_first(st_for(cfqg, wl_class, i));
3120 (!key_valid || queue->rb_key < lowest_key)) {
3121 lowest_key = queue->rb_key;
3131 choose_wl_class_and_type(struct cfq_data *cfqd, struct cfq_group *cfqg)
3135 struct cfq_rb_root *st;
3137 enum wl_class_t original_class = cfqd->serving_wl_class;
3138 u64 now = ktime_get_ns();
3140 /* Choose next priority. RT > BE > IDLE */
3141 if (cfq_group_busy_queues_wl(RT_WORKLOAD, cfqd, cfqg))
3142 cfqd->serving_wl_class = RT_WORKLOAD;
3143 else if (cfq_group_busy_queues_wl(BE_WORKLOAD, cfqd, cfqg))
3144 cfqd->serving_wl_class = BE_WORKLOAD;
3146 cfqd->serving_wl_class = IDLE_WORKLOAD;
3147 cfqd->workload_expires = now + jiffies_to_nsecs(1);
3151 if (original_class != cfqd->serving_wl_class)
3155 * For RT and BE, we have to choose also the type
3156 * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
3159 st = st_for(cfqg, cfqd->serving_wl_class, cfqd->serving_wl_type);
3163 * check workload expiration, and that we still have other queues ready
3165 if (count && !(now > cfqd->workload_expires))
3169 /* otherwise select new workload type */
3170 cfqd->serving_wl_type = cfq_choose_wl_type(cfqd, cfqg,
3171 cfqd->serving_wl_class);
3172 st = st_for(cfqg, cfqd->serving_wl_class, cfqd->serving_wl_type);
3176 * the workload slice is computed as a fraction of target latency
3177 * proportional to the number of queues in that workload, over
3178 * all the queues in the same priority class
3180 group_slice = cfq_group_slice(cfqd, cfqg);
3182 slice = div_u64(group_slice * count,
3183 max_t(unsigned, cfqg->busy_queues_avg[cfqd->serving_wl_class],
3184 cfq_group_busy_queues_wl(cfqd->serving_wl_class, cfqd,
3187 if (cfqd->serving_wl_type == ASYNC_WORKLOAD) {
3191 * Async queues are currently system wide. Just taking
3192 * proportion of queues with-in same group will lead to higher
3193 * async ratio system wide as generally root group is going
3194 * to have higher weight. A more accurate thing would be to
3195 * calculate system wide asnc/sync ratio.
3197 tmp = cfqd->cfq_target_latency *
3198 cfqg_busy_async_queues(cfqd, cfqg);
3199 tmp = div_u64(tmp, cfqd->busy_queues);
3200 slice = min_t(u64, slice, tmp);
3202 /* async workload slice is scaled down according to
3203 * the sync/async slice ratio. */
3204 slice = div64_u64(slice*cfqd->cfq_slice[0], cfqd->cfq_slice[1]);
3206 /* sync workload slice is at least 2 * cfq_slice_idle */
3207 slice = max(slice, 2 * cfqd->cfq_slice_idle);
3209 slice = max_t(u64, slice, CFQ_MIN_TT);
3210 cfq_log(cfqd, "workload slice:%llu", slice);
3211 cfqd->workload_expires = now + slice;
3214 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd)
3216 struct cfq_rb_root *st = &cfqd->grp_service_tree;
3217 struct cfq_group *cfqg;
3219 if (RB_EMPTY_ROOT(&st->rb))
3221 cfqg = cfq_rb_first_group(st);
3222 update_min_vdisktime(st);
3226 static void cfq_choose_cfqg(struct cfq_data *cfqd)
3228 struct cfq_group *cfqg = cfq_get_next_cfqg(cfqd);
3229 u64 now = ktime_get_ns();
3231 cfqd->serving_group = cfqg;
3233 /* Restore the workload type data */
3234 if (cfqg->saved_wl_slice) {
3235 cfqd->workload_expires = now + cfqg->saved_wl_slice;
3236 cfqd->serving_wl_type = cfqg->saved_wl_type;
3237 cfqd->serving_wl_class = cfqg->saved_wl_class;
3239 cfqd->workload_expires = now - 1;
3241 choose_wl_class_and_type(cfqd, cfqg);
3245 * Select a queue for service. If we have a current active queue,
3246 * check whether to continue servicing it, or retrieve and set a new one.
3248 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
3250 struct cfq_queue *cfqq, *new_cfqq = NULL;
3251 u64 now = ktime_get_ns();
3253 cfqq = cfqd->active_queue;
3257 if (!cfqd->rq_queued)
3261 * We were waiting for group to get backlogged. Expire the queue
3263 if (cfq_cfqq_wait_busy(cfqq) && !RB_EMPTY_ROOT(&cfqq->sort_list))
3267 * The active queue has run out of time, expire it and select new.
3269 if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq)) {
3271 * If slice had not expired at the completion of last request
3272 * we might not have turned on wait_busy flag. Don't expire
3273 * the queue yet. Allow the group to get backlogged.
3275 * The very fact that we have used the slice, that means we
3276 * have been idling all along on this queue and it should be
3277 * ok to wait for this request to complete.
3279 if (cfqq->cfqg->nr_cfqq == 1 && RB_EMPTY_ROOT(&cfqq->sort_list)
3280 && cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
3284 goto check_group_idle;
3288 * The active queue has requests and isn't expired, allow it to
3291 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3295 * If another queue has a request waiting within our mean seek
3296 * distance, let it run. The expire code will check for close
3297 * cooperators and put the close queue at the front of the service
3298 * tree. If possible, merge the expiring queue with the new cfqq.
3300 new_cfqq = cfq_close_cooperator(cfqd, cfqq);
3302 if (!cfqq->new_cfqq)
3303 cfq_setup_merge(cfqq, new_cfqq);
3308 * No requests pending. If the active queue still has requests in
3309 * flight or is idling for a new request, allow either of these
3310 * conditions to happen (or time out) before selecting a new queue.
3312 if (hrtimer_active(&cfqd->idle_slice_timer)) {
3318 * This is a deep seek queue, but the device is much faster than
3319 * the queue can deliver, don't idle
3321 if (CFQQ_SEEKY(cfqq) && cfq_cfqq_idle_window(cfqq) &&
3322 (cfq_cfqq_slice_new(cfqq) ||
3323 (cfqq->slice_end - now > now - cfqq->slice_start))) {
3324 cfq_clear_cfqq_deep(cfqq);
3325 cfq_clear_cfqq_idle_window(cfqq);
3328 if (cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
3334 * If group idle is enabled and there are requests dispatched from
3335 * this group, wait for requests to complete.
3338 if (cfqd->cfq_group_idle && cfqq->cfqg->nr_cfqq == 1 &&
3339 cfqq->cfqg->dispatched &&
3340 !cfq_io_thinktime_big(cfqd, &cfqq->cfqg->ttime, true)) {
3346 cfq_slice_expired(cfqd, 0);
3349 * Current queue expired. Check if we have to switch to a new
3353 cfq_choose_cfqg(cfqd);
3355 cfqq = cfq_set_active_queue(cfqd, new_cfqq);
3360 static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
3364 while (cfqq->next_rq) {
3365 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
3369 BUG_ON(!list_empty(&cfqq->fifo));
3371 /* By default cfqq is not expired if it is empty. Do it explicitly */
3372 __cfq_slice_expired(cfqq->cfqd, cfqq, 0);
3377 * Drain our current requests. Used for barriers and when switching
3378 * io schedulers on-the-fly.
3380 static int cfq_forced_dispatch(struct cfq_data *cfqd)
3382 struct cfq_queue *cfqq;
3385 /* Expire the timeslice of the current active queue first */
3386 cfq_slice_expired(cfqd, 0);
3387 while ((cfqq = cfq_get_next_queue_forced(cfqd)) != NULL) {
3388 __cfq_set_active_queue(cfqd, cfqq);
3389 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
3392 BUG_ON(cfqd->busy_queues);
3394 cfq_log(cfqd, "forced_dispatch=%d", dispatched);
3398 static inline bool cfq_slice_used_soon(struct cfq_data *cfqd,
3399 struct cfq_queue *cfqq)
3401 u64 now = ktime_get_ns();
3403 /* the queue hasn't finished any request, can't estimate */
3404 if (cfq_cfqq_slice_new(cfqq))
3406 if (now + cfqd->cfq_slice_idle * cfqq->dispatched > cfqq->slice_end)
3412 static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3414 unsigned int max_dispatch;
3416 if (cfq_cfqq_must_dispatch(cfqq))
3420 * Drain async requests before we start sync IO
3422 if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_flight[BLK_RW_ASYNC])
3426 * If this is an async queue and we have sync IO in flight, let it wait
3428 if (cfqd->rq_in_flight[BLK_RW_SYNC] && !cfq_cfqq_sync(cfqq))
3431 max_dispatch = max_t(unsigned int, cfqd->cfq_quantum / 2, 1);
3432 if (cfq_class_idle(cfqq))
3436 * Does this cfqq already have too much IO in flight?
3438 if (cfqq->dispatched >= max_dispatch) {
3439 bool promote_sync = false;
3441 * idle queue must always only have a single IO in flight
3443 if (cfq_class_idle(cfqq))
3447 * If there is only one sync queue
3448 * we can ignore async queue here and give the sync
3449 * queue no dispatch limit. The reason is a sync queue can
3450 * preempt async queue, limiting the sync queue doesn't make
3451 * sense. This is useful for aiostress test.
3453 if (cfq_cfqq_sync(cfqq) && cfqd->busy_sync_queues == 1)
3454 promote_sync = true;
3457 * We have other queues, don't allow more IO from this one
3459 if (cfqd->busy_queues > 1 && cfq_slice_used_soon(cfqd, cfqq) &&
3464 * Sole queue user, no limit
3466 if (cfqd->busy_queues == 1 || promote_sync)
3470 * Normally we start throttling cfqq when cfq_quantum/2
3471 * requests have been dispatched. But we can drive
3472 * deeper queue depths at the beginning of slice
3473 * subjected to upper limit of cfq_quantum.
3475 max_dispatch = cfqd->cfq_quantum;
3479 * Async queues must wait a bit before being allowed dispatch.
3480 * We also ramp up the dispatch depth gradually for async IO,
3481 * based on the last sync IO we serviced
3483 if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
3484 u64 last_sync = ktime_get_ns() - cfqd->last_delayed_sync;
3487 depth = div64_u64(last_sync, cfqd->cfq_slice[1]);
3488 if (!depth && !cfqq->dispatched)
3490 if (depth < max_dispatch)
3491 max_dispatch = depth;
3495 * If we're below the current max, allow a dispatch
3497 return cfqq->dispatched < max_dispatch;
3501 * Dispatch a request from cfqq, moving them to the request queue
3504 static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3508 BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
3510 rq = cfq_check_fifo(cfqq);
3512 cfq_mark_cfqq_must_dispatch(cfqq);
3514 if (!cfq_may_dispatch(cfqd, cfqq))
3518 * follow expired path, else get first next available
3523 cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq);
3526 * insert request into driver dispatch list
3528 cfq_dispatch_insert(cfqd->queue, rq);
3530 if (!cfqd->active_cic) {
3531 struct cfq_io_cq *cic = RQ_CIC(rq);
3533 atomic_long_inc(&cic->icq.ioc->refcount);
3534 cfqd->active_cic = cic;
3541 * Find the cfqq that we need to service and move a request from that to the
3544 static int cfq_dispatch_requests(struct request_queue *q, int force)
3546 struct cfq_data *cfqd = q->elevator->elevator_data;
3547 struct cfq_queue *cfqq;
3549 if (!cfqd->busy_queues)
3552 if (unlikely(force))
3553 return cfq_forced_dispatch(cfqd);
3555 cfqq = cfq_select_queue(cfqd);
3560 * Dispatch a request from this cfqq, if it is allowed
3562 if (!cfq_dispatch_request(cfqd, cfqq))
3565 cfqq->slice_dispatch++;
3566 cfq_clear_cfqq_must_dispatch(cfqq);
3569 * expire an async queue immediately if it has used up its slice. idle
3570 * queue always expire after 1 dispatch round.
3572 if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
3573 cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
3574 cfq_class_idle(cfqq))) {
3575 cfqq->slice_end = ktime_get_ns() + 1;
3576 cfq_slice_expired(cfqd, 0);
3579 cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
3584 * task holds one reference to the queue, dropped when task exits. each rq
3585 * in-flight on this queue also holds a reference, dropped when rq is freed.
3587 * Each cfq queue took a reference on the parent group. Drop it now.
3588 * queue lock must be held here.
3590 static void cfq_put_queue(struct cfq_queue *cfqq)
3592 struct cfq_data *cfqd = cfqq->cfqd;
3593 struct cfq_group *cfqg;
3595 BUG_ON(cfqq->ref <= 0);
3601 cfq_log_cfqq(cfqd, cfqq, "put_queue");
3602 BUG_ON(rb_first(&cfqq->sort_list));
3603 BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
3606 if (unlikely(cfqd->active_queue == cfqq)) {
3607 __cfq_slice_expired(cfqd, cfqq, 0);
3608 cfq_schedule_dispatch(cfqd);
3611 BUG_ON(cfq_cfqq_on_rr(cfqq));
3612 kmem_cache_free(cfq_pool, cfqq);
3616 static void cfq_put_cooperator(struct cfq_queue *cfqq)
3618 struct cfq_queue *__cfqq, *next;
3621 * If this queue was scheduled to merge with another queue, be
3622 * sure to drop the reference taken on that queue (and others in
3623 * the merge chain). See cfq_setup_merge and cfq_merge_cfqqs.
3625 __cfqq = cfqq->new_cfqq;
3627 if (__cfqq == cfqq) {
3628 WARN(1, "cfqq->new_cfqq loop detected\n");
3631 next = __cfqq->new_cfqq;
3632 cfq_put_queue(__cfqq);
3637 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3639 if (unlikely(cfqq == cfqd->active_queue)) {
3640 __cfq_slice_expired(cfqd, cfqq, 0);
3641 cfq_schedule_dispatch(cfqd);
3644 cfq_put_cooperator(cfqq);
3646 cfq_put_queue(cfqq);
3649 static void cfq_init_icq(struct io_cq *icq)
3651 struct cfq_io_cq *cic = icq_to_cic(icq);
3653 cic->ttime.last_end_request = ktime_get_ns();
3656 static void cfq_exit_icq(struct io_cq *icq)
3658 struct cfq_io_cq *cic = icq_to_cic(icq);
3659 struct cfq_data *cfqd = cic_to_cfqd(cic);
3661 if (cic_to_cfqq(cic, false)) {
3662 cfq_exit_cfqq(cfqd, cic_to_cfqq(cic, false));
3663 cic_set_cfqq(cic, NULL, false);
3666 if (cic_to_cfqq(cic, true)) {
3667 cfq_exit_cfqq(cfqd, cic_to_cfqq(cic, true));
3668 cic_set_cfqq(cic, NULL, true);
3672 static void cfq_init_prio_data(struct cfq_queue *cfqq, struct cfq_io_cq *cic)
3674 struct task_struct *tsk = current;
3677 if (!cfq_cfqq_prio_changed(cfqq))
3680 ioprio_class = IOPRIO_PRIO_CLASS(cic->ioprio);
3681 switch (ioprio_class) {
3683 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
3684 case IOPRIO_CLASS_NONE:
3686 * no prio set, inherit CPU scheduling settings
3688 cfqq->ioprio = task_nice_ioprio(tsk);
3689 cfqq->ioprio_class = task_nice_ioclass(tsk);
3691 case IOPRIO_CLASS_RT:
3692 cfqq->ioprio = IOPRIO_PRIO_DATA(cic->ioprio);
3693 cfqq->ioprio_class = IOPRIO_CLASS_RT;
3695 case IOPRIO_CLASS_BE:
3696 cfqq->ioprio = IOPRIO_PRIO_DATA(cic->ioprio);
3697 cfqq->ioprio_class = IOPRIO_CLASS_BE;
3699 case IOPRIO_CLASS_IDLE:
3700 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
3702 cfq_clear_cfqq_idle_window(cfqq);
3707 * keep track of original prio settings in case we have to temporarily
3708 * elevate the priority of this queue
3710 cfqq->org_ioprio = cfqq->ioprio;
3711 cfqq->org_ioprio_class = cfqq->ioprio_class;
3712 cfq_clear_cfqq_prio_changed(cfqq);
3715 static void check_ioprio_changed(struct cfq_io_cq *cic, struct bio *bio)
3717 int ioprio = cic->icq.ioc->ioprio;
3718 struct cfq_data *cfqd = cic_to_cfqd(cic);
3719 struct cfq_queue *cfqq;
3722 * Check whether ioprio has changed. The condition may trigger
3723 * spuriously on a newly created cic but there's no harm.
3725 if (unlikely(!cfqd) || likely(cic->ioprio == ioprio))
3728 cfqq = cic_to_cfqq(cic, false);
3730 cfq_put_queue(cfqq);
3731 cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic, bio);
3732 cic_set_cfqq(cic, cfqq, false);
3735 cfqq = cic_to_cfqq(cic, true);
3737 cfq_mark_cfqq_prio_changed(cfqq);
3739 cic->ioprio = ioprio;
3742 static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3743 pid_t pid, bool is_sync)
3745 RB_CLEAR_NODE(&cfqq->rb_node);
3746 RB_CLEAR_NODE(&cfqq->p_node);
3747 INIT_LIST_HEAD(&cfqq->fifo);
3752 cfq_mark_cfqq_prio_changed(cfqq);
3755 if (!cfq_class_idle(cfqq))
3756 cfq_mark_cfqq_idle_window(cfqq);
3757 cfq_mark_cfqq_sync(cfqq);
3762 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3763 static bool check_blkcg_changed(struct cfq_io_cq *cic, struct bio *bio)
3765 struct cfq_data *cfqd = cic_to_cfqd(cic);
3766 struct cfq_queue *cfqq;
3771 serial_nr = bio_blkcg(bio)->css.serial_nr;
3772 nonroot_cg = bio_blkcg(bio) != &blkcg_root;
3776 * Check whether blkcg has changed. The condition may trigger
3777 * spuriously on a newly created cic but there's no harm.
3779 if (unlikely(!cfqd) || likely(cic->blkcg_serial_nr == serial_nr))
3783 * Drop reference to queues. New queues will be assigned in new
3784 * group upon arrival of fresh requests.
3786 cfqq = cic_to_cfqq(cic, false);
3788 cfq_log_cfqq(cfqd, cfqq, "changed cgroup");
3789 cic_set_cfqq(cic, NULL, false);
3790 cfq_put_queue(cfqq);
3793 cfqq = cic_to_cfqq(cic, true);
3795 cfq_log_cfqq(cfqd, cfqq, "changed cgroup");
3796 cic_set_cfqq(cic, NULL, true);
3797 cfq_put_queue(cfqq);
3800 cic->blkcg_serial_nr = serial_nr;
3804 static inline bool check_blkcg_changed(struct cfq_io_cq *cic, struct bio *bio)
3808 #endif /* CONFIG_CFQ_GROUP_IOSCHED */
3810 static struct cfq_queue **
3811 cfq_async_queue_prio(struct cfq_group *cfqg, int ioprio_class, int ioprio)
3813 switch (ioprio_class) {
3814 case IOPRIO_CLASS_RT:
3815 return &cfqg->async_cfqq[0][ioprio];
3816 case IOPRIO_CLASS_NONE:
3817 ioprio = IOPRIO_NORM;
3819 case IOPRIO_CLASS_BE:
3820 return &cfqg->async_cfqq[1][ioprio];
3821 case IOPRIO_CLASS_IDLE:
3822 return &cfqg->async_idle_cfqq;
3828 static struct cfq_queue *
3829 cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct cfq_io_cq *cic,
3832 int ioprio_class = IOPRIO_PRIO_CLASS(cic->ioprio);
3833 int ioprio = IOPRIO_PRIO_DATA(cic->ioprio);
3834 struct cfq_queue **async_cfqq = NULL;
3835 struct cfq_queue *cfqq;
3836 struct cfq_group *cfqg;
3839 cfqg = cfq_lookup_cfqg(cfqd, bio_blkcg(bio));
3841 cfqq = &cfqd->oom_cfqq;
3846 if (!ioprio_valid(cic->ioprio)) {
3847 struct task_struct *tsk = current;
3848 ioprio = task_nice_ioprio(tsk);
3849 ioprio_class = task_nice_ioclass(tsk);
3851 async_cfqq = cfq_async_queue_prio(cfqg, ioprio_class, ioprio);
3857 cfqq = kmem_cache_alloc_node(cfq_pool,
3858 GFP_NOWAIT | __GFP_ZERO | __GFP_NOWARN,
3861 cfqq = &cfqd->oom_cfqq;
3865 /* cfq_init_cfqq() assumes cfqq->ioprio_class is initialized. */
3866 cfqq->ioprio_class = IOPRIO_CLASS_NONE;
3867 cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
3868 cfq_init_prio_data(cfqq, cic);
3869 cfq_link_cfqq_cfqg(cfqq, cfqg);
3870 cfq_log_cfqq(cfqd, cfqq, "alloced");
3873 /* a new async queue is created, pin and remember */
3884 __cfq_update_io_thinktime(struct cfq_ttime *ttime, u64 slice_idle)
3886 u64 elapsed = ktime_get_ns() - ttime->last_end_request;
3887 elapsed = min(elapsed, 2UL * slice_idle);
3889 ttime->ttime_samples = (7*ttime->ttime_samples + 256) / 8;
3890 ttime->ttime_total = div_u64(7*ttime->ttime_total + 256*elapsed, 8);
3891 ttime->ttime_mean = div64_ul(ttime->ttime_total + 128,
3892 ttime->ttime_samples);
3896 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3897 struct cfq_io_cq *cic)
3899 if (cfq_cfqq_sync(cfqq)) {
3900 __cfq_update_io_thinktime(&cic->ttime, cfqd->cfq_slice_idle);
3901 __cfq_update_io_thinktime(&cfqq->service_tree->ttime,
3902 cfqd->cfq_slice_idle);
3904 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3905 __cfq_update_io_thinktime(&cfqq->cfqg->ttime, cfqd->cfq_group_idle);
3910 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3914 sector_t n_sec = blk_rq_sectors(rq);
3915 if (cfqq->last_request_pos) {
3916 if (cfqq->last_request_pos < blk_rq_pos(rq))
3917 sdist = blk_rq_pos(rq) - cfqq->last_request_pos;
3919 sdist = cfqq->last_request_pos - blk_rq_pos(rq);
3922 cfqq->seek_history <<= 1;
3923 if (blk_queue_nonrot(cfqd->queue))
3924 cfqq->seek_history |= (n_sec < CFQQ_SECT_THR_NONROT);
3926 cfqq->seek_history |= (sdist > CFQQ_SEEK_THR);
3929 static inline bool req_noidle(struct request *req)
3931 return req_op(req) == REQ_OP_WRITE &&
3932 (req->cmd_flags & (REQ_SYNC | REQ_IDLE)) == REQ_SYNC;
3936 * Disable idle window if the process thinks too long or seeks so much that
3940 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3941 struct cfq_io_cq *cic)
3943 int old_idle, enable_idle;
3946 * Don't idle for async or idle io prio class
3948 if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
3951 enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
3953 if (cfqq->queued[0] + cfqq->queued[1] >= 4)
3954 cfq_mark_cfqq_deep(cfqq);
3956 if (cfqq->next_rq && req_noidle(cfqq->next_rq))
3958 else if (!atomic_read(&cic->icq.ioc->active_ref) ||
3959 !cfqd->cfq_slice_idle ||
3960 (!cfq_cfqq_deep(cfqq) && CFQQ_SEEKY(cfqq)))
3962 else if (sample_valid(cic->ttime.ttime_samples)) {
3963 if (cic->ttime.ttime_mean > cfqd->cfq_slice_idle)
3969 if (old_idle != enable_idle) {
3970 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
3972 cfq_mark_cfqq_idle_window(cfqq);
3974 cfq_clear_cfqq_idle_window(cfqq);
3979 * Check if new_cfqq should preempt the currently active queue. Return 0 for
3980 * no or if we aren't sure, a 1 will cause a preempt.
3983 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
3986 struct cfq_queue *cfqq;
3988 cfqq = cfqd->active_queue;
3992 if (cfq_class_idle(new_cfqq))
3995 if (cfq_class_idle(cfqq))
3999 * Don't allow a non-RT request to preempt an ongoing RT cfqq timeslice.
4001 if (cfq_class_rt(cfqq) && !cfq_class_rt(new_cfqq))
4005 * if the new request is sync, but the currently running queue is
4006 * not, let the sync request have priority.
4008 if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq) && !cfq_cfqq_must_dispatch(cfqq))
4012 * Treat ancestors of current cgroup the same way as current cgroup.
4013 * For anybody else we disallow preemption to guarantee service
4014 * fairness among cgroups.
4016 if (!cfqg_is_descendant(cfqq->cfqg, new_cfqq->cfqg))
4019 if (cfq_slice_used(cfqq))
4023 * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
4025 if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
4028 WARN_ON_ONCE(cfqq->ioprio_class != new_cfqq->ioprio_class);
4029 /* Allow preemption only if we are idling on sync-noidle tree */
4030 if (cfqd->serving_wl_type == SYNC_NOIDLE_WORKLOAD &&
4031 cfqq_type(new_cfqq) == SYNC_NOIDLE_WORKLOAD &&
4032 RB_EMPTY_ROOT(&cfqq->sort_list))
4036 * So both queues are sync. Let the new request get disk time if
4037 * it's a metadata request and the current queue is doing regular IO.
4039 if ((rq->cmd_flags & REQ_PRIO) && !cfqq->prio_pending)
4042 /* An idle queue should not be idle now for some reason */
4043 if (RB_EMPTY_ROOT(&cfqq->sort_list) && !cfq_should_idle(cfqd, cfqq))
4046 if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
4050 * if this request is as-good as one we would expect from the
4051 * current cfqq, let it preempt
4053 if (cfq_rq_close(cfqd, cfqq, rq))
4060 * cfqq preempts the active queue. if we allowed preempt with no slice left,
4061 * let it have half of its nominal slice.
4063 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
4065 enum wl_type_t old_type = cfqq_type(cfqd->active_queue);
4067 cfq_log_cfqq(cfqd, cfqq, "preempt");
4068 cfq_slice_expired(cfqd, 1);
4071 * workload type is changed, don't save slice, otherwise preempt
4074 if (old_type != cfqq_type(cfqq))
4075 cfqq->cfqg->saved_wl_slice = 0;
4078 * Put the new queue at the front of the of the current list,
4079 * so we know that it will be selected next.
4081 BUG_ON(!cfq_cfqq_on_rr(cfqq));
4083 cfq_service_tree_add(cfqd, cfqq, 1);
4085 cfqq->slice_end = 0;
4086 cfq_mark_cfqq_slice_new(cfqq);
4090 * Called when a new fs request (rq) is added (to cfqq). Check if there's
4091 * something we should do about it
4094 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
4097 struct cfq_io_cq *cic = RQ_CIC(rq);
4100 if (rq->cmd_flags & REQ_PRIO)
4101 cfqq->prio_pending++;
4103 cfq_update_io_thinktime(cfqd, cfqq, cic);
4104 cfq_update_io_seektime(cfqd, cfqq, rq);
4105 cfq_update_idle_window(cfqd, cfqq, cic);
4107 cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
4109 if (cfqq == cfqd->active_queue) {
4111 * Remember that we saw a request from this process, but
4112 * don't start queuing just yet. Otherwise we risk seeing lots
4113 * of tiny requests, because we disrupt the normal plugging
4114 * and merging. If the request is already larger than a single
4115 * page, let it rip immediately. For that case we assume that
4116 * merging is already done. Ditto for a busy system that
4117 * has other work pending, don't risk delaying until the
4118 * idle timer unplug to continue working.
4120 if (cfq_cfqq_wait_request(cfqq)) {
4121 if (blk_rq_bytes(rq) > PAGE_SIZE ||
4122 cfqd->busy_queues > 1) {
4123 cfq_del_timer(cfqd, cfqq);
4124 cfq_clear_cfqq_wait_request(cfqq);
4125 __blk_run_queue(cfqd->queue);
4127 cfqg_stats_update_idle_time(cfqq->cfqg);
4128 cfq_mark_cfqq_must_dispatch(cfqq);
4131 } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
4133 * not the active queue - expire current slice if it is
4134 * idle and has expired it's mean thinktime or this new queue
4135 * has some old slice time left and is of higher priority or
4136 * this new queue is RT and the current one is BE
4138 cfq_preempt_queue(cfqd, cfqq);
4139 __blk_run_queue(cfqd->queue);
4143 static void cfq_insert_request(struct request_queue *q, struct request *rq)
4145 struct cfq_data *cfqd = q->elevator->elevator_data;
4146 struct cfq_queue *cfqq = RQ_CFQQ(rq);
4148 cfq_log_cfqq(cfqd, cfqq, "insert_request");
4149 cfq_init_prio_data(cfqq, RQ_CIC(rq));
4151 rq->fifo_time = ktime_get_ns() + cfqd->cfq_fifo_expire[rq_is_sync(rq)];
4152 list_add_tail(&rq->queuelist, &cfqq->fifo);
4154 cfqg_stats_update_io_add(RQ_CFQG(rq), cfqd->serving_group,
4156 cfq_rq_enqueued(cfqd, cfqq, rq);
4160 * Update hw_tag based on peak queue depth over 50 samples under
4163 static void cfq_update_hw_tag(struct cfq_data *cfqd)
4165 struct cfq_queue *cfqq = cfqd->active_queue;
4167 if (cfqd->rq_in_driver > cfqd->hw_tag_est_depth)
4168 cfqd->hw_tag_est_depth = cfqd->rq_in_driver;
4170 if (cfqd->hw_tag == 1)
4173 if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
4174 cfqd->rq_in_driver <= CFQ_HW_QUEUE_MIN)
4178 * If active queue hasn't enough requests and can idle, cfq might not
4179 * dispatch sufficient requests to hardware. Don't zero hw_tag in this
4182 if (cfqq && cfq_cfqq_idle_window(cfqq) &&
4183 cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] <
4184 CFQ_HW_QUEUE_MIN && cfqd->rq_in_driver < CFQ_HW_QUEUE_MIN)
4187 if (cfqd->hw_tag_samples++ < 50)
4190 if (cfqd->hw_tag_est_depth >= CFQ_HW_QUEUE_MIN)
4196 static bool cfq_should_wait_busy(struct cfq_data *cfqd, struct cfq_queue *cfqq)
4198 struct cfq_io_cq *cic = cfqd->active_cic;
4199 u64 now = ktime_get_ns();
4201 /* If the queue already has requests, don't wait */
4202 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
4205 /* If there are other queues in the group, don't wait */
4206 if (cfqq->cfqg->nr_cfqq > 1)
4209 /* the only queue in the group, but think time is big */
4210 if (cfq_io_thinktime_big(cfqd, &cfqq->cfqg->ttime, true))
4213 if (cfq_slice_used(cfqq))
4216 /* if slice left is less than think time, wait busy */
4217 if (cic && sample_valid(cic->ttime.ttime_samples)
4218 && (cfqq->slice_end - now < cic->ttime.ttime_mean))
4222 * If think times is less than a jiffy than ttime_mean=0 and above
4223 * will not be true. It might happen that slice has not expired yet
4224 * but will expire soon (4-5 ns) during select_queue(). To cover the
4225 * case where think time is less than a jiffy, mark the queue wait
4226 * busy if only 1 jiffy is left in the slice.
4228 if (cfqq->slice_end - now <= jiffies_to_nsecs(1))
4234 static void cfq_completed_request(struct request_queue *q, struct request *rq)
4236 struct cfq_queue *cfqq = RQ_CFQQ(rq);
4237 struct cfq_data *cfqd = cfqq->cfqd;
4238 const int sync = rq_is_sync(rq);
4239 u64 now = ktime_get_ns();
4241 cfq_log_cfqq(cfqd, cfqq, "complete rqnoidle %d", req_noidle(rq));
4243 cfq_update_hw_tag(cfqd);
4245 WARN_ON(!cfqd->rq_in_driver);
4246 WARN_ON(!cfqq->dispatched);
4247 cfqd->rq_in_driver--;
4249 (RQ_CFQG(rq))->dispatched--;
4250 cfqg_stats_update_completion(cfqq->cfqg, rq_start_time_ns(rq),
4251 rq_io_start_time_ns(rq), rq->cmd_flags);
4253 cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]--;
4256 struct cfq_rb_root *st;
4258 RQ_CIC(rq)->ttime.last_end_request = now;
4260 if (cfq_cfqq_on_rr(cfqq))
4261 st = cfqq->service_tree;
4263 st = st_for(cfqq->cfqg, cfqq_class(cfqq),
4266 st->ttime.last_end_request = now;
4268 * We have to do this check in jiffies since start_time is in
4269 * jiffies and it is not trivial to convert to ns. If
4270 * cfq_fifo_expire[1] ever comes close to 1 jiffie, this test
4271 * will become problematic but so far we are fine (the default
4274 if (!time_after(rq->start_time +
4275 nsecs_to_jiffies(cfqd->cfq_fifo_expire[1]),
4277 cfqd->last_delayed_sync = now;
4280 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4281 cfqq->cfqg->ttime.last_end_request = now;
4285 * If this is the active queue, check if it needs to be expired,
4286 * or if we want to idle in case it has no pending requests.
4288 if (cfqd->active_queue == cfqq) {
4289 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
4291 if (cfq_cfqq_slice_new(cfqq)) {
4292 cfq_set_prio_slice(cfqd, cfqq);
4293 cfq_clear_cfqq_slice_new(cfqq);
4297 * Should we wait for next request to come in before we expire
4300 if (cfq_should_wait_busy(cfqd, cfqq)) {
4301 u64 extend_sl = cfqd->cfq_slice_idle;
4302 if (!cfqd->cfq_slice_idle)
4303 extend_sl = cfqd->cfq_group_idle;
4304 cfqq->slice_end = now + extend_sl;
4305 cfq_mark_cfqq_wait_busy(cfqq);
4306 cfq_log_cfqq(cfqd, cfqq, "will busy wait");
4310 * Idling is not enabled on:
4312 * - idle-priority queues
4314 * - queues with still some requests queued
4315 * - when there is a close cooperator
4317 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
4318 cfq_slice_expired(cfqd, 1);
4319 else if (sync && cfqq_empty &&
4320 !cfq_close_cooperator(cfqd, cfqq)) {
4321 cfq_arm_slice_timer(cfqd);
4325 if (!cfqd->rq_in_driver)
4326 cfq_schedule_dispatch(cfqd);
4329 static void cfqq_boost_on_prio(struct cfq_queue *cfqq, unsigned int op)
4332 * If REQ_PRIO is set, boost class and prio level, if it's below
4333 * BE/NORM. If prio is not set, restore the potentially boosted
4336 if (!(op & REQ_PRIO)) {
4337 cfqq->ioprio_class = cfqq->org_ioprio_class;
4338 cfqq->ioprio = cfqq->org_ioprio;
4340 if (cfq_class_idle(cfqq))
4341 cfqq->ioprio_class = IOPRIO_CLASS_BE;
4342 if (cfqq->ioprio > IOPRIO_NORM)
4343 cfqq->ioprio = IOPRIO_NORM;
4347 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
4349 if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
4350 cfq_mark_cfqq_must_alloc_slice(cfqq);
4351 return ELV_MQUEUE_MUST;
4354 return ELV_MQUEUE_MAY;
4357 static int cfq_may_queue(struct request_queue *q, unsigned int op)
4359 struct cfq_data *cfqd = q->elevator->elevator_data;
4360 struct task_struct *tsk = current;
4361 struct cfq_io_cq *cic;
4362 struct cfq_queue *cfqq;
4365 * don't force setup of a queue from here, as a call to may_queue
4366 * does not necessarily imply that a request actually will be queued.
4367 * so just lookup a possibly existing queue, or return 'may queue'
4370 cic = cfq_cic_lookup(cfqd, tsk->io_context);
4372 return ELV_MQUEUE_MAY;
4374 cfqq = cic_to_cfqq(cic, op_is_sync(op));
4376 cfq_init_prio_data(cfqq, cic);
4377 cfqq_boost_on_prio(cfqq, op);
4379 return __cfq_may_queue(cfqq);
4382 return ELV_MQUEUE_MAY;
4386 * queue lock held here
4388 static void cfq_put_request(struct request *rq)
4390 struct cfq_queue *cfqq = RQ_CFQQ(rq);
4393 const int rw = rq_data_dir(rq);
4395 BUG_ON(!cfqq->allocated[rw]);
4396 cfqq->allocated[rw]--;
4398 /* Put down rq reference on cfqg */
4399 cfqg_put(RQ_CFQG(rq));
4400 rq->elv.priv[0] = NULL;
4401 rq->elv.priv[1] = NULL;
4403 cfq_put_queue(cfqq);
4407 static struct cfq_queue *
4408 cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_cq *cic,
4409 struct cfq_queue *cfqq)
4411 cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq);
4412 cic_set_cfqq(cic, cfqq->new_cfqq, 1);
4413 cfq_mark_cfqq_coop(cfqq->new_cfqq);
4414 cfq_put_queue(cfqq);
4415 return cic_to_cfqq(cic, 1);
4419 * Returns NULL if a new cfqq should be allocated, or the old cfqq if this
4420 * was the last process referring to said cfqq.
4422 static struct cfq_queue *
4423 split_cfqq(struct cfq_io_cq *cic, struct cfq_queue *cfqq)
4425 if (cfqq_process_refs(cfqq) == 1) {
4426 cfqq->pid = current->pid;
4427 cfq_clear_cfqq_coop(cfqq);
4428 cfq_clear_cfqq_split_coop(cfqq);
4432 cic_set_cfqq(cic, NULL, 1);
4434 cfq_put_cooperator(cfqq);
4436 cfq_put_queue(cfqq);
4440 * Allocate cfq data structures associated with this request.
4443 cfq_set_request(struct request_queue *q, struct request *rq, struct bio *bio,
4446 struct cfq_data *cfqd = q->elevator->elevator_data;
4447 struct cfq_io_cq *cic = icq_to_cic(rq->elv.icq);
4448 const int rw = rq_data_dir(rq);
4449 const bool is_sync = rq_is_sync(rq);
4450 struct cfq_queue *cfqq;
4453 spin_lock_irq(q->queue_lock);
4455 check_ioprio_changed(cic, bio);
4456 disable_wbt = check_blkcg_changed(cic, bio);
4458 cfqq = cic_to_cfqq(cic, is_sync);
4459 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
4461 cfq_put_queue(cfqq);
4462 cfqq = cfq_get_queue(cfqd, is_sync, cic, bio);
4463 cic_set_cfqq(cic, cfqq, is_sync);
4466 * If the queue was seeky for too long, break it apart.
4468 if (cfq_cfqq_coop(cfqq) && cfq_cfqq_split_coop(cfqq)) {
4469 cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq");
4470 cfqq = split_cfqq(cic, cfqq);
4476 * Check to see if this queue is scheduled to merge with
4477 * another, closely cooperating queue. The merging of
4478 * queues happens here as it must be done in process context.
4479 * The reference on new_cfqq was taken in merge_cfqqs.
4482 cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq);
4485 cfqq->allocated[rw]++;
4488 cfqg_get(cfqq->cfqg);
4489 rq->elv.priv[0] = cfqq;
4490 rq->elv.priv[1] = cfqq->cfqg;
4491 spin_unlock_irq(q->queue_lock);
4494 wbt_disable_default(q);
4499 static void cfq_kick_queue(struct work_struct *work)
4501 struct cfq_data *cfqd =
4502 container_of(work, struct cfq_data, unplug_work);
4503 struct request_queue *q = cfqd->queue;
4505 spin_lock_irq(q->queue_lock);
4506 __blk_run_queue(cfqd->queue);
4507 spin_unlock_irq(q->queue_lock);
4511 * Timer running if the active_queue is currently idling inside its time slice
4513 static enum hrtimer_restart cfq_idle_slice_timer(struct hrtimer *timer)
4515 struct cfq_data *cfqd = container_of(timer, struct cfq_data,
4517 struct cfq_queue *cfqq;
4518 unsigned long flags;
4521 cfq_log(cfqd, "idle timer fired");
4523 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
4525 cfqq = cfqd->active_queue;
4530 * We saw a request before the queue expired, let it through
4532 if (cfq_cfqq_must_dispatch(cfqq))
4538 if (cfq_slice_used(cfqq))
4542 * only expire and reinvoke request handler, if there are
4543 * other queues with pending requests
4545 if (!cfqd->busy_queues)
4549 * not expired and it has a request pending, let it dispatch
4551 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
4555 * Queue depth flag is reset only when the idle didn't succeed
4557 cfq_clear_cfqq_deep(cfqq);
4560 cfq_slice_expired(cfqd, timed_out);
4562 cfq_schedule_dispatch(cfqd);
4564 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
4565 return HRTIMER_NORESTART;
4568 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
4570 hrtimer_cancel(&cfqd->idle_slice_timer);
4571 cancel_work_sync(&cfqd->unplug_work);
4574 static void cfq_exit_queue(struct elevator_queue *e)
4576 struct cfq_data *cfqd = e->elevator_data;
4577 struct request_queue *q = cfqd->queue;
4579 cfq_shutdown_timer_wq(cfqd);
4581 spin_lock_irq(q->queue_lock);
4583 if (cfqd->active_queue)
4584 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
4586 spin_unlock_irq(q->queue_lock);
4588 cfq_shutdown_timer_wq(cfqd);
4590 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4591 blkcg_deactivate_policy(q, &blkcg_policy_cfq);
4593 kfree(cfqd->root_group);
4598 static int cfq_init_queue(struct request_queue *q, struct elevator_type *e)
4600 struct cfq_data *cfqd;
4601 struct blkcg_gq *blkg __maybe_unused;
4603 struct elevator_queue *eq;
4605 eq = elevator_alloc(q, e);
4609 cfqd = kzalloc_node(sizeof(*cfqd), GFP_KERNEL, q->node);
4611 kobject_put(&eq->kobj);
4614 eq->elevator_data = cfqd;
4617 spin_lock_irq(q->queue_lock);
4619 spin_unlock_irq(q->queue_lock);
4621 /* Init root service tree */
4622 cfqd->grp_service_tree = CFQ_RB_ROOT;
4624 /* Init root group and prefer root group over other groups by default */
4625 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4626 ret = blkcg_activate_policy(q, &blkcg_policy_cfq);
4630 cfqd->root_group = blkg_to_cfqg(q->root_blkg);
4633 cfqd->root_group = kzalloc_node(sizeof(*cfqd->root_group),
4634 GFP_KERNEL, cfqd->queue->node);
4635 if (!cfqd->root_group)
4638 cfq_init_cfqg_base(cfqd->root_group);
4639 cfqd->root_group->weight = 2 * CFQ_WEIGHT_LEGACY_DFL;
4640 cfqd->root_group->leaf_weight = 2 * CFQ_WEIGHT_LEGACY_DFL;
4644 * Not strictly needed (since RB_ROOT just clears the node and we
4645 * zeroed cfqd on alloc), but better be safe in case someone decides
4646 * to add magic to the rb code
4648 for (i = 0; i < CFQ_PRIO_LISTS; i++)
4649 cfqd->prio_trees[i] = RB_ROOT;
4652 * Our fallback cfqq if cfq_get_queue() runs into OOM issues.
4653 * Grab a permanent reference to it, so that the normal code flow
4654 * will not attempt to free it. oom_cfqq is linked to root_group
4655 * but shouldn't hold a reference as it'll never be unlinked. Lose
4656 * the reference from linking right away.
4658 cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
4659 cfqd->oom_cfqq.ref++;
4661 spin_lock_irq(q->queue_lock);
4662 cfq_link_cfqq_cfqg(&cfqd->oom_cfqq, cfqd->root_group);
4663 cfqg_put(cfqd->root_group);
4664 spin_unlock_irq(q->queue_lock);
4666 hrtimer_init(&cfqd->idle_slice_timer, CLOCK_MONOTONIC,
4668 cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
4670 INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
4672 cfqd->cfq_quantum = cfq_quantum;
4673 cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
4674 cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
4675 cfqd->cfq_back_max = cfq_back_max;
4676 cfqd->cfq_back_penalty = cfq_back_penalty;
4677 cfqd->cfq_slice[0] = cfq_slice_async;
4678 cfqd->cfq_slice[1] = cfq_slice_sync;
4679 cfqd->cfq_target_latency = cfq_target_latency;
4680 cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
4681 cfqd->cfq_slice_idle = cfq_slice_idle;
4682 cfqd->cfq_group_idle = cfq_group_idle;
4683 cfqd->cfq_latency = 1;
4686 * we optimistically start assuming sync ops weren't delayed in last
4687 * second, in order to have larger depth for async operations.
4689 cfqd->last_delayed_sync = ktime_get_ns() - NSEC_PER_SEC;
4694 kobject_put(&eq->kobj);
4698 static void cfq_registered_queue(struct request_queue *q)
4700 struct elevator_queue *e = q->elevator;
4701 struct cfq_data *cfqd = e->elevator_data;
4704 * Default to IOPS mode with no idling for SSDs
4706 if (blk_queue_nonrot(q))
4707 cfqd->cfq_slice_idle = 0;
4711 * sysfs parts below -->
4714 cfq_var_show(unsigned int var, char *page)
4716 return sprintf(page, "%u\n", var);
4720 cfq_var_store(unsigned int *var, const char *page, size_t count)
4722 char *p = (char *) page;
4724 *var = simple_strtoul(p, &p, 10);
4728 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
4729 static ssize_t __FUNC(struct elevator_queue *e, char *page) \
4731 struct cfq_data *cfqd = e->elevator_data; \
4732 u64 __data = __VAR; \
4734 __data = div_u64(__data, NSEC_PER_MSEC); \
4735 return cfq_var_show(__data, (page)); \
4737 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
4738 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
4739 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
4740 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
4741 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
4742 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
4743 SHOW_FUNCTION(cfq_group_idle_show, cfqd->cfq_group_idle, 1);
4744 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
4745 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
4746 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
4747 SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
4748 SHOW_FUNCTION(cfq_target_latency_show, cfqd->cfq_target_latency, 1);
4749 #undef SHOW_FUNCTION
4751 #define USEC_SHOW_FUNCTION(__FUNC, __VAR) \
4752 static ssize_t __FUNC(struct elevator_queue *e, char *page) \
4754 struct cfq_data *cfqd = e->elevator_data; \
4755 u64 __data = __VAR; \
4756 __data = div_u64(__data, NSEC_PER_USEC); \
4757 return cfq_var_show(__data, (page)); \
4759 USEC_SHOW_FUNCTION(cfq_slice_idle_us_show, cfqd->cfq_slice_idle);
4760 USEC_SHOW_FUNCTION(cfq_group_idle_us_show, cfqd->cfq_group_idle);
4761 USEC_SHOW_FUNCTION(cfq_slice_sync_us_show, cfqd->cfq_slice[1]);
4762 USEC_SHOW_FUNCTION(cfq_slice_async_us_show, cfqd->cfq_slice[0]);
4763 USEC_SHOW_FUNCTION(cfq_target_latency_us_show, cfqd->cfq_target_latency);
4764 #undef USEC_SHOW_FUNCTION
4766 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
4767 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
4769 struct cfq_data *cfqd = e->elevator_data; \
4770 unsigned int __data; \
4771 int ret = cfq_var_store(&__data, (page), count); \
4772 if (__data < (MIN)) \
4774 else if (__data > (MAX)) \
4777 *(__PTR) = (u64)__data * NSEC_PER_MSEC; \
4779 *(__PTR) = __data; \
4782 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
4783 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
4785 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
4787 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
4788 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
4790 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
4791 STORE_FUNCTION(cfq_group_idle_store, &cfqd->cfq_group_idle, 0, UINT_MAX, 1);
4792 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
4793 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
4794 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
4796 STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
4797 STORE_FUNCTION(cfq_target_latency_store, &cfqd->cfq_target_latency, 1, UINT_MAX, 1);
4798 #undef STORE_FUNCTION
4800 #define USEC_STORE_FUNCTION(__FUNC, __PTR, MIN, MAX) \
4801 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
4803 struct cfq_data *cfqd = e->elevator_data; \
4804 unsigned int __data; \
4805 int ret = cfq_var_store(&__data, (page), count); \
4806 if (__data < (MIN)) \
4808 else if (__data > (MAX)) \
4810 *(__PTR) = (u64)__data * NSEC_PER_USEC; \
4813 USEC_STORE_FUNCTION(cfq_slice_idle_us_store, &cfqd->cfq_slice_idle, 0, UINT_MAX);
4814 USEC_STORE_FUNCTION(cfq_group_idle_us_store, &cfqd->cfq_group_idle, 0, UINT_MAX);
4815 USEC_STORE_FUNCTION(cfq_slice_sync_us_store, &cfqd->cfq_slice[1], 1, UINT_MAX);
4816 USEC_STORE_FUNCTION(cfq_slice_async_us_store, &cfqd->cfq_slice[0], 1, UINT_MAX);
4817 USEC_STORE_FUNCTION(cfq_target_latency_us_store, &cfqd->cfq_target_latency, 1, UINT_MAX);
4818 #undef USEC_STORE_FUNCTION
4820 #define CFQ_ATTR(name) \
4821 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
4823 static struct elv_fs_entry cfq_attrs[] = {
4825 CFQ_ATTR(fifo_expire_sync),
4826 CFQ_ATTR(fifo_expire_async),
4827 CFQ_ATTR(back_seek_max),
4828 CFQ_ATTR(back_seek_penalty),
4829 CFQ_ATTR(slice_sync),
4830 CFQ_ATTR(slice_sync_us),
4831 CFQ_ATTR(slice_async),
4832 CFQ_ATTR(slice_async_us),
4833 CFQ_ATTR(slice_async_rq),
4834 CFQ_ATTR(slice_idle),
4835 CFQ_ATTR(slice_idle_us),
4836 CFQ_ATTR(group_idle),
4837 CFQ_ATTR(group_idle_us),
4838 CFQ_ATTR(low_latency),
4839 CFQ_ATTR(target_latency),
4840 CFQ_ATTR(target_latency_us),
4844 static struct elevator_type iosched_cfq = {
4846 .elevator_merge_fn = cfq_merge,
4847 .elevator_merged_fn = cfq_merged_request,
4848 .elevator_merge_req_fn = cfq_merged_requests,
4849 .elevator_allow_bio_merge_fn = cfq_allow_bio_merge,
4850 .elevator_allow_rq_merge_fn = cfq_allow_rq_merge,
4851 .elevator_bio_merged_fn = cfq_bio_merged,
4852 .elevator_dispatch_fn = cfq_dispatch_requests,
4853 .elevator_add_req_fn = cfq_insert_request,
4854 .elevator_activate_req_fn = cfq_activate_request,
4855 .elevator_deactivate_req_fn = cfq_deactivate_request,
4856 .elevator_completed_req_fn = cfq_completed_request,
4857 .elevator_former_req_fn = elv_rb_former_request,
4858 .elevator_latter_req_fn = elv_rb_latter_request,
4859 .elevator_init_icq_fn = cfq_init_icq,
4860 .elevator_exit_icq_fn = cfq_exit_icq,
4861 .elevator_set_req_fn = cfq_set_request,
4862 .elevator_put_req_fn = cfq_put_request,
4863 .elevator_may_queue_fn = cfq_may_queue,
4864 .elevator_init_fn = cfq_init_queue,
4865 .elevator_exit_fn = cfq_exit_queue,
4866 .elevator_registered_fn = cfq_registered_queue,
4868 .icq_size = sizeof(struct cfq_io_cq),
4869 .icq_align = __alignof__(struct cfq_io_cq),
4870 .elevator_attrs = cfq_attrs,
4871 .elevator_name = "cfq",
4872 .elevator_owner = THIS_MODULE,
4875 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4876 static struct blkcg_policy blkcg_policy_cfq = {
4877 .dfl_cftypes = cfq_blkcg_files,
4878 .legacy_cftypes = cfq_blkcg_legacy_files,
4880 .cpd_alloc_fn = cfq_cpd_alloc,
4881 .cpd_init_fn = cfq_cpd_init,
4882 .cpd_free_fn = cfq_cpd_free,
4883 .cpd_bind_fn = cfq_cpd_bind,
4885 .pd_alloc_fn = cfq_pd_alloc,
4886 .pd_init_fn = cfq_pd_init,
4887 .pd_offline_fn = cfq_pd_offline,
4888 .pd_free_fn = cfq_pd_free,
4889 .pd_reset_stats_fn = cfq_pd_reset_stats,
4893 static int __init cfq_init(void)
4897 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4898 ret = blkcg_policy_register(&blkcg_policy_cfq);
4906 cfq_pool = KMEM_CACHE(cfq_queue, 0);
4910 ret = elv_register(&iosched_cfq);
4917 kmem_cache_destroy(cfq_pool);
4919 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4920 blkcg_policy_unregister(&blkcg_policy_cfq);
4925 static void __exit cfq_exit(void)
4927 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4928 blkcg_policy_unregister(&blkcg_policy_cfq);
4930 elv_unregister(&iosched_cfq);
4931 kmem_cache_destroy(cfq_pool);
4934 module_init(cfq_init);
4935 module_exit(cfq_exit);
4937 MODULE_AUTHOR("Jens Axboe");
4938 MODULE_LICENSE("GPL");
4939 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");