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;
1751 ret = blkg_conf_prep(blkcg, &blkcg_policy_cfq, buf, &ctx);
1756 cfqg = blkg_to_cfqg(ctx.blkg);
1757 cfqgd = blkcg_to_cfqgd(blkcg);
1758 if (!cfqg || !cfqgd)
1761 if (!ctx.v || (ctx.v >= CFQ_WEIGHT_MIN && ctx.v <= CFQ_WEIGHT_MAX)) {
1762 if (!is_leaf_weight) {
1763 cfqg->dev_weight = ctx.v;
1764 cfqg->new_weight = ctx.v ?: cfqgd->weight;
1766 cfqg->dev_leaf_weight = ctx.v;
1767 cfqg->new_leaf_weight = ctx.v ?: cfqgd->leaf_weight;
1773 blkg_conf_finish(&ctx);
1774 return ret ?: nbytes;
1777 static ssize_t cfqg_set_weight_device(struct kernfs_open_file *of,
1778 char *buf, size_t nbytes, loff_t off)
1780 return __cfqg_set_weight_device(of, buf, nbytes, off, false);
1783 static ssize_t cfqg_set_leaf_weight_device(struct kernfs_open_file *of,
1784 char *buf, size_t nbytes, loff_t off)
1786 return __cfqg_set_weight_device(of, buf, nbytes, off, true);
1789 static int __cfq_set_weight(struct cgroup_subsys_state *css, struct cftype *cft,
1790 u64 val, bool is_leaf_weight)
1792 struct blkcg *blkcg = css_to_blkcg(css);
1793 struct blkcg_gq *blkg;
1794 struct cfq_group_data *cfqgd;
1797 if (val < CFQ_WEIGHT_MIN || val > CFQ_WEIGHT_MAX)
1800 spin_lock_irq(&blkcg->lock);
1801 cfqgd = blkcg_to_cfqgd(blkcg);
1807 if (!is_leaf_weight)
1808 cfqgd->weight = val;
1810 cfqgd->leaf_weight = val;
1812 hlist_for_each_entry(blkg, &blkcg->blkg_list, blkcg_node) {
1813 struct cfq_group *cfqg = blkg_to_cfqg(blkg);
1818 if (!is_leaf_weight) {
1819 if (!cfqg->dev_weight)
1820 cfqg->new_weight = cfqgd->weight;
1822 if (!cfqg->dev_leaf_weight)
1823 cfqg->new_leaf_weight = cfqgd->leaf_weight;
1828 spin_unlock_irq(&blkcg->lock);
1832 static int cfq_set_weight(struct cgroup_subsys_state *css, struct cftype *cft,
1835 return __cfq_set_weight(css, cft, val, false);
1838 static int cfq_set_leaf_weight(struct cgroup_subsys_state *css,
1839 struct cftype *cft, u64 val)
1841 return __cfq_set_weight(css, cft, val, true);
1844 static int cfqg_print_stat(struct seq_file *sf, void *v)
1846 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), blkg_prfill_stat,
1847 &blkcg_policy_cfq, seq_cft(sf)->private, false);
1851 static int cfqg_print_rwstat(struct seq_file *sf, void *v)
1853 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), blkg_prfill_rwstat,
1854 &blkcg_policy_cfq, seq_cft(sf)->private, true);
1858 static u64 cfqg_prfill_stat_recursive(struct seq_file *sf,
1859 struct blkg_policy_data *pd, int off)
1861 u64 sum = blkg_stat_recursive_sum(pd_to_blkg(pd),
1862 &blkcg_policy_cfq, off);
1863 return __blkg_prfill_u64(sf, pd, sum);
1866 static u64 cfqg_prfill_rwstat_recursive(struct seq_file *sf,
1867 struct blkg_policy_data *pd, int off)
1869 struct blkg_rwstat sum = blkg_rwstat_recursive_sum(pd_to_blkg(pd),
1870 &blkcg_policy_cfq, off);
1871 return __blkg_prfill_rwstat(sf, pd, &sum);
1874 static int cfqg_print_stat_recursive(struct seq_file *sf, void *v)
1876 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1877 cfqg_prfill_stat_recursive, &blkcg_policy_cfq,
1878 seq_cft(sf)->private, false);
1882 static int cfqg_print_rwstat_recursive(struct seq_file *sf, void *v)
1884 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1885 cfqg_prfill_rwstat_recursive, &blkcg_policy_cfq,
1886 seq_cft(sf)->private, true);
1890 static u64 cfqg_prfill_sectors(struct seq_file *sf, struct blkg_policy_data *pd,
1893 u64 sum = blkg_rwstat_total(&pd->blkg->stat_bytes);
1895 return __blkg_prfill_u64(sf, pd, sum >> 9);
1898 static int cfqg_print_stat_sectors(struct seq_file *sf, void *v)
1900 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1901 cfqg_prfill_sectors, &blkcg_policy_cfq, 0, false);
1905 static u64 cfqg_prfill_sectors_recursive(struct seq_file *sf,
1906 struct blkg_policy_data *pd, int off)
1908 struct blkg_rwstat tmp = blkg_rwstat_recursive_sum(pd->blkg, NULL,
1909 offsetof(struct blkcg_gq, stat_bytes));
1910 u64 sum = atomic64_read(&tmp.aux_cnt[BLKG_RWSTAT_READ]) +
1911 atomic64_read(&tmp.aux_cnt[BLKG_RWSTAT_WRITE]);
1913 return __blkg_prfill_u64(sf, pd, sum >> 9);
1916 static int cfqg_print_stat_sectors_recursive(struct seq_file *sf, void *v)
1918 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1919 cfqg_prfill_sectors_recursive, &blkcg_policy_cfq, 0,
1924 #ifdef CONFIG_DEBUG_BLK_CGROUP
1925 static u64 cfqg_prfill_avg_queue_size(struct seq_file *sf,
1926 struct blkg_policy_data *pd, int off)
1928 struct cfq_group *cfqg = pd_to_cfqg(pd);
1929 u64 samples = blkg_stat_read(&cfqg->stats.avg_queue_size_samples);
1933 v = blkg_stat_read(&cfqg->stats.avg_queue_size_sum);
1934 v = div64_u64(v, samples);
1936 __blkg_prfill_u64(sf, pd, v);
1940 /* print avg_queue_size */
1941 static int cfqg_print_avg_queue_size(struct seq_file *sf, void *v)
1943 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1944 cfqg_prfill_avg_queue_size, &blkcg_policy_cfq,
1948 #endif /* CONFIG_DEBUG_BLK_CGROUP */
1950 static struct cftype cfq_blkcg_files[] = {
1951 /* on root, weight is mapped to leaf_weight */
1953 .name = "weight_device",
1954 .flags = CFTYPE_ONLY_ON_ROOT,
1955 .seq_show = cfqg_print_leaf_weight_device,
1956 .write = cfqg_set_leaf_weight_device,
1960 .flags = CFTYPE_ONLY_ON_ROOT,
1961 .seq_show = cfq_print_leaf_weight,
1962 .write_u64 = cfq_set_leaf_weight,
1965 /* no such mapping necessary for !roots */
1967 .name = "weight_device",
1968 .flags = CFTYPE_NOT_ON_ROOT,
1969 .seq_show = cfqg_print_weight_device,
1970 .write = cfqg_set_weight_device,
1974 .flags = CFTYPE_NOT_ON_ROOT,
1975 .seq_show = cfq_print_weight,
1976 .write_u64 = cfq_set_weight,
1980 .name = "leaf_weight_device",
1981 .seq_show = cfqg_print_leaf_weight_device,
1982 .write = cfqg_set_leaf_weight_device,
1985 .name = "leaf_weight",
1986 .seq_show = cfq_print_leaf_weight,
1987 .write_u64 = cfq_set_leaf_weight,
1990 /* statistics, covers only the tasks in the cfqg */
1993 .private = offsetof(struct cfq_group, stats.time),
1994 .seq_show = cfqg_print_stat,
1998 .seq_show = cfqg_print_stat_sectors,
2001 .name = "io_service_bytes",
2002 .private = (unsigned long)&blkcg_policy_cfq,
2003 .seq_show = blkg_print_stat_bytes,
2006 .name = "io_serviced",
2007 .private = (unsigned long)&blkcg_policy_cfq,
2008 .seq_show = blkg_print_stat_ios,
2011 .name = "io_service_time",
2012 .private = offsetof(struct cfq_group, stats.service_time),
2013 .seq_show = cfqg_print_rwstat,
2016 .name = "io_wait_time",
2017 .private = offsetof(struct cfq_group, stats.wait_time),
2018 .seq_show = cfqg_print_rwstat,
2021 .name = "io_merged",
2022 .private = offsetof(struct cfq_group, stats.merged),
2023 .seq_show = cfqg_print_rwstat,
2026 .name = "io_queued",
2027 .private = offsetof(struct cfq_group, stats.queued),
2028 .seq_show = cfqg_print_rwstat,
2031 /* the same statictics which cover the cfqg and its descendants */
2033 .name = "time_recursive",
2034 .private = offsetof(struct cfq_group, stats.time),
2035 .seq_show = cfqg_print_stat_recursive,
2038 .name = "sectors_recursive",
2039 .seq_show = cfqg_print_stat_sectors_recursive,
2042 .name = "io_service_bytes_recursive",
2043 .private = (unsigned long)&blkcg_policy_cfq,
2044 .seq_show = blkg_print_stat_bytes_recursive,
2047 .name = "io_serviced_recursive",
2048 .private = (unsigned long)&blkcg_policy_cfq,
2049 .seq_show = blkg_print_stat_ios_recursive,
2052 .name = "io_service_time_recursive",
2053 .private = offsetof(struct cfq_group, stats.service_time),
2054 .seq_show = cfqg_print_rwstat_recursive,
2057 .name = "io_wait_time_recursive",
2058 .private = offsetof(struct cfq_group, stats.wait_time),
2059 .seq_show = cfqg_print_rwstat_recursive,
2062 .name = "io_merged_recursive",
2063 .private = offsetof(struct cfq_group, stats.merged),
2064 .seq_show = cfqg_print_rwstat_recursive,
2067 .name = "io_queued_recursive",
2068 .private = offsetof(struct cfq_group, stats.queued),
2069 .seq_show = cfqg_print_rwstat_recursive,
2071 #ifdef CONFIG_DEBUG_BLK_CGROUP
2073 .name = "avg_queue_size",
2074 .seq_show = cfqg_print_avg_queue_size,
2077 .name = "group_wait_time",
2078 .private = offsetof(struct cfq_group, stats.group_wait_time),
2079 .seq_show = cfqg_print_stat,
2082 .name = "idle_time",
2083 .private = offsetof(struct cfq_group, stats.idle_time),
2084 .seq_show = cfqg_print_stat,
2087 .name = "empty_time",
2088 .private = offsetof(struct cfq_group, stats.empty_time),
2089 .seq_show = cfqg_print_stat,
2093 .private = offsetof(struct cfq_group, stats.dequeue),
2094 .seq_show = cfqg_print_stat,
2097 .name = "unaccounted_time",
2098 .private = offsetof(struct cfq_group, stats.unaccounted_time),
2099 .seq_show = cfqg_print_stat,
2101 #endif /* CONFIG_DEBUG_BLK_CGROUP */
2104 #else /* GROUP_IOSCHED */
2105 static struct cfq_group *cfq_lookup_cfqg(struct cfq_data *cfqd,
2106 struct blkcg *blkcg)
2108 return cfqd->root_group;
2112 cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) {
2116 #endif /* GROUP_IOSCHED */
2119 * The cfqd->service_trees holds all pending cfq_queue's that have
2120 * requests waiting to be processed. It is sorted in the order that
2121 * we will service the queues.
2123 static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2126 struct rb_node **p, *parent;
2127 struct cfq_queue *__cfqq;
2128 unsigned long rb_key;
2129 struct cfq_rb_root *st;
2133 st = st_for(cfqq->cfqg, cfqq_class(cfqq), cfqq_type(cfqq));
2134 if (cfq_class_idle(cfqq)) {
2135 rb_key = CFQ_IDLE_DELAY;
2136 parent = rb_last(&st->rb);
2137 if (parent && parent != &cfqq->rb_node) {
2138 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
2139 rb_key += __cfqq->rb_key;
2142 } else if (!add_front) {
2144 * Get our rb key offset. Subtract any residual slice
2145 * value carried from last service. A negative resid
2146 * count indicates slice overrun, and this should position
2147 * the next service time further away in the tree.
2149 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
2150 rb_key -= cfqq->slice_resid;
2151 cfqq->slice_resid = 0;
2154 __cfqq = cfq_rb_first(st);
2155 rb_key += __cfqq ? __cfqq->rb_key : jiffies;
2158 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
2161 * same position, nothing more to do
2163 if (rb_key == cfqq->rb_key && cfqq->service_tree == st)
2166 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
2167 cfqq->service_tree = NULL;
2172 cfqq->service_tree = st;
2173 p = &st->rb.rb_node;
2176 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
2179 * sort by key, that represents service time.
2181 if (time_before(rb_key, __cfqq->rb_key))
2182 p = &parent->rb_left;
2184 p = &parent->rb_right;
2190 st->left = &cfqq->rb_node;
2192 cfqq->rb_key = rb_key;
2193 rb_link_node(&cfqq->rb_node, parent, p);
2194 rb_insert_color(&cfqq->rb_node, &st->rb);
2196 if (add_front || !new_cfqq)
2198 cfq_group_notify_queue_add(cfqd, cfqq->cfqg);
2201 static struct cfq_queue *
2202 cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
2203 sector_t sector, struct rb_node **ret_parent,
2204 struct rb_node ***rb_link)
2206 struct rb_node **p, *parent;
2207 struct cfq_queue *cfqq = NULL;
2215 cfqq = rb_entry(parent, struct cfq_queue, p_node);
2218 * Sort strictly based on sector. Smallest to the left,
2219 * largest to the right.
2221 if (sector > blk_rq_pos(cfqq->next_rq))
2222 n = &(*p)->rb_right;
2223 else if (sector < blk_rq_pos(cfqq->next_rq))
2231 *ret_parent = parent;
2237 static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2239 struct rb_node **p, *parent;
2240 struct cfq_queue *__cfqq;
2243 rb_erase(&cfqq->p_node, cfqq->p_root);
2244 cfqq->p_root = NULL;
2247 if (cfq_class_idle(cfqq))
2252 cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
2253 __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
2254 blk_rq_pos(cfqq->next_rq), &parent, &p);
2256 rb_link_node(&cfqq->p_node, parent, p);
2257 rb_insert_color(&cfqq->p_node, cfqq->p_root);
2259 cfqq->p_root = NULL;
2263 * Update cfqq's position in the service tree.
2265 static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2268 * Resorting requires the cfqq to be on the RR list already.
2270 if (cfq_cfqq_on_rr(cfqq)) {
2271 cfq_service_tree_add(cfqd, cfqq, 0);
2272 cfq_prio_tree_add(cfqd, cfqq);
2277 * add to busy list of queues for service, trying to be fair in ordering
2278 * the pending list according to last request service
2280 static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2282 cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
2283 BUG_ON(cfq_cfqq_on_rr(cfqq));
2284 cfq_mark_cfqq_on_rr(cfqq);
2285 cfqd->busy_queues++;
2286 if (cfq_cfqq_sync(cfqq))
2287 cfqd->busy_sync_queues++;
2289 cfq_resort_rr_list(cfqd, cfqq);
2293 * Called when the cfqq no longer has requests pending, remove it from
2296 static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2298 cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
2299 BUG_ON(!cfq_cfqq_on_rr(cfqq));
2300 cfq_clear_cfqq_on_rr(cfqq);
2302 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
2303 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
2304 cfqq->service_tree = NULL;
2307 rb_erase(&cfqq->p_node, cfqq->p_root);
2308 cfqq->p_root = NULL;
2311 cfq_group_notify_queue_del(cfqd, cfqq->cfqg);
2312 BUG_ON(!cfqd->busy_queues);
2313 cfqd->busy_queues--;
2314 if (cfq_cfqq_sync(cfqq))
2315 cfqd->busy_sync_queues--;
2319 * rb tree support functions
2321 static void cfq_del_rq_rb(struct request *rq)
2323 struct cfq_queue *cfqq = RQ_CFQQ(rq);
2324 const int sync = rq_is_sync(rq);
2326 BUG_ON(!cfqq->queued[sync]);
2327 cfqq->queued[sync]--;
2329 elv_rb_del(&cfqq->sort_list, rq);
2331 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) {
2333 * Queue will be deleted from service tree when we actually
2334 * expire it later. Right now just remove it from prio tree
2338 rb_erase(&cfqq->p_node, cfqq->p_root);
2339 cfqq->p_root = NULL;
2344 static void cfq_add_rq_rb(struct request *rq)
2346 struct cfq_queue *cfqq = RQ_CFQQ(rq);
2347 struct cfq_data *cfqd = cfqq->cfqd;
2348 struct request *prev;
2350 cfqq->queued[rq_is_sync(rq)]++;
2352 elv_rb_add(&cfqq->sort_list, rq);
2354 if (!cfq_cfqq_on_rr(cfqq))
2355 cfq_add_cfqq_rr(cfqd, cfqq);
2358 * check if this request is a better next-serve candidate
2360 prev = cfqq->next_rq;
2361 cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq, cfqd->last_position);
2364 * adjust priority tree position, if ->next_rq changes
2366 if (prev != cfqq->next_rq)
2367 cfq_prio_tree_add(cfqd, cfqq);
2369 BUG_ON(!cfqq->next_rq);
2372 static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
2374 elv_rb_del(&cfqq->sort_list, rq);
2375 cfqq->queued[rq_is_sync(rq)]--;
2376 cfqg_stats_update_io_remove(RQ_CFQG(rq), rq->cmd_flags);
2378 cfqg_stats_update_io_add(RQ_CFQG(rq), cfqq->cfqd->serving_group,
2382 static struct request *
2383 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
2385 struct task_struct *tsk = current;
2386 struct cfq_io_cq *cic;
2387 struct cfq_queue *cfqq;
2389 cic = cfq_cic_lookup(cfqd, tsk->io_context);
2393 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
2395 return elv_rb_find(&cfqq->sort_list, bio_end_sector(bio));
2400 static void cfq_activate_request(struct request_queue *q, struct request *rq)
2402 struct cfq_data *cfqd = q->elevator->elevator_data;
2404 cfqd->rq_in_driver++;
2405 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
2406 cfqd->rq_in_driver);
2408 cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
2411 static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
2413 struct cfq_data *cfqd = q->elevator->elevator_data;
2415 WARN_ON(!cfqd->rq_in_driver);
2416 cfqd->rq_in_driver--;
2417 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
2418 cfqd->rq_in_driver);
2421 static void cfq_remove_request(struct request *rq)
2423 struct cfq_queue *cfqq = RQ_CFQQ(rq);
2425 if (cfqq->next_rq == rq)
2426 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
2428 list_del_init(&rq->queuelist);
2431 cfqq->cfqd->rq_queued--;
2432 cfqg_stats_update_io_remove(RQ_CFQG(rq), rq->cmd_flags);
2433 if (rq->cmd_flags & REQ_PRIO) {
2434 WARN_ON(!cfqq->prio_pending);
2435 cfqq->prio_pending--;
2439 static int cfq_merge(struct request_queue *q, struct request **req,
2442 struct cfq_data *cfqd = q->elevator->elevator_data;
2443 struct request *__rq;
2445 __rq = cfq_find_rq_fmerge(cfqd, bio);
2446 if (__rq && elv_rq_merge_ok(__rq, bio)) {
2448 return ELEVATOR_FRONT_MERGE;
2451 return ELEVATOR_NO_MERGE;
2454 static void cfq_merged_request(struct request_queue *q, struct request *req,
2457 if (type == ELEVATOR_FRONT_MERGE) {
2458 struct cfq_queue *cfqq = RQ_CFQQ(req);
2460 cfq_reposition_rq_rb(cfqq, req);
2464 static void cfq_bio_merged(struct request_queue *q, struct request *req,
2467 cfqg_stats_update_io_merged(RQ_CFQG(req), bio->bi_rw);
2471 cfq_merged_requests(struct request_queue *q, struct request *rq,
2472 struct request *next)
2474 struct cfq_queue *cfqq = RQ_CFQQ(rq);
2475 struct cfq_data *cfqd = q->elevator->elevator_data;
2478 * reposition in fifo if next is older than rq
2480 if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
2481 time_before(next->fifo_time, rq->fifo_time) &&
2482 cfqq == RQ_CFQQ(next)) {
2483 list_move(&rq->queuelist, &next->queuelist);
2484 rq->fifo_time = next->fifo_time;
2487 if (cfqq->next_rq == next)
2489 cfq_remove_request(next);
2490 cfqg_stats_update_io_merged(RQ_CFQG(rq), next->cmd_flags);
2492 cfqq = RQ_CFQQ(next);
2494 * all requests of this queue are merged to other queues, delete it
2495 * from the service tree. If it's the active_queue,
2496 * cfq_dispatch_requests() will choose to expire it or do idle
2498 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list) &&
2499 cfqq != cfqd->active_queue)
2500 cfq_del_cfqq_rr(cfqd, cfqq);
2503 static int cfq_allow_merge(struct request_queue *q, struct request *rq,
2506 struct cfq_data *cfqd = q->elevator->elevator_data;
2507 struct cfq_io_cq *cic;
2508 struct cfq_queue *cfqq;
2511 * Disallow merge of a sync bio into an async request.
2513 if (cfq_bio_sync(bio) && !rq_is_sync(rq))
2517 * Lookup the cfqq that this bio will be queued with and allow
2518 * merge only if rq is queued there.
2520 cic = cfq_cic_lookup(cfqd, current->io_context);
2524 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
2525 return cfqq == RQ_CFQQ(rq);
2528 static inline void cfq_del_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2530 del_timer(&cfqd->idle_slice_timer);
2531 cfqg_stats_update_idle_time(cfqq->cfqg);
2534 static void __cfq_set_active_queue(struct cfq_data *cfqd,
2535 struct cfq_queue *cfqq)
2538 cfq_log_cfqq(cfqd, cfqq, "set_active wl_class:%d wl_type:%d",
2539 cfqd->serving_wl_class, cfqd->serving_wl_type);
2540 cfqg_stats_update_avg_queue_size(cfqq->cfqg);
2541 cfqq->slice_start = 0;
2542 cfqq->dispatch_start = jiffies;
2543 cfqq->allocated_slice = 0;
2544 cfqq->slice_end = 0;
2545 cfqq->slice_dispatch = 0;
2546 cfqq->nr_sectors = 0;
2548 cfq_clear_cfqq_wait_request(cfqq);
2549 cfq_clear_cfqq_must_dispatch(cfqq);
2550 cfq_clear_cfqq_must_alloc_slice(cfqq);
2551 cfq_clear_cfqq_fifo_expire(cfqq);
2552 cfq_mark_cfqq_slice_new(cfqq);
2554 cfq_del_timer(cfqd, cfqq);
2557 cfqd->active_queue = cfqq;
2561 * current cfqq expired its slice (or was too idle), select new one
2564 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2567 cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
2569 if (cfq_cfqq_wait_request(cfqq))
2570 cfq_del_timer(cfqd, cfqq);
2572 cfq_clear_cfqq_wait_request(cfqq);
2573 cfq_clear_cfqq_wait_busy(cfqq);
2576 * If this cfqq is shared between multiple processes, check to
2577 * make sure that those processes are still issuing I/Os within
2578 * the mean seek distance. If not, it may be time to break the
2579 * queues apart again.
2581 if (cfq_cfqq_coop(cfqq) && CFQQ_SEEKY(cfqq))
2582 cfq_mark_cfqq_split_coop(cfqq);
2585 * store what was left of this slice, if the queue idled/timed out
2588 if (cfq_cfqq_slice_new(cfqq))
2589 cfqq->slice_resid = cfq_scaled_cfqq_slice(cfqd, cfqq);
2591 cfqq->slice_resid = cfqq->slice_end - jiffies;
2592 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
2595 cfq_group_served(cfqd, cfqq->cfqg, cfqq);
2597 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
2598 cfq_del_cfqq_rr(cfqd, cfqq);
2600 cfq_resort_rr_list(cfqd, cfqq);
2602 if (cfqq == cfqd->active_queue)
2603 cfqd->active_queue = NULL;
2605 if (cfqd->active_cic) {
2606 put_io_context(cfqd->active_cic->icq.ioc);
2607 cfqd->active_cic = NULL;
2611 static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out)
2613 struct cfq_queue *cfqq = cfqd->active_queue;
2616 __cfq_slice_expired(cfqd, cfqq, timed_out);
2620 * Get next queue for service. Unless we have a queue preemption,
2621 * we'll simply select the first cfqq in the service tree.
2623 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
2625 struct cfq_rb_root *st = st_for(cfqd->serving_group,
2626 cfqd->serving_wl_class, cfqd->serving_wl_type);
2628 if (!cfqd->rq_queued)
2631 /* There is nothing to dispatch */
2634 if (RB_EMPTY_ROOT(&st->rb))
2636 return cfq_rb_first(st);
2639 static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd)
2641 struct cfq_group *cfqg;
2642 struct cfq_queue *cfqq;
2644 struct cfq_rb_root *st;
2646 if (!cfqd->rq_queued)
2649 cfqg = cfq_get_next_cfqg(cfqd);
2653 for_each_cfqg_st(cfqg, i, j, st)
2654 if ((cfqq = cfq_rb_first(st)) != NULL)
2660 * Get and set a new active queue for service.
2662 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
2663 struct cfq_queue *cfqq)
2666 cfqq = cfq_get_next_queue(cfqd);
2668 __cfq_set_active_queue(cfqd, cfqq);
2672 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
2675 if (blk_rq_pos(rq) >= cfqd->last_position)
2676 return blk_rq_pos(rq) - cfqd->last_position;
2678 return cfqd->last_position - blk_rq_pos(rq);
2681 static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2684 return cfq_dist_from_last(cfqd, rq) <= CFQQ_CLOSE_THR;
2687 static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
2688 struct cfq_queue *cur_cfqq)
2690 struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
2691 struct rb_node *parent, *node;
2692 struct cfq_queue *__cfqq;
2693 sector_t sector = cfqd->last_position;
2695 if (RB_EMPTY_ROOT(root))
2699 * First, if we find a request starting at the end of the last
2700 * request, choose it.
2702 __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
2707 * If the exact sector wasn't found, the parent of the NULL leaf
2708 * will contain the closest sector.
2710 __cfqq = rb_entry(parent, struct cfq_queue, p_node);
2711 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
2714 if (blk_rq_pos(__cfqq->next_rq) < sector)
2715 node = rb_next(&__cfqq->p_node);
2717 node = rb_prev(&__cfqq->p_node);
2721 __cfqq = rb_entry(node, struct cfq_queue, p_node);
2722 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
2730 * cur_cfqq - passed in so that we don't decide that the current queue is
2731 * closely cooperating with itself.
2733 * So, basically we're assuming that that cur_cfqq has dispatched at least
2734 * one request, and that cfqd->last_position reflects a position on the disk
2735 * associated with the I/O issued by cur_cfqq. I'm not sure this is a valid
2738 static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
2739 struct cfq_queue *cur_cfqq)
2741 struct cfq_queue *cfqq;
2743 if (cfq_class_idle(cur_cfqq))
2745 if (!cfq_cfqq_sync(cur_cfqq))
2747 if (CFQQ_SEEKY(cur_cfqq))
2751 * Don't search priority tree if it's the only queue in the group.
2753 if (cur_cfqq->cfqg->nr_cfqq == 1)
2757 * We should notice if some of the queues are cooperating, eg
2758 * working closely on the same area of the disk. In that case,
2759 * we can group them together and don't waste time idling.
2761 cfqq = cfqq_close(cfqd, cur_cfqq);
2765 /* If new queue belongs to different cfq_group, don't choose it */
2766 if (cur_cfqq->cfqg != cfqq->cfqg)
2770 * It only makes sense to merge sync queues.
2772 if (!cfq_cfqq_sync(cfqq))
2774 if (CFQQ_SEEKY(cfqq))
2778 * Do not merge queues of different priority classes
2780 if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq))
2787 * Determine whether we should enforce idle window for this queue.
2790 static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2792 enum wl_class_t wl_class = cfqq_class(cfqq);
2793 struct cfq_rb_root *st = cfqq->service_tree;
2798 if (!cfqd->cfq_slice_idle)
2801 /* We never do for idle class queues. */
2802 if (wl_class == IDLE_WORKLOAD)
2805 /* We do for queues that were marked with idle window flag. */
2806 if (cfq_cfqq_idle_window(cfqq) &&
2807 !(blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag))
2811 * Otherwise, we do only if they are the last ones
2812 * in their service tree.
2814 if (st->count == 1 && cfq_cfqq_sync(cfqq) &&
2815 !cfq_io_thinktime_big(cfqd, &st->ttime, false))
2817 cfq_log_cfqq(cfqd, cfqq, "Not idling. st->count:%d", st->count);
2821 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
2823 struct cfq_queue *cfqq = cfqd->active_queue;
2824 struct cfq_io_cq *cic;
2825 unsigned long sl, group_idle = 0;
2828 * SSD device without seek penalty, disable idling. But only do so
2829 * for devices that support queuing, otherwise we still have a problem
2830 * with sync vs async workloads.
2832 if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
2835 WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
2836 WARN_ON(cfq_cfqq_slice_new(cfqq));
2839 * idle is disabled, either manually or by past process history
2841 if (!cfq_should_idle(cfqd, cfqq)) {
2842 /* no queue idling. Check for group idling */
2843 if (cfqd->cfq_group_idle)
2844 group_idle = cfqd->cfq_group_idle;
2850 * still active requests from this queue, don't idle
2852 if (cfqq->dispatched)
2856 * task has exited, don't wait
2858 cic = cfqd->active_cic;
2859 if (!cic || !atomic_read(&cic->icq.ioc->active_ref))
2863 * If our average think time is larger than the remaining time
2864 * slice, then don't idle. This avoids overrunning the allotted
2867 if (sample_valid(cic->ttime.ttime_samples) &&
2868 (cfqq->slice_end - jiffies < cic->ttime.ttime_mean)) {
2869 cfq_log_cfqq(cfqd, cfqq, "Not idling. think_time:%lu",
2870 cic->ttime.ttime_mean);
2874 /* There are other queues in the group, don't do group idle */
2875 if (group_idle && cfqq->cfqg->nr_cfqq > 1)
2878 cfq_mark_cfqq_wait_request(cfqq);
2881 sl = cfqd->cfq_group_idle;
2883 sl = cfqd->cfq_slice_idle;
2885 mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
2886 cfqg_stats_set_start_idle_time(cfqq->cfqg);
2887 cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu group_idle: %d", sl,
2888 group_idle ? 1 : 0);
2892 * Move request from internal lists to the request queue dispatch list.
2894 static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
2896 struct cfq_data *cfqd = q->elevator->elevator_data;
2897 struct cfq_queue *cfqq = RQ_CFQQ(rq);
2899 cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
2901 cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
2902 cfq_remove_request(rq);
2904 (RQ_CFQG(rq))->dispatched++;
2905 elv_dispatch_sort(q, rq);
2907 cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]++;
2908 cfqq->nr_sectors += blk_rq_sectors(rq);
2912 * return expired entry, or NULL to just start from scratch in rbtree
2914 static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
2916 struct request *rq = NULL;
2918 if (cfq_cfqq_fifo_expire(cfqq))
2921 cfq_mark_cfqq_fifo_expire(cfqq);
2923 if (list_empty(&cfqq->fifo))
2926 rq = rq_entry_fifo(cfqq->fifo.next);
2927 if (time_before(jiffies, rq->fifo_time))
2930 cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq);
2935 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2937 const int base_rq = cfqd->cfq_slice_async_rq;
2939 WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
2941 return 2 * base_rq * (IOPRIO_BE_NR - cfqq->ioprio);
2945 * Must be called with the queue_lock held.
2947 static int cfqq_process_refs(struct cfq_queue *cfqq)
2949 int process_refs, io_refs;
2951 io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE];
2952 process_refs = cfqq->ref - io_refs;
2953 BUG_ON(process_refs < 0);
2954 return process_refs;
2957 static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq)
2959 int process_refs, new_process_refs;
2960 struct cfq_queue *__cfqq;
2963 * If there are no process references on the new_cfqq, then it is
2964 * unsafe to follow the ->new_cfqq chain as other cfqq's in the
2965 * chain may have dropped their last reference (not just their
2966 * last process reference).
2968 if (!cfqq_process_refs(new_cfqq))
2971 /* Avoid a circular list and skip interim queue merges */
2972 while ((__cfqq = new_cfqq->new_cfqq)) {
2978 process_refs = cfqq_process_refs(cfqq);
2979 new_process_refs = cfqq_process_refs(new_cfqq);
2981 * If the process for the cfqq has gone away, there is no
2982 * sense in merging the queues.
2984 if (process_refs == 0 || new_process_refs == 0)
2988 * Merge in the direction of the lesser amount of work.
2990 if (new_process_refs >= process_refs) {
2991 cfqq->new_cfqq = new_cfqq;
2992 new_cfqq->ref += process_refs;
2994 new_cfqq->new_cfqq = cfqq;
2995 cfqq->ref += new_process_refs;
2999 static enum wl_type_t cfq_choose_wl_type(struct cfq_data *cfqd,
3000 struct cfq_group *cfqg, enum wl_class_t wl_class)
3002 struct cfq_queue *queue;
3004 bool key_valid = false;
3005 unsigned long lowest_key = 0;
3006 enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD;
3008 for (i = 0; i <= SYNC_WORKLOAD; ++i) {
3009 /* select the one with lowest rb_key */
3010 queue = cfq_rb_first(st_for(cfqg, wl_class, i));
3012 (!key_valid || time_before(queue->rb_key, lowest_key))) {
3013 lowest_key = queue->rb_key;
3023 choose_wl_class_and_type(struct cfq_data *cfqd, struct cfq_group *cfqg)
3027 struct cfq_rb_root *st;
3028 unsigned group_slice;
3029 enum wl_class_t original_class = cfqd->serving_wl_class;
3031 /* Choose next priority. RT > BE > IDLE */
3032 if (cfq_group_busy_queues_wl(RT_WORKLOAD, cfqd, cfqg))
3033 cfqd->serving_wl_class = RT_WORKLOAD;
3034 else if (cfq_group_busy_queues_wl(BE_WORKLOAD, cfqd, cfqg))
3035 cfqd->serving_wl_class = BE_WORKLOAD;
3037 cfqd->serving_wl_class = IDLE_WORKLOAD;
3038 cfqd->workload_expires = jiffies + 1;
3042 if (original_class != cfqd->serving_wl_class)
3046 * For RT and BE, we have to choose also the type
3047 * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
3050 st = st_for(cfqg, cfqd->serving_wl_class, cfqd->serving_wl_type);
3054 * check workload expiration, and that we still have other queues ready
3056 if (count && !time_after(jiffies, cfqd->workload_expires))
3060 /* otherwise select new workload type */
3061 cfqd->serving_wl_type = cfq_choose_wl_type(cfqd, cfqg,
3062 cfqd->serving_wl_class);
3063 st = st_for(cfqg, cfqd->serving_wl_class, cfqd->serving_wl_type);
3067 * the workload slice is computed as a fraction of target latency
3068 * proportional to the number of queues in that workload, over
3069 * all the queues in the same priority class
3071 group_slice = cfq_group_slice(cfqd, cfqg);
3073 slice = group_slice * count /
3074 max_t(unsigned, cfqg->busy_queues_avg[cfqd->serving_wl_class],
3075 cfq_group_busy_queues_wl(cfqd->serving_wl_class, cfqd,
3078 if (cfqd->serving_wl_type == ASYNC_WORKLOAD) {
3082 * Async queues are currently system wide. Just taking
3083 * proportion of queues with-in same group will lead to higher
3084 * async ratio system wide as generally root group is going
3085 * to have higher weight. A more accurate thing would be to
3086 * calculate system wide asnc/sync ratio.
3088 tmp = cfqd->cfq_target_latency *
3089 cfqg_busy_async_queues(cfqd, cfqg);
3090 tmp = tmp/cfqd->busy_queues;
3091 slice = min_t(unsigned, slice, tmp);
3093 /* async workload slice is scaled down according to
3094 * the sync/async slice ratio. */
3095 slice = slice * cfqd->cfq_slice[0] / cfqd->cfq_slice[1];
3097 /* sync workload slice is at least 2 * cfq_slice_idle */
3098 slice = max(slice, 2 * cfqd->cfq_slice_idle);
3100 slice = max_t(unsigned, slice, CFQ_MIN_TT);
3101 cfq_log(cfqd, "workload slice:%d", slice);
3102 cfqd->workload_expires = jiffies + slice;
3105 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd)
3107 struct cfq_rb_root *st = &cfqd->grp_service_tree;
3108 struct cfq_group *cfqg;
3110 if (RB_EMPTY_ROOT(&st->rb))
3112 cfqg = cfq_rb_first_group(st);
3113 update_min_vdisktime(st);
3117 static void cfq_choose_cfqg(struct cfq_data *cfqd)
3119 struct cfq_group *cfqg = cfq_get_next_cfqg(cfqd);
3121 cfqd->serving_group = cfqg;
3123 /* Restore the workload type data */
3124 if (cfqg->saved_wl_slice) {
3125 cfqd->workload_expires = jiffies + cfqg->saved_wl_slice;
3126 cfqd->serving_wl_type = cfqg->saved_wl_type;
3127 cfqd->serving_wl_class = cfqg->saved_wl_class;
3129 cfqd->workload_expires = jiffies - 1;
3131 choose_wl_class_and_type(cfqd, cfqg);
3135 * Select a queue for service. If we have a current active queue,
3136 * check whether to continue servicing it, or retrieve and set a new one.
3138 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
3140 struct cfq_queue *cfqq, *new_cfqq = NULL;
3142 cfqq = cfqd->active_queue;
3146 if (!cfqd->rq_queued)
3150 * We were waiting for group to get backlogged. Expire the queue
3152 if (cfq_cfqq_wait_busy(cfqq) && !RB_EMPTY_ROOT(&cfqq->sort_list))
3156 * The active queue has run out of time, expire it and select new.
3158 if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq)) {
3160 * If slice had not expired at the completion of last request
3161 * we might not have turned on wait_busy flag. Don't expire
3162 * the queue yet. Allow the group to get backlogged.
3164 * The very fact that we have used the slice, that means we
3165 * have been idling all along on this queue and it should be
3166 * ok to wait for this request to complete.
3168 if (cfqq->cfqg->nr_cfqq == 1 && RB_EMPTY_ROOT(&cfqq->sort_list)
3169 && cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
3173 goto check_group_idle;
3177 * The active queue has requests and isn't expired, allow it to
3180 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3184 * If another queue has a request waiting within our mean seek
3185 * distance, let it run. The expire code will check for close
3186 * cooperators and put the close queue at the front of the service
3187 * tree. If possible, merge the expiring queue with the new cfqq.
3189 new_cfqq = cfq_close_cooperator(cfqd, cfqq);
3191 if (!cfqq->new_cfqq)
3192 cfq_setup_merge(cfqq, new_cfqq);
3197 * No requests pending. If the active queue still has requests in
3198 * flight or is idling for a new request, allow either of these
3199 * conditions to happen (or time out) before selecting a new queue.
3201 if (timer_pending(&cfqd->idle_slice_timer)) {
3207 * This is a deep seek queue, but the device is much faster than
3208 * the queue can deliver, don't idle
3210 if (CFQQ_SEEKY(cfqq) && cfq_cfqq_idle_window(cfqq) &&
3211 (cfq_cfqq_slice_new(cfqq) ||
3212 (cfqq->slice_end - jiffies > jiffies - cfqq->slice_start))) {
3213 cfq_clear_cfqq_deep(cfqq);
3214 cfq_clear_cfqq_idle_window(cfqq);
3217 if (cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
3223 * If group idle is enabled and there are requests dispatched from
3224 * this group, wait for requests to complete.
3227 if (cfqd->cfq_group_idle && cfqq->cfqg->nr_cfqq == 1 &&
3228 cfqq->cfqg->dispatched &&
3229 !cfq_io_thinktime_big(cfqd, &cfqq->cfqg->ttime, true)) {
3235 cfq_slice_expired(cfqd, 0);
3238 * Current queue expired. Check if we have to switch to a new
3242 cfq_choose_cfqg(cfqd);
3244 cfqq = cfq_set_active_queue(cfqd, new_cfqq);
3249 static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
3253 while (cfqq->next_rq) {
3254 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
3258 BUG_ON(!list_empty(&cfqq->fifo));
3260 /* By default cfqq is not expired if it is empty. Do it explicitly */
3261 __cfq_slice_expired(cfqq->cfqd, cfqq, 0);
3266 * Drain our current requests. Used for barriers and when switching
3267 * io schedulers on-the-fly.
3269 static int cfq_forced_dispatch(struct cfq_data *cfqd)
3271 struct cfq_queue *cfqq;
3274 /* Expire the timeslice of the current active queue first */
3275 cfq_slice_expired(cfqd, 0);
3276 while ((cfqq = cfq_get_next_queue_forced(cfqd)) != NULL) {
3277 __cfq_set_active_queue(cfqd, cfqq);
3278 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
3281 BUG_ON(cfqd->busy_queues);
3283 cfq_log(cfqd, "forced_dispatch=%d", dispatched);
3287 static inline bool cfq_slice_used_soon(struct cfq_data *cfqd,
3288 struct cfq_queue *cfqq)
3290 /* the queue hasn't finished any request, can't estimate */
3291 if (cfq_cfqq_slice_new(cfqq))
3293 if (time_after(jiffies + cfqd->cfq_slice_idle * cfqq->dispatched,
3300 static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3302 unsigned int max_dispatch;
3305 * Drain async requests before we start sync IO
3307 if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_flight[BLK_RW_ASYNC])
3311 * If this is an async queue and we have sync IO in flight, let it wait
3313 if (cfqd->rq_in_flight[BLK_RW_SYNC] && !cfq_cfqq_sync(cfqq))
3316 max_dispatch = max_t(unsigned int, cfqd->cfq_quantum / 2, 1);
3317 if (cfq_class_idle(cfqq))
3321 * Does this cfqq already have too much IO in flight?
3323 if (cfqq->dispatched >= max_dispatch) {
3324 bool promote_sync = false;
3326 * idle queue must always only have a single IO in flight
3328 if (cfq_class_idle(cfqq))
3332 * If there is only one sync queue
3333 * we can ignore async queue here and give the sync
3334 * queue no dispatch limit. The reason is a sync queue can
3335 * preempt async queue, limiting the sync queue doesn't make
3336 * sense. This is useful for aiostress test.
3338 if (cfq_cfqq_sync(cfqq) && cfqd->busy_sync_queues == 1)
3339 promote_sync = true;
3342 * We have other queues, don't allow more IO from this one
3344 if (cfqd->busy_queues > 1 && cfq_slice_used_soon(cfqd, cfqq) &&
3349 * Sole queue user, no limit
3351 if (cfqd->busy_queues == 1 || promote_sync)
3355 * Normally we start throttling cfqq when cfq_quantum/2
3356 * requests have been dispatched. But we can drive
3357 * deeper queue depths at the beginning of slice
3358 * subjected to upper limit of cfq_quantum.
3360 max_dispatch = cfqd->cfq_quantum;
3364 * Async queues must wait a bit before being allowed dispatch.
3365 * We also ramp up the dispatch depth gradually for async IO,
3366 * based on the last sync IO we serviced
3368 if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
3369 unsigned long last_sync = jiffies - cfqd->last_delayed_sync;
3372 depth = last_sync / cfqd->cfq_slice[1];
3373 if (!depth && !cfqq->dispatched)
3375 if (depth < max_dispatch)
3376 max_dispatch = depth;
3380 * If we're below the current max, allow a dispatch
3382 return cfqq->dispatched < max_dispatch;
3386 * Dispatch a request from cfqq, moving them to the request queue
3389 static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3393 BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
3395 if (!cfq_may_dispatch(cfqd, cfqq))
3399 * follow expired path, else get first next available
3401 rq = cfq_check_fifo(cfqq);
3406 * insert request into driver dispatch list
3408 cfq_dispatch_insert(cfqd->queue, rq);
3410 if (!cfqd->active_cic) {
3411 struct cfq_io_cq *cic = RQ_CIC(rq);
3413 atomic_long_inc(&cic->icq.ioc->refcount);
3414 cfqd->active_cic = cic;
3421 * Find the cfqq that we need to service and move a request from that to the
3424 static int cfq_dispatch_requests(struct request_queue *q, int force)
3426 struct cfq_data *cfqd = q->elevator->elevator_data;
3427 struct cfq_queue *cfqq;
3429 if (!cfqd->busy_queues)
3432 if (unlikely(force))
3433 return cfq_forced_dispatch(cfqd);
3435 cfqq = cfq_select_queue(cfqd);
3440 * Dispatch a request from this cfqq, if it is allowed
3442 if (!cfq_dispatch_request(cfqd, cfqq))
3445 cfqq->slice_dispatch++;
3446 cfq_clear_cfqq_must_dispatch(cfqq);
3449 * expire an async queue immediately if it has used up its slice. idle
3450 * queue always expire after 1 dispatch round.
3452 if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
3453 cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
3454 cfq_class_idle(cfqq))) {
3455 cfqq->slice_end = jiffies + 1;
3456 cfq_slice_expired(cfqd, 0);
3459 cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
3464 * task holds one reference to the queue, dropped when task exits. each rq
3465 * in-flight on this queue also holds a reference, dropped when rq is freed.
3467 * Each cfq queue took a reference on the parent group. Drop it now.
3468 * queue lock must be held here.
3470 static void cfq_put_queue(struct cfq_queue *cfqq)
3472 struct cfq_data *cfqd = cfqq->cfqd;
3473 struct cfq_group *cfqg;
3475 BUG_ON(cfqq->ref <= 0);
3481 cfq_log_cfqq(cfqd, cfqq, "put_queue");
3482 BUG_ON(rb_first(&cfqq->sort_list));
3483 BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
3486 if (unlikely(cfqd->active_queue == cfqq)) {
3487 __cfq_slice_expired(cfqd, cfqq, 0);
3488 cfq_schedule_dispatch(cfqd);
3491 BUG_ON(cfq_cfqq_on_rr(cfqq));
3492 kmem_cache_free(cfq_pool, cfqq);
3496 static void cfq_put_cooperator(struct cfq_queue *cfqq)
3498 struct cfq_queue *__cfqq, *next;
3501 * If this queue was scheduled to merge with another queue, be
3502 * sure to drop the reference taken on that queue (and others in
3503 * the merge chain). See cfq_setup_merge and cfq_merge_cfqqs.
3505 __cfqq = cfqq->new_cfqq;
3507 if (__cfqq == cfqq) {
3508 WARN(1, "cfqq->new_cfqq loop detected\n");
3511 next = __cfqq->new_cfqq;
3512 cfq_put_queue(__cfqq);
3517 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3519 if (unlikely(cfqq == cfqd->active_queue)) {
3520 __cfq_slice_expired(cfqd, cfqq, 0);
3521 cfq_schedule_dispatch(cfqd);
3524 cfq_put_cooperator(cfqq);
3526 cfq_put_queue(cfqq);
3529 static void cfq_init_icq(struct io_cq *icq)
3531 struct cfq_io_cq *cic = icq_to_cic(icq);
3533 cic->ttime.last_end_request = jiffies;
3536 static void cfq_exit_icq(struct io_cq *icq)
3538 struct cfq_io_cq *cic = icq_to_cic(icq);
3539 struct cfq_data *cfqd = cic_to_cfqd(cic);
3541 if (cic_to_cfqq(cic, false)) {
3542 cfq_exit_cfqq(cfqd, cic_to_cfqq(cic, false));
3543 cic_set_cfqq(cic, NULL, false);
3546 if (cic_to_cfqq(cic, true)) {
3547 cfq_exit_cfqq(cfqd, cic_to_cfqq(cic, true));
3548 cic_set_cfqq(cic, NULL, true);
3552 static void cfq_init_prio_data(struct cfq_queue *cfqq, struct cfq_io_cq *cic)
3554 struct task_struct *tsk = current;
3557 if (!cfq_cfqq_prio_changed(cfqq))
3560 ioprio_class = IOPRIO_PRIO_CLASS(cic->ioprio);
3561 switch (ioprio_class) {
3563 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
3564 case IOPRIO_CLASS_NONE:
3566 * no prio set, inherit CPU scheduling settings
3568 cfqq->ioprio = task_nice_ioprio(tsk);
3569 cfqq->ioprio_class = task_nice_ioclass(tsk);
3571 case IOPRIO_CLASS_RT:
3572 cfqq->ioprio = IOPRIO_PRIO_DATA(cic->ioprio);
3573 cfqq->ioprio_class = IOPRIO_CLASS_RT;
3575 case IOPRIO_CLASS_BE:
3576 cfqq->ioprio = IOPRIO_PRIO_DATA(cic->ioprio);
3577 cfqq->ioprio_class = IOPRIO_CLASS_BE;
3579 case IOPRIO_CLASS_IDLE:
3580 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
3582 cfq_clear_cfqq_idle_window(cfqq);
3587 * keep track of original prio settings in case we have to temporarily
3588 * elevate the priority of this queue
3590 cfqq->org_ioprio = cfqq->ioprio;
3591 cfq_clear_cfqq_prio_changed(cfqq);
3594 static void check_ioprio_changed(struct cfq_io_cq *cic, struct bio *bio)
3596 int ioprio = cic->icq.ioc->ioprio;
3597 struct cfq_data *cfqd = cic_to_cfqd(cic);
3598 struct cfq_queue *cfqq;
3601 * Check whether ioprio has changed. The condition may trigger
3602 * spuriously on a newly created cic but there's no harm.
3604 if (unlikely(!cfqd) || likely(cic->ioprio == ioprio))
3607 cfqq = cic_to_cfqq(cic, false);
3609 cfq_put_queue(cfqq);
3610 cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic, bio);
3611 cic_set_cfqq(cic, cfqq, false);
3614 cfqq = cic_to_cfqq(cic, true);
3616 cfq_mark_cfqq_prio_changed(cfqq);
3618 cic->ioprio = ioprio;
3621 static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3622 pid_t pid, bool is_sync)
3624 RB_CLEAR_NODE(&cfqq->rb_node);
3625 RB_CLEAR_NODE(&cfqq->p_node);
3626 INIT_LIST_HEAD(&cfqq->fifo);
3631 cfq_mark_cfqq_prio_changed(cfqq);
3634 if (!cfq_class_idle(cfqq))
3635 cfq_mark_cfqq_idle_window(cfqq);
3636 cfq_mark_cfqq_sync(cfqq);
3641 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3642 static void check_blkcg_changed(struct cfq_io_cq *cic, struct bio *bio)
3644 struct cfq_data *cfqd = cic_to_cfqd(cic);
3645 struct cfq_queue *cfqq;
3649 serial_nr = bio_blkcg(bio)->css.serial_nr;
3653 * Check whether blkcg has changed. The condition may trigger
3654 * spuriously on a newly created cic but there's no harm.
3656 if (unlikely(!cfqd) || likely(cic->blkcg_serial_nr == serial_nr))
3660 * Drop reference to queues. New queues will be assigned in new
3661 * group upon arrival of fresh requests.
3663 cfqq = cic_to_cfqq(cic, false);
3665 cfq_log_cfqq(cfqd, cfqq, "changed cgroup");
3666 cic_set_cfqq(cic, NULL, false);
3667 cfq_put_queue(cfqq);
3670 cfqq = cic_to_cfqq(cic, true);
3672 cfq_log_cfqq(cfqd, cfqq, "changed cgroup");
3673 cic_set_cfqq(cic, NULL, true);
3674 cfq_put_queue(cfqq);
3677 cic->blkcg_serial_nr = serial_nr;
3680 static inline void check_blkcg_changed(struct cfq_io_cq *cic, struct bio *bio) { }
3681 #endif /* CONFIG_CFQ_GROUP_IOSCHED */
3683 static struct cfq_queue **
3684 cfq_async_queue_prio(struct cfq_group *cfqg, int ioprio_class, int ioprio)
3686 switch (ioprio_class) {
3687 case IOPRIO_CLASS_RT:
3688 return &cfqg->async_cfqq[0][ioprio];
3689 case IOPRIO_CLASS_NONE:
3690 ioprio = IOPRIO_NORM;
3692 case IOPRIO_CLASS_BE:
3693 return &cfqg->async_cfqq[1][ioprio];
3694 case IOPRIO_CLASS_IDLE:
3695 return &cfqg->async_idle_cfqq;
3701 static struct cfq_queue *
3702 cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct cfq_io_cq *cic,
3705 int ioprio_class = IOPRIO_PRIO_CLASS(cic->ioprio);
3706 int ioprio = IOPRIO_PRIO_DATA(cic->ioprio);
3707 struct cfq_queue **async_cfqq = NULL;
3708 struct cfq_queue *cfqq;
3709 struct cfq_group *cfqg;
3712 cfqg = cfq_lookup_cfqg(cfqd, bio_blkcg(bio));
3714 cfqq = &cfqd->oom_cfqq;
3719 if (!ioprio_valid(cic->ioprio)) {
3720 struct task_struct *tsk = current;
3721 ioprio = task_nice_ioprio(tsk);
3722 ioprio_class = task_nice_ioclass(tsk);
3724 async_cfqq = cfq_async_queue_prio(cfqg, ioprio_class, ioprio);
3730 cfqq = kmem_cache_alloc_node(cfq_pool, GFP_NOWAIT | __GFP_ZERO,
3733 cfqq = &cfqd->oom_cfqq;
3737 cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
3738 cfq_init_prio_data(cfqq, cic);
3739 cfq_link_cfqq_cfqg(cfqq, cfqg);
3740 cfq_log_cfqq(cfqd, cfqq, "alloced");
3743 /* a new async queue is created, pin and remember */
3754 __cfq_update_io_thinktime(struct cfq_ttime *ttime, unsigned long slice_idle)
3756 unsigned long elapsed = jiffies - ttime->last_end_request;
3757 elapsed = min(elapsed, 2UL * slice_idle);
3759 ttime->ttime_samples = (7*ttime->ttime_samples + 256) / 8;
3760 ttime->ttime_total = (7*ttime->ttime_total + 256*elapsed) / 8;
3761 ttime->ttime_mean = (ttime->ttime_total + 128) / ttime->ttime_samples;
3765 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3766 struct cfq_io_cq *cic)
3768 if (cfq_cfqq_sync(cfqq)) {
3769 __cfq_update_io_thinktime(&cic->ttime, cfqd->cfq_slice_idle);
3770 __cfq_update_io_thinktime(&cfqq->service_tree->ttime,
3771 cfqd->cfq_slice_idle);
3773 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3774 __cfq_update_io_thinktime(&cfqq->cfqg->ttime, cfqd->cfq_group_idle);
3779 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3783 sector_t n_sec = blk_rq_sectors(rq);
3784 if (cfqq->last_request_pos) {
3785 if (cfqq->last_request_pos < blk_rq_pos(rq))
3786 sdist = blk_rq_pos(rq) - cfqq->last_request_pos;
3788 sdist = cfqq->last_request_pos - blk_rq_pos(rq);
3791 cfqq->seek_history <<= 1;
3792 if (blk_queue_nonrot(cfqd->queue))
3793 cfqq->seek_history |= (n_sec < CFQQ_SECT_THR_NONROT);
3795 cfqq->seek_history |= (sdist > CFQQ_SEEK_THR);
3799 * Disable idle window if the process thinks too long or seeks so much that
3803 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3804 struct cfq_io_cq *cic)
3806 int old_idle, enable_idle;
3809 * Don't idle for async or idle io prio class
3811 if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
3814 enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
3816 if (cfqq->queued[0] + cfqq->queued[1] >= 4)
3817 cfq_mark_cfqq_deep(cfqq);
3819 if (cfqq->next_rq && (cfqq->next_rq->cmd_flags & REQ_NOIDLE))
3821 else if (!atomic_read(&cic->icq.ioc->active_ref) ||
3822 !cfqd->cfq_slice_idle ||
3823 (!cfq_cfqq_deep(cfqq) && CFQQ_SEEKY(cfqq)))
3825 else if (sample_valid(cic->ttime.ttime_samples)) {
3826 if (cic->ttime.ttime_mean > cfqd->cfq_slice_idle)
3832 if (old_idle != enable_idle) {
3833 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
3835 cfq_mark_cfqq_idle_window(cfqq);
3837 cfq_clear_cfqq_idle_window(cfqq);
3842 * Check if new_cfqq should preempt the currently active queue. Return 0 for
3843 * no or if we aren't sure, a 1 will cause a preempt.
3846 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
3849 struct cfq_queue *cfqq;
3851 cfqq = cfqd->active_queue;
3855 if (cfq_class_idle(new_cfqq))
3858 if (cfq_class_idle(cfqq))
3862 * Don't allow a non-RT request to preempt an ongoing RT cfqq timeslice.
3864 if (cfq_class_rt(cfqq) && !cfq_class_rt(new_cfqq))
3868 * if the new request is sync, but the currently running queue is
3869 * not, let the sync request have priority.
3871 if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
3874 if (new_cfqq->cfqg != cfqq->cfqg)
3877 if (cfq_slice_used(cfqq))
3880 /* Allow preemption only if we are idling on sync-noidle tree */
3881 if (cfqd->serving_wl_type == SYNC_NOIDLE_WORKLOAD &&
3882 cfqq_type(new_cfqq) == SYNC_NOIDLE_WORKLOAD &&
3883 new_cfqq->service_tree->count == 2 &&
3884 RB_EMPTY_ROOT(&cfqq->sort_list))
3888 * So both queues are sync. Let the new request get disk time if
3889 * it's a metadata request and the current queue is doing regular IO.
3891 if ((rq->cmd_flags & REQ_PRIO) && !cfqq->prio_pending)
3895 * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
3897 if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
3900 /* An idle queue should not be idle now for some reason */
3901 if (RB_EMPTY_ROOT(&cfqq->sort_list) && !cfq_should_idle(cfqd, cfqq))
3904 if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
3908 * if this request is as-good as one we would expect from the
3909 * current cfqq, let it preempt
3911 if (cfq_rq_close(cfqd, cfqq, rq))
3918 * cfqq preempts the active queue. if we allowed preempt with no slice left,
3919 * let it have half of its nominal slice.
3921 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3923 enum wl_type_t old_type = cfqq_type(cfqd->active_queue);
3925 cfq_log_cfqq(cfqd, cfqq, "preempt");
3926 cfq_slice_expired(cfqd, 1);
3929 * workload type is changed, don't save slice, otherwise preempt
3932 if (old_type != cfqq_type(cfqq))
3933 cfqq->cfqg->saved_wl_slice = 0;
3936 * Put the new queue at the front of the of the current list,
3937 * so we know that it will be selected next.
3939 BUG_ON(!cfq_cfqq_on_rr(cfqq));
3941 cfq_service_tree_add(cfqd, cfqq, 1);
3943 cfqq->slice_end = 0;
3944 cfq_mark_cfqq_slice_new(cfqq);
3948 * Called when a new fs request (rq) is added (to cfqq). Check if there's
3949 * something we should do about it
3952 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3955 struct cfq_io_cq *cic = RQ_CIC(rq);
3958 if (rq->cmd_flags & REQ_PRIO)
3959 cfqq->prio_pending++;
3961 cfq_update_io_thinktime(cfqd, cfqq, cic);
3962 cfq_update_io_seektime(cfqd, cfqq, rq);
3963 cfq_update_idle_window(cfqd, cfqq, cic);
3965 cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
3967 if (cfqq == cfqd->active_queue) {
3969 * Remember that we saw a request from this process, but
3970 * don't start queuing just yet. Otherwise we risk seeing lots
3971 * of tiny requests, because we disrupt the normal plugging
3972 * and merging. If the request is already larger than a single
3973 * page, let it rip immediately. For that case we assume that
3974 * merging is already done. Ditto for a busy system that
3975 * has other work pending, don't risk delaying until the
3976 * idle timer unplug to continue working.
3978 if (cfq_cfqq_wait_request(cfqq)) {
3979 if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
3980 cfqd->busy_queues > 1) {
3981 cfq_del_timer(cfqd, cfqq);
3982 cfq_clear_cfqq_wait_request(cfqq);
3983 __blk_run_queue(cfqd->queue);
3985 cfqg_stats_update_idle_time(cfqq->cfqg);
3986 cfq_mark_cfqq_must_dispatch(cfqq);
3989 } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
3991 * not the active queue - expire current slice if it is
3992 * idle and has expired it's mean thinktime or this new queue
3993 * has some old slice time left and is of higher priority or
3994 * this new queue is RT and the current one is BE
3996 cfq_preempt_queue(cfqd, cfqq);
3997 __blk_run_queue(cfqd->queue);
4001 static void cfq_insert_request(struct request_queue *q, struct request *rq)
4003 struct cfq_data *cfqd = q->elevator->elevator_data;
4004 struct cfq_queue *cfqq = RQ_CFQQ(rq);
4006 cfq_log_cfqq(cfqd, cfqq, "insert_request");
4007 cfq_init_prio_data(cfqq, RQ_CIC(rq));
4009 rq->fifo_time = jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)];
4010 list_add_tail(&rq->queuelist, &cfqq->fifo);
4012 cfqg_stats_update_io_add(RQ_CFQG(rq), cfqd->serving_group,
4014 cfq_rq_enqueued(cfqd, cfqq, rq);
4018 * Update hw_tag based on peak queue depth over 50 samples under
4021 static void cfq_update_hw_tag(struct cfq_data *cfqd)
4023 struct cfq_queue *cfqq = cfqd->active_queue;
4025 if (cfqd->rq_in_driver > cfqd->hw_tag_est_depth)
4026 cfqd->hw_tag_est_depth = cfqd->rq_in_driver;
4028 if (cfqd->hw_tag == 1)
4031 if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
4032 cfqd->rq_in_driver <= CFQ_HW_QUEUE_MIN)
4036 * If active queue hasn't enough requests and can idle, cfq might not
4037 * dispatch sufficient requests to hardware. Don't zero hw_tag in this
4040 if (cfqq && cfq_cfqq_idle_window(cfqq) &&
4041 cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] <
4042 CFQ_HW_QUEUE_MIN && cfqd->rq_in_driver < CFQ_HW_QUEUE_MIN)
4045 if (cfqd->hw_tag_samples++ < 50)
4048 if (cfqd->hw_tag_est_depth >= CFQ_HW_QUEUE_MIN)
4054 static bool cfq_should_wait_busy(struct cfq_data *cfqd, struct cfq_queue *cfqq)
4056 struct cfq_io_cq *cic = cfqd->active_cic;
4058 /* If the queue already has requests, don't wait */
4059 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
4062 /* If there are other queues in the group, don't wait */
4063 if (cfqq->cfqg->nr_cfqq > 1)
4066 /* the only queue in the group, but think time is big */
4067 if (cfq_io_thinktime_big(cfqd, &cfqq->cfqg->ttime, true))
4070 if (cfq_slice_used(cfqq))
4073 /* if slice left is less than think time, wait busy */
4074 if (cic && sample_valid(cic->ttime.ttime_samples)
4075 && (cfqq->slice_end - jiffies < cic->ttime.ttime_mean))
4079 * If think times is less than a jiffy than ttime_mean=0 and above
4080 * will not be true. It might happen that slice has not expired yet
4081 * but will expire soon (4-5 ns) during select_queue(). To cover the
4082 * case where think time is less than a jiffy, mark the queue wait
4083 * busy if only 1 jiffy is left in the slice.
4085 if (cfqq->slice_end - jiffies == 1)
4091 static void cfq_completed_request(struct request_queue *q, struct request *rq)
4093 struct cfq_queue *cfqq = RQ_CFQQ(rq);
4094 struct cfq_data *cfqd = cfqq->cfqd;
4095 const int sync = rq_is_sync(rq);
4099 cfq_log_cfqq(cfqd, cfqq, "complete rqnoidle %d",
4100 !!(rq->cmd_flags & REQ_NOIDLE));
4102 cfq_update_hw_tag(cfqd);
4104 WARN_ON(!cfqd->rq_in_driver);
4105 WARN_ON(!cfqq->dispatched);
4106 cfqd->rq_in_driver--;
4108 (RQ_CFQG(rq))->dispatched--;
4109 cfqg_stats_update_completion(cfqq->cfqg, rq_start_time_ns(rq),
4110 rq_io_start_time_ns(rq), rq->cmd_flags);
4112 cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]--;
4115 struct cfq_rb_root *st;
4117 RQ_CIC(rq)->ttime.last_end_request = now;
4119 if (cfq_cfqq_on_rr(cfqq))
4120 st = cfqq->service_tree;
4122 st = st_for(cfqq->cfqg, cfqq_class(cfqq),
4125 st->ttime.last_end_request = now;
4126 if (!time_after(rq->start_time + cfqd->cfq_fifo_expire[1], now))
4127 cfqd->last_delayed_sync = now;
4130 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4131 cfqq->cfqg->ttime.last_end_request = now;
4135 * If this is the active queue, check if it needs to be expired,
4136 * or if we want to idle in case it has no pending requests.
4138 if (cfqd->active_queue == cfqq) {
4139 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
4141 if (cfq_cfqq_slice_new(cfqq)) {
4142 cfq_set_prio_slice(cfqd, cfqq);
4143 cfq_clear_cfqq_slice_new(cfqq);
4147 * Should we wait for next request to come in before we expire
4150 if (cfq_should_wait_busy(cfqd, cfqq)) {
4151 unsigned long extend_sl = cfqd->cfq_slice_idle;
4152 if (!cfqd->cfq_slice_idle)
4153 extend_sl = cfqd->cfq_group_idle;
4154 cfqq->slice_end = jiffies + extend_sl;
4155 cfq_mark_cfqq_wait_busy(cfqq);
4156 cfq_log_cfqq(cfqd, cfqq, "will busy wait");
4160 * Idling is not enabled on:
4162 * - idle-priority queues
4164 * - queues with still some requests queued
4165 * - when there is a close cooperator
4167 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
4168 cfq_slice_expired(cfqd, 1);
4169 else if (sync && cfqq_empty &&
4170 !cfq_close_cooperator(cfqd, cfqq)) {
4171 cfq_arm_slice_timer(cfqd);
4175 if (!cfqd->rq_in_driver)
4176 cfq_schedule_dispatch(cfqd);
4179 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
4181 if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
4182 cfq_mark_cfqq_must_alloc_slice(cfqq);
4183 return ELV_MQUEUE_MUST;
4186 return ELV_MQUEUE_MAY;
4189 static int cfq_may_queue(struct request_queue *q, int rw)
4191 struct cfq_data *cfqd = q->elevator->elevator_data;
4192 struct task_struct *tsk = current;
4193 struct cfq_io_cq *cic;
4194 struct cfq_queue *cfqq;
4197 * don't force setup of a queue from here, as a call to may_queue
4198 * does not necessarily imply that a request actually will be queued.
4199 * so just lookup a possibly existing queue, or return 'may queue'
4202 cic = cfq_cic_lookup(cfqd, tsk->io_context);
4204 return ELV_MQUEUE_MAY;
4206 cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
4208 cfq_init_prio_data(cfqq, cic);
4210 return __cfq_may_queue(cfqq);
4213 return ELV_MQUEUE_MAY;
4217 * queue lock held here
4219 static void cfq_put_request(struct request *rq)
4221 struct cfq_queue *cfqq = RQ_CFQQ(rq);
4224 const int rw = rq_data_dir(rq);
4226 BUG_ON(!cfqq->allocated[rw]);
4227 cfqq->allocated[rw]--;
4229 /* Put down rq reference on cfqg */
4230 cfqg_put(RQ_CFQG(rq));
4231 rq->elv.priv[0] = NULL;
4232 rq->elv.priv[1] = NULL;
4234 cfq_put_queue(cfqq);
4238 static struct cfq_queue *
4239 cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_cq *cic,
4240 struct cfq_queue *cfqq)
4242 cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq);
4243 cic_set_cfqq(cic, cfqq->new_cfqq, 1);
4244 cfq_mark_cfqq_coop(cfqq->new_cfqq);
4245 cfq_put_queue(cfqq);
4246 return cic_to_cfqq(cic, 1);
4250 * Returns NULL if a new cfqq should be allocated, or the old cfqq if this
4251 * was the last process referring to said cfqq.
4253 static struct cfq_queue *
4254 split_cfqq(struct cfq_io_cq *cic, struct cfq_queue *cfqq)
4256 if (cfqq_process_refs(cfqq) == 1) {
4257 cfqq->pid = current->pid;
4258 cfq_clear_cfqq_coop(cfqq);
4259 cfq_clear_cfqq_split_coop(cfqq);
4263 cic_set_cfqq(cic, NULL, 1);
4265 cfq_put_cooperator(cfqq);
4267 cfq_put_queue(cfqq);
4271 * Allocate cfq data structures associated with this request.
4274 cfq_set_request(struct request_queue *q, struct request *rq, struct bio *bio,
4277 struct cfq_data *cfqd = q->elevator->elevator_data;
4278 struct cfq_io_cq *cic = icq_to_cic(rq->elv.icq);
4279 const int rw = rq_data_dir(rq);
4280 const bool is_sync = rq_is_sync(rq);
4281 struct cfq_queue *cfqq;
4283 spin_lock_irq(q->queue_lock);
4285 check_ioprio_changed(cic, bio);
4286 check_blkcg_changed(cic, bio);
4288 cfqq = cic_to_cfqq(cic, is_sync);
4289 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
4291 cfq_put_queue(cfqq);
4292 cfqq = cfq_get_queue(cfqd, is_sync, cic, bio);
4293 cic_set_cfqq(cic, cfqq, is_sync);
4296 * If the queue was seeky for too long, break it apart.
4298 if (cfq_cfqq_coop(cfqq) && cfq_cfqq_split_coop(cfqq)) {
4299 cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq");
4300 cfqq = split_cfqq(cic, cfqq);
4306 * Check to see if this queue is scheduled to merge with
4307 * another, closely cooperating queue. The merging of
4308 * queues happens here as it must be done in process context.
4309 * The reference on new_cfqq was taken in merge_cfqqs.
4312 cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq);
4315 cfqq->allocated[rw]++;
4318 cfqg_get(cfqq->cfqg);
4319 rq->elv.priv[0] = cfqq;
4320 rq->elv.priv[1] = cfqq->cfqg;
4321 spin_unlock_irq(q->queue_lock);
4325 static void cfq_kick_queue(struct work_struct *work)
4327 struct cfq_data *cfqd =
4328 container_of(work, struct cfq_data, unplug_work);
4329 struct request_queue *q = cfqd->queue;
4331 spin_lock_irq(q->queue_lock);
4332 __blk_run_queue(cfqd->queue);
4333 spin_unlock_irq(q->queue_lock);
4337 * Timer running if the active_queue is currently idling inside its time slice
4339 static void cfq_idle_slice_timer(unsigned long data)
4341 struct cfq_data *cfqd = (struct cfq_data *) data;
4342 struct cfq_queue *cfqq;
4343 unsigned long flags;
4346 cfq_log(cfqd, "idle timer fired");
4348 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
4350 cfqq = cfqd->active_queue;
4355 * We saw a request before the queue expired, let it through
4357 if (cfq_cfqq_must_dispatch(cfqq))
4363 if (cfq_slice_used(cfqq))
4367 * only expire and reinvoke request handler, if there are
4368 * other queues with pending requests
4370 if (!cfqd->busy_queues)
4374 * not expired and it has a request pending, let it dispatch
4376 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
4380 * Queue depth flag is reset only when the idle didn't succeed
4382 cfq_clear_cfqq_deep(cfqq);
4385 cfq_slice_expired(cfqd, timed_out);
4387 cfq_schedule_dispatch(cfqd);
4389 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
4392 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
4394 del_timer_sync(&cfqd->idle_slice_timer);
4395 cancel_work_sync(&cfqd->unplug_work);
4398 static void cfq_exit_queue(struct elevator_queue *e)
4400 struct cfq_data *cfqd = e->elevator_data;
4401 struct request_queue *q = cfqd->queue;
4403 cfq_shutdown_timer_wq(cfqd);
4405 spin_lock_irq(q->queue_lock);
4407 if (cfqd->active_queue)
4408 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
4410 spin_unlock_irq(q->queue_lock);
4412 cfq_shutdown_timer_wq(cfqd);
4414 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4415 blkcg_deactivate_policy(q, &blkcg_policy_cfq);
4417 kfree(cfqd->root_group);
4422 static int cfq_init_queue(struct request_queue *q, struct elevator_type *e)
4424 struct cfq_data *cfqd;
4425 struct blkcg_gq *blkg __maybe_unused;
4427 struct elevator_queue *eq;
4429 eq = elevator_alloc(q, e);
4433 cfqd = kzalloc_node(sizeof(*cfqd), GFP_KERNEL, q->node);
4435 kobject_put(&eq->kobj);
4438 eq->elevator_data = cfqd;
4441 spin_lock_irq(q->queue_lock);
4443 spin_unlock_irq(q->queue_lock);
4445 /* Init root service tree */
4446 cfqd->grp_service_tree = CFQ_RB_ROOT;
4448 /* Init root group and prefer root group over other groups by default */
4449 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4450 ret = blkcg_activate_policy(q, &blkcg_policy_cfq);
4454 cfqd->root_group = blkg_to_cfqg(q->root_blkg);
4457 cfqd->root_group = kzalloc_node(sizeof(*cfqd->root_group),
4458 GFP_KERNEL, cfqd->queue->node);
4459 if (!cfqd->root_group)
4462 cfq_init_cfqg_base(cfqd->root_group);
4464 cfqd->root_group->weight = 2 * CFQ_WEIGHT_DEFAULT;
4465 cfqd->root_group->leaf_weight = 2 * CFQ_WEIGHT_DEFAULT;
4468 * Not strictly needed (since RB_ROOT just clears the node and we
4469 * zeroed cfqd on alloc), but better be safe in case someone decides
4470 * to add magic to the rb code
4472 for (i = 0; i < CFQ_PRIO_LISTS; i++)
4473 cfqd->prio_trees[i] = RB_ROOT;
4476 * Our fallback cfqq if cfq_get_queue() runs into OOM issues.
4477 * Grab a permanent reference to it, so that the normal code flow
4478 * will not attempt to free it. oom_cfqq is linked to root_group
4479 * but shouldn't hold a reference as it'll never be unlinked. Lose
4480 * the reference from linking right away.
4482 cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
4483 cfqd->oom_cfqq.ref++;
4485 spin_lock_irq(q->queue_lock);
4486 cfq_link_cfqq_cfqg(&cfqd->oom_cfqq, cfqd->root_group);
4487 cfqg_put(cfqd->root_group);
4488 spin_unlock_irq(q->queue_lock);
4490 init_timer(&cfqd->idle_slice_timer);
4491 cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
4492 cfqd->idle_slice_timer.data = (unsigned long) cfqd;
4494 INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
4496 cfqd->cfq_quantum = cfq_quantum;
4497 cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
4498 cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
4499 cfqd->cfq_back_max = cfq_back_max;
4500 cfqd->cfq_back_penalty = cfq_back_penalty;
4501 cfqd->cfq_slice[0] = cfq_slice_async;
4502 cfqd->cfq_slice[1] = cfq_slice_sync;
4503 cfqd->cfq_target_latency = cfq_target_latency;
4504 cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
4505 cfqd->cfq_slice_idle = cfq_slice_idle;
4506 cfqd->cfq_group_idle = cfq_group_idle;
4507 cfqd->cfq_latency = 1;
4510 * we optimistically start assuming sync ops weren't delayed in last
4511 * second, in order to have larger depth for async operations.
4513 cfqd->last_delayed_sync = jiffies - HZ;
4518 kobject_put(&eq->kobj);
4522 static void cfq_registered_queue(struct request_queue *q)
4524 struct elevator_queue *e = q->elevator;
4525 struct cfq_data *cfqd = e->elevator_data;
4528 * Default to IOPS mode with no idling for SSDs
4530 if (blk_queue_nonrot(q))
4531 cfqd->cfq_slice_idle = 0;
4535 * sysfs parts below -->
4538 cfq_var_show(unsigned int var, char *page)
4540 return sprintf(page, "%u\n", var);
4544 cfq_var_store(unsigned int *var, const char *page, size_t count)
4546 char *p = (char *) page;
4548 *var = simple_strtoul(p, &p, 10);
4552 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
4553 static ssize_t __FUNC(struct elevator_queue *e, char *page) \
4555 struct cfq_data *cfqd = e->elevator_data; \
4556 unsigned int __data = __VAR; \
4558 __data = jiffies_to_msecs(__data); \
4559 return cfq_var_show(__data, (page)); \
4561 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
4562 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
4563 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
4564 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
4565 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
4566 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
4567 SHOW_FUNCTION(cfq_group_idle_show, cfqd->cfq_group_idle, 1);
4568 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
4569 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
4570 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
4571 SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
4572 SHOW_FUNCTION(cfq_target_latency_show, cfqd->cfq_target_latency, 1);
4573 #undef SHOW_FUNCTION
4575 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
4576 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
4578 struct cfq_data *cfqd = e->elevator_data; \
4579 unsigned int __data; \
4580 int ret = cfq_var_store(&__data, (page), count); \
4581 if (__data < (MIN)) \
4583 else if (__data > (MAX)) \
4586 *(__PTR) = msecs_to_jiffies(__data); \
4588 *(__PTR) = __data; \
4591 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
4592 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
4594 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
4596 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
4597 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
4599 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
4600 STORE_FUNCTION(cfq_group_idle_store, &cfqd->cfq_group_idle, 0, UINT_MAX, 1);
4601 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
4602 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
4603 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
4605 STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
4606 STORE_FUNCTION(cfq_target_latency_store, &cfqd->cfq_target_latency, 1, UINT_MAX, 1);
4607 #undef STORE_FUNCTION
4609 #define CFQ_ATTR(name) \
4610 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
4612 static struct elv_fs_entry cfq_attrs[] = {
4614 CFQ_ATTR(fifo_expire_sync),
4615 CFQ_ATTR(fifo_expire_async),
4616 CFQ_ATTR(back_seek_max),
4617 CFQ_ATTR(back_seek_penalty),
4618 CFQ_ATTR(slice_sync),
4619 CFQ_ATTR(slice_async),
4620 CFQ_ATTR(slice_async_rq),
4621 CFQ_ATTR(slice_idle),
4622 CFQ_ATTR(group_idle),
4623 CFQ_ATTR(low_latency),
4624 CFQ_ATTR(target_latency),
4628 static struct elevator_type iosched_cfq = {
4630 .elevator_merge_fn = cfq_merge,
4631 .elevator_merged_fn = cfq_merged_request,
4632 .elevator_merge_req_fn = cfq_merged_requests,
4633 .elevator_allow_merge_fn = cfq_allow_merge,
4634 .elevator_bio_merged_fn = cfq_bio_merged,
4635 .elevator_dispatch_fn = cfq_dispatch_requests,
4636 .elevator_add_req_fn = cfq_insert_request,
4637 .elevator_activate_req_fn = cfq_activate_request,
4638 .elevator_deactivate_req_fn = cfq_deactivate_request,
4639 .elevator_completed_req_fn = cfq_completed_request,
4640 .elevator_former_req_fn = elv_rb_former_request,
4641 .elevator_latter_req_fn = elv_rb_latter_request,
4642 .elevator_init_icq_fn = cfq_init_icq,
4643 .elevator_exit_icq_fn = cfq_exit_icq,
4644 .elevator_set_req_fn = cfq_set_request,
4645 .elevator_put_req_fn = cfq_put_request,
4646 .elevator_may_queue_fn = cfq_may_queue,
4647 .elevator_init_fn = cfq_init_queue,
4648 .elevator_exit_fn = cfq_exit_queue,
4649 .elevator_registered_fn = cfq_registered_queue,
4651 .icq_size = sizeof(struct cfq_io_cq),
4652 .icq_align = __alignof__(struct cfq_io_cq),
4653 .elevator_attrs = cfq_attrs,
4654 .elevator_name = "cfq",
4655 .elevator_owner = THIS_MODULE,
4658 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4659 static struct blkcg_policy blkcg_policy_cfq = {
4660 .cftypes = cfq_blkcg_files,
4662 .cpd_alloc_fn = cfq_cpd_alloc,
4663 .cpd_init_fn = cfq_cpd_init,
4664 .cpd_free_fn = cfq_cpd_free,
4666 .pd_alloc_fn = cfq_pd_alloc,
4667 .pd_init_fn = cfq_pd_init,
4668 .pd_offline_fn = cfq_pd_offline,
4669 .pd_free_fn = cfq_pd_free,
4670 .pd_reset_stats_fn = cfq_pd_reset_stats,
4674 static int __init cfq_init(void)
4679 * could be 0 on HZ < 1000 setups
4681 if (!cfq_slice_async)
4682 cfq_slice_async = 1;
4683 if (!cfq_slice_idle)
4686 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4687 if (!cfq_group_idle)
4690 ret = blkcg_policy_register(&blkcg_policy_cfq);
4698 cfq_pool = KMEM_CACHE(cfq_queue, 0);
4702 ret = elv_register(&iosched_cfq);
4709 kmem_cache_destroy(cfq_pool);
4711 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4712 blkcg_policy_unregister(&blkcg_policy_cfq);
4717 static void __exit cfq_exit(void)
4719 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4720 blkcg_policy_unregister(&blkcg_policy_cfq);
4722 elv_unregister(&iosched_cfq);
4723 kmem_cache_destroy(cfq_pool);
4726 module_init(cfq_init);
4727 module_exit(cfq_exit);
4729 MODULE_AUTHOR("Jens Axboe");
4730 MODULE_LICENSE("GPL");
4731 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");