2 * CFQ, or complete fairness queueing, disk scheduler.
4 * Based on ideas from a previously unfinished io
5 * scheduler (round robin per-process disk scheduling) and Andrea Arcangeli.
7 * Copyright (C) 2003 Jens Axboe <axboe@kernel.dk>
9 #include <linux/module.h>
10 #include <linux/slab.h>
11 #include <linux/blkdev.h>
12 #include <linux/elevator.h>
13 #include <linux/jiffies.h>
14 #include <linux/rbtree.h>
15 #include <linux/ioprio.h>
16 #include <linux/blktrace_api.h>
17 #include <linux/blk-cgroup.h>
23 /* max queue in one round of service */
24 static const int cfq_quantum = 8;
25 static const int cfq_fifo_expire[2] = { HZ / 4, HZ / 8 };
26 /* maximum backwards seek, in KiB */
27 static const int cfq_back_max = 16 * 1024;
28 /* penalty of a backwards seek */
29 static const int cfq_back_penalty = 2;
30 static const int cfq_slice_sync = HZ / 10;
31 static int cfq_slice_async = HZ / 25;
32 static const int cfq_slice_async_rq = 2;
33 static int cfq_slice_idle = HZ / 125;
34 static int cfq_group_idle = HZ / 125;
35 static const int cfq_target_latency = HZ * 3/10; /* 300 ms */
36 static const int cfq_hist_divisor = 4;
39 * offset from end of service tree
41 #define CFQ_IDLE_DELAY (HZ / 5)
44 * below this threshold, we consider thinktime immediate
46 #define CFQ_MIN_TT (2)
48 #define CFQ_SLICE_SCALE (5)
49 #define CFQ_HW_QUEUE_MIN (5)
50 #define CFQ_SERVICE_SHIFT 12
52 #define CFQQ_SEEK_THR (sector_t)(8 * 100)
53 #define CFQQ_CLOSE_THR (sector_t)(8 * 1024)
54 #define CFQQ_SECT_THR_NONROT (sector_t)(2 * 32)
55 #define CFQQ_SEEKY(cfqq) (hweight32(cfqq->seek_history) > 32/8)
57 #define RQ_CIC(rq) icq_to_cic((rq)->elv.icq)
58 #define RQ_CFQQ(rq) (struct cfq_queue *) ((rq)->elv.priv[0])
59 #define RQ_CFQG(rq) (struct cfq_group *) ((rq)->elv.priv[1])
61 static struct kmem_cache *cfq_pool;
63 #define CFQ_PRIO_LISTS IOPRIO_BE_NR
64 #define cfq_class_idle(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
65 #define cfq_class_rt(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
67 #define sample_valid(samples) ((samples) > 80)
68 #define rb_entry_cfqg(node) rb_entry((node), struct cfq_group, rb_node)
70 /* blkio-related constants */
71 #define CFQ_WEIGHT_MIN 10
72 #define CFQ_WEIGHT_MAX 1000
73 #define CFQ_WEIGHT_DEFAULT 500
76 unsigned long last_end_request;
78 unsigned long ttime_total;
79 unsigned long ttime_samples;
80 unsigned long ttime_mean;
84 * Most of our rbtree usage is for sorting with min extraction, so
85 * if we cache the leftmost node we don't have to walk down the tree
86 * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should
87 * move this into the elevator for the rq sorting as well.
94 struct cfq_ttime ttime;
96 #define CFQ_RB_ROOT (struct cfq_rb_root) { .rb = RB_ROOT, \
97 .ttime = {.last_end_request = jiffies,},}
100 * Per process-grouping structure
103 /* reference count */
105 /* various state flags, see below */
107 /* parent cfq_data */
108 struct cfq_data *cfqd;
109 /* service_tree member */
110 struct rb_node rb_node;
111 /* service_tree key */
112 unsigned long rb_key;
113 /* prio tree member */
114 struct rb_node p_node;
115 /* prio tree root we belong to, if any */
116 struct rb_root *p_root;
117 /* sorted list of pending requests */
118 struct rb_root sort_list;
119 /* if fifo isn't expired, next request to serve */
120 struct request *next_rq;
121 /* requests queued in sort_list */
123 /* currently allocated requests */
125 /* fifo list of requests in sort_list */
126 struct list_head fifo;
128 /* time when queue got scheduled in to dispatch first request. */
129 unsigned long dispatch_start;
130 unsigned int allocated_slice;
131 unsigned int slice_dispatch;
132 /* time when first request from queue completed and slice started. */
133 unsigned long slice_start;
134 unsigned long slice_end;
137 /* pending priority requests */
139 /* number of requests that are on the dispatch list or inside driver */
142 /* io prio of this group */
143 unsigned short ioprio, org_ioprio;
144 unsigned short ioprio_class;
149 sector_t last_request_pos;
151 struct cfq_rb_root *service_tree;
152 struct cfq_queue *new_cfqq;
153 struct cfq_group *cfqg;
154 /* Number of sectors dispatched from queue in single dispatch round */
155 unsigned long nr_sectors;
159 * First index in the service_trees.
160 * IDLE is handled separately, so it has negative index
170 * Second index in the service_trees.
174 SYNC_NOIDLE_WORKLOAD = 1,
179 #ifdef CONFIG_CFQ_GROUP_IOSCHED
180 /* number of ios merged */
181 struct blkg_rwstat merged;
182 /* total time spent on device in ns, may not be accurate w/ queueing */
183 struct blkg_rwstat service_time;
184 /* total time spent waiting in scheduler queue in ns */
185 struct blkg_rwstat wait_time;
186 /* number of IOs queued up */
187 struct blkg_rwstat queued;
188 /* total disk time and nr sectors dispatched by this group */
189 struct blkg_stat time;
190 #ifdef CONFIG_DEBUG_BLK_CGROUP
191 /* time not charged to this cgroup */
192 struct blkg_stat unaccounted_time;
193 /* sum of number of ios queued across all samples */
194 struct blkg_stat avg_queue_size_sum;
195 /* count of samples taken for average */
196 struct blkg_stat avg_queue_size_samples;
197 /* how many times this group has been removed from service tree */
198 struct blkg_stat dequeue;
199 /* total time spent waiting for it to be assigned a timeslice. */
200 struct blkg_stat group_wait_time;
201 /* time spent idling for this blkcg_gq */
202 struct blkg_stat idle_time;
203 /* total time with empty current active q with other requests queued */
204 struct blkg_stat empty_time;
205 /* fields after this shouldn't be cleared on stat reset */
206 uint64_t start_group_wait_time;
207 uint64_t start_idle_time;
208 uint64_t start_empty_time;
210 #endif /* CONFIG_DEBUG_BLK_CGROUP */
211 #endif /* CONFIG_CFQ_GROUP_IOSCHED */
214 /* Per-cgroup data */
215 struct cfq_group_data {
216 /* must be the first member */
217 struct blkcg_policy_data cpd;
220 unsigned int leaf_weight;
223 /* This is per cgroup per device grouping structure */
225 /* must be the first member */
226 struct blkg_policy_data pd;
228 /* group service_tree member */
229 struct rb_node rb_node;
231 /* group service_tree key */
235 * The number of active cfqgs and sum of their weights under this
236 * cfqg. This covers this cfqg's leaf_weight and all children's
237 * weights, but does not cover weights of further descendants.
239 * If a cfqg is on the service tree, it's active. An active cfqg
240 * also activates its parent and contributes to the children_weight
244 unsigned int children_weight;
247 * vfraction is the fraction of vdisktime that the tasks in this
248 * cfqg are entitled to. This is determined by compounding the
249 * ratios walking up from this cfqg to the root.
251 * It is in fixed point w/ CFQ_SERVICE_SHIFT and the sum of all
252 * vfractions on a service tree is approximately 1. The sum may
253 * deviate a bit due to rounding errors and fluctuations caused by
254 * cfqgs entering and leaving the service tree.
256 unsigned int vfraction;
259 * There are two weights - (internal) weight is the weight of this
260 * cfqg against the sibling cfqgs. leaf_weight is the wight of
261 * this cfqg against the child cfqgs. For the root cfqg, both
262 * weights are kept in sync for backward compatibility.
265 unsigned int new_weight;
266 unsigned int dev_weight;
268 unsigned int leaf_weight;
269 unsigned int new_leaf_weight;
270 unsigned int dev_leaf_weight;
272 /* number of cfqq currently on this group */
276 * Per group busy queues average. Useful for workload slice calc. We
277 * create the array for each prio class but at run time it is used
278 * only for RT and BE class and slot for IDLE class remains unused.
279 * This is primarily done to avoid confusion and a gcc warning.
281 unsigned int busy_queues_avg[CFQ_PRIO_NR];
283 * rr lists of queues with requests. We maintain service trees for
284 * RT and BE classes. These trees are subdivided in subclasses
285 * of SYNC, SYNC_NOIDLE and ASYNC based on workload type. For IDLE
286 * class there is no subclassification and all the cfq queues go on
287 * a single tree service_tree_idle.
288 * Counts are embedded in the cfq_rb_root
290 struct cfq_rb_root service_trees[2][3];
291 struct cfq_rb_root service_tree_idle;
293 unsigned long saved_wl_slice;
294 enum wl_type_t saved_wl_type;
295 enum wl_class_t saved_wl_class;
297 /* number of requests that are on the dispatch list or inside driver */
299 struct cfq_ttime ttime;
300 struct cfqg_stats stats; /* stats for this cfqg */
302 /* async queue for each priority case */
303 struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
304 struct cfq_queue *async_idle_cfqq;
309 struct io_cq icq; /* must be the first member */
310 struct cfq_queue *cfqq[2];
311 struct cfq_ttime ttime;
312 int ioprio; /* the current ioprio */
313 #ifdef CONFIG_CFQ_GROUP_IOSCHED
314 uint64_t blkcg_serial_nr; /* the current blkcg serial */
319 * Per block device queue structure
322 struct request_queue *queue;
323 /* Root service tree for cfq_groups */
324 struct cfq_rb_root grp_service_tree;
325 struct cfq_group *root_group;
328 * The priority currently being served
330 enum wl_class_t serving_wl_class;
331 enum wl_type_t serving_wl_type;
332 unsigned long workload_expires;
333 struct cfq_group *serving_group;
336 * Each priority tree is sorted by next_request position. These
337 * trees are used when determining if two or more queues are
338 * interleaving requests (see cfq_close_cooperator).
340 struct rb_root prio_trees[CFQ_PRIO_LISTS];
342 unsigned int busy_queues;
343 unsigned int busy_sync_queues;
349 * queue-depth detection
355 * -1 => indeterminate, (cfq will behave as if NCQ is present, to allow better detection)
356 * 1 => NCQ is present (hw_tag_est_depth is the estimated max depth)
359 int hw_tag_est_depth;
360 unsigned int hw_tag_samples;
363 * idle window management
365 struct timer_list idle_slice_timer;
366 struct work_struct unplug_work;
368 struct cfq_queue *active_queue;
369 struct cfq_io_cq *active_cic;
371 sector_t last_position;
374 * tunables, see top of file
376 unsigned int cfq_quantum;
377 unsigned int cfq_fifo_expire[2];
378 unsigned int cfq_back_penalty;
379 unsigned int cfq_back_max;
380 unsigned int cfq_slice[2];
381 unsigned int cfq_slice_async_rq;
382 unsigned int cfq_slice_idle;
383 unsigned int cfq_group_idle;
384 unsigned int cfq_latency;
385 unsigned int cfq_target_latency;
388 * Fallback dummy cfqq for extreme OOM conditions
390 struct cfq_queue oom_cfqq;
392 unsigned long last_delayed_sync;
395 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd);
396 static void cfq_put_queue(struct cfq_queue *cfqq);
398 static struct cfq_rb_root *st_for(struct cfq_group *cfqg,
399 enum wl_class_t class,
405 if (class == IDLE_WORKLOAD)
406 return &cfqg->service_tree_idle;
408 return &cfqg->service_trees[class][type];
411 enum cfqq_state_flags {
412 CFQ_CFQQ_FLAG_on_rr = 0, /* on round-robin busy list */
413 CFQ_CFQQ_FLAG_wait_request, /* waiting for a request */
414 CFQ_CFQQ_FLAG_must_dispatch, /* must be allowed a dispatch */
415 CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
416 CFQ_CFQQ_FLAG_fifo_expire, /* FIFO checked in this slice */
417 CFQ_CFQQ_FLAG_idle_window, /* slice idling enabled */
418 CFQ_CFQQ_FLAG_prio_changed, /* task priority has changed */
419 CFQ_CFQQ_FLAG_slice_new, /* no requests dispatched in slice */
420 CFQ_CFQQ_FLAG_sync, /* synchronous queue */
421 CFQ_CFQQ_FLAG_coop, /* cfqq is shared */
422 CFQ_CFQQ_FLAG_split_coop, /* shared cfqq will be splitted */
423 CFQ_CFQQ_FLAG_deep, /* sync cfqq experienced large depth */
424 CFQ_CFQQ_FLAG_wait_busy, /* Waiting for next request */
427 #define CFQ_CFQQ_FNS(name) \
428 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \
430 (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name); \
432 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \
434 (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \
436 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \
438 return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \
442 CFQ_CFQQ_FNS(wait_request);
443 CFQ_CFQQ_FNS(must_dispatch);
444 CFQ_CFQQ_FNS(must_alloc_slice);
445 CFQ_CFQQ_FNS(fifo_expire);
446 CFQ_CFQQ_FNS(idle_window);
447 CFQ_CFQQ_FNS(prio_changed);
448 CFQ_CFQQ_FNS(slice_new);
451 CFQ_CFQQ_FNS(split_coop);
453 CFQ_CFQQ_FNS(wait_busy);
456 #if defined(CONFIG_CFQ_GROUP_IOSCHED) && defined(CONFIG_DEBUG_BLK_CGROUP)
458 /* cfqg stats flags */
459 enum cfqg_stats_flags {
460 CFQG_stats_waiting = 0,
465 #define CFQG_FLAG_FNS(name) \
466 static inline void cfqg_stats_mark_##name(struct cfqg_stats *stats) \
468 stats->flags |= (1 << CFQG_stats_##name); \
470 static inline void cfqg_stats_clear_##name(struct cfqg_stats *stats) \
472 stats->flags &= ~(1 << CFQG_stats_##name); \
474 static inline int cfqg_stats_##name(struct cfqg_stats *stats) \
476 return (stats->flags & (1 << CFQG_stats_##name)) != 0; \
479 CFQG_FLAG_FNS(waiting)
480 CFQG_FLAG_FNS(idling)
484 /* This should be called with the queue_lock held. */
485 static void cfqg_stats_update_group_wait_time(struct cfqg_stats *stats)
487 unsigned long long now;
489 if (!cfqg_stats_waiting(stats))
493 if (time_after64(now, stats->start_group_wait_time))
494 blkg_stat_add(&stats->group_wait_time,
495 now - stats->start_group_wait_time);
496 cfqg_stats_clear_waiting(stats);
499 /* This should be called with the queue_lock held. */
500 static void cfqg_stats_set_start_group_wait_time(struct cfq_group *cfqg,
501 struct cfq_group *curr_cfqg)
503 struct cfqg_stats *stats = &cfqg->stats;
505 if (cfqg_stats_waiting(stats))
507 if (cfqg == curr_cfqg)
509 stats->start_group_wait_time = sched_clock();
510 cfqg_stats_mark_waiting(stats);
513 /* This should be called with the queue_lock held. */
514 static void cfqg_stats_end_empty_time(struct cfqg_stats *stats)
516 unsigned long long now;
518 if (!cfqg_stats_empty(stats))
522 if (time_after64(now, stats->start_empty_time))
523 blkg_stat_add(&stats->empty_time,
524 now - stats->start_empty_time);
525 cfqg_stats_clear_empty(stats);
528 static void cfqg_stats_update_dequeue(struct cfq_group *cfqg)
530 blkg_stat_add(&cfqg->stats.dequeue, 1);
533 static void cfqg_stats_set_start_empty_time(struct cfq_group *cfqg)
535 struct cfqg_stats *stats = &cfqg->stats;
537 if (blkg_rwstat_total(&stats->queued))
541 * group is already marked empty. This can happen if cfqq got new
542 * request in parent group and moved to this group while being added
543 * to service tree. Just ignore the event and move on.
545 if (cfqg_stats_empty(stats))
548 stats->start_empty_time = sched_clock();
549 cfqg_stats_mark_empty(stats);
552 static void cfqg_stats_update_idle_time(struct cfq_group *cfqg)
554 struct cfqg_stats *stats = &cfqg->stats;
556 if (cfqg_stats_idling(stats)) {
557 unsigned long long now = sched_clock();
559 if (time_after64(now, stats->start_idle_time))
560 blkg_stat_add(&stats->idle_time,
561 now - stats->start_idle_time);
562 cfqg_stats_clear_idling(stats);
566 static void cfqg_stats_set_start_idle_time(struct cfq_group *cfqg)
568 struct cfqg_stats *stats = &cfqg->stats;
570 BUG_ON(cfqg_stats_idling(stats));
572 stats->start_idle_time = sched_clock();
573 cfqg_stats_mark_idling(stats);
576 static void cfqg_stats_update_avg_queue_size(struct cfq_group *cfqg)
578 struct cfqg_stats *stats = &cfqg->stats;
580 blkg_stat_add(&stats->avg_queue_size_sum,
581 blkg_rwstat_total(&stats->queued));
582 blkg_stat_add(&stats->avg_queue_size_samples, 1);
583 cfqg_stats_update_group_wait_time(stats);
586 #else /* CONFIG_CFQ_GROUP_IOSCHED && CONFIG_DEBUG_BLK_CGROUP */
588 static inline void cfqg_stats_set_start_group_wait_time(struct cfq_group *cfqg, struct cfq_group *curr_cfqg) { }
589 static inline void cfqg_stats_end_empty_time(struct cfqg_stats *stats) { }
590 static inline void cfqg_stats_update_dequeue(struct cfq_group *cfqg) { }
591 static inline void cfqg_stats_set_start_empty_time(struct cfq_group *cfqg) { }
592 static inline void cfqg_stats_update_idle_time(struct cfq_group *cfqg) { }
593 static inline void cfqg_stats_set_start_idle_time(struct cfq_group *cfqg) { }
594 static inline void cfqg_stats_update_avg_queue_size(struct cfq_group *cfqg) { }
596 #endif /* CONFIG_CFQ_GROUP_IOSCHED && CONFIG_DEBUG_BLK_CGROUP */
598 #ifdef CONFIG_CFQ_GROUP_IOSCHED
600 static inline struct cfq_group *pd_to_cfqg(struct blkg_policy_data *pd)
602 return pd ? container_of(pd, struct cfq_group, pd) : NULL;
605 static struct cfq_group_data
606 *cpd_to_cfqgd(struct blkcg_policy_data *cpd)
608 return cpd ? container_of(cpd, struct cfq_group_data, cpd) : NULL;
611 static inline struct blkcg_gq *cfqg_to_blkg(struct cfq_group *cfqg)
613 return pd_to_blkg(&cfqg->pd);
616 static struct blkcg_policy blkcg_policy_cfq;
618 static inline struct cfq_group *blkg_to_cfqg(struct blkcg_gq *blkg)
620 return pd_to_cfqg(blkg_to_pd(blkg, &blkcg_policy_cfq));
623 static struct cfq_group_data *blkcg_to_cfqgd(struct blkcg *blkcg)
625 return cpd_to_cfqgd(blkcg_to_cpd(blkcg, &blkcg_policy_cfq));
628 static inline struct cfq_group *cfqg_parent(struct cfq_group *cfqg)
630 struct blkcg_gq *pblkg = cfqg_to_blkg(cfqg)->parent;
632 return pblkg ? blkg_to_cfqg(pblkg) : NULL;
635 static inline void cfqg_get(struct cfq_group *cfqg)
637 return blkg_get(cfqg_to_blkg(cfqg));
640 static inline void cfqg_put(struct cfq_group *cfqg)
642 return blkg_put(cfqg_to_blkg(cfqg));
645 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...) do { \
648 blkg_path(cfqg_to_blkg((cfqq)->cfqg), __pbuf, sizeof(__pbuf)); \
649 blk_add_trace_msg((cfqd)->queue, "cfq%d%c%c %s " fmt, (cfqq)->pid, \
650 cfq_cfqq_sync((cfqq)) ? 'S' : 'A', \
651 cfqq_type((cfqq)) == SYNC_NOIDLE_WORKLOAD ? 'N' : ' ',\
655 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...) do { \
658 blkg_path(cfqg_to_blkg(cfqg), __pbuf, sizeof(__pbuf)); \
659 blk_add_trace_msg((cfqd)->queue, "%s " fmt, __pbuf, ##args); \
662 static inline void cfqg_stats_update_io_add(struct cfq_group *cfqg,
663 struct cfq_group *curr_cfqg, int rw)
665 blkg_rwstat_add(&cfqg->stats.queued, rw, 1);
666 cfqg_stats_end_empty_time(&cfqg->stats);
667 cfqg_stats_set_start_group_wait_time(cfqg, curr_cfqg);
670 static inline void cfqg_stats_update_timeslice_used(struct cfq_group *cfqg,
671 unsigned long time, unsigned long unaccounted_time)
673 blkg_stat_add(&cfqg->stats.time, time);
674 #ifdef CONFIG_DEBUG_BLK_CGROUP
675 blkg_stat_add(&cfqg->stats.unaccounted_time, unaccounted_time);
679 static inline void cfqg_stats_update_io_remove(struct cfq_group *cfqg, int rw)
681 blkg_rwstat_add(&cfqg->stats.queued, rw, -1);
684 static inline void cfqg_stats_update_io_merged(struct cfq_group *cfqg, int rw)
686 blkg_rwstat_add(&cfqg->stats.merged, rw, 1);
689 static inline void cfqg_stats_update_completion(struct cfq_group *cfqg,
690 uint64_t start_time, uint64_t io_start_time, int rw)
692 struct cfqg_stats *stats = &cfqg->stats;
693 unsigned long long now = sched_clock();
695 if (time_after64(now, io_start_time))
696 blkg_rwstat_add(&stats->service_time, rw, now - io_start_time);
697 if (time_after64(io_start_time, start_time))
698 blkg_rwstat_add(&stats->wait_time, rw,
699 io_start_time - start_time);
703 static void cfqg_stats_reset(struct cfqg_stats *stats)
705 /* queued stats shouldn't be cleared */
706 blkg_rwstat_reset(&stats->merged);
707 blkg_rwstat_reset(&stats->service_time);
708 blkg_rwstat_reset(&stats->wait_time);
709 blkg_stat_reset(&stats->time);
710 #ifdef CONFIG_DEBUG_BLK_CGROUP
711 blkg_stat_reset(&stats->unaccounted_time);
712 blkg_stat_reset(&stats->avg_queue_size_sum);
713 blkg_stat_reset(&stats->avg_queue_size_samples);
714 blkg_stat_reset(&stats->dequeue);
715 blkg_stat_reset(&stats->group_wait_time);
716 blkg_stat_reset(&stats->idle_time);
717 blkg_stat_reset(&stats->empty_time);
722 static void cfqg_stats_add_aux(struct cfqg_stats *to, struct cfqg_stats *from)
724 /* queued stats shouldn't be cleared */
725 blkg_rwstat_add_aux(&to->merged, &from->merged);
726 blkg_rwstat_add_aux(&to->service_time, &from->service_time);
727 blkg_rwstat_add_aux(&to->wait_time, &from->wait_time);
728 blkg_stat_add_aux(&from->time, &from->time);
729 #ifdef CONFIG_DEBUG_BLK_CGROUP
730 blkg_stat_add_aux(&to->unaccounted_time, &from->unaccounted_time);
731 blkg_stat_add_aux(&to->avg_queue_size_sum, &from->avg_queue_size_sum);
732 blkg_stat_add_aux(&to->avg_queue_size_samples, &from->avg_queue_size_samples);
733 blkg_stat_add_aux(&to->dequeue, &from->dequeue);
734 blkg_stat_add_aux(&to->group_wait_time, &from->group_wait_time);
735 blkg_stat_add_aux(&to->idle_time, &from->idle_time);
736 blkg_stat_add_aux(&to->empty_time, &from->empty_time);
741 * Transfer @cfqg's stats to its parent's aux counts so that the ancestors'
742 * recursive stats can still account for the amount used by this cfqg after
745 static void cfqg_stats_xfer_dead(struct cfq_group *cfqg)
747 struct cfq_group *parent = cfqg_parent(cfqg);
749 lockdep_assert_held(cfqg_to_blkg(cfqg)->q->queue_lock);
751 if (unlikely(!parent))
754 cfqg_stats_add_aux(&parent->stats, &cfqg->stats);
755 cfqg_stats_reset(&cfqg->stats);
758 #else /* CONFIG_CFQ_GROUP_IOSCHED */
760 static inline struct cfq_group *cfqg_parent(struct cfq_group *cfqg) { return NULL; }
761 static inline void cfqg_get(struct cfq_group *cfqg) { }
762 static inline void cfqg_put(struct cfq_group *cfqg) { }
764 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
765 blk_add_trace_msg((cfqd)->queue, "cfq%d%c%c " fmt, (cfqq)->pid, \
766 cfq_cfqq_sync((cfqq)) ? 'S' : 'A', \
767 cfqq_type((cfqq)) == SYNC_NOIDLE_WORKLOAD ? 'N' : ' ',\
769 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...) do {} while (0)
771 static inline void cfqg_stats_update_io_add(struct cfq_group *cfqg,
772 struct cfq_group *curr_cfqg, int rw) { }
773 static inline void cfqg_stats_update_timeslice_used(struct cfq_group *cfqg,
774 unsigned long time, unsigned long unaccounted_time) { }
775 static inline void cfqg_stats_update_io_remove(struct cfq_group *cfqg, int rw) { }
776 static inline void cfqg_stats_update_io_merged(struct cfq_group *cfqg, int rw) { }
777 static inline void cfqg_stats_update_completion(struct cfq_group *cfqg,
778 uint64_t start_time, uint64_t io_start_time, int rw) { }
780 #endif /* CONFIG_CFQ_GROUP_IOSCHED */
782 #define cfq_log(cfqd, fmt, args...) \
783 blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
785 /* Traverses through cfq group service trees */
786 #define for_each_cfqg_st(cfqg, i, j, st) \
787 for (i = 0; i <= IDLE_WORKLOAD; i++) \
788 for (j = 0, st = i < IDLE_WORKLOAD ? &cfqg->service_trees[i][j]\
789 : &cfqg->service_tree_idle; \
790 (i < IDLE_WORKLOAD && j <= SYNC_WORKLOAD) || \
791 (i == IDLE_WORKLOAD && j == 0); \
792 j++, st = i < IDLE_WORKLOAD ? \
793 &cfqg->service_trees[i][j]: NULL) \
795 static inline bool cfq_io_thinktime_big(struct cfq_data *cfqd,
796 struct cfq_ttime *ttime, bool group_idle)
799 if (!sample_valid(ttime->ttime_samples))
802 slice = cfqd->cfq_group_idle;
804 slice = cfqd->cfq_slice_idle;
805 return ttime->ttime_mean > slice;
808 static inline bool iops_mode(struct cfq_data *cfqd)
811 * If we are not idling on queues and it is a NCQ drive, parallel
812 * execution of requests is on and measuring time is not possible
813 * in most of the cases until and unless we drive shallower queue
814 * depths and that becomes a performance bottleneck. In such cases
815 * switch to start providing fairness in terms of number of IOs.
817 if (!cfqd->cfq_slice_idle && cfqd->hw_tag)
823 static inline enum wl_class_t cfqq_class(struct cfq_queue *cfqq)
825 if (cfq_class_idle(cfqq))
826 return IDLE_WORKLOAD;
827 if (cfq_class_rt(cfqq))
833 static enum wl_type_t cfqq_type(struct cfq_queue *cfqq)
835 if (!cfq_cfqq_sync(cfqq))
836 return ASYNC_WORKLOAD;
837 if (!cfq_cfqq_idle_window(cfqq))
838 return SYNC_NOIDLE_WORKLOAD;
839 return SYNC_WORKLOAD;
842 static inline int cfq_group_busy_queues_wl(enum wl_class_t wl_class,
843 struct cfq_data *cfqd,
844 struct cfq_group *cfqg)
846 if (wl_class == IDLE_WORKLOAD)
847 return cfqg->service_tree_idle.count;
849 return cfqg->service_trees[wl_class][ASYNC_WORKLOAD].count +
850 cfqg->service_trees[wl_class][SYNC_NOIDLE_WORKLOAD].count +
851 cfqg->service_trees[wl_class][SYNC_WORKLOAD].count;
854 static inline int cfqg_busy_async_queues(struct cfq_data *cfqd,
855 struct cfq_group *cfqg)
857 return cfqg->service_trees[RT_WORKLOAD][ASYNC_WORKLOAD].count +
858 cfqg->service_trees[BE_WORKLOAD][ASYNC_WORKLOAD].count;
861 static void cfq_dispatch_insert(struct request_queue *, struct request *);
862 static struct cfq_queue *cfq_get_queue(struct cfq_data *cfqd, bool is_sync,
863 struct cfq_io_cq *cic, struct bio *bio);
865 static inline struct cfq_io_cq *icq_to_cic(struct io_cq *icq)
867 /* cic->icq is the first member, %NULL will convert to %NULL */
868 return container_of(icq, struct cfq_io_cq, icq);
871 static inline struct cfq_io_cq *cfq_cic_lookup(struct cfq_data *cfqd,
872 struct io_context *ioc)
875 return icq_to_cic(ioc_lookup_icq(ioc, cfqd->queue));
879 static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_cq *cic, bool is_sync)
881 return cic->cfqq[is_sync];
884 static inline void cic_set_cfqq(struct cfq_io_cq *cic, struct cfq_queue *cfqq,
887 cic->cfqq[is_sync] = cfqq;
890 static inline struct cfq_data *cic_to_cfqd(struct cfq_io_cq *cic)
892 return cic->icq.q->elevator->elevator_data;
896 * We regard a request as SYNC, if it's either a read or has the SYNC bit
897 * set (in which case it could also be direct WRITE).
899 static inline bool cfq_bio_sync(struct bio *bio)
901 return bio_data_dir(bio) == READ || (bio->bi_rw & REQ_SYNC);
905 * scheduler run of queue, if there are requests pending and no one in the
906 * driver that will restart queueing
908 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
910 if (cfqd->busy_queues) {
911 cfq_log(cfqd, "schedule dispatch");
912 kblockd_schedule_work(&cfqd->unplug_work);
917 * Scale schedule slice based on io priority. Use the sync time slice only
918 * if a queue is marked sync and has sync io queued. A sync queue with async
919 * io only, should not get full sync slice length.
921 static inline int cfq_prio_slice(struct cfq_data *cfqd, bool sync,
924 const int base_slice = cfqd->cfq_slice[sync];
926 WARN_ON(prio >= IOPRIO_BE_NR);
928 return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
932 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
934 return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
938 * cfqg_scale_charge - scale disk time charge according to cfqg weight
939 * @charge: disk time being charged
940 * @vfraction: vfraction of the cfqg, fixed point w/ CFQ_SERVICE_SHIFT
942 * Scale @charge according to @vfraction, which is in range (0, 1]. The
943 * scaling is inversely proportional.
945 * scaled = charge / vfraction
947 * The result is also in fixed point w/ CFQ_SERVICE_SHIFT.
949 static inline u64 cfqg_scale_charge(unsigned long charge,
950 unsigned int vfraction)
952 u64 c = charge << CFQ_SERVICE_SHIFT; /* make it fixed point */
954 /* charge / vfraction */
955 c <<= CFQ_SERVICE_SHIFT;
956 do_div(c, vfraction);
960 static inline u64 max_vdisktime(u64 min_vdisktime, u64 vdisktime)
962 s64 delta = (s64)(vdisktime - min_vdisktime);
964 min_vdisktime = vdisktime;
966 return min_vdisktime;
969 static inline u64 min_vdisktime(u64 min_vdisktime, u64 vdisktime)
971 s64 delta = (s64)(vdisktime - min_vdisktime);
973 min_vdisktime = vdisktime;
975 return min_vdisktime;
978 static void update_min_vdisktime(struct cfq_rb_root *st)
980 struct cfq_group *cfqg;
983 cfqg = rb_entry_cfqg(st->left);
984 st->min_vdisktime = max_vdisktime(st->min_vdisktime,
990 * get averaged number of queues of RT/BE priority.
991 * average is updated, with a formula that gives more weight to higher numbers,
992 * to quickly follows sudden increases and decrease slowly
995 static inline unsigned cfq_group_get_avg_queues(struct cfq_data *cfqd,
996 struct cfq_group *cfqg, bool rt)
998 unsigned min_q, max_q;
999 unsigned mult = cfq_hist_divisor - 1;
1000 unsigned round = cfq_hist_divisor / 2;
1001 unsigned busy = cfq_group_busy_queues_wl(rt, cfqd, cfqg);
1003 min_q = min(cfqg->busy_queues_avg[rt], busy);
1004 max_q = max(cfqg->busy_queues_avg[rt], busy);
1005 cfqg->busy_queues_avg[rt] = (mult * max_q + min_q + round) /
1007 return cfqg->busy_queues_avg[rt];
1010 static inline unsigned
1011 cfq_group_slice(struct cfq_data *cfqd, struct cfq_group *cfqg)
1013 return cfqd->cfq_target_latency * cfqg->vfraction >> CFQ_SERVICE_SHIFT;
1016 static inline unsigned
1017 cfq_scaled_cfqq_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1019 unsigned slice = cfq_prio_to_slice(cfqd, cfqq);
1020 if (cfqd->cfq_latency) {
1022 * interested queues (we consider only the ones with the same
1023 * priority class in the cfq group)
1025 unsigned iq = cfq_group_get_avg_queues(cfqd, cfqq->cfqg,
1026 cfq_class_rt(cfqq));
1027 unsigned sync_slice = cfqd->cfq_slice[1];
1028 unsigned expect_latency = sync_slice * iq;
1029 unsigned group_slice = cfq_group_slice(cfqd, cfqq->cfqg);
1031 if (expect_latency > group_slice) {
1032 unsigned base_low_slice = 2 * cfqd->cfq_slice_idle;
1033 /* scale low_slice according to IO priority
1034 * and sync vs async */
1035 unsigned low_slice =
1036 min(slice, base_low_slice * slice / sync_slice);
1037 /* the adapted slice value is scaled to fit all iqs
1038 * into the target latency */
1039 slice = max(slice * group_slice / expect_latency,
1047 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1049 unsigned slice = cfq_scaled_cfqq_slice(cfqd, cfqq);
1051 cfqq->slice_start = jiffies;
1052 cfqq->slice_end = jiffies + slice;
1053 cfqq->allocated_slice = slice;
1054 cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
1058 * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
1059 * isn't valid until the first request from the dispatch is activated
1060 * and the slice time set.
1062 static inline bool cfq_slice_used(struct cfq_queue *cfqq)
1064 if (cfq_cfqq_slice_new(cfqq))
1066 if (time_before(jiffies, cfqq->slice_end))
1073 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
1074 * We choose the request that is closest to the head right now. Distance
1075 * behind the head is penalized and only allowed to a certain extent.
1077 static struct request *
1078 cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2, sector_t last)
1080 sector_t s1, s2, d1 = 0, d2 = 0;
1081 unsigned long back_max;
1082 #define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */
1083 #define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */
1084 unsigned wrap = 0; /* bit mask: requests behind the disk head? */
1086 if (rq1 == NULL || rq1 == rq2)
1091 if (rq_is_sync(rq1) != rq_is_sync(rq2))
1092 return rq_is_sync(rq1) ? rq1 : rq2;
1094 if ((rq1->cmd_flags ^ rq2->cmd_flags) & REQ_PRIO)
1095 return rq1->cmd_flags & REQ_PRIO ? rq1 : rq2;
1097 s1 = blk_rq_pos(rq1);
1098 s2 = blk_rq_pos(rq2);
1101 * by definition, 1KiB is 2 sectors
1103 back_max = cfqd->cfq_back_max * 2;
1106 * Strict one way elevator _except_ in the case where we allow
1107 * short backward seeks which are biased as twice the cost of a
1108 * similar forward seek.
1112 else if (s1 + back_max >= last)
1113 d1 = (last - s1) * cfqd->cfq_back_penalty;
1115 wrap |= CFQ_RQ1_WRAP;
1119 else if (s2 + back_max >= last)
1120 d2 = (last - s2) * cfqd->cfq_back_penalty;
1122 wrap |= CFQ_RQ2_WRAP;
1124 /* Found required data */
1127 * By doing switch() on the bit mask "wrap" we avoid having to
1128 * check two variables for all permutations: --> faster!
1131 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
1147 case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
1150 * Since both rqs are wrapped,
1151 * start with the one that's further behind head
1152 * (--> only *one* back seek required),
1153 * since back seek takes more time than forward.
1163 * The below is leftmost cache rbtree addon
1165 static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
1167 /* Service tree is empty */
1172 root->left = rb_first(&root->rb);
1175 return rb_entry(root->left, struct cfq_queue, rb_node);
1180 static struct cfq_group *cfq_rb_first_group(struct cfq_rb_root *root)
1183 root->left = rb_first(&root->rb);
1186 return rb_entry_cfqg(root->left);
1191 static void rb_erase_init(struct rb_node *n, struct rb_root *root)
1197 static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
1199 if (root->left == n)
1201 rb_erase_init(n, &root->rb);
1206 * would be nice to take fifo expire time into account as well
1208 static struct request *
1209 cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1210 struct request *last)
1212 struct rb_node *rbnext = rb_next(&last->rb_node);
1213 struct rb_node *rbprev = rb_prev(&last->rb_node);
1214 struct request *next = NULL, *prev = NULL;
1216 BUG_ON(RB_EMPTY_NODE(&last->rb_node));
1219 prev = rb_entry_rq(rbprev);
1222 next = rb_entry_rq(rbnext);
1224 rbnext = rb_first(&cfqq->sort_list);
1225 if (rbnext && rbnext != &last->rb_node)
1226 next = rb_entry_rq(rbnext);
1229 return cfq_choose_req(cfqd, next, prev, blk_rq_pos(last));
1232 static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
1233 struct cfq_queue *cfqq)
1236 * just an approximation, should be ok.
1238 return (cfqq->cfqg->nr_cfqq - 1) * (cfq_prio_slice(cfqd, 1, 0) -
1239 cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
1243 cfqg_key(struct cfq_rb_root *st, struct cfq_group *cfqg)
1245 return cfqg->vdisktime - st->min_vdisktime;
1249 __cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
1251 struct rb_node **node = &st->rb.rb_node;
1252 struct rb_node *parent = NULL;
1253 struct cfq_group *__cfqg;
1254 s64 key = cfqg_key(st, cfqg);
1257 while (*node != NULL) {
1259 __cfqg = rb_entry_cfqg(parent);
1261 if (key < cfqg_key(st, __cfqg))
1262 node = &parent->rb_left;
1264 node = &parent->rb_right;
1270 st->left = &cfqg->rb_node;
1272 rb_link_node(&cfqg->rb_node, parent, node);
1273 rb_insert_color(&cfqg->rb_node, &st->rb);
1277 * This has to be called only on activation of cfqg
1280 cfq_update_group_weight(struct cfq_group *cfqg)
1282 if (cfqg->new_weight) {
1283 cfqg->weight = cfqg->new_weight;
1284 cfqg->new_weight = 0;
1289 cfq_update_group_leaf_weight(struct cfq_group *cfqg)
1291 BUG_ON(!RB_EMPTY_NODE(&cfqg->rb_node));
1293 if (cfqg->new_leaf_weight) {
1294 cfqg->leaf_weight = cfqg->new_leaf_weight;
1295 cfqg->new_leaf_weight = 0;
1300 cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
1302 unsigned int vfr = 1 << CFQ_SERVICE_SHIFT; /* start with 1 */
1303 struct cfq_group *pos = cfqg;
1304 struct cfq_group *parent;
1307 /* add to the service tree */
1308 BUG_ON(!RB_EMPTY_NODE(&cfqg->rb_node));
1311 * Update leaf_weight. We cannot update weight at this point
1312 * because cfqg might already have been activated and is
1313 * contributing its current weight to the parent's child_weight.
1315 cfq_update_group_leaf_weight(cfqg);
1316 __cfq_group_service_tree_add(st, cfqg);
1319 * Activate @cfqg and calculate the portion of vfraction @cfqg is
1320 * entitled to. vfraction is calculated by walking the tree
1321 * towards the root calculating the fraction it has at each level.
1322 * The compounded ratio is how much vfraction @cfqg owns.
1324 * Start with the proportion tasks in this cfqg has against active
1325 * children cfqgs - its leaf_weight against children_weight.
1327 propagate = !pos->nr_active++;
1328 pos->children_weight += pos->leaf_weight;
1329 vfr = vfr * pos->leaf_weight / pos->children_weight;
1332 * Compound ->weight walking up the tree. Both activation and
1333 * vfraction calculation are done in the same loop. Propagation
1334 * stops once an already activated node is met. vfraction
1335 * calculation should always continue to the root.
1337 while ((parent = cfqg_parent(pos))) {
1339 cfq_update_group_weight(pos);
1340 propagate = !parent->nr_active++;
1341 parent->children_weight += pos->weight;
1343 vfr = vfr * pos->weight / parent->children_weight;
1347 cfqg->vfraction = max_t(unsigned, vfr, 1);
1351 cfq_group_notify_queue_add(struct cfq_data *cfqd, struct cfq_group *cfqg)
1353 struct cfq_rb_root *st = &cfqd->grp_service_tree;
1354 struct cfq_group *__cfqg;
1358 if (!RB_EMPTY_NODE(&cfqg->rb_node))
1362 * Currently put the group at the end. Later implement something
1363 * so that groups get lesser vtime based on their weights, so that
1364 * if group does not loose all if it was not continuously backlogged.
1366 n = rb_last(&st->rb);
1368 __cfqg = rb_entry_cfqg(n);
1369 cfqg->vdisktime = __cfqg->vdisktime + CFQ_IDLE_DELAY;
1371 cfqg->vdisktime = st->min_vdisktime;
1372 cfq_group_service_tree_add(st, cfqg);
1376 cfq_group_service_tree_del(struct cfq_rb_root *st, struct cfq_group *cfqg)
1378 struct cfq_group *pos = cfqg;
1382 * Undo activation from cfq_group_service_tree_add(). Deactivate
1383 * @cfqg and propagate deactivation upwards.
1385 propagate = !--pos->nr_active;
1386 pos->children_weight -= pos->leaf_weight;
1389 struct cfq_group *parent = cfqg_parent(pos);
1391 /* @pos has 0 nr_active at this point */
1392 WARN_ON_ONCE(pos->children_weight);
1398 propagate = !--parent->nr_active;
1399 parent->children_weight -= pos->weight;
1403 /* remove from the service tree */
1404 if (!RB_EMPTY_NODE(&cfqg->rb_node))
1405 cfq_rb_erase(&cfqg->rb_node, st);
1409 cfq_group_notify_queue_del(struct cfq_data *cfqd, struct cfq_group *cfqg)
1411 struct cfq_rb_root *st = &cfqd->grp_service_tree;
1413 BUG_ON(cfqg->nr_cfqq < 1);
1416 /* If there are other cfq queues under this group, don't delete it */
1420 cfq_log_cfqg(cfqd, cfqg, "del_from_rr group");
1421 cfq_group_service_tree_del(st, cfqg);
1422 cfqg->saved_wl_slice = 0;
1423 cfqg_stats_update_dequeue(cfqg);
1426 static inline unsigned int cfq_cfqq_slice_usage(struct cfq_queue *cfqq,
1427 unsigned int *unaccounted_time)
1429 unsigned int slice_used;
1432 * Queue got expired before even a single request completed or
1433 * got expired immediately after first request completion.
1435 if (!cfqq->slice_start || cfqq->slice_start == jiffies) {
1437 * Also charge the seek time incurred to the group, otherwise
1438 * if there are mutiple queues in the group, each can dispatch
1439 * a single request on seeky media and cause lots of seek time
1440 * and group will never know it.
1442 slice_used = max_t(unsigned, (jiffies - cfqq->dispatch_start),
1445 slice_used = jiffies - cfqq->slice_start;
1446 if (slice_used > cfqq->allocated_slice) {
1447 *unaccounted_time = slice_used - cfqq->allocated_slice;
1448 slice_used = cfqq->allocated_slice;
1450 if (time_after(cfqq->slice_start, cfqq->dispatch_start))
1451 *unaccounted_time += cfqq->slice_start -
1452 cfqq->dispatch_start;
1458 static void cfq_group_served(struct cfq_data *cfqd, struct cfq_group *cfqg,
1459 struct cfq_queue *cfqq)
1461 struct cfq_rb_root *st = &cfqd->grp_service_tree;
1462 unsigned int used_sl, charge, unaccounted_sl = 0;
1463 int nr_sync = cfqg->nr_cfqq - cfqg_busy_async_queues(cfqd, cfqg)
1464 - cfqg->service_tree_idle.count;
1467 BUG_ON(nr_sync < 0);
1468 used_sl = charge = cfq_cfqq_slice_usage(cfqq, &unaccounted_sl);
1470 if (iops_mode(cfqd))
1471 charge = cfqq->slice_dispatch;
1472 else if (!cfq_cfqq_sync(cfqq) && !nr_sync)
1473 charge = cfqq->allocated_slice;
1476 * Can't update vdisktime while on service tree and cfqg->vfraction
1477 * is valid only while on it. Cache vfr, leave the service tree,
1478 * update vdisktime and go back on. The re-addition to the tree
1479 * will also update the weights as necessary.
1481 vfr = cfqg->vfraction;
1482 cfq_group_service_tree_del(st, cfqg);
1483 cfqg->vdisktime += cfqg_scale_charge(charge, vfr);
1484 cfq_group_service_tree_add(st, cfqg);
1486 /* This group is being expired. Save the context */
1487 if (time_after(cfqd->workload_expires, jiffies)) {
1488 cfqg->saved_wl_slice = cfqd->workload_expires
1490 cfqg->saved_wl_type = cfqd->serving_wl_type;
1491 cfqg->saved_wl_class = cfqd->serving_wl_class;
1493 cfqg->saved_wl_slice = 0;
1495 cfq_log_cfqg(cfqd, cfqg, "served: vt=%llu min_vt=%llu", cfqg->vdisktime,
1497 cfq_log_cfqq(cfqq->cfqd, cfqq,
1498 "sl_used=%u disp=%u charge=%u iops=%u sect=%lu",
1499 used_sl, cfqq->slice_dispatch, charge,
1500 iops_mode(cfqd), cfqq->nr_sectors);
1501 cfqg_stats_update_timeslice_used(cfqg, used_sl, unaccounted_sl);
1502 cfqg_stats_set_start_empty_time(cfqg);
1506 * cfq_init_cfqg_base - initialize base part of a cfq_group
1507 * @cfqg: cfq_group to initialize
1509 * Initialize the base part which is used whether %CONFIG_CFQ_GROUP_IOSCHED
1510 * is enabled or not.
1512 static void cfq_init_cfqg_base(struct cfq_group *cfqg)
1514 struct cfq_rb_root *st;
1517 for_each_cfqg_st(cfqg, i, j, st)
1519 RB_CLEAR_NODE(&cfqg->rb_node);
1521 cfqg->ttime.last_end_request = jiffies;
1524 #ifdef CONFIG_CFQ_GROUP_IOSCHED
1525 static void cfqg_stats_exit(struct cfqg_stats *stats)
1527 blkg_rwstat_exit(&stats->merged);
1528 blkg_rwstat_exit(&stats->service_time);
1529 blkg_rwstat_exit(&stats->wait_time);
1530 blkg_rwstat_exit(&stats->queued);
1531 blkg_stat_exit(&stats->time);
1532 #ifdef CONFIG_DEBUG_BLK_CGROUP
1533 blkg_stat_exit(&stats->unaccounted_time);
1534 blkg_stat_exit(&stats->avg_queue_size_sum);
1535 blkg_stat_exit(&stats->avg_queue_size_samples);
1536 blkg_stat_exit(&stats->dequeue);
1537 blkg_stat_exit(&stats->group_wait_time);
1538 blkg_stat_exit(&stats->idle_time);
1539 blkg_stat_exit(&stats->empty_time);
1543 static int cfqg_stats_init(struct cfqg_stats *stats, gfp_t gfp)
1545 if (blkg_rwstat_init(&stats->merged, gfp) ||
1546 blkg_rwstat_init(&stats->service_time, gfp) ||
1547 blkg_rwstat_init(&stats->wait_time, gfp) ||
1548 blkg_rwstat_init(&stats->queued, gfp) ||
1549 blkg_stat_init(&stats->time, gfp))
1552 #ifdef CONFIG_DEBUG_BLK_CGROUP
1553 if (blkg_stat_init(&stats->unaccounted_time, gfp) ||
1554 blkg_stat_init(&stats->avg_queue_size_sum, gfp) ||
1555 blkg_stat_init(&stats->avg_queue_size_samples, gfp) ||
1556 blkg_stat_init(&stats->dequeue, gfp) ||
1557 blkg_stat_init(&stats->group_wait_time, gfp) ||
1558 blkg_stat_init(&stats->idle_time, gfp) ||
1559 blkg_stat_init(&stats->empty_time, gfp))
1564 cfqg_stats_exit(stats);
1568 static struct blkcg_policy_data *cfq_cpd_alloc(gfp_t gfp)
1570 struct cfq_group_data *cgd;
1572 cgd = kzalloc(sizeof(*cgd), GFP_KERNEL);
1578 static void cfq_cpd_init(struct blkcg_policy_data *cpd)
1580 struct cfq_group_data *cgd = cpd_to_cfqgd(cpd);
1582 if (cpd_to_blkcg(cpd) == &blkcg_root) {
1583 cgd->weight = 2 * CFQ_WEIGHT_DEFAULT;
1584 cgd->leaf_weight = 2 * CFQ_WEIGHT_DEFAULT;
1586 cgd->weight = CFQ_WEIGHT_DEFAULT;
1587 cgd->leaf_weight = CFQ_WEIGHT_DEFAULT;
1591 static void cfq_cpd_free(struct blkcg_policy_data *cpd)
1593 kfree(cpd_to_cfqgd(cpd));
1596 static struct blkg_policy_data *cfq_pd_alloc(gfp_t gfp, int node)
1598 struct cfq_group *cfqg;
1600 cfqg = kzalloc_node(sizeof(*cfqg), gfp, node);
1604 cfq_init_cfqg_base(cfqg);
1605 if (cfqg_stats_init(&cfqg->stats, gfp)) {
1613 static void cfq_pd_init(struct blkg_policy_data *pd)
1615 struct cfq_group *cfqg = pd_to_cfqg(pd);
1616 struct cfq_group_data *cgd = blkcg_to_cfqgd(pd->blkg->blkcg);
1618 cfqg->weight = cgd->weight;
1619 cfqg->leaf_weight = cgd->leaf_weight;
1622 static void cfq_pd_offline(struct blkg_policy_data *pd)
1624 struct cfq_group *cfqg = pd_to_cfqg(pd);
1627 for (i = 0; i < IOPRIO_BE_NR; i++) {
1628 if (cfqg->async_cfqq[0][i])
1629 cfq_put_queue(cfqg->async_cfqq[0][i]);
1630 if (cfqg->async_cfqq[1][i])
1631 cfq_put_queue(cfqg->async_cfqq[1][i]);
1634 if (cfqg->async_idle_cfqq)
1635 cfq_put_queue(cfqg->async_idle_cfqq);
1638 * @blkg is going offline and will be ignored by
1639 * blkg_[rw]stat_recursive_sum(). Transfer stats to the parent so
1640 * that they don't get lost. If IOs complete after this point, the
1641 * stats for them will be lost. Oh well...
1643 cfqg_stats_xfer_dead(cfqg);
1646 static void cfq_pd_free(struct blkg_policy_data *pd)
1648 struct cfq_group *cfqg = pd_to_cfqg(pd);
1650 cfqg_stats_exit(&cfqg->stats);
1654 static void cfq_pd_reset_stats(struct blkg_policy_data *pd)
1656 struct cfq_group *cfqg = pd_to_cfqg(pd);
1658 cfqg_stats_reset(&cfqg->stats);
1661 static struct cfq_group *cfq_lookup_cfqg(struct cfq_data *cfqd,
1662 struct blkcg *blkcg)
1664 struct blkcg_gq *blkg;
1666 blkg = blkg_lookup(blkcg, cfqd->queue);
1668 return blkg_to_cfqg(blkg);
1672 static void cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg)
1675 /* cfqq reference on cfqg */
1679 static u64 cfqg_prfill_weight_device(struct seq_file *sf,
1680 struct blkg_policy_data *pd, int off)
1682 struct cfq_group *cfqg = pd_to_cfqg(pd);
1684 if (!cfqg->dev_weight)
1686 return __blkg_prfill_u64(sf, pd, cfqg->dev_weight);
1689 static int cfqg_print_weight_device(struct seq_file *sf, void *v)
1691 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1692 cfqg_prfill_weight_device, &blkcg_policy_cfq,
1697 static u64 cfqg_prfill_leaf_weight_device(struct seq_file *sf,
1698 struct blkg_policy_data *pd, int off)
1700 struct cfq_group *cfqg = pd_to_cfqg(pd);
1702 if (!cfqg->dev_leaf_weight)
1704 return __blkg_prfill_u64(sf, pd, cfqg->dev_leaf_weight);
1707 static int cfqg_print_leaf_weight_device(struct seq_file *sf, void *v)
1709 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1710 cfqg_prfill_leaf_weight_device, &blkcg_policy_cfq,
1715 static int cfq_print_weight(struct seq_file *sf, void *v)
1717 struct blkcg *blkcg = css_to_blkcg(seq_css(sf));
1718 struct cfq_group_data *cgd = blkcg_to_cfqgd(blkcg);
1719 unsigned int val = 0;
1724 seq_printf(sf, "%u\n", val);
1728 static int cfq_print_leaf_weight(struct seq_file *sf, void *v)
1730 struct blkcg *blkcg = css_to_blkcg(seq_css(sf));
1731 struct cfq_group_data *cgd = blkcg_to_cfqgd(blkcg);
1732 unsigned int val = 0;
1735 val = cgd->leaf_weight;
1737 seq_printf(sf, "%u\n", val);
1741 static ssize_t __cfqg_set_weight_device(struct kernfs_open_file *of,
1742 char *buf, size_t nbytes, loff_t off,
1743 bool is_leaf_weight)
1745 struct blkcg *blkcg = css_to_blkcg(of_css(of));
1746 struct blkg_conf_ctx ctx;
1747 struct cfq_group *cfqg;
1748 struct cfq_group_data *cfqgd;
1752 ret = blkg_conf_prep(blkcg, &blkcg_policy_cfq, buf, &ctx);
1757 if (sscanf(ctx.body, "%llu", &v) != 1)
1760 cfqg = blkg_to_cfqg(ctx.blkg);
1761 cfqgd = blkcg_to_cfqgd(blkcg);
1764 if (!v || (v >= CFQ_WEIGHT_MIN && v <= CFQ_WEIGHT_MAX)) {
1765 if (!is_leaf_weight) {
1766 cfqg->dev_weight = v;
1767 cfqg->new_weight = v ?: cfqgd->weight;
1769 cfqg->dev_leaf_weight = v;
1770 cfqg->new_leaf_weight = v ?: cfqgd->leaf_weight;
1775 blkg_conf_finish(&ctx);
1776 return ret ?: nbytes;
1779 static ssize_t cfqg_set_weight_device(struct kernfs_open_file *of,
1780 char *buf, size_t nbytes, loff_t off)
1782 return __cfqg_set_weight_device(of, buf, nbytes, off, false);
1785 static ssize_t cfqg_set_leaf_weight_device(struct kernfs_open_file *of,
1786 char *buf, size_t nbytes, loff_t off)
1788 return __cfqg_set_weight_device(of, buf, nbytes, off, true);
1791 static int __cfq_set_weight(struct cgroup_subsys_state *css, u64 val,
1792 bool is_leaf_weight)
1794 struct blkcg *blkcg = css_to_blkcg(css);
1795 struct blkcg_gq *blkg;
1796 struct cfq_group_data *cfqgd;
1799 if (val < CFQ_WEIGHT_MIN || val > CFQ_WEIGHT_MAX)
1802 spin_lock_irq(&blkcg->lock);
1803 cfqgd = blkcg_to_cfqgd(blkcg);
1809 if (!is_leaf_weight)
1810 cfqgd->weight = val;
1812 cfqgd->leaf_weight = val;
1814 hlist_for_each_entry(blkg, &blkcg->blkg_list, blkcg_node) {
1815 struct cfq_group *cfqg = blkg_to_cfqg(blkg);
1820 if (!is_leaf_weight) {
1821 if (!cfqg->dev_weight)
1822 cfqg->new_weight = cfqgd->weight;
1824 if (!cfqg->dev_leaf_weight)
1825 cfqg->new_leaf_weight = cfqgd->leaf_weight;
1830 spin_unlock_irq(&blkcg->lock);
1834 static int cfq_set_weight(struct cgroup_subsys_state *css, struct cftype *cft,
1837 return __cfq_set_weight(css, val, false);
1840 static int cfq_set_leaf_weight(struct cgroup_subsys_state *css,
1841 struct cftype *cft, u64 val)
1843 return __cfq_set_weight(css, val, true);
1846 static int cfqg_print_stat(struct seq_file *sf, void *v)
1848 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), blkg_prfill_stat,
1849 &blkcg_policy_cfq, seq_cft(sf)->private, false);
1853 static int cfqg_print_rwstat(struct seq_file *sf, void *v)
1855 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), blkg_prfill_rwstat,
1856 &blkcg_policy_cfq, seq_cft(sf)->private, true);
1860 static u64 cfqg_prfill_stat_recursive(struct seq_file *sf,
1861 struct blkg_policy_data *pd, int off)
1863 u64 sum = blkg_stat_recursive_sum(pd_to_blkg(pd),
1864 &blkcg_policy_cfq, off);
1865 return __blkg_prfill_u64(sf, pd, sum);
1868 static u64 cfqg_prfill_rwstat_recursive(struct seq_file *sf,
1869 struct blkg_policy_data *pd, int off)
1871 struct blkg_rwstat sum = blkg_rwstat_recursive_sum(pd_to_blkg(pd),
1872 &blkcg_policy_cfq, off);
1873 return __blkg_prfill_rwstat(sf, pd, &sum);
1876 static int cfqg_print_stat_recursive(struct seq_file *sf, void *v)
1878 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1879 cfqg_prfill_stat_recursive, &blkcg_policy_cfq,
1880 seq_cft(sf)->private, false);
1884 static int cfqg_print_rwstat_recursive(struct seq_file *sf, void *v)
1886 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1887 cfqg_prfill_rwstat_recursive, &blkcg_policy_cfq,
1888 seq_cft(sf)->private, true);
1892 static u64 cfqg_prfill_sectors(struct seq_file *sf, struct blkg_policy_data *pd,
1895 u64 sum = blkg_rwstat_total(&pd->blkg->stat_bytes);
1897 return __blkg_prfill_u64(sf, pd, sum >> 9);
1900 static int cfqg_print_stat_sectors(struct seq_file *sf, void *v)
1902 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1903 cfqg_prfill_sectors, &blkcg_policy_cfq, 0, false);
1907 static u64 cfqg_prfill_sectors_recursive(struct seq_file *sf,
1908 struct blkg_policy_data *pd, int off)
1910 struct blkg_rwstat tmp = blkg_rwstat_recursive_sum(pd->blkg, NULL,
1911 offsetof(struct blkcg_gq, stat_bytes));
1912 u64 sum = atomic64_read(&tmp.aux_cnt[BLKG_RWSTAT_READ]) +
1913 atomic64_read(&tmp.aux_cnt[BLKG_RWSTAT_WRITE]);
1915 return __blkg_prfill_u64(sf, pd, sum >> 9);
1918 static int cfqg_print_stat_sectors_recursive(struct seq_file *sf, void *v)
1920 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1921 cfqg_prfill_sectors_recursive, &blkcg_policy_cfq, 0,
1926 #ifdef CONFIG_DEBUG_BLK_CGROUP
1927 static u64 cfqg_prfill_avg_queue_size(struct seq_file *sf,
1928 struct blkg_policy_data *pd, int off)
1930 struct cfq_group *cfqg = pd_to_cfqg(pd);
1931 u64 samples = blkg_stat_read(&cfqg->stats.avg_queue_size_samples);
1935 v = blkg_stat_read(&cfqg->stats.avg_queue_size_sum);
1936 v = div64_u64(v, samples);
1938 __blkg_prfill_u64(sf, pd, v);
1942 /* print avg_queue_size */
1943 static int cfqg_print_avg_queue_size(struct seq_file *sf, void *v)
1945 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1946 cfqg_prfill_avg_queue_size, &blkcg_policy_cfq,
1950 #endif /* CONFIG_DEBUG_BLK_CGROUP */
1952 static struct cftype cfq_blkcg_legacy_files[] = {
1953 /* on root, weight is mapped to leaf_weight */
1955 .name = "weight_device",
1956 .flags = CFTYPE_ONLY_ON_ROOT,
1957 .seq_show = cfqg_print_leaf_weight_device,
1958 .write = cfqg_set_leaf_weight_device,
1962 .flags = CFTYPE_ONLY_ON_ROOT,
1963 .seq_show = cfq_print_leaf_weight,
1964 .write_u64 = cfq_set_leaf_weight,
1967 /* no such mapping necessary for !roots */
1969 .name = "weight_device",
1970 .flags = CFTYPE_NOT_ON_ROOT,
1971 .seq_show = cfqg_print_weight_device,
1972 .write = cfqg_set_weight_device,
1976 .flags = CFTYPE_NOT_ON_ROOT,
1977 .seq_show = cfq_print_weight,
1978 .write_u64 = cfq_set_weight,
1982 .name = "leaf_weight_device",
1983 .seq_show = cfqg_print_leaf_weight_device,
1984 .write = cfqg_set_leaf_weight_device,
1987 .name = "leaf_weight",
1988 .seq_show = cfq_print_leaf_weight,
1989 .write_u64 = cfq_set_leaf_weight,
1992 /* statistics, covers only the tasks in the cfqg */
1995 .private = offsetof(struct cfq_group, stats.time),
1996 .seq_show = cfqg_print_stat,
2000 .seq_show = cfqg_print_stat_sectors,
2003 .name = "io_service_bytes",
2004 .private = (unsigned long)&blkcg_policy_cfq,
2005 .seq_show = blkg_print_stat_bytes,
2008 .name = "io_serviced",
2009 .private = (unsigned long)&blkcg_policy_cfq,
2010 .seq_show = blkg_print_stat_ios,
2013 .name = "io_service_time",
2014 .private = offsetof(struct cfq_group, stats.service_time),
2015 .seq_show = cfqg_print_rwstat,
2018 .name = "io_wait_time",
2019 .private = offsetof(struct cfq_group, stats.wait_time),
2020 .seq_show = cfqg_print_rwstat,
2023 .name = "io_merged",
2024 .private = offsetof(struct cfq_group, stats.merged),
2025 .seq_show = cfqg_print_rwstat,
2028 .name = "io_queued",
2029 .private = offsetof(struct cfq_group, stats.queued),
2030 .seq_show = cfqg_print_rwstat,
2033 /* the same statictics which cover the cfqg and its descendants */
2035 .name = "time_recursive",
2036 .private = offsetof(struct cfq_group, stats.time),
2037 .seq_show = cfqg_print_stat_recursive,
2040 .name = "sectors_recursive",
2041 .seq_show = cfqg_print_stat_sectors_recursive,
2044 .name = "io_service_bytes_recursive",
2045 .private = (unsigned long)&blkcg_policy_cfq,
2046 .seq_show = blkg_print_stat_bytes_recursive,
2049 .name = "io_serviced_recursive",
2050 .private = (unsigned long)&blkcg_policy_cfq,
2051 .seq_show = blkg_print_stat_ios_recursive,
2054 .name = "io_service_time_recursive",
2055 .private = offsetof(struct cfq_group, stats.service_time),
2056 .seq_show = cfqg_print_rwstat_recursive,
2059 .name = "io_wait_time_recursive",
2060 .private = offsetof(struct cfq_group, stats.wait_time),
2061 .seq_show = cfqg_print_rwstat_recursive,
2064 .name = "io_merged_recursive",
2065 .private = offsetof(struct cfq_group, stats.merged),
2066 .seq_show = cfqg_print_rwstat_recursive,
2069 .name = "io_queued_recursive",
2070 .private = offsetof(struct cfq_group, stats.queued),
2071 .seq_show = cfqg_print_rwstat_recursive,
2073 #ifdef CONFIG_DEBUG_BLK_CGROUP
2075 .name = "avg_queue_size",
2076 .seq_show = cfqg_print_avg_queue_size,
2079 .name = "group_wait_time",
2080 .private = offsetof(struct cfq_group, stats.group_wait_time),
2081 .seq_show = cfqg_print_stat,
2084 .name = "idle_time",
2085 .private = offsetof(struct cfq_group, stats.idle_time),
2086 .seq_show = cfqg_print_stat,
2089 .name = "empty_time",
2090 .private = offsetof(struct cfq_group, stats.empty_time),
2091 .seq_show = cfqg_print_stat,
2095 .private = offsetof(struct cfq_group, stats.dequeue),
2096 .seq_show = cfqg_print_stat,
2099 .name = "unaccounted_time",
2100 .private = offsetof(struct cfq_group, stats.unaccounted_time),
2101 .seq_show = cfqg_print_stat,
2103 #endif /* CONFIG_DEBUG_BLK_CGROUP */
2106 #else /* GROUP_IOSCHED */
2107 static struct cfq_group *cfq_lookup_cfqg(struct cfq_data *cfqd,
2108 struct blkcg *blkcg)
2110 return cfqd->root_group;
2114 cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) {
2118 #endif /* GROUP_IOSCHED */
2121 * The cfqd->service_trees holds all pending cfq_queue's that have
2122 * requests waiting to be processed. It is sorted in the order that
2123 * we will service the queues.
2125 static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2128 struct rb_node **p, *parent;
2129 struct cfq_queue *__cfqq;
2130 unsigned long rb_key;
2131 struct cfq_rb_root *st;
2135 st = st_for(cfqq->cfqg, cfqq_class(cfqq), cfqq_type(cfqq));
2136 if (cfq_class_idle(cfqq)) {
2137 rb_key = CFQ_IDLE_DELAY;
2138 parent = rb_last(&st->rb);
2139 if (parent && parent != &cfqq->rb_node) {
2140 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
2141 rb_key += __cfqq->rb_key;
2144 } else if (!add_front) {
2146 * Get our rb key offset. Subtract any residual slice
2147 * value carried from last service. A negative resid
2148 * count indicates slice overrun, and this should position
2149 * the next service time further away in the tree.
2151 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
2152 rb_key -= cfqq->slice_resid;
2153 cfqq->slice_resid = 0;
2156 __cfqq = cfq_rb_first(st);
2157 rb_key += __cfqq ? __cfqq->rb_key : jiffies;
2160 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
2163 * same position, nothing more to do
2165 if (rb_key == cfqq->rb_key && cfqq->service_tree == st)
2168 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
2169 cfqq->service_tree = NULL;
2174 cfqq->service_tree = st;
2175 p = &st->rb.rb_node;
2178 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
2181 * sort by key, that represents service time.
2183 if (time_before(rb_key, __cfqq->rb_key))
2184 p = &parent->rb_left;
2186 p = &parent->rb_right;
2192 st->left = &cfqq->rb_node;
2194 cfqq->rb_key = rb_key;
2195 rb_link_node(&cfqq->rb_node, parent, p);
2196 rb_insert_color(&cfqq->rb_node, &st->rb);
2198 if (add_front || !new_cfqq)
2200 cfq_group_notify_queue_add(cfqd, cfqq->cfqg);
2203 static struct cfq_queue *
2204 cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
2205 sector_t sector, struct rb_node **ret_parent,
2206 struct rb_node ***rb_link)
2208 struct rb_node **p, *parent;
2209 struct cfq_queue *cfqq = NULL;
2217 cfqq = rb_entry(parent, struct cfq_queue, p_node);
2220 * Sort strictly based on sector. Smallest to the left,
2221 * largest to the right.
2223 if (sector > blk_rq_pos(cfqq->next_rq))
2224 n = &(*p)->rb_right;
2225 else if (sector < blk_rq_pos(cfqq->next_rq))
2233 *ret_parent = parent;
2239 static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2241 struct rb_node **p, *parent;
2242 struct cfq_queue *__cfqq;
2245 rb_erase(&cfqq->p_node, cfqq->p_root);
2246 cfqq->p_root = NULL;
2249 if (cfq_class_idle(cfqq))
2254 cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
2255 __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
2256 blk_rq_pos(cfqq->next_rq), &parent, &p);
2258 rb_link_node(&cfqq->p_node, parent, p);
2259 rb_insert_color(&cfqq->p_node, cfqq->p_root);
2261 cfqq->p_root = NULL;
2265 * Update cfqq's position in the service tree.
2267 static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2270 * Resorting requires the cfqq to be on the RR list already.
2272 if (cfq_cfqq_on_rr(cfqq)) {
2273 cfq_service_tree_add(cfqd, cfqq, 0);
2274 cfq_prio_tree_add(cfqd, cfqq);
2279 * add to busy list of queues for service, trying to be fair in ordering
2280 * the pending list according to last request service
2282 static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2284 cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
2285 BUG_ON(cfq_cfqq_on_rr(cfqq));
2286 cfq_mark_cfqq_on_rr(cfqq);
2287 cfqd->busy_queues++;
2288 if (cfq_cfqq_sync(cfqq))
2289 cfqd->busy_sync_queues++;
2291 cfq_resort_rr_list(cfqd, cfqq);
2295 * Called when the cfqq no longer has requests pending, remove it from
2298 static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2300 cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
2301 BUG_ON(!cfq_cfqq_on_rr(cfqq));
2302 cfq_clear_cfqq_on_rr(cfqq);
2304 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
2305 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
2306 cfqq->service_tree = NULL;
2309 rb_erase(&cfqq->p_node, cfqq->p_root);
2310 cfqq->p_root = NULL;
2313 cfq_group_notify_queue_del(cfqd, cfqq->cfqg);
2314 BUG_ON(!cfqd->busy_queues);
2315 cfqd->busy_queues--;
2316 if (cfq_cfqq_sync(cfqq))
2317 cfqd->busy_sync_queues--;
2321 * rb tree support functions
2323 static void cfq_del_rq_rb(struct request *rq)
2325 struct cfq_queue *cfqq = RQ_CFQQ(rq);
2326 const int sync = rq_is_sync(rq);
2328 BUG_ON(!cfqq->queued[sync]);
2329 cfqq->queued[sync]--;
2331 elv_rb_del(&cfqq->sort_list, rq);
2333 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) {
2335 * Queue will be deleted from service tree when we actually
2336 * expire it later. Right now just remove it from prio tree
2340 rb_erase(&cfqq->p_node, cfqq->p_root);
2341 cfqq->p_root = NULL;
2346 static void cfq_add_rq_rb(struct request *rq)
2348 struct cfq_queue *cfqq = RQ_CFQQ(rq);
2349 struct cfq_data *cfqd = cfqq->cfqd;
2350 struct request *prev;
2352 cfqq->queued[rq_is_sync(rq)]++;
2354 elv_rb_add(&cfqq->sort_list, rq);
2356 if (!cfq_cfqq_on_rr(cfqq))
2357 cfq_add_cfqq_rr(cfqd, cfqq);
2360 * check if this request is a better next-serve candidate
2362 prev = cfqq->next_rq;
2363 cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq, cfqd->last_position);
2366 * adjust priority tree position, if ->next_rq changes
2368 if (prev != cfqq->next_rq)
2369 cfq_prio_tree_add(cfqd, cfqq);
2371 BUG_ON(!cfqq->next_rq);
2374 static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
2376 elv_rb_del(&cfqq->sort_list, rq);
2377 cfqq->queued[rq_is_sync(rq)]--;
2378 cfqg_stats_update_io_remove(RQ_CFQG(rq), rq->cmd_flags);
2380 cfqg_stats_update_io_add(RQ_CFQG(rq), cfqq->cfqd->serving_group,
2384 static struct request *
2385 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
2387 struct task_struct *tsk = current;
2388 struct cfq_io_cq *cic;
2389 struct cfq_queue *cfqq;
2391 cic = cfq_cic_lookup(cfqd, tsk->io_context);
2395 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
2397 return elv_rb_find(&cfqq->sort_list, bio_end_sector(bio));
2402 static void cfq_activate_request(struct request_queue *q, struct request *rq)
2404 struct cfq_data *cfqd = q->elevator->elevator_data;
2406 cfqd->rq_in_driver++;
2407 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
2408 cfqd->rq_in_driver);
2410 cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
2413 static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
2415 struct cfq_data *cfqd = q->elevator->elevator_data;
2417 WARN_ON(!cfqd->rq_in_driver);
2418 cfqd->rq_in_driver--;
2419 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
2420 cfqd->rq_in_driver);
2423 static void cfq_remove_request(struct request *rq)
2425 struct cfq_queue *cfqq = RQ_CFQQ(rq);
2427 if (cfqq->next_rq == rq)
2428 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
2430 list_del_init(&rq->queuelist);
2433 cfqq->cfqd->rq_queued--;
2434 cfqg_stats_update_io_remove(RQ_CFQG(rq), rq->cmd_flags);
2435 if (rq->cmd_flags & REQ_PRIO) {
2436 WARN_ON(!cfqq->prio_pending);
2437 cfqq->prio_pending--;
2441 static int cfq_merge(struct request_queue *q, struct request **req,
2444 struct cfq_data *cfqd = q->elevator->elevator_data;
2445 struct request *__rq;
2447 __rq = cfq_find_rq_fmerge(cfqd, bio);
2448 if (__rq && elv_rq_merge_ok(__rq, bio)) {
2450 return ELEVATOR_FRONT_MERGE;
2453 return ELEVATOR_NO_MERGE;
2456 static void cfq_merged_request(struct request_queue *q, struct request *req,
2459 if (type == ELEVATOR_FRONT_MERGE) {
2460 struct cfq_queue *cfqq = RQ_CFQQ(req);
2462 cfq_reposition_rq_rb(cfqq, req);
2466 static void cfq_bio_merged(struct request_queue *q, struct request *req,
2469 cfqg_stats_update_io_merged(RQ_CFQG(req), bio->bi_rw);
2473 cfq_merged_requests(struct request_queue *q, struct request *rq,
2474 struct request *next)
2476 struct cfq_queue *cfqq = RQ_CFQQ(rq);
2477 struct cfq_data *cfqd = q->elevator->elevator_data;
2480 * reposition in fifo if next is older than rq
2482 if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
2483 time_before(next->fifo_time, rq->fifo_time) &&
2484 cfqq == RQ_CFQQ(next)) {
2485 list_move(&rq->queuelist, &next->queuelist);
2486 rq->fifo_time = next->fifo_time;
2489 if (cfqq->next_rq == next)
2491 cfq_remove_request(next);
2492 cfqg_stats_update_io_merged(RQ_CFQG(rq), next->cmd_flags);
2494 cfqq = RQ_CFQQ(next);
2496 * all requests of this queue are merged to other queues, delete it
2497 * from the service tree. If it's the active_queue,
2498 * cfq_dispatch_requests() will choose to expire it or do idle
2500 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list) &&
2501 cfqq != cfqd->active_queue)
2502 cfq_del_cfqq_rr(cfqd, cfqq);
2505 static int cfq_allow_merge(struct request_queue *q, struct request *rq,
2508 struct cfq_data *cfqd = q->elevator->elevator_data;
2509 struct cfq_io_cq *cic;
2510 struct cfq_queue *cfqq;
2513 * Disallow merge of a sync bio into an async request.
2515 if (cfq_bio_sync(bio) && !rq_is_sync(rq))
2519 * Lookup the cfqq that this bio will be queued with and allow
2520 * merge only if rq is queued there.
2522 cic = cfq_cic_lookup(cfqd, current->io_context);
2526 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
2527 return cfqq == RQ_CFQQ(rq);
2530 static inline void cfq_del_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2532 del_timer(&cfqd->idle_slice_timer);
2533 cfqg_stats_update_idle_time(cfqq->cfqg);
2536 static void __cfq_set_active_queue(struct cfq_data *cfqd,
2537 struct cfq_queue *cfqq)
2540 cfq_log_cfqq(cfqd, cfqq, "set_active wl_class:%d wl_type:%d",
2541 cfqd->serving_wl_class, cfqd->serving_wl_type);
2542 cfqg_stats_update_avg_queue_size(cfqq->cfqg);
2543 cfqq->slice_start = 0;
2544 cfqq->dispatch_start = jiffies;
2545 cfqq->allocated_slice = 0;
2546 cfqq->slice_end = 0;
2547 cfqq->slice_dispatch = 0;
2548 cfqq->nr_sectors = 0;
2550 cfq_clear_cfqq_wait_request(cfqq);
2551 cfq_clear_cfqq_must_dispatch(cfqq);
2552 cfq_clear_cfqq_must_alloc_slice(cfqq);
2553 cfq_clear_cfqq_fifo_expire(cfqq);
2554 cfq_mark_cfqq_slice_new(cfqq);
2556 cfq_del_timer(cfqd, cfqq);
2559 cfqd->active_queue = cfqq;
2563 * current cfqq expired its slice (or was too idle), select new one
2566 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2569 cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
2571 if (cfq_cfqq_wait_request(cfqq))
2572 cfq_del_timer(cfqd, cfqq);
2574 cfq_clear_cfqq_wait_request(cfqq);
2575 cfq_clear_cfqq_wait_busy(cfqq);
2578 * If this cfqq is shared between multiple processes, check to
2579 * make sure that those processes are still issuing I/Os within
2580 * the mean seek distance. If not, it may be time to break the
2581 * queues apart again.
2583 if (cfq_cfqq_coop(cfqq) && CFQQ_SEEKY(cfqq))
2584 cfq_mark_cfqq_split_coop(cfqq);
2587 * store what was left of this slice, if the queue idled/timed out
2590 if (cfq_cfqq_slice_new(cfqq))
2591 cfqq->slice_resid = cfq_scaled_cfqq_slice(cfqd, cfqq);
2593 cfqq->slice_resid = cfqq->slice_end - jiffies;
2594 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
2597 cfq_group_served(cfqd, cfqq->cfqg, cfqq);
2599 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
2600 cfq_del_cfqq_rr(cfqd, cfqq);
2602 cfq_resort_rr_list(cfqd, cfqq);
2604 if (cfqq == cfqd->active_queue)
2605 cfqd->active_queue = NULL;
2607 if (cfqd->active_cic) {
2608 put_io_context(cfqd->active_cic->icq.ioc);
2609 cfqd->active_cic = NULL;
2613 static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out)
2615 struct cfq_queue *cfqq = cfqd->active_queue;
2618 __cfq_slice_expired(cfqd, cfqq, timed_out);
2622 * Get next queue for service. Unless we have a queue preemption,
2623 * we'll simply select the first cfqq in the service tree.
2625 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
2627 struct cfq_rb_root *st = st_for(cfqd->serving_group,
2628 cfqd->serving_wl_class, cfqd->serving_wl_type);
2630 if (!cfqd->rq_queued)
2633 /* There is nothing to dispatch */
2636 if (RB_EMPTY_ROOT(&st->rb))
2638 return cfq_rb_first(st);
2641 static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd)
2643 struct cfq_group *cfqg;
2644 struct cfq_queue *cfqq;
2646 struct cfq_rb_root *st;
2648 if (!cfqd->rq_queued)
2651 cfqg = cfq_get_next_cfqg(cfqd);
2655 for_each_cfqg_st(cfqg, i, j, st)
2656 if ((cfqq = cfq_rb_first(st)) != NULL)
2662 * Get and set a new active queue for service.
2664 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
2665 struct cfq_queue *cfqq)
2668 cfqq = cfq_get_next_queue(cfqd);
2670 __cfq_set_active_queue(cfqd, cfqq);
2674 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
2677 if (blk_rq_pos(rq) >= cfqd->last_position)
2678 return blk_rq_pos(rq) - cfqd->last_position;
2680 return cfqd->last_position - blk_rq_pos(rq);
2683 static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2686 return cfq_dist_from_last(cfqd, rq) <= CFQQ_CLOSE_THR;
2689 static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
2690 struct cfq_queue *cur_cfqq)
2692 struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
2693 struct rb_node *parent, *node;
2694 struct cfq_queue *__cfqq;
2695 sector_t sector = cfqd->last_position;
2697 if (RB_EMPTY_ROOT(root))
2701 * First, if we find a request starting at the end of the last
2702 * request, choose it.
2704 __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
2709 * If the exact sector wasn't found, the parent of the NULL leaf
2710 * will contain the closest sector.
2712 __cfqq = rb_entry(parent, struct cfq_queue, p_node);
2713 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
2716 if (blk_rq_pos(__cfqq->next_rq) < sector)
2717 node = rb_next(&__cfqq->p_node);
2719 node = rb_prev(&__cfqq->p_node);
2723 __cfqq = rb_entry(node, struct cfq_queue, p_node);
2724 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
2732 * cur_cfqq - passed in so that we don't decide that the current queue is
2733 * closely cooperating with itself.
2735 * So, basically we're assuming that that cur_cfqq has dispatched at least
2736 * one request, and that cfqd->last_position reflects a position on the disk
2737 * associated with the I/O issued by cur_cfqq. I'm not sure this is a valid
2740 static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
2741 struct cfq_queue *cur_cfqq)
2743 struct cfq_queue *cfqq;
2745 if (cfq_class_idle(cur_cfqq))
2747 if (!cfq_cfqq_sync(cur_cfqq))
2749 if (CFQQ_SEEKY(cur_cfqq))
2753 * Don't search priority tree if it's the only queue in the group.
2755 if (cur_cfqq->cfqg->nr_cfqq == 1)
2759 * We should notice if some of the queues are cooperating, eg
2760 * working closely on the same area of the disk. In that case,
2761 * we can group them together and don't waste time idling.
2763 cfqq = cfqq_close(cfqd, cur_cfqq);
2767 /* If new queue belongs to different cfq_group, don't choose it */
2768 if (cur_cfqq->cfqg != cfqq->cfqg)
2772 * It only makes sense to merge sync queues.
2774 if (!cfq_cfqq_sync(cfqq))
2776 if (CFQQ_SEEKY(cfqq))
2780 * Do not merge queues of different priority classes
2782 if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq))
2789 * Determine whether we should enforce idle window for this queue.
2792 static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2794 enum wl_class_t wl_class = cfqq_class(cfqq);
2795 struct cfq_rb_root *st = cfqq->service_tree;
2800 if (!cfqd->cfq_slice_idle)
2803 /* We never do for idle class queues. */
2804 if (wl_class == IDLE_WORKLOAD)
2807 /* We do for queues that were marked with idle window flag. */
2808 if (cfq_cfqq_idle_window(cfqq) &&
2809 !(blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag))
2813 * Otherwise, we do only if they are the last ones
2814 * in their service tree.
2816 if (st->count == 1 && cfq_cfqq_sync(cfqq) &&
2817 !cfq_io_thinktime_big(cfqd, &st->ttime, false))
2819 cfq_log_cfqq(cfqd, cfqq, "Not idling. st->count:%d", st->count);
2823 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
2825 struct cfq_queue *cfqq = cfqd->active_queue;
2826 struct cfq_io_cq *cic;
2827 unsigned long sl, group_idle = 0;
2830 * SSD device without seek penalty, disable idling. But only do so
2831 * for devices that support queuing, otherwise we still have a problem
2832 * with sync vs async workloads.
2834 if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
2837 WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
2838 WARN_ON(cfq_cfqq_slice_new(cfqq));
2841 * idle is disabled, either manually or by past process history
2843 if (!cfq_should_idle(cfqd, cfqq)) {
2844 /* no queue idling. Check for group idling */
2845 if (cfqd->cfq_group_idle)
2846 group_idle = cfqd->cfq_group_idle;
2852 * still active requests from this queue, don't idle
2854 if (cfqq->dispatched)
2858 * task has exited, don't wait
2860 cic = cfqd->active_cic;
2861 if (!cic || !atomic_read(&cic->icq.ioc->active_ref))
2865 * If our average think time is larger than the remaining time
2866 * slice, then don't idle. This avoids overrunning the allotted
2869 if (sample_valid(cic->ttime.ttime_samples) &&
2870 (cfqq->slice_end - jiffies < cic->ttime.ttime_mean)) {
2871 cfq_log_cfqq(cfqd, cfqq, "Not idling. think_time:%lu",
2872 cic->ttime.ttime_mean);
2876 /* There are other queues in the group, don't do group idle */
2877 if (group_idle && cfqq->cfqg->nr_cfqq > 1)
2880 cfq_mark_cfqq_wait_request(cfqq);
2883 sl = cfqd->cfq_group_idle;
2885 sl = cfqd->cfq_slice_idle;
2887 mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
2888 cfqg_stats_set_start_idle_time(cfqq->cfqg);
2889 cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu group_idle: %d", sl,
2890 group_idle ? 1 : 0);
2894 * Move request from internal lists to the request queue dispatch list.
2896 static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
2898 struct cfq_data *cfqd = q->elevator->elevator_data;
2899 struct cfq_queue *cfqq = RQ_CFQQ(rq);
2901 cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
2903 cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
2904 cfq_remove_request(rq);
2906 (RQ_CFQG(rq))->dispatched++;
2907 elv_dispatch_sort(q, rq);
2909 cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]++;
2910 cfqq->nr_sectors += blk_rq_sectors(rq);
2914 * return expired entry, or NULL to just start from scratch in rbtree
2916 static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
2918 struct request *rq = NULL;
2920 if (cfq_cfqq_fifo_expire(cfqq))
2923 cfq_mark_cfqq_fifo_expire(cfqq);
2925 if (list_empty(&cfqq->fifo))
2928 rq = rq_entry_fifo(cfqq->fifo.next);
2929 if (time_before(jiffies, rq->fifo_time))
2932 cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq);
2937 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2939 const int base_rq = cfqd->cfq_slice_async_rq;
2941 WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
2943 return 2 * base_rq * (IOPRIO_BE_NR - cfqq->ioprio);
2947 * Must be called with the queue_lock held.
2949 static int cfqq_process_refs(struct cfq_queue *cfqq)
2951 int process_refs, io_refs;
2953 io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE];
2954 process_refs = cfqq->ref - io_refs;
2955 BUG_ON(process_refs < 0);
2956 return process_refs;
2959 static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq)
2961 int process_refs, new_process_refs;
2962 struct cfq_queue *__cfqq;
2965 * If there are no process references on the new_cfqq, then it is
2966 * unsafe to follow the ->new_cfqq chain as other cfqq's in the
2967 * chain may have dropped their last reference (not just their
2968 * last process reference).
2970 if (!cfqq_process_refs(new_cfqq))
2973 /* Avoid a circular list and skip interim queue merges */
2974 while ((__cfqq = new_cfqq->new_cfqq)) {
2980 process_refs = cfqq_process_refs(cfqq);
2981 new_process_refs = cfqq_process_refs(new_cfqq);
2983 * If the process for the cfqq has gone away, there is no
2984 * sense in merging the queues.
2986 if (process_refs == 0 || new_process_refs == 0)
2990 * Merge in the direction of the lesser amount of work.
2992 if (new_process_refs >= process_refs) {
2993 cfqq->new_cfqq = new_cfqq;
2994 new_cfqq->ref += process_refs;
2996 new_cfqq->new_cfqq = cfqq;
2997 cfqq->ref += new_process_refs;
3001 static enum wl_type_t cfq_choose_wl_type(struct cfq_data *cfqd,
3002 struct cfq_group *cfqg, enum wl_class_t wl_class)
3004 struct cfq_queue *queue;
3006 bool key_valid = false;
3007 unsigned long lowest_key = 0;
3008 enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD;
3010 for (i = 0; i <= SYNC_WORKLOAD; ++i) {
3011 /* select the one with lowest rb_key */
3012 queue = cfq_rb_first(st_for(cfqg, wl_class, i));
3014 (!key_valid || time_before(queue->rb_key, lowest_key))) {
3015 lowest_key = queue->rb_key;
3025 choose_wl_class_and_type(struct cfq_data *cfqd, struct cfq_group *cfqg)
3029 struct cfq_rb_root *st;
3030 unsigned group_slice;
3031 enum wl_class_t original_class = cfqd->serving_wl_class;
3033 /* Choose next priority. RT > BE > IDLE */
3034 if (cfq_group_busy_queues_wl(RT_WORKLOAD, cfqd, cfqg))
3035 cfqd->serving_wl_class = RT_WORKLOAD;
3036 else if (cfq_group_busy_queues_wl(BE_WORKLOAD, cfqd, cfqg))
3037 cfqd->serving_wl_class = BE_WORKLOAD;
3039 cfqd->serving_wl_class = IDLE_WORKLOAD;
3040 cfqd->workload_expires = jiffies + 1;
3044 if (original_class != cfqd->serving_wl_class)
3048 * For RT and BE, we have to choose also the type
3049 * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
3052 st = st_for(cfqg, cfqd->serving_wl_class, cfqd->serving_wl_type);
3056 * check workload expiration, and that we still have other queues ready
3058 if (count && !time_after(jiffies, cfqd->workload_expires))
3062 /* otherwise select new workload type */
3063 cfqd->serving_wl_type = cfq_choose_wl_type(cfqd, cfqg,
3064 cfqd->serving_wl_class);
3065 st = st_for(cfqg, cfqd->serving_wl_class, cfqd->serving_wl_type);
3069 * the workload slice is computed as a fraction of target latency
3070 * proportional to the number of queues in that workload, over
3071 * all the queues in the same priority class
3073 group_slice = cfq_group_slice(cfqd, cfqg);
3075 slice = group_slice * count /
3076 max_t(unsigned, cfqg->busy_queues_avg[cfqd->serving_wl_class],
3077 cfq_group_busy_queues_wl(cfqd->serving_wl_class, cfqd,
3080 if (cfqd->serving_wl_type == ASYNC_WORKLOAD) {
3084 * Async queues are currently system wide. Just taking
3085 * proportion of queues with-in same group will lead to higher
3086 * async ratio system wide as generally root group is going
3087 * to have higher weight. A more accurate thing would be to
3088 * calculate system wide asnc/sync ratio.
3090 tmp = cfqd->cfq_target_latency *
3091 cfqg_busy_async_queues(cfqd, cfqg);
3092 tmp = tmp/cfqd->busy_queues;
3093 slice = min_t(unsigned, slice, tmp);
3095 /* async workload slice is scaled down according to
3096 * the sync/async slice ratio. */
3097 slice = slice * cfqd->cfq_slice[0] / cfqd->cfq_slice[1];
3099 /* sync workload slice is at least 2 * cfq_slice_idle */
3100 slice = max(slice, 2 * cfqd->cfq_slice_idle);
3102 slice = max_t(unsigned, slice, CFQ_MIN_TT);
3103 cfq_log(cfqd, "workload slice:%d", slice);
3104 cfqd->workload_expires = jiffies + slice;
3107 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd)
3109 struct cfq_rb_root *st = &cfqd->grp_service_tree;
3110 struct cfq_group *cfqg;
3112 if (RB_EMPTY_ROOT(&st->rb))
3114 cfqg = cfq_rb_first_group(st);
3115 update_min_vdisktime(st);
3119 static void cfq_choose_cfqg(struct cfq_data *cfqd)
3121 struct cfq_group *cfqg = cfq_get_next_cfqg(cfqd);
3123 cfqd->serving_group = cfqg;
3125 /* Restore the workload type data */
3126 if (cfqg->saved_wl_slice) {
3127 cfqd->workload_expires = jiffies + cfqg->saved_wl_slice;
3128 cfqd->serving_wl_type = cfqg->saved_wl_type;
3129 cfqd->serving_wl_class = cfqg->saved_wl_class;
3131 cfqd->workload_expires = jiffies - 1;
3133 choose_wl_class_and_type(cfqd, cfqg);
3137 * Select a queue for service. If we have a current active queue,
3138 * check whether to continue servicing it, or retrieve and set a new one.
3140 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
3142 struct cfq_queue *cfqq, *new_cfqq = NULL;
3144 cfqq = cfqd->active_queue;
3148 if (!cfqd->rq_queued)
3152 * We were waiting for group to get backlogged. Expire the queue
3154 if (cfq_cfqq_wait_busy(cfqq) && !RB_EMPTY_ROOT(&cfqq->sort_list))
3158 * The active queue has run out of time, expire it and select new.
3160 if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq)) {
3162 * If slice had not expired at the completion of last request
3163 * we might not have turned on wait_busy flag. Don't expire
3164 * the queue yet. Allow the group to get backlogged.
3166 * The very fact that we have used the slice, that means we
3167 * have been idling all along on this queue and it should be
3168 * ok to wait for this request to complete.
3170 if (cfqq->cfqg->nr_cfqq == 1 && RB_EMPTY_ROOT(&cfqq->sort_list)
3171 && cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
3175 goto check_group_idle;
3179 * The active queue has requests and isn't expired, allow it to
3182 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3186 * If another queue has a request waiting within our mean seek
3187 * distance, let it run. The expire code will check for close
3188 * cooperators and put the close queue at the front of the service
3189 * tree. If possible, merge the expiring queue with the new cfqq.
3191 new_cfqq = cfq_close_cooperator(cfqd, cfqq);
3193 if (!cfqq->new_cfqq)
3194 cfq_setup_merge(cfqq, new_cfqq);
3199 * No requests pending. If the active queue still has requests in
3200 * flight or is idling for a new request, allow either of these
3201 * conditions to happen (or time out) before selecting a new queue.
3203 if (timer_pending(&cfqd->idle_slice_timer)) {
3209 * This is a deep seek queue, but the device is much faster than
3210 * the queue can deliver, don't idle
3212 if (CFQQ_SEEKY(cfqq) && cfq_cfqq_idle_window(cfqq) &&
3213 (cfq_cfqq_slice_new(cfqq) ||
3214 (cfqq->slice_end - jiffies > jiffies - cfqq->slice_start))) {
3215 cfq_clear_cfqq_deep(cfqq);
3216 cfq_clear_cfqq_idle_window(cfqq);
3219 if (cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
3225 * If group idle is enabled and there are requests dispatched from
3226 * this group, wait for requests to complete.
3229 if (cfqd->cfq_group_idle && cfqq->cfqg->nr_cfqq == 1 &&
3230 cfqq->cfqg->dispatched &&
3231 !cfq_io_thinktime_big(cfqd, &cfqq->cfqg->ttime, true)) {
3237 cfq_slice_expired(cfqd, 0);
3240 * Current queue expired. Check if we have to switch to a new
3244 cfq_choose_cfqg(cfqd);
3246 cfqq = cfq_set_active_queue(cfqd, new_cfqq);
3251 static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
3255 while (cfqq->next_rq) {
3256 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
3260 BUG_ON(!list_empty(&cfqq->fifo));
3262 /* By default cfqq is not expired if it is empty. Do it explicitly */
3263 __cfq_slice_expired(cfqq->cfqd, cfqq, 0);
3268 * Drain our current requests. Used for barriers and when switching
3269 * io schedulers on-the-fly.
3271 static int cfq_forced_dispatch(struct cfq_data *cfqd)
3273 struct cfq_queue *cfqq;
3276 /* Expire the timeslice of the current active queue first */
3277 cfq_slice_expired(cfqd, 0);
3278 while ((cfqq = cfq_get_next_queue_forced(cfqd)) != NULL) {
3279 __cfq_set_active_queue(cfqd, cfqq);
3280 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
3283 BUG_ON(cfqd->busy_queues);
3285 cfq_log(cfqd, "forced_dispatch=%d", dispatched);
3289 static inline bool cfq_slice_used_soon(struct cfq_data *cfqd,
3290 struct cfq_queue *cfqq)
3292 /* the queue hasn't finished any request, can't estimate */
3293 if (cfq_cfqq_slice_new(cfqq))
3295 if (time_after(jiffies + cfqd->cfq_slice_idle * cfqq->dispatched,
3302 static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3304 unsigned int max_dispatch;
3307 * Drain async requests before we start sync IO
3309 if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_flight[BLK_RW_ASYNC])
3313 * If this is an async queue and we have sync IO in flight, let it wait
3315 if (cfqd->rq_in_flight[BLK_RW_SYNC] && !cfq_cfqq_sync(cfqq))
3318 max_dispatch = max_t(unsigned int, cfqd->cfq_quantum / 2, 1);
3319 if (cfq_class_idle(cfqq))
3323 * Does this cfqq already have too much IO in flight?
3325 if (cfqq->dispatched >= max_dispatch) {
3326 bool promote_sync = false;
3328 * idle queue must always only have a single IO in flight
3330 if (cfq_class_idle(cfqq))
3334 * If there is only one sync queue
3335 * we can ignore async queue here and give the sync
3336 * queue no dispatch limit. The reason is a sync queue can
3337 * preempt async queue, limiting the sync queue doesn't make
3338 * sense. This is useful for aiostress test.
3340 if (cfq_cfqq_sync(cfqq) && cfqd->busy_sync_queues == 1)
3341 promote_sync = true;
3344 * We have other queues, don't allow more IO from this one
3346 if (cfqd->busy_queues > 1 && cfq_slice_used_soon(cfqd, cfqq) &&
3351 * Sole queue user, no limit
3353 if (cfqd->busy_queues == 1 || promote_sync)
3357 * Normally we start throttling cfqq when cfq_quantum/2
3358 * requests have been dispatched. But we can drive
3359 * deeper queue depths at the beginning of slice
3360 * subjected to upper limit of cfq_quantum.
3362 max_dispatch = cfqd->cfq_quantum;
3366 * Async queues must wait a bit before being allowed dispatch.
3367 * We also ramp up the dispatch depth gradually for async IO,
3368 * based on the last sync IO we serviced
3370 if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
3371 unsigned long last_sync = jiffies - cfqd->last_delayed_sync;
3374 depth = last_sync / cfqd->cfq_slice[1];
3375 if (!depth && !cfqq->dispatched)
3377 if (depth < max_dispatch)
3378 max_dispatch = depth;
3382 * If we're below the current max, allow a dispatch
3384 return cfqq->dispatched < max_dispatch;
3388 * Dispatch a request from cfqq, moving them to the request queue
3391 static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3395 BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
3397 if (!cfq_may_dispatch(cfqd, cfqq))
3401 * follow expired path, else get first next available
3403 rq = cfq_check_fifo(cfqq);
3408 * insert request into driver dispatch list
3410 cfq_dispatch_insert(cfqd->queue, rq);
3412 if (!cfqd->active_cic) {
3413 struct cfq_io_cq *cic = RQ_CIC(rq);
3415 atomic_long_inc(&cic->icq.ioc->refcount);
3416 cfqd->active_cic = cic;
3423 * Find the cfqq that we need to service and move a request from that to the
3426 static int cfq_dispatch_requests(struct request_queue *q, int force)
3428 struct cfq_data *cfqd = q->elevator->elevator_data;
3429 struct cfq_queue *cfqq;
3431 if (!cfqd->busy_queues)
3434 if (unlikely(force))
3435 return cfq_forced_dispatch(cfqd);
3437 cfqq = cfq_select_queue(cfqd);
3442 * Dispatch a request from this cfqq, if it is allowed
3444 if (!cfq_dispatch_request(cfqd, cfqq))
3447 cfqq->slice_dispatch++;
3448 cfq_clear_cfqq_must_dispatch(cfqq);
3451 * expire an async queue immediately if it has used up its slice. idle
3452 * queue always expire after 1 dispatch round.
3454 if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
3455 cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
3456 cfq_class_idle(cfqq))) {
3457 cfqq->slice_end = jiffies + 1;
3458 cfq_slice_expired(cfqd, 0);
3461 cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
3466 * task holds one reference to the queue, dropped when task exits. each rq
3467 * in-flight on this queue also holds a reference, dropped when rq is freed.
3469 * Each cfq queue took a reference on the parent group. Drop it now.
3470 * queue lock must be held here.
3472 static void cfq_put_queue(struct cfq_queue *cfqq)
3474 struct cfq_data *cfqd = cfqq->cfqd;
3475 struct cfq_group *cfqg;
3477 BUG_ON(cfqq->ref <= 0);
3483 cfq_log_cfqq(cfqd, cfqq, "put_queue");
3484 BUG_ON(rb_first(&cfqq->sort_list));
3485 BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
3488 if (unlikely(cfqd->active_queue == cfqq)) {
3489 __cfq_slice_expired(cfqd, cfqq, 0);
3490 cfq_schedule_dispatch(cfqd);
3493 BUG_ON(cfq_cfqq_on_rr(cfqq));
3494 kmem_cache_free(cfq_pool, cfqq);
3498 static void cfq_put_cooperator(struct cfq_queue *cfqq)
3500 struct cfq_queue *__cfqq, *next;
3503 * If this queue was scheduled to merge with another queue, be
3504 * sure to drop the reference taken on that queue (and others in
3505 * the merge chain). See cfq_setup_merge and cfq_merge_cfqqs.
3507 __cfqq = cfqq->new_cfqq;
3509 if (__cfqq == cfqq) {
3510 WARN(1, "cfqq->new_cfqq loop detected\n");
3513 next = __cfqq->new_cfqq;
3514 cfq_put_queue(__cfqq);
3519 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3521 if (unlikely(cfqq == cfqd->active_queue)) {
3522 __cfq_slice_expired(cfqd, cfqq, 0);
3523 cfq_schedule_dispatch(cfqd);
3526 cfq_put_cooperator(cfqq);
3528 cfq_put_queue(cfqq);
3531 static void cfq_init_icq(struct io_cq *icq)
3533 struct cfq_io_cq *cic = icq_to_cic(icq);
3535 cic->ttime.last_end_request = jiffies;
3538 static void cfq_exit_icq(struct io_cq *icq)
3540 struct cfq_io_cq *cic = icq_to_cic(icq);
3541 struct cfq_data *cfqd = cic_to_cfqd(cic);
3543 if (cic_to_cfqq(cic, false)) {
3544 cfq_exit_cfqq(cfqd, cic_to_cfqq(cic, false));
3545 cic_set_cfqq(cic, NULL, false);
3548 if (cic_to_cfqq(cic, true)) {
3549 cfq_exit_cfqq(cfqd, cic_to_cfqq(cic, true));
3550 cic_set_cfqq(cic, NULL, true);
3554 static void cfq_init_prio_data(struct cfq_queue *cfqq, struct cfq_io_cq *cic)
3556 struct task_struct *tsk = current;
3559 if (!cfq_cfqq_prio_changed(cfqq))
3562 ioprio_class = IOPRIO_PRIO_CLASS(cic->ioprio);
3563 switch (ioprio_class) {
3565 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
3566 case IOPRIO_CLASS_NONE:
3568 * no prio set, inherit CPU scheduling settings
3570 cfqq->ioprio = task_nice_ioprio(tsk);
3571 cfqq->ioprio_class = task_nice_ioclass(tsk);
3573 case IOPRIO_CLASS_RT:
3574 cfqq->ioprio = IOPRIO_PRIO_DATA(cic->ioprio);
3575 cfqq->ioprio_class = IOPRIO_CLASS_RT;
3577 case IOPRIO_CLASS_BE:
3578 cfqq->ioprio = IOPRIO_PRIO_DATA(cic->ioprio);
3579 cfqq->ioprio_class = IOPRIO_CLASS_BE;
3581 case IOPRIO_CLASS_IDLE:
3582 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
3584 cfq_clear_cfqq_idle_window(cfqq);
3589 * keep track of original prio settings in case we have to temporarily
3590 * elevate the priority of this queue
3592 cfqq->org_ioprio = cfqq->ioprio;
3593 cfq_clear_cfqq_prio_changed(cfqq);
3596 static void check_ioprio_changed(struct cfq_io_cq *cic, struct bio *bio)
3598 int ioprio = cic->icq.ioc->ioprio;
3599 struct cfq_data *cfqd = cic_to_cfqd(cic);
3600 struct cfq_queue *cfqq;
3603 * Check whether ioprio has changed. The condition may trigger
3604 * spuriously on a newly created cic but there's no harm.
3606 if (unlikely(!cfqd) || likely(cic->ioprio == ioprio))
3609 cfqq = cic_to_cfqq(cic, false);
3611 cfq_put_queue(cfqq);
3612 cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic, bio);
3613 cic_set_cfqq(cic, cfqq, false);
3616 cfqq = cic_to_cfqq(cic, true);
3618 cfq_mark_cfqq_prio_changed(cfqq);
3620 cic->ioprio = ioprio;
3623 static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3624 pid_t pid, bool is_sync)
3626 RB_CLEAR_NODE(&cfqq->rb_node);
3627 RB_CLEAR_NODE(&cfqq->p_node);
3628 INIT_LIST_HEAD(&cfqq->fifo);
3633 cfq_mark_cfqq_prio_changed(cfqq);
3636 if (!cfq_class_idle(cfqq))
3637 cfq_mark_cfqq_idle_window(cfqq);
3638 cfq_mark_cfqq_sync(cfqq);
3643 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3644 static void check_blkcg_changed(struct cfq_io_cq *cic, struct bio *bio)
3646 struct cfq_data *cfqd = cic_to_cfqd(cic);
3647 struct cfq_queue *cfqq;
3651 serial_nr = bio_blkcg(bio)->css.serial_nr;
3655 * Check whether blkcg has changed. The condition may trigger
3656 * spuriously on a newly created cic but there's no harm.
3658 if (unlikely(!cfqd) || likely(cic->blkcg_serial_nr == serial_nr))
3662 * Drop reference to queues. New queues will be assigned in new
3663 * group upon arrival of fresh requests.
3665 cfqq = cic_to_cfqq(cic, false);
3667 cfq_log_cfqq(cfqd, cfqq, "changed cgroup");
3668 cic_set_cfqq(cic, NULL, false);
3669 cfq_put_queue(cfqq);
3672 cfqq = cic_to_cfqq(cic, true);
3674 cfq_log_cfqq(cfqd, cfqq, "changed cgroup");
3675 cic_set_cfqq(cic, NULL, true);
3676 cfq_put_queue(cfqq);
3679 cic->blkcg_serial_nr = serial_nr;
3682 static inline void check_blkcg_changed(struct cfq_io_cq *cic, struct bio *bio) { }
3683 #endif /* CONFIG_CFQ_GROUP_IOSCHED */
3685 static struct cfq_queue **
3686 cfq_async_queue_prio(struct cfq_group *cfqg, int ioprio_class, int ioprio)
3688 switch (ioprio_class) {
3689 case IOPRIO_CLASS_RT:
3690 return &cfqg->async_cfqq[0][ioprio];
3691 case IOPRIO_CLASS_NONE:
3692 ioprio = IOPRIO_NORM;
3694 case IOPRIO_CLASS_BE:
3695 return &cfqg->async_cfqq[1][ioprio];
3696 case IOPRIO_CLASS_IDLE:
3697 return &cfqg->async_idle_cfqq;
3703 static struct cfq_queue *
3704 cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct cfq_io_cq *cic,
3707 int ioprio_class = IOPRIO_PRIO_CLASS(cic->ioprio);
3708 int ioprio = IOPRIO_PRIO_DATA(cic->ioprio);
3709 struct cfq_queue **async_cfqq = NULL;
3710 struct cfq_queue *cfqq;
3711 struct cfq_group *cfqg;
3714 cfqg = cfq_lookup_cfqg(cfqd, bio_blkcg(bio));
3716 cfqq = &cfqd->oom_cfqq;
3721 if (!ioprio_valid(cic->ioprio)) {
3722 struct task_struct *tsk = current;
3723 ioprio = task_nice_ioprio(tsk);
3724 ioprio_class = task_nice_ioclass(tsk);
3726 async_cfqq = cfq_async_queue_prio(cfqg, ioprio_class, ioprio);
3732 cfqq = kmem_cache_alloc_node(cfq_pool, GFP_NOWAIT | __GFP_ZERO,
3735 cfqq = &cfqd->oom_cfqq;
3739 cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
3740 cfq_init_prio_data(cfqq, cic);
3741 cfq_link_cfqq_cfqg(cfqq, cfqg);
3742 cfq_log_cfqq(cfqd, cfqq, "alloced");
3745 /* a new async queue is created, pin and remember */
3756 __cfq_update_io_thinktime(struct cfq_ttime *ttime, unsigned long slice_idle)
3758 unsigned long elapsed = jiffies - ttime->last_end_request;
3759 elapsed = min(elapsed, 2UL * slice_idle);
3761 ttime->ttime_samples = (7*ttime->ttime_samples + 256) / 8;
3762 ttime->ttime_total = (7*ttime->ttime_total + 256*elapsed) / 8;
3763 ttime->ttime_mean = (ttime->ttime_total + 128) / ttime->ttime_samples;
3767 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3768 struct cfq_io_cq *cic)
3770 if (cfq_cfqq_sync(cfqq)) {
3771 __cfq_update_io_thinktime(&cic->ttime, cfqd->cfq_slice_idle);
3772 __cfq_update_io_thinktime(&cfqq->service_tree->ttime,
3773 cfqd->cfq_slice_idle);
3775 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3776 __cfq_update_io_thinktime(&cfqq->cfqg->ttime, cfqd->cfq_group_idle);
3781 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3785 sector_t n_sec = blk_rq_sectors(rq);
3786 if (cfqq->last_request_pos) {
3787 if (cfqq->last_request_pos < blk_rq_pos(rq))
3788 sdist = blk_rq_pos(rq) - cfqq->last_request_pos;
3790 sdist = cfqq->last_request_pos - blk_rq_pos(rq);
3793 cfqq->seek_history <<= 1;
3794 if (blk_queue_nonrot(cfqd->queue))
3795 cfqq->seek_history |= (n_sec < CFQQ_SECT_THR_NONROT);
3797 cfqq->seek_history |= (sdist > CFQQ_SEEK_THR);
3801 * Disable idle window if the process thinks too long or seeks so much that
3805 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3806 struct cfq_io_cq *cic)
3808 int old_idle, enable_idle;
3811 * Don't idle for async or idle io prio class
3813 if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
3816 enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
3818 if (cfqq->queued[0] + cfqq->queued[1] >= 4)
3819 cfq_mark_cfqq_deep(cfqq);
3821 if (cfqq->next_rq && (cfqq->next_rq->cmd_flags & REQ_NOIDLE))
3823 else if (!atomic_read(&cic->icq.ioc->active_ref) ||
3824 !cfqd->cfq_slice_idle ||
3825 (!cfq_cfqq_deep(cfqq) && CFQQ_SEEKY(cfqq)))
3827 else if (sample_valid(cic->ttime.ttime_samples)) {
3828 if (cic->ttime.ttime_mean > cfqd->cfq_slice_idle)
3834 if (old_idle != enable_idle) {
3835 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
3837 cfq_mark_cfqq_idle_window(cfqq);
3839 cfq_clear_cfqq_idle_window(cfqq);
3844 * Check if new_cfqq should preempt the currently active queue. Return 0 for
3845 * no or if we aren't sure, a 1 will cause a preempt.
3848 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
3851 struct cfq_queue *cfqq;
3853 cfqq = cfqd->active_queue;
3857 if (cfq_class_idle(new_cfqq))
3860 if (cfq_class_idle(cfqq))
3864 * Don't allow a non-RT request to preempt an ongoing RT cfqq timeslice.
3866 if (cfq_class_rt(cfqq) && !cfq_class_rt(new_cfqq))
3870 * if the new request is sync, but the currently running queue is
3871 * not, let the sync request have priority.
3873 if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
3876 if (new_cfqq->cfqg != cfqq->cfqg)
3879 if (cfq_slice_used(cfqq))
3882 /* Allow preemption only if we are idling on sync-noidle tree */
3883 if (cfqd->serving_wl_type == SYNC_NOIDLE_WORKLOAD &&
3884 cfqq_type(new_cfqq) == SYNC_NOIDLE_WORKLOAD &&
3885 new_cfqq->service_tree->count == 2 &&
3886 RB_EMPTY_ROOT(&cfqq->sort_list))
3890 * So both queues are sync. Let the new request get disk time if
3891 * it's a metadata request and the current queue is doing regular IO.
3893 if ((rq->cmd_flags & REQ_PRIO) && !cfqq->prio_pending)
3897 * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
3899 if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
3902 /* An idle queue should not be idle now for some reason */
3903 if (RB_EMPTY_ROOT(&cfqq->sort_list) && !cfq_should_idle(cfqd, cfqq))
3906 if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
3910 * if this request is as-good as one we would expect from the
3911 * current cfqq, let it preempt
3913 if (cfq_rq_close(cfqd, cfqq, rq))
3920 * cfqq preempts the active queue. if we allowed preempt with no slice left,
3921 * let it have half of its nominal slice.
3923 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3925 enum wl_type_t old_type = cfqq_type(cfqd->active_queue);
3927 cfq_log_cfqq(cfqd, cfqq, "preempt");
3928 cfq_slice_expired(cfqd, 1);
3931 * workload type is changed, don't save slice, otherwise preempt
3934 if (old_type != cfqq_type(cfqq))
3935 cfqq->cfqg->saved_wl_slice = 0;
3938 * Put the new queue at the front of the of the current list,
3939 * so we know that it will be selected next.
3941 BUG_ON(!cfq_cfqq_on_rr(cfqq));
3943 cfq_service_tree_add(cfqd, cfqq, 1);
3945 cfqq->slice_end = 0;
3946 cfq_mark_cfqq_slice_new(cfqq);
3950 * Called when a new fs request (rq) is added (to cfqq). Check if there's
3951 * something we should do about it
3954 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3957 struct cfq_io_cq *cic = RQ_CIC(rq);
3960 if (rq->cmd_flags & REQ_PRIO)
3961 cfqq->prio_pending++;
3963 cfq_update_io_thinktime(cfqd, cfqq, cic);
3964 cfq_update_io_seektime(cfqd, cfqq, rq);
3965 cfq_update_idle_window(cfqd, cfqq, cic);
3967 cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
3969 if (cfqq == cfqd->active_queue) {
3971 * Remember that we saw a request from this process, but
3972 * don't start queuing just yet. Otherwise we risk seeing lots
3973 * of tiny requests, because we disrupt the normal plugging
3974 * and merging. If the request is already larger than a single
3975 * page, let it rip immediately. For that case we assume that
3976 * merging is already done. Ditto for a busy system that
3977 * has other work pending, don't risk delaying until the
3978 * idle timer unplug to continue working.
3980 if (cfq_cfqq_wait_request(cfqq)) {
3981 if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
3982 cfqd->busy_queues > 1) {
3983 cfq_del_timer(cfqd, cfqq);
3984 cfq_clear_cfqq_wait_request(cfqq);
3985 __blk_run_queue(cfqd->queue);
3987 cfqg_stats_update_idle_time(cfqq->cfqg);
3988 cfq_mark_cfqq_must_dispatch(cfqq);
3991 } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
3993 * not the active queue - expire current slice if it is
3994 * idle and has expired it's mean thinktime or this new queue
3995 * has some old slice time left and is of higher priority or
3996 * this new queue is RT and the current one is BE
3998 cfq_preempt_queue(cfqd, cfqq);
3999 __blk_run_queue(cfqd->queue);
4003 static void cfq_insert_request(struct request_queue *q, struct request *rq)
4005 struct cfq_data *cfqd = q->elevator->elevator_data;
4006 struct cfq_queue *cfqq = RQ_CFQQ(rq);
4008 cfq_log_cfqq(cfqd, cfqq, "insert_request");
4009 cfq_init_prio_data(cfqq, RQ_CIC(rq));
4011 rq->fifo_time = jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)];
4012 list_add_tail(&rq->queuelist, &cfqq->fifo);
4014 cfqg_stats_update_io_add(RQ_CFQG(rq), cfqd->serving_group,
4016 cfq_rq_enqueued(cfqd, cfqq, rq);
4020 * Update hw_tag based on peak queue depth over 50 samples under
4023 static void cfq_update_hw_tag(struct cfq_data *cfqd)
4025 struct cfq_queue *cfqq = cfqd->active_queue;
4027 if (cfqd->rq_in_driver > cfqd->hw_tag_est_depth)
4028 cfqd->hw_tag_est_depth = cfqd->rq_in_driver;
4030 if (cfqd->hw_tag == 1)
4033 if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
4034 cfqd->rq_in_driver <= CFQ_HW_QUEUE_MIN)
4038 * If active queue hasn't enough requests and can idle, cfq might not
4039 * dispatch sufficient requests to hardware. Don't zero hw_tag in this
4042 if (cfqq && cfq_cfqq_idle_window(cfqq) &&
4043 cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] <
4044 CFQ_HW_QUEUE_MIN && cfqd->rq_in_driver < CFQ_HW_QUEUE_MIN)
4047 if (cfqd->hw_tag_samples++ < 50)
4050 if (cfqd->hw_tag_est_depth >= CFQ_HW_QUEUE_MIN)
4056 static bool cfq_should_wait_busy(struct cfq_data *cfqd, struct cfq_queue *cfqq)
4058 struct cfq_io_cq *cic = cfqd->active_cic;
4060 /* If the queue already has requests, don't wait */
4061 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
4064 /* If there are other queues in the group, don't wait */
4065 if (cfqq->cfqg->nr_cfqq > 1)
4068 /* the only queue in the group, but think time is big */
4069 if (cfq_io_thinktime_big(cfqd, &cfqq->cfqg->ttime, true))
4072 if (cfq_slice_used(cfqq))
4075 /* if slice left is less than think time, wait busy */
4076 if (cic && sample_valid(cic->ttime.ttime_samples)
4077 && (cfqq->slice_end - jiffies < cic->ttime.ttime_mean))
4081 * If think times is less than a jiffy than ttime_mean=0 and above
4082 * will not be true. It might happen that slice has not expired yet
4083 * but will expire soon (4-5 ns) during select_queue(). To cover the
4084 * case where think time is less than a jiffy, mark the queue wait
4085 * busy if only 1 jiffy is left in the slice.
4087 if (cfqq->slice_end - jiffies == 1)
4093 static void cfq_completed_request(struct request_queue *q, struct request *rq)
4095 struct cfq_queue *cfqq = RQ_CFQQ(rq);
4096 struct cfq_data *cfqd = cfqq->cfqd;
4097 const int sync = rq_is_sync(rq);
4101 cfq_log_cfqq(cfqd, cfqq, "complete rqnoidle %d",
4102 !!(rq->cmd_flags & REQ_NOIDLE));
4104 cfq_update_hw_tag(cfqd);
4106 WARN_ON(!cfqd->rq_in_driver);
4107 WARN_ON(!cfqq->dispatched);
4108 cfqd->rq_in_driver--;
4110 (RQ_CFQG(rq))->dispatched--;
4111 cfqg_stats_update_completion(cfqq->cfqg, rq_start_time_ns(rq),
4112 rq_io_start_time_ns(rq), rq->cmd_flags);
4114 cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]--;
4117 struct cfq_rb_root *st;
4119 RQ_CIC(rq)->ttime.last_end_request = now;
4121 if (cfq_cfqq_on_rr(cfqq))
4122 st = cfqq->service_tree;
4124 st = st_for(cfqq->cfqg, cfqq_class(cfqq),
4127 st->ttime.last_end_request = now;
4128 if (!time_after(rq->start_time + cfqd->cfq_fifo_expire[1], now))
4129 cfqd->last_delayed_sync = now;
4132 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4133 cfqq->cfqg->ttime.last_end_request = now;
4137 * If this is the active queue, check if it needs to be expired,
4138 * or if we want to idle in case it has no pending requests.
4140 if (cfqd->active_queue == cfqq) {
4141 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
4143 if (cfq_cfqq_slice_new(cfqq)) {
4144 cfq_set_prio_slice(cfqd, cfqq);
4145 cfq_clear_cfqq_slice_new(cfqq);
4149 * Should we wait for next request to come in before we expire
4152 if (cfq_should_wait_busy(cfqd, cfqq)) {
4153 unsigned long extend_sl = cfqd->cfq_slice_idle;
4154 if (!cfqd->cfq_slice_idle)
4155 extend_sl = cfqd->cfq_group_idle;
4156 cfqq->slice_end = jiffies + extend_sl;
4157 cfq_mark_cfqq_wait_busy(cfqq);
4158 cfq_log_cfqq(cfqd, cfqq, "will busy wait");
4162 * Idling is not enabled on:
4164 * - idle-priority queues
4166 * - queues with still some requests queued
4167 * - when there is a close cooperator
4169 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
4170 cfq_slice_expired(cfqd, 1);
4171 else if (sync && cfqq_empty &&
4172 !cfq_close_cooperator(cfqd, cfqq)) {
4173 cfq_arm_slice_timer(cfqd);
4177 if (!cfqd->rq_in_driver)
4178 cfq_schedule_dispatch(cfqd);
4181 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
4183 if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
4184 cfq_mark_cfqq_must_alloc_slice(cfqq);
4185 return ELV_MQUEUE_MUST;
4188 return ELV_MQUEUE_MAY;
4191 static int cfq_may_queue(struct request_queue *q, int rw)
4193 struct cfq_data *cfqd = q->elevator->elevator_data;
4194 struct task_struct *tsk = current;
4195 struct cfq_io_cq *cic;
4196 struct cfq_queue *cfqq;
4199 * don't force setup of a queue from here, as a call to may_queue
4200 * does not necessarily imply that a request actually will be queued.
4201 * so just lookup a possibly existing queue, or return 'may queue'
4204 cic = cfq_cic_lookup(cfqd, tsk->io_context);
4206 return ELV_MQUEUE_MAY;
4208 cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
4210 cfq_init_prio_data(cfqq, cic);
4212 return __cfq_may_queue(cfqq);
4215 return ELV_MQUEUE_MAY;
4219 * queue lock held here
4221 static void cfq_put_request(struct request *rq)
4223 struct cfq_queue *cfqq = RQ_CFQQ(rq);
4226 const int rw = rq_data_dir(rq);
4228 BUG_ON(!cfqq->allocated[rw]);
4229 cfqq->allocated[rw]--;
4231 /* Put down rq reference on cfqg */
4232 cfqg_put(RQ_CFQG(rq));
4233 rq->elv.priv[0] = NULL;
4234 rq->elv.priv[1] = NULL;
4236 cfq_put_queue(cfqq);
4240 static struct cfq_queue *
4241 cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_cq *cic,
4242 struct cfq_queue *cfqq)
4244 cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq);
4245 cic_set_cfqq(cic, cfqq->new_cfqq, 1);
4246 cfq_mark_cfqq_coop(cfqq->new_cfqq);
4247 cfq_put_queue(cfqq);
4248 return cic_to_cfqq(cic, 1);
4252 * Returns NULL if a new cfqq should be allocated, or the old cfqq if this
4253 * was the last process referring to said cfqq.
4255 static struct cfq_queue *
4256 split_cfqq(struct cfq_io_cq *cic, struct cfq_queue *cfqq)
4258 if (cfqq_process_refs(cfqq) == 1) {
4259 cfqq->pid = current->pid;
4260 cfq_clear_cfqq_coop(cfqq);
4261 cfq_clear_cfqq_split_coop(cfqq);
4265 cic_set_cfqq(cic, NULL, 1);
4267 cfq_put_cooperator(cfqq);
4269 cfq_put_queue(cfqq);
4273 * Allocate cfq data structures associated with this request.
4276 cfq_set_request(struct request_queue *q, struct request *rq, struct bio *bio,
4279 struct cfq_data *cfqd = q->elevator->elevator_data;
4280 struct cfq_io_cq *cic = icq_to_cic(rq->elv.icq);
4281 const int rw = rq_data_dir(rq);
4282 const bool is_sync = rq_is_sync(rq);
4283 struct cfq_queue *cfqq;
4285 spin_lock_irq(q->queue_lock);
4287 check_ioprio_changed(cic, bio);
4288 check_blkcg_changed(cic, bio);
4290 cfqq = cic_to_cfqq(cic, is_sync);
4291 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
4293 cfq_put_queue(cfqq);
4294 cfqq = cfq_get_queue(cfqd, is_sync, cic, bio);
4295 cic_set_cfqq(cic, cfqq, is_sync);
4298 * If the queue was seeky for too long, break it apart.
4300 if (cfq_cfqq_coop(cfqq) && cfq_cfqq_split_coop(cfqq)) {
4301 cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq");
4302 cfqq = split_cfqq(cic, cfqq);
4308 * Check to see if this queue is scheduled to merge with
4309 * another, closely cooperating queue. The merging of
4310 * queues happens here as it must be done in process context.
4311 * The reference on new_cfqq was taken in merge_cfqqs.
4314 cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq);
4317 cfqq->allocated[rw]++;
4320 cfqg_get(cfqq->cfqg);
4321 rq->elv.priv[0] = cfqq;
4322 rq->elv.priv[1] = cfqq->cfqg;
4323 spin_unlock_irq(q->queue_lock);
4327 static void cfq_kick_queue(struct work_struct *work)
4329 struct cfq_data *cfqd =
4330 container_of(work, struct cfq_data, unplug_work);
4331 struct request_queue *q = cfqd->queue;
4333 spin_lock_irq(q->queue_lock);
4334 __blk_run_queue(cfqd->queue);
4335 spin_unlock_irq(q->queue_lock);
4339 * Timer running if the active_queue is currently idling inside its time slice
4341 static void cfq_idle_slice_timer(unsigned long data)
4343 struct cfq_data *cfqd = (struct cfq_data *) data;
4344 struct cfq_queue *cfqq;
4345 unsigned long flags;
4348 cfq_log(cfqd, "idle timer fired");
4350 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
4352 cfqq = cfqd->active_queue;
4357 * We saw a request before the queue expired, let it through
4359 if (cfq_cfqq_must_dispatch(cfqq))
4365 if (cfq_slice_used(cfqq))
4369 * only expire and reinvoke request handler, if there are
4370 * other queues with pending requests
4372 if (!cfqd->busy_queues)
4376 * not expired and it has a request pending, let it dispatch
4378 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
4382 * Queue depth flag is reset only when the idle didn't succeed
4384 cfq_clear_cfqq_deep(cfqq);
4387 cfq_slice_expired(cfqd, timed_out);
4389 cfq_schedule_dispatch(cfqd);
4391 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
4394 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
4396 del_timer_sync(&cfqd->idle_slice_timer);
4397 cancel_work_sync(&cfqd->unplug_work);
4400 static void cfq_exit_queue(struct elevator_queue *e)
4402 struct cfq_data *cfqd = e->elevator_data;
4403 struct request_queue *q = cfqd->queue;
4405 cfq_shutdown_timer_wq(cfqd);
4407 spin_lock_irq(q->queue_lock);
4409 if (cfqd->active_queue)
4410 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
4412 spin_unlock_irq(q->queue_lock);
4414 cfq_shutdown_timer_wq(cfqd);
4416 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4417 blkcg_deactivate_policy(q, &blkcg_policy_cfq);
4419 kfree(cfqd->root_group);
4424 static int cfq_init_queue(struct request_queue *q, struct elevator_type *e)
4426 struct cfq_data *cfqd;
4427 struct blkcg_gq *blkg __maybe_unused;
4429 struct elevator_queue *eq;
4431 eq = elevator_alloc(q, e);
4435 cfqd = kzalloc_node(sizeof(*cfqd), GFP_KERNEL, q->node);
4437 kobject_put(&eq->kobj);
4440 eq->elevator_data = cfqd;
4443 spin_lock_irq(q->queue_lock);
4445 spin_unlock_irq(q->queue_lock);
4447 /* Init root service tree */
4448 cfqd->grp_service_tree = CFQ_RB_ROOT;
4450 /* Init root group and prefer root group over other groups by default */
4451 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4452 ret = blkcg_activate_policy(q, &blkcg_policy_cfq);
4456 cfqd->root_group = blkg_to_cfqg(q->root_blkg);
4459 cfqd->root_group = kzalloc_node(sizeof(*cfqd->root_group),
4460 GFP_KERNEL, cfqd->queue->node);
4461 if (!cfqd->root_group)
4464 cfq_init_cfqg_base(cfqd->root_group);
4466 cfqd->root_group->weight = 2 * CFQ_WEIGHT_DEFAULT;
4467 cfqd->root_group->leaf_weight = 2 * CFQ_WEIGHT_DEFAULT;
4470 * Not strictly needed (since RB_ROOT just clears the node and we
4471 * zeroed cfqd on alloc), but better be safe in case someone decides
4472 * to add magic to the rb code
4474 for (i = 0; i < CFQ_PRIO_LISTS; i++)
4475 cfqd->prio_trees[i] = RB_ROOT;
4478 * Our fallback cfqq if cfq_get_queue() runs into OOM issues.
4479 * Grab a permanent reference to it, so that the normal code flow
4480 * will not attempt to free it. oom_cfqq is linked to root_group
4481 * but shouldn't hold a reference as it'll never be unlinked. Lose
4482 * the reference from linking right away.
4484 cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
4485 cfqd->oom_cfqq.ref++;
4487 spin_lock_irq(q->queue_lock);
4488 cfq_link_cfqq_cfqg(&cfqd->oom_cfqq, cfqd->root_group);
4489 cfqg_put(cfqd->root_group);
4490 spin_unlock_irq(q->queue_lock);
4492 init_timer(&cfqd->idle_slice_timer);
4493 cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
4494 cfqd->idle_slice_timer.data = (unsigned long) cfqd;
4496 INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
4498 cfqd->cfq_quantum = cfq_quantum;
4499 cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
4500 cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
4501 cfqd->cfq_back_max = cfq_back_max;
4502 cfqd->cfq_back_penalty = cfq_back_penalty;
4503 cfqd->cfq_slice[0] = cfq_slice_async;
4504 cfqd->cfq_slice[1] = cfq_slice_sync;
4505 cfqd->cfq_target_latency = cfq_target_latency;
4506 cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
4507 cfqd->cfq_slice_idle = cfq_slice_idle;
4508 cfqd->cfq_group_idle = cfq_group_idle;
4509 cfqd->cfq_latency = 1;
4512 * we optimistically start assuming sync ops weren't delayed in last
4513 * second, in order to have larger depth for async operations.
4515 cfqd->last_delayed_sync = jiffies - HZ;
4520 kobject_put(&eq->kobj);
4524 static void cfq_registered_queue(struct request_queue *q)
4526 struct elevator_queue *e = q->elevator;
4527 struct cfq_data *cfqd = e->elevator_data;
4530 * Default to IOPS mode with no idling for SSDs
4532 if (blk_queue_nonrot(q))
4533 cfqd->cfq_slice_idle = 0;
4537 * sysfs parts below -->
4540 cfq_var_show(unsigned int var, char *page)
4542 return sprintf(page, "%u\n", var);
4546 cfq_var_store(unsigned int *var, const char *page, size_t count)
4548 char *p = (char *) page;
4550 *var = simple_strtoul(p, &p, 10);
4554 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
4555 static ssize_t __FUNC(struct elevator_queue *e, char *page) \
4557 struct cfq_data *cfqd = e->elevator_data; \
4558 unsigned int __data = __VAR; \
4560 __data = jiffies_to_msecs(__data); \
4561 return cfq_var_show(__data, (page)); \
4563 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
4564 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
4565 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
4566 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
4567 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
4568 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
4569 SHOW_FUNCTION(cfq_group_idle_show, cfqd->cfq_group_idle, 1);
4570 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
4571 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
4572 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
4573 SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
4574 SHOW_FUNCTION(cfq_target_latency_show, cfqd->cfq_target_latency, 1);
4575 #undef SHOW_FUNCTION
4577 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
4578 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
4580 struct cfq_data *cfqd = e->elevator_data; \
4581 unsigned int __data; \
4582 int ret = cfq_var_store(&__data, (page), count); \
4583 if (__data < (MIN)) \
4585 else if (__data > (MAX)) \
4588 *(__PTR) = msecs_to_jiffies(__data); \
4590 *(__PTR) = __data; \
4593 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
4594 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
4596 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
4598 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
4599 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
4601 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
4602 STORE_FUNCTION(cfq_group_idle_store, &cfqd->cfq_group_idle, 0, UINT_MAX, 1);
4603 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
4604 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
4605 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
4607 STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
4608 STORE_FUNCTION(cfq_target_latency_store, &cfqd->cfq_target_latency, 1, UINT_MAX, 1);
4609 #undef STORE_FUNCTION
4611 #define CFQ_ATTR(name) \
4612 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
4614 static struct elv_fs_entry cfq_attrs[] = {
4616 CFQ_ATTR(fifo_expire_sync),
4617 CFQ_ATTR(fifo_expire_async),
4618 CFQ_ATTR(back_seek_max),
4619 CFQ_ATTR(back_seek_penalty),
4620 CFQ_ATTR(slice_sync),
4621 CFQ_ATTR(slice_async),
4622 CFQ_ATTR(slice_async_rq),
4623 CFQ_ATTR(slice_idle),
4624 CFQ_ATTR(group_idle),
4625 CFQ_ATTR(low_latency),
4626 CFQ_ATTR(target_latency),
4630 static struct elevator_type iosched_cfq = {
4632 .elevator_merge_fn = cfq_merge,
4633 .elevator_merged_fn = cfq_merged_request,
4634 .elevator_merge_req_fn = cfq_merged_requests,
4635 .elevator_allow_merge_fn = cfq_allow_merge,
4636 .elevator_bio_merged_fn = cfq_bio_merged,
4637 .elevator_dispatch_fn = cfq_dispatch_requests,
4638 .elevator_add_req_fn = cfq_insert_request,
4639 .elevator_activate_req_fn = cfq_activate_request,
4640 .elevator_deactivate_req_fn = cfq_deactivate_request,
4641 .elevator_completed_req_fn = cfq_completed_request,
4642 .elevator_former_req_fn = elv_rb_former_request,
4643 .elevator_latter_req_fn = elv_rb_latter_request,
4644 .elevator_init_icq_fn = cfq_init_icq,
4645 .elevator_exit_icq_fn = cfq_exit_icq,
4646 .elevator_set_req_fn = cfq_set_request,
4647 .elevator_put_req_fn = cfq_put_request,
4648 .elevator_may_queue_fn = cfq_may_queue,
4649 .elevator_init_fn = cfq_init_queue,
4650 .elevator_exit_fn = cfq_exit_queue,
4651 .elevator_registered_fn = cfq_registered_queue,
4653 .icq_size = sizeof(struct cfq_io_cq),
4654 .icq_align = __alignof__(struct cfq_io_cq),
4655 .elevator_attrs = cfq_attrs,
4656 .elevator_name = "cfq",
4657 .elevator_owner = THIS_MODULE,
4660 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4661 static struct blkcg_policy blkcg_policy_cfq = {
4662 .legacy_cftypes = cfq_blkcg_legacy_files,
4664 .cpd_alloc_fn = cfq_cpd_alloc,
4665 .cpd_init_fn = cfq_cpd_init,
4666 .cpd_free_fn = cfq_cpd_free,
4668 .pd_alloc_fn = cfq_pd_alloc,
4669 .pd_init_fn = cfq_pd_init,
4670 .pd_offline_fn = cfq_pd_offline,
4671 .pd_free_fn = cfq_pd_free,
4672 .pd_reset_stats_fn = cfq_pd_reset_stats,
4676 static int __init cfq_init(void)
4681 * could be 0 on HZ < 1000 setups
4683 if (!cfq_slice_async)
4684 cfq_slice_async = 1;
4685 if (!cfq_slice_idle)
4688 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4689 if (!cfq_group_idle)
4692 ret = blkcg_policy_register(&blkcg_policy_cfq);
4700 cfq_pool = KMEM_CACHE(cfq_queue, 0);
4704 ret = elv_register(&iosched_cfq);
4711 kmem_cache_destroy(cfq_pool);
4713 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4714 blkcg_policy_unregister(&blkcg_policy_cfq);
4719 static void __exit cfq_exit(void)
4721 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4722 blkcg_policy_unregister(&blkcg_policy_cfq);
4724 elv_unregister(&iosched_cfq);
4725 kmem_cache_destroy(cfq_pool);
4728 module_init(cfq_init);
4729 module_exit(cfq_exit);
4731 MODULE_AUTHOR("Jens Axboe");
4732 MODULE_LICENSE("GPL");
4733 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");