2 * CFQ, or complete fairness queueing, disk scheduler.
4 * Based on ideas from a previously unfinished io
5 * scheduler (round robin per-process disk scheduling) and Andrea Arcangeli.
7 * Copyright (C) 2003 Jens Axboe <axboe@kernel.dk>
9 #include <linux/module.h>
10 #include <linux/slab.h>
11 #include <linux/blkdev.h>
12 #include <linux/elevator.h>
13 #include <linux/jiffies.h>
14 #include <linux/rbtree.h>
15 #include <linux/ioprio.h>
16 #include <linux/blktrace_api.h>
18 #include "blk-cgroup.h"
20 static struct blkio_policy_type blkio_policy_cfq;
25 /* max queue in one round of service */
26 static const int cfq_quantum = 8;
27 static const int cfq_fifo_expire[2] = { HZ / 4, HZ / 8 };
28 /* maximum backwards seek, in KiB */
29 static const int cfq_back_max = 16 * 1024;
30 /* penalty of a backwards seek */
31 static const int cfq_back_penalty = 2;
32 static const int cfq_slice_sync = HZ / 10;
33 static int cfq_slice_async = HZ / 25;
34 static const int cfq_slice_async_rq = 2;
35 static int cfq_slice_idle = HZ / 125;
36 static int cfq_group_idle = HZ / 125;
37 static const int cfq_target_latency = HZ * 3/10; /* 300 ms */
38 static const int cfq_hist_divisor = 4;
41 * offset from end of service tree
43 #define CFQ_IDLE_DELAY (HZ / 5)
46 * below this threshold, we consider thinktime immediate
48 #define CFQ_MIN_TT (2)
50 #define CFQ_SLICE_SCALE (5)
51 #define CFQ_HW_QUEUE_MIN (5)
52 #define CFQ_SERVICE_SHIFT 12
54 #define CFQQ_SEEK_THR (sector_t)(8 * 100)
55 #define CFQQ_CLOSE_THR (sector_t)(8 * 1024)
56 #define CFQQ_SECT_THR_NONROT (sector_t)(2 * 32)
57 #define CFQQ_SEEKY(cfqq) (hweight32(cfqq->seek_history) > 32/8)
59 #define RQ_CIC(rq) icq_to_cic((rq)->elv.icq)
60 #define RQ_CFQQ(rq) (struct cfq_queue *) ((rq)->elv.priv[0])
61 #define RQ_CFQG(rq) (struct cfq_group *) ((rq)->elv.priv[1])
63 static struct kmem_cache *cfq_pool;
65 #define CFQ_PRIO_LISTS IOPRIO_BE_NR
66 #define cfq_class_idle(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
67 #define cfq_class_rt(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
69 #define sample_valid(samples) ((samples) > 80)
70 #define rb_entry_cfqg(node) rb_entry((node), struct cfq_group, rb_node)
73 unsigned long last_end_request;
75 unsigned long ttime_total;
76 unsigned long ttime_samples;
77 unsigned long ttime_mean;
81 * Most of our rbtree usage is for sorting with min extraction, so
82 * if we cache the leftmost node we don't have to walk down the tree
83 * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should
84 * move this into the elevator for the rq sorting as well.
90 unsigned total_weight;
92 struct cfq_ttime ttime;
94 #define CFQ_RB_ROOT (struct cfq_rb_root) { .rb = RB_ROOT, \
95 .ttime = {.last_end_request = jiffies,},}
98 * Per process-grouping structure
101 /* reference count */
103 /* various state flags, see below */
105 /* parent cfq_data */
106 struct cfq_data *cfqd;
107 /* service_tree member */
108 struct rb_node rb_node;
109 /* service_tree key */
110 unsigned long rb_key;
111 /* prio tree member */
112 struct rb_node p_node;
113 /* prio tree root we belong to, if any */
114 struct rb_root *p_root;
115 /* sorted list of pending requests */
116 struct rb_root sort_list;
117 /* if fifo isn't expired, next request to serve */
118 struct request *next_rq;
119 /* requests queued in sort_list */
121 /* currently allocated requests */
123 /* fifo list of requests in sort_list */
124 struct list_head fifo;
126 /* time when queue got scheduled in to dispatch first request. */
127 unsigned long dispatch_start;
128 unsigned int allocated_slice;
129 unsigned int slice_dispatch;
130 /* time when first request from queue completed and slice started. */
131 unsigned long slice_start;
132 unsigned long slice_end;
135 /* pending priority requests */
137 /* number of requests that are on the dispatch list or inside driver */
140 /* io prio of this group */
141 unsigned short ioprio, org_ioprio;
142 unsigned short ioprio_class;
147 sector_t last_request_pos;
149 struct cfq_rb_root *service_tree;
150 struct cfq_queue *new_cfqq;
151 struct cfq_group *cfqg;
152 /* Number of sectors dispatched from queue in single dispatch round */
153 unsigned long nr_sectors;
157 * First index in the service_trees.
158 * IDLE is handled separately, so it has negative index
168 * Second index in the service_trees.
172 SYNC_NOIDLE_WORKLOAD = 1,
177 #ifdef CONFIG_CFQ_GROUP_IOSCHED
178 /* total bytes transferred */
179 struct blkg_rwstat service_bytes;
180 /* total IOs serviced, post merge */
181 struct blkg_rwstat serviced;
182 /* number of ios merged */
183 struct blkg_rwstat merged;
184 /* total time spent on device in ns, may not be accurate w/ queueing */
185 struct blkg_rwstat service_time;
186 /* total time spent waiting in scheduler queue in ns */
187 struct blkg_rwstat wait_time;
188 /* number of IOs queued up */
189 struct blkg_rwstat queued;
190 /* total sectors transferred */
191 struct blkg_stat sectors;
192 /* total disk time and nr sectors dispatched by this group */
193 struct blkg_stat time;
194 #ifdef CONFIG_DEBUG_BLK_CGROUP
195 /* time not charged to this cgroup */
196 struct blkg_stat unaccounted_time;
197 /* sum of number of ios queued across all samples */
198 struct blkg_stat avg_queue_size_sum;
199 /* count of samples taken for average */
200 struct blkg_stat avg_queue_size_samples;
201 /* how many times this group has been removed from service tree */
202 struct blkg_stat dequeue;
203 /* total time spent waiting for it to be assigned a timeslice. */
204 struct blkg_stat group_wait_time;
205 /* time spent idling for this blkio_group */
206 struct blkg_stat idle_time;
207 /* total time with empty current active q with other requests queued */
208 struct blkg_stat empty_time;
209 /* fields after this shouldn't be cleared on stat reset */
210 uint64_t start_group_wait_time;
211 uint64_t start_idle_time;
212 uint64_t start_empty_time;
214 #endif /* CONFIG_DEBUG_BLK_CGROUP */
215 #endif /* CONFIG_CFQ_GROUP_IOSCHED */
218 /* This is per cgroup per device grouping structure */
220 /* group service_tree member */
221 struct rb_node rb_node;
223 /* group service_tree key */
226 unsigned int new_weight;
229 /* number of cfqq currently on this group */
233 * Per group busy queues average. Useful for workload slice calc. We
234 * create the array for each prio class but at run time it is used
235 * only for RT and BE class and slot for IDLE class remains unused.
236 * This is primarily done to avoid confusion and a gcc warning.
238 unsigned int busy_queues_avg[CFQ_PRIO_NR];
240 * rr lists of queues with requests. We maintain service trees for
241 * RT and BE classes. These trees are subdivided in subclasses
242 * of SYNC, SYNC_NOIDLE and ASYNC based on workload type. For IDLE
243 * class there is no subclassification and all the cfq queues go on
244 * a single tree service_tree_idle.
245 * Counts are embedded in the cfq_rb_root
247 struct cfq_rb_root service_trees[2][3];
248 struct cfq_rb_root service_tree_idle;
250 unsigned long saved_workload_slice;
251 enum wl_type_t saved_workload;
252 enum wl_prio_t saved_serving_prio;
254 /* number of requests that are on the dispatch list or inside driver */
256 struct cfq_ttime ttime;
257 struct cfqg_stats stats;
261 struct io_cq icq; /* must be the first member */
262 struct cfq_queue *cfqq[2];
263 struct cfq_ttime ttime;
264 int ioprio; /* the current ioprio */
265 #ifdef CONFIG_CFQ_GROUP_IOSCHED
266 uint64_t blkcg_id; /* the current blkcg ID */
271 * Per block device queue structure
274 struct request_queue *queue;
275 /* Root service tree for cfq_groups */
276 struct cfq_rb_root grp_service_tree;
277 struct cfq_group *root_group;
280 * The priority currently being served
282 enum wl_prio_t serving_prio;
283 enum wl_type_t serving_type;
284 unsigned long workload_expires;
285 struct cfq_group *serving_group;
288 * Each priority tree is sorted by next_request position. These
289 * trees are used when determining if two or more queues are
290 * interleaving requests (see cfq_close_cooperator).
292 struct rb_root prio_trees[CFQ_PRIO_LISTS];
294 unsigned int busy_queues;
295 unsigned int busy_sync_queues;
301 * queue-depth detection
307 * -1 => indeterminate, (cfq will behave as if NCQ is present, to allow better detection)
308 * 1 => NCQ is present (hw_tag_est_depth is the estimated max depth)
311 int hw_tag_est_depth;
312 unsigned int hw_tag_samples;
315 * idle window management
317 struct timer_list idle_slice_timer;
318 struct work_struct unplug_work;
320 struct cfq_queue *active_queue;
321 struct cfq_io_cq *active_cic;
324 * async queue for each priority case
326 struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
327 struct cfq_queue *async_idle_cfqq;
329 sector_t last_position;
332 * tunables, see top of file
334 unsigned int cfq_quantum;
335 unsigned int cfq_fifo_expire[2];
336 unsigned int cfq_back_penalty;
337 unsigned int cfq_back_max;
338 unsigned int cfq_slice[2];
339 unsigned int cfq_slice_async_rq;
340 unsigned int cfq_slice_idle;
341 unsigned int cfq_group_idle;
342 unsigned int cfq_latency;
345 * Fallback dummy cfqq for extreme OOM conditions
347 struct cfq_queue oom_cfqq;
349 unsigned long last_delayed_sync;
352 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd);
354 static struct cfq_rb_root *service_tree_for(struct cfq_group *cfqg,
361 if (prio == IDLE_WORKLOAD)
362 return &cfqg->service_tree_idle;
364 return &cfqg->service_trees[prio][type];
367 enum cfqq_state_flags {
368 CFQ_CFQQ_FLAG_on_rr = 0, /* on round-robin busy list */
369 CFQ_CFQQ_FLAG_wait_request, /* waiting for a request */
370 CFQ_CFQQ_FLAG_must_dispatch, /* must be allowed a dispatch */
371 CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
372 CFQ_CFQQ_FLAG_fifo_expire, /* FIFO checked in this slice */
373 CFQ_CFQQ_FLAG_idle_window, /* slice idling enabled */
374 CFQ_CFQQ_FLAG_prio_changed, /* task priority has changed */
375 CFQ_CFQQ_FLAG_slice_new, /* no requests dispatched in slice */
376 CFQ_CFQQ_FLAG_sync, /* synchronous queue */
377 CFQ_CFQQ_FLAG_coop, /* cfqq is shared */
378 CFQ_CFQQ_FLAG_split_coop, /* shared cfqq will be splitted */
379 CFQ_CFQQ_FLAG_deep, /* sync cfqq experienced large depth */
380 CFQ_CFQQ_FLAG_wait_busy, /* Waiting for next request */
383 #define CFQ_CFQQ_FNS(name) \
384 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \
386 (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name); \
388 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \
390 (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \
392 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \
394 return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \
398 CFQ_CFQQ_FNS(wait_request);
399 CFQ_CFQQ_FNS(must_dispatch);
400 CFQ_CFQQ_FNS(must_alloc_slice);
401 CFQ_CFQQ_FNS(fifo_expire);
402 CFQ_CFQQ_FNS(idle_window);
403 CFQ_CFQQ_FNS(prio_changed);
404 CFQ_CFQQ_FNS(slice_new);
407 CFQ_CFQQ_FNS(split_coop);
409 CFQ_CFQQ_FNS(wait_busy);
412 #if defined(CONFIG_CFQ_GROUP_IOSCHED) && defined(CONFIG_DEBUG_BLK_CGROUP)
414 /* cfqg stats flags */
415 enum cfqg_stats_flags {
416 CFQG_stats_waiting = 0,
421 #define CFQG_FLAG_FNS(name) \
422 static inline void cfqg_stats_mark_##name(struct cfqg_stats *stats) \
424 stats->flags |= (1 << CFQG_stats_##name); \
426 static inline void cfqg_stats_clear_##name(struct cfqg_stats *stats) \
428 stats->flags &= ~(1 << CFQG_stats_##name); \
430 static inline int cfqg_stats_##name(struct cfqg_stats *stats) \
432 return (stats->flags & (1 << CFQG_stats_##name)) != 0; \
435 CFQG_FLAG_FNS(waiting)
436 CFQG_FLAG_FNS(idling)
440 /* This should be called with the queue_lock held. */
441 static void cfqg_stats_update_group_wait_time(struct cfqg_stats *stats)
443 unsigned long long now;
445 if (!cfqg_stats_waiting(stats))
449 if (time_after64(now, stats->start_group_wait_time))
450 blkg_stat_add(&stats->group_wait_time,
451 now - stats->start_group_wait_time);
452 cfqg_stats_clear_waiting(stats);
455 /* This should be called with the queue_lock held. */
456 static void cfqg_stats_set_start_group_wait_time(struct cfq_group *cfqg,
457 struct cfq_group *curr_cfqg)
459 struct cfqg_stats *stats = &cfqg->stats;
461 if (cfqg_stats_waiting(stats))
463 if (cfqg == curr_cfqg)
465 stats->start_group_wait_time = sched_clock();
466 cfqg_stats_mark_waiting(stats);
469 /* This should be called with the queue_lock held. */
470 static void cfqg_stats_end_empty_time(struct cfqg_stats *stats)
472 unsigned long long now;
474 if (!cfqg_stats_empty(stats))
478 if (time_after64(now, stats->start_empty_time))
479 blkg_stat_add(&stats->empty_time,
480 now - stats->start_empty_time);
481 cfqg_stats_clear_empty(stats);
484 static void cfqg_stats_update_dequeue(struct cfq_group *cfqg)
486 blkg_stat_add(&cfqg->stats.dequeue, 1);
489 static void cfqg_stats_set_start_empty_time(struct cfq_group *cfqg)
491 struct cfqg_stats *stats = &cfqg->stats;
493 if (blkg_rwstat_sum(&stats->queued))
497 * group is already marked empty. This can happen if cfqq got new
498 * request in parent group and moved to this group while being added
499 * to service tree. Just ignore the event and move on.
501 if (cfqg_stats_empty(stats))
504 stats->start_empty_time = sched_clock();
505 cfqg_stats_mark_empty(stats);
508 static void cfqg_stats_update_idle_time(struct cfq_group *cfqg)
510 struct cfqg_stats *stats = &cfqg->stats;
512 if (cfqg_stats_idling(stats)) {
513 unsigned long long now = sched_clock();
515 if (time_after64(now, stats->start_idle_time))
516 blkg_stat_add(&stats->idle_time,
517 now - stats->start_idle_time);
518 cfqg_stats_clear_idling(stats);
522 static void cfqg_stats_set_start_idle_time(struct cfq_group *cfqg)
524 struct cfqg_stats *stats = &cfqg->stats;
526 BUG_ON(cfqg_stats_idling(stats));
528 stats->start_idle_time = sched_clock();
529 cfqg_stats_mark_idling(stats);
532 static void cfqg_stats_update_avg_queue_size(struct cfq_group *cfqg)
534 struct cfqg_stats *stats = &cfqg->stats;
536 blkg_stat_add(&stats->avg_queue_size_sum,
537 blkg_rwstat_sum(&stats->queued));
538 blkg_stat_add(&stats->avg_queue_size_samples, 1);
539 cfqg_stats_update_group_wait_time(stats);
542 #else /* CONFIG_CFQ_GROUP_IOSCHED && CONFIG_DEBUG_BLK_CGROUP */
544 static void cfqg_stats_set_start_group_wait_time(struct cfq_group *cfqg,
545 struct cfq_group *curr_cfqg) { }
546 static void cfqg_stats_end_empty_time(struct cfqg_stats *stats) { }
547 static void cfqg_stats_update_dequeue(struct cfq_group *cfqg) { }
548 static void cfqg_stats_set_start_empty_time(struct cfq_group *cfqg) { }
549 static void cfqg_stats_update_idle_time(struct cfq_group *cfqg) { }
550 static void cfqg_stats_set_start_idle_time(struct cfq_group *cfqg) { }
551 static void cfqg_stats_update_avg_queue_size(struct cfq_group *cfqg) { }
553 #endif /* CONFIG_CFQ_GROUP_IOSCHED && CONFIG_DEBUG_BLK_CGROUP */
555 #ifdef CONFIG_CFQ_GROUP_IOSCHED
557 static inline struct cfq_group *blkg_to_cfqg(struct blkio_group *blkg)
559 return blkg_to_pdata(blkg, &blkio_policy_cfq);
562 static inline struct blkio_group *cfqg_to_blkg(struct cfq_group *cfqg)
564 return pdata_to_blkg(cfqg);
567 static inline void cfqg_get(struct cfq_group *cfqg)
569 return blkg_get(cfqg_to_blkg(cfqg));
572 static inline void cfqg_put(struct cfq_group *cfqg)
574 return blkg_put(cfqg_to_blkg(cfqg));
577 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
578 blk_add_trace_msg((cfqd)->queue, "cfq%d%c %s " fmt, (cfqq)->pid, \
579 cfq_cfqq_sync((cfqq)) ? 'S' : 'A', \
580 blkg_path(cfqg_to_blkg((cfqq)->cfqg)), ##args)
582 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...) \
583 blk_add_trace_msg((cfqd)->queue, "%s " fmt, \
584 blkg_path(cfqg_to_blkg((cfqg))), ##args) \
586 static inline void cfqg_stats_update_io_add(struct cfq_group *cfqg,
587 struct cfq_group *curr_cfqg, int rw)
589 blkg_rwstat_add(&cfqg->stats.queued, rw, 1);
590 cfqg_stats_end_empty_time(&cfqg->stats);
591 cfqg_stats_set_start_group_wait_time(cfqg, curr_cfqg);
594 static inline void cfqg_stats_update_timeslice_used(struct cfq_group *cfqg,
595 unsigned long time, unsigned long unaccounted_time)
597 blkg_stat_add(&cfqg->stats.time, time);
598 #ifdef CONFIG_DEBUG_BLK_CGROUP
599 blkg_stat_add(&cfqg->stats.unaccounted_time, unaccounted_time);
603 static inline void cfqg_stats_update_io_remove(struct cfq_group *cfqg, int rw)
605 blkg_rwstat_add(&cfqg->stats.queued, rw, -1);
608 static inline void cfqg_stats_update_io_merged(struct cfq_group *cfqg, int rw)
610 blkg_rwstat_add(&cfqg->stats.merged, rw, 1);
613 static inline void cfqg_stats_update_dispatch(struct cfq_group *cfqg,
614 uint64_t bytes, int rw)
616 blkg_stat_add(&cfqg->stats.sectors, bytes >> 9);
617 blkg_rwstat_add(&cfqg->stats.serviced, rw, 1);
618 blkg_rwstat_add(&cfqg->stats.service_bytes, rw, bytes);
621 static inline void cfqg_stats_update_completion(struct cfq_group *cfqg,
622 uint64_t start_time, uint64_t io_start_time, int rw)
624 struct cfqg_stats *stats = &cfqg->stats;
625 unsigned long long now = sched_clock();
627 if (time_after64(now, io_start_time))
628 blkg_rwstat_add(&stats->service_time, rw, now - io_start_time);
629 if (time_after64(io_start_time, start_time))
630 blkg_rwstat_add(&stats->wait_time, rw,
631 io_start_time - start_time);
634 static void cfqg_stats_reset(struct blkio_group *blkg)
636 struct cfq_group *cfqg = blkg_to_cfqg(blkg);
637 struct cfqg_stats *stats = &cfqg->stats;
639 /* queued stats shouldn't be cleared */
640 blkg_rwstat_reset(&stats->service_bytes);
641 blkg_rwstat_reset(&stats->serviced);
642 blkg_rwstat_reset(&stats->merged);
643 blkg_rwstat_reset(&stats->service_time);
644 blkg_rwstat_reset(&stats->wait_time);
645 blkg_stat_reset(&stats->time);
646 #ifdef CONFIG_DEBUG_BLK_CGROUP
647 blkg_stat_reset(&stats->unaccounted_time);
648 blkg_stat_reset(&stats->avg_queue_size_sum);
649 blkg_stat_reset(&stats->avg_queue_size_samples);
650 blkg_stat_reset(&stats->dequeue);
651 blkg_stat_reset(&stats->group_wait_time);
652 blkg_stat_reset(&stats->idle_time);
653 blkg_stat_reset(&stats->empty_time);
657 #else /* CONFIG_CFQ_GROUP_IOSCHED */
659 static inline struct cfq_group *blkg_to_cfqg(struct blkio_group *blkg) { return NULL; }
660 static inline struct blkio_group *cfqg_to_blkg(struct cfq_group *cfqg) { return NULL; }
661 static inline void cfqg_get(struct cfq_group *cfqg) { }
662 static inline void cfqg_put(struct cfq_group *cfqg) { }
664 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
665 blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args)
666 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...) do {} while (0)
668 static inline void cfqg_stats_update_io_add(struct cfq_group *cfqg,
669 struct cfq_group *curr_cfqg, int rw) { }
670 static inline void cfqg_stats_update_timeslice_used(struct cfq_group *cfqg,
671 unsigned long time, unsigned long unaccounted_time) { }
672 static inline void cfqg_stats_update_io_remove(struct cfq_group *cfqg, int rw) { }
673 static inline void cfqg_stats_update_io_merged(struct cfq_group *cfqg, int rw) { }
674 static inline void cfqg_stats_update_dispatch(struct cfq_group *cfqg,
675 uint64_t bytes, int rw) { }
676 static inline void cfqg_stats_update_completion(struct cfq_group *cfqg,
677 uint64_t start_time, uint64_t io_start_time, int rw) { }
679 #endif /* CONFIG_CFQ_GROUP_IOSCHED */
681 #define cfq_log(cfqd, fmt, args...) \
682 blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
684 /* Traverses through cfq group service trees */
685 #define for_each_cfqg_st(cfqg, i, j, st) \
686 for (i = 0; i <= IDLE_WORKLOAD; i++) \
687 for (j = 0, st = i < IDLE_WORKLOAD ? &cfqg->service_trees[i][j]\
688 : &cfqg->service_tree_idle; \
689 (i < IDLE_WORKLOAD && j <= SYNC_WORKLOAD) || \
690 (i == IDLE_WORKLOAD && j == 0); \
691 j++, st = i < IDLE_WORKLOAD ? \
692 &cfqg->service_trees[i][j]: NULL) \
694 static inline bool cfq_io_thinktime_big(struct cfq_data *cfqd,
695 struct cfq_ttime *ttime, bool group_idle)
698 if (!sample_valid(ttime->ttime_samples))
701 slice = cfqd->cfq_group_idle;
703 slice = cfqd->cfq_slice_idle;
704 return ttime->ttime_mean > slice;
707 static inline bool iops_mode(struct cfq_data *cfqd)
710 * If we are not idling on queues and it is a NCQ drive, parallel
711 * execution of requests is on and measuring time is not possible
712 * in most of the cases until and unless we drive shallower queue
713 * depths and that becomes a performance bottleneck. In such cases
714 * switch to start providing fairness in terms of number of IOs.
716 if (!cfqd->cfq_slice_idle && cfqd->hw_tag)
722 static inline enum wl_prio_t cfqq_prio(struct cfq_queue *cfqq)
724 if (cfq_class_idle(cfqq))
725 return IDLE_WORKLOAD;
726 if (cfq_class_rt(cfqq))
732 static enum wl_type_t cfqq_type(struct cfq_queue *cfqq)
734 if (!cfq_cfqq_sync(cfqq))
735 return ASYNC_WORKLOAD;
736 if (!cfq_cfqq_idle_window(cfqq))
737 return SYNC_NOIDLE_WORKLOAD;
738 return SYNC_WORKLOAD;
741 static inline int cfq_group_busy_queues_wl(enum wl_prio_t wl,
742 struct cfq_data *cfqd,
743 struct cfq_group *cfqg)
745 if (wl == IDLE_WORKLOAD)
746 return cfqg->service_tree_idle.count;
748 return cfqg->service_trees[wl][ASYNC_WORKLOAD].count
749 + cfqg->service_trees[wl][SYNC_NOIDLE_WORKLOAD].count
750 + cfqg->service_trees[wl][SYNC_WORKLOAD].count;
753 static inline int cfqg_busy_async_queues(struct cfq_data *cfqd,
754 struct cfq_group *cfqg)
756 return cfqg->service_trees[RT_WORKLOAD][ASYNC_WORKLOAD].count
757 + cfqg->service_trees[BE_WORKLOAD][ASYNC_WORKLOAD].count;
760 static void cfq_dispatch_insert(struct request_queue *, struct request *);
761 static struct cfq_queue *cfq_get_queue(struct cfq_data *cfqd, bool is_sync,
762 struct cfq_io_cq *cic, struct bio *bio,
765 static inline struct cfq_io_cq *icq_to_cic(struct io_cq *icq)
767 /* cic->icq is the first member, %NULL will convert to %NULL */
768 return container_of(icq, struct cfq_io_cq, icq);
771 static inline struct cfq_io_cq *cfq_cic_lookup(struct cfq_data *cfqd,
772 struct io_context *ioc)
775 return icq_to_cic(ioc_lookup_icq(ioc, cfqd->queue));
779 static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_cq *cic, bool is_sync)
781 return cic->cfqq[is_sync];
784 static inline void cic_set_cfqq(struct cfq_io_cq *cic, struct cfq_queue *cfqq,
787 cic->cfqq[is_sync] = cfqq;
790 static inline struct cfq_data *cic_to_cfqd(struct cfq_io_cq *cic)
792 return cic->icq.q->elevator->elevator_data;
796 * We regard a request as SYNC, if it's either a read or has the SYNC bit
797 * set (in which case it could also be direct WRITE).
799 static inline bool cfq_bio_sync(struct bio *bio)
801 return bio_data_dir(bio) == READ || (bio->bi_rw & REQ_SYNC);
805 * scheduler run of queue, if there are requests pending and no one in the
806 * driver that will restart queueing
808 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
810 if (cfqd->busy_queues) {
811 cfq_log(cfqd, "schedule dispatch");
812 kblockd_schedule_work(cfqd->queue, &cfqd->unplug_work);
817 * Scale schedule slice based on io priority. Use the sync time slice only
818 * if a queue is marked sync and has sync io queued. A sync queue with async
819 * io only, should not get full sync slice length.
821 static inline int cfq_prio_slice(struct cfq_data *cfqd, bool sync,
824 const int base_slice = cfqd->cfq_slice[sync];
826 WARN_ON(prio >= IOPRIO_BE_NR);
828 return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
832 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
834 return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
837 static inline u64 cfq_scale_slice(unsigned long delta, struct cfq_group *cfqg)
839 u64 d = delta << CFQ_SERVICE_SHIFT;
841 d = d * BLKIO_WEIGHT_DEFAULT;
842 do_div(d, cfqg->weight);
846 static inline u64 max_vdisktime(u64 min_vdisktime, u64 vdisktime)
848 s64 delta = (s64)(vdisktime - min_vdisktime);
850 min_vdisktime = vdisktime;
852 return min_vdisktime;
855 static inline u64 min_vdisktime(u64 min_vdisktime, u64 vdisktime)
857 s64 delta = (s64)(vdisktime - min_vdisktime);
859 min_vdisktime = vdisktime;
861 return min_vdisktime;
864 static void update_min_vdisktime(struct cfq_rb_root *st)
866 struct cfq_group *cfqg;
869 cfqg = rb_entry_cfqg(st->left);
870 st->min_vdisktime = max_vdisktime(st->min_vdisktime,
876 * get averaged number of queues of RT/BE priority.
877 * average is updated, with a formula that gives more weight to higher numbers,
878 * to quickly follows sudden increases and decrease slowly
881 static inline unsigned cfq_group_get_avg_queues(struct cfq_data *cfqd,
882 struct cfq_group *cfqg, bool rt)
884 unsigned min_q, max_q;
885 unsigned mult = cfq_hist_divisor - 1;
886 unsigned round = cfq_hist_divisor / 2;
887 unsigned busy = cfq_group_busy_queues_wl(rt, cfqd, cfqg);
889 min_q = min(cfqg->busy_queues_avg[rt], busy);
890 max_q = max(cfqg->busy_queues_avg[rt], busy);
891 cfqg->busy_queues_avg[rt] = (mult * max_q + min_q + round) /
893 return cfqg->busy_queues_avg[rt];
896 static inline unsigned
897 cfq_group_slice(struct cfq_data *cfqd, struct cfq_group *cfqg)
899 struct cfq_rb_root *st = &cfqd->grp_service_tree;
901 return cfq_target_latency * cfqg->weight / st->total_weight;
904 static inline unsigned
905 cfq_scaled_cfqq_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
907 unsigned slice = cfq_prio_to_slice(cfqd, cfqq);
908 if (cfqd->cfq_latency) {
910 * interested queues (we consider only the ones with the same
911 * priority class in the cfq group)
913 unsigned iq = cfq_group_get_avg_queues(cfqd, cfqq->cfqg,
915 unsigned sync_slice = cfqd->cfq_slice[1];
916 unsigned expect_latency = sync_slice * iq;
917 unsigned group_slice = cfq_group_slice(cfqd, cfqq->cfqg);
919 if (expect_latency > group_slice) {
920 unsigned base_low_slice = 2 * cfqd->cfq_slice_idle;
921 /* scale low_slice according to IO priority
922 * and sync vs async */
924 min(slice, base_low_slice * slice / sync_slice);
925 /* the adapted slice value is scaled to fit all iqs
926 * into the target latency */
927 slice = max(slice * group_slice / expect_latency,
935 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
937 unsigned slice = cfq_scaled_cfqq_slice(cfqd, cfqq);
939 cfqq->slice_start = jiffies;
940 cfqq->slice_end = jiffies + slice;
941 cfqq->allocated_slice = slice;
942 cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
946 * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
947 * isn't valid until the first request from the dispatch is activated
948 * and the slice time set.
950 static inline bool cfq_slice_used(struct cfq_queue *cfqq)
952 if (cfq_cfqq_slice_new(cfqq))
954 if (time_before(jiffies, cfqq->slice_end))
961 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
962 * We choose the request that is closest to the head right now. Distance
963 * behind the head is penalized and only allowed to a certain extent.
965 static struct request *
966 cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2, sector_t last)
968 sector_t s1, s2, d1 = 0, d2 = 0;
969 unsigned long back_max;
970 #define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */
971 #define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */
972 unsigned wrap = 0; /* bit mask: requests behind the disk head? */
974 if (rq1 == NULL || rq1 == rq2)
979 if (rq_is_sync(rq1) != rq_is_sync(rq2))
980 return rq_is_sync(rq1) ? rq1 : rq2;
982 if ((rq1->cmd_flags ^ rq2->cmd_flags) & REQ_PRIO)
983 return rq1->cmd_flags & REQ_PRIO ? rq1 : rq2;
985 s1 = blk_rq_pos(rq1);
986 s2 = blk_rq_pos(rq2);
989 * by definition, 1KiB is 2 sectors
991 back_max = cfqd->cfq_back_max * 2;
994 * Strict one way elevator _except_ in the case where we allow
995 * short backward seeks which are biased as twice the cost of a
996 * similar forward seek.
1000 else if (s1 + back_max >= last)
1001 d1 = (last - s1) * cfqd->cfq_back_penalty;
1003 wrap |= CFQ_RQ1_WRAP;
1007 else if (s2 + back_max >= last)
1008 d2 = (last - s2) * cfqd->cfq_back_penalty;
1010 wrap |= CFQ_RQ2_WRAP;
1012 /* Found required data */
1015 * By doing switch() on the bit mask "wrap" we avoid having to
1016 * check two variables for all permutations: --> faster!
1019 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
1035 case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
1038 * Since both rqs are wrapped,
1039 * start with the one that's further behind head
1040 * (--> only *one* back seek required),
1041 * since back seek takes more time than forward.
1051 * The below is leftmost cache rbtree addon
1053 static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
1055 /* Service tree is empty */
1060 root->left = rb_first(&root->rb);
1063 return rb_entry(root->left, struct cfq_queue, rb_node);
1068 static struct cfq_group *cfq_rb_first_group(struct cfq_rb_root *root)
1071 root->left = rb_first(&root->rb);
1074 return rb_entry_cfqg(root->left);
1079 static void rb_erase_init(struct rb_node *n, struct rb_root *root)
1085 static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
1087 if (root->left == n)
1089 rb_erase_init(n, &root->rb);
1094 * would be nice to take fifo expire time into account as well
1096 static struct request *
1097 cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1098 struct request *last)
1100 struct rb_node *rbnext = rb_next(&last->rb_node);
1101 struct rb_node *rbprev = rb_prev(&last->rb_node);
1102 struct request *next = NULL, *prev = NULL;
1104 BUG_ON(RB_EMPTY_NODE(&last->rb_node));
1107 prev = rb_entry_rq(rbprev);
1110 next = rb_entry_rq(rbnext);
1112 rbnext = rb_first(&cfqq->sort_list);
1113 if (rbnext && rbnext != &last->rb_node)
1114 next = rb_entry_rq(rbnext);
1117 return cfq_choose_req(cfqd, next, prev, blk_rq_pos(last));
1120 static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
1121 struct cfq_queue *cfqq)
1124 * just an approximation, should be ok.
1126 return (cfqq->cfqg->nr_cfqq - 1) * (cfq_prio_slice(cfqd, 1, 0) -
1127 cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
1131 cfqg_key(struct cfq_rb_root *st, struct cfq_group *cfqg)
1133 return cfqg->vdisktime - st->min_vdisktime;
1137 __cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
1139 struct rb_node **node = &st->rb.rb_node;
1140 struct rb_node *parent = NULL;
1141 struct cfq_group *__cfqg;
1142 s64 key = cfqg_key(st, cfqg);
1145 while (*node != NULL) {
1147 __cfqg = rb_entry_cfqg(parent);
1149 if (key < cfqg_key(st, __cfqg))
1150 node = &parent->rb_left;
1152 node = &parent->rb_right;
1158 st->left = &cfqg->rb_node;
1160 rb_link_node(&cfqg->rb_node, parent, node);
1161 rb_insert_color(&cfqg->rb_node, &st->rb);
1165 cfq_update_group_weight(struct cfq_group *cfqg)
1167 BUG_ON(!RB_EMPTY_NODE(&cfqg->rb_node));
1168 if (cfqg->needs_update) {
1169 cfqg->weight = cfqg->new_weight;
1170 cfqg->needs_update = false;
1175 cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
1177 BUG_ON(!RB_EMPTY_NODE(&cfqg->rb_node));
1179 cfq_update_group_weight(cfqg);
1180 __cfq_group_service_tree_add(st, cfqg);
1181 st->total_weight += cfqg->weight;
1185 cfq_group_notify_queue_add(struct cfq_data *cfqd, struct cfq_group *cfqg)
1187 struct cfq_rb_root *st = &cfqd->grp_service_tree;
1188 struct cfq_group *__cfqg;
1192 if (!RB_EMPTY_NODE(&cfqg->rb_node))
1196 * Currently put the group at the end. Later implement something
1197 * so that groups get lesser vtime based on their weights, so that
1198 * if group does not loose all if it was not continuously backlogged.
1200 n = rb_last(&st->rb);
1202 __cfqg = rb_entry_cfqg(n);
1203 cfqg->vdisktime = __cfqg->vdisktime + CFQ_IDLE_DELAY;
1205 cfqg->vdisktime = st->min_vdisktime;
1206 cfq_group_service_tree_add(st, cfqg);
1210 cfq_group_service_tree_del(struct cfq_rb_root *st, struct cfq_group *cfqg)
1212 st->total_weight -= cfqg->weight;
1213 if (!RB_EMPTY_NODE(&cfqg->rb_node))
1214 cfq_rb_erase(&cfqg->rb_node, st);
1218 cfq_group_notify_queue_del(struct cfq_data *cfqd, struct cfq_group *cfqg)
1220 struct cfq_rb_root *st = &cfqd->grp_service_tree;
1222 BUG_ON(cfqg->nr_cfqq < 1);
1225 /* If there are other cfq queues under this group, don't delete it */
1229 cfq_log_cfqg(cfqd, cfqg, "del_from_rr group");
1230 cfq_group_service_tree_del(st, cfqg);
1231 cfqg->saved_workload_slice = 0;
1232 cfqg_stats_update_dequeue(cfqg);
1235 static inline unsigned int cfq_cfqq_slice_usage(struct cfq_queue *cfqq,
1236 unsigned int *unaccounted_time)
1238 unsigned int slice_used;
1241 * Queue got expired before even a single request completed or
1242 * got expired immediately after first request completion.
1244 if (!cfqq->slice_start || cfqq->slice_start == jiffies) {
1246 * Also charge the seek time incurred to the group, otherwise
1247 * if there are mutiple queues in the group, each can dispatch
1248 * a single request on seeky media and cause lots of seek time
1249 * and group will never know it.
1251 slice_used = max_t(unsigned, (jiffies - cfqq->dispatch_start),
1254 slice_used = jiffies - cfqq->slice_start;
1255 if (slice_used > cfqq->allocated_slice) {
1256 *unaccounted_time = slice_used - cfqq->allocated_slice;
1257 slice_used = cfqq->allocated_slice;
1259 if (time_after(cfqq->slice_start, cfqq->dispatch_start))
1260 *unaccounted_time += cfqq->slice_start -
1261 cfqq->dispatch_start;
1267 static void cfq_group_served(struct cfq_data *cfqd, struct cfq_group *cfqg,
1268 struct cfq_queue *cfqq)
1270 struct cfq_rb_root *st = &cfqd->grp_service_tree;
1271 unsigned int used_sl, charge, unaccounted_sl = 0;
1272 int nr_sync = cfqg->nr_cfqq - cfqg_busy_async_queues(cfqd, cfqg)
1273 - cfqg->service_tree_idle.count;
1275 BUG_ON(nr_sync < 0);
1276 used_sl = charge = cfq_cfqq_slice_usage(cfqq, &unaccounted_sl);
1278 if (iops_mode(cfqd))
1279 charge = cfqq->slice_dispatch;
1280 else if (!cfq_cfqq_sync(cfqq) && !nr_sync)
1281 charge = cfqq->allocated_slice;
1283 /* Can't update vdisktime while group is on service tree */
1284 cfq_group_service_tree_del(st, cfqg);
1285 cfqg->vdisktime += cfq_scale_slice(charge, cfqg);
1286 /* If a new weight was requested, update now, off tree */
1287 cfq_group_service_tree_add(st, cfqg);
1289 /* This group is being expired. Save the context */
1290 if (time_after(cfqd->workload_expires, jiffies)) {
1291 cfqg->saved_workload_slice = cfqd->workload_expires
1293 cfqg->saved_workload = cfqd->serving_type;
1294 cfqg->saved_serving_prio = cfqd->serving_prio;
1296 cfqg->saved_workload_slice = 0;
1298 cfq_log_cfqg(cfqd, cfqg, "served: vt=%llu min_vt=%llu", cfqg->vdisktime,
1300 cfq_log_cfqq(cfqq->cfqd, cfqq,
1301 "sl_used=%u disp=%u charge=%u iops=%u sect=%lu",
1302 used_sl, cfqq->slice_dispatch, charge,
1303 iops_mode(cfqd), cfqq->nr_sectors);
1304 cfqg_stats_update_timeslice_used(cfqg, used_sl, unaccounted_sl);
1305 cfqg_stats_set_start_empty_time(cfqg);
1309 * cfq_init_cfqg_base - initialize base part of a cfq_group
1310 * @cfqg: cfq_group to initialize
1312 * Initialize the base part which is used whether %CONFIG_CFQ_GROUP_IOSCHED
1313 * is enabled or not.
1315 static void cfq_init_cfqg_base(struct cfq_group *cfqg)
1317 struct cfq_rb_root *st;
1320 for_each_cfqg_st(cfqg, i, j, st)
1322 RB_CLEAR_NODE(&cfqg->rb_node);
1324 cfqg->ttime.last_end_request = jiffies;
1327 #ifdef CONFIG_CFQ_GROUP_IOSCHED
1328 static void cfq_update_blkio_group_weight(struct blkio_group *blkg,
1329 unsigned int weight)
1331 struct cfq_group *cfqg = blkg_to_cfqg(blkg);
1333 cfqg->new_weight = weight;
1334 cfqg->needs_update = true;
1337 static void cfq_init_blkio_group(struct blkio_group *blkg)
1339 struct cfq_group *cfqg = blkg_to_cfqg(blkg);
1341 cfq_init_cfqg_base(cfqg);
1342 cfqg->weight = blkg->blkcg->weight;
1346 * Search for the cfq group current task belongs to. request_queue lock must
1349 static struct cfq_group *cfq_lookup_create_cfqg(struct cfq_data *cfqd,
1350 struct blkio_cgroup *blkcg)
1352 struct request_queue *q = cfqd->queue;
1353 struct cfq_group *cfqg = NULL;
1355 /* avoid lookup for the common case where there's no blkio cgroup */
1356 if (blkcg == &blkio_root_cgroup) {
1357 cfqg = cfqd->root_group;
1359 struct blkio_group *blkg;
1361 blkg = blkg_lookup_create(blkcg, q, false);
1363 cfqg = blkg_to_cfqg(blkg);
1369 static void cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg)
1371 /* Currently, all async queues are mapped to root group */
1372 if (!cfq_cfqq_sync(cfqq))
1373 cfqg = cfqq->cfqd->root_group;
1376 /* cfqq reference on cfqg */
1380 static u64 blkg_prfill_weight_device(struct seq_file *sf,
1381 struct blkg_policy_data *pd, int off)
1383 if (!pd->conf.weight)
1385 return __blkg_prfill_u64(sf, pd, pd->conf.weight);
1388 static int blkcg_print_weight_device(struct cgroup *cgrp, struct cftype *cft,
1389 struct seq_file *sf)
1391 blkcg_print_blkgs(sf, cgroup_to_blkio_cgroup(cgrp),
1392 blkg_prfill_weight_device, BLKIO_POLICY_PROP, 0,
1397 static int blkcg_print_weight(struct cgroup *cgrp, struct cftype *cft,
1398 struct seq_file *sf)
1400 seq_printf(sf, "%u\n", cgroup_to_blkio_cgroup(cgrp)->weight);
1404 static int blkcg_set_weight_device(struct cgroup *cgrp, struct cftype *cft,
1407 struct blkio_cgroup *blkcg = cgroup_to_blkio_cgroup(cgrp);
1408 struct blkg_policy_data *pd;
1409 struct blkg_conf_ctx ctx;
1412 ret = blkg_conf_prep(blkcg, buf, &ctx);
1417 pd = ctx.blkg->pd[BLKIO_POLICY_PROP];
1418 if (pd && (!ctx.v || (ctx.v >= BLKIO_WEIGHT_MIN &&
1419 ctx.v <= BLKIO_WEIGHT_MAX))) {
1420 pd->conf.weight = ctx.v;
1421 cfq_update_blkio_group_weight(ctx.blkg, ctx.v ?: blkcg->weight);
1425 blkg_conf_finish(&ctx);
1429 static int blkcg_set_weight(struct cgroup *cgrp, struct cftype *cft, u64 val)
1431 struct blkio_cgroup *blkcg = cgroup_to_blkio_cgroup(cgrp);
1432 struct blkio_group *blkg;
1433 struct hlist_node *n;
1435 if (val < BLKIO_WEIGHT_MIN || val > BLKIO_WEIGHT_MAX)
1438 spin_lock_irq(&blkcg->lock);
1439 blkcg->weight = (unsigned int)val;
1441 hlist_for_each_entry(blkg, n, &blkcg->blkg_list, blkcg_node) {
1442 struct blkg_policy_data *pd = blkg->pd[BLKIO_POLICY_PROP];
1444 if (pd && !pd->conf.weight)
1445 cfq_update_blkio_group_weight(blkg, blkcg->weight);
1448 spin_unlock_irq(&blkcg->lock);
1452 #ifdef CONFIG_DEBUG_BLK_CGROUP
1453 static u64 cfqg_prfill_avg_queue_size(struct seq_file *sf,
1454 struct blkg_policy_data *pd, int off)
1456 struct cfq_group *cfqg = (void *)pd->pdata;
1457 u64 samples = blkg_stat_read(&cfqg->stats.avg_queue_size_samples);
1461 v = blkg_stat_read(&cfqg->stats.avg_queue_size_sum);
1464 __blkg_prfill_u64(sf, pd, v);
1468 /* print avg_queue_size */
1469 static int cfqg_print_avg_queue_size(struct cgroup *cgrp, struct cftype *cft,
1470 struct seq_file *sf)
1472 struct blkio_cgroup *blkcg = cgroup_to_blkio_cgroup(cgrp);
1474 blkcg_print_blkgs(sf, blkcg, cfqg_prfill_avg_queue_size,
1475 BLKIO_POLICY_PROP, 0, false);
1478 #endif /* CONFIG_DEBUG_BLK_CGROUP */
1480 static struct cftype cfq_blkcg_files[] = {
1482 .name = "weight_device",
1483 .read_seq_string = blkcg_print_weight_device,
1484 .write_string = blkcg_set_weight_device,
1485 .max_write_len = 256,
1489 .read_seq_string = blkcg_print_weight,
1490 .write_u64 = blkcg_set_weight,
1494 .private = BLKCG_STAT_PRIV(BLKIO_POLICY_PROP,
1495 offsetof(struct cfq_group, stats.time)),
1496 .read_seq_string = blkcg_print_stat,
1500 .private = BLKCG_STAT_PRIV(BLKIO_POLICY_PROP,
1501 offsetof(struct cfq_group, stats.sectors)),
1502 .read_seq_string = blkcg_print_stat,
1505 .name = "io_service_bytes",
1506 .private = BLKCG_STAT_PRIV(BLKIO_POLICY_PROP,
1507 offsetof(struct cfq_group, stats.service_bytes)),
1508 .read_seq_string = blkcg_print_rwstat,
1511 .name = "io_serviced",
1512 .private = BLKCG_STAT_PRIV(BLKIO_POLICY_PROP,
1513 offsetof(struct cfq_group, stats.serviced)),
1514 .read_seq_string = blkcg_print_rwstat,
1517 .name = "io_service_time",
1518 .private = BLKCG_STAT_PRIV(BLKIO_POLICY_PROP,
1519 offsetof(struct cfq_group, stats.service_time)),
1520 .read_seq_string = blkcg_print_rwstat,
1523 .name = "io_wait_time",
1524 .private = BLKCG_STAT_PRIV(BLKIO_POLICY_PROP,
1525 offsetof(struct cfq_group, stats.wait_time)),
1526 .read_seq_string = blkcg_print_rwstat,
1529 .name = "io_merged",
1530 .private = BLKCG_STAT_PRIV(BLKIO_POLICY_PROP,
1531 offsetof(struct cfq_group, stats.merged)),
1532 .read_seq_string = blkcg_print_rwstat,
1535 .name = "io_queued",
1536 .private = BLKCG_STAT_PRIV(BLKIO_POLICY_PROP,
1537 offsetof(struct cfq_group, stats.queued)),
1538 .read_seq_string = blkcg_print_rwstat,
1540 #ifdef CONFIG_DEBUG_BLK_CGROUP
1542 .name = "avg_queue_size",
1543 .read_seq_string = cfqg_print_avg_queue_size,
1546 .name = "group_wait_time",
1547 .private = BLKCG_STAT_PRIV(BLKIO_POLICY_PROP,
1548 offsetof(struct cfq_group, stats.group_wait_time)),
1549 .read_seq_string = blkcg_print_stat,
1552 .name = "idle_time",
1553 .private = BLKCG_STAT_PRIV(BLKIO_POLICY_PROP,
1554 offsetof(struct cfq_group, stats.idle_time)),
1555 .read_seq_string = blkcg_print_stat,
1558 .name = "empty_time",
1559 .private = BLKCG_STAT_PRIV(BLKIO_POLICY_PROP,
1560 offsetof(struct cfq_group, stats.empty_time)),
1561 .read_seq_string = blkcg_print_stat,
1565 .private = BLKCG_STAT_PRIV(BLKIO_POLICY_PROP,
1566 offsetof(struct cfq_group, stats.dequeue)),
1567 .read_seq_string = blkcg_print_stat,
1570 .name = "unaccounted_time",
1571 .private = BLKCG_STAT_PRIV(BLKIO_POLICY_PROP,
1572 offsetof(struct cfq_group, stats.unaccounted_time)),
1573 .read_seq_string = blkcg_print_stat,
1575 #endif /* CONFIG_DEBUG_BLK_CGROUP */
1578 #else /* GROUP_IOSCHED */
1579 static struct cfq_group *cfq_lookup_create_cfqg(struct cfq_data *cfqd,
1580 struct blkio_cgroup *blkcg)
1582 return cfqd->root_group;
1586 cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) {
1590 #endif /* GROUP_IOSCHED */
1593 * The cfqd->service_trees holds all pending cfq_queue's that have
1594 * requests waiting to be processed. It is sorted in the order that
1595 * we will service the queues.
1597 static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1600 struct rb_node **p, *parent;
1601 struct cfq_queue *__cfqq;
1602 unsigned long rb_key;
1603 struct cfq_rb_root *service_tree;
1607 service_tree = service_tree_for(cfqq->cfqg, cfqq_prio(cfqq),
1609 if (cfq_class_idle(cfqq)) {
1610 rb_key = CFQ_IDLE_DELAY;
1611 parent = rb_last(&service_tree->rb);
1612 if (parent && parent != &cfqq->rb_node) {
1613 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1614 rb_key += __cfqq->rb_key;
1617 } else if (!add_front) {
1619 * Get our rb key offset. Subtract any residual slice
1620 * value carried from last service. A negative resid
1621 * count indicates slice overrun, and this should position
1622 * the next service time further away in the tree.
1624 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
1625 rb_key -= cfqq->slice_resid;
1626 cfqq->slice_resid = 0;
1629 __cfqq = cfq_rb_first(service_tree);
1630 rb_key += __cfqq ? __cfqq->rb_key : jiffies;
1633 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1636 * same position, nothing more to do
1638 if (rb_key == cfqq->rb_key &&
1639 cfqq->service_tree == service_tree)
1642 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1643 cfqq->service_tree = NULL;
1648 cfqq->service_tree = service_tree;
1649 p = &service_tree->rb.rb_node;
1654 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1657 * sort by key, that represents service time.
1659 if (time_before(rb_key, __cfqq->rb_key))
1662 n = &(*p)->rb_right;
1670 service_tree->left = &cfqq->rb_node;
1672 cfqq->rb_key = rb_key;
1673 rb_link_node(&cfqq->rb_node, parent, p);
1674 rb_insert_color(&cfqq->rb_node, &service_tree->rb);
1675 service_tree->count++;
1676 if (add_front || !new_cfqq)
1678 cfq_group_notify_queue_add(cfqd, cfqq->cfqg);
1681 static struct cfq_queue *
1682 cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
1683 sector_t sector, struct rb_node **ret_parent,
1684 struct rb_node ***rb_link)
1686 struct rb_node **p, *parent;
1687 struct cfq_queue *cfqq = NULL;
1695 cfqq = rb_entry(parent, struct cfq_queue, p_node);
1698 * Sort strictly based on sector. Smallest to the left,
1699 * largest to the right.
1701 if (sector > blk_rq_pos(cfqq->next_rq))
1702 n = &(*p)->rb_right;
1703 else if (sector < blk_rq_pos(cfqq->next_rq))
1711 *ret_parent = parent;
1717 static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1719 struct rb_node **p, *parent;
1720 struct cfq_queue *__cfqq;
1723 rb_erase(&cfqq->p_node, cfqq->p_root);
1724 cfqq->p_root = NULL;
1727 if (cfq_class_idle(cfqq))
1732 cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
1733 __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
1734 blk_rq_pos(cfqq->next_rq), &parent, &p);
1736 rb_link_node(&cfqq->p_node, parent, p);
1737 rb_insert_color(&cfqq->p_node, cfqq->p_root);
1739 cfqq->p_root = NULL;
1743 * Update cfqq's position in the service tree.
1745 static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1748 * Resorting requires the cfqq to be on the RR list already.
1750 if (cfq_cfqq_on_rr(cfqq)) {
1751 cfq_service_tree_add(cfqd, cfqq, 0);
1752 cfq_prio_tree_add(cfqd, cfqq);
1757 * add to busy list of queues for service, trying to be fair in ordering
1758 * the pending list according to last request service
1760 static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1762 cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
1763 BUG_ON(cfq_cfqq_on_rr(cfqq));
1764 cfq_mark_cfqq_on_rr(cfqq);
1765 cfqd->busy_queues++;
1766 if (cfq_cfqq_sync(cfqq))
1767 cfqd->busy_sync_queues++;
1769 cfq_resort_rr_list(cfqd, cfqq);
1773 * Called when the cfqq no longer has requests pending, remove it from
1776 static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1778 cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
1779 BUG_ON(!cfq_cfqq_on_rr(cfqq));
1780 cfq_clear_cfqq_on_rr(cfqq);
1782 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1783 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1784 cfqq->service_tree = NULL;
1787 rb_erase(&cfqq->p_node, cfqq->p_root);
1788 cfqq->p_root = NULL;
1791 cfq_group_notify_queue_del(cfqd, cfqq->cfqg);
1792 BUG_ON(!cfqd->busy_queues);
1793 cfqd->busy_queues--;
1794 if (cfq_cfqq_sync(cfqq))
1795 cfqd->busy_sync_queues--;
1799 * rb tree support functions
1801 static void cfq_del_rq_rb(struct request *rq)
1803 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1804 const int sync = rq_is_sync(rq);
1806 BUG_ON(!cfqq->queued[sync]);
1807 cfqq->queued[sync]--;
1809 elv_rb_del(&cfqq->sort_list, rq);
1811 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) {
1813 * Queue will be deleted from service tree when we actually
1814 * expire it later. Right now just remove it from prio tree
1818 rb_erase(&cfqq->p_node, cfqq->p_root);
1819 cfqq->p_root = NULL;
1824 static void cfq_add_rq_rb(struct request *rq)
1826 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1827 struct cfq_data *cfqd = cfqq->cfqd;
1828 struct request *prev;
1830 cfqq->queued[rq_is_sync(rq)]++;
1832 elv_rb_add(&cfqq->sort_list, rq);
1834 if (!cfq_cfqq_on_rr(cfqq))
1835 cfq_add_cfqq_rr(cfqd, cfqq);
1838 * check if this request is a better next-serve candidate
1840 prev = cfqq->next_rq;
1841 cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq, cfqd->last_position);
1844 * adjust priority tree position, if ->next_rq changes
1846 if (prev != cfqq->next_rq)
1847 cfq_prio_tree_add(cfqd, cfqq);
1849 BUG_ON(!cfqq->next_rq);
1852 static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
1854 elv_rb_del(&cfqq->sort_list, rq);
1855 cfqq->queued[rq_is_sync(rq)]--;
1856 cfqg_stats_update_io_remove(RQ_CFQG(rq), rq->cmd_flags);
1858 cfqg_stats_update_io_add(RQ_CFQG(rq), cfqq->cfqd->serving_group,
1862 static struct request *
1863 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
1865 struct task_struct *tsk = current;
1866 struct cfq_io_cq *cic;
1867 struct cfq_queue *cfqq;
1869 cic = cfq_cic_lookup(cfqd, tsk->io_context);
1873 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
1875 sector_t sector = bio->bi_sector + bio_sectors(bio);
1877 return elv_rb_find(&cfqq->sort_list, sector);
1883 static void cfq_activate_request(struct request_queue *q, struct request *rq)
1885 struct cfq_data *cfqd = q->elevator->elevator_data;
1887 cfqd->rq_in_driver++;
1888 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
1889 cfqd->rq_in_driver);
1891 cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
1894 static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
1896 struct cfq_data *cfqd = q->elevator->elevator_data;
1898 WARN_ON(!cfqd->rq_in_driver);
1899 cfqd->rq_in_driver--;
1900 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
1901 cfqd->rq_in_driver);
1904 static void cfq_remove_request(struct request *rq)
1906 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1908 if (cfqq->next_rq == rq)
1909 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
1911 list_del_init(&rq->queuelist);
1914 cfqq->cfqd->rq_queued--;
1915 cfqg_stats_update_io_remove(RQ_CFQG(rq), rq->cmd_flags);
1916 if (rq->cmd_flags & REQ_PRIO) {
1917 WARN_ON(!cfqq->prio_pending);
1918 cfqq->prio_pending--;
1922 static int cfq_merge(struct request_queue *q, struct request **req,
1925 struct cfq_data *cfqd = q->elevator->elevator_data;
1926 struct request *__rq;
1928 __rq = cfq_find_rq_fmerge(cfqd, bio);
1929 if (__rq && elv_rq_merge_ok(__rq, bio)) {
1931 return ELEVATOR_FRONT_MERGE;
1934 return ELEVATOR_NO_MERGE;
1937 static void cfq_merged_request(struct request_queue *q, struct request *req,
1940 if (type == ELEVATOR_FRONT_MERGE) {
1941 struct cfq_queue *cfqq = RQ_CFQQ(req);
1943 cfq_reposition_rq_rb(cfqq, req);
1947 static void cfq_bio_merged(struct request_queue *q, struct request *req,
1950 cfqg_stats_update_io_merged(RQ_CFQG(req), bio->bi_rw);
1954 cfq_merged_requests(struct request_queue *q, struct request *rq,
1955 struct request *next)
1957 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1958 struct cfq_data *cfqd = q->elevator->elevator_data;
1961 * reposition in fifo if next is older than rq
1963 if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
1964 time_before(rq_fifo_time(next), rq_fifo_time(rq))) {
1965 list_move(&rq->queuelist, &next->queuelist);
1966 rq_set_fifo_time(rq, rq_fifo_time(next));
1969 if (cfqq->next_rq == next)
1971 cfq_remove_request(next);
1972 cfqg_stats_update_io_merged(RQ_CFQG(rq), next->cmd_flags);
1974 cfqq = RQ_CFQQ(next);
1976 * all requests of this queue are merged to other queues, delete it
1977 * from the service tree. If it's the active_queue,
1978 * cfq_dispatch_requests() will choose to expire it or do idle
1980 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list) &&
1981 cfqq != cfqd->active_queue)
1982 cfq_del_cfqq_rr(cfqd, cfqq);
1985 static int cfq_allow_merge(struct request_queue *q, struct request *rq,
1988 struct cfq_data *cfqd = q->elevator->elevator_data;
1989 struct cfq_io_cq *cic;
1990 struct cfq_queue *cfqq;
1993 * Disallow merge of a sync bio into an async request.
1995 if (cfq_bio_sync(bio) && !rq_is_sync(rq))
1999 * Lookup the cfqq that this bio will be queued with and allow
2000 * merge only if rq is queued there.
2002 cic = cfq_cic_lookup(cfqd, current->io_context);
2006 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
2007 return cfqq == RQ_CFQQ(rq);
2010 static inline void cfq_del_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2012 del_timer(&cfqd->idle_slice_timer);
2013 cfqg_stats_update_idle_time(cfqq->cfqg);
2016 static void __cfq_set_active_queue(struct cfq_data *cfqd,
2017 struct cfq_queue *cfqq)
2020 cfq_log_cfqq(cfqd, cfqq, "set_active wl_prio:%d wl_type:%d",
2021 cfqd->serving_prio, cfqd->serving_type);
2022 cfqg_stats_update_avg_queue_size(cfqq->cfqg);
2023 cfqq->slice_start = 0;
2024 cfqq->dispatch_start = jiffies;
2025 cfqq->allocated_slice = 0;
2026 cfqq->slice_end = 0;
2027 cfqq->slice_dispatch = 0;
2028 cfqq->nr_sectors = 0;
2030 cfq_clear_cfqq_wait_request(cfqq);
2031 cfq_clear_cfqq_must_dispatch(cfqq);
2032 cfq_clear_cfqq_must_alloc_slice(cfqq);
2033 cfq_clear_cfqq_fifo_expire(cfqq);
2034 cfq_mark_cfqq_slice_new(cfqq);
2036 cfq_del_timer(cfqd, cfqq);
2039 cfqd->active_queue = cfqq;
2043 * current cfqq expired its slice (or was too idle), select new one
2046 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2049 cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
2051 if (cfq_cfqq_wait_request(cfqq))
2052 cfq_del_timer(cfqd, cfqq);
2054 cfq_clear_cfqq_wait_request(cfqq);
2055 cfq_clear_cfqq_wait_busy(cfqq);
2058 * If this cfqq is shared between multiple processes, check to
2059 * make sure that those processes are still issuing I/Os within
2060 * the mean seek distance. If not, it may be time to break the
2061 * queues apart again.
2063 if (cfq_cfqq_coop(cfqq) && CFQQ_SEEKY(cfqq))
2064 cfq_mark_cfqq_split_coop(cfqq);
2067 * store what was left of this slice, if the queue idled/timed out
2070 if (cfq_cfqq_slice_new(cfqq))
2071 cfqq->slice_resid = cfq_scaled_cfqq_slice(cfqd, cfqq);
2073 cfqq->slice_resid = cfqq->slice_end - jiffies;
2074 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
2077 cfq_group_served(cfqd, cfqq->cfqg, cfqq);
2079 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
2080 cfq_del_cfqq_rr(cfqd, cfqq);
2082 cfq_resort_rr_list(cfqd, cfqq);
2084 if (cfqq == cfqd->active_queue)
2085 cfqd->active_queue = NULL;
2087 if (cfqd->active_cic) {
2088 put_io_context(cfqd->active_cic->icq.ioc);
2089 cfqd->active_cic = NULL;
2093 static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out)
2095 struct cfq_queue *cfqq = cfqd->active_queue;
2098 __cfq_slice_expired(cfqd, cfqq, timed_out);
2102 * Get next queue for service. Unless we have a queue preemption,
2103 * we'll simply select the first cfqq in the service tree.
2105 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
2107 struct cfq_rb_root *service_tree =
2108 service_tree_for(cfqd->serving_group, cfqd->serving_prio,
2109 cfqd->serving_type);
2111 if (!cfqd->rq_queued)
2114 /* There is nothing to dispatch */
2117 if (RB_EMPTY_ROOT(&service_tree->rb))
2119 return cfq_rb_first(service_tree);
2122 static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd)
2124 struct cfq_group *cfqg;
2125 struct cfq_queue *cfqq;
2127 struct cfq_rb_root *st;
2129 if (!cfqd->rq_queued)
2132 cfqg = cfq_get_next_cfqg(cfqd);
2136 for_each_cfqg_st(cfqg, i, j, st)
2137 if ((cfqq = cfq_rb_first(st)) != NULL)
2143 * Get and set a new active queue for service.
2145 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
2146 struct cfq_queue *cfqq)
2149 cfqq = cfq_get_next_queue(cfqd);
2151 __cfq_set_active_queue(cfqd, cfqq);
2155 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
2158 if (blk_rq_pos(rq) >= cfqd->last_position)
2159 return blk_rq_pos(rq) - cfqd->last_position;
2161 return cfqd->last_position - blk_rq_pos(rq);
2164 static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2167 return cfq_dist_from_last(cfqd, rq) <= CFQQ_CLOSE_THR;
2170 static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
2171 struct cfq_queue *cur_cfqq)
2173 struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
2174 struct rb_node *parent, *node;
2175 struct cfq_queue *__cfqq;
2176 sector_t sector = cfqd->last_position;
2178 if (RB_EMPTY_ROOT(root))
2182 * First, if we find a request starting at the end of the last
2183 * request, choose it.
2185 __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
2190 * If the exact sector wasn't found, the parent of the NULL leaf
2191 * will contain the closest sector.
2193 __cfqq = rb_entry(parent, struct cfq_queue, p_node);
2194 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
2197 if (blk_rq_pos(__cfqq->next_rq) < sector)
2198 node = rb_next(&__cfqq->p_node);
2200 node = rb_prev(&__cfqq->p_node);
2204 __cfqq = rb_entry(node, struct cfq_queue, p_node);
2205 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
2213 * cur_cfqq - passed in so that we don't decide that the current queue is
2214 * closely cooperating with itself.
2216 * So, basically we're assuming that that cur_cfqq has dispatched at least
2217 * one request, and that cfqd->last_position reflects a position on the disk
2218 * associated with the I/O issued by cur_cfqq. I'm not sure this is a valid
2221 static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
2222 struct cfq_queue *cur_cfqq)
2224 struct cfq_queue *cfqq;
2226 if (cfq_class_idle(cur_cfqq))
2228 if (!cfq_cfqq_sync(cur_cfqq))
2230 if (CFQQ_SEEKY(cur_cfqq))
2234 * Don't search priority tree if it's the only queue in the group.
2236 if (cur_cfqq->cfqg->nr_cfqq == 1)
2240 * We should notice if some of the queues are cooperating, eg
2241 * working closely on the same area of the disk. In that case,
2242 * we can group them together and don't waste time idling.
2244 cfqq = cfqq_close(cfqd, cur_cfqq);
2248 /* If new queue belongs to different cfq_group, don't choose it */
2249 if (cur_cfqq->cfqg != cfqq->cfqg)
2253 * It only makes sense to merge sync queues.
2255 if (!cfq_cfqq_sync(cfqq))
2257 if (CFQQ_SEEKY(cfqq))
2261 * Do not merge queues of different priority classes
2263 if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq))
2270 * Determine whether we should enforce idle window for this queue.
2273 static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2275 enum wl_prio_t prio = cfqq_prio(cfqq);
2276 struct cfq_rb_root *service_tree = cfqq->service_tree;
2278 BUG_ON(!service_tree);
2279 BUG_ON(!service_tree->count);
2281 if (!cfqd->cfq_slice_idle)
2284 /* We never do for idle class queues. */
2285 if (prio == IDLE_WORKLOAD)
2288 /* We do for queues that were marked with idle window flag. */
2289 if (cfq_cfqq_idle_window(cfqq) &&
2290 !(blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag))
2294 * Otherwise, we do only if they are the last ones
2295 * in their service tree.
2297 if (service_tree->count == 1 && cfq_cfqq_sync(cfqq) &&
2298 !cfq_io_thinktime_big(cfqd, &service_tree->ttime, false))
2300 cfq_log_cfqq(cfqd, cfqq, "Not idling. st->count:%d",
2301 service_tree->count);
2305 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
2307 struct cfq_queue *cfqq = cfqd->active_queue;
2308 struct cfq_io_cq *cic;
2309 unsigned long sl, group_idle = 0;
2312 * SSD device without seek penalty, disable idling. But only do so
2313 * for devices that support queuing, otherwise we still have a problem
2314 * with sync vs async workloads.
2316 if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
2319 WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
2320 WARN_ON(cfq_cfqq_slice_new(cfqq));
2323 * idle is disabled, either manually or by past process history
2325 if (!cfq_should_idle(cfqd, cfqq)) {
2326 /* no queue idling. Check for group idling */
2327 if (cfqd->cfq_group_idle)
2328 group_idle = cfqd->cfq_group_idle;
2334 * still active requests from this queue, don't idle
2336 if (cfqq->dispatched)
2340 * task has exited, don't wait
2342 cic = cfqd->active_cic;
2343 if (!cic || !atomic_read(&cic->icq.ioc->active_ref))
2347 * If our average think time is larger than the remaining time
2348 * slice, then don't idle. This avoids overrunning the allotted
2351 if (sample_valid(cic->ttime.ttime_samples) &&
2352 (cfqq->slice_end - jiffies < cic->ttime.ttime_mean)) {
2353 cfq_log_cfqq(cfqd, cfqq, "Not idling. think_time:%lu",
2354 cic->ttime.ttime_mean);
2358 /* There are other queues in the group, don't do group idle */
2359 if (group_idle && cfqq->cfqg->nr_cfqq > 1)
2362 cfq_mark_cfqq_wait_request(cfqq);
2365 sl = cfqd->cfq_group_idle;
2367 sl = cfqd->cfq_slice_idle;
2369 mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
2370 cfqg_stats_set_start_idle_time(cfqq->cfqg);
2371 cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu group_idle: %d", sl,
2372 group_idle ? 1 : 0);
2376 * Move request from internal lists to the request queue dispatch list.
2378 static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
2380 struct cfq_data *cfqd = q->elevator->elevator_data;
2381 struct cfq_queue *cfqq = RQ_CFQQ(rq);
2383 cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
2385 cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
2386 cfq_remove_request(rq);
2388 (RQ_CFQG(rq))->dispatched++;
2389 elv_dispatch_sort(q, rq);
2391 cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]++;
2392 cfqq->nr_sectors += blk_rq_sectors(rq);
2393 cfqg_stats_update_dispatch(cfqq->cfqg, blk_rq_bytes(rq), rq->cmd_flags);
2397 * return expired entry, or NULL to just start from scratch in rbtree
2399 static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
2401 struct request *rq = NULL;
2403 if (cfq_cfqq_fifo_expire(cfqq))
2406 cfq_mark_cfqq_fifo_expire(cfqq);
2408 if (list_empty(&cfqq->fifo))
2411 rq = rq_entry_fifo(cfqq->fifo.next);
2412 if (time_before(jiffies, rq_fifo_time(rq)))
2415 cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq);
2420 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2422 const int base_rq = cfqd->cfq_slice_async_rq;
2424 WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
2426 return 2 * base_rq * (IOPRIO_BE_NR - cfqq->ioprio);
2430 * Must be called with the queue_lock held.
2432 static int cfqq_process_refs(struct cfq_queue *cfqq)
2434 int process_refs, io_refs;
2436 io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE];
2437 process_refs = cfqq->ref - io_refs;
2438 BUG_ON(process_refs < 0);
2439 return process_refs;
2442 static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq)
2444 int process_refs, new_process_refs;
2445 struct cfq_queue *__cfqq;
2448 * If there are no process references on the new_cfqq, then it is
2449 * unsafe to follow the ->new_cfqq chain as other cfqq's in the
2450 * chain may have dropped their last reference (not just their
2451 * last process reference).
2453 if (!cfqq_process_refs(new_cfqq))
2456 /* Avoid a circular list and skip interim queue merges */
2457 while ((__cfqq = new_cfqq->new_cfqq)) {
2463 process_refs = cfqq_process_refs(cfqq);
2464 new_process_refs = cfqq_process_refs(new_cfqq);
2466 * If the process for the cfqq has gone away, there is no
2467 * sense in merging the queues.
2469 if (process_refs == 0 || new_process_refs == 0)
2473 * Merge in the direction of the lesser amount of work.
2475 if (new_process_refs >= process_refs) {
2476 cfqq->new_cfqq = new_cfqq;
2477 new_cfqq->ref += process_refs;
2479 new_cfqq->new_cfqq = cfqq;
2480 cfqq->ref += new_process_refs;
2484 static enum wl_type_t cfq_choose_wl(struct cfq_data *cfqd,
2485 struct cfq_group *cfqg, enum wl_prio_t prio)
2487 struct cfq_queue *queue;
2489 bool key_valid = false;
2490 unsigned long lowest_key = 0;
2491 enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD;
2493 for (i = 0; i <= SYNC_WORKLOAD; ++i) {
2494 /* select the one with lowest rb_key */
2495 queue = cfq_rb_first(service_tree_for(cfqg, prio, i));
2497 (!key_valid || time_before(queue->rb_key, lowest_key))) {
2498 lowest_key = queue->rb_key;
2507 static void choose_service_tree(struct cfq_data *cfqd, struct cfq_group *cfqg)
2511 struct cfq_rb_root *st;
2512 unsigned group_slice;
2513 enum wl_prio_t original_prio = cfqd->serving_prio;
2515 /* Choose next priority. RT > BE > IDLE */
2516 if (cfq_group_busy_queues_wl(RT_WORKLOAD, cfqd, cfqg))
2517 cfqd->serving_prio = RT_WORKLOAD;
2518 else if (cfq_group_busy_queues_wl(BE_WORKLOAD, cfqd, cfqg))
2519 cfqd->serving_prio = BE_WORKLOAD;
2521 cfqd->serving_prio = IDLE_WORKLOAD;
2522 cfqd->workload_expires = jiffies + 1;
2526 if (original_prio != cfqd->serving_prio)
2530 * For RT and BE, we have to choose also the type
2531 * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
2534 st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
2538 * check workload expiration, and that we still have other queues ready
2540 if (count && !time_after(jiffies, cfqd->workload_expires))
2544 /* otherwise select new workload type */
2545 cfqd->serving_type =
2546 cfq_choose_wl(cfqd, cfqg, cfqd->serving_prio);
2547 st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
2551 * the workload slice is computed as a fraction of target latency
2552 * proportional to the number of queues in that workload, over
2553 * all the queues in the same priority class
2555 group_slice = cfq_group_slice(cfqd, cfqg);
2557 slice = group_slice * count /
2558 max_t(unsigned, cfqg->busy_queues_avg[cfqd->serving_prio],
2559 cfq_group_busy_queues_wl(cfqd->serving_prio, cfqd, cfqg));
2561 if (cfqd->serving_type == ASYNC_WORKLOAD) {
2565 * Async queues are currently system wide. Just taking
2566 * proportion of queues with-in same group will lead to higher
2567 * async ratio system wide as generally root group is going
2568 * to have higher weight. A more accurate thing would be to
2569 * calculate system wide asnc/sync ratio.
2571 tmp = cfq_target_latency * cfqg_busy_async_queues(cfqd, cfqg);
2572 tmp = tmp/cfqd->busy_queues;
2573 slice = min_t(unsigned, slice, tmp);
2575 /* async workload slice is scaled down according to
2576 * the sync/async slice ratio. */
2577 slice = slice * cfqd->cfq_slice[0] / cfqd->cfq_slice[1];
2579 /* sync workload slice is at least 2 * cfq_slice_idle */
2580 slice = max(slice, 2 * cfqd->cfq_slice_idle);
2582 slice = max_t(unsigned, slice, CFQ_MIN_TT);
2583 cfq_log(cfqd, "workload slice:%d", slice);
2584 cfqd->workload_expires = jiffies + slice;
2587 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd)
2589 struct cfq_rb_root *st = &cfqd->grp_service_tree;
2590 struct cfq_group *cfqg;
2592 if (RB_EMPTY_ROOT(&st->rb))
2594 cfqg = cfq_rb_first_group(st);
2595 update_min_vdisktime(st);
2599 static void cfq_choose_cfqg(struct cfq_data *cfqd)
2601 struct cfq_group *cfqg = cfq_get_next_cfqg(cfqd);
2603 cfqd->serving_group = cfqg;
2605 /* Restore the workload type data */
2606 if (cfqg->saved_workload_slice) {
2607 cfqd->workload_expires = jiffies + cfqg->saved_workload_slice;
2608 cfqd->serving_type = cfqg->saved_workload;
2609 cfqd->serving_prio = cfqg->saved_serving_prio;
2611 cfqd->workload_expires = jiffies - 1;
2613 choose_service_tree(cfqd, cfqg);
2617 * Select a queue for service. If we have a current active queue,
2618 * check whether to continue servicing it, or retrieve and set a new one.
2620 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
2622 struct cfq_queue *cfqq, *new_cfqq = NULL;
2624 cfqq = cfqd->active_queue;
2628 if (!cfqd->rq_queued)
2632 * We were waiting for group to get backlogged. Expire the queue
2634 if (cfq_cfqq_wait_busy(cfqq) && !RB_EMPTY_ROOT(&cfqq->sort_list))
2638 * The active queue has run out of time, expire it and select new.
2640 if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq)) {
2642 * If slice had not expired at the completion of last request
2643 * we might not have turned on wait_busy flag. Don't expire
2644 * the queue yet. Allow the group to get backlogged.
2646 * The very fact that we have used the slice, that means we
2647 * have been idling all along on this queue and it should be
2648 * ok to wait for this request to complete.
2650 if (cfqq->cfqg->nr_cfqq == 1 && RB_EMPTY_ROOT(&cfqq->sort_list)
2651 && cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2655 goto check_group_idle;
2659 * The active queue has requests and isn't expired, allow it to
2662 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
2666 * If another queue has a request waiting within our mean seek
2667 * distance, let it run. The expire code will check for close
2668 * cooperators and put the close queue at the front of the service
2669 * tree. If possible, merge the expiring queue with the new cfqq.
2671 new_cfqq = cfq_close_cooperator(cfqd, cfqq);
2673 if (!cfqq->new_cfqq)
2674 cfq_setup_merge(cfqq, new_cfqq);
2679 * No requests pending. If the active queue still has requests in
2680 * flight or is idling for a new request, allow either of these
2681 * conditions to happen (or time out) before selecting a new queue.
2683 if (timer_pending(&cfqd->idle_slice_timer)) {
2689 * This is a deep seek queue, but the device is much faster than
2690 * the queue can deliver, don't idle
2692 if (CFQQ_SEEKY(cfqq) && cfq_cfqq_idle_window(cfqq) &&
2693 (cfq_cfqq_slice_new(cfqq) ||
2694 (cfqq->slice_end - jiffies > jiffies - cfqq->slice_start))) {
2695 cfq_clear_cfqq_deep(cfqq);
2696 cfq_clear_cfqq_idle_window(cfqq);
2699 if (cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2705 * If group idle is enabled and there are requests dispatched from
2706 * this group, wait for requests to complete.
2709 if (cfqd->cfq_group_idle && cfqq->cfqg->nr_cfqq == 1 &&
2710 cfqq->cfqg->dispatched &&
2711 !cfq_io_thinktime_big(cfqd, &cfqq->cfqg->ttime, true)) {
2717 cfq_slice_expired(cfqd, 0);
2720 * Current queue expired. Check if we have to switch to a new
2724 cfq_choose_cfqg(cfqd);
2726 cfqq = cfq_set_active_queue(cfqd, new_cfqq);
2731 static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
2735 while (cfqq->next_rq) {
2736 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
2740 BUG_ON(!list_empty(&cfqq->fifo));
2742 /* By default cfqq is not expired if it is empty. Do it explicitly */
2743 __cfq_slice_expired(cfqq->cfqd, cfqq, 0);
2748 * Drain our current requests. Used for barriers and when switching
2749 * io schedulers on-the-fly.
2751 static int cfq_forced_dispatch(struct cfq_data *cfqd)
2753 struct cfq_queue *cfqq;
2756 /* Expire the timeslice of the current active queue first */
2757 cfq_slice_expired(cfqd, 0);
2758 while ((cfqq = cfq_get_next_queue_forced(cfqd)) != NULL) {
2759 __cfq_set_active_queue(cfqd, cfqq);
2760 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
2763 BUG_ON(cfqd->busy_queues);
2765 cfq_log(cfqd, "forced_dispatch=%d", dispatched);
2769 static inline bool cfq_slice_used_soon(struct cfq_data *cfqd,
2770 struct cfq_queue *cfqq)
2772 /* the queue hasn't finished any request, can't estimate */
2773 if (cfq_cfqq_slice_new(cfqq))
2775 if (time_after(jiffies + cfqd->cfq_slice_idle * cfqq->dispatched,
2782 static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2784 unsigned int max_dispatch;
2787 * Drain async requests before we start sync IO
2789 if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_flight[BLK_RW_ASYNC])
2793 * If this is an async queue and we have sync IO in flight, let it wait
2795 if (cfqd->rq_in_flight[BLK_RW_SYNC] && !cfq_cfqq_sync(cfqq))
2798 max_dispatch = max_t(unsigned int, cfqd->cfq_quantum / 2, 1);
2799 if (cfq_class_idle(cfqq))
2803 * Does this cfqq already have too much IO in flight?
2805 if (cfqq->dispatched >= max_dispatch) {
2806 bool promote_sync = false;
2808 * idle queue must always only have a single IO in flight
2810 if (cfq_class_idle(cfqq))
2814 * If there is only one sync queue
2815 * we can ignore async queue here and give the sync
2816 * queue no dispatch limit. The reason is a sync queue can
2817 * preempt async queue, limiting the sync queue doesn't make
2818 * sense. This is useful for aiostress test.
2820 if (cfq_cfqq_sync(cfqq) && cfqd->busy_sync_queues == 1)
2821 promote_sync = true;
2824 * We have other queues, don't allow more IO from this one
2826 if (cfqd->busy_queues > 1 && cfq_slice_used_soon(cfqd, cfqq) &&
2831 * Sole queue user, no limit
2833 if (cfqd->busy_queues == 1 || promote_sync)
2837 * Normally we start throttling cfqq when cfq_quantum/2
2838 * requests have been dispatched. But we can drive
2839 * deeper queue depths at the beginning of slice
2840 * subjected to upper limit of cfq_quantum.
2842 max_dispatch = cfqd->cfq_quantum;
2846 * Async queues must wait a bit before being allowed dispatch.
2847 * We also ramp up the dispatch depth gradually for async IO,
2848 * based on the last sync IO we serviced
2850 if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
2851 unsigned long last_sync = jiffies - cfqd->last_delayed_sync;
2854 depth = last_sync / cfqd->cfq_slice[1];
2855 if (!depth && !cfqq->dispatched)
2857 if (depth < max_dispatch)
2858 max_dispatch = depth;
2862 * If we're below the current max, allow a dispatch
2864 return cfqq->dispatched < max_dispatch;
2868 * Dispatch a request from cfqq, moving them to the request queue
2871 static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2875 BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
2877 if (!cfq_may_dispatch(cfqd, cfqq))
2881 * follow expired path, else get first next available
2883 rq = cfq_check_fifo(cfqq);
2888 * insert request into driver dispatch list
2890 cfq_dispatch_insert(cfqd->queue, rq);
2892 if (!cfqd->active_cic) {
2893 struct cfq_io_cq *cic = RQ_CIC(rq);
2895 atomic_long_inc(&cic->icq.ioc->refcount);
2896 cfqd->active_cic = cic;
2903 * Find the cfqq that we need to service and move a request from that to the
2906 static int cfq_dispatch_requests(struct request_queue *q, int force)
2908 struct cfq_data *cfqd = q->elevator->elevator_data;
2909 struct cfq_queue *cfqq;
2911 if (!cfqd->busy_queues)
2914 if (unlikely(force))
2915 return cfq_forced_dispatch(cfqd);
2917 cfqq = cfq_select_queue(cfqd);
2922 * Dispatch a request from this cfqq, if it is allowed
2924 if (!cfq_dispatch_request(cfqd, cfqq))
2927 cfqq->slice_dispatch++;
2928 cfq_clear_cfqq_must_dispatch(cfqq);
2931 * expire an async queue immediately if it has used up its slice. idle
2932 * queue always expire after 1 dispatch round.
2934 if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
2935 cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
2936 cfq_class_idle(cfqq))) {
2937 cfqq->slice_end = jiffies + 1;
2938 cfq_slice_expired(cfqd, 0);
2941 cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
2946 * task holds one reference to the queue, dropped when task exits. each rq
2947 * in-flight on this queue also holds a reference, dropped when rq is freed.
2949 * Each cfq queue took a reference on the parent group. Drop it now.
2950 * queue lock must be held here.
2952 static void cfq_put_queue(struct cfq_queue *cfqq)
2954 struct cfq_data *cfqd = cfqq->cfqd;
2955 struct cfq_group *cfqg;
2957 BUG_ON(cfqq->ref <= 0);
2963 cfq_log_cfqq(cfqd, cfqq, "put_queue");
2964 BUG_ON(rb_first(&cfqq->sort_list));
2965 BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
2968 if (unlikely(cfqd->active_queue == cfqq)) {
2969 __cfq_slice_expired(cfqd, cfqq, 0);
2970 cfq_schedule_dispatch(cfqd);
2973 BUG_ON(cfq_cfqq_on_rr(cfqq));
2974 kmem_cache_free(cfq_pool, cfqq);
2978 static void cfq_put_cooperator(struct cfq_queue *cfqq)
2980 struct cfq_queue *__cfqq, *next;
2983 * If this queue was scheduled to merge with another queue, be
2984 * sure to drop the reference taken on that queue (and others in
2985 * the merge chain). See cfq_setup_merge and cfq_merge_cfqqs.
2987 __cfqq = cfqq->new_cfqq;
2989 if (__cfqq == cfqq) {
2990 WARN(1, "cfqq->new_cfqq loop detected\n");
2993 next = __cfqq->new_cfqq;
2994 cfq_put_queue(__cfqq);
2999 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3001 if (unlikely(cfqq == cfqd->active_queue)) {
3002 __cfq_slice_expired(cfqd, cfqq, 0);
3003 cfq_schedule_dispatch(cfqd);
3006 cfq_put_cooperator(cfqq);
3008 cfq_put_queue(cfqq);
3011 static void cfq_init_icq(struct io_cq *icq)
3013 struct cfq_io_cq *cic = icq_to_cic(icq);
3015 cic->ttime.last_end_request = jiffies;
3018 static void cfq_exit_icq(struct io_cq *icq)
3020 struct cfq_io_cq *cic = icq_to_cic(icq);
3021 struct cfq_data *cfqd = cic_to_cfqd(cic);
3023 if (cic->cfqq[BLK_RW_ASYNC]) {
3024 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]);
3025 cic->cfqq[BLK_RW_ASYNC] = NULL;
3028 if (cic->cfqq[BLK_RW_SYNC]) {
3029 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]);
3030 cic->cfqq[BLK_RW_SYNC] = NULL;
3034 static void cfq_init_prio_data(struct cfq_queue *cfqq, struct cfq_io_cq *cic)
3036 struct task_struct *tsk = current;
3039 if (!cfq_cfqq_prio_changed(cfqq))
3042 ioprio_class = IOPRIO_PRIO_CLASS(cic->ioprio);
3043 switch (ioprio_class) {
3045 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
3046 case IOPRIO_CLASS_NONE:
3048 * no prio set, inherit CPU scheduling settings
3050 cfqq->ioprio = task_nice_ioprio(tsk);
3051 cfqq->ioprio_class = task_nice_ioclass(tsk);
3053 case IOPRIO_CLASS_RT:
3054 cfqq->ioprio = IOPRIO_PRIO_DATA(cic->ioprio);
3055 cfqq->ioprio_class = IOPRIO_CLASS_RT;
3057 case IOPRIO_CLASS_BE:
3058 cfqq->ioprio = IOPRIO_PRIO_DATA(cic->ioprio);
3059 cfqq->ioprio_class = IOPRIO_CLASS_BE;
3061 case IOPRIO_CLASS_IDLE:
3062 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
3064 cfq_clear_cfqq_idle_window(cfqq);
3069 * keep track of original prio settings in case we have to temporarily
3070 * elevate the priority of this queue
3072 cfqq->org_ioprio = cfqq->ioprio;
3073 cfq_clear_cfqq_prio_changed(cfqq);
3076 static void check_ioprio_changed(struct cfq_io_cq *cic, struct bio *bio)
3078 int ioprio = cic->icq.ioc->ioprio;
3079 struct cfq_data *cfqd = cic_to_cfqd(cic);
3080 struct cfq_queue *cfqq;
3083 * Check whether ioprio has changed. The condition may trigger
3084 * spuriously on a newly created cic but there's no harm.
3086 if (unlikely(!cfqd) || likely(cic->ioprio == ioprio))
3089 cfqq = cic->cfqq[BLK_RW_ASYNC];
3091 struct cfq_queue *new_cfqq;
3092 new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic, bio,
3095 cic->cfqq[BLK_RW_ASYNC] = new_cfqq;
3096 cfq_put_queue(cfqq);
3100 cfqq = cic->cfqq[BLK_RW_SYNC];
3102 cfq_mark_cfqq_prio_changed(cfqq);
3104 cic->ioprio = ioprio;
3107 static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3108 pid_t pid, bool is_sync)
3110 RB_CLEAR_NODE(&cfqq->rb_node);
3111 RB_CLEAR_NODE(&cfqq->p_node);
3112 INIT_LIST_HEAD(&cfqq->fifo);
3117 cfq_mark_cfqq_prio_changed(cfqq);
3120 if (!cfq_class_idle(cfqq))
3121 cfq_mark_cfqq_idle_window(cfqq);
3122 cfq_mark_cfqq_sync(cfqq);
3127 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3128 static void check_blkcg_changed(struct cfq_io_cq *cic, struct bio *bio)
3130 struct cfq_data *cfqd = cic_to_cfqd(cic);
3131 struct cfq_queue *sync_cfqq;
3135 id = bio_blkio_cgroup(bio)->id;
3139 * Check whether blkcg has changed. The condition may trigger
3140 * spuriously on a newly created cic but there's no harm.
3142 if (unlikely(!cfqd) || likely(cic->blkcg_id == id))
3145 sync_cfqq = cic_to_cfqq(cic, 1);
3148 * Drop reference to sync queue. A new sync queue will be
3149 * assigned in new group upon arrival of a fresh request.
3151 cfq_log_cfqq(cfqd, sync_cfqq, "changed cgroup");
3152 cic_set_cfqq(cic, NULL, 1);
3153 cfq_put_queue(sync_cfqq);
3159 static inline void check_blkcg_changed(struct cfq_io_cq *cic, struct bio *bio) { }
3160 #endif /* CONFIG_CFQ_GROUP_IOSCHED */
3162 static struct cfq_queue *
3163 cfq_find_alloc_queue(struct cfq_data *cfqd, bool is_sync, struct cfq_io_cq *cic,
3164 struct bio *bio, gfp_t gfp_mask)
3166 struct blkio_cgroup *blkcg;
3167 struct cfq_queue *cfqq, *new_cfqq = NULL;
3168 struct cfq_group *cfqg;
3173 blkcg = bio_blkio_cgroup(bio);
3174 cfqg = cfq_lookup_create_cfqg(cfqd, blkcg);
3175 cfqq = cic_to_cfqq(cic, is_sync);
3178 * Always try a new alloc if we fell back to the OOM cfqq
3179 * originally, since it should just be a temporary situation.
3181 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
3186 } else if (gfp_mask & __GFP_WAIT) {
3188 spin_unlock_irq(cfqd->queue->queue_lock);
3189 new_cfqq = kmem_cache_alloc_node(cfq_pool,
3190 gfp_mask | __GFP_ZERO,
3192 spin_lock_irq(cfqd->queue->queue_lock);
3196 cfqq = kmem_cache_alloc_node(cfq_pool,
3197 gfp_mask | __GFP_ZERO,
3202 cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
3203 cfq_init_prio_data(cfqq, cic);
3204 cfq_link_cfqq_cfqg(cfqq, cfqg);
3205 cfq_log_cfqq(cfqd, cfqq, "alloced");
3207 cfqq = &cfqd->oom_cfqq;
3211 kmem_cache_free(cfq_pool, new_cfqq);
3217 static struct cfq_queue **
3218 cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
3220 switch (ioprio_class) {
3221 case IOPRIO_CLASS_RT:
3222 return &cfqd->async_cfqq[0][ioprio];
3223 case IOPRIO_CLASS_NONE:
3224 ioprio = IOPRIO_NORM;
3226 case IOPRIO_CLASS_BE:
3227 return &cfqd->async_cfqq[1][ioprio];
3228 case IOPRIO_CLASS_IDLE:
3229 return &cfqd->async_idle_cfqq;
3235 static struct cfq_queue *
3236 cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct cfq_io_cq *cic,
3237 struct bio *bio, gfp_t gfp_mask)
3239 const int ioprio_class = IOPRIO_PRIO_CLASS(cic->ioprio);
3240 const int ioprio = IOPRIO_PRIO_DATA(cic->ioprio);
3241 struct cfq_queue **async_cfqq = NULL;
3242 struct cfq_queue *cfqq = NULL;
3245 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
3250 cfqq = cfq_find_alloc_queue(cfqd, is_sync, cic, bio, gfp_mask);
3253 * pin the queue now that it's allocated, scheduler exit will prune it
3255 if (!is_sync && !(*async_cfqq)) {
3265 __cfq_update_io_thinktime(struct cfq_ttime *ttime, unsigned long slice_idle)
3267 unsigned long elapsed = jiffies - ttime->last_end_request;
3268 elapsed = min(elapsed, 2UL * slice_idle);
3270 ttime->ttime_samples = (7*ttime->ttime_samples + 256) / 8;
3271 ttime->ttime_total = (7*ttime->ttime_total + 256*elapsed) / 8;
3272 ttime->ttime_mean = (ttime->ttime_total + 128) / ttime->ttime_samples;
3276 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3277 struct cfq_io_cq *cic)
3279 if (cfq_cfqq_sync(cfqq)) {
3280 __cfq_update_io_thinktime(&cic->ttime, cfqd->cfq_slice_idle);
3281 __cfq_update_io_thinktime(&cfqq->service_tree->ttime,
3282 cfqd->cfq_slice_idle);
3284 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3285 __cfq_update_io_thinktime(&cfqq->cfqg->ttime, cfqd->cfq_group_idle);
3290 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3294 sector_t n_sec = blk_rq_sectors(rq);
3295 if (cfqq->last_request_pos) {
3296 if (cfqq->last_request_pos < blk_rq_pos(rq))
3297 sdist = blk_rq_pos(rq) - cfqq->last_request_pos;
3299 sdist = cfqq->last_request_pos - blk_rq_pos(rq);
3302 cfqq->seek_history <<= 1;
3303 if (blk_queue_nonrot(cfqd->queue))
3304 cfqq->seek_history |= (n_sec < CFQQ_SECT_THR_NONROT);
3306 cfqq->seek_history |= (sdist > CFQQ_SEEK_THR);
3310 * Disable idle window if the process thinks too long or seeks so much that
3314 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3315 struct cfq_io_cq *cic)
3317 int old_idle, enable_idle;
3320 * Don't idle for async or idle io prio class
3322 if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
3325 enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
3327 if (cfqq->queued[0] + cfqq->queued[1] >= 4)
3328 cfq_mark_cfqq_deep(cfqq);
3330 if (cfqq->next_rq && (cfqq->next_rq->cmd_flags & REQ_NOIDLE))
3332 else if (!atomic_read(&cic->icq.ioc->active_ref) ||
3333 !cfqd->cfq_slice_idle ||
3334 (!cfq_cfqq_deep(cfqq) && CFQQ_SEEKY(cfqq)))
3336 else if (sample_valid(cic->ttime.ttime_samples)) {
3337 if (cic->ttime.ttime_mean > cfqd->cfq_slice_idle)
3343 if (old_idle != enable_idle) {
3344 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
3346 cfq_mark_cfqq_idle_window(cfqq);
3348 cfq_clear_cfqq_idle_window(cfqq);
3353 * Check if new_cfqq should preempt the currently active queue. Return 0 for
3354 * no or if we aren't sure, a 1 will cause a preempt.
3357 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
3360 struct cfq_queue *cfqq;
3362 cfqq = cfqd->active_queue;
3366 if (cfq_class_idle(new_cfqq))
3369 if (cfq_class_idle(cfqq))
3373 * Don't allow a non-RT request to preempt an ongoing RT cfqq timeslice.
3375 if (cfq_class_rt(cfqq) && !cfq_class_rt(new_cfqq))
3379 * if the new request is sync, but the currently running queue is
3380 * not, let the sync request have priority.
3382 if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
3385 if (new_cfqq->cfqg != cfqq->cfqg)
3388 if (cfq_slice_used(cfqq))
3391 /* Allow preemption only if we are idling on sync-noidle tree */
3392 if (cfqd->serving_type == SYNC_NOIDLE_WORKLOAD &&
3393 cfqq_type(new_cfqq) == SYNC_NOIDLE_WORKLOAD &&
3394 new_cfqq->service_tree->count == 2 &&
3395 RB_EMPTY_ROOT(&cfqq->sort_list))
3399 * So both queues are sync. Let the new request get disk time if
3400 * it's a metadata request and the current queue is doing regular IO.
3402 if ((rq->cmd_flags & REQ_PRIO) && !cfqq->prio_pending)
3406 * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
3408 if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
3411 /* An idle queue should not be idle now for some reason */
3412 if (RB_EMPTY_ROOT(&cfqq->sort_list) && !cfq_should_idle(cfqd, cfqq))
3415 if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
3419 * if this request is as-good as one we would expect from the
3420 * current cfqq, let it preempt
3422 if (cfq_rq_close(cfqd, cfqq, rq))
3429 * cfqq preempts the active queue. if we allowed preempt with no slice left,
3430 * let it have half of its nominal slice.
3432 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3434 enum wl_type_t old_type = cfqq_type(cfqd->active_queue);
3436 cfq_log_cfqq(cfqd, cfqq, "preempt");
3437 cfq_slice_expired(cfqd, 1);
3440 * workload type is changed, don't save slice, otherwise preempt
3443 if (old_type != cfqq_type(cfqq))
3444 cfqq->cfqg->saved_workload_slice = 0;
3447 * Put the new queue at the front of the of the current list,
3448 * so we know that it will be selected next.
3450 BUG_ON(!cfq_cfqq_on_rr(cfqq));
3452 cfq_service_tree_add(cfqd, cfqq, 1);
3454 cfqq->slice_end = 0;
3455 cfq_mark_cfqq_slice_new(cfqq);
3459 * Called when a new fs request (rq) is added (to cfqq). Check if there's
3460 * something we should do about it
3463 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3466 struct cfq_io_cq *cic = RQ_CIC(rq);
3469 if (rq->cmd_flags & REQ_PRIO)
3470 cfqq->prio_pending++;
3472 cfq_update_io_thinktime(cfqd, cfqq, cic);
3473 cfq_update_io_seektime(cfqd, cfqq, rq);
3474 cfq_update_idle_window(cfqd, cfqq, cic);
3476 cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
3478 if (cfqq == cfqd->active_queue) {
3480 * Remember that we saw a request from this process, but
3481 * don't start queuing just yet. Otherwise we risk seeing lots
3482 * of tiny requests, because we disrupt the normal plugging
3483 * and merging. If the request is already larger than a single
3484 * page, let it rip immediately. For that case we assume that
3485 * merging is already done. Ditto for a busy system that
3486 * has other work pending, don't risk delaying until the
3487 * idle timer unplug to continue working.
3489 if (cfq_cfqq_wait_request(cfqq)) {
3490 if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
3491 cfqd->busy_queues > 1) {
3492 cfq_del_timer(cfqd, cfqq);
3493 cfq_clear_cfqq_wait_request(cfqq);
3494 __blk_run_queue(cfqd->queue);
3496 cfqg_stats_update_idle_time(cfqq->cfqg);
3497 cfq_mark_cfqq_must_dispatch(cfqq);
3500 } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
3502 * not the active queue - expire current slice if it is
3503 * idle and has expired it's mean thinktime or this new queue
3504 * has some old slice time left and is of higher priority or
3505 * this new queue is RT and the current one is BE
3507 cfq_preempt_queue(cfqd, cfqq);
3508 __blk_run_queue(cfqd->queue);
3512 static void cfq_insert_request(struct request_queue *q, struct request *rq)
3514 struct cfq_data *cfqd = q->elevator->elevator_data;
3515 struct cfq_queue *cfqq = RQ_CFQQ(rq);
3517 cfq_log_cfqq(cfqd, cfqq, "insert_request");
3518 cfq_init_prio_data(cfqq, RQ_CIC(rq));
3520 rq_set_fifo_time(rq, jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)]);
3521 list_add_tail(&rq->queuelist, &cfqq->fifo);
3523 cfqg_stats_update_io_add(RQ_CFQG(rq), cfqd->serving_group,
3525 cfq_rq_enqueued(cfqd, cfqq, rq);
3529 * Update hw_tag based on peak queue depth over 50 samples under
3532 static void cfq_update_hw_tag(struct cfq_data *cfqd)
3534 struct cfq_queue *cfqq = cfqd->active_queue;
3536 if (cfqd->rq_in_driver > cfqd->hw_tag_est_depth)
3537 cfqd->hw_tag_est_depth = cfqd->rq_in_driver;
3539 if (cfqd->hw_tag == 1)
3542 if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
3543 cfqd->rq_in_driver <= CFQ_HW_QUEUE_MIN)
3547 * If active queue hasn't enough requests and can idle, cfq might not
3548 * dispatch sufficient requests to hardware. Don't zero hw_tag in this
3551 if (cfqq && cfq_cfqq_idle_window(cfqq) &&
3552 cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] <
3553 CFQ_HW_QUEUE_MIN && cfqd->rq_in_driver < CFQ_HW_QUEUE_MIN)
3556 if (cfqd->hw_tag_samples++ < 50)
3559 if (cfqd->hw_tag_est_depth >= CFQ_HW_QUEUE_MIN)
3565 static bool cfq_should_wait_busy(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3567 struct cfq_io_cq *cic = cfqd->active_cic;
3569 /* If the queue already has requests, don't wait */
3570 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3573 /* If there are other queues in the group, don't wait */
3574 if (cfqq->cfqg->nr_cfqq > 1)
3577 /* the only queue in the group, but think time is big */
3578 if (cfq_io_thinktime_big(cfqd, &cfqq->cfqg->ttime, true))
3581 if (cfq_slice_used(cfqq))
3584 /* if slice left is less than think time, wait busy */
3585 if (cic && sample_valid(cic->ttime.ttime_samples)
3586 && (cfqq->slice_end - jiffies < cic->ttime.ttime_mean))
3590 * If think times is less than a jiffy than ttime_mean=0 and above
3591 * will not be true. It might happen that slice has not expired yet
3592 * but will expire soon (4-5 ns) during select_queue(). To cover the
3593 * case where think time is less than a jiffy, mark the queue wait
3594 * busy if only 1 jiffy is left in the slice.
3596 if (cfqq->slice_end - jiffies == 1)
3602 static void cfq_completed_request(struct request_queue *q, struct request *rq)
3604 struct cfq_queue *cfqq = RQ_CFQQ(rq);
3605 struct cfq_data *cfqd = cfqq->cfqd;
3606 const int sync = rq_is_sync(rq);
3610 cfq_log_cfqq(cfqd, cfqq, "complete rqnoidle %d",
3611 !!(rq->cmd_flags & REQ_NOIDLE));
3613 cfq_update_hw_tag(cfqd);
3615 WARN_ON(!cfqd->rq_in_driver);
3616 WARN_ON(!cfqq->dispatched);
3617 cfqd->rq_in_driver--;
3619 (RQ_CFQG(rq))->dispatched--;
3620 cfqg_stats_update_completion(cfqq->cfqg, rq_start_time_ns(rq),
3621 rq_io_start_time_ns(rq), rq->cmd_flags);
3623 cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]--;
3626 struct cfq_rb_root *service_tree;
3628 RQ_CIC(rq)->ttime.last_end_request = now;
3630 if (cfq_cfqq_on_rr(cfqq))
3631 service_tree = cfqq->service_tree;
3633 service_tree = service_tree_for(cfqq->cfqg,
3634 cfqq_prio(cfqq), cfqq_type(cfqq));
3635 service_tree->ttime.last_end_request = now;
3636 if (!time_after(rq->start_time + cfqd->cfq_fifo_expire[1], now))
3637 cfqd->last_delayed_sync = now;
3640 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3641 cfqq->cfqg->ttime.last_end_request = now;
3645 * If this is the active queue, check if it needs to be expired,
3646 * or if we want to idle in case it has no pending requests.
3648 if (cfqd->active_queue == cfqq) {
3649 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
3651 if (cfq_cfqq_slice_new(cfqq)) {
3652 cfq_set_prio_slice(cfqd, cfqq);
3653 cfq_clear_cfqq_slice_new(cfqq);
3657 * Should we wait for next request to come in before we expire
3660 if (cfq_should_wait_busy(cfqd, cfqq)) {
3661 unsigned long extend_sl = cfqd->cfq_slice_idle;
3662 if (!cfqd->cfq_slice_idle)
3663 extend_sl = cfqd->cfq_group_idle;
3664 cfqq->slice_end = jiffies + extend_sl;
3665 cfq_mark_cfqq_wait_busy(cfqq);
3666 cfq_log_cfqq(cfqd, cfqq, "will busy wait");
3670 * Idling is not enabled on:
3672 * - idle-priority queues
3674 * - queues with still some requests queued
3675 * - when there is a close cooperator
3677 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
3678 cfq_slice_expired(cfqd, 1);
3679 else if (sync && cfqq_empty &&
3680 !cfq_close_cooperator(cfqd, cfqq)) {
3681 cfq_arm_slice_timer(cfqd);
3685 if (!cfqd->rq_in_driver)
3686 cfq_schedule_dispatch(cfqd);
3689 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
3691 if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
3692 cfq_mark_cfqq_must_alloc_slice(cfqq);
3693 return ELV_MQUEUE_MUST;
3696 return ELV_MQUEUE_MAY;
3699 static int cfq_may_queue(struct request_queue *q, int rw)
3701 struct cfq_data *cfqd = q->elevator->elevator_data;
3702 struct task_struct *tsk = current;
3703 struct cfq_io_cq *cic;
3704 struct cfq_queue *cfqq;
3707 * don't force setup of a queue from here, as a call to may_queue
3708 * does not necessarily imply that a request actually will be queued.
3709 * so just lookup a possibly existing queue, or return 'may queue'
3712 cic = cfq_cic_lookup(cfqd, tsk->io_context);
3714 return ELV_MQUEUE_MAY;
3716 cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
3718 cfq_init_prio_data(cfqq, cic);
3720 return __cfq_may_queue(cfqq);
3723 return ELV_MQUEUE_MAY;
3727 * queue lock held here
3729 static void cfq_put_request(struct request *rq)
3731 struct cfq_queue *cfqq = RQ_CFQQ(rq);
3734 const int rw = rq_data_dir(rq);
3736 BUG_ON(!cfqq->allocated[rw]);
3737 cfqq->allocated[rw]--;
3739 /* Put down rq reference on cfqg */
3740 cfqg_put(RQ_CFQG(rq));
3741 rq->elv.priv[0] = NULL;
3742 rq->elv.priv[1] = NULL;
3744 cfq_put_queue(cfqq);
3748 static struct cfq_queue *
3749 cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_cq *cic,
3750 struct cfq_queue *cfqq)
3752 cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq);
3753 cic_set_cfqq(cic, cfqq->new_cfqq, 1);
3754 cfq_mark_cfqq_coop(cfqq->new_cfqq);
3755 cfq_put_queue(cfqq);
3756 return cic_to_cfqq(cic, 1);
3760 * Returns NULL if a new cfqq should be allocated, or the old cfqq if this
3761 * was the last process referring to said cfqq.
3763 static struct cfq_queue *
3764 split_cfqq(struct cfq_io_cq *cic, struct cfq_queue *cfqq)
3766 if (cfqq_process_refs(cfqq) == 1) {
3767 cfqq->pid = current->pid;
3768 cfq_clear_cfqq_coop(cfqq);
3769 cfq_clear_cfqq_split_coop(cfqq);
3773 cic_set_cfqq(cic, NULL, 1);
3775 cfq_put_cooperator(cfqq);
3777 cfq_put_queue(cfqq);
3781 * Allocate cfq data structures associated with this request.
3784 cfq_set_request(struct request_queue *q, struct request *rq, struct bio *bio,
3787 struct cfq_data *cfqd = q->elevator->elevator_data;
3788 struct cfq_io_cq *cic = icq_to_cic(rq->elv.icq);
3789 const int rw = rq_data_dir(rq);
3790 const bool is_sync = rq_is_sync(rq);
3791 struct cfq_queue *cfqq;
3793 might_sleep_if(gfp_mask & __GFP_WAIT);
3795 spin_lock_irq(q->queue_lock);
3797 check_ioprio_changed(cic, bio);
3798 check_blkcg_changed(cic, bio);
3800 cfqq = cic_to_cfqq(cic, is_sync);
3801 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
3802 cfqq = cfq_get_queue(cfqd, is_sync, cic, bio, gfp_mask);
3803 cic_set_cfqq(cic, cfqq, is_sync);
3806 * If the queue was seeky for too long, break it apart.
3808 if (cfq_cfqq_coop(cfqq) && cfq_cfqq_split_coop(cfqq)) {
3809 cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq");
3810 cfqq = split_cfqq(cic, cfqq);
3816 * Check to see if this queue is scheduled to merge with
3817 * another, closely cooperating queue. The merging of
3818 * queues happens here as it must be done in process context.
3819 * The reference on new_cfqq was taken in merge_cfqqs.
3822 cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq);
3825 cfqq->allocated[rw]++;
3828 cfqg_get(cfqq->cfqg);
3829 rq->elv.priv[0] = cfqq;
3830 rq->elv.priv[1] = cfqq->cfqg;
3831 spin_unlock_irq(q->queue_lock);
3835 static void cfq_kick_queue(struct work_struct *work)
3837 struct cfq_data *cfqd =
3838 container_of(work, struct cfq_data, unplug_work);
3839 struct request_queue *q = cfqd->queue;
3841 spin_lock_irq(q->queue_lock);
3842 __blk_run_queue(cfqd->queue);
3843 spin_unlock_irq(q->queue_lock);
3847 * Timer running if the active_queue is currently idling inside its time slice
3849 static void cfq_idle_slice_timer(unsigned long data)
3851 struct cfq_data *cfqd = (struct cfq_data *) data;
3852 struct cfq_queue *cfqq;
3853 unsigned long flags;
3856 cfq_log(cfqd, "idle timer fired");
3858 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
3860 cfqq = cfqd->active_queue;
3865 * We saw a request before the queue expired, let it through
3867 if (cfq_cfqq_must_dispatch(cfqq))
3873 if (cfq_slice_used(cfqq))
3877 * only expire and reinvoke request handler, if there are
3878 * other queues with pending requests
3880 if (!cfqd->busy_queues)
3884 * not expired and it has a request pending, let it dispatch
3886 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3890 * Queue depth flag is reset only when the idle didn't succeed
3892 cfq_clear_cfqq_deep(cfqq);
3895 cfq_slice_expired(cfqd, timed_out);
3897 cfq_schedule_dispatch(cfqd);
3899 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
3902 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
3904 del_timer_sync(&cfqd->idle_slice_timer);
3905 cancel_work_sync(&cfqd->unplug_work);
3908 static void cfq_put_async_queues(struct cfq_data *cfqd)
3912 for (i = 0; i < IOPRIO_BE_NR; i++) {
3913 if (cfqd->async_cfqq[0][i])
3914 cfq_put_queue(cfqd->async_cfqq[0][i]);
3915 if (cfqd->async_cfqq[1][i])
3916 cfq_put_queue(cfqd->async_cfqq[1][i]);
3919 if (cfqd->async_idle_cfqq)
3920 cfq_put_queue(cfqd->async_idle_cfqq);
3923 static void cfq_exit_queue(struct elevator_queue *e)
3925 struct cfq_data *cfqd = e->elevator_data;
3926 struct request_queue *q = cfqd->queue;
3928 cfq_shutdown_timer_wq(cfqd);
3930 spin_lock_irq(q->queue_lock);
3932 if (cfqd->active_queue)
3933 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
3935 cfq_put_async_queues(cfqd);
3937 spin_unlock_irq(q->queue_lock);
3939 cfq_shutdown_timer_wq(cfqd);
3941 #ifndef CONFIG_CFQ_GROUP_IOSCHED
3942 kfree(cfqd->root_group);
3944 update_root_blkg_pd(q, BLKIO_POLICY_PROP);
3948 static int cfq_init_queue(struct request_queue *q)
3950 struct cfq_data *cfqd;
3951 struct blkio_group *blkg __maybe_unused;
3954 cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
3959 q->elevator->elevator_data = cfqd;
3961 /* Init root service tree */
3962 cfqd->grp_service_tree = CFQ_RB_ROOT;
3964 /* Init root group and prefer root group over other groups by default */
3965 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3967 spin_lock_irq(q->queue_lock);
3969 blkg = blkg_lookup_create(&blkio_root_cgroup, q, true);
3971 cfqd->root_group = blkg_to_cfqg(blkg);
3973 spin_unlock_irq(q->queue_lock);
3976 cfqd->root_group = kzalloc_node(sizeof(*cfqd->root_group),
3977 GFP_KERNEL, cfqd->queue->node);
3978 if (cfqd->root_group)
3979 cfq_init_cfqg_base(cfqd->root_group);
3981 if (!cfqd->root_group) {
3986 cfqd->root_group->weight = 2*BLKIO_WEIGHT_DEFAULT;
3989 * Not strictly needed (since RB_ROOT just clears the node and we
3990 * zeroed cfqd on alloc), but better be safe in case someone decides
3991 * to add magic to the rb code
3993 for (i = 0; i < CFQ_PRIO_LISTS; i++)
3994 cfqd->prio_trees[i] = RB_ROOT;
3997 * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
3998 * Grab a permanent reference to it, so that the normal code flow
3999 * will not attempt to free it. oom_cfqq is linked to root_group
4000 * but shouldn't hold a reference as it'll never be unlinked. Lose
4001 * the reference from linking right away.
4003 cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
4004 cfqd->oom_cfqq.ref++;
4006 spin_lock_irq(q->queue_lock);
4007 cfq_link_cfqq_cfqg(&cfqd->oom_cfqq, cfqd->root_group);
4008 cfqg_put(cfqd->root_group);
4009 spin_unlock_irq(q->queue_lock);
4011 init_timer(&cfqd->idle_slice_timer);
4012 cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
4013 cfqd->idle_slice_timer.data = (unsigned long) cfqd;
4015 INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
4017 cfqd->cfq_quantum = cfq_quantum;
4018 cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
4019 cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
4020 cfqd->cfq_back_max = cfq_back_max;
4021 cfqd->cfq_back_penalty = cfq_back_penalty;
4022 cfqd->cfq_slice[0] = cfq_slice_async;
4023 cfqd->cfq_slice[1] = cfq_slice_sync;
4024 cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
4025 cfqd->cfq_slice_idle = cfq_slice_idle;
4026 cfqd->cfq_group_idle = cfq_group_idle;
4027 cfqd->cfq_latency = 1;
4030 * we optimistically start assuming sync ops weren't delayed in last
4031 * second, in order to have larger depth for async operations.
4033 cfqd->last_delayed_sync = jiffies - HZ;
4038 * sysfs parts below -->
4041 cfq_var_show(unsigned int var, char *page)
4043 return sprintf(page, "%d\n", var);
4047 cfq_var_store(unsigned int *var, const char *page, size_t count)
4049 char *p = (char *) page;
4051 *var = simple_strtoul(p, &p, 10);
4055 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
4056 static ssize_t __FUNC(struct elevator_queue *e, char *page) \
4058 struct cfq_data *cfqd = e->elevator_data; \
4059 unsigned int __data = __VAR; \
4061 __data = jiffies_to_msecs(__data); \
4062 return cfq_var_show(__data, (page)); \
4064 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
4065 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
4066 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
4067 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
4068 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
4069 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
4070 SHOW_FUNCTION(cfq_group_idle_show, cfqd->cfq_group_idle, 1);
4071 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
4072 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
4073 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
4074 SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
4075 #undef SHOW_FUNCTION
4077 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
4078 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
4080 struct cfq_data *cfqd = e->elevator_data; \
4081 unsigned int __data; \
4082 int ret = cfq_var_store(&__data, (page), count); \
4083 if (__data < (MIN)) \
4085 else if (__data > (MAX)) \
4088 *(__PTR) = msecs_to_jiffies(__data); \
4090 *(__PTR) = __data; \
4093 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
4094 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
4096 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
4098 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
4099 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
4101 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
4102 STORE_FUNCTION(cfq_group_idle_store, &cfqd->cfq_group_idle, 0, UINT_MAX, 1);
4103 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
4104 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
4105 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
4107 STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
4108 #undef STORE_FUNCTION
4110 #define CFQ_ATTR(name) \
4111 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
4113 static struct elv_fs_entry cfq_attrs[] = {
4115 CFQ_ATTR(fifo_expire_sync),
4116 CFQ_ATTR(fifo_expire_async),
4117 CFQ_ATTR(back_seek_max),
4118 CFQ_ATTR(back_seek_penalty),
4119 CFQ_ATTR(slice_sync),
4120 CFQ_ATTR(slice_async),
4121 CFQ_ATTR(slice_async_rq),
4122 CFQ_ATTR(slice_idle),
4123 CFQ_ATTR(group_idle),
4124 CFQ_ATTR(low_latency),
4128 static struct elevator_type iosched_cfq = {
4130 .elevator_merge_fn = cfq_merge,
4131 .elevator_merged_fn = cfq_merged_request,
4132 .elevator_merge_req_fn = cfq_merged_requests,
4133 .elevator_allow_merge_fn = cfq_allow_merge,
4134 .elevator_bio_merged_fn = cfq_bio_merged,
4135 .elevator_dispatch_fn = cfq_dispatch_requests,
4136 .elevator_add_req_fn = cfq_insert_request,
4137 .elevator_activate_req_fn = cfq_activate_request,
4138 .elevator_deactivate_req_fn = cfq_deactivate_request,
4139 .elevator_completed_req_fn = cfq_completed_request,
4140 .elevator_former_req_fn = elv_rb_former_request,
4141 .elevator_latter_req_fn = elv_rb_latter_request,
4142 .elevator_init_icq_fn = cfq_init_icq,
4143 .elevator_exit_icq_fn = cfq_exit_icq,
4144 .elevator_set_req_fn = cfq_set_request,
4145 .elevator_put_req_fn = cfq_put_request,
4146 .elevator_may_queue_fn = cfq_may_queue,
4147 .elevator_init_fn = cfq_init_queue,
4148 .elevator_exit_fn = cfq_exit_queue,
4150 .icq_size = sizeof(struct cfq_io_cq),
4151 .icq_align = __alignof__(struct cfq_io_cq),
4152 .elevator_attrs = cfq_attrs,
4153 .elevator_name = "cfq",
4154 .elevator_owner = THIS_MODULE,
4157 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4158 static struct blkio_policy_type blkio_policy_cfq = {
4160 .blkio_init_group_fn = cfq_init_blkio_group,
4161 .blkio_reset_group_stats_fn = cfqg_stats_reset,
4163 .plid = BLKIO_POLICY_PROP,
4164 .pdata_size = sizeof(struct cfq_group),
4165 .cftypes = cfq_blkcg_files,
4169 static int __init cfq_init(void)
4174 * could be 0 on HZ < 1000 setups
4176 if (!cfq_slice_async)
4177 cfq_slice_async = 1;
4178 if (!cfq_slice_idle)
4181 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4182 if (!cfq_group_idle)
4187 cfq_pool = KMEM_CACHE(cfq_queue, 0);
4191 ret = elv_register(&iosched_cfq);
4193 kmem_cache_destroy(cfq_pool);
4197 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4198 blkio_policy_register(&blkio_policy_cfq);
4203 static void __exit cfq_exit(void)
4205 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4206 blkio_policy_unregister(&blkio_policy_cfq);
4208 elv_unregister(&iosched_cfq);
4209 kmem_cache_destroy(cfq_pool);
4212 module_init(cfq_init);
4213 module_exit(cfq_exit);
4215 MODULE_AUTHOR("Jens Axboe");
4216 MODULE_LICENSE("GPL");
4217 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");