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"
23 /* max queue in one round of service */
24 static const int cfq_quantum = 8;
25 static const int cfq_fifo_expire[2] = { HZ / 4, HZ / 8 };
26 /* maximum backwards seek, in KiB */
27 static const int cfq_back_max = 16 * 1024;
28 /* penalty of a backwards seek */
29 static const int cfq_back_penalty = 2;
30 static const int cfq_slice_sync = HZ / 10;
31 static int cfq_slice_async = HZ / 25;
32 static const int cfq_slice_async_rq = 2;
33 static int cfq_slice_idle = HZ / 125;
34 static int cfq_group_idle = HZ / 125;
35 static const int cfq_target_latency = HZ * 3/10; /* 300 ms */
36 static const int cfq_hist_divisor = 4;
39 * offset from end of service tree
41 #define CFQ_IDLE_DELAY (HZ / 5)
44 * below this threshold, we consider thinktime immediate
46 #define CFQ_MIN_TT (2)
48 #define CFQ_SLICE_SCALE (5)
49 #define CFQ_HW_QUEUE_MIN (5)
50 #define CFQ_SERVICE_SHIFT 12
52 #define CFQQ_SEEK_THR (sector_t)(8 * 100)
53 #define CFQQ_CLOSE_THR (sector_t)(8 * 1024)
54 #define CFQQ_SECT_THR_NONROT (sector_t)(2 * 32)
55 #define CFQQ_SEEKY(cfqq) (hweight32(cfqq->seek_history) > 32/8)
57 #define RQ_CIC(rq) icq_to_cic((rq)->elv.icq)
58 #define RQ_CFQQ(rq) (struct cfq_queue *) ((rq)->elv.priv[0])
59 #define RQ_CFQG(rq) (struct cfq_group *) ((rq)->elv.priv[1])
61 static struct kmem_cache *cfq_pool;
63 #define CFQ_PRIO_LISTS IOPRIO_BE_NR
64 #define cfq_class_idle(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
65 #define cfq_class_rt(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
67 #define sample_valid(samples) ((samples) > 80)
68 #define rb_entry_cfqg(node) rb_entry((node), struct cfq_group, rb_node)
71 unsigned long last_end_request;
73 unsigned long ttime_total;
74 unsigned long ttime_samples;
75 unsigned long ttime_mean;
79 * Most of our rbtree usage is for sorting with min extraction, so
80 * if we cache the leftmost node we don't have to walk down the tree
81 * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should
82 * move this into the elevator for the rq sorting as well.
88 unsigned total_weight;
90 struct cfq_ttime ttime;
92 #define CFQ_RB_ROOT (struct cfq_rb_root) { .rb = RB_ROOT, \
93 .ttime = {.last_end_request = jiffies,},}
96 * Per process-grouping structure
101 /* various state flags, see below */
103 /* parent cfq_data */
104 struct cfq_data *cfqd;
105 /* service_tree member */
106 struct rb_node rb_node;
107 /* service_tree key */
108 unsigned long rb_key;
109 /* prio tree member */
110 struct rb_node p_node;
111 /* prio tree root we belong to, if any */
112 struct rb_root *p_root;
113 /* sorted list of pending requests */
114 struct rb_root sort_list;
115 /* if fifo isn't expired, next request to serve */
116 struct request *next_rq;
117 /* requests queued in sort_list */
119 /* currently allocated requests */
121 /* fifo list of requests in sort_list */
122 struct list_head fifo;
124 /* time when queue got scheduled in to dispatch first request. */
125 unsigned long dispatch_start;
126 unsigned int allocated_slice;
127 unsigned int slice_dispatch;
128 /* time when first request from queue completed and slice started. */
129 unsigned long slice_start;
130 unsigned long slice_end;
133 /* pending priority requests */
135 /* number of requests that are on the dispatch list or inside driver */
138 /* io prio of this group */
139 unsigned short ioprio, org_ioprio;
140 unsigned short ioprio_class;
145 sector_t last_request_pos;
147 struct cfq_rb_root *service_tree;
148 struct cfq_queue *new_cfqq;
149 struct cfq_group *cfqg;
150 /* Number of sectors dispatched from queue in single dispatch round */
151 unsigned long nr_sectors;
155 * First index in the service_trees.
156 * IDLE is handled separately, so it has negative index
166 * Second index in the service_trees.
170 SYNC_NOIDLE_WORKLOAD = 1,
175 #ifdef CONFIG_CFQ_GROUP_IOSCHED
176 /* total bytes transferred */
177 struct blkg_rwstat service_bytes;
178 /* total IOs serviced, post merge */
179 struct blkg_rwstat serviced;
180 /* number of ios merged */
181 struct blkg_rwstat merged;
182 /* total time spent on device in ns, may not be accurate w/ queueing */
183 struct blkg_rwstat service_time;
184 /* total time spent waiting in scheduler queue in ns */
185 struct blkg_rwstat wait_time;
186 /* number of IOs queued up */
187 struct blkg_rwstat queued;
188 /* total sectors transferred */
189 struct blkg_stat sectors;
190 /* total disk time and nr sectors dispatched by this group */
191 struct blkg_stat time;
192 #ifdef CONFIG_DEBUG_BLK_CGROUP
193 /* time not charged to this cgroup */
194 struct blkg_stat unaccounted_time;
195 /* sum of number of ios queued across all samples */
196 struct blkg_stat avg_queue_size_sum;
197 /* count of samples taken for average */
198 struct blkg_stat avg_queue_size_samples;
199 /* how many times this group has been removed from service tree */
200 struct blkg_stat dequeue;
201 /* total time spent waiting for it to be assigned a timeslice. */
202 struct blkg_stat group_wait_time;
203 /* time spent idling for this blkcg_gq */
204 struct blkg_stat idle_time;
205 /* total time with empty current active q with other requests queued */
206 struct blkg_stat empty_time;
207 /* fields after this shouldn't be cleared on stat reset */
208 uint64_t start_group_wait_time;
209 uint64_t start_idle_time;
210 uint64_t start_empty_time;
212 #endif /* CONFIG_DEBUG_BLK_CGROUP */
213 #endif /* CONFIG_CFQ_GROUP_IOSCHED */
216 /* This is per cgroup per device grouping structure */
218 /* must be the first member */
219 struct blkg_policy_data pd;
221 /* group service_tree member */
222 struct rb_node rb_node;
224 /* group service_tree key */
227 unsigned int new_weight;
228 unsigned int dev_weight;
230 /* number of cfqq currently on this group */
234 * Per group busy queues average. Useful for workload slice calc. We
235 * create the array for each prio class but at run time it is used
236 * only for RT and BE class and slot for IDLE class remains unused.
237 * This is primarily done to avoid confusion and a gcc warning.
239 unsigned int busy_queues_avg[CFQ_PRIO_NR];
241 * rr lists of queues with requests. We maintain service trees for
242 * RT and BE classes. These trees are subdivided in subclasses
243 * of SYNC, SYNC_NOIDLE and ASYNC based on workload type. For IDLE
244 * class there is no subclassification and all the cfq queues go on
245 * a single tree service_tree_idle.
246 * Counts are embedded in the cfq_rb_root
248 struct cfq_rb_root service_trees[2][3];
249 struct cfq_rb_root service_tree_idle;
251 unsigned long saved_wl_slice;
252 enum wl_type_t saved_wl_type;
253 enum wl_class_t saved_wl_class;
255 /* number of requests that are on the dispatch list or inside driver */
257 struct cfq_ttime ttime;
258 struct cfqg_stats stats;
262 struct io_cq icq; /* must be the first member */
263 struct cfq_queue *cfqq[2];
264 struct cfq_ttime ttime;
265 int ioprio; /* the current ioprio */
266 #ifdef CONFIG_CFQ_GROUP_IOSCHED
267 uint64_t blkcg_id; /* the current blkcg ID */
272 * Per block device queue structure
275 struct request_queue *queue;
276 /* Root service tree for cfq_groups */
277 struct cfq_rb_root grp_service_tree;
278 struct cfq_group *root_group;
281 * The priority currently being served
283 enum wl_class_t serving_wl_class;
284 enum wl_type_t serving_wl_type;
285 unsigned long workload_expires;
286 struct cfq_group *serving_group;
289 * Each priority tree is sorted by next_request position. These
290 * trees are used when determining if two or more queues are
291 * interleaving requests (see cfq_close_cooperator).
293 struct rb_root prio_trees[CFQ_PRIO_LISTS];
295 unsigned int busy_queues;
296 unsigned int busy_sync_queues;
302 * queue-depth detection
308 * -1 => indeterminate, (cfq will behave as if NCQ is present, to allow better detection)
309 * 1 => NCQ is present (hw_tag_est_depth is the estimated max depth)
312 int hw_tag_est_depth;
313 unsigned int hw_tag_samples;
316 * idle window management
318 struct timer_list idle_slice_timer;
319 struct work_struct unplug_work;
321 struct cfq_queue *active_queue;
322 struct cfq_io_cq *active_cic;
325 * async queue for each priority case
327 struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
328 struct cfq_queue *async_idle_cfqq;
330 sector_t last_position;
333 * tunables, see top of file
335 unsigned int cfq_quantum;
336 unsigned int cfq_fifo_expire[2];
337 unsigned int cfq_back_penalty;
338 unsigned int cfq_back_max;
339 unsigned int cfq_slice[2];
340 unsigned int cfq_slice_async_rq;
341 unsigned int cfq_slice_idle;
342 unsigned int cfq_group_idle;
343 unsigned int cfq_latency;
344 unsigned int cfq_target_latency;
347 * Fallback dummy cfqq for extreme OOM conditions
349 struct cfq_queue oom_cfqq;
351 unsigned long last_delayed_sync;
354 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd);
356 static struct cfq_rb_root *service_tree_for(struct cfq_group *cfqg,
357 enum wl_class_t class,
363 if (class == IDLE_WORKLOAD)
364 return &cfqg->service_tree_idle;
366 return &cfqg->service_trees[class][type];
369 enum cfqq_state_flags {
370 CFQ_CFQQ_FLAG_on_rr = 0, /* on round-robin busy list */
371 CFQ_CFQQ_FLAG_wait_request, /* waiting for a request */
372 CFQ_CFQQ_FLAG_must_dispatch, /* must be allowed a dispatch */
373 CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
374 CFQ_CFQQ_FLAG_fifo_expire, /* FIFO checked in this slice */
375 CFQ_CFQQ_FLAG_idle_window, /* slice idling enabled */
376 CFQ_CFQQ_FLAG_prio_changed, /* task priority has changed */
377 CFQ_CFQQ_FLAG_slice_new, /* no requests dispatched in slice */
378 CFQ_CFQQ_FLAG_sync, /* synchronous queue */
379 CFQ_CFQQ_FLAG_coop, /* cfqq is shared */
380 CFQ_CFQQ_FLAG_split_coop, /* shared cfqq will be splitted */
381 CFQ_CFQQ_FLAG_deep, /* sync cfqq experienced large depth */
382 CFQ_CFQQ_FLAG_wait_busy, /* Waiting for next request */
385 #define CFQ_CFQQ_FNS(name) \
386 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \
388 (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name); \
390 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \
392 (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \
394 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \
396 return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \
400 CFQ_CFQQ_FNS(wait_request);
401 CFQ_CFQQ_FNS(must_dispatch);
402 CFQ_CFQQ_FNS(must_alloc_slice);
403 CFQ_CFQQ_FNS(fifo_expire);
404 CFQ_CFQQ_FNS(idle_window);
405 CFQ_CFQQ_FNS(prio_changed);
406 CFQ_CFQQ_FNS(slice_new);
409 CFQ_CFQQ_FNS(split_coop);
411 CFQ_CFQQ_FNS(wait_busy);
414 static inline struct cfq_group *pd_to_cfqg(struct blkg_policy_data *pd)
416 return pd ? container_of(pd, struct cfq_group, pd) : NULL;
419 static inline struct blkcg_gq *cfqg_to_blkg(struct cfq_group *cfqg)
421 return pd_to_blkg(&cfqg->pd);
424 #if defined(CONFIG_CFQ_GROUP_IOSCHED) && defined(CONFIG_DEBUG_BLK_CGROUP)
426 /* cfqg stats flags */
427 enum cfqg_stats_flags {
428 CFQG_stats_waiting = 0,
433 #define CFQG_FLAG_FNS(name) \
434 static inline void cfqg_stats_mark_##name(struct cfqg_stats *stats) \
436 stats->flags |= (1 << CFQG_stats_##name); \
438 static inline void cfqg_stats_clear_##name(struct cfqg_stats *stats) \
440 stats->flags &= ~(1 << CFQG_stats_##name); \
442 static inline int cfqg_stats_##name(struct cfqg_stats *stats) \
444 return (stats->flags & (1 << CFQG_stats_##name)) != 0; \
447 CFQG_FLAG_FNS(waiting)
448 CFQG_FLAG_FNS(idling)
452 /* This should be called with the queue_lock held. */
453 static void cfqg_stats_update_group_wait_time(struct cfqg_stats *stats)
455 unsigned long long now;
457 if (!cfqg_stats_waiting(stats))
461 if (time_after64(now, stats->start_group_wait_time))
462 blkg_stat_add(&stats->group_wait_time,
463 now - stats->start_group_wait_time);
464 cfqg_stats_clear_waiting(stats);
467 /* This should be called with the queue_lock held. */
468 static void cfqg_stats_set_start_group_wait_time(struct cfq_group *cfqg,
469 struct cfq_group *curr_cfqg)
471 struct cfqg_stats *stats = &cfqg->stats;
473 if (cfqg_stats_waiting(stats))
475 if (cfqg == curr_cfqg)
477 stats->start_group_wait_time = sched_clock();
478 cfqg_stats_mark_waiting(stats);
481 /* This should be called with the queue_lock held. */
482 static void cfqg_stats_end_empty_time(struct cfqg_stats *stats)
484 unsigned long long now;
486 if (!cfqg_stats_empty(stats))
490 if (time_after64(now, stats->start_empty_time))
491 blkg_stat_add(&stats->empty_time,
492 now - stats->start_empty_time);
493 cfqg_stats_clear_empty(stats);
496 static void cfqg_stats_update_dequeue(struct cfq_group *cfqg)
498 blkg_stat_add(&cfqg->stats.dequeue, 1);
501 static void cfqg_stats_set_start_empty_time(struct cfq_group *cfqg)
503 struct cfqg_stats *stats = &cfqg->stats;
505 if (blkg_rwstat_sum(&stats->queued))
509 * group is already marked empty. This can happen if cfqq got new
510 * request in parent group and moved to this group while being added
511 * to service tree. Just ignore the event and move on.
513 if (cfqg_stats_empty(stats))
516 stats->start_empty_time = sched_clock();
517 cfqg_stats_mark_empty(stats);
520 static void cfqg_stats_update_idle_time(struct cfq_group *cfqg)
522 struct cfqg_stats *stats = &cfqg->stats;
524 if (cfqg_stats_idling(stats)) {
525 unsigned long long now = sched_clock();
527 if (time_after64(now, stats->start_idle_time))
528 blkg_stat_add(&stats->idle_time,
529 now - stats->start_idle_time);
530 cfqg_stats_clear_idling(stats);
534 static void cfqg_stats_set_start_idle_time(struct cfq_group *cfqg)
536 struct cfqg_stats *stats = &cfqg->stats;
538 BUG_ON(cfqg_stats_idling(stats));
540 stats->start_idle_time = sched_clock();
541 cfqg_stats_mark_idling(stats);
544 static void cfqg_stats_update_avg_queue_size(struct cfq_group *cfqg)
546 struct cfqg_stats *stats = &cfqg->stats;
548 blkg_stat_add(&stats->avg_queue_size_sum,
549 blkg_rwstat_sum(&stats->queued));
550 blkg_stat_add(&stats->avg_queue_size_samples, 1);
551 cfqg_stats_update_group_wait_time(stats);
554 #else /* CONFIG_CFQ_GROUP_IOSCHED && CONFIG_DEBUG_BLK_CGROUP */
556 static inline void cfqg_stats_set_start_group_wait_time(struct cfq_group *cfqg, struct cfq_group *curr_cfqg) { }
557 static inline void cfqg_stats_end_empty_time(struct cfqg_stats *stats) { }
558 static inline void cfqg_stats_update_dequeue(struct cfq_group *cfqg) { }
559 static inline void cfqg_stats_set_start_empty_time(struct cfq_group *cfqg) { }
560 static inline void cfqg_stats_update_idle_time(struct cfq_group *cfqg) { }
561 static inline void cfqg_stats_set_start_idle_time(struct cfq_group *cfqg) { }
562 static inline void cfqg_stats_update_avg_queue_size(struct cfq_group *cfqg) { }
564 #endif /* CONFIG_CFQ_GROUP_IOSCHED && CONFIG_DEBUG_BLK_CGROUP */
566 #ifdef CONFIG_CFQ_GROUP_IOSCHED
568 static struct blkcg_policy blkcg_policy_cfq;
570 static inline struct cfq_group *blkg_to_cfqg(struct blkcg_gq *blkg)
572 return pd_to_cfqg(blkg_to_pd(blkg, &blkcg_policy_cfq));
575 static inline void cfqg_get(struct cfq_group *cfqg)
577 return blkg_get(cfqg_to_blkg(cfqg));
580 static inline void cfqg_put(struct cfq_group *cfqg)
582 return blkg_put(cfqg_to_blkg(cfqg));
585 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...) do { \
588 blkg_path(cfqg_to_blkg((cfqq)->cfqg), __pbuf, sizeof(__pbuf)); \
589 blk_add_trace_msg((cfqd)->queue, "cfq%d%c %s " fmt, (cfqq)->pid, \
590 cfq_cfqq_sync((cfqq)) ? 'S' : 'A', \
594 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...) do { \
597 blkg_path(cfqg_to_blkg(cfqg), __pbuf, sizeof(__pbuf)); \
598 blk_add_trace_msg((cfqd)->queue, "%s " fmt, __pbuf, ##args); \
601 static inline void cfqg_stats_update_io_add(struct cfq_group *cfqg,
602 struct cfq_group *curr_cfqg, int rw)
604 blkg_rwstat_add(&cfqg->stats.queued, rw, 1);
605 cfqg_stats_end_empty_time(&cfqg->stats);
606 cfqg_stats_set_start_group_wait_time(cfqg, curr_cfqg);
609 static inline void cfqg_stats_update_timeslice_used(struct cfq_group *cfqg,
610 unsigned long time, unsigned long unaccounted_time)
612 blkg_stat_add(&cfqg->stats.time, time);
613 #ifdef CONFIG_DEBUG_BLK_CGROUP
614 blkg_stat_add(&cfqg->stats.unaccounted_time, unaccounted_time);
618 static inline void cfqg_stats_update_io_remove(struct cfq_group *cfqg, int rw)
620 blkg_rwstat_add(&cfqg->stats.queued, rw, -1);
623 static inline void cfqg_stats_update_io_merged(struct cfq_group *cfqg, int rw)
625 blkg_rwstat_add(&cfqg->stats.merged, rw, 1);
628 static inline void cfqg_stats_update_dispatch(struct cfq_group *cfqg,
629 uint64_t bytes, int rw)
631 blkg_stat_add(&cfqg->stats.sectors, bytes >> 9);
632 blkg_rwstat_add(&cfqg->stats.serviced, rw, 1);
633 blkg_rwstat_add(&cfqg->stats.service_bytes, rw, bytes);
636 static inline void cfqg_stats_update_completion(struct cfq_group *cfqg,
637 uint64_t start_time, uint64_t io_start_time, int rw)
639 struct cfqg_stats *stats = &cfqg->stats;
640 unsigned long long now = sched_clock();
642 if (time_after64(now, io_start_time))
643 blkg_rwstat_add(&stats->service_time, rw, now - io_start_time);
644 if (time_after64(io_start_time, start_time))
645 blkg_rwstat_add(&stats->wait_time, rw,
646 io_start_time - start_time);
649 static void cfq_pd_reset_stats(struct blkcg_gq *blkg)
651 struct cfq_group *cfqg = blkg_to_cfqg(blkg);
652 struct cfqg_stats *stats = &cfqg->stats;
654 /* queued stats shouldn't be cleared */
655 blkg_rwstat_reset(&stats->service_bytes);
656 blkg_rwstat_reset(&stats->serviced);
657 blkg_rwstat_reset(&stats->merged);
658 blkg_rwstat_reset(&stats->service_time);
659 blkg_rwstat_reset(&stats->wait_time);
660 blkg_stat_reset(&stats->time);
661 #ifdef CONFIG_DEBUG_BLK_CGROUP
662 blkg_stat_reset(&stats->unaccounted_time);
663 blkg_stat_reset(&stats->avg_queue_size_sum);
664 blkg_stat_reset(&stats->avg_queue_size_samples);
665 blkg_stat_reset(&stats->dequeue);
666 blkg_stat_reset(&stats->group_wait_time);
667 blkg_stat_reset(&stats->idle_time);
668 blkg_stat_reset(&stats->empty_time);
672 #else /* CONFIG_CFQ_GROUP_IOSCHED */
674 static inline void cfqg_get(struct cfq_group *cfqg) { }
675 static inline void cfqg_put(struct cfq_group *cfqg) { }
677 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
678 blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args)
679 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...) do {} while (0)
681 static inline void cfqg_stats_update_io_add(struct cfq_group *cfqg,
682 struct cfq_group *curr_cfqg, int rw) { }
683 static inline void cfqg_stats_update_timeslice_used(struct cfq_group *cfqg,
684 unsigned long time, unsigned long unaccounted_time) { }
685 static inline void cfqg_stats_update_io_remove(struct cfq_group *cfqg, int rw) { }
686 static inline void cfqg_stats_update_io_merged(struct cfq_group *cfqg, int rw) { }
687 static inline void cfqg_stats_update_dispatch(struct cfq_group *cfqg,
688 uint64_t bytes, int rw) { }
689 static inline void cfqg_stats_update_completion(struct cfq_group *cfqg,
690 uint64_t start_time, uint64_t io_start_time, int rw) { }
692 #endif /* CONFIG_CFQ_GROUP_IOSCHED */
694 #define cfq_log(cfqd, fmt, args...) \
695 blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
697 /* Traverses through cfq group service trees */
698 #define for_each_cfqg_st(cfqg, i, j, st) \
699 for (i = 0; i <= IDLE_WORKLOAD; i++) \
700 for (j = 0, st = i < IDLE_WORKLOAD ? &cfqg->service_trees[i][j]\
701 : &cfqg->service_tree_idle; \
702 (i < IDLE_WORKLOAD && j <= SYNC_WORKLOAD) || \
703 (i == IDLE_WORKLOAD && j == 0); \
704 j++, st = i < IDLE_WORKLOAD ? \
705 &cfqg->service_trees[i][j]: NULL) \
707 static inline bool cfq_io_thinktime_big(struct cfq_data *cfqd,
708 struct cfq_ttime *ttime, bool group_idle)
711 if (!sample_valid(ttime->ttime_samples))
714 slice = cfqd->cfq_group_idle;
716 slice = cfqd->cfq_slice_idle;
717 return ttime->ttime_mean > slice;
720 static inline bool iops_mode(struct cfq_data *cfqd)
723 * If we are not idling on queues and it is a NCQ drive, parallel
724 * execution of requests is on and measuring time is not possible
725 * in most of the cases until and unless we drive shallower queue
726 * depths and that becomes a performance bottleneck. In such cases
727 * switch to start providing fairness in terms of number of IOs.
729 if (!cfqd->cfq_slice_idle && cfqd->hw_tag)
735 static inline enum wl_class_t cfqq_class(struct cfq_queue *cfqq)
737 if (cfq_class_idle(cfqq))
738 return IDLE_WORKLOAD;
739 if (cfq_class_rt(cfqq))
745 static enum wl_type_t cfqq_type(struct cfq_queue *cfqq)
747 if (!cfq_cfqq_sync(cfqq))
748 return ASYNC_WORKLOAD;
749 if (!cfq_cfqq_idle_window(cfqq))
750 return SYNC_NOIDLE_WORKLOAD;
751 return SYNC_WORKLOAD;
754 static inline int cfq_group_busy_queues_wl(enum wl_class_t wl_class,
755 struct cfq_data *cfqd,
756 struct cfq_group *cfqg)
758 if (wl_class == IDLE_WORKLOAD)
759 return cfqg->service_tree_idle.count;
761 return cfqg->service_trees[wl_class][ASYNC_WORKLOAD].count
762 + cfqg->service_trees[wl_class][SYNC_NOIDLE_WORKLOAD].count
763 + cfqg->service_trees[wl_class][SYNC_WORKLOAD].count;
766 static inline int cfqg_busy_async_queues(struct cfq_data *cfqd,
767 struct cfq_group *cfqg)
769 return cfqg->service_trees[RT_WORKLOAD][ASYNC_WORKLOAD].count
770 + cfqg->service_trees[BE_WORKLOAD][ASYNC_WORKLOAD].count;
773 static void cfq_dispatch_insert(struct request_queue *, struct request *);
774 static struct cfq_queue *cfq_get_queue(struct cfq_data *cfqd, bool is_sync,
775 struct cfq_io_cq *cic, struct bio *bio,
778 static inline struct cfq_io_cq *icq_to_cic(struct io_cq *icq)
780 /* cic->icq is the first member, %NULL will convert to %NULL */
781 return container_of(icq, struct cfq_io_cq, icq);
784 static inline struct cfq_io_cq *cfq_cic_lookup(struct cfq_data *cfqd,
785 struct io_context *ioc)
788 return icq_to_cic(ioc_lookup_icq(ioc, cfqd->queue));
792 static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_cq *cic, bool is_sync)
794 return cic->cfqq[is_sync];
797 static inline void cic_set_cfqq(struct cfq_io_cq *cic, struct cfq_queue *cfqq,
800 cic->cfqq[is_sync] = cfqq;
803 static inline struct cfq_data *cic_to_cfqd(struct cfq_io_cq *cic)
805 return cic->icq.q->elevator->elevator_data;
809 * We regard a request as SYNC, if it's either a read or has the SYNC bit
810 * set (in which case it could also be direct WRITE).
812 static inline bool cfq_bio_sync(struct bio *bio)
814 return bio_data_dir(bio) == READ || (bio->bi_rw & REQ_SYNC);
818 * scheduler run of queue, if there are requests pending and no one in the
819 * driver that will restart queueing
821 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
823 if (cfqd->busy_queues) {
824 cfq_log(cfqd, "schedule dispatch");
825 kblockd_schedule_work(cfqd->queue, &cfqd->unplug_work);
830 * Scale schedule slice based on io priority. Use the sync time slice only
831 * if a queue is marked sync and has sync io queued. A sync queue with async
832 * io only, should not get full sync slice length.
834 static inline int cfq_prio_slice(struct cfq_data *cfqd, bool sync,
837 const int base_slice = cfqd->cfq_slice[sync];
839 WARN_ON(prio >= IOPRIO_BE_NR);
841 return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
845 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
847 return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
850 static inline u64 cfq_scale_slice(unsigned long delta, struct cfq_group *cfqg)
852 u64 d = delta << CFQ_SERVICE_SHIFT;
854 d = d * CFQ_WEIGHT_DEFAULT;
855 do_div(d, cfqg->weight);
859 static inline u64 max_vdisktime(u64 min_vdisktime, u64 vdisktime)
861 s64 delta = (s64)(vdisktime - min_vdisktime);
863 min_vdisktime = vdisktime;
865 return min_vdisktime;
868 static inline u64 min_vdisktime(u64 min_vdisktime, u64 vdisktime)
870 s64 delta = (s64)(vdisktime - min_vdisktime);
872 min_vdisktime = vdisktime;
874 return min_vdisktime;
877 static void update_min_vdisktime(struct cfq_rb_root *st)
879 struct cfq_group *cfqg;
882 cfqg = rb_entry_cfqg(st->left);
883 st->min_vdisktime = max_vdisktime(st->min_vdisktime,
889 * get averaged number of queues of RT/BE priority.
890 * average is updated, with a formula that gives more weight to higher numbers,
891 * to quickly follows sudden increases and decrease slowly
894 static inline unsigned cfq_group_get_avg_queues(struct cfq_data *cfqd,
895 struct cfq_group *cfqg, bool rt)
897 unsigned min_q, max_q;
898 unsigned mult = cfq_hist_divisor - 1;
899 unsigned round = cfq_hist_divisor / 2;
900 unsigned busy = cfq_group_busy_queues_wl(rt, cfqd, cfqg);
902 min_q = min(cfqg->busy_queues_avg[rt], busy);
903 max_q = max(cfqg->busy_queues_avg[rt], busy);
904 cfqg->busy_queues_avg[rt] = (mult * max_q + min_q + round) /
906 return cfqg->busy_queues_avg[rt];
909 static inline unsigned
910 cfq_group_slice(struct cfq_data *cfqd, struct cfq_group *cfqg)
912 struct cfq_rb_root *st = &cfqd->grp_service_tree;
914 return cfqd->cfq_target_latency * cfqg->weight / st->total_weight;
917 static inline unsigned
918 cfq_scaled_cfqq_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
920 unsigned slice = cfq_prio_to_slice(cfqd, cfqq);
921 if (cfqd->cfq_latency) {
923 * interested queues (we consider only the ones with the same
924 * priority class in the cfq group)
926 unsigned iq = cfq_group_get_avg_queues(cfqd, cfqq->cfqg,
928 unsigned sync_slice = cfqd->cfq_slice[1];
929 unsigned expect_latency = sync_slice * iq;
930 unsigned group_slice = cfq_group_slice(cfqd, cfqq->cfqg);
932 if (expect_latency > group_slice) {
933 unsigned base_low_slice = 2 * cfqd->cfq_slice_idle;
934 /* scale low_slice according to IO priority
935 * and sync vs async */
937 min(slice, base_low_slice * slice / sync_slice);
938 /* the adapted slice value is scaled to fit all iqs
939 * into the target latency */
940 slice = max(slice * group_slice / expect_latency,
948 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
950 unsigned slice = cfq_scaled_cfqq_slice(cfqd, cfqq);
952 cfqq->slice_start = jiffies;
953 cfqq->slice_end = jiffies + slice;
954 cfqq->allocated_slice = slice;
955 cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
959 * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
960 * isn't valid until the first request from the dispatch is activated
961 * and the slice time set.
963 static inline bool cfq_slice_used(struct cfq_queue *cfqq)
965 if (cfq_cfqq_slice_new(cfqq))
967 if (time_before(jiffies, cfqq->slice_end))
974 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
975 * We choose the request that is closest to the head right now. Distance
976 * behind the head is penalized and only allowed to a certain extent.
978 static struct request *
979 cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2, sector_t last)
981 sector_t s1, s2, d1 = 0, d2 = 0;
982 unsigned long back_max;
983 #define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */
984 #define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */
985 unsigned wrap = 0; /* bit mask: requests behind the disk head? */
987 if (rq1 == NULL || rq1 == rq2)
992 if (rq_is_sync(rq1) != rq_is_sync(rq2))
993 return rq_is_sync(rq1) ? rq1 : rq2;
995 if ((rq1->cmd_flags ^ rq2->cmd_flags) & REQ_PRIO)
996 return rq1->cmd_flags & REQ_PRIO ? rq1 : rq2;
998 s1 = blk_rq_pos(rq1);
999 s2 = blk_rq_pos(rq2);
1002 * by definition, 1KiB is 2 sectors
1004 back_max = cfqd->cfq_back_max * 2;
1007 * Strict one way elevator _except_ in the case where we allow
1008 * short backward seeks which are biased as twice the cost of a
1009 * similar forward seek.
1013 else if (s1 + back_max >= last)
1014 d1 = (last - s1) * cfqd->cfq_back_penalty;
1016 wrap |= CFQ_RQ1_WRAP;
1020 else if (s2 + back_max >= last)
1021 d2 = (last - s2) * cfqd->cfq_back_penalty;
1023 wrap |= CFQ_RQ2_WRAP;
1025 /* Found required data */
1028 * By doing switch() on the bit mask "wrap" we avoid having to
1029 * check two variables for all permutations: --> faster!
1032 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
1048 case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
1051 * Since both rqs are wrapped,
1052 * start with the one that's further behind head
1053 * (--> only *one* back seek required),
1054 * since back seek takes more time than forward.
1064 * The below is leftmost cache rbtree addon
1066 static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
1068 /* Service tree is empty */
1073 root->left = rb_first(&root->rb);
1076 return rb_entry(root->left, struct cfq_queue, rb_node);
1081 static struct cfq_group *cfq_rb_first_group(struct cfq_rb_root *root)
1084 root->left = rb_first(&root->rb);
1087 return rb_entry_cfqg(root->left);
1092 static void rb_erase_init(struct rb_node *n, struct rb_root *root)
1098 static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
1100 if (root->left == n)
1102 rb_erase_init(n, &root->rb);
1107 * would be nice to take fifo expire time into account as well
1109 static struct request *
1110 cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1111 struct request *last)
1113 struct rb_node *rbnext = rb_next(&last->rb_node);
1114 struct rb_node *rbprev = rb_prev(&last->rb_node);
1115 struct request *next = NULL, *prev = NULL;
1117 BUG_ON(RB_EMPTY_NODE(&last->rb_node));
1120 prev = rb_entry_rq(rbprev);
1123 next = rb_entry_rq(rbnext);
1125 rbnext = rb_first(&cfqq->sort_list);
1126 if (rbnext && rbnext != &last->rb_node)
1127 next = rb_entry_rq(rbnext);
1130 return cfq_choose_req(cfqd, next, prev, blk_rq_pos(last));
1133 static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
1134 struct cfq_queue *cfqq)
1137 * just an approximation, should be ok.
1139 return (cfqq->cfqg->nr_cfqq - 1) * (cfq_prio_slice(cfqd, 1, 0) -
1140 cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
1144 cfqg_key(struct cfq_rb_root *st, struct cfq_group *cfqg)
1146 return cfqg->vdisktime - st->min_vdisktime;
1150 __cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
1152 struct rb_node **node = &st->rb.rb_node;
1153 struct rb_node *parent = NULL;
1154 struct cfq_group *__cfqg;
1155 s64 key = cfqg_key(st, cfqg);
1158 while (*node != NULL) {
1160 __cfqg = rb_entry_cfqg(parent);
1162 if (key < cfqg_key(st, __cfqg))
1163 node = &parent->rb_left;
1165 node = &parent->rb_right;
1171 st->left = &cfqg->rb_node;
1173 rb_link_node(&cfqg->rb_node, parent, node);
1174 rb_insert_color(&cfqg->rb_node, &st->rb);
1178 cfq_update_group_weight(struct cfq_group *cfqg)
1180 BUG_ON(!RB_EMPTY_NODE(&cfqg->rb_node));
1181 if (cfqg->new_weight) {
1182 cfqg->weight = cfqg->new_weight;
1183 cfqg->new_weight = 0;
1188 cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
1190 BUG_ON(!RB_EMPTY_NODE(&cfqg->rb_node));
1192 cfq_update_group_weight(cfqg);
1193 __cfq_group_service_tree_add(st, cfqg);
1194 st->total_weight += cfqg->weight;
1198 cfq_group_notify_queue_add(struct cfq_data *cfqd, struct cfq_group *cfqg)
1200 struct cfq_rb_root *st = &cfqd->grp_service_tree;
1201 struct cfq_group *__cfqg;
1205 if (!RB_EMPTY_NODE(&cfqg->rb_node))
1209 * Currently put the group at the end. Later implement something
1210 * so that groups get lesser vtime based on their weights, so that
1211 * if group does not loose all if it was not continuously backlogged.
1213 n = rb_last(&st->rb);
1215 __cfqg = rb_entry_cfqg(n);
1216 cfqg->vdisktime = __cfqg->vdisktime + CFQ_IDLE_DELAY;
1218 cfqg->vdisktime = st->min_vdisktime;
1219 cfq_group_service_tree_add(st, cfqg);
1223 cfq_group_service_tree_del(struct cfq_rb_root *st, struct cfq_group *cfqg)
1225 st->total_weight -= cfqg->weight;
1226 if (!RB_EMPTY_NODE(&cfqg->rb_node))
1227 cfq_rb_erase(&cfqg->rb_node, st);
1231 cfq_group_notify_queue_del(struct cfq_data *cfqd, struct cfq_group *cfqg)
1233 struct cfq_rb_root *st = &cfqd->grp_service_tree;
1235 BUG_ON(cfqg->nr_cfqq < 1);
1238 /* If there are other cfq queues under this group, don't delete it */
1242 cfq_log_cfqg(cfqd, cfqg, "del_from_rr group");
1243 cfq_group_service_tree_del(st, cfqg);
1244 cfqg->saved_wl_slice = 0;
1245 cfqg_stats_update_dequeue(cfqg);
1248 static inline unsigned int cfq_cfqq_slice_usage(struct cfq_queue *cfqq,
1249 unsigned int *unaccounted_time)
1251 unsigned int slice_used;
1254 * Queue got expired before even a single request completed or
1255 * got expired immediately after first request completion.
1257 if (!cfqq->slice_start || cfqq->slice_start == jiffies) {
1259 * Also charge the seek time incurred to the group, otherwise
1260 * if there are mutiple queues in the group, each can dispatch
1261 * a single request on seeky media and cause lots of seek time
1262 * and group will never know it.
1264 slice_used = max_t(unsigned, (jiffies - cfqq->dispatch_start),
1267 slice_used = jiffies - cfqq->slice_start;
1268 if (slice_used > cfqq->allocated_slice) {
1269 *unaccounted_time = slice_used - cfqq->allocated_slice;
1270 slice_used = cfqq->allocated_slice;
1272 if (time_after(cfqq->slice_start, cfqq->dispatch_start))
1273 *unaccounted_time += cfqq->slice_start -
1274 cfqq->dispatch_start;
1280 static void cfq_group_served(struct cfq_data *cfqd, struct cfq_group *cfqg,
1281 struct cfq_queue *cfqq)
1283 struct cfq_rb_root *st = &cfqd->grp_service_tree;
1284 unsigned int used_sl, charge, unaccounted_sl = 0;
1285 int nr_sync = cfqg->nr_cfqq - cfqg_busy_async_queues(cfqd, cfqg)
1286 - cfqg->service_tree_idle.count;
1288 BUG_ON(nr_sync < 0);
1289 used_sl = charge = cfq_cfqq_slice_usage(cfqq, &unaccounted_sl);
1291 if (iops_mode(cfqd))
1292 charge = cfqq->slice_dispatch;
1293 else if (!cfq_cfqq_sync(cfqq) && !nr_sync)
1294 charge = cfqq->allocated_slice;
1296 /* Can't update vdisktime while group is on service tree */
1297 cfq_group_service_tree_del(st, cfqg);
1298 cfqg->vdisktime += cfq_scale_slice(charge, cfqg);
1299 /* If a new weight was requested, update now, off tree */
1300 cfq_group_service_tree_add(st, cfqg);
1302 /* This group is being expired. Save the context */
1303 if (time_after(cfqd->workload_expires, jiffies)) {
1304 cfqg->saved_wl_slice = cfqd->workload_expires
1306 cfqg->saved_wl_type = cfqd->serving_wl_type;
1307 cfqg->saved_wl_class = cfqd->serving_wl_class;
1309 cfqg->saved_wl_slice = 0;
1311 cfq_log_cfqg(cfqd, cfqg, "served: vt=%llu min_vt=%llu", cfqg->vdisktime,
1313 cfq_log_cfqq(cfqq->cfqd, cfqq,
1314 "sl_used=%u disp=%u charge=%u iops=%u sect=%lu",
1315 used_sl, cfqq->slice_dispatch, charge,
1316 iops_mode(cfqd), cfqq->nr_sectors);
1317 cfqg_stats_update_timeslice_used(cfqg, used_sl, unaccounted_sl);
1318 cfqg_stats_set_start_empty_time(cfqg);
1322 * cfq_init_cfqg_base - initialize base part of a cfq_group
1323 * @cfqg: cfq_group to initialize
1325 * Initialize the base part which is used whether %CONFIG_CFQ_GROUP_IOSCHED
1326 * is enabled or not.
1328 static void cfq_init_cfqg_base(struct cfq_group *cfqg)
1330 struct cfq_rb_root *st;
1333 for_each_cfqg_st(cfqg, i, j, st)
1335 RB_CLEAR_NODE(&cfqg->rb_node);
1337 cfqg->ttime.last_end_request = jiffies;
1340 #ifdef CONFIG_CFQ_GROUP_IOSCHED
1341 static void cfq_pd_init(struct blkcg_gq *blkg)
1343 struct cfq_group *cfqg = blkg_to_cfqg(blkg);
1345 cfq_init_cfqg_base(cfqg);
1346 cfqg->weight = blkg->blkcg->cfq_weight;
1350 * Search for the cfq group current task belongs to. request_queue lock must
1353 static struct cfq_group *cfq_lookup_create_cfqg(struct cfq_data *cfqd,
1354 struct blkcg *blkcg)
1356 struct request_queue *q = cfqd->queue;
1357 struct cfq_group *cfqg = NULL;
1359 /* avoid lookup for the common case where there's no blkcg */
1360 if (blkcg == &blkcg_root) {
1361 cfqg = cfqd->root_group;
1363 struct blkcg_gq *blkg;
1365 blkg = blkg_lookup_create(blkcg, q);
1367 cfqg = blkg_to_cfqg(blkg);
1373 static void cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg)
1375 /* Currently, all async queues are mapped to root group */
1376 if (!cfq_cfqq_sync(cfqq))
1377 cfqg = cfqq->cfqd->root_group;
1380 /* cfqq reference on cfqg */
1384 static u64 cfqg_prfill_weight_device(struct seq_file *sf,
1385 struct blkg_policy_data *pd, int off)
1387 struct cfq_group *cfqg = pd_to_cfqg(pd);
1389 if (!cfqg->dev_weight)
1391 return __blkg_prfill_u64(sf, pd, cfqg->dev_weight);
1394 static int cfqg_print_weight_device(struct cgroup *cgrp, struct cftype *cft,
1395 struct seq_file *sf)
1397 blkcg_print_blkgs(sf, cgroup_to_blkcg(cgrp),
1398 cfqg_prfill_weight_device, &blkcg_policy_cfq, 0,
1403 static int cfq_print_weight(struct cgroup *cgrp, struct cftype *cft,
1404 struct seq_file *sf)
1406 seq_printf(sf, "%u\n", cgroup_to_blkcg(cgrp)->cfq_weight);
1410 static int cfqg_set_weight_device(struct cgroup *cgrp, struct cftype *cft,
1413 struct blkcg *blkcg = cgroup_to_blkcg(cgrp);
1414 struct blkg_conf_ctx ctx;
1415 struct cfq_group *cfqg;
1418 ret = blkg_conf_prep(blkcg, &blkcg_policy_cfq, buf, &ctx);
1423 cfqg = blkg_to_cfqg(ctx.blkg);
1424 if (!ctx.v || (ctx.v >= CFQ_WEIGHT_MIN && ctx.v <= CFQ_WEIGHT_MAX)) {
1425 cfqg->dev_weight = ctx.v;
1426 cfqg->new_weight = cfqg->dev_weight ?: blkcg->cfq_weight;
1430 blkg_conf_finish(&ctx);
1434 static int cfq_set_weight(struct cgroup *cgrp, struct cftype *cft, u64 val)
1436 struct blkcg *blkcg = cgroup_to_blkcg(cgrp);
1437 struct blkcg_gq *blkg;
1438 struct hlist_node *n;
1440 if (val < CFQ_WEIGHT_MIN || val > CFQ_WEIGHT_MAX)
1443 spin_lock_irq(&blkcg->lock);
1444 blkcg->cfq_weight = (unsigned int)val;
1446 hlist_for_each_entry(blkg, n, &blkcg->blkg_list, blkcg_node) {
1447 struct cfq_group *cfqg = blkg_to_cfqg(blkg);
1449 if (cfqg && !cfqg->dev_weight)
1450 cfqg->new_weight = blkcg->cfq_weight;
1453 spin_unlock_irq(&blkcg->lock);
1457 static int cfqg_print_stat(struct cgroup *cgrp, struct cftype *cft,
1458 struct seq_file *sf)
1460 struct blkcg *blkcg = cgroup_to_blkcg(cgrp);
1462 blkcg_print_blkgs(sf, blkcg, blkg_prfill_stat, &blkcg_policy_cfq,
1463 cft->private, false);
1467 static int cfqg_print_rwstat(struct cgroup *cgrp, struct cftype *cft,
1468 struct seq_file *sf)
1470 struct blkcg *blkcg = cgroup_to_blkcg(cgrp);
1472 blkcg_print_blkgs(sf, blkcg, blkg_prfill_rwstat, &blkcg_policy_cfq,
1473 cft->private, true);
1477 #ifdef CONFIG_DEBUG_BLK_CGROUP
1478 static u64 cfqg_prfill_avg_queue_size(struct seq_file *sf,
1479 struct blkg_policy_data *pd, int off)
1481 struct cfq_group *cfqg = pd_to_cfqg(pd);
1482 u64 samples = blkg_stat_read(&cfqg->stats.avg_queue_size_samples);
1486 v = blkg_stat_read(&cfqg->stats.avg_queue_size_sum);
1489 __blkg_prfill_u64(sf, pd, v);
1493 /* print avg_queue_size */
1494 static int cfqg_print_avg_queue_size(struct cgroup *cgrp, struct cftype *cft,
1495 struct seq_file *sf)
1497 struct blkcg *blkcg = cgroup_to_blkcg(cgrp);
1499 blkcg_print_blkgs(sf, blkcg, cfqg_prfill_avg_queue_size,
1500 &blkcg_policy_cfq, 0, false);
1503 #endif /* CONFIG_DEBUG_BLK_CGROUP */
1505 static struct cftype cfq_blkcg_files[] = {
1507 .name = "weight_device",
1508 .read_seq_string = cfqg_print_weight_device,
1509 .write_string = cfqg_set_weight_device,
1510 .max_write_len = 256,
1514 .read_seq_string = cfq_print_weight,
1515 .write_u64 = cfq_set_weight,
1519 .private = offsetof(struct cfq_group, stats.time),
1520 .read_seq_string = cfqg_print_stat,
1524 .private = offsetof(struct cfq_group, stats.sectors),
1525 .read_seq_string = cfqg_print_stat,
1528 .name = "io_service_bytes",
1529 .private = offsetof(struct cfq_group, stats.service_bytes),
1530 .read_seq_string = cfqg_print_rwstat,
1533 .name = "io_serviced",
1534 .private = offsetof(struct cfq_group, stats.serviced),
1535 .read_seq_string = cfqg_print_rwstat,
1538 .name = "io_service_time",
1539 .private = offsetof(struct cfq_group, stats.service_time),
1540 .read_seq_string = cfqg_print_rwstat,
1543 .name = "io_wait_time",
1544 .private = offsetof(struct cfq_group, stats.wait_time),
1545 .read_seq_string = cfqg_print_rwstat,
1548 .name = "io_merged",
1549 .private = offsetof(struct cfq_group, stats.merged),
1550 .read_seq_string = cfqg_print_rwstat,
1553 .name = "io_queued",
1554 .private = offsetof(struct cfq_group, stats.queued),
1555 .read_seq_string = cfqg_print_rwstat,
1557 #ifdef CONFIG_DEBUG_BLK_CGROUP
1559 .name = "avg_queue_size",
1560 .read_seq_string = cfqg_print_avg_queue_size,
1563 .name = "group_wait_time",
1564 .private = offsetof(struct cfq_group, stats.group_wait_time),
1565 .read_seq_string = cfqg_print_stat,
1568 .name = "idle_time",
1569 .private = offsetof(struct cfq_group, stats.idle_time),
1570 .read_seq_string = cfqg_print_stat,
1573 .name = "empty_time",
1574 .private = offsetof(struct cfq_group, stats.empty_time),
1575 .read_seq_string = cfqg_print_stat,
1579 .private = offsetof(struct cfq_group, stats.dequeue),
1580 .read_seq_string = cfqg_print_stat,
1583 .name = "unaccounted_time",
1584 .private = offsetof(struct cfq_group, stats.unaccounted_time),
1585 .read_seq_string = cfqg_print_stat,
1587 #endif /* CONFIG_DEBUG_BLK_CGROUP */
1590 #else /* GROUP_IOSCHED */
1591 static struct cfq_group *cfq_lookup_create_cfqg(struct cfq_data *cfqd,
1592 struct blkcg *blkcg)
1594 return cfqd->root_group;
1598 cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) {
1602 #endif /* GROUP_IOSCHED */
1605 * The cfqd->service_trees holds all pending cfq_queue's that have
1606 * requests waiting to be processed. It is sorted in the order that
1607 * we will service the queues.
1609 static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1612 struct rb_node **p, *parent;
1613 struct cfq_queue *__cfqq;
1614 unsigned long rb_key;
1615 struct cfq_rb_root *service_tree;
1619 service_tree = service_tree_for(cfqq->cfqg, cfqq_class(cfqq),
1621 if (cfq_class_idle(cfqq)) {
1622 rb_key = CFQ_IDLE_DELAY;
1623 parent = rb_last(&service_tree->rb);
1624 if (parent && parent != &cfqq->rb_node) {
1625 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1626 rb_key += __cfqq->rb_key;
1629 } else if (!add_front) {
1631 * Get our rb key offset. Subtract any residual slice
1632 * value carried from last service. A negative resid
1633 * count indicates slice overrun, and this should position
1634 * the next service time further away in the tree.
1636 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
1637 rb_key -= cfqq->slice_resid;
1638 cfqq->slice_resid = 0;
1641 __cfqq = cfq_rb_first(service_tree);
1642 rb_key += __cfqq ? __cfqq->rb_key : jiffies;
1645 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1648 * same position, nothing more to do
1650 if (rb_key == cfqq->rb_key &&
1651 cfqq->service_tree == service_tree)
1654 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1655 cfqq->service_tree = NULL;
1660 cfqq->service_tree = service_tree;
1661 p = &service_tree->rb.rb_node;
1666 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1669 * sort by key, that represents service time.
1671 if (time_before(rb_key, __cfqq->rb_key))
1674 n = &(*p)->rb_right;
1682 service_tree->left = &cfqq->rb_node;
1684 cfqq->rb_key = rb_key;
1685 rb_link_node(&cfqq->rb_node, parent, p);
1686 rb_insert_color(&cfqq->rb_node, &service_tree->rb);
1687 service_tree->count++;
1688 if (add_front || !new_cfqq)
1690 cfq_group_notify_queue_add(cfqd, cfqq->cfqg);
1693 static struct cfq_queue *
1694 cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
1695 sector_t sector, struct rb_node **ret_parent,
1696 struct rb_node ***rb_link)
1698 struct rb_node **p, *parent;
1699 struct cfq_queue *cfqq = NULL;
1707 cfqq = rb_entry(parent, struct cfq_queue, p_node);
1710 * Sort strictly based on sector. Smallest to the left,
1711 * largest to the right.
1713 if (sector > blk_rq_pos(cfqq->next_rq))
1714 n = &(*p)->rb_right;
1715 else if (sector < blk_rq_pos(cfqq->next_rq))
1723 *ret_parent = parent;
1729 static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1731 struct rb_node **p, *parent;
1732 struct cfq_queue *__cfqq;
1735 rb_erase(&cfqq->p_node, cfqq->p_root);
1736 cfqq->p_root = NULL;
1739 if (cfq_class_idle(cfqq))
1744 cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
1745 __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
1746 blk_rq_pos(cfqq->next_rq), &parent, &p);
1748 rb_link_node(&cfqq->p_node, parent, p);
1749 rb_insert_color(&cfqq->p_node, cfqq->p_root);
1751 cfqq->p_root = NULL;
1755 * Update cfqq's position in the service tree.
1757 static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1760 * Resorting requires the cfqq to be on the RR list already.
1762 if (cfq_cfqq_on_rr(cfqq)) {
1763 cfq_service_tree_add(cfqd, cfqq, 0);
1764 cfq_prio_tree_add(cfqd, cfqq);
1769 * add to busy list of queues for service, trying to be fair in ordering
1770 * the pending list according to last request service
1772 static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1774 cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
1775 BUG_ON(cfq_cfqq_on_rr(cfqq));
1776 cfq_mark_cfqq_on_rr(cfqq);
1777 cfqd->busy_queues++;
1778 if (cfq_cfqq_sync(cfqq))
1779 cfqd->busy_sync_queues++;
1781 cfq_resort_rr_list(cfqd, cfqq);
1785 * Called when the cfqq no longer has requests pending, remove it from
1788 static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1790 cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
1791 BUG_ON(!cfq_cfqq_on_rr(cfqq));
1792 cfq_clear_cfqq_on_rr(cfqq);
1794 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1795 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1796 cfqq->service_tree = NULL;
1799 rb_erase(&cfqq->p_node, cfqq->p_root);
1800 cfqq->p_root = NULL;
1803 cfq_group_notify_queue_del(cfqd, cfqq->cfqg);
1804 BUG_ON(!cfqd->busy_queues);
1805 cfqd->busy_queues--;
1806 if (cfq_cfqq_sync(cfqq))
1807 cfqd->busy_sync_queues--;
1811 * rb tree support functions
1813 static void cfq_del_rq_rb(struct request *rq)
1815 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1816 const int sync = rq_is_sync(rq);
1818 BUG_ON(!cfqq->queued[sync]);
1819 cfqq->queued[sync]--;
1821 elv_rb_del(&cfqq->sort_list, rq);
1823 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) {
1825 * Queue will be deleted from service tree when we actually
1826 * expire it later. Right now just remove it from prio tree
1830 rb_erase(&cfqq->p_node, cfqq->p_root);
1831 cfqq->p_root = NULL;
1836 static void cfq_add_rq_rb(struct request *rq)
1838 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1839 struct cfq_data *cfqd = cfqq->cfqd;
1840 struct request *prev;
1842 cfqq->queued[rq_is_sync(rq)]++;
1844 elv_rb_add(&cfqq->sort_list, rq);
1846 if (!cfq_cfqq_on_rr(cfqq))
1847 cfq_add_cfqq_rr(cfqd, cfqq);
1850 * check if this request is a better next-serve candidate
1852 prev = cfqq->next_rq;
1853 cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq, cfqd->last_position);
1856 * adjust priority tree position, if ->next_rq changes
1858 if (prev != cfqq->next_rq)
1859 cfq_prio_tree_add(cfqd, cfqq);
1861 BUG_ON(!cfqq->next_rq);
1864 static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
1866 elv_rb_del(&cfqq->sort_list, rq);
1867 cfqq->queued[rq_is_sync(rq)]--;
1868 cfqg_stats_update_io_remove(RQ_CFQG(rq), rq->cmd_flags);
1870 cfqg_stats_update_io_add(RQ_CFQG(rq), cfqq->cfqd->serving_group,
1874 static struct request *
1875 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
1877 struct task_struct *tsk = current;
1878 struct cfq_io_cq *cic;
1879 struct cfq_queue *cfqq;
1881 cic = cfq_cic_lookup(cfqd, tsk->io_context);
1885 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
1887 sector_t sector = bio->bi_sector + bio_sectors(bio);
1889 return elv_rb_find(&cfqq->sort_list, sector);
1895 static void cfq_activate_request(struct request_queue *q, struct request *rq)
1897 struct cfq_data *cfqd = q->elevator->elevator_data;
1899 cfqd->rq_in_driver++;
1900 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
1901 cfqd->rq_in_driver);
1903 cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
1906 static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
1908 struct cfq_data *cfqd = q->elevator->elevator_data;
1910 WARN_ON(!cfqd->rq_in_driver);
1911 cfqd->rq_in_driver--;
1912 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
1913 cfqd->rq_in_driver);
1916 static void cfq_remove_request(struct request *rq)
1918 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1920 if (cfqq->next_rq == rq)
1921 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
1923 list_del_init(&rq->queuelist);
1926 cfqq->cfqd->rq_queued--;
1927 cfqg_stats_update_io_remove(RQ_CFQG(rq), rq->cmd_flags);
1928 if (rq->cmd_flags & REQ_PRIO) {
1929 WARN_ON(!cfqq->prio_pending);
1930 cfqq->prio_pending--;
1934 static int cfq_merge(struct request_queue *q, struct request **req,
1937 struct cfq_data *cfqd = q->elevator->elevator_data;
1938 struct request *__rq;
1940 __rq = cfq_find_rq_fmerge(cfqd, bio);
1941 if (__rq && elv_rq_merge_ok(__rq, bio)) {
1943 return ELEVATOR_FRONT_MERGE;
1946 return ELEVATOR_NO_MERGE;
1949 static void cfq_merged_request(struct request_queue *q, struct request *req,
1952 if (type == ELEVATOR_FRONT_MERGE) {
1953 struct cfq_queue *cfqq = RQ_CFQQ(req);
1955 cfq_reposition_rq_rb(cfqq, req);
1959 static void cfq_bio_merged(struct request_queue *q, struct request *req,
1962 cfqg_stats_update_io_merged(RQ_CFQG(req), bio->bi_rw);
1966 cfq_merged_requests(struct request_queue *q, struct request *rq,
1967 struct request *next)
1969 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1970 struct cfq_data *cfqd = q->elevator->elevator_data;
1973 * reposition in fifo if next is older than rq
1975 if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
1976 time_before(rq_fifo_time(next), rq_fifo_time(rq)) &&
1977 cfqq == RQ_CFQQ(next)) {
1978 list_move(&rq->queuelist, &next->queuelist);
1979 rq_set_fifo_time(rq, rq_fifo_time(next));
1982 if (cfqq->next_rq == next)
1984 cfq_remove_request(next);
1985 cfqg_stats_update_io_merged(RQ_CFQG(rq), next->cmd_flags);
1987 cfqq = RQ_CFQQ(next);
1989 * all requests of this queue are merged to other queues, delete it
1990 * from the service tree. If it's the active_queue,
1991 * cfq_dispatch_requests() will choose to expire it or do idle
1993 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list) &&
1994 cfqq != cfqd->active_queue)
1995 cfq_del_cfqq_rr(cfqd, cfqq);
1998 static int cfq_allow_merge(struct request_queue *q, struct request *rq,
2001 struct cfq_data *cfqd = q->elevator->elevator_data;
2002 struct cfq_io_cq *cic;
2003 struct cfq_queue *cfqq;
2006 * Disallow merge of a sync bio into an async request.
2008 if (cfq_bio_sync(bio) && !rq_is_sync(rq))
2012 * Lookup the cfqq that this bio will be queued with and allow
2013 * merge only if rq is queued there.
2015 cic = cfq_cic_lookup(cfqd, current->io_context);
2019 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
2020 return cfqq == RQ_CFQQ(rq);
2023 static inline void cfq_del_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2025 del_timer(&cfqd->idle_slice_timer);
2026 cfqg_stats_update_idle_time(cfqq->cfqg);
2029 static void __cfq_set_active_queue(struct cfq_data *cfqd,
2030 struct cfq_queue *cfqq)
2033 cfq_log_cfqq(cfqd, cfqq, "set_active wl_class:%d wl_type:%d",
2034 cfqd->serving_wl_class, cfqd->serving_wl_type);
2035 cfqg_stats_update_avg_queue_size(cfqq->cfqg);
2036 cfqq->slice_start = 0;
2037 cfqq->dispatch_start = jiffies;
2038 cfqq->allocated_slice = 0;
2039 cfqq->slice_end = 0;
2040 cfqq->slice_dispatch = 0;
2041 cfqq->nr_sectors = 0;
2043 cfq_clear_cfqq_wait_request(cfqq);
2044 cfq_clear_cfqq_must_dispatch(cfqq);
2045 cfq_clear_cfqq_must_alloc_slice(cfqq);
2046 cfq_clear_cfqq_fifo_expire(cfqq);
2047 cfq_mark_cfqq_slice_new(cfqq);
2049 cfq_del_timer(cfqd, cfqq);
2052 cfqd->active_queue = cfqq;
2056 * current cfqq expired its slice (or was too idle), select new one
2059 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2062 cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
2064 if (cfq_cfqq_wait_request(cfqq))
2065 cfq_del_timer(cfqd, cfqq);
2067 cfq_clear_cfqq_wait_request(cfqq);
2068 cfq_clear_cfqq_wait_busy(cfqq);
2071 * If this cfqq is shared between multiple processes, check to
2072 * make sure that those processes are still issuing I/Os within
2073 * the mean seek distance. If not, it may be time to break the
2074 * queues apart again.
2076 if (cfq_cfqq_coop(cfqq) && CFQQ_SEEKY(cfqq))
2077 cfq_mark_cfqq_split_coop(cfqq);
2080 * store what was left of this slice, if the queue idled/timed out
2083 if (cfq_cfqq_slice_new(cfqq))
2084 cfqq->slice_resid = cfq_scaled_cfqq_slice(cfqd, cfqq);
2086 cfqq->slice_resid = cfqq->slice_end - jiffies;
2087 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
2090 cfq_group_served(cfqd, cfqq->cfqg, cfqq);
2092 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
2093 cfq_del_cfqq_rr(cfqd, cfqq);
2095 cfq_resort_rr_list(cfqd, cfqq);
2097 if (cfqq == cfqd->active_queue)
2098 cfqd->active_queue = NULL;
2100 if (cfqd->active_cic) {
2101 put_io_context(cfqd->active_cic->icq.ioc);
2102 cfqd->active_cic = NULL;
2106 static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out)
2108 struct cfq_queue *cfqq = cfqd->active_queue;
2111 __cfq_slice_expired(cfqd, cfqq, timed_out);
2115 * Get next queue for service. Unless we have a queue preemption,
2116 * we'll simply select the first cfqq in the service tree.
2118 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
2120 struct cfq_rb_root *service_tree =
2121 service_tree_for(cfqd->serving_group, cfqd->serving_wl_class,
2122 cfqd->serving_wl_type);
2124 if (!cfqd->rq_queued)
2127 /* There is nothing to dispatch */
2130 if (RB_EMPTY_ROOT(&service_tree->rb))
2132 return cfq_rb_first(service_tree);
2135 static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd)
2137 struct cfq_group *cfqg;
2138 struct cfq_queue *cfqq;
2140 struct cfq_rb_root *st;
2142 if (!cfqd->rq_queued)
2145 cfqg = cfq_get_next_cfqg(cfqd);
2149 for_each_cfqg_st(cfqg, i, j, st)
2150 if ((cfqq = cfq_rb_first(st)) != NULL)
2156 * Get and set a new active queue for service.
2158 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
2159 struct cfq_queue *cfqq)
2162 cfqq = cfq_get_next_queue(cfqd);
2164 __cfq_set_active_queue(cfqd, cfqq);
2168 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
2171 if (blk_rq_pos(rq) >= cfqd->last_position)
2172 return blk_rq_pos(rq) - cfqd->last_position;
2174 return cfqd->last_position - blk_rq_pos(rq);
2177 static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2180 return cfq_dist_from_last(cfqd, rq) <= CFQQ_CLOSE_THR;
2183 static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
2184 struct cfq_queue *cur_cfqq)
2186 struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
2187 struct rb_node *parent, *node;
2188 struct cfq_queue *__cfqq;
2189 sector_t sector = cfqd->last_position;
2191 if (RB_EMPTY_ROOT(root))
2195 * First, if we find a request starting at the end of the last
2196 * request, choose it.
2198 __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
2203 * If the exact sector wasn't found, the parent of the NULL leaf
2204 * will contain the closest sector.
2206 __cfqq = rb_entry(parent, struct cfq_queue, p_node);
2207 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
2210 if (blk_rq_pos(__cfqq->next_rq) < sector)
2211 node = rb_next(&__cfqq->p_node);
2213 node = rb_prev(&__cfqq->p_node);
2217 __cfqq = rb_entry(node, struct cfq_queue, p_node);
2218 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
2226 * cur_cfqq - passed in so that we don't decide that the current queue is
2227 * closely cooperating with itself.
2229 * So, basically we're assuming that that cur_cfqq has dispatched at least
2230 * one request, and that cfqd->last_position reflects a position on the disk
2231 * associated with the I/O issued by cur_cfqq. I'm not sure this is a valid
2234 static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
2235 struct cfq_queue *cur_cfqq)
2237 struct cfq_queue *cfqq;
2239 if (cfq_class_idle(cur_cfqq))
2241 if (!cfq_cfqq_sync(cur_cfqq))
2243 if (CFQQ_SEEKY(cur_cfqq))
2247 * Don't search priority tree if it's the only queue in the group.
2249 if (cur_cfqq->cfqg->nr_cfqq == 1)
2253 * We should notice if some of the queues are cooperating, eg
2254 * working closely on the same area of the disk. In that case,
2255 * we can group them together and don't waste time idling.
2257 cfqq = cfqq_close(cfqd, cur_cfqq);
2261 /* If new queue belongs to different cfq_group, don't choose it */
2262 if (cur_cfqq->cfqg != cfqq->cfqg)
2266 * It only makes sense to merge sync queues.
2268 if (!cfq_cfqq_sync(cfqq))
2270 if (CFQQ_SEEKY(cfqq))
2274 * Do not merge queues of different priority classes
2276 if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq))
2283 * Determine whether we should enforce idle window for this queue.
2286 static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2288 enum wl_class_t wl_class = cfqq_class(cfqq);
2289 struct cfq_rb_root *service_tree = cfqq->service_tree;
2291 BUG_ON(!service_tree);
2292 BUG_ON(!service_tree->count);
2294 if (!cfqd->cfq_slice_idle)
2297 /* We never do for idle class queues. */
2298 if (wl_class == IDLE_WORKLOAD)
2301 /* We do for queues that were marked with idle window flag. */
2302 if (cfq_cfqq_idle_window(cfqq) &&
2303 !(blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag))
2307 * Otherwise, we do only if they are the last ones
2308 * in their service tree.
2310 if (service_tree->count == 1 && cfq_cfqq_sync(cfqq) &&
2311 !cfq_io_thinktime_big(cfqd, &service_tree->ttime, false))
2313 cfq_log_cfqq(cfqd, cfqq, "Not idling. st->count:%d",
2314 service_tree->count);
2318 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
2320 struct cfq_queue *cfqq = cfqd->active_queue;
2321 struct cfq_io_cq *cic;
2322 unsigned long sl, group_idle = 0;
2325 * SSD device without seek penalty, disable idling. But only do so
2326 * for devices that support queuing, otherwise we still have a problem
2327 * with sync vs async workloads.
2329 if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
2332 WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
2333 WARN_ON(cfq_cfqq_slice_new(cfqq));
2336 * idle is disabled, either manually or by past process history
2338 if (!cfq_should_idle(cfqd, cfqq)) {
2339 /* no queue idling. Check for group idling */
2340 if (cfqd->cfq_group_idle)
2341 group_idle = cfqd->cfq_group_idle;
2347 * still active requests from this queue, don't idle
2349 if (cfqq->dispatched)
2353 * task has exited, don't wait
2355 cic = cfqd->active_cic;
2356 if (!cic || !atomic_read(&cic->icq.ioc->active_ref))
2360 * If our average think time is larger than the remaining time
2361 * slice, then don't idle. This avoids overrunning the allotted
2364 if (sample_valid(cic->ttime.ttime_samples) &&
2365 (cfqq->slice_end - jiffies < cic->ttime.ttime_mean)) {
2366 cfq_log_cfqq(cfqd, cfqq, "Not idling. think_time:%lu",
2367 cic->ttime.ttime_mean);
2371 /* There are other queues in the group, don't do group idle */
2372 if (group_idle && cfqq->cfqg->nr_cfqq > 1)
2375 cfq_mark_cfqq_wait_request(cfqq);
2378 sl = cfqd->cfq_group_idle;
2380 sl = cfqd->cfq_slice_idle;
2382 mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
2383 cfqg_stats_set_start_idle_time(cfqq->cfqg);
2384 cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu group_idle: %d", sl,
2385 group_idle ? 1 : 0);
2389 * Move request from internal lists to the request queue dispatch list.
2391 static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
2393 struct cfq_data *cfqd = q->elevator->elevator_data;
2394 struct cfq_queue *cfqq = RQ_CFQQ(rq);
2396 cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
2398 cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
2399 cfq_remove_request(rq);
2401 (RQ_CFQG(rq))->dispatched++;
2402 elv_dispatch_sort(q, rq);
2404 cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]++;
2405 cfqq->nr_sectors += blk_rq_sectors(rq);
2406 cfqg_stats_update_dispatch(cfqq->cfqg, blk_rq_bytes(rq), rq->cmd_flags);
2410 * return expired entry, or NULL to just start from scratch in rbtree
2412 static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
2414 struct request *rq = NULL;
2416 if (cfq_cfqq_fifo_expire(cfqq))
2419 cfq_mark_cfqq_fifo_expire(cfqq);
2421 if (list_empty(&cfqq->fifo))
2424 rq = rq_entry_fifo(cfqq->fifo.next);
2425 if (time_before(jiffies, rq_fifo_time(rq)))
2428 cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq);
2433 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2435 const int base_rq = cfqd->cfq_slice_async_rq;
2437 WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
2439 return 2 * base_rq * (IOPRIO_BE_NR - cfqq->ioprio);
2443 * Must be called with the queue_lock held.
2445 static int cfqq_process_refs(struct cfq_queue *cfqq)
2447 int process_refs, io_refs;
2449 io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE];
2450 process_refs = cfqq->ref - io_refs;
2451 BUG_ON(process_refs < 0);
2452 return process_refs;
2455 static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq)
2457 int process_refs, new_process_refs;
2458 struct cfq_queue *__cfqq;
2461 * If there are no process references on the new_cfqq, then it is
2462 * unsafe to follow the ->new_cfqq chain as other cfqq's in the
2463 * chain may have dropped their last reference (not just their
2464 * last process reference).
2466 if (!cfqq_process_refs(new_cfqq))
2469 /* Avoid a circular list and skip interim queue merges */
2470 while ((__cfqq = new_cfqq->new_cfqq)) {
2476 process_refs = cfqq_process_refs(cfqq);
2477 new_process_refs = cfqq_process_refs(new_cfqq);
2479 * If the process for the cfqq has gone away, there is no
2480 * sense in merging the queues.
2482 if (process_refs == 0 || new_process_refs == 0)
2486 * Merge in the direction of the lesser amount of work.
2488 if (new_process_refs >= process_refs) {
2489 cfqq->new_cfqq = new_cfqq;
2490 new_cfqq->ref += process_refs;
2492 new_cfqq->new_cfqq = cfqq;
2493 cfqq->ref += new_process_refs;
2497 static enum wl_type_t cfq_choose_wl(struct cfq_data *cfqd,
2498 struct cfq_group *cfqg, enum wl_class_t wl_class)
2500 struct cfq_queue *queue;
2502 bool key_valid = false;
2503 unsigned long lowest_key = 0;
2504 enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD;
2506 for (i = 0; i <= SYNC_WORKLOAD; ++i) {
2507 /* select the one with lowest rb_key */
2508 queue = cfq_rb_first(service_tree_for(cfqg, wl_class, i));
2510 (!key_valid || time_before(queue->rb_key, lowest_key))) {
2511 lowest_key = queue->rb_key;
2520 static void choose_service_tree(struct cfq_data *cfqd, struct cfq_group *cfqg)
2524 struct cfq_rb_root *st;
2525 unsigned group_slice;
2526 enum wl_class_t original_class = cfqd->serving_wl_class;
2528 /* Choose next priority. RT > BE > IDLE */
2529 if (cfq_group_busy_queues_wl(RT_WORKLOAD, cfqd, cfqg))
2530 cfqd->serving_wl_class = RT_WORKLOAD;
2531 else if (cfq_group_busy_queues_wl(BE_WORKLOAD, cfqd, cfqg))
2532 cfqd->serving_wl_class = BE_WORKLOAD;
2534 cfqd->serving_wl_class = IDLE_WORKLOAD;
2535 cfqd->workload_expires = jiffies + 1;
2539 if (original_class != cfqd->serving_wl_class)
2543 * For RT and BE, we have to choose also the type
2544 * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
2547 st = service_tree_for(cfqg, cfqd->serving_wl_class,
2548 cfqd->serving_wl_type);
2552 * check workload expiration, and that we still have other queues ready
2554 if (count && !time_after(jiffies, cfqd->workload_expires))
2558 /* otherwise select new workload type */
2559 cfqd->serving_wl_type = cfq_choose_wl(cfqd, cfqg,
2560 cfqd->serving_wl_class);
2561 st = service_tree_for(cfqg, cfqd->serving_wl_class,
2562 cfqd->serving_wl_type);
2566 * the workload slice is computed as a fraction of target latency
2567 * proportional to the number of queues in that workload, over
2568 * all the queues in the same priority class
2570 group_slice = cfq_group_slice(cfqd, cfqg);
2572 slice = group_slice * count /
2573 max_t(unsigned, cfqg->busy_queues_avg[cfqd->serving_wl_class],
2574 cfq_group_busy_queues_wl(cfqd->serving_wl_class, cfqd,
2577 if (cfqd->serving_wl_type == ASYNC_WORKLOAD) {
2581 * Async queues are currently system wide. Just taking
2582 * proportion of queues with-in same group will lead to higher
2583 * async ratio system wide as generally root group is going
2584 * to have higher weight. A more accurate thing would be to
2585 * calculate system wide asnc/sync ratio.
2587 tmp = cfqd->cfq_target_latency *
2588 cfqg_busy_async_queues(cfqd, cfqg);
2589 tmp = tmp/cfqd->busy_queues;
2590 slice = min_t(unsigned, slice, tmp);
2592 /* async workload slice is scaled down according to
2593 * the sync/async slice ratio. */
2594 slice = slice * cfqd->cfq_slice[0] / cfqd->cfq_slice[1];
2596 /* sync workload slice is at least 2 * cfq_slice_idle */
2597 slice = max(slice, 2 * cfqd->cfq_slice_idle);
2599 slice = max_t(unsigned, slice, CFQ_MIN_TT);
2600 cfq_log(cfqd, "workload slice:%d", slice);
2601 cfqd->workload_expires = jiffies + slice;
2604 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd)
2606 struct cfq_rb_root *st = &cfqd->grp_service_tree;
2607 struct cfq_group *cfqg;
2609 if (RB_EMPTY_ROOT(&st->rb))
2611 cfqg = cfq_rb_first_group(st);
2612 update_min_vdisktime(st);
2616 static void cfq_choose_cfqg(struct cfq_data *cfqd)
2618 struct cfq_group *cfqg = cfq_get_next_cfqg(cfqd);
2620 cfqd->serving_group = cfqg;
2622 /* Restore the workload type data */
2623 if (cfqg->saved_wl_slice) {
2624 cfqd->workload_expires = jiffies + cfqg->saved_wl_slice;
2625 cfqd->serving_wl_type = cfqg->saved_wl_type;
2626 cfqd->serving_wl_class = cfqg->saved_wl_class;
2628 cfqd->workload_expires = jiffies - 1;
2630 choose_service_tree(cfqd, cfqg);
2634 * Select a queue for service. If we have a current active queue,
2635 * check whether to continue servicing it, or retrieve and set a new one.
2637 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
2639 struct cfq_queue *cfqq, *new_cfqq = NULL;
2641 cfqq = cfqd->active_queue;
2645 if (!cfqd->rq_queued)
2649 * We were waiting for group to get backlogged. Expire the queue
2651 if (cfq_cfqq_wait_busy(cfqq) && !RB_EMPTY_ROOT(&cfqq->sort_list))
2655 * The active queue has run out of time, expire it and select new.
2657 if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq)) {
2659 * If slice had not expired at the completion of last request
2660 * we might not have turned on wait_busy flag. Don't expire
2661 * the queue yet. Allow the group to get backlogged.
2663 * The very fact that we have used the slice, that means we
2664 * have been idling all along on this queue and it should be
2665 * ok to wait for this request to complete.
2667 if (cfqq->cfqg->nr_cfqq == 1 && RB_EMPTY_ROOT(&cfqq->sort_list)
2668 && cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2672 goto check_group_idle;
2676 * The active queue has requests and isn't expired, allow it to
2679 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
2683 * If another queue has a request waiting within our mean seek
2684 * distance, let it run. The expire code will check for close
2685 * cooperators and put the close queue at the front of the service
2686 * tree. If possible, merge the expiring queue with the new cfqq.
2688 new_cfqq = cfq_close_cooperator(cfqd, cfqq);
2690 if (!cfqq->new_cfqq)
2691 cfq_setup_merge(cfqq, new_cfqq);
2696 * No requests pending. If the active queue still has requests in
2697 * flight or is idling for a new request, allow either of these
2698 * conditions to happen (or time out) before selecting a new queue.
2700 if (timer_pending(&cfqd->idle_slice_timer)) {
2706 * This is a deep seek queue, but the device is much faster than
2707 * the queue can deliver, don't idle
2709 if (CFQQ_SEEKY(cfqq) && cfq_cfqq_idle_window(cfqq) &&
2710 (cfq_cfqq_slice_new(cfqq) ||
2711 (cfqq->slice_end - jiffies > jiffies - cfqq->slice_start))) {
2712 cfq_clear_cfqq_deep(cfqq);
2713 cfq_clear_cfqq_idle_window(cfqq);
2716 if (cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2722 * If group idle is enabled and there are requests dispatched from
2723 * this group, wait for requests to complete.
2726 if (cfqd->cfq_group_idle && cfqq->cfqg->nr_cfqq == 1 &&
2727 cfqq->cfqg->dispatched &&
2728 !cfq_io_thinktime_big(cfqd, &cfqq->cfqg->ttime, true)) {
2734 cfq_slice_expired(cfqd, 0);
2737 * Current queue expired. Check if we have to switch to a new
2741 cfq_choose_cfqg(cfqd);
2743 cfqq = cfq_set_active_queue(cfqd, new_cfqq);
2748 static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
2752 while (cfqq->next_rq) {
2753 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
2757 BUG_ON(!list_empty(&cfqq->fifo));
2759 /* By default cfqq is not expired if it is empty. Do it explicitly */
2760 __cfq_slice_expired(cfqq->cfqd, cfqq, 0);
2765 * Drain our current requests. Used for barriers and when switching
2766 * io schedulers on-the-fly.
2768 static int cfq_forced_dispatch(struct cfq_data *cfqd)
2770 struct cfq_queue *cfqq;
2773 /* Expire the timeslice of the current active queue first */
2774 cfq_slice_expired(cfqd, 0);
2775 while ((cfqq = cfq_get_next_queue_forced(cfqd)) != NULL) {
2776 __cfq_set_active_queue(cfqd, cfqq);
2777 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
2780 BUG_ON(cfqd->busy_queues);
2782 cfq_log(cfqd, "forced_dispatch=%d", dispatched);
2786 static inline bool cfq_slice_used_soon(struct cfq_data *cfqd,
2787 struct cfq_queue *cfqq)
2789 /* the queue hasn't finished any request, can't estimate */
2790 if (cfq_cfqq_slice_new(cfqq))
2792 if (time_after(jiffies + cfqd->cfq_slice_idle * cfqq->dispatched,
2799 static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2801 unsigned int max_dispatch;
2804 * Drain async requests before we start sync IO
2806 if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_flight[BLK_RW_ASYNC])
2810 * If this is an async queue and we have sync IO in flight, let it wait
2812 if (cfqd->rq_in_flight[BLK_RW_SYNC] && !cfq_cfqq_sync(cfqq))
2815 max_dispatch = max_t(unsigned int, cfqd->cfq_quantum / 2, 1);
2816 if (cfq_class_idle(cfqq))
2820 * Does this cfqq already have too much IO in flight?
2822 if (cfqq->dispatched >= max_dispatch) {
2823 bool promote_sync = false;
2825 * idle queue must always only have a single IO in flight
2827 if (cfq_class_idle(cfqq))
2831 * If there is only one sync queue
2832 * we can ignore async queue here and give the sync
2833 * queue no dispatch limit. The reason is a sync queue can
2834 * preempt async queue, limiting the sync queue doesn't make
2835 * sense. This is useful for aiostress test.
2837 if (cfq_cfqq_sync(cfqq) && cfqd->busy_sync_queues == 1)
2838 promote_sync = true;
2841 * We have other queues, don't allow more IO from this one
2843 if (cfqd->busy_queues > 1 && cfq_slice_used_soon(cfqd, cfqq) &&
2848 * Sole queue user, no limit
2850 if (cfqd->busy_queues == 1 || promote_sync)
2854 * Normally we start throttling cfqq when cfq_quantum/2
2855 * requests have been dispatched. But we can drive
2856 * deeper queue depths at the beginning of slice
2857 * subjected to upper limit of cfq_quantum.
2859 max_dispatch = cfqd->cfq_quantum;
2863 * Async queues must wait a bit before being allowed dispatch.
2864 * We also ramp up the dispatch depth gradually for async IO,
2865 * based on the last sync IO we serviced
2867 if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
2868 unsigned long last_sync = jiffies - cfqd->last_delayed_sync;
2871 depth = last_sync / cfqd->cfq_slice[1];
2872 if (!depth && !cfqq->dispatched)
2874 if (depth < max_dispatch)
2875 max_dispatch = depth;
2879 * If we're below the current max, allow a dispatch
2881 return cfqq->dispatched < max_dispatch;
2885 * Dispatch a request from cfqq, moving them to the request queue
2888 static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2892 BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
2894 if (!cfq_may_dispatch(cfqd, cfqq))
2898 * follow expired path, else get first next available
2900 rq = cfq_check_fifo(cfqq);
2905 * insert request into driver dispatch list
2907 cfq_dispatch_insert(cfqd->queue, rq);
2909 if (!cfqd->active_cic) {
2910 struct cfq_io_cq *cic = RQ_CIC(rq);
2912 atomic_long_inc(&cic->icq.ioc->refcount);
2913 cfqd->active_cic = cic;
2920 * Find the cfqq that we need to service and move a request from that to the
2923 static int cfq_dispatch_requests(struct request_queue *q, int force)
2925 struct cfq_data *cfqd = q->elevator->elevator_data;
2926 struct cfq_queue *cfqq;
2928 if (!cfqd->busy_queues)
2931 if (unlikely(force))
2932 return cfq_forced_dispatch(cfqd);
2934 cfqq = cfq_select_queue(cfqd);
2939 * Dispatch a request from this cfqq, if it is allowed
2941 if (!cfq_dispatch_request(cfqd, cfqq))
2944 cfqq->slice_dispatch++;
2945 cfq_clear_cfqq_must_dispatch(cfqq);
2948 * expire an async queue immediately if it has used up its slice. idle
2949 * queue always expire after 1 dispatch round.
2951 if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
2952 cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
2953 cfq_class_idle(cfqq))) {
2954 cfqq->slice_end = jiffies + 1;
2955 cfq_slice_expired(cfqd, 0);
2958 cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
2963 * task holds one reference to the queue, dropped when task exits. each rq
2964 * in-flight on this queue also holds a reference, dropped when rq is freed.
2966 * Each cfq queue took a reference on the parent group. Drop it now.
2967 * queue lock must be held here.
2969 static void cfq_put_queue(struct cfq_queue *cfqq)
2971 struct cfq_data *cfqd = cfqq->cfqd;
2972 struct cfq_group *cfqg;
2974 BUG_ON(cfqq->ref <= 0);
2980 cfq_log_cfqq(cfqd, cfqq, "put_queue");
2981 BUG_ON(rb_first(&cfqq->sort_list));
2982 BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
2985 if (unlikely(cfqd->active_queue == cfqq)) {
2986 __cfq_slice_expired(cfqd, cfqq, 0);
2987 cfq_schedule_dispatch(cfqd);
2990 BUG_ON(cfq_cfqq_on_rr(cfqq));
2991 kmem_cache_free(cfq_pool, cfqq);
2995 static void cfq_put_cooperator(struct cfq_queue *cfqq)
2997 struct cfq_queue *__cfqq, *next;
3000 * If this queue was scheduled to merge with another queue, be
3001 * sure to drop the reference taken on that queue (and others in
3002 * the merge chain). See cfq_setup_merge and cfq_merge_cfqqs.
3004 __cfqq = cfqq->new_cfqq;
3006 if (__cfqq == cfqq) {
3007 WARN(1, "cfqq->new_cfqq loop detected\n");
3010 next = __cfqq->new_cfqq;
3011 cfq_put_queue(__cfqq);
3016 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3018 if (unlikely(cfqq == cfqd->active_queue)) {
3019 __cfq_slice_expired(cfqd, cfqq, 0);
3020 cfq_schedule_dispatch(cfqd);
3023 cfq_put_cooperator(cfqq);
3025 cfq_put_queue(cfqq);
3028 static void cfq_init_icq(struct io_cq *icq)
3030 struct cfq_io_cq *cic = icq_to_cic(icq);
3032 cic->ttime.last_end_request = jiffies;
3035 static void cfq_exit_icq(struct io_cq *icq)
3037 struct cfq_io_cq *cic = icq_to_cic(icq);
3038 struct cfq_data *cfqd = cic_to_cfqd(cic);
3040 if (cic->cfqq[BLK_RW_ASYNC]) {
3041 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]);
3042 cic->cfqq[BLK_RW_ASYNC] = NULL;
3045 if (cic->cfqq[BLK_RW_SYNC]) {
3046 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]);
3047 cic->cfqq[BLK_RW_SYNC] = NULL;
3051 static void cfq_init_prio_data(struct cfq_queue *cfqq, struct cfq_io_cq *cic)
3053 struct task_struct *tsk = current;
3056 if (!cfq_cfqq_prio_changed(cfqq))
3059 ioprio_class = IOPRIO_PRIO_CLASS(cic->ioprio);
3060 switch (ioprio_class) {
3062 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
3063 case IOPRIO_CLASS_NONE:
3065 * no prio set, inherit CPU scheduling settings
3067 cfqq->ioprio = task_nice_ioprio(tsk);
3068 cfqq->ioprio_class = task_nice_ioclass(tsk);
3070 case IOPRIO_CLASS_RT:
3071 cfqq->ioprio = IOPRIO_PRIO_DATA(cic->ioprio);
3072 cfqq->ioprio_class = IOPRIO_CLASS_RT;
3074 case IOPRIO_CLASS_BE:
3075 cfqq->ioprio = IOPRIO_PRIO_DATA(cic->ioprio);
3076 cfqq->ioprio_class = IOPRIO_CLASS_BE;
3078 case IOPRIO_CLASS_IDLE:
3079 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
3081 cfq_clear_cfqq_idle_window(cfqq);
3086 * keep track of original prio settings in case we have to temporarily
3087 * elevate the priority of this queue
3089 cfqq->org_ioprio = cfqq->ioprio;
3090 cfq_clear_cfqq_prio_changed(cfqq);
3093 static void check_ioprio_changed(struct cfq_io_cq *cic, struct bio *bio)
3095 int ioprio = cic->icq.ioc->ioprio;
3096 struct cfq_data *cfqd = cic_to_cfqd(cic);
3097 struct cfq_queue *cfqq;
3100 * Check whether ioprio has changed. The condition may trigger
3101 * spuriously on a newly created cic but there's no harm.
3103 if (unlikely(!cfqd) || likely(cic->ioprio == ioprio))
3106 cfqq = cic->cfqq[BLK_RW_ASYNC];
3108 struct cfq_queue *new_cfqq;
3109 new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic, bio,
3112 cic->cfqq[BLK_RW_ASYNC] = new_cfqq;
3113 cfq_put_queue(cfqq);
3117 cfqq = cic->cfqq[BLK_RW_SYNC];
3119 cfq_mark_cfqq_prio_changed(cfqq);
3121 cic->ioprio = ioprio;
3124 static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3125 pid_t pid, bool is_sync)
3127 RB_CLEAR_NODE(&cfqq->rb_node);
3128 RB_CLEAR_NODE(&cfqq->p_node);
3129 INIT_LIST_HEAD(&cfqq->fifo);
3134 cfq_mark_cfqq_prio_changed(cfqq);
3137 if (!cfq_class_idle(cfqq))
3138 cfq_mark_cfqq_idle_window(cfqq);
3139 cfq_mark_cfqq_sync(cfqq);
3144 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3145 static void check_blkcg_changed(struct cfq_io_cq *cic, struct bio *bio)
3147 struct cfq_data *cfqd = cic_to_cfqd(cic);
3148 struct cfq_queue *sync_cfqq;
3152 id = bio_blkcg(bio)->id;
3156 * Check whether blkcg has changed. The condition may trigger
3157 * spuriously on a newly created cic but there's no harm.
3159 if (unlikely(!cfqd) || likely(cic->blkcg_id == id))
3162 sync_cfqq = cic_to_cfqq(cic, 1);
3165 * Drop reference to sync queue. A new sync queue will be
3166 * assigned in new group upon arrival of a fresh request.
3168 cfq_log_cfqq(cfqd, sync_cfqq, "changed cgroup");
3169 cic_set_cfqq(cic, NULL, 1);
3170 cfq_put_queue(sync_cfqq);
3176 static inline void check_blkcg_changed(struct cfq_io_cq *cic, struct bio *bio) { }
3177 #endif /* CONFIG_CFQ_GROUP_IOSCHED */
3179 static struct cfq_queue *
3180 cfq_find_alloc_queue(struct cfq_data *cfqd, bool is_sync, struct cfq_io_cq *cic,
3181 struct bio *bio, gfp_t gfp_mask)
3183 struct blkcg *blkcg;
3184 struct cfq_queue *cfqq, *new_cfqq = NULL;
3185 struct cfq_group *cfqg;
3190 blkcg = bio_blkcg(bio);
3191 cfqg = cfq_lookup_create_cfqg(cfqd, blkcg);
3192 cfqq = cic_to_cfqq(cic, is_sync);
3195 * Always try a new alloc if we fell back to the OOM cfqq
3196 * originally, since it should just be a temporary situation.
3198 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
3203 } else if (gfp_mask & __GFP_WAIT) {
3205 spin_unlock_irq(cfqd->queue->queue_lock);
3206 new_cfqq = kmem_cache_alloc_node(cfq_pool,
3207 gfp_mask | __GFP_ZERO,
3209 spin_lock_irq(cfqd->queue->queue_lock);
3213 cfqq = kmem_cache_alloc_node(cfq_pool,
3214 gfp_mask | __GFP_ZERO,
3219 cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
3220 cfq_init_prio_data(cfqq, cic);
3221 cfq_link_cfqq_cfqg(cfqq, cfqg);
3222 cfq_log_cfqq(cfqd, cfqq, "alloced");
3224 cfqq = &cfqd->oom_cfqq;
3228 kmem_cache_free(cfq_pool, new_cfqq);
3234 static struct cfq_queue **
3235 cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
3237 switch (ioprio_class) {
3238 case IOPRIO_CLASS_RT:
3239 return &cfqd->async_cfqq[0][ioprio];
3240 case IOPRIO_CLASS_NONE:
3241 ioprio = IOPRIO_NORM;
3243 case IOPRIO_CLASS_BE:
3244 return &cfqd->async_cfqq[1][ioprio];
3245 case IOPRIO_CLASS_IDLE:
3246 return &cfqd->async_idle_cfqq;
3252 static struct cfq_queue *
3253 cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct cfq_io_cq *cic,
3254 struct bio *bio, gfp_t gfp_mask)
3256 const int ioprio_class = IOPRIO_PRIO_CLASS(cic->ioprio);
3257 const int ioprio = IOPRIO_PRIO_DATA(cic->ioprio);
3258 struct cfq_queue **async_cfqq = NULL;
3259 struct cfq_queue *cfqq = NULL;
3262 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
3267 cfqq = cfq_find_alloc_queue(cfqd, is_sync, cic, bio, gfp_mask);
3270 * pin the queue now that it's allocated, scheduler exit will prune it
3272 if (!is_sync && !(*async_cfqq)) {
3282 __cfq_update_io_thinktime(struct cfq_ttime *ttime, unsigned long slice_idle)
3284 unsigned long elapsed = jiffies - ttime->last_end_request;
3285 elapsed = min(elapsed, 2UL * slice_idle);
3287 ttime->ttime_samples = (7*ttime->ttime_samples + 256) / 8;
3288 ttime->ttime_total = (7*ttime->ttime_total + 256*elapsed) / 8;
3289 ttime->ttime_mean = (ttime->ttime_total + 128) / ttime->ttime_samples;
3293 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3294 struct cfq_io_cq *cic)
3296 if (cfq_cfqq_sync(cfqq)) {
3297 __cfq_update_io_thinktime(&cic->ttime, cfqd->cfq_slice_idle);
3298 __cfq_update_io_thinktime(&cfqq->service_tree->ttime,
3299 cfqd->cfq_slice_idle);
3301 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3302 __cfq_update_io_thinktime(&cfqq->cfqg->ttime, cfqd->cfq_group_idle);
3307 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3311 sector_t n_sec = blk_rq_sectors(rq);
3312 if (cfqq->last_request_pos) {
3313 if (cfqq->last_request_pos < blk_rq_pos(rq))
3314 sdist = blk_rq_pos(rq) - cfqq->last_request_pos;
3316 sdist = cfqq->last_request_pos - blk_rq_pos(rq);
3319 cfqq->seek_history <<= 1;
3320 if (blk_queue_nonrot(cfqd->queue))
3321 cfqq->seek_history |= (n_sec < CFQQ_SECT_THR_NONROT);
3323 cfqq->seek_history |= (sdist > CFQQ_SEEK_THR);
3327 * Disable idle window if the process thinks too long or seeks so much that
3331 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3332 struct cfq_io_cq *cic)
3334 int old_idle, enable_idle;
3337 * Don't idle for async or idle io prio class
3339 if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
3342 enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
3344 if (cfqq->queued[0] + cfqq->queued[1] >= 4)
3345 cfq_mark_cfqq_deep(cfqq);
3347 if (cfqq->next_rq && (cfqq->next_rq->cmd_flags & REQ_NOIDLE))
3349 else if (!atomic_read(&cic->icq.ioc->active_ref) ||
3350 !cfqd->cfq_slice_idle ||
3351 (!cfq_cfqq_deep(cfqq) && CFQQ_SEEKY(cfqq)))
3353 else if (sample_valid(cic->ttime.ttime_samples)) {
3354 if (cic->ttime.ttime_mean > cfqd->cfq_slice_idle)
3360 if (old_idle != enable_idle) {
3361 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
3363 cfq_mark_cfqq_idle_window(cfqq);
3365 cfq_clear_cfqq_idle_window(cfqq);
3370 * Check if new_cfqq should preempt the currently active queue. Return 0 for
3371 * no or if we aren't sure, a 1 will cause a preempt.
3374 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
3377 struct cfq_queue *cfqq;
3379 cfqq = cfqd->active_queue;
3383 if (cfq_class_idle(new_cfqq))
3386 if (cfq_class_idle(cfqq))
3390 * Don't allow a non-RT request to preempt an ongoing RT cfqq timeslice.
3392 if (cfq_class_rt(cfqq) && !cfq_class_rt(new_cfqq))
3396 * if the new request is sync, but the currently running queue is
3397 * not, let the sync request have priority.
3399 if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
3402 if (new_cfqq->cfqg != cfqq->cfqg)
3405 if (cfq_slice_used(cfqq))
3408 /* Allow preemption only if we are idling on sync-noidle tree */
3409 if (cfqd->serving_wl_type == SYNC_NOIDLE_WORKLOAD &&
3410 cfqq_type(new_cfqq) == SYNC_NOIDLE_WORKLOAD &&
3411 new_cfqq->service_tree->count == 2 &&
3412 RB_EMPTY_ROOT(&cfqq->sort_list))
3416 * So both queues are sync. Let the new request get disk time if
3417 * it's a metadata request and the current queue is doing regular IO.
3419 if ((rq->cmd_flags & REQ_PRIO) && !cfqq->prio_pending)
3423 * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
3425 if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
3428 /* An idle queue should not be idle now for some reason */
3429 if (RB_EMPTY_ROOT(&cfqq->sort_list) && !cfq_should_idle(cfqd, cfqq))
3432 if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
3436 * if this request is as-good as one we would expect from the
3437 * current cfqq, let it preempt
3439 if (cfq_rq_close(cfqd, cfqq, rq))
3446 * cfqq preempts the active queue. if we allowed preempt with no slice left,
3447 * let it have half of its nominal slice.
3449 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3451 enum wl_type_t old_type = cfqq_type(cfqd->active_queue);
3453 cfq_log_cfqq(cfqd, cfqq, "preempt");
3454 cfq_slice_expired(cfqd, 1);
3457 * workload type is changed, don't save slice, otherwise preempt
3460 if (old_type != cfqq_type(cfqq))
3461 cfqq->cfqg->saved_wl_slice = 0;
3464 * Put the new queue at the front of the of the current list,
3465 * so we know that it will be selected next.
3467 BUG_ON(!cfq_cfqq_on_rr(cfqq));
3469 cfq_service_tree_add(cfqd, cfqq, 1);
3471 cfqq->slice_end = 0;
3472 cfq_mark_cfqq_slice_new(cfqq);
3476 * Called when a new fs request (rq) is added (to cfqq). Check if there's
3477 * something we should do about it
3480 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3483 struct cfq_io_cq *cic = RQ_CIC(rq);
3486 if (rq->cmd_flags & REQ_PRIO)
3487 cfqq->prio_pending++;
3489 cfq_update_io_thinktime(cfqd, cfqq, cic);
3490 cfq_update_io_seektime(cfqd, cfqq, rq);
3491 cfq_update_idle_window(cfqd, cfqq, cic);
3493 cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
3495 if (cfqq == cfqd->active_queue) {
3497 * Remember that we saw a request from this process, but
3498 * don't start queuing just yet. Otherwise we risk seeing lots
3499 * of tiny requests, because we disrupt the normal plugging
3500 * and merging. If the request is already larger than a single
3501 * page, let it rip immediately. For that case we assume that
3502 * merging is already done. Ditto for a busy system that
3503 * has other work pending, don't risk delaying until the
3504 * idle timer unplug to continue working.
3506 if (cfq_cfqq_wait_request(cfqq)) {
3507 if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
3508 cfqd->busy_queues > 1) {
3509 cfq_del_timer(cfqd, cfqq);
3510 cfq_clear_cfqq_wait_request(cfqq);
3511 __blk_run_queue(cfqd->queue);
3513 cfqg_stats_update_idle_time(cfqq->cfqg);
3514 cfq_mark_cfqq_must_dispatch(cfqq);
3517 } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
3519 * not the active queue - expire current slice if it is
3520 * idle and has expired it's mean thinktime or this new queue
3521 * has some old slice time left and is of higher priority or
3522 * this new queue is RT and the current one is BE
3524 cfq_preempt_queue(cfqd, cfqq);
3525 __blk_run_queue(cfqd->queue);
3529 static void cfq_insert_request(struct request_queue *q, struct request *rq)
3531 struct cfq_data *cfqd = q->elevator->elevator_data;
3532 struct cfq_queue *cfqq = RQ_CFQQ(rq);
3534 cfq_log_cfqq(cfqd, cfqq, "insert_request");
3535 cfq_init_prio_data(cfqq, RQ_CIC(rq));
3537 rq_set_fifo_time(rq, jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)]);
3538 list_add_tail(&rq->queuelist, &cfqq->fifo);
3540 cfqg_stats_update_io_add(RQ_CFQG(rq), cfqd->serving_group,
3542 cfq_rq_enqueued(cfqd, cfqq, rq);
3546 * Update hw_tag based on peak queue depth over 50 samples under
3549 static void cfq_update_hw_tag(struct cfq_data *cfqd)
3551 struct cfq_queue *cfqq = cfqd->active_queue;
3553 if (cfqd->rq_in_driver > cfqd->hw_tag_est_depth)
3554 cfqd->hw_tag_est_depth = cfqd->rq_in_driver;
3556 if (cfqd->hw_tag == 1)
3559 if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
3560 cfqd->rq_in_driver <= CFQ_HW_QUEUE_MIN)
3564 * If active queue hasn't enough requests and can idle, cfq might not
3565 * dispatch sufficient requests to hardware. Don't zero hw_tag in this
3568 if (cfqq && cfq_cfqq_idle_window(cfqq) &&
3569 cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] <
3570 CFQ_HW_QUEUE_MIN && cfqd->rq_in_driver < CFQ_HW_QUEUE_MIN)
3573 if (cfqd->hw_tag_samples++ < 50)
3576 if (cfqd->hw_tag_est_depth >= CFQ_HW_QUEUE_MIN)
3582 static bool cfq_should_wait_busy(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3584 struct cfq_io_cq *cic = cfqd->active_cic;
3586 /* If the queue already has requests, don't wait */
3587 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3590 /* If there are other queues in the group, don't wait */
3591 if (cfqq->cfqg->nr_cfqq > 1)
3594 /* the only queue in the group, but think time is big */
3595 if (cfq_io_thinktime_big(cfqd, &cfqq->cfqg->ttime, true))
3598 if (cfq_slice_used(cfqq))
3601 /* if slice left is less than think time, wait busy */
3602 if (cic && sample_valid(cic->ttime.ttime_samples)
3603 && (cfqq->slice_end - jiffies < cic->ttime.ttime_mean))
3607 * If think times is less than a jiffy than ttime_mean=0 and above
3608 * will not be true. It might happen that slice has not expired yet
3609 * but will expire soon (4-5 ns) during select_queue(). To cover the
3610 * case where think time is less than a jiffy, mark the queue wait
3611 * busy if only 1 jiffy is left in the slice.
3613 if (cfqq->slice_end - jiffies == 1)
3619 static void cfq_completed_request(struct request_queue *q, struct request *rq)
3621 struct cfq_queue *cfqq = RQ_CFQQ(rq);
3622 struct cfq_data *cfqd = cfqq->cfqd;
3623 const int sync = rq_is_sync(rq);
3627 cfq_log_cfqq(cfqd, cfqq, "complete rqnoidle %d",
3628 !!(rq->cmd_flags & REQ_NOIDLE));
3630 cfq_update_hw_tag(cfqd);
3632 WARN_ON(!cfqd->rq_in_driver);
3633 WARN_ON(!cfqq->dispatched);
3634 cfqd->rq_in_driver--;
3636 (RQ_CFQG(rq))->dispatched--;
3637 cfqg_stats_update_completion(cfqq->cfqg, rq_start_time_ns(rq),
3638 rq_io_start_time_ns(rq), rq->cmd_flags);
3640 cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]--;
3643 struct cfq_rb_root *service_tree;
3645 RQ_CIC(rq)->ttime.last_end_request = now;
3647 if (cfq_cfqq_on_rr(cfqq))
3648 service_tree = cfqq->service_tree;
3650 service_tree = service_tree_for(cfqq->cfqg,
3651 cfqq_class(cfqq), cfqq_type(cfqq));
3652 service_tree->ttime.last_end_request = now;
3653 if (!time_after(rq->start_time + cfqd->cfq_fifo_expire[1], now))
3654 cfqd->last_delayed_sync = now;
3657 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3658 cfqq->cfqg->ttime.last_end_request = now;
3662 * If this is the active queue, check if it needs to be expired,
3663 * or if we want to idle in case it has no pending requests.
3665 if (cfqd->active_queue == cfqq) {
3666 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
3668 if (cfq_cfqq_slice_new(cfqq)) {
3669 cfq_set_prio_slice(cfqd, cfqq);
3670 cfq_clear_cfqq_slice_new(cfqq);
3674 * Should we wait for next request to come in before we expire
3677 if (cfq_should_wait_busy(cfqd, cfqq)) {
3678 unsigned long extend_sl = cfqd->cfq_slice_idle;
3679 if (!cfqd->cfq_slice_idle)
3680 extend_sl = cfqd->cfq_group_idle;
3681 cfqq->slice_end = jiffies + extend_sl;
3682 cfq_mark_cfqq_wait_busy(cfqq);
3683 cfq_log_cfqq(cfqd, cfqq, "will busy wait");
3687 * Idling is not enabled on:
3689 * - idle-priority queues
3691 * - queues with still some requests queued
3692 * - when there is a close cooperator
3694 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
3695 cfq_slice_expired(cfqd, 1);
3696 else if (sync && cfqq_empty &&
3697 !cfq_close_cooperator(cfqd, cfqq)) {
3698 cfq_arm_slice_timer(cfqd);
3702 if (!cfqd->rq_in_driver)
3703 cfq_schedule_dispatch(cfqd);
3706 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
3708 if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
3709 cfq_mark_cfqq_must_alloc_slice(cfqq);
3710 return ELV_MQUEUE_MUST;
3713 return ELV_MQUEUE_MAY;
3716 static int cfq_may_queue(struct request_queue *q, int rw)
3718 struct cfq_data *cfqd = q->elevator->elevator_data;
3719 struct task_struct *tsk = current;
3720 struct cfq_io_cq *cic;
3721 struct cfq_queue *cfqq;
3724 * don't force setup of a queue from here, as a call to may_queue
3725 * does not necessarily imply that a request actually will be queued.
3726 * so just lookup a possibly existing queue, or return 'may queue'
3729 cic = cfq_cic_lookup(cfqd, tsk->io_context);
3731 return ELV_MQUEUE_MAY;
3733 cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
3735 cfq_init_prio_data(cfqq, cic);
3737 return __cfq_may_queue(cfqq);
3740 return ELV_MQUEUE_MAY;
3744 * queue lock held here
3746 static void cfq_put_request(struct request *rq)
3748 struct cfq_queue *cfqq = RQ_CFQQ(rq);
3751 const int rw = rq_data_dir(rq);
3753 BUG_ON(!cfqq->allocated[rw]);
3754 cfqq->allocated[rw]--;
3756 /* Put down rq reference on cfqg */
3757 cfqg_put(RQ_CFQG(rq));
3758 rq->elv.priv[0] = NULL;
3759 rq->elv.priv[1] = NULL;
3761 cfq_put_queue(cfqq);
3765 static struct cfq_queue *
3766 cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_cq *cic,
3767 struct cfq_queue *cfqq)
3769 cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq);
3770 cic_set_cfqq(cic, cfqq->new_cfqq, 1);
3771 cfq_mark_cfqq_coop(cfqq->new_cfqq);
3772 cfq_put_queue(cfqq);
3773 return cic_to_cfqq(cic, 1);
3777 * Returns NULL if a new cfqq should be allocated, or the old cfqq if this
3778 * was the last process referring to said cfqq.
3780 static struct cfq_queue *
3781 split_cfqq(struct cfq_io_cq *cic, struct cfq_queue *cfqq)
3783 if (cfqq_process_refs(cfqq) == 1) {
3784 cfqq->pid = current->pid;
3785 cfq_clear_cfqq_coop(cfqq);
3786 cfq_clear_cfqq_split_coop(cfqq);
3790 cic_set_cfqq(cic, NULL, 1);
3792 cfq_put_cooperator(cfqq);
3794 cfq_put_queue(cfqq);
3798 * Allocate cfq data structures associated with this request.
3801 cfq_set_request(struct request_queue *q, struct request *rq, struct bio *bio,
3804 struct cfq_data *cfqd = q->elevator->elevator_data;
3805 struct cfq_io_cq *cic = icq_to_cic(rq->elv.icq);
3806 const int rw = rq_data_dir(rq);
3807 const bool is_sync = rq_is_sync(rq);
3808 struct cfq_queue *cfqq;
3810 might_sleep_if(gfp_mask & __GFP_WAIT);
3812 spin_lock_irq(q->queue_lock);
3814 check_ioprio_changed(cic, bio);
3815 check_blkcg_changed(cic, bio);
3817 cfqq = cic_to_cfqq(cic, is_sync);
3818 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
3819 cfqq = cfq_get_queue(cfqd, is_sync, cic, bio, gfp_mask);
3820 cic_set_cfqq(cic, cfqq, is_sync);
3823 * If the queue was seeky for too long, break it apart.
3825 if (cfq_cfqq_coop(cfqq) && cfq_cfqq_split_coop(cfqq)) {
3826 cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq");
3827 cfqq = split_cfqq(cic, cfqq);
3833 * Check to see if this queue is scheduled to merge with
3834 * another, closely cooperating queue. The merging of
3835 * queues happens here as it must be done in process context.
3836 * The reference on new_cfqq was taken in merge_cfqqs.
3839 cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq);
3842 cfqq->allocated[rw]++;
3845 cfqg_get(cfqq->cfqg);
3846 rq->elv.priv[0] = cfqq;
3847 rq->elv.priv[1] = cfqq->cfqg;
3848 spin_unlock_irq(q->queue_lock);
3852 static void cfq_kick_queue(struct work_struct *work)
3854 struct cfq_data *cfqd =
3855 container_of(work, struct cfq_data, unplug_work);
3856 struct request_queue *q = cfqd->queue;
3858 spin_lock_irq(q->queue_lock);
3859 __blk_run_queue(cfqd->queue);
3860 spin_unlock_irq(q->queue_lock);
3864 * Timer running if the active_queue is currently idling inside its time slice
3866 static void cfq_idle_slice_timer(unsigned long data)
3868 struct cfq_data *cfqd = (struct cfq_data *) data;
3869 struct cfq_queue *cfqq;
3870 unsigned long flags;
3873 cfq_log(cfqd, "idle timer fired");
3875 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
3877 cfqq = cfqd->active_queue;
3882 * We saw a request before the queue expired, let it through
3884 if (cfq_cfqq_must_dispatch(cfqq))
3890 if (cfq_slice_used(cfqq))
3894 * only expire and reinvoke request handler, if there are
3895 * other queues with pending requests
3897 if (!cfqd->busy_queues)
3901 * not expired and it has a request pending, let it dispatch
3903 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3907 * Queue depth flag is reset only when the idle didn't succeed
3909 cfq_clear_cfqq_deep(cfqq);
3912 cfq_slice_expired(cfqd, timed_out);
3914 cfq_schedule_dispatch(cfqd);
3916 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
3919 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
3921 del_timer_sync(&cfqd->idle_slice_timer);
3922 cancel_work_sync(&cfqd->unplug_work);
3925 static void cfq_put_async_queues(struct cfq_data *cfqd)
3929 for (i = 0; i < IOPRIO_BE_NR; i++) {
3930 if (cfqd->async_cfqq[0][i])
3931 cfq_put_queue(cfqd->async_cfqq[0][i]);
3932 if (cfqd->async_cfqq[1][i])
3933 cfq_put_queue(cfqd->async_cfqq[1][i]);
3936 if (cfqd->async_idle_cfqq)
3937 cfq_put_queue(cfqd->async_idle_cfqq);
3940 static void cfq_exit_queue(struct elevator_queue *e)
3942 struct cfq_data *cfqd = e->elevator_data;
3943 struct request_queue *q = cfqd->queue;
3945 cfq_shutdown_timer_wq(cfqd);
3947 spin_lock_irq(q->queue_lock);
3949 if (cfqd->active_queue)
3950 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
3952 cfq_put_async_queues(cfqd);
3954 spin_unlock_irq(q->queue_lock);
3956 cfq_shutdown_timer_wq(cfqd);
3958 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3959 blkcg_deactivate_policy(q, &blkcg_policy_cfq);
3961 kfree(cfqd->root_group);
3966 static int cfq_init_queue(struct request_queue *q)
3968 struct cfq_data *cfqd;
3969 struct blkcg_gq *blkg __maybe_unused;
3972 cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
3977 q->elevator->elevator_data = cfqd;
3979 /* Init root service tree */
3980 cfqd->grp_service_tree = CFQ_RB_ROOT;
3982 /* Init root group and prefer root group over other groups by default */
3983 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3984 ret = blkcg_activate_policy(q, &blkcg_policy_cfq);
3988 cfqd->root_group = blkg_to_cfqg(q->root_blkg);
3991 cfqd->root_group = kzalloc_node(sizeof(*cfqd->root_group),
3992 GFP_KERNEL, cfqd->queue->node);
3993 if (!cfqd->root_group)
3996 cfq_init_cfqg_base(cfqd->root_group);
3998 cfqd->root_group->weight = 2 * CFQ_WEIGHT_DEFAULT;
4001 * Not strictly needed (since RB_ROOT just clears the node and we
4002 * zeroed cfqd on alloc), but better be safe in case someone decides
4003 * to add magic to the rb code
4005 for (i = 0; i < CFQ_PRIO_LISTS; i++)
4006 cfqd->prio_trees[i] = RB_ROOT;
4009 * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
4010 * Grab a permanent reference to it, so that the normal code flow
4011 * will not attempt to free it. oom_cfqq is linked to root_group
4012 * but shouldn't hold a reference as it'll never be unlinked. Lose
4013 * the reference from linking right away.
4015 cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
4016 cfqd->oom_cfqq.ref++;
4018 spin_lock_irq(q->queue_lock);
4019 cfq_link_cfqq_cfqg(&cfqd->oom_cfqq, cfqd->root_group);
4020 cfqg_put(cfqd->root_group);
4021 spin_unlock_irq(q->queue_lock);
4023 init_timer(&cfqd->idle_slice_timer);
4024 cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
4025 cfqd->idle_slice_timer.data = (unsigned long) cfqd;
4027 INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
4029 cfqd->cfq_quantum = cfq_quantum;
4030 cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
4031 cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
4032 cfqd->cfq_back_max = cfq_back_max;
4033 cfqd->cfq_back_penalty = cfq_back_penalty;
4034 cfqd->cfq_slice[0] = cfq_slice_async;
4035 cfqd->cfq_slice[1] = cfq_slice_sync;
4036 cfqd->cfq_target_latency = cfq_target_latency;
4037 cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
4038 cfqd->cfq_slice_idle = cfq_slice_idle;
4039 cfqd->cfq_group_idle = cfq_group_idle;
4040 cfqd->cfq_latency = 1;
4043 * we optimistically start assuming sync ops weren't delayed in last
4044 * second, in order to have larger depth for async operations.
4046 cfqd->last_delayed_sync = jiffies - HZ;
4055 * sysfs parts below -->
4058 cfq_var_show(unsigned int var, char *page)
4060 return sprintf(page, "%d\n", var);
4064 cfq_var_store(unsigned int *var, const char *page, size_t count)
4066 char *p = (char *) page;
4068 *var = simple_strtoul(p, &p, 10);
4072 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
4073 static ssize_t __FUNC(struct elevator_queue *e, char *page) \
4075 struct cfq_data *cfqd = e->elevator_data; \
4076 unsigned int __data = __VAR; \
4078 __data = jiffies_to_msecs(__data); \
4079 return cfq_var_show(__data, (page)); \
4081 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
4082 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
4083 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
4084 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
4085 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
4086 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
4087 SHOW_FUNCTION(cfq_group_idle_show, cfqd->cfq_group_idle, 1);
4088 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
4089 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
4090 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
4091 SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
4092 SHOW_FUNCTION(cfq_target_latency_show, cfqd->cfq_target_latency, 1);
4093 #undef SHOW_FUNCTION
4095 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
4096 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
4098 struct cfq_data *cfqd = e->elevator_data; \
4099 unsigned int __data; \
4100 int ret = cfq_var_store(&__data, (page), count); \
4101 if (__data < (MIN)) \
4103 else if (__data > (MAX)) \
4106 *(__PTR) = msecs_to_jiffies(__data); \
4108 *(__PTR) = __data; \
4111 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
4112 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
4114 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
4116 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
4117 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
4119 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
4120 STORE_FUNCTION(cfq_group_idle_store, &cfqd->cfq_group_idle, 0, UINT_MAX, 1);
4121 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
4122 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
4123 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
4125 STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
4126 STORE_FUNCTION(cfq_target_latency_store, &cfqd->cfq_target_latency, 1, UINT_MAX, 1);
4127 #undef STORE_FUNCTION
4129 #define CFQ_ATTR(name) \
4130 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
4132 static struct elv_fs_entry cfq_attrs[] = {
4134 CFQ_ATTR(fifo_expire_sync),
4135 CFQ_ATTR(fifo_expire_async),
4136 CFQ_ATTR(back_seek_max),
4137 CFQ_ATTR(back_seek_penalty),
4138 CFQ_ATTR(slice_sync),
4139 CFQ_ATTR(slice_async),
4140 CFQ_ATTR(slice_async_rq),
4141 CFQ_ATTR(slice_idle),
4142 CFQ_ATTR(group_idle),
4143 CFQ_ATTR(low_latency),
4144 CFQ_ATTR(target_latency),
4148 static struct elevator_type iosched_cfq = {
4150 .elevator_merge_fn = cfq_merge,
4151 .elevator_merged_fn = cfq_merged_request,
4152 .elevator_merge_req_fn = cfq_merged_requests,
4153 .elevator_allow_merge_fn = cfq_allow_merge,
4154 .elevator_bio_merged_fn = cfq_bio_merged,
4155 .elevator_dispatch_fn = cfq_dispatch_requests,
4156 .elevator_add_req_fn = cfq_insert_request,
4157 .elevator_activate_req_fn = cfq_activate_request,
4158 .elevator_deactivate_req_fn = cfq_deactivate_request,
4159 .elevator_completed_req_fn = cfq_completed_request,
4160 .elevator_former_req_fn = elv_rb_former_request,
4161 .elevator_latter_req_fn = elv_rb_latter_request,
4162 .elevator_init_icq_fn = cfq_init_icq,
4163 .elevator_exit_icq_fn = cfq_exit_icq,
4164 .elevator_set_req_fn = cfq_set_request,
4165 .elevator_put_req_fn = cfq_put_request,
4166 .elevator_may_queue_fn = cfq_may_queue,
4167 .elevator_init_fn = cfq_init_queue,
4168 .elevator_exit_fn = cfq_exit_queue,
4170 .icq_size = sizeof(struct cfq_io_cq),
4171 .icq_align = __alignof__(struct cfq_io_cq),
4172 .elevator_attrs = cfq_attrs,
4173 .elevator_name = "cfq",
4174 .elevator_owner = THIS_MODULE,
4177 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4178 static struct blkcg_policy blkcg_policy_cfq = {
4179 .pd_size = sizeof(struct cfq_group),
4180 .cftypes = cfq_blkcg_files,
4182 .pd_init_fn = cfq_pd_init,
4183 .pd_reset_stats_fn = cfq_pd_reset_stats,
4187 static int __init cfq_init(void)
4192 * could be 0 on HZ < 1000 setups
4194 if (!cfq_slice_async)
4195 cfq_slice_async = 1;
4196 if (!cfq_slice_idle)
4199 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4200 if (!cfq_group_idle)
4203 ret = blkcg_policy_register(&blkcg_policy_cfq);
4211 cfq_pool = KMEM_CACHE(cfq_queue, 0);
4215 ret = elv_register(&iosched_cfq);
4222 kmem_cache_destroy(cfq_pool);
4224 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4225 blkcg_policy_unregister(&blkcg_policy_cfq);
4230 static void __exit cfq_exit(void)
4232 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4233 blkcg_policy_unregister(&blkcg_policy_cfq);
4235 elv_unregister(&iosched_cfq);
4236 kmem_cache_destroy(cfq_pool);
4239 module_init(cfq_init);
4240 module_exit(cfq_exit);
4242 MODULE_AUTHOR("Jens Axboe");
4243 MODULE_LICENSE("GPL");
4244 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");