2 * CFQ, or complete fairness queueing, disk scheduler.
4 * Based on ideas from a previously unfinished io
5 * scheduler (round robin per-process disk scheduling) and Andrea Arcangeli.
7 * Copyright (C) 2003 Jens Axboe <axboe@kernel.dk>
9 #include <linux/module.h>
10 #include <linux/slab.h>
11 #include <linux/blkdev.h>
12 #include <linux/elevator.h>
13 #include <linux/ktime.h>
14 #include <linux/rbtree.h>
15 #include <linux/ioprio.h>
16 #include <linux/blktrace_api.h>
17 #include <linux/blk-cgroup.h>
23 /* max queue in one round of service */
24 static const int cfq_quantum = 8;
25 static const u64 cfq_fifo_expire[2] = { NSEC_PER_SEC / 4, NSEC_PER_SEC / 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 u64 cfq_slice_sync = NSEC_PER_SEC / 10;
31 static u64 cfq_slice_async = NSEC_PER_SEC / 25;
32 static const int cfq_slice_async_rq = 2;
33 static u64 cfq_slice_idle = NSEC_PER_SEC / 125;
34 static u64 cfq_group_idle = NSEC_PER_SEC / 125;
35 static const u64 cfq_target_latency = (u64)NSEC_PER_SEC * 3/10; /* 300 ms */
36 static const int cfq_hist_divisor = 4;
39 * offset from end of service tree
41 #define CFQ_IDLE_DELAY (NSEC_PER_SEC / 5)
44 * below this threshold, we consider thinktime immediate
46 #define CFQ_MIN_TT (2 * NSEC_PER_SEC / HZ)
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_LEGACY_MIN 10
72 #define CFQ_WEIGHT_LEGACY_DFL 500
73 #define CFQ_WEIGHT_LEGACY_MAX 1000
80 unsigned long ttime_samples;
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 = ktime_get_ns(),},}
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 */
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. */
132 /* time when first request from queue completed and slice started. */
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, org_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;
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 u64 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 hrtimer 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_back_penalty;
378 unsigned int cfq_back_max;
379 unsigned int cfq_slice_async_rq;
380 unsigned int cfq_latency;
381 u64 cfq_fifo_expire[2];
385 u64 cfq_target_latency;
388 * Fallback dummy cfqq for extreme OOM conditions
390 struct cfq_queue oom_cfqq;
392 u64 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 bool cfqg_is_descendant(struct cfq_group *cfqg,
636 struct cfq_group *ancestor)
638 return cgroup_is_descendant(cfqg_to_blkg(cfqg)->blkcg->css.cgroup,
639 cfqg_to_blkg(ancestor)->blkcg->css.cgroup);
642 static inline void cfqg_get(struct cfq_group *cfqg)
644 return blkg_get(cfqg_to_blkg(cfqg));
647 static inline void cfqg_put(struct cfq_group *cfqg)
649 return blkg_put(cfqg_to_blkg(cfqg));
652 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...) do { \
655 blkg_path(cfqg_to_blkg((cfqq)->cfqg), __pbuf, sizeof(__pbuf)); \
656 blk_add_trace_msg((cfqd)->queue, "cfq%d%c%c %s " fmt, (cfqq)->pid, \
657 cfq_cfqq_sync((cfqq)) ? 'S' : 'A', \
658 cfqq_type((cfqq)) == SYNC_NOIDLE_WORKLOAD ? 'N' : ' ',\
662 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...) do { \
665 blkg_path(cfqg_to_blkg(cfqg), __pbuf, sizeof(__pbuf)); \
666 blk_add_trace_msg((cfqd)->queue, "%s " fmt, __pbuf, ##args); \
669 static inline void cfqg_stats_update_io_add(struct cfq_group *cfqg,
670 struct cfq_group *curr_cfqg, int op,
673 blkg_rwstat_add(&cfqg->stats.queued, op, op_flags, 1);
674 cfqg_stats_end_empty_time(&cfqg->stats);
675 cfqg_stats_set_start_group_wait_time(cfqg, curr_cfqg);
678 static inline void cfqg_stats_update_timeslice_used(struct cfq_group *cfqg,
679 uint64_t time, unsigned long unaccounted_time)
681 blkg_stat_add(&cfqg->stats.time, time);
682 #ifdef CONFIG_DEBUG_BLK_CGROUP
683 blkg_stat_add(&cfqg->stats.unaccounted_time, unaccounted_time);
687 static inline void cfqg_stats_update_io_remove(struct cfq_group *cfqg, int op,
690 blkg_rwstat_add(&cfqg->stats.queued, op, op_flags, -1);
693 static inline void cfqg_stats_update_io_merged(struct cfq_group *cfqg, int op,
696 blkg_rwstat_add(&cfqg->stats.merged, op, op_flags, 1);
699 static inline void cfqg_stats_update_completion(struct cfq_group *cfqg,
700 uint64_t start_time, uint64_t io_start_time, int op,
703 struct cfqg_stats *stats = &cfqg->stats;
704 unsigned long long now = sched_clock();
706 if (time_after64(now, io_start_time))
707 blkg_rwstat_add(&stats->service_time, op, op_flags,
708 now - io_start_time);
709 if (time_after64(io_start_time, start_time))
710 blkg_rwstat_add(&stats->wait_time, op, op_flags,
711 io_start_time - start_time);
715 static void cfqg_stats_reset(struct cfqg_stats *stats)
717 /* queued stats shouldn't be cleared */
718 blkg_rwstat_reset(&stats->merged);
719 blkg_rwstat_reset(&stats->service_time);
720 blkg_rwstat_reset(&stats->wait_time);
721 blkg_stat_reset(&stats->time);
722 #ifdef CONFIG_DEBUG_BLK_CGROUP
723 blkg_stat_reset(&stats->unaccounted_time);
724 blkg_stat_reset(&stats->avg_queue_size_sum);
725 blkg_stat_reset(&stats->avg_queue_size_samples);
726 blkg_stat_reset(&stats->dequeue);
727 blkg_stat_reset(&stats->group_wait_time);
728 blkg_stat_reset(&stats->idle_time);
729 blkg_stat_reset(&stats->empty_time);
734 static void cfqg_stats_add_aux(struct cfqg_stats *to, struct cfqg_stats *from)
736 /* queued stats shouldn't be cleared */
737 blkg_rwstat_add_aux(&to->merged, &from->merged);
738 blkg_rwstat_add_aux(&to->service_time, &from->service_time);
739 blkg_rwstat_add_aux(&to->wait_time, &from->wait_time);
740 blkg_stat_add_aux(&from->time, &from->time);
741 #ifdef CONFIG_DEBUG_BLK_CGROUP
742 blkg_stat_add_aux(&to->unaccounted_time, &from->unaccounted_time);
743 blkg_stat_add_aux(&to->avg_queue_size_sum, &from->avg_queue_size_sum);
744 blkg_stat_add_aux(&to->avg_queue_size_samples, &from->avg_queue_size_samples);
745 blkg_stat_add_aux(&to->dequeue, &from->dequeue);
746 blkg_stat_add_aux(&to->group_wait_time, &from->group_wait_time);
747 blkg_stat_add_aux(&to->idle_time, &from->idle_time);
748 blkg_stat_add_aux(&to->empty_time, &from->empty_time);
753 * Transfer @cfqg's stats to its parent's aux counts so that the ancestors'
754 * recursive stats can still account for the amount used by this cfqg after
757 static void cfqg_stats_xfer_dead(struct cfq_group *cfqg)
759 struct cfq_group *parent = cfqg_parent(cfqg);
761 lockdep_assert_held(cfqg_to_blkg(cfqg)->q->queue_lock);
763 if (unlikely(!parent))
766 cfqg_stats_add_aux(&parent->stats, &cfqg->stats);
767 cfqg_stats_reset(&cfqg->stats);
770 #else /* CONFIG_CFQ_GROUP_IOSCHED */
772 static inline struct cfq_group *cfqg_parent(struct cfq_group *cfqg) { return NULL; }
773 static inline bool cfqg_is_descendant(struct cfq_group *cfqg,
774 struct cfq_group *ancestor)
778 static inline void cfqg_get(struct cfq_group *cfqg) { }
779 static inline void cfqg_put(struct cfq_group *cfqg) { }
781 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
782 blk_add_trace_msg((cfqd)->queue, "cfq%d%c%c " fmt, (cfqq)->pid, \
783 cfq_cfqq_sync((cfqq)) ? 'S' : 'A', \
784 cfqq_type((cfqq)) == SYNC_NOIDLE_WORKLOAD ? 'N' : ' ',\
786 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...) do {} while (0)
788 static inline void cfqg_stats_update_io_add(struct cfq_group *cfqg,
789 struct cfq_group *curr_cfqg, int op, int op_flags) { }
790 static inline void cfqg_stats_update_timeslice_used(struct cfq_group *cfqg,
791 uint64_t time, unsigned long unaccounted_time) { }
792 static inline void cfqg_stats_update_io_remove(struct cfq_group *cfqg, int op,
794 static inline void cfqg_stats_update_io_merged(struct cfq_group *cfqg, int op,
796 static inline void cfqg_stats_update_completion(struct cfq_group *cfqg,
797 uint64_t start_time, uint64_t io_start_time, int op,
800 #endif /* CONFIG_CFQ_GROUP_IOSCHED */
802 #define cfq_log(cfqd, fmt, args...) \
803 blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
805 /* Traverses through cfq group service trees */
806 #define for_each_cfqg_st(cfqg, i, j, st) \
807 for (i = 0; i <= IDLE_WORKLOAD; i++) \
808 for (j = 0, st = i < IDLE_WORKLOAD ? &cfqg->service_trees[i][j]\
809 : &cfqg->service_tree_idle; \
810 (i < IDLE_WORKLOAD && j <= SYNC_WORKLOAD) || \
811 (i == IDLE_WORKLOAD && j == 0); \
812 j++, st = i < IDLE_WORKLOAD ? \
813 &cfqg->service_trees[i][j]: NULL) \
815 static inline bool cfq_io_thinktime_big(struct cfq_data *cfqd,
816 struct cfq_ttime *ttime, bool group_idle)
819 if (!sample_valid(ttime->ttime_samples))
822 slice = cfqd->cfq_group_idle;
824 slice = cfqd->cfq_slice_idle;
825 return ttime->ttime_mean > slice;
828 static inline bool iops_mode(struct cfq_data *cfqd)
831 * If we are not idling on queues and it is a NCQ drive, parallel
832 * execution of requests is on and measuring time is not possible
833 * in most of the cases until and unless we drive shallower queue
834 * depths and that becomes a performance bottleneck. In such cases
835 * switch to start providing fairness in terms of number of IOs.
837 if (!cfqd->cfq_slice_idle && cfqd->hw_tag)
843 static inline enum wl_class_t cfqq_class(struct cfq_queue *cfqq)
845 if (cfq_class_idle(cfqq))
846 return IDLE_WORKLOAD;
847 if (cfq_class_rt(cfqq))
853 static enum wl_type_t cfqq_type(struct cfq_queue *cfqq)
855 if (!cfq_cfqq_sync(cfqq))
856 return ASYNC_WORKLOAD;
857 if (!cfq_cfqq_idle_window(cfqq))
858 return SYNC_NOIDLE_WORKLOAD;
859 return SYNC_WORKLOAD;
862 static inline int cfq_group_busy_queues_wl(enum wl_class_t wl_class,
863 struct cfq_data *cfqd,
864 struct cfq_group *cfqg)
866 if (wl_class == IDLE_WORKLOAD)
867 return cfqg->service_tree_idle.count;
869 return cfqg->service_trees[wl_class][ASYNC_WORKLOAD].count +
870 cfqg->service_trees[wl_class][SYNC_NOIDLE_WORKLOAD].count +
871 cfqg->service_trees[wl_class][SYNC_WORKLOAD].count;
874 static inline int cfqg_busy_async_queues(struct cfq_data *cfqd,
875 struct cfq_group *cfqg)
877 return cfqg->service_trees[RT_WORKLOAD][ASYNC_WORKLOAD].count +
878 cfqg->service_trees[BE_WORKLOAD][ASYNC_WORKLOAD].count;
881 static void cfq_dispatch_insert(struct request_queue *, struct request *);
882 static struct cfq_queue *cfq_get_queue(struct cfq_data *cfqd, bool is_sync,
883 struct cfq_io_cq *cic, struct bio *bio);
885 static inline struct cfq_io_cq *icq_to_cic(struct io_cq *icq)
887 /* cic->icq is the first member, %NULL will convert to %NULL */
888 return container_of(icq, struct cfq_io_cq, icq);
891 static inline struct cfq_io_cq *cfq_cic_lookup(struct cfq_data *cfqd,
892 struct io_context *ioc)
895 return icq_to_cic(ioc_lookup_icq(ioc, cfqd->queue));
899 static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_cq *cic, bool is_sync)
901 return cic->cfqq[is_sync];
904 static inline void cic_set_cfqq(struct cfq_io_cq *cic, struct cfq_queue *cfqq,
907 cic->cfqq[is_sync] = cfqq;
910 static inline struct cfq_data *cic_to_cfqd(struct cfq_io_cq *cic)
912 return cic->icq.q->elevator->elevator_data;
916 * We regard a request as SYNC, if it's either a read or has the SYNC bit
917 * set (in which case it could also be direct WRITE).
919 static inline bool cfq_bio_sync(struct bio *bio)
921 return bio_data_dir(bio) == READ || (bio->bi_rw & REQ_SYNC);
925 * scheduler run of queue, if there are requests pending and no one in the
926 * driver that will restart queueing
928 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
930 if (cfqd->busy_queues) {
931 cfq_log(cfqd, "schedule dispatch");
932 kblockd_schedule_work(&cfqd->unplug_work);
937 * Scale schedule slice based on io priority. Use the sync time slice only
938 * if a queue is marked sync and has sync io queued. A sync queue with async
939 * io only, should not get full sync slice length.
941 static inline u64 cfq_prio_slice(struct cfq_data *cfqd, bool sync,
944 u64 base_slice = cfqd->cfq_slice[sync];
945 u64 slice = div_u64(base_slice, CFQ_SLICE_SCALE);
947 WARN_ON(prio >= IOPRIO_BE_NR);
949 return base_slice + (slice * (4 - prio));
953 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
955 return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
959 * cfqg_scale_charge - scale disk time charge according to cfqg weight
960 * @charge: disk time being charged
961 * @vfraction: vfraction of the cfqg, fixed point w/ CFQ_SERVICE_SHIFT
963 * Scale @charge according to @vfraction, which is in range (0, 1]. The
964 * scaling is inversely proportional.
966 * scaled = charge / vfraction
968 * The result is also in fixed point w/ CFQ_SERVICE_SHIFT.
970 static inline u64 cfqg_scale_charge(u64 charge,
971 unsigned int vfraction)
973 u64 c = charge << CFQ_SERVICE_SHIFT; /* make it fixed point */
975 /* charge / vfraction */
976 c <<= CFQ_SERVICE_SHIFT;
977 return div_u64(c, vfraction);
980 static inline u64 max_vdisktime(u64 min_vdisktime, u64 vdisktime)
982 s64 delta = (s64)(vdisktime - min_vdisktime);
984 min_vdisktime = vdisktime;
986 return min_vdisktime;
989 static inline u64 min_vdisktime(u64 min_vdisktime, u64 vdisktime)
991 s64 delta = (s64)(vdisktime - min_vdisktime);
993 min_vdisktime = vdisktime;
995 return min_vdisktime;
998 static void update_min_vdisktime(struct cfq_rb_root *st)
1000 struct cfq_group *cfqg;
1003 cfqg = rb_entry_cfqg(st->left);
1004 st->min_vdisktime = max_vdisktime(st->min_vdisktime,
1010 * get averaged number of queues of RT/BE priority.
1011 * average is updated, with a formula that gives more weight to higher numbers,
1012 * to quickly follows sudden increases and decrease slowly
1015 static inline unsigned cfq_group_get_avg_queues(struct cfq_data *cfqd,
1016 struct cfq_group *cfqg, bool rt)
1018 unsigned min_q, max_q;
1019 unsigned mult = cfq_hist_divisor - 1;
1020 unsigned round = cfq_hist_divisor / 2;
1021 unsigned busy = cfq_group_busy_queues_wl(rt, cfqd, cfqg);
1023 min_q = min(cfqg->busy_queues_avg[rt], busy);
1024 max_q = max(cfqg->busy_queues_avg[rt], busy);
1025 cfqg->busy_queues_avg[rt] = (mult * max_q + min_q + round) /
1027 return cfqg->busy_queues_avg[rt];
1031 cfq_group_slice(struct cfq_data *cfqd, struct cfq_group *cfqg)
1033 return cfqd->cfq_target_latency * cfqg->vfraction >> CFQ_SERVICE_SHIFT;
1037 cfq_scaled_cfqq_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1039 u64 slice = cfq_prio_to_slice(cfqd, cfqq);
1040 if (cfqd->cfq_latency) {
1042 * interested queues (we consider only the ones with the same
1043 * priority class in the cfq group)
1045 unsigned iq = cfq_group_get_avg_queues(cfqd, cfqq->cfqg,
1046 cfq_class_rt(cfqq));
1047 u64 sync_slice = cfqd->cfq_slice[1];
1048 u64 expect_latency = sync_slice * iq;
1049 u64 group_slice = cfq_group_slice(cfqd, cfqq->cfqg);
1051 if (expect_latency > group_slice) {
1052 u64 base_low_slice = 2 * cfqd->cfq_slice_idle;
1055 /* scale low_slice according to IO priority
1056 * and sync vs async */
1057 low_slice = div64_u64(base_low_slice*slice, sync_slice);
1058 low_slice = min(slice, low_slice);
1059 /* the adapted slice value is scaled to fit all iqs
1060 * into the target latency */
1061 slice = div64_u64(slice*group_slice, expect_latency);
1062 slice = max(slice, low_slice);
1069 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1071 u64 slice = cfq_scaled_cfqq_slice(cfqd, cfqq);
1072 u64 now = ktime_get_ns();
1074 cfqq->slice_start = now;
1075 cfqq->slice_end = now + slice;
1076 cfqq->allocated_slice = slice;
1077 cfq_log_cfqq(cfqd, cfqq, "set_slice=%llu", cfqq->slice_end - now);
1081 * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
1082 * isn't valid until the first request from the dispatch is activated
1083 * and the slice time set.
1085 static inline bool cfq_slice_used(struct cfq_queue *cfqq)
1087 if (cfq_cfqq_slice_new(cfqq))
1089 if (ktime_get_ns() < cfqq->slice_end)
1096 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
1097 * We choose the request that is closest to the head right now. Distance
1098 * behind the head is penalized and only allowed to a certain extent.
1100 static struct request *
1101 cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2, sector_t last)
1103 sector_t s1, s2, d1 = 0, d2 = 0;
1104 unsigned long back_max;
1105 #define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */
1106 #define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */
1107 unsigned wrap = 0; /* bit mask: requests behind the disk head? */
1109 if (rq1 == NULL || rq1 == rq2)
1114 if (rq_is_sync(rq1) != rq_is_sync(rq2))
1115 return rq_is_sync(rq1) ? rq1 : rq2;
1117 if ((rq1->cmd_flags ^ rq2->cmd_flags) & REQ_PRIO)
1118 return rq1->cmd_flags & REQ_PRIO ? rq1 : rq2;
1120 s1 = blk_rq_pos(rq1);
1121 s2 = blk_rq_pos(rq2);
1124 * by definition, 1KiB is 2 sectors
1126 back_max = cfqd->cfq_back_max * 2;
1129 * Strict one way elevator _except_ in the case where we allow
1130 * short backward seeks which are biased as twice the cost of a
1131 * similar forward seek.
1135 else if (s1 + back_max >= last)
1136 d1 = (last - s1) * cfqd->cfq_back_penalty;
1138 wrap |= CFQ_RQ1_WRAP;
1142 else if (s2 + back_max >= last)
1143 d2 = (last - s2) * cfqd->cfq_back_penalty;
1145 wrap |= CFQ_RQ2_WRAP;
1147 /* Found required data */
1150 * By doing switch() on the bit mask "wrap" we avoid having to
1151 * check two variables for all permutations: --> faster!
1154 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
1170 case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
1173 * Since both rqs are wrapped,
1174 * start with the one that's further behind head
1175 * (--> only *one* back seek required),
1176 * since back seek takes more time than forward.
1186 * The below is leftmost cache rbtree addon
1188 static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
1190 /* Service tree is empty */
1195 root->left = rb_first(&root->rb);
1198 return rb_entry(root->left, struct cfq_queue, rb_node);
1203 static struct cfq_group *cfq_rb_first_group(struct cfq_rb_root *root)
1206 root->left = rb_first(&root->rb);
1209 return rb_entry_cfqg(root->left);
1214 static void rb_erase_init(struct rb_node *n, struct rb_root *root)
1220 static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
1222 if (root->left == n)
1224 rb_erase_init(n, &root->rb);
1229 * would be nice to take fifo expire time into account as well
1231 static struct request *
1232 cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1233 struct request *last)
1235 struct rb_node *rbnext = rb_next(&last->rb_node);
1236 struct rb_node *rbprev = rb_prev(&last->rb_node);
1237 struct request *next = NULL, *prev = NULL;
1239 BUG_ON(RB_EMPTY_NODE(&last->rb_node));
1242 prev = rb_entry_rq(rbprev);
1245 next = rb_entry_rq(rbnext);
1247 rbnext = rb_first(&cfqq->sort_list);
1248 if (rbnext && rbnext != &last->rb_node)
1249 next = rb_entry_rq(rbnext);
1252 return cfq_choose_req(cfqd, next, prev, blk_rq_pos(last));
1255 static u64 cfq_slice_offset(struct cfq_data *cfqd,
1256 struct cfq_queue *cfqq)
1259 * just an approximation, should be ok.
1261 return (cfqq->cfqg->nr_cfqq - 1) * (cfq_prio_slice(cfqd, 1, 0) -
1262 cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
1266 cfqg_key(struct cfq_rb_root *st, struct cfq_group *cfqg)
1268 return cfqg->vdisktime - st->min_vdisktime;
1272 __cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
1274 struct rb_node **node = &st->rb.rb_node;
1275 struct rb_node *parent = NULL;
1276 struct cfq_group *__cfqg;
1277 s64 key = cfqg_key(st, cfqg);
1280 while (*node != NULL) {
1282 __cfqg = rb_entry_cfqg(parent);
1284 if (key < cfqg_key(st, __cfqg))
1285 node = &parent->rb_left;
1287 node = &parent->rb_right;
1293 st->left = &cfqg->rb_node;
1295 rb_link_node(&cfqg->rb_node, parent, node);
1296 rb_insert_color(&cfqg->rb_node, &st->rb);
1300 * This has to be called only on activation of cfqg
1303 cfq_update_group_weight(struct cfq_group *cfqg)
1305 if (cfqg->new_weight) {
1306 cfqg->weight = cfqg->new_weight;
1307 cfqg->new_weight = 0;
1312 cfq_update_group_leaf_weight(struct cfq_group *cfqg)
1314 BUG_ON(!RB_EMPTY_NODE(&cfqg->rb_node));
1316 if (cfqg->new_leaf_weight) {
1317 cfqg->leaf_weight = cfqg->new_leaf_weight;
1318 cfqg->new_leaf_weight = 0;
1323 cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
1325 unsigned int vfr = 1 << CFQ_SERVICE_SHIFT; /* start with 1 */
1326 struct cfq_group *pos = cfqg;
1327 struct cfq_group *parent;
1330 /* add to the service tree */
1331 BUG_ON(!RB_EMPTY_NODE(&cfqg->rb_node));
1334 * Update leaf_weight. We cannot update weight at this point
1335 * because cfqg might already have been activated and is
1336 * contributing its current weight to the parent's child_weight.
1338 cfq_update_group_leaf_weight(cfqg);
1339 __cfq_group_service_tree_add(st, cfqg);
1342 * Activate @cfqg and calculate the portion of vfraction @cfqg is
1343 * entitled to. vfraction is calculated by walking the tree
1344 * towards the root calculating the fraction it has at each level.
1345 * The compounded ratio is how much vfraction @cfqg owns.
1347 * Start with the proportion tasks in this cfqg has against active
1348 * children cfqgs - its leaf_weight against children_weight.
1350 propagate = !pos->nr_active++;
1351 pos->children_weight += pos->leaf_weight;
1352 vfr = vfr * pos->leaf_weight / pos->children_weight;
1355 * Compound ->weight walking up the tree. Both activation and
1356 * vfraction calculation are done in the same loop. Propagation
1357 * stops once an already activated node is met. vfraction
1358 * calculation should always continue to the root.
1360 while ((parent = cfqg_parent(pos))) {
1362 cfq_update_group_weight(pos);
1363 propagate = !parent->nr_active++;
1364 parent->children_weight += pos->weight;
1366 vfr = vfr * pos->weight / parent->children_weight;
1370 cfqg->vfraction = max_t(unsigned, vfr, 1);
1374 cfq_group_notify_queue_add(struct cfq_data *cfqd, struct cfq_group *cfqg)
1376 struct cfq_rb_root *st = &cfqd->grp_service_tree;
1377 struct cfq_group *__cfqg;
1381 if (!RB_EMPTY_NODE(&cfqg->rb_node))
1385 * Currently put the group at the end. Later implement something
1386 * so that groups get lesser vtime based on their weights, so that
1387 * if group does not loose all if it was not continuously backlogged.
1389 n = rb_last(&st->rb);
1391 __cfqg = rb_entry_cfqg(n);
1392 cfqg->vdisktime = __cfqg->vdisktime + CFQ_IDLE_DELAY;
1394 cfqg->vdisktime = st->min_vdisktime;
1395 cfq_group_service_tree_add(st, cfqg);
1399 cfq_group_service_tree_del(struct cfq_rb_root *st, struct cfq_group *cfqg)
1401 struct cfq_group *pos = cfqg;
1405 * Undo activation from cfq_group_service_tree_add(). Deactivate
1406 * @cfqg and propagate deactivation upwards.
1408 propagate = !--pos->nr_active;
1409 pos->children_weight -= pos->leaf_weight;
1412 struct cfq_group *parent = cfqg_parent(pos);
1414 /* @pos has 0 nr_active at this point */
1415 WARN_ON_ONCE(pos->children_weight);
1421 propagate = !--parent->nr_active;
1422 parent->children_weight -= pos->weight;
1426 /* remove from the service tree */
1427 if (!RB_EMPTY_NODE(&cfqg->rb_node))
1428 cfq_rb_erase(&cfqg->rb_node, st);
1432 cfq_group_notify_queue_del(struct cfq_data *cfqd, struct cfq_group *cfqg)
1434 struct cfq_rb_root *st = &cfqd->grp_service_tree;
1436 BUG_ON(cfqg->nr_cfqq < 1);
1439 /* If there are other cfq queues under this group, don't delete it */
1443 cfq_log_cfqg(cfqd, cfqg, "del_from_rr group");
1444 cfq_group_service_tree_del(st, cfqg);
1445 cfqg->saved_wl_slice = 0;
1446 cfqg_stats_update_dequeue(cfqg);
1449 static inline u64 cfq_cfqq_slice_usage(struct cfq_queue *cfqq,
1450 u64 *unaccounted_time)
1453 u64 now = ktime_get_ns();
1456 * Queue got expired before even a single request completed or
1457 * got expired immediately after first request completion.
1459 if (!cfqq->slice_start || cfqq->slice_start == now) {
1461 * Also charge the seek time incurred to the group, otherwise
1462 * if there are mutiple queues in the group, each can dispatch
1463 * a single request on seeky media and cause lots of seek time
1464 * and group will never know it.
1466 slice_used = max_t(u64, (now - cfqq->dispatch_start), 1);
1468 slice_used = now - cfqq->slice_start;
1469 if (slice_used > cfqq->allocated_slice) {
1470 *unaccounted_time = slice_used - cfqq->allocated_slice;
1471 slice_used = cfqq->allocated_slice;
1473 if (cfqq->slice_start > cfqq->dispatch_start)
1474 *unaccounted_time += cfqq->slice_start -
1475 cfqq->dispatch_start;
1481 static void cfq_group_served(struct cfq_data *cfqd, struct cfq_group *cfqg,
1482 struct cfq_queue *cfqq)
1484 struct cfq_rb_root *st = &cfqd->grp_service_tree;
1485 u64 used_sl, charge, unaccounted_sl = 0;
1486 int nr_sync = cfqg->nr_cfqq - cfqg_busy_async_queues(cfqd, cfqg)
1487 - cfqg->service_tree_idle.count;
1489 u64 now = ktime_get_ns();
1491 BUG_ON(nr_sync < 0);
1492 used_sl = charge = cfq_cfqq_slice_usage(cfqq, &unaccounted_sl);
1494 if (iops_mode(cfqd))
1495 charge = cfqq->slice_dispatch;
1496 else if (!cfq_cfqq_sync(cfqq) && !nr_sync)
1497 charge = cfqq->allocated_slice;
1500 * Can't update vdisktime while on service tree and cfqg->vfraction
1501 * is valid only while on it. Cache vfr, leave the service tree,
1502 * update vdisktime and go back on. The re-addition to the tree
1503 * will also update the weights as necessary.
1505 vfr = cfqg->vfraction;
1506 cfq_group_service_tree_del(st, cfqg);
1507 cfqg->vdisktime += cfqg_scale_charge(charge, vfr);
1508 cfq_group_service_tree_add(st, cfqg);
1510 /* This group is being expired. Save the context */
1511 if (cfqd->workload_expires > now) {
1512 cfqg->saved_wl_slice = cfqd->workload_expires - now;
1513 cfqg->saved_wl_type = cfqd->serving_wl_type;
1514 cfqg->saved_wl_class = cfqd->serving_wl_class;
1516 cfqg->saved_wl_slice = 0;
1518 cfq_log_cfqg(cfqd, cfqg, "served: vt=%llu min_vt=%llu", cfqg->vdisktime,
1520 cfq_log_cfqq(cfqq->cfqd, cfqq,
1521 "sl_used=%llu disp=%llu charge=%llu iops=%u sect=%lu",
1522 used_sl, cfqq->slice_dispatch, charge,
1523 iops_mode(cfqd), cfqq->nr_sectors);
1524 cfqg_stats_update_timeslice_used(cfqg, used_sl, unaccounted_sl);
1525 cfqg_stats_set_start_empty_time(cfqg);
1529 * cfq_init_cfqg_base - initialize base part of a cfq_group
1530 * @cfqg: cfq_group to initialize
1532 * Initialize the base part which is used whether %CONFIG_CFQ_GROUP_IOSCHED
1533 * is enabled or not.
1535 static void cfq_init_cfqg_base(struct cfq_group *cfqg)
1537 struct cfq_rb_root *st;
1540 for_each_cfqg_st(cfqg, i, j, st)
1542 RB_CLEAR_NODE(&cfqg->rb_node);
1544 cfqg->ttime.last_end_request = ktime_get_ns();
1547 #ifdef CONFIG_CFQ_GROUP_IOSCHED
1548 static int __cfq_set_weight(struct cgroup_subsys_state *css, u64 val,
1549 bool on_dfl, bool reset_dev, bool is_leaf_weight);
1551 static void cfqg_stats_exit(struct cfqg_stats *stats)
1553 blkg_rwstat_exit(&stats->merged);
1554 blkg_rwstat_exit(&stats->service_time);
1555 blkg_rwstat_exit(&stats->wait_time);
1556 blkg_rwstat_exit(&stats->queued);
1557 blkg_stat_exit(&stats->time);
1558 #ifdef CONFIG_DEBUG_BLK_CGROUP
1559 blkg_stat_exit(&stats->unaccounted_time);
1560 blkg_stat_exit(&stats->avg_queue_size_sum);
1561 blkg_stat_exit(&stats->avg_queue_size_samples);
1562 blkg_stat_exit(&stats->dequeue);
1563 blkg_stat_exit(&stats->group_wait_time);
1564 blkg_stat_exit(&stats->idle_time);
1565 blkg_stat_exit(&stats->empty_time);
1569 static int cfqg_stats_init(struct cfqg_stats *stats, gfp_t gfp)
1571 if (blkg_rwstat_init(&stats->merged, gfp) ||
1572 blkg_rwstat_init(&stats->service_time, gfp) ||
1573 blkg_rwstat_init(&stats->wait_time, gfp) ||
1574 blkg_rwstat_init(&stats->queued, gfp) ||
1575 blkg_stat_init(&stats->time, gfp))
1578 #ifdef CONFIG_DEBUG_BLK_CGROUP
1579 if (blkg_stat_init(&stats->unaccounted_time, gfp) ||
1580 blkg_stat_init(&stats->avg_queue_size_sum, gfp) ||
1581 blkg_stat_init(&stats->avg_queue_size_samples, gfp) ||
1582 blkg_stat_init(&stats->dequeue, gfp) ||
1583 blkg_stat_init(&stats->group_wait_time, gfp) ||
1584 blkg_stat_init(&stats->idle_time, gfp) ||
1585 blkg_stat_init(&stats->empty_time, gfp))
1590 cfqg_stats_exit(stats);
1594 static struct blkcg_policy_data *cfq_cpd_alloc(gfp_t gfp)
1596 struct cfq_group_data *cgd;
1598 cgd = kzalloc(sizeof(*cgd), GFP_KERNEL);
1604 static void cfq_cpd_init(struct blkcg_policy_data *cpd)
1606 struct cfq_group_data *cgd = cpd_to_cfqgd(cpd);
1607 unsigned int weight = cgroup_subsys_on_dfl(io_cgrp_subsys) ?
1608 CGROUP_WEIGHT_DFL : CFQ_WEIGHT_LEGACY_DFL;
1610 if (cpd_to_blkcg(cpd) == &blkcg_root)
1613 cgd->weight = weight;
1614 cgd->leaf_weight = weight;
1617 static void cfq_cpd_free(struct blkcg_policy_data *cpd)
1619 kfree(cpd_to_cfqgd(cpd));
1622 static void cfq_cpd_bind(struct blkcg_policy_data *cpd)
1624 struct blkcg *blkcg = cpd_to_blkcg(cpd);
1625 bool on_dfl = cgroup_subsys_on_dfl(io_cgrp_subsys);
1626 unsigned int weight = on_dfl ? CGROUP_WEIGHT_DFL : CFQ_WEIGHT_LEGACY_DFL;
1628 if (blkcg == &blkcg_root)
1631 WARN_ON_ONCE(__cfq_set_weight(&blkcg->css, weight, on_dfl, true, false));
1632 WARN_ON_ONCE(__cfq_set_weight(&blkcg->css, weight, on_dfl, true, true));
1635 static struct blkg_policy_data *cfq_pd_alloc(gfp_t gfp, int node)
1637 struct cfq_group *cfqg;
1639 cfqg = kzalloc_node(sizeof(*cfqg), gfp, node);
1643 cfq_init_cfqg_base(cfqg);
1644 if (cfqg_stats_init(&cfqg->stats, gfp)) {
1652 static void cfq_pd_init(struct blkg_policy_data *pd)
1654 struct cfq_group *cfqg = pd_to_cfqg(pd);
1655 struct cfq_group_data *cgd = blkcg_to_cfqgd(pd->blkg->blkcg);
1657 cfqg->weight = cgd->weight;
1658 cfqg->leaf_weight = cgd->leaf_weight;
1661 static void cfq_pd_offline(struct blkg_policy_data *pd)
1663 struct cfq_group *cfqg = pd_to_cfqg(pd);
1666 for (i = 0; i < IOPRIO_BE_NR; i++) {
1667 if (cfqg->async_cfqq[0][i])
1668 cfq_put_queue(cfqg->async_cfqq[0][i]);
1669 if (cfqg->async_cfqq[1][i])
1670 cfq_put_queue(cfqg->async_cfqq[1][i]);
1673 if (cfqg->async_idle_cfqq)
1674 cfq_put_queue(cfqg->async_idle_cfqq);
1677 * @blkg is going offline and will be ignored by
1678 * blkg_[rw]stat_recursive_sum(). Transfer stats to the parent so
1679 * that they don't get lost. If IOs complete after this point, the
1680 * stats for them will be lost. Oh well...
1682 cfqg_stats_xfer_dead(cfqg);
1685 static void cfq_pd_free(struct blkg_policy_data *pd)
1687 struct cfq_group *cfqg = pd_to_cfqg(pd);
1689 cfqg_stats_exit(&cfqg->stats);
1693 static void cfq_pd_reset_stats(struct blkg_policy_data *pd)
1695 struct cfq_group *cfqg = pd_to_cfqg(pd);
1697 cfqg_stats_reset(&cfqg->stats);
1700 static struct cfq_group *cfq_lookup_cfqg(struct cfq_data *cfqd,
1701 struct blkcg *blkcg)
1703 struct blkcg_gq *blkg;
1705 blkg = blkg_lookup(blkcg, cfqd->queue);
1707 return blkg_to_cfqg(blkg);
1711 static void cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg)
1714 /* cfqq reference on cfqg */
1718 static u64 cfqg_prfill_weight_device(struct seq_file *sf,
1719 struct blkg_policy_data *pd, int off)
1721 struct cfq_group *cfqg = pd_to_cfqg(pd);
1723 if (!cfqg->dev_weight)
1725 return __blkg_prfill_u64(sf, pd, cfqg->dev_weight);
1728 static int cfqg_print_weight_device(struct seq_file *sf, void *v)
1730 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1731 cfqg_prfill_weight_device, &blkcg_policy_cfq,
1736 static u64 cfqg_prfill_leaf_weight_device(struct seq_file *sf,
1737 struct blkg_policy_data *pd, int off)
1739 struct cfq_group *cfqg = pd_to_cfqg(pd);
1741 if (!cfqg->dev_leaf_weight)
1743 return __blkg_prfill_u64(sf, pd, cfqg->dev_leaf_weight);
1746 static int cfqg_print_leaf_weight_device(struct seq_file *sf, void *v)
1748 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1749 cfqg_prfill_leaf_weight_device, &blkcg_policy_cfq,
1754 static int cfq_print_weight(struct seq_file *sf, void *v)
1756 struct blkcg *blkcg = css_to_blkcg(seq_css(sf));
1757 struct cfq_group_data *cgd = blkcg_to_cfqgd(blkcg);
1758 unsigned int val = 0;
1763 seq_printf(sf, "%u\n", val);
1767 static int cfq_print_leaf_weight(struct seq_file *sf, void *v)
1769 struct blkcg *blkcg = css_to_blkcg(seq_css(sf));
1770 struct cfq_group_data *cgd = blkcg_to_cfqgd(blkcg);
1771 unsigned int val = 0;
1774 val = cgd->leaf_weight;
1776 seq_printf(sf, "%u\n", val);
1780 static ssize_t __cfqg_set_weight_device(struct kernfs_open_file *of,
1781 char *buf, size_t nbytes, loff_t off,
1782 bool on_dfl, bool is_leaf_weight)
1784 unsigned int min = on_dfl ? CGROUP_WEIGHT_MIN : CFQ_WEIGHT_LEGACY_MIN;
1785 unsigned int max = on_dfl ? CGROUP_WEIGHT_MAX : CFQ_WEIGHT_LEGACY_MAX;
1786 struct blkcg *blkcg = css_to_blkcg(of_css(of));
1787 struct blkg_conf_ctx ctx;
1788 struct cfq_group *cfqg;
1789 struct cfq_group_data *cfqgd;
1793 ret = blkg_conf_prep(blkcg, &blkcg_policy_cfq, buf, &ctx);
1797 if (sscanf(ctx.body, "%llu", &v) == 1) {
1798 /* require "default" on dfl */
1802 } else if (!strcmp(strim(ctx.body), "default")) {
1809 cfqg = blkg_to_cfqg(ctx.blkg);
1810 cfqgd = blkcg_to_cfqgd(blkcg);
1813 if (!v || (v >= min && v <= max)) {
1814 if (!is_leaf_weight) {
1815 cfqg->dev_weight = v;
1816 cfqg->new_weight = v ?: cfqgd->weight;
1818 cfqg->dev_leaf_weight = v;
1819 cfqg->new_leaf_weight = v ?: cfqgd->leaf_weight;
1824 blkg_conf_finish(&ctx);
1825 return ret ?: nbytes;
1828 static ssize_t cfqg_set_weight_device(struct kernfs_open_file *of,
1829 char *buf, size_t nbytes, loff_t off)
1831 return __cfqg_set_weight_device(of, buf, nbytes, off, false, false);
1834 static ssize_t cfqg_set_leaf_weight_device(struct kernfs_open_file *of,
1835 char *buf, size_t nbytes, loff_t off)
1837 return __cfqg_set_weight_device(of, buf, nbytes, off, false, true);
1840 static int __cfq_set_weight(struct cgroup_subsys_state *css, u64 val,
1841 bool on_dfl, bool reset_dev, bool is_leaf_weight)
1843 unsigned int min = on_dfl ? CGROUP_WEIGHT_MIN : CFQ_WEIGHT_LEGACY_MIN;
1844 unsigned int max = on_dfl ? CGROUP_WEIGHT_MAX : CFQ_WEIGHT_LEGACY_MAX;
1845 struct blkcg *blkcg = css_to_blkcg(css);
1846 struct blkcg_gq *blkg;
1847 struct cfq_group_data *cfqgd;
1850 if (val < min || val > max)
1853 spin_lock_irq(&blkcg->lock);
1854 cfqgd = blkcg_to_cfqgd(blkcg);
1860 if (!is_leaf_weight)
1861 cfqgd->weight = val;
1863 cfqgd->leaf_weight = val;
1865 hlist_for_each_entry(blkg, &blkcg->blkg_list, blkcg_node) {
1866 struct cfq_group *cfqg = blkg_to_cfqg(blkg);
1871 if (!is_leaf_weight) {
1873 cfqg->dev_weight = 0;
1874 if (!cfqg->dev_weight)
1875 cfqg->new_weight = cfqgd->weight;
1878 cfqg->dev_leaf_weight = 0;
1879 if (!cfqg->dev_leaf_weight)
1880 cfqg->new_leaf_weight = cfqgd->leaf_weight;
1885 spin_unlock_irq(&blkcg->lock);
1889 static int cfq_set_weight(struct cgroup_subsys_state *css, struct cftype *cft,
1892 return __cfq_set_weight(css, val, false, false, false);
1895 static int cfq_set_leaf_weight(struct cgroup_subsys_state *css,
1896 struct cftype *cft, u64 val)
1898 return __cfq_set_weight(css, val, false, false, true);
1901 static int cfqg_print_stat(struct seq_file *sf, void *v)
1903 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), blkg_prfill_stat,
1904 &blkcg_policy_cfq, seq_cft(sf)->private, false);
1908 static int cfqg_print_rwstat(struct seq_file *sf, void *v)
1910 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), blkg_prfill_rwstat,
1911 &blkcg_policy_cfq, seq_cft(sf)->private, true);
1915 static u64 cfqg_prfill_stat_recursive(struct seq_file *sf,
1916 struct blkg_policy_data *pd, int off)
1918 u64 sum = blkg_stat_recursive_sum(pd_to_blkg(pd),
1919 &blkcg_policy_cfq, off);
1920 return __blkg_prfill_u64(sf, pd, sum);
1923 static u64 cfqg_prfill_rwstat_recursive(struct seq_file *sf,
1924 struct blkg_policy_data *pd, int off)
1926 struct blkg_rwstat sum = blkg_rwstat_recursive_sum(pd_to_blkg(pd),
1927 &blkcg_policy_cfq, off);
1928 return __blkg_prfill_rwstat(sf, pd, &sum);
1931 static int cfqg_print_stat_recursive(struct seq_file *sf, void *v)
1933 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1934 cfqg_prfill_stat_recursive, &blkcg_policy_cfq,
1935 seq_cft(sf)->private, false);
1939 static int cfqg_print_rwstat_recursive(struct seq_file *sf, void *v)
1941 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1942 cfqg_prfill_rwstat_recursive, &blkcg_policy_cfq,
1943 seq_cft(sf)->private, true);
1947 static u64 cfqg_prfill_sectors(struct seq_file *sf, struct blkg_policy_data *pd,
1950 u64 sum = blkg_rwstat_total(&pd->blkg->stat_bytes);
1952 return __blkg_prfill_u64(sf, pd, sum >> 9);
1955 static int cfqg_print_stat_sectors(struct seq_file *sf, void *v)
1957 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1958 cfqg_prfill_sectors, &blkcg_policy_cfq, 0, false);
1962 static u64 cfqg_prfill_sectors_recursive(struct seq_file *sf,
1963 struct blkg_policy_data *pd, int off)
1965 struct blkg_rwstat tmp = blkg_rwstat_recursive_sum(pd->blkg, NULL,
1966 offsetof(struct blkcg_gq, stat_bytes));
1967 u64 sum = atomic64_read(&tmp.aux_cnt[BLKG_RWSTAT_READ]) +
1968 atomic64_read(&tmp.aux_cnt[BLKG_RWSTAT_WRITE]);
1970 return __blkg_prfill_u64(sf, pd, sum >> 9);
1973 static int cfqg_print_stat_sectors_recursive(struct seq_file *sf, void *v)
1975 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1976 cfqg_prfill_sectors_recursive, &blkcg_policy_cfq, 0,
1981 #ifdef CONFIG_DEBUG_BLK_CGROUP
1982 static u64 cfqg_prfill_avg_queue_size(struct seq_file *sf,
1983 struct blkg_policy_data *pd, int off)
1985 struct cfq_group *cfqg = pd_to_cfqg(pd);
1986 u64 samples = blkg_stat_read(&cfqg->stats.avg_queue_size_samples);
1990 v = blkg_stat_read(&cfqg->stats.avg_queue_size_sum);
1991 v = div64_u64(v, samples);
1993 __blkg_prfill_u64(sf, pd, v);
1997 /* print avg_queue_size */
1998 static int cfqg_print_avg_queue_size(struct seq_file *sf, void *v)
2000 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
2001 cfqg_prfill_avg_queue_size, &blkcg_policy_cfq,
2005 #endif /* CONFIG_DEBUG_BLK_CGROUP */
2007 static struct cftype cfq_blkcg_legacy_files[] = {
2008 /* on root, weight is mapped to leaf_weight */
2010 .name = "weight_device",
2011 .flags = CFTYPE_ONLY_ON_ROOT,
2012 .seq_show = cfqg_print_leaf_weight_device,
2013 .write = cfqg_set_leaf_weight_device,
2017 .flags = CFTYPE_ONLY_ON_ROOT,
2018 .seq_show = cfq_print_leaf_weight,
2019 .write_u64 = cfq_set_leaf_weight,
2022 /* no such mapping necessary for !roots */
2024 .name = "weight_device",
2025 .flags = CFTYPE_NOT_ON_ROOT,
2026 .seq_show = cfqg_print_weight_device,
2027 .write = cfqg_set_weight_device,
2031 .flags = CFTYPE_NOT_ON_ROOT,
2032 .seq_show = cfq_print_weight,
2033 .write_u64 = cfq_set_weight,
2037 .name = "leaf_weight_device",
2038 .seq_show = cfqg_print_leaf_weight_device,
2039 .write = cfqg_set_leaf_weight_device,
2042 .name = "leaf_weight",
2043 .seq_show = cfq_print_leaf_weight,
2044 .write_u64 = cfq_set_leaf_weight,
2047 /* statistics, covers only the tasks in the cfqg */
2050 .private = offsetof(struct cfq_group, stats.time),
2051 .seq_show = cfqg_print_stat,
2055 .seq_show = cfqg_print_stat_sectors,
2058 .name = "io_service_bytes",
2059 .private = (unsigned long)&blkcg_policy_cfq,
2060 .seq_show = blkg_print_stat_bytes,
2063 .name = "io_serviced",
2064 .private = (unsigned long)&blkcg_policy_cfq,
2065 .seq_show = blkg_print_stat_ios,
2068 .name = "io_service_time",
2069 .private = offsetof(struct cfq_group, stats.service_time),
2070 .seq_show = cfqg_print_rwstat,
2073 .name = "io_wait_time",
2074 .private = offsetof(struct cfq_group, stats.wait_time),
2075 .seq_show = cfqg_print_rwstat,
2078 .name = "io_merged",
2079 .private = offsetof(struct cfq_group, stats.merged),
2080 .seq_show = cfqg_print_rwstat,
2083 .name = "io_queued",
2084 .private = offsetof(struct cfq_group, stats.queued),
2085 .seq_show = cfqg_print_rwstat,
2088 /* the same statictics which cover the cfqg and its descendants */
2090 .name = "time_recursive",
2091 .private = offsetof(struct cfq_group, stats.time),
2092 .seq_show = cfqg_print_stat_recursive,
2095 .name = "sectors_recursive",
2096 .seq_show = cfqg_print_stat_sectors_recursive,
2099 .name = "io_service_bytes_recursive",
2100 .private = (unsigned long)&blkcg_policy_cfq,
2101 .seq_show = blkg_print_stat_bytes_recursive,
2104 .name = "io_serviced_recursive",
2105 .private = (unsigned long)&blkcg_policy_cfq,
2106 .seq_show = blkg_print_stat_ios_recursive,
2109 .name = "io_service_time_recursive",
2110 .private = offsetof(struct cfq_group, stats.service_time),
2111 .seq_show = cfqg_print_rwstat_recursive,
2114 .name = "io_wait_time_recursive",
2115 .private = offsetof(struct cfq_group, stats.wait_time),
2116 .seq_show = cfqg_print_rwstat_recursive,
2119 .name = "io_merged_recursive",
2120 .private = offsetof(struct cfq_group, stats.merged),
2121 .seq_show = cfqg_print_rwstat_recursive,
2124 .name = "io_queued_recursive",
2125 .private = offsetof(struct cfq_group, stats.queued),
2126 .seq_show = cfqg_print_rwstat_recursive,
2128 #ifdef CONFIG_DEBUG_BLK_CGROUP
2130 .name = "avg_queue_size",
2131 .seq_show = cfqg_print_avg_queue_size,
2134 .name = "group_wait_time",
2135 .private = offsetof(struct cfq_group, stats.group_wait_time),
2136 .seq_show = cfqg_print_stat,
2139 .name = "idle_time",
2140 .private = offsetof(struct cfq_group, stats.idle_time),
2141 .seq_show = cfqg_print_stat,
2144 .name = "empty_time",
2145 .private = offsetof(struct cfq_group, stats.empty_time),
2146 .seq_show = cfqg_print_stat,
2150 .private = offsetof(struct cfq_group, stats.dequeue),
2151 .seq_show = cfqg_print_stat,
2154 .name = "unaccounted_time",
2155 .private = offsetof(struct cfq_group, stats.unaccounted_time),
2156 .seq_show = cfqg_print_stat,
2158 #endif /* CONFIG_DEBUG_BLK_CGROUP */
2162 static int cfq_print_weight_on_dfl(struct seq_file *sf, void *v)
2164 struct blkcg *blkcg = css_to_blkcg(seq_css(sf));
2165 struct cfq_group_data *cgd = blkcg_to_cfqgd(blkcg);
2167 seq_printf(sf, "default %u\n", cgd->weight);
2168 blkcg_print_blkgs(sf, blkcg, cfqg_prfill_weight_device,
2169 &blkcg_policy_cfq, 0, false);
2173 static ssize_t cfq_set_weight_on_dfl(struct kernfs_open_file *of,
2174 char *buf, size_t nbytes, loff_t off)
2182 /* "WEIGHT" or "default WEIGHT" sets the default weight */
2183 v = simple_strtoull(buf, &endp, 0);
2184 if (*endp == '\0' || sscanf(buf, "default %llu", &v) == 1) {
2185 ret = __cfq_set_weight(of_css(of), v, true, false, false);
2186 return ret ?: nbytes;
2189 /* "MAJ:MIN WEIGHT" */
2190 return __cfqg_set_weight_device(of, buf, nbytes, off, true, false);
2193 static struct cftype cfq_blkcg_files[] = {
2196 .flags = CFTYPE_NOT_ON_ROOT,
2197 .seq_show = cfq_print_weight_on_dfl,
2198 .write = cfq_set_weight_on_dfl,
2203 #else /* GROUP_IOSCHED */
2204 static struct cfq_group *cfq_lookup_cfqg(struct cfq_data *cfqd,
2205 struct blkcg *blkcg)
2207 return cfqd->root_group;
2211 cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) {
2215 #endif /* GROUP_IOSCHED */
2218 * The cfqd->service_trees holds all pending cfq_queue's that have
2219 * requests waiting to be processed. It is sorted in the order that
2220 * we will service the queues.
2222 static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2225 struct rb_node **p, *parent;
2226 struct cfq_queue *__cfqq;
2228 struct cfq_rb_root *st;
2231 u64 now = ktime_get_ns();
2233 st = st_for(cfqq->cfqg, cfqq_class(cfqq), cfqq_type(cfqq));
2234 if (cfq_class_idle(cfqq)) {
2235 rb_key = CFQ_IDLE_DELAY;
2236 parent = rb_last(&st->rb);
2237 if (parent && parent != &cfqq->rb_node) {
2238 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
2239 rb_key += __cfqq->rb_key;
2242 } else if (!add_front) {
2244 * Get our rb key offset. Subtract any residual slice
2245 * value carried from last service. A negative resid
2246 * count indicates slice overrun, and this should position
2247 * the next service time further away in the tree.
2249 rb_key = cfq_slice_offset(cfqd, cfqq) + now;
2250 rb_key -= cfqq->slice_resid;
2251 cfqq->slice_resid = 0;
2253 rb_key = -NSEC_PER_SEC;
2254 __cfqq = cfq_rb_first(st);
2255 rb_key += __cfqq ? __cfqq->rb_key : now;
2258 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
2261 * same position, nothing more to do
2263 if (rb_key == cfqq->rb_key && cfqq->service_tree == st)
2266 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
2267 cfqq->service_tree = NULL;
2272 cfqq->service_tree = st;
2273 p = &st->rb.rb_node;
2276 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
2279 * sort by key, that represents service time.
2281 if (rb_key < __cfqq->rb_key)
2282 p = &parent->rb_left;
2284 p = &parent->rb_right;
2290 st->left = &cfqq->rb_node;
2292 cfqq->rb_key = rb_key;
2293 rb_link_node(&cfqq->rb_node, parent, p);
2294 rb_insert_color(&cfqq->rb_node, &st->rb);
2296 if (add_front || !new_cfqq)
2298 cfq_group_notify_queue_add(cfqd, cfqq->cfqg);
2301 static struct cfq_queue *
2302 cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
2303 sector_t sector, struct rb_node **ret_parent,
2304 struct rb_node ***rb_link)
2306 struct rb_node **p, *parent;
2307 struct cfq_queue *cfqq = NULL;
2315 cfqq = rb_entry(parent, struct cfq_queue, p_node);
2318 * Sort strictly based on sector. Smallest to the left,
2319 * largest to the right.
2321 if (sector > blk_rq_pos(cfqq->next_rq))
2322 n = &(*p)->rb_right;
2323 else if (sector < blk_rq_pos(cfqq->next_rq))
2331 *ret_parent = parent;
2337 static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2339 struct rb_node **p, *parent;
2340 struct cfq_queue *__cfqq;
2343 rb_erase(&cfqq->p_node, cfqq->p_root);
2344 cfqq->p_root = NULL;
2347 if (cfq_class_idle(cfqq))
2352 cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
2353 __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
2354 blk_rq_pos(cfqq->next_rq), &parent, &p);
2356 rb_link_node(&cfqq->p_node, parent, p);
2357 rb_insert_color(&cfqq->p_node, cfqq->p_root);
2359 cfqq->p_root = NULL;
2363 * Update cfqq's position in the service tree.
2365 static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2368 * Resorting requires the cfqq to be on the RR list already.
2370 if (cfq_cfqq_on_rr(cfqq)) {
2371 cfq_service_tree_add(cfqd, cfqq, 0);
2372 cfq_prio_tree_add(cfqd, cfqq);
2377 * add to busy list of queues for service, trying to be fair in ordering
2378 * the pending list according to last request service
2380 static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2382 cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
2383 BUG_ON(cfq_cfqq_on_rr(cfqq));
2384 cfq_mark_cfqq_on_rr(cfqq);
2385 cfqd->busy_queues++;
2386 if (cfq_cfqq_sync(cfqq))
2387 cfqd->busy_sync_queues++;
2389 cfq_resort_rr_list(cfqd, cfqq);
2393 * Called when the cfqq no longer has requests pending, remove it from
2396 static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2398 cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
2399 BUG_ON(!cfq_cfqq_on_rr(cfqq));
2400 cfq_clear_cfqq_on_rr(cfqq);
2402 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
2403 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
2404 cfqq->service_tree = NULL;
2407 rb_erase(&cfqq->p_node, cfqq->p_root);
2408 cfqq->p_root = NULL;
2411 cfq_group_notify_queue_del(cfqd, cfqq->cfqg);
2412 BUG_ON(!cfqd->busy_queues);
2413 cfqd->busy_queues--;
2414 if (cfq_cfqq_sync(cfqq))
2415 cfqd->busy_sync_queues--;
2419 * rb tree support functions
2421 static void cfq_del_rq_rb(struct request *rq)
2423 struct cfq_queue *cfqq = RQ_CFQQ(rq);
2424 const int sync = rq_is_sync(rq);
2426 BUG_ON(!cfqq->queued[sync]);
2427 cfqq->queued[sync]--;
2429 elv_rb_del(&cfqq->sort_list, rq);
2431 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) {
2433 * Queue will be deleted from service tree when we actually
2434 * expire it later. Right now just remove it from prio tree
2438 rb_erase(&cfqq->p_node, cfqq->p_root);
2439 cfqq->p_root = NULL;
2444 static void cfq_add_rq_rb(struct request *rq)
2446 struct cfq_queue *cfqq = RQ_CFQQ(rq);
2447 struct cfq_data *cfqd = cfqq->cfqd;
2448 struct request *prev;
2450 cfqq->queued[rq_is_sync(rq)]++;
2452 elv_rb_add(&cfqq->sort_list, rq);
2454 if (!cfq_cfqq_on_rr(cfqq))
2455 cfq_add_cfqq_rr(cfqd, cfqq);
2458 * check if this request is a better next-serve candidate
2460 prev = cfqq->next_rq;
2461 cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq, cfqd->last_position);
2464 * adjust priority tree position, if ->next_rq changes
2466 if (prev != cfqq->next_rq)
2467 cfq_prio_tree_add(cfqd, cfqq);
2469 BUG_ON(!cfqq->next_rq);
2472 static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
2474 elv_rb_del(&cfqq->sort_list, rq);
2475 cfqq->queued[rq_is_sync(rq)]--;
2476 cfqg_stats_update_io_remove(RQ_CFQG(rq), req_op(rq), rq->cmd_flags);
2478 cfqg_stats_update_io_add(RQ_CFQG(rq), cfqq->cfqd->serving_group,
2479 req_op(rq), rq->cmd_flags);
2482 static struct request *
2483 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
2485 struct task_struct *tsk = current;
2486 struct cfq_io_cq *cic;
2487 struct cfq_queue *cfqq;
2489 cic = cfq_cic_lookup(cfqd, tsk->io_context);
2493 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
2495 return elv_rb_find(&cfqq->sort_list, bio_end_sector(bio));
2500 static void cfq_activate_request(struct request_queue *q, struct request *rq)
2502 struct cfq_data *cfqd = q->elevator->elevator_data;
2504 cfqd->rq_in_driver++;
2505 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
2506 cfqd->rq_in_driver);
2508 cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
2511 static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
2513 struct cfq_data *cfqd = q->elevator->elevator_data;
2515 WARN_ON(!cfqd->rq_in_driver);
2516 cfqd->rq_in_driver--;
2517 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
2518 cfqd->rq_in_driver);
2521 static void cfq_remove_request(struct request *rq)
2523 struct cfq_queue *cfqq = RQ_CFQQ(rq);
2525 if (cfqq->next_rq == rq)
2526 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
2528 list_del_init(&rq->queuelist);
2531 cfqq->cfqd->rq_queued--;
2532 cfqg_stats_update_io_remove(RQ_CFQG(rq), req_op(rq), rq->cmd_flags);
2533 if (rq->cmd_flags & REQ_PRIO) {
2534 WARN_ON(!cfqq->prio_pending);
2535 cfqq->prio_pending--;
2539 static int cfq_merge(struct request_queue *q, struct request **req,
2542 struct cfq_data *cfqd = q->elevator->elevator_data;
2543 struct request *__rq;
2545 __rq = cfq_find_rq_fmerge(cfqd, bio);
2546 if (__rq && elv_rq_merge_ok(__rq, bio)) {
2548 return ELEVATOR_FRONT_MERGE;
2551 return ELEVATOR_NO_MERGE;
2554 static void cfq_merged_request(struct request_queue *q, struct request *req,
2557 if (type == ELEVATOR_FRONT_MERGE) {
2558 struct cfq_queue *cfqq = RQ_CFQQ(req);
2560 cfq_reposition_rq_rb(cfqq, req);
2564 static void cfq_bio_merged(struct request_queue *q, struct request *req,
2567 cfqg_stats_update_io_merged(RQ_CFQG(req), bio_op(bio), bio->bi_rw);
2571 cfq_merged_requests(struct request_queue *q, struct request *rq,
2572 struct request *next)
2574 struct cfq_queue *cfqq = RQ_CFQQ(rq);
2575 struct cfq_data *cfqd = q->elevator->elevator_data;
2578 * reposition in fifo if next is older than rq
2580 if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
2581 next->fifo_time < rq->fifo_time &&
2582 cfqq == RQ_CFQQ(next)) {
2583 list_move(&rq->queuelist, &next->queuelist);
2584 rq->fifo_time = next->fifo_time;
2587 if (cfqq->next_rq == next)
2589 cfq_remove_request(next);
2590 cfqg_stats_update_io_merged(RQ_CFQG(rq), req_op(next), next->cmd_flags);
2592 cfqq = RQ_CFQQ(next);
2594 * all requests of this queue are merged to other queues, delete it
2595 * from the service tree. If it's the active_queue,
2596 * cfq_dispatch_requests() will choose to expire it or do idle
2598 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list) &&
2599 cfqq != cfqd->active_queue)
2600 cfq_del_cfqq_rr(cfqd, cfqq);
2603 static int cfq_allow_merge(struct request_queue *q, struct request *rq,
2606 struct cfq_data *cfqd = q->elevator->elevator_data;
2607 struct cfq_io_cq *cic;
2608 struct cfq_queue *cfqq;
2611 * Disallow merge of a sync bio into an async request.
2613 if (cfq_bio_sync(bio) && !rq_is_sync(rq))
2617 * Lookup the cfqq that this bio will be queued with and allow
2618 * merge only if rq is queued there.
2620 cic = cfq_cic_lookup(cfqd, current->io_context);
2624 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
2625 return cfqq == RQ_CFQQ(rq);
2628 static inline void cfq_del_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2630 hrtimer_try_to_cancel(&cfqd->idle_slice_timer);
2631 cfqg_stats_update_idle_time(cfqq->cfqg);
2634 static void __cfq_set_active_queue(struct cfq_data *cfqd,
2635 struct cfq_queue *cfqq)
2638 cfq_log_cfqq(cfqd, cfqq, "set_active wl_class:%d wl_type:%d",
2639 cfqd->serving_wl_class, cfqd->serving_wl_type);
2640 cfqg_stats_update_avg_queue_size(cfqq->cfqg);
2641 cfqq->slice_start = 0;
2642 cfqq->dispatch_start = ktime_get_ns();
2643 cfqq->allocated_slice = 0;
2644 cfqq->slice_end = 0;
2645 cfqq->slice_dispatch = 0;
2646 cfqq->nr_sectors = 0;
2648 cfq_clear_cfqq_wait_request(cfqq);
2649 cfq_clear_cfqq_must_dispatch(cfqq);
2650 cfq_clear_cfqq_must_alloc_slice(cfqq);
2651 cfq_clear_cfqq_fifo_expire(cfqq);
2652 cfq_mark_cfqq_slice_new(cfqq);
2654 cfq_del_timer(cfqd, cfqq);
2657 cfqd->active_queue = cfqq;
2661 * current cfqq expired its slice (or was too idle), select new one
2664 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2667 cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
2669 if (cfq_cfqq_wait_request(cfqq))
2670 cfq_del_timer(cfqd, cfqq);
2672 cfq_clear_cfqq_wait_request(cfqq);
2673 cfq_clear_cfqq_wait_busy(cfqq);
2676 * If this cfqq is shared between multiple processes, check to
2677 * make sure that those processes are still issuing I/Os within
2678 * the mean seek distance. If not, it may be time to break the
2679 * queues apart again.
2681 if (cfq_cfqq_coop(cfqq) && CFQQ_SEEKY(cfqq))
2682 cfq_mark_cfqq_split_coop(cfqq);
2685 * store what was left of this slice, if the queue idled/timed out
2688 if (cfq_cfqq_slice_new(cfqq))
2689 cfqq->slice_resid = cfq_scaled_cfqq_slice(cfqd, cfqq);
2691 cfqq->slice_resid = cfqq->slice_end - ktime_get_ns();
2692 cfq_log_cfqq(cfqd, cfqq, "resid=%lld", cfqq->slice_resid);
2695 cfq_group_served(cfqd, cfqq->cfqg, cfqq);
2697 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
2698 cfq_del_cfqq_rr(cfqd, cfqq);
2700 cfq_resort_rr_list(cfqd, cfqq);
2702 if (cfqq == cfqd->active_queue)
2703 cfqd->active_queue = NULL;
2705 if (cfqd->active_cic) {
2706 put_io_context(cfqd->active_cic->icq.ioc);
2707 cfqd->active_cic = NULL;
2711 static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out)
2713 struct cfq_queue *cfqq = cfqd->active_queue;
2716 __cfq_slice_expired(cfqd, cfqq, timed_out);
2720 * Get next queue for service. Unless we have a queue preemption,
2721 * we'll simply select the first cfqq in the service tree.
2723 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
2725 struct cfq_rb_root *st = st_for(cfqd->serving_group,
2726 cfqd->serving_wl_class, cfqd->serving_wl_type);
2728 if (!cfqd->rq_queued)
2731 /* There is nothing to dispatch */
2734 if (RB_EMPTY_ROOT(&st->rb))
2736 return cfq_rb_first(st);
2739 static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd)
2741 struct cfq_group *cfqg;
2742 struct cfq_queue *cfqq;
2744 struct cfq_rb_root *st;
2746 if (!cfqd->rq_queued)
2749 cfqg = cfq_get_next_cfqg(cfqd);
2753 for_each_cfqg_st(cfqg, i, j, st)
2754 if ((cfqq = cfq_rb_first(st)) != NULL)
2760 * Get and set a new active queue for service.
2762 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
2763 struct cfq_queue *cfqq)
2766 cfqq = cfq_get_next_queue(cfqd);
2768 __cfq_set_active_queue(cfqd, cfqq);
2772 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
2775 if (blk_rq_pos(rq) >= cfqd->last_position)
2776 return blk_rq_pos(rq) - cfqd->last_position;
2778 return cfqd->last_position - blk_rq_pos(rq);
2781 static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2784 return cfq_dist_from_last(cfqd, rq) <= CFQQ_CLOSE_THR;
2787 static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
2788 struct cfq_queue *cur_cfqq)
2790 struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
2791 struct rb_node *parent, *node;
2792 struct cfq_queue *__cfqq;
2793 sector_t sector = cfqd->last_position;
2795 if (RB_EMPTY_ROOT(root))
2799 * First, if we find a request starting at the end of the last
2800 * request, choose it.
2802 __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
2807 * If the exact sector wasn't found, the parent of the NULL leaf
2808 * will contain the closest sector.
2810 __cfqq = rb_entry(parent, struct cfq_queue, p_node);
2811 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
2814 if (blk_rq_pos(__cfqq->next_rq) < sector)
2815 node = rb_next(&__cfqq->p_node);
2817 node = rb_prev(&__cfqq->p_node);
2821 __cfqq = rb_entry(node, struct cfq_queue, p_node);
2822 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
2830 * cur_cfqq - passed in so that we don't decide that the current queue is
2831 * closely cooperating with itself.
2833 * So, basically we're assuming that that cur_cfqq has dispatched at least
2834 * one request, and that cfqd->last_position reflects a position on the disk
2835 * associated with the I/O issued by cur_cfqq. I'm not sure this is a valid
2838 static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
2839 struct cfq_queue *cur_cfqq)
2841 struct cfq_queue *cfqq;
2843 if (cfq_class_idle(cur_cfqq))
2845 if (!cfq_cfqq_sync(cur_cfqq))
2847 if (CFQQ_SEEKY(cur_cfqq))
2851 * Don't search priority tree if it's the only queue in the group.
2853 if (cur_cfqq->cfqg->nr_cfqq == 1)
2857 * We should notice if some of the queues are cooperating, eg
2858 * working closely on the same area of the disk. In that case,
2859 * we can group them together and don't waste time idling.
2861 cfqq = cfqq_close(cfqd, cur_cfqq);
2865 /* If new queue belongs to different cfq_group, don't choose it */
2866 if (cur_cfqq->cfqg != cfqq->cfqg)
2870 * It only makes sense to merge sync queues.
2872 if (!cfq_cfqq_sync(cfqq))
2874 if (CFQQ_SEEKY(cfqq))
2878 * Do not merge queues of different priority classes
2880 if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq))
2887 * Determine whether we should enforce idle window for this queue.
2890 static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2892 enum wl_class_t wl_class = cfqq_class(cfqq);
2893 struct cfq_rb_root *st = cfqq->service_tree;
2898 if (!cfqd->cfq_slice_idle)
2901 /* We never do for idle class queues. */
2902 if (wl_class == IDLE_WORKLOAD)
2905 /* We do for queues that were marked with idle window flag. */
2906 if (cfq_cfqq_idle_window(cfqq) &&
2907 !(blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag))
2911 * Otherwise, we do only if they are the last ones
2912 * in their service tree.
2914 if (st->count == 1 && cfq_cfqq_sync(cfqq) &&
2915 !cfq_io_thinktime_big(cfqd, &st->ttime, false))
2917 cfq_log_cfqq(cfqd, cfqq, "Not idling. st->count:%d", st->count);
2921 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
2923 struct cfq_queue *cfqq = cfqd->active_queue;
2924 struct cfq_rb_root *st = cfqq->service_tree;
2925 struct cfq_io_cq *cic;
2926 u64 sl, group_idle = 0;
2927 u64 now = ktime_get_ns();
2930 * SSD device without seek penalty, disable idling. But only do so
2931 * for devices that support queuing, otherwise we still have a problem
2932 * with sync vs async workloads.
2934 if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
2937 WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
2938 WARN_ON(cfq_cfqq_slice_new(cfqq));
2941 * idle is disabled, either manually or by past process history
2943 if (!cfq_should_idle(cfqd, cfqq)) {
2944 /* no queue idling. Check for group idling */
2945 if (cfqd->cfq_group_idle)
2946 group_idle = cfqd->cfq_group_idle;
2952 * still active requests from this queue, don't idle
2954 if (cfqq->dispatched)
2958 * task has exited, don't wait
2960 cic = cfqd->active_cic;
2961 if (!cic || !atomic_read(&cic->icq.ioc->active_ref))
2965 * If our average think time is larger than the remaining time
2966 * slice, then don't idle. This avoids overrunning the allotted
2969 if (sample_valid(cic->ttime.ttime_samples) &&
2970 (cfqq->slice_end - now < cic->ttime.ttime_mean)) {
2971 cfq_log_cfqq(cfqd, cfqq, "Not idling. think_time:%llu",
2972 cic->ttime.ttime_mean);
2977 * There are other queues in the group or this is the only group and
2978 * it has too big thinktime, don't do group idle.
2981 (cfqq->cfqg->nr_cfqq > 1 ||
2982 cfq_io_thinktime_big(cfqd, &st->ttime, true)))
2985 cfq_mark_cfqq_wait_request(cfqq);
2988 sl = cfqd->cfq_group_idle;
2990 sl = cfqd->cfq_slice_idle;
2992 hrtimer_start(&cfqd->idle_slice_timer, ns_to_ktime(sl),
2994 cfqg_stats_set_start_idle_time(cfqq->cfqg);
2995 cfq_log_cfqq(cfqd, cfqq, "arm_idle: %llu group_idle: %d", sl,
2996 group_idle ? 1 : 0);
3000 * Move request from internal lists to the request queue dispatch list.
3002 static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
3004 struct cfq_data *cfqd = q->elevator->elevator_data;
3005 struct cfq_queue *cfqq = RQ_CFQQ(rq);
3007 cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
3009 cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
3010 cfq_remove_request(rq);
3012 (RQ_CFQG(rq))->dispatched++;
3013 elv_dispatch_sort(q, rq);
3015 cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]++;
3016 cfqq->nr_sectors += blk_rq_sectors(rq);
3020 * return expired entry, or NULL to just start from scratch in rbtree
3022 static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
3024 struct request *rq = NULL;
3026 if (cfq_cfqq_fifo_expire(cfqq))
3029 cfq_mark_cfqq_fifo_expire(cfqq);
3031 if (list_empty(&cfqq->fifo))
3034 rq = rq_entry_fifo(cfqq->fifo.next);
3035 if (ktime_get_ns() < rq->fifo_time)
3038 cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq);
3043 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3045 const int base_rq = cfqd->cfq_slice_async_rq;
3047 WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
3049 return 2 * base_rq * (IOPRIO_BE_NR - cfqq->ioprio);
3053 * Must be called with the queue_lock held.
3055 static int cfqq_process_refs(struct cfq_queue *cfqq)
3057 int process_refs, io_refs;
3059 io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE];
3060 process_refs = cfqq->ref - io_refs;
3061 BUG_ON(process_refs < 0);
3062 return process_refs;
3065 static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq)
3067 int process_refs, new_process_refs;
3068 struct cfq_queue *__cfqq;
3071 * If there are no process references on the new_cfqq, then it is
3072 * unsafe to follow the ->new_cfqq chain as other cfqq's in the
3073 * chain may have dropped their last reference (not just their
3074 * last process reference).
3076 if (!cfqq_process_refs(new_cfqq))
3079 /* Avoid a circular list and skip interim queue merges */
3080 while ((__cfqq = new_cfqq->new_cfqq)) {
3086 process_refs = cfqq_process_refs(cfqq);
3087 new_process_refs = cfqq_process_refs(new_cfqq);
3089 * If the process for the cfqq has gone away, there is no
3090 * sense in merging the queues.
3092 if (process_refs == 0 || new_process_refs == 0)
3096 * Merge in the direction of the lesser amount of work.
3098 if (new_process_refs >= process_refs) {
3099 cfqq->new_cfqq = new_cfqq;
3100 new_cfqq->ref += process_refs;
3102 new_cfqq->new_cfqq = cfqq;
3103 cfqq->ref += new_process_refs;
3107 static enum wl_type_t cfq_choose_wl_type(struct cfq_data *cfqd,
3108 struct cfq_group *cfqg, enum wl_class_t wl_class)
3110 struct cfq_queue *queue;
3112 bool key_valid = false;
3114 enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD;
3116 for (i = 0; i <= SYNC_WORKLOAD; ++i) {
3117 /* select the one with lowest rb_key */
3118 queue = cfq_rb_first(st_for(cfqg, wl_class, i));
3120 (!key_valid || queue->rb_key < lowest_key)) {
3121 lowest_key = queue->rb_key;
3131 choose_wl_class_and_type(struct cfq_data *cfqd, struct cfq_group *cfqg)
3135 struct cfq_rb_root *st;
3137 enum wl_class_t original_class = cfqd->serving_wl_class;
3138 u64 now = ktime_get_ns();
3140 /* Choose next priority. RT > BE > IDLE */
3141 if (cfq_group_busy_queues_wl(RT_WORKLOAD, cfqd, cfqg))
3142 cfqd->serving_wl_class = RT_WORKLOAD;
3143 else if (cfq_group_busy_queues_wl(BE_WORKLOAD, cfqd, cfqg))
3144 cfqd->serving_wl_class = BE_WORKLOAD;
3146 cfqd->serving_wl_class = IDLE_WORKLOAD;
3147 cfqd->workload_expires = now + jiffies_to_nsecs(1);
3151 if (original_class != cfqd->serving_wl_class)
3155 * For RT and BE, we have to choose also the type
3156 * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
3159 st = st_for(cfqg, cfqd->serving_wl_class, cfqd->serving_wl_type);
3163 * check workload expiration, and that we still have other queues ready
3165 if (count && !(now > cfqd->workload_expires))
3169 /* otherwise select new workload type */
3170 cfqd->serving_wl_type = cfq_choose_wl_type(cfqd, cfqg,
3171 cfqd->serving_wl_class);
3172 st = st_for(cfqg, cfqd->serving_wl_class, cfqd->serving_wl_type);
3176 * the workload slice is computed as a fraction of target latency
3177 * proportional to the number of queues in that workload, over
3178 * all the queues in the same priority class
3180 group_slice = cfq_group_slice(cfqd, cfqg);
3182 slice = div_u64(group_slice * count,
3183 max_t(unsigned, cfqg->busy_queues_avg[cfqd->serving_wl_class],
3184 cfq_group_busy_queues_wl(cfqd->serving_wl_class, cfqd,
3187 if (cfqd->serving_wl_type == ASYNC_WORKLOAD) {
3191 * Async queues are currently system wide. Just taking
3192 * proportion of queues with-in same group will lead to higher
3193 * async ratio system wide as generally root group is going
3194 * to have higher weight. A more accurate thing would be to
3195 * calculate system wide asnc/sync ratio.
3197 tmp = cfqd->cfq_target_latency *
3198 cfqg_busy_async_queues(cfqd, cfqg);
3199 tmp = div_u64(tmp, cfqd->busy_queues);
3200 slice = min_t(u64, slice, tmp);
3202 /* async workload slice is scaled down according to
3203 * the sync/async slice ratio. */
3204 slice = div64_u64(slice*cfqd->cfq_slice[0], cfqd->cfq_slice[1]);
3206 /* sync workload slice is at least 2 * cfq_slice_idle */
3207 slice = max(slice, 2 * cfqd->cfq_slice_idle);
3209 slice = max_t(u64, slice, CFQ_MIN_TT);
3210 cfq_log(cfqd, "workload slice:%llu", slice);
3211 cfqd->workload_expires = now + slice;
3214 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd)
3216 struct cfq_rb_root *st = &cfqd->grp_service_tree;
3217 struct cfq_group *cfqg;
3219 if (RB_EMPTY_ROOT(&st->rb))
3221 cfqg = cfq_rb_first_group(st);
3222 update_min_vdisktime(st);
3226 static void cfq_choose_cfqg(struct cfq_data *cfqd)
3228 struct cfq_group *cfqg = cfq_get_next_cfqg(cfqd);
3229 u64 now = ktime_get_ns();
3231 cfqd->serving_group = cfqg;
3233 /* Restore the workload type data */
3234 if (cfqg->saved_wl_slice) {
3235 cfqd->workload_expires = now + cfqg->saved_wl_slice;
3236 cfqd->serving_wl_type = cfqg->saved_wl_type;
3237 cfqd->serving_wl_class = cfqg->saved_wl_class;
3239 cfqd->workload_expires = now - 1;
3241 choose_wl_class_and_type(cfqd, cfqg);
3245 * Select a queue for service. If we have a current active queue,
3246 * check whether to continue servicing it, or retrieve and set a new one.
3248 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
3250 struct cfq_queue *cfqq, *new_cfqq = NULL;
3251 u64 now = ktime_get_ns();
3253 cfqq = cfqd->active_queue;
3257 if (!cfqd->rq_queued)
3261 * We were waiting for group to get backlogged. Expire the queue
3263 if (cfq_cfqq_wait_busy(cfqq) && !RB_EMPTY_ROOT(&cfqq->sort_list))
3267 * The active queue has run out of time, expire it and select new.
3269 if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq)) {
3271 * If slice had not expired at the completion of last request
3272 * we might not have turned on wait_busy flag. Don't expire
3273 * the queue yet. Allow the group to get backlogged.
3275 * The very fact that we have used the slice, that means we
3276 * have been idling all along on this queue and it should be
3277 * ok to wait for this request to complete.
3279 if (cfqq->cfqg->nr_cfqq == 1 && RB_EMPTY_ROOT(&cfqq->sort_list)
3280 && cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
3284 goto check_group_idle;
3288 * The active queue has requests and isn't expired, allow it to
3291 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3295 * If another queue has a request waiting within our mean seek
3296 * distance, let it run. The expire code will check for close
3297 * cooperators and put the close queue at the front of the service
3298 * tree. If possible, merge the expiring queue with the new cfqq.
3300 new_cfqq = cfq_close_cooperator(cfqd, cfqq);
3302 if (!cfqq->new_cfqq)
3303 cfq_setup_merge(cfqq, new_cfqq);
3308 * No requests pending. If the active queue still has requests in
3309 * flight or is idling for a new request, allow either of these
3310 * conditions to happen (or time out) before selecting a new queue.
3312 if (hrtimer_active(&cfqd->idle_slice_timer)) {
3318 * This is a deep seek queue, but the device is much faster than
3319 * the queue can deliver, don't idle
3321 if (CFQQ_SEEKY(cfqq) && cfq_cfqq_idle_window(cfqq) &&
3322 (cfq_cfqq_slice_new(cfqq) ||
3323 (cfqq->slice_end - now > now - cfqq->slice_start))) {
3324 cfq_clear_cfqq_deep(cfqq);
3325 cfq_clear_cfqq_idle_window(cfqq);
3328 if (cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
3334 * If group idle is enabled and there are requests dispatched from
3335 * this group, wait for requests to complete.
3338 if (cfqd->cfq_group_idle && cfqq->cfqg->nr_cfqq == 1 &&
3339 cfqq->cfqg->dispatched &&
3340 !cfq_io_thinktime_big(cfqd, &cfqq->cfqg->ttime, true)) {
3346 cfq_slice_expired(cfqd, 0);
3349 * Current queue expired. Check if we have to switch to a new
3353 cfq_choose_cfqg(cfqd);
3355 cfqq = cfq_set_active_queue(cfqd, new_cfqq);
3360 static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
3364 while (cfqq->next_rq) {
3365 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
3369 BUG_ON(!list_empty(&cfqq->fifo));
3371 /* By default cfqq is not expired if it is empty. Do it explicitly */
3372 __cfq_slice_expired(cfqq->cfqd, cfqq, 0);
3377 * Drain our current requests. Used for barriers and when switching
3378 * io schedulers on-the-fly.
3380 static int cfq_forced_dispatch(struct cfq_data *cfqd)
3382 struct cfq_queue *cfqq;
3385 /* Expire the timeslice of the current active queue first */
3386 cfq_slice_expired(cfqd, 0);
3387 while ((cfqq = cfq_get_next_queue_forced(cfqd)) != NULL) {
3388 __cfq_set_active_queue(cfqd, cfqq);
3389 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
3392 BUG_ON(cfqd->busy_queues);
3394 cfq_log(cfqd, "forced_dispatch=%d", dispatched);
3398 static inline bool cfq_slice_used_soon(struct cfq_data *cfqd,
3399 struct cfq_queue *cfqq)
3401 u64 now = ktime_get_ns();
3403 /* the queue hasn't finished any request, can't estimate */
3404 if (cfq_cfqq_slice_new(cfqq))
3406 if (now + cfqd->cfq_slice_idle * cfqq->dispatched > cfqq->slice_end)
3412 static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3414 unsigned int max_dispatch;
3417 * Drain async requests before we start sync IO
3419 if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_flight[BLK_RW_ASYNC])
3423 * If this is an async queue and we have sync IO in flight, let it wait
3425 if (cfqd->rq_in_flight[BLK_RW_SYNC] && !cfq_cfqq_sync(cfqq))
3428 max_dispatch = max_t(unsigned int, cfqd->cfq_quantum / 2, 1);
3429 if (cfq_class_idle(cfqq))
3433 * Does this cfqq already have too much IO in flight?
3435 if (cfqq->dispatched >= max_dispatch) {
3436 bool promote_sync = false;
3438 * idle queue must always only have a single IO in flight
3440 if (cfq_class_idle(cfqq))
3444 * If there is only one sync queue
3445 * we can ignore async queue here and give the sync
3446 * queue no dispatch limit. The reason is a sync queue can
3447 * preempt async queue, limiting the sync queue doesn't make
3448 * sense. This is useful for aiostress test.
3450 if (cfq_cfqq_sync(cfqq) && cfqd->busy_sync_queues == 1)
3451 promote_sync = true;
3454 * We have other queues, don't allow more IO from this one
3456 if (cfqd->busy_queues > 1 && cfq_slice_used_soon(cfqd, cfqq) &&
3461 * Sole queue user, no limit
3463 if (cfqd->busy_queues == 1 || promote_sync)
3467 * Normally we start throttling cfqq when cfq_quantum/2
3468 * requests have been dispatched. But we can drive
3469 * deeper queue depths at the beginning of slice
3470 * subjected to upper limit of cfq_quantum.
3472 max_dispatch = cfqd->cfq_quantum;
3476 * Async queues must wait a bit before being allowed dispatch.
3477 * We also ramp up the dispatch depth gradually for async IO,
3478 * based on the last sync IO we serviced
3480 if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
3481 u64 last_sync = ktime_get_ns() - cfqd->last_delayed_sync;
3484 depth = div64_u64(last_sync, cfqd->cfq_slice[1]);
3485 if (!depth && !cfqq->dispatched)
3487 if (depth < max_dispatch)
3488 max_dispatch = depth;
3492 * If we're below the current max, allow a dispatch
3494 return cfqq->dispatched < max_dispatch;
3498 * Dispatch a request from cfqq, moving them to the request queue
3501 static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3505 BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
3507 if (!cfq_may_dispatch(cfqd, cfqq))
3511 * follow expired path, else get first next available
3513 rq = cfq_check_fifo(cfqq);
3518 * insert request into driver dispatch list
3520 cfq_dispatch_insert(cfqd->queue, rq);
3522 if (!cfqd->active_cic) {
3523 struct cfq_io_cq *cic = RQ_CIC(rq);
3525 atomic_long_inc(&cic->icq.ioc->refcount);
3526 cfqd->active_cic = cic;
3533 * Find the cfqq that we need to service and move a request from that to the
3536 static int cfq_dispatch_requests(struct request_queue *q, int force)
3538 struct cfq_data *cfqd = q->elevator->elevator_data;
3539 struct cfq_queue *cfqq;
3541 if (!cfqd->busy_queues)
3544 if (unlikely(force))
3545 return cfq_forced_dispatch(cfqd);
3547 cfqq = cfq_select_queue(cfqd);
3552 * Dispatch a request from this cfqq, if it is allowed
3554 if (!cfq_dispatch_request(cfqd, cfqq))
3557 cfqq->slice_dispatch++;
3558 cfq_clear_cfqq_must_dispatch(cfqq);
3561 * expire an async queue immediately if it has used up its slice. idle
3562 * queue always expire after 1 dispatch round.
3564 if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
3565 cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
3566 cfq_class_idle(cfqq))) {
3567 cfqq->slice_end = ktime_get_ns() + 1;
3568 cfq_slice_expired(cfqd, 0);
3571 cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
3576 * task holds one reference to the queue, dropped when task exits. each rq
3577 * in-flight on this queue also holds a reference, dropped when rq is freed.
3579 * Each cfq queue took a reference on the parent group. Drop it now.
3580 * queue lock must be held here.
3582 static void cfq_put_queue(struct cfq_queue *cfqq)
3584 struct cfq_data *cfqd = cfqq->cfqd;
3585 struct cfq_group *cfqg;
3587 BUG_ON(cfqq->ref <= 0);
3593 cfq_log_cfqq(cfqd, cfqq, "put_queue");
3594 BUG_ON(rb_first(&cfqq->sort_list));
3595 BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
3598 if (unlikely(cfqd->active_queue == cfqq)) {
3599 __cfq_slice_expired(cfqd, cfqq, 0);
3600 cfq_schedule_dispatch(cfqd);
3603 BUG_ON(cfq_cfqq_on_rr(cfqq));
3604 kmem_cache_free(cfq_pool, cfqq);
3608 static void cfq_put_cooperator(struct cfq_queue *cfqq)
3610 struct cfq_queue *__cfqq, *next;
3613 * If this queue was scheduled to merge with another queue, be
3614 * sure to drop the reference taken on that queue (and others in
3615 * the merge chain). See cfq_setup_merge and cfq_merge_cfqqs.
3617 __cfqq = cfqq->new_cfqq;
3619 if (__cfqq == cfqq) {
3620 WARN(1, "cfqq->new_cfqq loop detected\n");
3623 next = __cfqq->new_cfqq;
3624 cfq_put_queue(__cfqq);
3629 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3631 if (unlikely(cfqq == cfqd->active_queue)) {
3632 __cfq_slice_expired(cfqd, cfqq, 0);
3633 cfq_schedule_dispatch(cfqd);
3636 cfq_put_cooperator(cfqq);
3638 cfq_put_queue(cfqq);
3641 static void cfq_init_icq(struct io_cq *icq)
3643 struct cfq_io_cq *cic = icq_to_cic(icq);
3645 cic->ttime.last_end_request = ktime_get_ns();
3648 static void cfq_exit_icq(struct io_cq *icq)
3650 struct cfq_io_cq *cic = icq_to_cic(icq);
3651 struct cfq_data *cfqd = cic_to_cfqd(cic);
3653 if (cic_to_cfqq(cic, false)) {
3654 cfq_exit_cfqq(cfqd, cic_to_cfqq(cic, false));
3655 cic_set_cfqq(cic, NULL, false);
3658 if (cic_to_cfqq(cic, true)) {
3659 cfq_exit_cfqq(cfqd, cic_to_cfqq(cic, true));
3660 cic_set_cfqq(cic, NULL, true);
3664 static void cfq_init_prio_data(struct cfq_queue *cfqq, struct cfq_io_cq *cic)
3666 struct task_struct *tsk = current;
3669 if (!cfq_cfqq_prio_changed(cfqq))
3672 ioprio_class = IOPRIO_PRIO_CLASS(cic->ioprio);
3673 switch (ioprio_class) {
3675 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
3676 case IOPRIO_CLASS_NONE:
3678 * no prio set, inherit CPU scheduling settings
3680 cfqq->ioprio = task_nice_ioprio(tsk);
3681 cfqq->ioprio_class = task_nice_ioclass(tsk);
3683 case IOPRIO_CLASS_RT:
3684 cfqq->ioprio = IOPRIO_PRIO_DATA(cic->ioprio);
3685 cfqq->ioprio_class = IOPRIO_CLASS_RT;
3687 case IOPRIO_CLASS_BE:
3688 cfqq->ioprio = IOPRIO_PRIO_DATA(cic->ioprio);
3689 cfqq->ioprio_class = IOPRIO_CLASS_BE;
3691 case IOPRIO_CLASS_IDLE:
3692 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
3694 cfq_clear_cfqq_idle_window(cfqq);
3699 * keep track of original prio settings in case we have to temporarily
3700 * elevate the priority of this queue
3702 cfqq->org_ioprio = cfqq->ioprio;
3703 cfqq->org_ioprio_class = cfqq->ioprio_class;
3704 cfq_clear_cfqq_prio_changed(cfqq);
3707 static void check_ioprio_changed(struct cfq_io_cq *cic, struct bio *bio)
3709 int ioprio = cic->icq.ioc->ioprio;
3710 struct cfq_data *cfqd = cic_to_cfqd(cic);
3711 struct cfq_queue *cfqq;
3714 * Check whether ioprio has changed. The condition may trigger
3715 * spuriously on a newly created cic but there's no harm.
3717 if (unlikely(!cfqd) || likely(cic->ioprio == ioprio))
3720 cfqq = cic_to_cfqq(cic, false);
3722 cfq_put_queue(cfqq);
3723 cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic, bio);
3724 cic_set_cfqq(cic, cfqq, false);
3727 cfqq = cic_to_cfqq(cic, true);
3729 cfq_mark_cfqq_prio_changed(cfqq);
3731 cic->ioprio = ioprio;
3734 static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3735 pid_t pid, bool is_sync)
3737 RB_CLEAR_NODE(&cfqq->rb_node);
3738 RB_CLEAR_NODE(&cfqq->p_node);
3739 INIT_LIST_HEAD(&cfqq->fifo);
3744 cfq_mark_cfqq_prio_changed(cfqq);
3747 if (!cfq_class_idle(cfqq))
3748 cfq_mark_cfqq_idle_window(cfqq);
3749 cfq_mark_cfqq_sync(cfqq);
3754 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3755 static void check_blkcg_changed(struct cfq_io_cq *cic, struct bio *bio)
3757 struct cfq_data *cfqd = cic_to_cfqd(cic);
3758 struct cfq_queue *cfqq;
3762 serial_nr = bio_blkcg(bio)->css.serial_nr;
3766 * Check whether blkcg has changed. The condition may trigger
3767 * spuriously on a newly created cic but there's no harm.
3769 if (unlikely(!cfqd) || likely(cic->blkcg_serial_nr == serial_nr))
3773 * Drop reference to queues. New queues will be assigned in new
3774 * group upon arrival of fresh requests.
3776 cfqq = cic_to_cfqq(cic, false);
3778 cfq_log_cfqq(cfqd, cfqq, "changed cgroup");
3779 cic_set_cfqq(cic, NULL, false);
3780 cfq_put_queue(cfqq);
3783 cfqq = cic_to_cfqq(cic, true);
3785 cfq_log_cfqq(cfqd, cfqq, "changed cgroup");
3786 cic_set_cfqq(cic, NULL, true);
3787 cfq_put_queue(cfqq);
3790 cic->blkcg_serial_nr = serial_nr;
3793 static inline void check_blkcg_changed(struct cfq_io_cq *cic, struct bio *bio) { }
3794 #endif /* CONFIG_CFQ_GROUP_IOSCHED */
3796 static struct cfq_queue **
3797 cfq_async_queue_prio(struct cfq_group *cfqg, int ioprio_class, int ioprio)
3799 switch (ioprio_class) {
3800 case IOPRIO_CLASS_RT:
3801 return &cfqg->async_cfqq[0][ioprio];
3802 case IOPRIO_CLASS_NONE:
3803 ioprio = IOPRIO_NORM;
3805 case IOPRIO_CLASS_BE:
3806 return &cfqg->async_cfqq[1][ioprio];
3807 case IOPRIO_CLASS_IDLE:
3808 return &cfqg->async_idle_cfqq;
3814 static struct cfq_queue *
3815 cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct cfq_io_cq *cic,
3818 int ioprio_class = IOPRIO_PRIO_CLASS(cic->ioprio);
3819 int ioprio = IOPRIO_PRIO_DATA(cic->ioprio);
3820 struct cfq_queue **async_cfqq = NULL;
3821 struct cfq_queue *cfqq;
3822 struct cfq_group *cfqg;
3825 cfqg = cfq_lookup_cfqg(cfqd, bio_blkcg(bio));
3827 cfqq = &cfqd->oom_cfqq;
3832 if (!ioprio_valid(cic->ioprio)) {
3833 struct task_struct *tsk = current;
3834 ioprio = task_nice_ioprio(tsk);
3835 ioprio_class = task_nice_ioclass(tsk);
3837 async_cfqq = cfq_async_queue_prio(cfqg, ioprio_class, ioprio);
3843 cfqq = kmem_cache_alloc_node(cfq_pool, GFP_NOWAIT | __GFP_ZERO,
3846 cfqq = &cfqd->oom_cfqq;
3850 cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
3851 cfq_init_prio_data(cfqq, cic);
3852 cfq_link_cfqq_cfqg(cfqq, cfqg);
3853 cfq_log_cfqq(cfqd, cfqq, "alloced");
3856 /* a new async queue is created, pin and remember */
3867 __cfq_update_io_thinktime(struct cfq_ttime *ttime, u64 slice_idle)
3869 u64 elapsed = ktime_get_ns() - ttime->last_end_request;
3870 elapsed = min(elapsed, 2UL * slice_idle);
3872 ttime->ttime_samples = (7*ttime->ttime_samples + 256) / 8;
3873 ttime->ttime_total = div_u64(7*ttime->ttime_total + 256*elapsed, 8);
3874 ttime->ttime_mean = div64_ul(ttime->ttime_total + 128,
3875 ttime->ttime_samples);
3879 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3880 struct cfq_io_cq *cic)
3882 if (cfq_cfqq_sync(cfqq)) {
3883 __cfq_update_io_thinktime(&cic->ttime, cfqd->cfq_slice_idle);
3884 __cfq_update_io_thinktime(&cfqq->service_tree->ttime,
3885 cfqd->cfq_slice_idle);
3887 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3888 __cfq_update_io_thinktime(&cfqq->cfqg->ttime, cfqd->cfq_group_idle);
3893 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3897 sector_t n_sec = blk_rq_sectors(rq);
3898 if (cfqq->last_request_pos) {
3899 if (cfqq->last_request_pos < blk_rq_pos(rq))
3900 sdist = blk_rq_pos(rq) - cfqq->last_request_pos;
3902 sdist = cfqq->last_request_pos - blk_rq_pos(rq);
3905 cfqq->seek_history <<= 1;
3906 if (blk_queue_nonrot(cfqd->queue))
3907 cfqq->seek_history |= (n_sec < CFQQ_SECT_THR_NONROT);
3909 cfqq->seek_history |= (sdist > CFQQ_SEEK_THR);
3913 * Disable idle window if the process thinks too long or seeks so much that
3917 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3918 struct cfq_io_cq *cic)
3920 int old_idle, enable_idle;
3923 * Don't idle for async or idle io prio class
3925 if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
3928 enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
3930 if (cfqq->queued[0] + cfqq->queued[1] >= 4)
3931 cfq_mark_cfqq_deep(cfqq);
3933 if (cfqq->next_rq && (cfqq->next_rq->cmd_flags & REQ_NOIDLE))
3935 else if (!atomic_read(&cic->icq.ioc->active_ref) ||
3936 !cfqd->cfq_slice_idle ||
3937 (!cfq_cfqq_deep(cfqq) && CFQQ_SEEKY(cfqq)))
3939 else if (sample_valid(cic->ttime.ttime_samples)) {
3940 if (cic->ttime.ttime_mean > cfqd->cfq_slice_idle)
3946 if (old_idle != enable_idle) {
3947 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
3949 cfq_mark_cfqq_idle_window(cfqq);
3951 cfq_clear_cfqq_idle_window(cfqq);
3956 * Check if new_cfqq should preempt the currently active queue. Return 0 for
3957 * no or if we aren't sure, a 1 will cause a preempt.
3960 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
3963 struct cfq_queue *cfqq;
3965 cfqq = cfqd->active_queue;
3969 if (cfq_class_idle(new_cfqq))
3972 if (cfq_class_idle(cfqq))
3976 * Don't allow a non-RT request to preempt an ongoing RT cfqq timeslice.
3978 if (cfq_class_rt(cfqq) && !cfq_class_rt(new_cfqq))
3982 * if the new request is sync, but the currently running queue is
3983 * not, let the sync request have priority.
3985 if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
3989 * Treat ancestors of current cgroup the same way as current cgroup.
3990 * For anybody else we disallow preemption to guarantee service
3991 * fairness among cgroups.
3993 if (!cfqg_is_descendant(cfqq->cfqg, new_cfqq->cfqg))
3996 if (cfq_slice_used(cfqq))
4000 * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
4002 if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
4005 WARN_ON_ONCE(cfqq->ioprio_class != new_cfqq->ioprio_class);
4006 /* Allow preemption only if we are idling on sync-noidle tree */
4007 if (cfqd->serving_wl_type == SYNC_NOIDLE_WORKLOAD &&
4008 cfqq_type(new_cfqq) == SYNC_NOIDLE_WORKLOAD &&
4009 RB_EMPTY_ROOT(&cfqq->sort_list))
4013 * So both queues are sync. Let the new request get disk time if
4014 * it's a metadata request and the current queue is doing regular IO.
4016 if ((rq->cmd_flags & REQ_PRIO) && !cfqq->prio_pending)
4019 /* An idle queue should not be idle now for some reason */
4020 if (RB_EMPTY_ROOT(&cfqq->sort_list) && !cfq_should_idle(cfqd, cfqq))
4023 if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
4027 * if this request is as-good as one we would expect from the
4028 * current cfqq, let it preempt
4030 if (cfq_rq_close(cfqd, cfqq, rq))
4037 * cfqq preempts the active queue. if we allowed preempt with no slice left,
4038 * let it have half of its nominal slice.
4040 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
4042 enum wl_type_t old_type = cfqq_type(cfqd->active_queue);
4044 cfq_log_cfqq(cfqd, cfqq, "preempt");
4045 cfq_slice_expired(cfqd, 1);
4048 * workload type is changed, don't save slice, otherwise preempt
4051 if (old_type != cfqq_type(cfqq))
4052 cfqq->cfqg->saved_wl_slice = 0;
4055 * Put the new queue at the front of the of the current list,
4056 * so we know that it will be selected next.
4058 BUG_ON(!cfq_cfqq_on_rr(cfqq));
4060 cfq_service_tree_add(cfqd, cfqq, 1);
4062 cfqq->slice_end = 0;
4063 cfq_mark_cfqq_slice_new(cfqq);
4067 * Called when a new fs request (rq) is added (to cfqq). Check if there's
4068 * something we should do about it
4071 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
4074 struct cfq_io_cq *cic = RQ_CIC(rq);
4077 if (rq->cmd_flags & REQ_PRIO)
4078 cfqq->prio_pending++;
4080 cfq_update_io_thinktime(cfqd, cfqq, cic);
4081 cfq_update_io_seektime(cfqd, cfqq, rq);
4082 cfq_update_idle_window(cfqd, cfqq, cic);
4084 cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
4086 if (cfqq == cfqd->active_queue) {
4088 * Remember that we saw a request from this process, but
4089 * don't start queuing just yet. Otherwise we risk seeing lots
4090 * of tiny requests, because we disrupt the normal plugging
4091 * and merging. If the request is already larger than a single
4092 * page, let it rip immediately. For that case we assume that
4093 * merging is already done. Ditto for a busy system that
4094 * has other work pending, don't risk delaying until the
4095 * idle timer unplug to continue working.
4097 if (cfq_cfqq_wait_request(cfqq)) {
4098 if (blk_rq_bytes(rq) > PAGE_SIZE ||
4099 cfqd->busy_queues > 1) {
4100 cfq_del_timer(cfqd, cfqq);
4101 cfq_clear_cfqq_wait_request(cfqq);
4102 __blk_run_queue(cfqd->queue);
4104 cfqg_stats_update_idle_time(cfqq->cfqg);
4105 cfq_mark_cfqq_must_dispatch(cfqq);
4108 } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
4110 * not the active queue - expire current slice if it is
4111 * idle and has expired it's mean thinktime or this new queue
4112 * has some old slice time left and is of higher priority or
4113 * this new queue is RT and the current one is BE
4115 cfq_preempt_queue(cfqd, cfqq);
4116 __blk_run_queue(cfqd->queue);
4120 static void cfq_insert_request(struct request_queue *q, struct request *rq)
4122 struct cfq_data *cfqd = q->elevator->elevator_data;
4123 struct cfq_queue *cfqq = RQ_CFQQ(rq);
4125 cfq_log_cfqq(cfqd, cfqq, "insert_request");
4126 cfq_init_prio_data(cfqq, RQ_CIC(rq));
4128 rq->fifo_time = ktime_get_ns() + cfqd->cfq_fifo_expire[rq_is_sync(rq)];
4129 list_add_tail(&rq->queuelist, &cfqq->fifo);
4131 cfqg_stats_update_io_add(RQ_CFQG(rq), cfqd->serving_group, req_op(rq),
4133 cfq_rq_enqueued(cfqd, cfqq, rq);
4137 * Update hw_tag based on peak queue depth over 50 samples under
4140 static void cfq_update_hw_tag(struct cfq_data *cfqd)
4142 struct cfq_queue *cfqq = cfqd->active_queue;
4144 if (cfqd->rq_in_driver > cfqd->hw_tag_est_depth)
4145 cfqd->hw_tag_est_depth = cfqd->rq_in_driver;
4147 if (cfqd->hw_tag == 1)
4150 if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
4151 cfqd->rq_in_driver <= CFQ_HW_QUEUE_MIN)
4155 * If active queue hasn't enough requests and can idle, cfq might not
4156 * dispatch sufficient requests to hardware. Don't zero hw_tag in this
4159 if (cfqq && cfq_cfqq_idle_window(cfqq) &&
4160 cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] <
4161 CFQ_HW_QUEUE_MIN && cfqd->rq_in_driver < CFQ_HW_QUEUE_MIN)
4164 if (cfqd->hw_tag_samples++ < 50)
4167 if (cfqd->hw_tag_est_depth >= CFQ_HW_QUEUE_MIN)
4173 static bool cfq_should_wait_busy(struct cfq_data *cfqd, struct cfq_queue *cfqq)
4175 struct cfq_io_cq *cic = cfqd->active_cic;
4176 u64 now = ktime_get_ns();
4178 /* If the queue already has requests, don't wait */
4179 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
4182 /* If there are other queues in the group, don't wait */
4183 if (cfqq->cfqg->nr_cfqq > 1)
4186 /* the only queue in the group, but think time is big */
4187 if (cfq_io_thinktime_big(cfqd, &cfqq->cfqg->ttime, true))
4190 if (cfq_slice_used(cfqq))
4193 /* if slice left is less than think time, wait busy */
4194 if (cic && sample_valid(cic->ttime.ttime_samples)
4195 && (cfqq->slice_end - now < cic->ttime.ttime_mean))
4199 * If think times is less than a jiffy than ttime_mean=0 and above
4200 * will not be true. It might happen that slice has not expired yet
4201 * but will expire soon (4-5 ns) during select_queue(). To cover the
4202 * case where think time is less than a jiffy, mark the queue wait
4203 * busy if only 1 jiffy is left in the slice.
4205 if (cfqq->slice_end - now <= jiffies_to_nsecs(1))
4211 static void cfq_completed_request(struct request_queue *q, struct request *rq)
4213 struct cfq_queue *cfqq = RQ_CFQQ(rq);
4214 struct cfq_data *cfqd = cfqq->cfqd;
4215 const int sync = rq_is_sync(rq);
4216 u64 now = ktime_get_ns();
4218 cfq_log_cfqq(cfqd, cfqq, "complete rqnoidle %d",
4219 !!(rq->cmd_flags & REQ_NOIDLE));
4221 cfq_update_hw_tag(cfqd);
4223 WARN_ON(!cfqd->rq_in_driver);
4224 WARN_ON(!cfqq->dispatched);
4225 cfqd->rq_in_driver--;
4227 (RQ_CFQG(rq))->dispatched--;
4228 cfqg_stats_update_completion(cfqq->cfqg, rq_start_time_ns(rq),
4229 rq_io_start_time_ns(rq), req_op(rq),
4232 cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]--;
4235 struct cfq_rb_root *st;
4237 RQ_CIC(rq)->ttime.last_end_request = now;
4239 if (cfq_cfqq_on_rr(cfqq))
4240 st = cfqq->service_tree;
4242 st = st_for(cfqq->cfqg, cfqq_class(cfqq),
4245 st->ttime.last_end_request = now;
4247 * We have to do this check in jiffies since start_time is in
4248 * jiffies and it is not trivial to convert to ns. If
4249 * cfq_fifo_expire[1] ever comes close to 1 jiffie, this test
4250 * will become problematic but so far we are fine (the default
4253 if (!time_after(rq->start_time +
4254 nsecs_to_jiffies(cfqd->cfq_fifo_expire[1]),
4256 cfqd->last_delayed_sync = now;
4259 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4260 cfqq->cfqg->ttime.last_end_request = now;
4264 * If this is the active queue, check if it needs to be expired,
4265 * or if we want to idle in case it has no pending requests.
4267 if (cfqd->active_queue == cfqq) {
4268 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
4270 if (cfq_cfqq_slice_new(cfqq)) {
4271 cfq_set_prio_slice(cfqd, cfqq);
4272 cfq_clear_cfqq_slice_new(cfqq);
4276 * Should we wait for next request to come in before we expire
4279 if (cfq_should_wait_busy(cfqd, cfqq)) {
4280 u64 extend_sl = cfqd->cfq_slice_idle;
4281 if (!cfqd->cfq_slice_idle)
4282 extend_sl = cfqd->cfq_group_idle;
4283 cfqq->slice_end = now + extend_sl;
4284 cfq_mark_cfqq_wait_busy(cfqq);
4285 cfq_log_cfqq(cfqd, cfqq, "will busy wait");
4289 * Idling is not enabled on:
4291 * - idle-priority queues
4293 * - queues with still some requests queued
4294 * - when there is a close cooperator
4296 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
4297 cfq_slice_expired(cfqd, 1);
4298 else if (sync && cfqq_empty &&
4299 !cfq_close_cooperator(cfqd, cfqq)) {
4300 cfq_arm_slice_timer(cfqd);
4304 if (!cfqd->rq_in_driver)
4305 cfq_schedule_dispatch(cfqd);
4308 static void cfqq_boost_on_prio(struct cfq_queue *cfqq, int op_flags)
4311 * If REQ_PRIO is set, boost class and prio level, if it's below
4312 * BE/NORM. If prio is not set, restore the potentially boosted
4315 if (!(op_flags & REQ_PRIO)) {
4316 cfqq->ioprio_class = cfqq->org_ioprio_class;
4317 cfqq->ioprio = cfqq->org_ioprio;
4319 if (cfq_class_idle(cfqq))
4320 cfqq->ioprio_class = IOPRIO_CLASS_BE;
4321 if (cfqq->ioprio > IOPRIO_NORM)
4322 cfqq->ioprio = IOPRIO_NORM;
4326 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
4328 if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
4329 cfq_mark_cfqq_must_alloc_slice(cfqq);
4330 return ELV_MQUEUE_MUST;
4333 return ELV_MQUEUE_MAY;
4336 static int cfq_may_queue(struct request_queue *q, int op, int op_flags)
4338 struct cfq_data *cfqd = q->elevator->elevator_data;
4339 struct task_struct *tsk = current;
4340 struct cfq_io_cq *cic;
4341 struct cfq_queue *cfqq;
4344 * don't force setup of a queue from here, as a call to may_queue
4345 * does not necessarily imply that a request actually will be queued.
4346 * so just lookup a possibly existing queue, or return 'may queue'
4349 cic = cfq_cic_lookup(cfqd, tsk->io_context);
4351 return ELV_MQUEUE_MAY;
4353 cfqq = cic_to_cfqq(cic, rw_is_sync(op, op_flags));
4355 cfq_init_prio_data(cfqq, cic);
4356 cfqq_boost_on_prio(cfqq, op_flags);
4358 return __cfq_may_queue(cfqq);
4361 return ELV_MQUEUE_MAY;
4365 * queue lock held here
4367 static void cfq_put_request(struct request *rq)
4369 struct cfq_queue *cfqq = RQ_CFQQ(rq);
4372 const int rw = rq_data_dir(rq);
4374 BUG_ON(!cfqq->allocated[rw]);
4375 cfqq->allocated[rw]--;
4377 /* Put down rq reference on cfqg */
4378 cfqg_put(RQ_CFQG(rq));
4379 rq->elv.priv[0] = NULL;
4380 rq->elv.priv[1] = NULL;
4382 cfq_put_queue(cfqq);
4386 static struct cfq_queue *
4387 cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_cq *cic,
4388 struct cfq_queue *cfqq)
4390 cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq);
4391 cic_set_cfqq(cic, cfqq->new_cfqq, 1);
4392 cfq_mark_cfqq_coop(cfqq->new_cfqq);
4393 cfq_put_queue(cfqq);
4394 return cic_to_cfqq(cic, 1);
4398 * Returns NULL if a new cfqq should be allocated, or the old cfqq if this
4399 * was the last process referring to said cfqq.
4401 static struct cfq_queue *
4402 split_cfqq(struct cfq_io_cq *cic, struct cfq_queue *cfqq)
4404 if (cfqq_process_refs(cfqq) == 1) {
4405 cfqq->pid = current->pid;
4406 cfq_clear_cfqq_coop(cfqq);
4407 cfq_clear_cfqq_split_coop(cfqq);
4411 cic_set_cfqq(cic, NULL, 1);
4413 cfq_put_cooperator(cfqq);
4415 cfq_put_queue(cfqq);
4419 * Allocate cfq data structures associated with this request.
4422 cfq_set_request(struct request_queue *q, struct request *rq, struct bio *bio,
4425 struct cfq_data *cfqd = q->elevator->elevator_data;
4426 struct cfq_io_cq *cic = icq_to_cic(rq->elv.icq);
4427 const int rw = rq_data_dir(rq);
4428 const bool is_sync = rq_is_sync(rq);
4429 struct cfq_queue *cfqq;
4431 spin_lock_irq(q->queue_lock);
4433 check_ioprio_changed(cic, bio);
4434 check_blkcg_changed(cic, bio);
4436 cfqq = cic_to_cfqq(cic, is_sync);
4437 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
4439 cfq_put_queue(cfqq);
4440 cfqq = cfq_get_queue(cfqd, is_sync, cic, bio);
4441 cic_set_cfqq(cic, cfqq, is_sync);
4444 * If the queue was seeky for too long, break it apart.
4446 if (cfq_cfqq_coop(cfqq) && cfq_cfqq_split_coop(cfqq)) {
4447 cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq");
4448 cfqq = split_cfqq(cic, cfqq);
4454 * Check to see if this queue is scheduled to merge with
4455 * another, closely cooperating queue. The merging of
4456 * queues happens here as it must be done in process context.
4457 * The reference on new_cfqq was taken in merge_cfqqs.
4460 cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq);
4463 cfqq->allocated[rw]++;
4466 cfqg_get(cfqq->cfqg);
4467 rq->elv.priv[0] = cfqq;
4468 rq->elv.priv[1] = cfqq->cfqg;
4469 spin_unlock_irq(q->queue_lock);
4473 static void cfq_kick_queue(struct work_struct *work)
4475 struct cfq_data *cfqd =
4476 container_of(work, struct cfq_data, unplug_work);
4477 struct request_queue *q = cfqd->queue;
4479 spin_lock_irq(q->queue_lock);
4480 __blk_run_queue(cfqd->queue);
4481 spin_unlock_irq(q->queue_lock);
4485 * Timer running if the active_queue is currently idling inside its time slice
4487 static enum hrtimer_restart cfq_idle_slice_timer(struct hrtimer *timer)
4489 struct cfq_data *cfqd = container_of(timer, struct cfq_data,
4491 struct cfq_queue *cfqq;
4492 unsigned long flags;
4495 cfq_log(cfqd, "idle timer fired");
4497 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
4499 cfqq = cfqd->active_queue;
4504 * We saw a request before the queue expired, let it through
4506 if (cfq_cfqq_must_dispatch(cfqq))
4512 if (cfq_slice_used(cfqq))
4516 * only expire and reinvoke request handler, if there are
4517 * other queues with pending requests
4519 if (!cfqd->busy_queues)
4523 * not expired and it has a request pending, let it dispatch
4525 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
4529 * Queue depth flag is reset only when the idle didn't succeed
4531 cfq_clear_cfqq_deep(cfqq);
4534 cfq_slice_expired(cfqd, timed_out);
4536 cfq_schedule_dispatch(cfqd);
4538 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
4539 return HRTIMER_NORESTART;
4542 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
4544 hrtimer_cancel(&cfqd->idle_slice_timer);
4545 cancel_work_sync(&cfqd->unplug_work);
4548 static void cfq_exit_queue(struct elevator_queue *e)
4550 struct cfq_data *cfqd = e->elevator_data;
4551 struct request_queue *q = cfqd->queue;
4553 cfq_shutdown_timer_wq(cfqd);
4555 spin_lock_irq(q->queue_lock);
4557 if (cfqd->active_queue)
4558 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
4560 spin_unlock_irq(q->queue_lock);
4562 cfq_shutdown_timer_wq(cfqd);
4564 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4565 blkcg_deactivate_policy(q, &blkcg_policy_cfq);
4567 kfree(cfqd->root_group);
4572 static int cfq_init_queue(struct request_queue *q, struct elevator_type *e)
4574 struct cfq_data *cfqd;
4575 struct blkcg_gq *blkg __maybe_unused;
4577 struct elevator_queue *eq;
4579 eq = elevator_alloc(q, e);
4583 cfqd = kzalloc_node(sizeof(*cfqd), GFP_KERNEL, q->node);
4585 kobject_put(&eq->kobj);
4588 eq->elevator_data = cfqd;
4591 spin_lock_irq(q->queue_lock);
4593 spin_unlock_irq(q->queue_lock);
4595 /* Init root service tree */
4596 cfqd->grp_service_tree = CFQ_RB_ROOT;
4598 /* Init root group and prefer root group over other groups by default */
4599 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4600 ret = blkcg_activate_policy(q, &blkcg_policy_cfq);
4604 cfqd->root_group = blkg_to_cfqg(q->root_blkg);
4607 cfqd->root_group = kzalloc_node(sizeof(*cfqd->root_group),
4608 GFP_KERNEL, cfqd->queue->node);
4609 if (!cfqd->root_group)
4612 cfq_init_cfqg_base(cfqd->root_group);
4613 cfqd->root_group->weight = 2 * CFQ_WEIGHT_LEGACY_DFL;
4614 cfqd->root_group->leaf_weight = 2 * CFQ_WEIGHT_LEGACY_DFL;
4618 * Not strictly needed (since RB_ROOT just clears the node and we
4619 * zeroed cfqd on alloc), but better be safe in case someone decides
4620 * to add magic to the rb code
4622 for (i = 0; i < CFQ_PRIO_LISTS; i++)
4623 cfqd->prio_trees[i] = RB_ROOT;
4626 * Our fallback cfqq if cfq_get_queue() runs into OOM issues.
4627 * Grab a permanent reference to it, so that the normal code flow
4628 * will not attempt to free it. oom_cfqq is linked to root_group
4629 * but shouldn't hold a reference as it'll never be unlinked. Lose
4630 * the reference from linking right away.
4632 cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
4633 cfqd->oom_cfqq.ref++;
4635 spin_lock_irq(q->queue_lock);
4636 cfq_link_cfqq_cfqg(&cfqd->oom_cfqq, cfqd->root_group);
4637 cfqg_put(cfqd->root_group);
4638 spin_unlock_irq(q->queue_lock);
4640 hrtimer_init(&cfqd->idle_slice_timer, CLOCK_MONOTONIC,
4642 cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
4644 INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
4646 cfqd->cfq_quantum = cfq_quantum;
4647 cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
4648 cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
4649 cfqd->cfq_back_max = cfq_back_max;
4650 cfqd->cfq_back_penalty = cfq_back_penalty;
4651 cfqd->cfq_slice[0] = cfq_slice_async;
4652 cfqd->cfq_slice[1] = cfq_slice_sync;
4653 cfqd->cfq_target_latency = cfq_target_latency;
4654 cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
4655 cfqd->cfq_slice_idle = cfq_slice_idle;
4656 cfqd->cfq_group_idle = cfq_group_idle;
4657 cfqd->cfq_latency = 1;
4660 * we optimistically start assuming sync ops weren't delayed in last
4661 * second, in order to have larger depth for async operations.
4663 cfqd->last_delayed_sync = ktime_get_ns() - NSEC_PER_SEC;
4668 kobject_put(&eq->kobj);
4672 static void cfq_registered_queue(struct request_queue *q)
4674 struct elevator_queue *e = q->elevator;
4675 struct cfq_data *cfqd = e->elevator_data;
4678 * Default to IOPS mode with no idling for SSDs
4680 if (blk_queue_nonrot(q))
4681 cfqd->cfq_slice_idle = 0;
4685 * sysfs parts below -->
4688 cfq_var_show(unsigned int var, char *page)
4690 return sprintf(page, "%u\n", var);
4694 cfq_var_store(unsigned int *var, const char *page, size_t count)
4696 char *p = (char *) page;
4698 *var = simple_strtoul(p, &p, 10);
4702 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
4703 static ssize_t __FUNC(struct elevator_queue *e, char *page) \
4705 struct cfq_data *cfqd = e->elevator_data; \
4706 u64 __data = __VAR; \
4708 __data = div_u64(__data, NSEC_PER_MSEC); \
4709 return cfq_var_show(__data, (page)); \
4711 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
4712 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
4713 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
4714 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
4715 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
4716 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
4717 SHOW_FUNCTION(cfq_group_idle_show, cfqd->cfq_group_idle, 1);
4718 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
4719 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
4720 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
4721 SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
4722 SHOW_FUNCTION(cfq_target_latency_show, cfqd->cfq_target_latency, 1);
4723 #undef SHOW_FUNCTION
4725 #define USEC_SHOW_FUNCTION(__FUNC, __VAR) \
4726 static ssize_t __FUNC(struct elevator_queue *e, char *page) \
4728 struct cfq_data *cfqd = e->elevator_data; \
4729 u64 __data = __VAR; \
4730 __data = div_u64(__data, NSEC_PER_USEC); \
4731 return cfq_var_show(__data, (page)); \
4733 USEC_SHOW_FUNCTION(cfq_slice_idle_us_show, cfqd->cfq_slice_idle);
4734 USEC_SHOW_FUNCTION(cfq_group_idle_us_show, cfqd->cfq_group_idle);
4735 USEC_SHOW_FUNCTION(cfq_slice_sync_us_show, cfqd->cfq_slice[1]);
4736 USEC_SHOW_FUNCTION(cfq_slice_async_us_show, cfqd->cfq_slice[0]);
4737 USEC_SHOW_FUNCTION(cfq_target_latency_us_show, cfqd->cfq_target_latency);
4738 #undef USEC_SHOW_FUNCTION
4740 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
4741 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
4743 struct cfq_data *cfqd = e->elevator_data; \
4744 unsigned int __data; \
4745 int ret = cfq_var_store(&__data, (page), count); \
4746 if (__data < (MIN)) \
4748 else if (__data > (MAX)) \
4751 *(__PTR) = (u64)__data * NSEC_PER_MSEC; \
4753 *(__PTR) = __data; \
4756 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
4757 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
4759 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
4761 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
4762 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
4764 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
4765 STORE_FUNCTION(cfq_group_idle_store, &cfqd->cfq_group_idle, 0, UINT_MAX, 1);
4766 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
4767 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
4768 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
4770 STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
4771 STORE_FUNCTION(cfq_target_latency_store, &cfqd->cfq_target_latency, 1, UINT_MAX, 1);
4772 #undef STORE_FUNCTION
4774 #define USEC_STORE_FUNCTION(__FUNC, __PTR, MIN, MAX) \
4775 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
4777 struct cfq_data *cfqd = e->elevator_data; \
4778 unsigned int __data; \
4779 int ret = cfq_var_store(&__data, (page), count); \
4780 if (__data < (MIN)) \
4782 else if (__data > (MAX)) \
4784 *(__PTR) = (u64)__data * NSEC_PER_USEC; \
4787 USEC_STORE_FUNCTION(cfq_slice_idle_us_store, &cfqd->cfq_slice_idle, 0, UINT_MAX);
4788 USEC_STORE_FUNCTION(cfq_group_idle_us_store, &cfqd->cfq_group_idle, 0, UINT_MAX);
4789 USEC_STORE_FUNCTION(cfq_slice_sync_us_store, &cfqd->cfq_slice[1], 1, UINT_MAX);
4790 USEC_STORE_FUNCTION(cfq_slice_async_us_store, &cfqd->cfq_slice[0], 1, UINT_MAX);
4791 USEC_STORE_FUNCTION(cfq_target_latency_us_store, &cfqd->cfq_target_latency, 1, UINT_MAX);
4792 #undef USEC_STORE_FUNCTION
4794 #define CFQ_ATTR(name) \
4795 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
4797 static struct elv_fs_entry cfq_attrs[] = {
4799 CFQ_ATTR(fifo_expire_sync),
4800 CFQ_ATTR(fifo_expire_async),
4801 CFQ_ATTR(back_seek_max),
4802 CFQ_ATTR(back_seek_penalty),
4803 CFQ_ATTR(slice_sync),
4804 CFQ_ATTR(slice_sync_us),
4805 CFQ_ATTR(slice_async),
4806 CFQ_ATTR(slice_async_us),
4807 CFQ_ATTR(slice_async_rq),
4808 CFQ_ATTR(slice_idle),
4809 CFQ_ATTR(slice_idle_us),
4810 CFQ_ATTR(group_idle),
4811 CFQ_ATTR(group_idle_us),
4812 CFQ_ATTR(low_latency),
4813 CFQ_ATTR(target_latency),
4814 CFQ_ATTR(target_latency_us),
4818 static struct elevator_type iosched_cfq = {
4820 .elevator_merge_fn = cfq_merge,
4821 .elevator_merged_fn = cfq_merged_request,
4822 .elevator_merge_req_fn = cfq_merged_requests,
4823 .elevator_allow_merge_fn = cfq_allow_merge,
4824 .elevator_bio_merged_fn = cfq_bio_merged,
4825 .elevator_dispatch_fn = cfq_dispatch_requests,
4826 .elevator_add_req_fn = cfq_insert_request,
4827 .elevator_activate_req_fn = cfq_activate_request,
4828 .elevator_deactivate_req_fn = cfq_deactivate_request,
4829 .elevator_completed_req_fn = cfq_completed_request,
4830 .elevator_former_req_fn = elv_rb_former_request,
4831 .elevator_latter_req_fn = elv_rb_latter_request,
4832 .elevator_init_icq_fn = cfq_init_icq,
4833 .elevator_exit_icq_fn = cfq_exit_icq,
4834 .elevator_set_req_fn = cfq_set_request,
4835 .elevator_put_req_fn = cfq_put_request,
4836 .elevator_may_queue_fn = cfq_may_queue,
4837 .elevator_init_fn = cfq_init_queue,
4838 .elevator_exit_fn = cfq_exit_queue,
4839 .elevator_registered_fn = cfq_registered_queue,
4841 .icq_size = sizeof(struct cfq_io_cq),
4842 .icq_align = __alignof__(struct cfq_io_cq),
4843 .elevator_attrs = cfq_attrs,
4844 .elevator_name = "cfq",
4845 .elevator_owner = THIS_MODULE,
4848 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4849 static struct blkcg_policy blkcg_policy_cfq = {
4850 .dfl_cftypes = cfq_blkcg_files,
4851 .legacy_cftypes = cfq_blkcg_legacy_files,
4853 .cpd_alloc_fn = cfq_cpd_alloc,
4854 .cpd_init_fn = cfq_cpd_init,
4855 .cpd_free_fn = cfq_cpd_free,
4856 .cpd_bind_fn = cfq_cpd_bind,
4858 .pd_alloc_fn = cfq_pd_alloc,
4859 .pd_init_fn = cfq_pd_init,
4860 .pd_offline_fn = cfq_pd_offline,
4861 .pd_free_fn = cfq_pd_free,
4862 .pd_reset_stats_fn = cfq_pd_reset_stats,
4866 static int __init cfq_init(void)
4870 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4871 ret = blkcg_policy_register(&blkcg_policy_cfq);
4879 cfq_pool = KMEM_CACHE(cfq_queue, 0);
4883 ret = elv_register(&iosched_cfq);
4890 kmem_cache_destroy(cfq_pool);
4892 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4893 blkcg_policy_unregister(&blkcg_policy_cfq);
4898 static void __exit cfq_exit(void)
4900 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4901 blkcg_policy_unregister(&blkcg_policy_cfq);
4903 elv_unregister(&iosched_cfq);
4904 kmem_cache_destroy(cfq_pool);
4907 module_init(cfq_init);
4908 module_exit(cfq_exit);
4910 MODULE_AUTHOR("Jens Axboe");
4911 MODULE_LICENSE("GPL");
4912 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");