2 * CFQ, or complete fairness queueing, disk scheduler.
4 * Based on ideas from a previously unfinished io
5 * scheduler (round robin per-process disk scheduling) and Andrea Arcangeli.
7 * Copyright (C) 2003 Jens Axboe <axboe@kernel.dk>
9 #include <linux/module.h>
10 #include <linux/slab.h>
11 #include <linux/blkdev.h>
12 #include <linux/elevator.h>
13 #include <linux/jiffies.h>
14 #include <linux/rbtree.h>
15 #include <linux/ioprio.h>
16 #include <linux/blktrace_api.h>
17 #include <linux/blk-cgroup.h>
23 /* max queue in one round of service */
24 static const int cfq_quantum
= 8;
25 static const int cfq_fifo_expire
[2] = { HZ
/ 4, HZ
/ 8 };
26 /* maximum backwards seek, in KiB */
27 static const int cfq_back_max
= 16 * 1024;
28 /* penalty of a backwards seek */
29 static const int cfq_back_penalty
= 2;
30 static const int cfq_slice_sync
= HZ
/ 10;
31 static int cfq_slice_async
= HZ
/ 25;
32 static const int cfq_slice_async_rq
= 2;
33 static int cfq_slice_idle
= HZ
/ 125;
34 static int cfq_group_idle
= HZ
/ 125;
35 static const int cfq_target_latency
= HZ
* 3/10; /* 300 ms */
36 static const int cfq_hist_divisor
= 4;
39 * offset from end of service tree
41 #define CFQ_IDLE_DELAY (HZ / 5)
44 * below this threshold, we consider thinktime immediate
46 #define CFQ_MIN_TT (2)
48 #define CFQ_SLICE_SCALE (5)
49 #define CFQ_HW_QUEUE_MIN (5)
50 #define CFQ_SERVICE_SHIFT 12
52 #define CFQQ_SEEK_THR (sector_t)(8 * 100)
53 #define CFQQ_CLOSE_THR (sector_t)(8 * 1024)
54 #define CFQQ_SECT_THR_NONROT (sector_t)(2 * 32)
55 #define CFQQ_SEEKY(cfqq) (hweight32(cfqq->seek_history) > 32/8)
57 #define RQ_CIC(rq) icq_to_cic((rq)->elv.icq)
58 #define RQ_CFQQ(rq) (struct cfq_queue *) ((rq)->elv.priv[0])
59 #define RQ_CFQG(rq) (struct cfq_group *) ((rq)->elv.priv[1])
61 static struct kmem_cache
*cfq_pool
;
63 #define CFQ_PRIO_LISTS IOPRIO_BE_NR
64 #define cfq_class_idle(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
65 #define cfq_class_rt(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
67 #define sample_valid(samples) ((samples) > 80)
68 #define rb_entry_cfqg(node) rb_entry((node), struct cfq_group, rb_node)
70 /* blkio-related constants */
71 #define CFQ_WEIGHT_MIN 10
72 #define CFQ_WEIGHT_MAX 1000
73 #define CFQ_WEIGHT_DEFAULT 500
76 unsigned long last_end_request
;
78 unsigned long ttime_total
;
79 unsigned long ttime_samples
;
80 unsigned long ttime_mean
;
84 * Most of our rbtree usage is for sorting with min extraction, so
85 * if we cache the leftmost node we don't have to walk down the tree
86 * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should
87 * move this into the elevator for the rq sorting as well.
94 struct cfq_ttime ttime
;
96 #define CFQ_RB_ROOT (struct cfq_rb_root) { .rb = RB_ROOT, \
97 .ttime = {.last_end_request = jiffies,},}
100 * Per process-grouping structure
103 /* reference count */
105 /* various state flags, see below */
107 /* parent cfq_data */
108 struct cfq_data
*cfqd
;
109 /* service_tree member */
110 struct rb_node rb_node
;
111 /* service_tree key */
112 unsigned long rb_key
;
113 /* prio tree member */
114 struct rb_node p_node
;
115 /* prio tree root we belong to, if any */
116 struct rb_root
*p_root
;
117 /* sorted list of pending requests */
118 struct rb_root sort_list
;
119 /* if fifo isn't expired, next request to serve */
120 struct request
*next_rq
;
121 /* requests queued in sort_list */
123 /* currently allocated requests */
125 /* fifo list of requests in sort_list */
126 struct list_head fifo
;
128 /* time when queue got scheduled in to dispatch first request. */
129 unsigned long dispatch_start
;
130 unsigned int allocated_slice
;
131 unsigned int slice_dispatch
;
132 /* time when first request from queue completed and slice started. */
133 unsigned long slice_start
;
134 unsigned long slice_end
;
137 /* pending priority requests */
139 /* number of requests that are on the dispatch list or inside driver */
142 /* io prio of this group */
143 unsigned short ioprio
, org_ioprio
;
144 unsigned short ioprio_class
;
149 sector_t last_request_pos
;
151 struct cfq_rb_root
*service_tree
;
152 struct cfq_queue
*new_cfqq
;
153 struct cfq_group
*cfqg
;
154 /* Number of sectors dispatched from queue in single dispatch round */
155 unsigned long nr_sectors
;
159 * First index in the service_trees.
160 * IDLE is handled separately, so it has negative index
170 * Second index in the service_trees.
174 SYNC_NOIDLE_WORKLOAD
= 1,
179 #ifdef CONFIG_CFQ_GROUP_IOSCHED
180 /* total bytes transferred */
181 struct blkg_rwstat service_bytes
;
182 /* total IOs serviced, post merge */
183 struct blkg_rwstat serviced
;
184 /* number of ios merged */
185 struct blkg_rwstat merged
;
186 /* total time spent on device in ns, may not be accurate w/ queueing */
187 struct blkg_rwstat service_time
;
188 /* total time spent waiting in scheduler queue in ns */
189 struct blkg_rwstat wait_time
;
190 /* number of IOs queued up */
191 struct blkg_rwstat queued
;
192 /* total sectors transferred */
193 struct blkg_stat sectors
;
194 /* total disk time and nr sectors dispatched by this group */
195 struct blkg_stat time
;
196 #ifdef CONFIG_DEBUG_BLK_CGROUP
197 /* time not charged to this cgroup */
198 struct blkg_stat unaccounted_time
;
199 /* sum of number of ios queued across all samples */
200 struct blkg_stat avg_queue_size_sum
;
201 /* count of samples taken for average */
202 struct blkg_stat avg_queue_size_samples
;
203 /* how many times this group has been removed from service tree */
204 struct blkg_stat dequeue
;
205 /* total time spent waiting for it to be assigned a timeslice. */
206 struct blkg_stat group_wait_time
;
207 /* time spent idling for this blkcg_gq */
208 struct blkg_stat idle_time
;
209 /* total time with empty current active q with other requests queued */
210 struct blkg_stat empty_time
;
211 /* fields after this shouldn't be cleared on stat reset */
212 uint64_t start_group_wait_time
;
213 uint64_t start_idle_time
;
214 uint64_t start_empty_time
;
216 #endif /* CONFIG_DEBUG_BLK_CGROUP */
217 #endif /* CONFIG_CFQ_GROUP_IOSCHED */
220 /* Per-cgroup data */
221 struct cfq_group_data
{
222 /* must be the first member */
223 struct blkcg_policy_data pd
;
226 unsigned int leaf_weight
;
229 /* This is per cgroup per device grouping structure */
231 /* must be the first member */
232 struct blkg_policy_data pd
;
234 /* group service_tree member */
235 struct rb_node rb_node
;
237 /* group service_tree key */
241 * The number of active cfqgs and sum of their weights under this
242 * cfqg. This covers this cfqg's leaf_weight and all children's
243 * weights, but does not cover weights of further descendants.
245 * If a cfqg is on the service tree, it's active. An active cfqg
246 * also activates its parent and contributes to the children_weight
250 unsigned int children_weight
;
253 * vfraction is the fraction of vdisktime that the tasks in this
254 * cfqg are entitled to. This is determined by compounding the
255 * ratios walking up from this cfqg to the root.
257 * It is in fixed point w/ CFQ_SERVICE_SHIFT and the sum of all
258 * vfractions on a service tree is approximately 1. The sum may
259 * deviate a bit due to rounding errors and fluctuations caused by
260 * cfqgs entering and leaving the service tree.
262 unsigned int vfraction
;
265 * There are two weights - (internal) weight is the weight of this
266 * cfqg against the sibling cfqgs. leaf_weight is the wight of
267 * this cfqg against the child cfqgs. For the root cfqg, both
268 * weights are kept in sync for backward compatibility.
271 unsigned int new_weight
;
272 unsigned int dev_weight
;
274 unsigned int leaf_weight
;
275 unsigned int new_leaf_weight
;
276 unsigned int dev_leaf_weight
;
278 /* number of cfqq currently on this group */
282 * Per group busy queues average. Useful for workload slice calc. We
283 * create the array for each prio class but at run time it is used
284 * only for RT and BE class and slot for IDLE class remains unused.
285 * This is primarily done to avoid confusion and a gcc warning.
287 unsigned int busy_queues_avg
[CFQ_PRIO_NR
];
289 * rr lists of queues with requests. We maintain service trees for
290 * RT and BE classes. These trees are subdivided in subclasses
291 * of SYNC, SYNC_NOIDLE and ASYNC based on workload type. For IDLE
292 * class there is no subclassification and all the cfq queues go on
293 * a single tree service_tree_idle.
294 * Counts are embedded in the cfq_rb_root
296 struct cfq_rb_root service_trees
[2][3];
297 struct cfq_rb_root service_tree_idle
;
299 unsigned long saved_wl_slice
;
300 enum wl_type_t saved_wl_type
;
301 enum wl_class_t saved_wl_class
;
303 /* number of requests that are on the dispatch list or inside driver */
305 struct cfq_ttime ttime
;
306 struct cfqg_stats stats
; /* stats for this cfqg */
307 struct cfqg_stats dead_stats
; /* stats pushed from dead children */
311 struct io_cq icq
; /* must be the first member */
312 struct cfq_queue
*cfqq
[2];
313 struct cfq_ttime ttime
;
314 int ioprio
; /* the current ioprio */
315 #ifdef CONFIG_CFQ_GROUP_IOSCHED
316 uint64_t blkcg_serial_nr
; /* the current blkcg serial */
321 * Per block device queue structure
324 struct request_queue
*queue
;
325 /* Root service tree for cfq_groups */
326 struct cfq_rb_root grp_service_tree
;
327 struct cfq_group
*root_group
;
330 * The priority currently being served
332 enum wl_class_t serving_wl_class
;
333 enum wl_type_t serving_wl_type
;
334 unsigned long workload_expires
;
335 struct cfq_group
*serving_group
;
338 * Each priority tree is sorted by next_request position. These
339 * trees are used when determining if two or more queues are
340 * interleaving requests (see cfq_close_cooperator).
342 struct rb_root prio_trees
[CFQ_PRIO_LISTS
];
344 unsigned int busy_queues
;
345 unsigned int busy_sync_queues
;
351 * queue-depth detection
357 * -1 => indeterminate, (cfq will behave as if NCQ is present, to allow better detection)
358 * 1 => NCQ is present (hw_tag_est_depth is the estimated max depth)
361 int hw_tag_est_depth
;
362 unsigned int hw_tag_samples
;
365 * idle window management
367 struct timer_list idle_slice_timer
;
368 struct work_struct unplug_work
;
370 struct cfq_queue
*active_queue
;
371 struct cfq_io_cq
*active_cic
;
374 * async queue for each priority case
376 struct cfq_queue
*async_cfqq
[2][IOPRIO_BE_NR
];
377 struct cfq_queue
*async_idle_cfqq
;
379 sector_t last_position
;
382 * tunables, see top of file
384 unsigned int cfq_quantum
;
385 unsigned int cfq_fifo_expire
[2];
386 unsigned int cfq_back_penalty
;
387 unsigned int cfq_back_max
;
388 unsigned int cfq_slice
[2];
389 unsigned int cfq_slice_async_rq
;
390 unsigned int cfq_slice_idle
;
391 unsigned int cfq_group_idle
;
392 unsigned int cfq_latency
;
393 unsigned int cfq_target_latency
;
396 * Fallback dummy cfqq for extreme OOM conditions
398 struct cfq_queue oom_cfqq
;
400 unsigned long last_delayed_sync
;
403 static struct cfq_group
*cfq_get_next_cfqg(struct cfq_data
*cfqd
);
405 static struct cfq_rb_root
*st_for(struct cfq_group
*cfqg
,
406 enum wl_class_t
class,
412 if (class == IDLE_WORKLOAD
)
413 return &cfqg
->service_tree_idle
;
415 return &cfqg
->service_trees
[class][type
];
418 enum cfqq_state_flags
{
419 CFQ_CFQQ_FLAG_on_rr
= 0, /* on round-robin busy list */
420 CFQ_CFQQ_FLAG_wait_request
, /* waiting for a request */
421 CFQ_CFQQ_FLAG_must_dispatch
, /* must be allowed a dispatch */
422 CFQ_CFQQ_FLAG_must_alloc_slice
, /* per-slice must_alloc flag */
423 CFQ_CFQQ_FLAG_fifo_expire
, /* FIFO checked in this slice */
424 CFQ_CFQQ_FLAG_idle_window
, /* slice idling enabled */
425 CFQ_CFQQ_FLAG_prio_changed
, /* task priority has changed */
426 CFQ_CFQQ_FLAG_slice_new
, /* no requests dispatched in slice */
427 CFQ_CFQQ_FLAG_sync
, /* synchronous queue */
428 CFQ_CFQQ_FLAG_coop
, /* cfqq is shared */
429 CFQ_CFQQ_FLAG_split_coop
, /* shared cfqq will be splitted */
430 CFQ_CFQQ_FLAG_deep
, /* sync cfqq experienced large depth */
431 CFQ_CFQQ_FLAG_wait_busy
, /* Waiting for next request */
434 #define CFQ_CFQQ_FNS(name) \
435 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \
437 (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name); \
439 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \
441 (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \
443 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \
445 return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \
449 CFQ_CFQQ_FNS(wait_request
);
450 CFQ_CFQQ_FNS(must_dispatch
);
451 CFQ_CFQQ_FNS(must_alloc_slice
);
452 CFQ_CFQQ_FNS(fifo_expire
);
453 CFQ_CFQQ_FNS(idle_window
);
454 CFQ_CFQQ_FNS(prio_changed
);
455 CFQ_CFQQ_FNS(slice_new
);
458 CFQ_CFQQ_FNS(split_coop
);
460 CFQ_CFQQ_FNS(wait_busy
);
463 #if defined(CONFIG_CFQ_GROUP_IOSCHED) && defined(CONFIG_DEBUG_BLK_CGROUP)
465 /* cfqg stats flags */
466 enum cfqg_stats_flags
{
467 CFQG_stats_waiting
= 0,
472 #define CFQG_FLAG_FNS(name) \
473 static inline void cfqg_stats_mark_##name(struct cfqg_stats *stats) \
475 stats->flags |= (1 << CFQG_stats_##name); \
477 static inline void cfqg_stats_clear_##name(struct cfqg_stats *stats) \
479 stats->flags &= ~(1 << CFQG_stats_##name); \
481 static inline int cfqg_stats_##name(struct cfqg_stats *stats) \
483 return (stats->flags & (1 << CFQG_stats_##name)) != 0; \
486 CFQG_FLAG_FNS(waiting)
487 CFQG_FLAG_FNS(idling
)
491 /* This should be called with the queue_lock held. */
492 static void cfqg_stats_update_group_wait_time(struct cfqg_stats
*stats
)
494 unsigned long long now
;
496 if (!cfqg_stats_waiting(stats
))
500 if (time_after64(now
, stats
->start_group_wait_time
))
501 blkg_stat_add(&stats
->group_wait_time
,
502 now
- stats
->start_group_wait_time
);
503 cfqg_stats_clear_waiting(stats
);
506 /* This should be called with the queue_lock held. */
507 static void cfqg_stats_set_start_group_wait_time(struct cfq_group
*cfqg
,
508 struct cfq_group
*curr_cfqg
)
510 struct cfqg_stats
*stats
= &cfqg
->stats
;
512 if (cfqg_stats_waiting(stats
))
514 if (cfqg
== curr_cfqg
)
516 stats
->start_group_wait_time
= sched_clock();
517 cfqg_stats_mark_waiting(stats
);
520 /* This should be called with the queue_lock held. */
521 static void cfqg_stats_end_empty_time(struct cfqg_stats
*stats
)
523 unsigned long long now
;
525 if (!cfqg_stats_empty(stats
))
529 if (time_after64(now
, stats
->start_empty_time
))
530 blkg_stat_add(&stats
->empty_time
,
531 now
- stats
->start_empty_time
);
532 cfqg_stats_clear_empty(stats
);
535 static void cfqg_stats_update_dequeue(struct cfq_group
*cfqg
)
537 blkg_stat_add(&cfqg
->stats
.dequeue
, 1);
540 static void cfqg_stats_set_start_empty_time(struct cfq_group
*cfqg
)
542 struct cfqg_stats
*stats
= &cfqg
->stats
;
544 if (blkg_rwstat_total(&stats
->queued
))
548 * group is already marked empty. This can happen if cfqq got new
549 * request in parent group and moved to this group while being added
550 * to service tree. Just ignore the event and move on.
552 if (cfqg_stats_empty(stats
))
555 stats
->start_empty_time
= sched_clock();
556 cfqg_stats_mark_empty(stats
);
559 static void cfqg_stats_update_idle_time(struct cfq_group
*cfqg
)
561 struct cfqg_stats
*stats
= &cfqg
->stats
;
563 if (cfqg_stats_idling(stats
)) {
564 unsigned long long now
= sched_clock();
566 if (time_after64(now
, stats
->start_idle_time
))
567 blkg_stat_add(&stats
->idle_time
,
568 now
- stats
->start_idle_time
);
569 cfqg_stats_clear_idling(stats
);
573 static void cfqg_stats_set_start_idle_time(struct cfq_group
*cfqg
)
575 struct cfqg_stats
*stats
= &cfqg
->stats
;
577 BUG_ON(cfqg_stats_idling(stats
));
579 stats
->start_idle_time
= sched_clock();
580 cfqg_stats_mark_idling(stats
);
583 static void cfqg_stats_update_avg_queue_size(struct cfq_group
*cfqg
)
585 struct cfqg_stats
*stats
= &cfqg
->stats
;
587 blkg_stat_add(&stats
->avg_queue_size_sum
,
588 blkg_rwstat_total(&stats
->queued
));
589 blkg_stat_add(&stats
->avg_queue_size_samples
, 1);
590 cfqg_stats_update_group_wait_time(stats
);
593 #else /* CONFIG_CFQ_GROUP_IOSCHED && CONFIG_DEBUG_BLK_CGROUP */
595 static inline void cfqg_stats_set_start_group_wait_time(struct cfq_group
*cfqg
, struct cfq_group
*curr_cfqg
) { }
596 static inline void cfqg_stats_end_empty_time(struct cfqg_stats
*stats
) { }
597 static inline void cfqg_stats_update_dequeue(struct cfq_group
*cfqg
) { }
598 static inline void cfqg_stats_set_start_empty_time(struct cfq_group
*cfqg
) { }
599 static inline void cfqg_stats_update_idle_time(struct cfq_group
*cfqg
) { }
600 static inline void cfqg_stats_set_start_idle_time(struct cfq_group
*cfqg
) { }
601 static inline void cfqg_stats_update_avg_queue_size(struct cfq_group
*cfqg
) { }
603 #endif /* CONFIG_CFQ_GROUP_IOSCHED && CONFIG_DEBUG_BLK_CGROUP */
605 #ifdef CONFIG_CFQ_GROUP_IOSCHED
607 static inline struct cfq_group
*pd_to_cfqg(struct blkg_policy_data
*pd
)
609 return pd
? container_of(pd
, struct cfq_group
, pd
) : NULL
;
612 static struct cfq_group_data
613 *cpd_to_cfqgd(struct blkcg_policy_data
*cpd
)
615 return cpd
? container_of(cpd
, struct cfq_group_data
, pd
) : NULL
;
618 static inline struct blkcg_gq
*cfqg_to_blkg(struct cfq_group
*cfqg
)
620 return pd_to_blkg(&cfqg
->pd
);
623 static struct blkcg_policy blkcg_policy_cfq
;
625 static inline struct cfq_group
*blkg_to_cfqg(struct blkcg_gq
*blkg
)
627 return pd_to_cfqg(blkg_to_pd(blkg
, &blkcg_policy_cfq
));
630 static struct cfq_group_data
*blkcg_to_cfqgd(struct blkcg
*blkcg
)
632 return cpd_to_cfqgd(blkcg_to_cpd(blkcg
, &blkcg_policy_cfq
));
635 static inline struct cfq_group
*cfqg_parent(struct cfq_group
*cfqg
)
637 struct blkcg_gq
*pblkg
= cfqg_to_blkg(cfqg
)->parent
;
639 return pblkg
? blkg_to_cfqg(pblkg
) : NULL
;
642 static inline void cfqg_get(struct cfq_group
*cfqg
)
644 return blkg_get(cfqg_to_blkg(cfqg
));
647 static inline void cfqg_put(struct cfq_group
*cfqg
)
649 return blkg_put(cfqg_to_blkg(cfqg
));
652 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...) do { \
655 blkg_path(cfqg_to_blkg((cfqq)->cfqg), __pbuf, sizeof(__pbuf)); \
656 blk_add_trace_msg((cfqd)->queue, "cfq%d%c%c %s " fmt, (cfqq)->pid, \
657 cfq_cfqq_sync((cfqq)) ? 'S' : 'A', \
658 cfqq_type((cfqq)) == SYNC_NOIDLE_WORKLOAD ? 'N' : ' ',\
662 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...) do { \
665 blkg_path(cfqg_to_blkg(cfqg), __pbuf, sizeof(__pbuf)); \
666 blk_add_trace_msg((cfqd)->queue, "%s " fmt, __pbuf, ##args); \
669 static inline void cfqg_stats_update_io_add(struct cfq_group
*cfqg
,
670 struct cfq_group
*curr_cfqg
, int rw
)
672 blkg_rwstat_add(&cfqg
->stats
.queued
, rw
, 1);
673 cfqg_stats_end_empty_time(&cfqg
->stats
);
674 cfqg_stats_set_start_group_wait_time(cfqg
, curr_cfqg
);
677 static inline void cfqg_stats_update_timeslice_used(struct cfq_group
*cfqg
,
678 unsigned long time
, unsigned long unaccounted_time
)
680 blkg_stat_add(&cfqg
->stats
.time
, time
);
681 #ifdef CONFIG_DEBUG_BLK_CGROUP
682 blkg_stat_add(&cfqg
->stats
.unaccounted_time
, unaccounted_time
);
686 static inline void cfqg_stats_update_io_remove(struct cfq_group
*cfqg
, int rw
)
688 blkg_rwstat_add(&cfqg
->stats
.queued
, rw
, -1);
691 static inline void cfqg_stats_update_io_merged(struct cfq_group
*cfqg
, int rw
)
693 blkg_rwstat_add(&cfqg
->stats
.merged
, rw
, 1);
696 static inline void cfqg_stats_update_dispatch(struct cfq_group
*cfqg
,
697 uint64_t bytes
, int rw
)
699 blkg_stat_add(&cfqg
->stats
.sectors
, bytes
>> 9);
700 blkg_rwstat_add(&cfqg
->stats
.serviced
, rw
, 1);
701 blkg_rwstat_add(&cfqg
->stats
.service_bytes
, rw
, bytes
);
704 static inline void cfqg_stats_update_completion(struct cfq_group
*cfqg
,
705 uint64_t start_time
, uint64_t io_start_time
, int rw
)
707 struct cfqg_stats
*stats
= &cfqg
->stats
;
708 unsigned long long now
= sched_clock();
710 if (time_after64(now
, io_start_time
))
711 blkg_rwstat_add(&stats
->service_time
, rw
, now
- io_start_time
);
712 if (time_after64(io_start_time
, start_time
))
713 blkg_rwstat_add(&stats
->wait_time
, rw
,
714 io_start_time
- start_time
);
718 static void cfqg_stats_reset(struct cfqg_stats
*stats
)
720 /* queued stats shouldn't be cleared */
721 blkg_rwstat_reset(&stats
->service_bytes
);
722 blkg_rwstat_reset(&stats
->serviced
);
723 blkg_rwstat_reset(&stats
->merged
);
724 blkg_rwstat_reset(&stats
->service_time
);
725 blkg_rwstat_reset(&stats
->wait_time
);
726 blkg_stat_reset(&stats
->time
);
727 #ifdef CONFIG_DEBUG_BLK_CGROUP
728 blkg_stat_reset(&stats
->unaccounted_time
);
729 blkg_stat_reset(&stats
->avg_queue_size_sum
);
730 blkg_stat_reset(&stats
->avg_queue_size_samples
);
731 blkg_stat_reset(&stats
->dequeue
);
732 blkg_stat_reset(&stats
->group_wait_time
);
733 blkg_stat_reset(&stats
->idle_time
);
734 blkg_stat_reset(&stats
->empty_time
);
739 static void cfqg_stats_merge(struct cfqg_stats
*to
, struct cfqg_stats
*from
)
741 /* queued stats shouldn't be cleared */
742 blkg_rwstat_merge(&to
->service_bytes
, &from
->service_bytes
);
743 blkg_rwstat_merge(&to
->serviced
, &from
->serviced
);
744 blkg_rwstat_merge(&to
->merged
, &from
->merged
);
745 blkg_rwstat_merge(&to
->service_time
, &from
->service_time
);
746 blkg_rwstat_merge(&to
->wait_time
, &from
->wait_time
);
747 blkg_stat_merge(&from
->time
, &from
->time
);
748 #ifdef CONFIG_DEBUG_BLK_CGROUP
749 blkg_stat_merge(&to
->unaccounted_time
, &from
->unaccounted_time
);
750 blkg_stat_merge(&to
->avg_queue_size_sum
, &from
->avg_queue_size_sum
);
751 blkg_stat_merge(&to
->avg_queue_size_samples
, &from
->avg_queue_size_samples
);
752 blkg_stat_merge(&to
->dequeue
, &from
->dequeue
);
753 blkg_stat_merge(&to
->group_wait_time
, &from
->group_wait_time
);
754 blkg_stat_merge(&to
->idle_time
, &from
->idle_time
);
755 blkg_stat_merge(&to
->empty_time
, &from
->empty_time
);
760 * Transfer @cfqg's stats to its parent's dead_stats so that the ancestors'
761 * recursive stats can still account for the amount used by this cfqg after
764 static void cfqg_stats_xfer_dead(struct cfq_group
*cfqg
)
766 struct cfq_group
*parent
= cfqg_parent(cfqg
);
768 lockdep_assert_held(cfqg_to_blkg(cfqg
)->q
->queue_lock
);
770 if (unlikely(!parent
))
773 cfqg_stats_merge(&parent
->dead_stats
, &cfqg
->stats
);
774 cfqg_stats_merge(&parent
->dead_stats
, &cfqg
->dead_stats
);
775 cfqg_stats_reset(&cfqg
->stats
);
776 cfqg_stats_reset(&cfqg
->dead_stats
);
779 #else /* CONFIG_CFQ_GROUP_IOSCHED */
781 static inline struct cfq_group
*cfqg_parent(struct cfq_group
*cfqg
) { return NULL
; }
782 static inline void cfqg_get(struct cfq_group
*cfqg
) { }
783 static inline void cfqg_put(struct cfq_group
*cfqg
) { }
785 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
786 blk_add_trace_msg((cfqd)->queue, "cfq%d%c%c " fmt, (cfqq)->pid, \
787 cfq_cfqq_sync((cfqq)) ? 'S' : 'A', \
788 cfqq_type((cfqq)) == SYNC_NOIDLE_WORKLOAD ? 'N' : ' ',\
790 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...) do {} while (0)
792 static inline void cfqg_stats_update_io_add(struct cfq_group
*cfqg
,
793 struct cfq_group
*curr_cfqg
, int rw
) { }
794 static inline void cfqg_stats_update_timeslice_used(struct cfq_group
*cfqg
,
795 unsigned long time
, unsigned long unaccounted_time
) { }
796 static inline void cfqg_stats_update_io_remove(struct cfq_group
*cfqg
, int rw
) { }
797 static inline void cfqg_stats_update_io_merged(struct cfq_group
*cfqg
, int rw
) { }
798 static inline void cfqg_stats_update_dispatch(struct cfq_group
*cfqg
,
799 uint64_t bytes
, int rw
) { }
800 static inline void cfqg_stats_update_completion(struct cfq_group
*cfqg
,
801 uint64_t start_time
, uint64_t io_start_time
, int rw
) { }
803 #endif /* CONFIG_CFQ_GROUP_IOSCHED */
805 #define cfq_log(cfqd, fmt, args...) \
806 blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
808 /* Traverses through cfq group service trees */
809 #define for_each_cfqg_st(cfqg, i, j, st) \
810 for (i = 0; i <= IDLE_WORKLOAD; i++) \
811 for (j = 0, st = i < IDLE_WORKLOAD ? &cfqg->service_trees[i][j]\
812 : &cfqg->service_tree_idle; \
813 (i < IDLE_WORKLOAD && j <= SYNC_WORKLOAD) || \
814 (i == IDLE_WORKLOAD && j == 0); \
815 j++, st = i < IDLE_WORKLOAD ? \
816 &cfqg->service_trees[i][j]: NULL) \
818 static inline bool cfq_io_thinktime_big(struct cfq_data *cfqd,
819 struct cfq_ttime
*ttime
, bool group_idle
)
822 if (!sample_valid(ttime
->ttime_samples
))
825 slice
= cfqd
->cfq_group_idle
;
827 slice
= cfqd
->cfq_slice_idle
;
828 return ttime
->ttime_mean
> slice
;
831 static inline bool iops_mode(struct cfq_data
*cfqd
)
834 * If we are not idling on queues and it is a NCQ drive, parallel
835 * execution of requests is on and measuring time is not possible
836 * in most of the cases until and unless we drive shallower queue
837 * depths and that becomes a performance bottleneck. In such cases
838 * switch to start providing fairness in terms of number of IOs.
840 if (!cfqd
->cfq_slice_idle
&& cfqd
->hw_tag
)
846 static inline enum wl_class_t
cfqq_class(struct cfq_queue
*cfqq
)
848 if (cfq_class_idle(cfqq
))
849 return IDLE_WORKLOAD
;
850 if (cfq_class_rt(cfqq
))
856 static enum wl_type_t
cfqq_type(struct cfq_queue
*cfqq
)
858 if (!cfq_cfqq_sync(cfqq
))
859 return ASYNC_WORKLOAD
;
860 if (!cfq_cfqq_idle_window(cfqq
))
861 return SYNC_NOIDLE_WORKLOAD
;
862 return SYNC_WORKLOAD
;
865 static inline int cfq_group_busy_queues_wl(enum wl_class_t wl_class
,
866 struct cfq_data
*cfqd
,
867 struct cfq_group
*cfqg
)
869 if (wl_class
== IDLE_WORKLOAD
)
870 return cfqg
->service_tree_idle
.count
;
872 return cfqg
->service_trees
[wl_class
][ASYNC_WORKLOAD
].count
+
873 cfqg
->service_trees
[wl_class
][SYNC_NOIDLE_WORKLOAD
].count
+
874 cfqg
->service_trees
[wl_class
][SYNC_WORKLOAD
].count
;
877 static inline int cfqg_busy_async_queues(struct cfq_data
*cfqd
,
878 struct cfq_group
*cfqg
)
880 return cfqg
->service_trees
[RT_WORKLOAD
][ASYNC_WORKLOAD
].count
+
881 cfqg
->service_trees
[BE_WORKLOAD
][ASYNC_WORKLOAD
].count
;
884 static void cfq_dispatch_insert(struct request_queue
*, struct request
*);
885 static struct cfq_queue
*cfq_get_queue(struct cfq_data
*cfqd
, bool is_sync
,
886 struct cfq_io_cq
*cic
, struct bio
*bio
);
888 static inline struct cfq_io_cq
*icq_to_cic(struct io_cq
*icq
)
890 /* cic->icq is the first member, %NULL will convert to %NULL */
891 return container_of(icq
, struct cfq_io_cq
, icq
);
894 static inline struct cfq_io_cq
*cfq_cic_lookup(struct cfq_data
*cfqd
,
895 struct io_context
*ioc
)
898 return icq_to_cic(ioc_lookup_icq(ioc
, cfqd
->queue
));
902 static inline struct cfq_queue
*cic_to_cfqq(struct cfq_io_cq
*cic
, bool is_sync
)
904 return cic
->cfqq
[is_sync
];
907 static inline void cic_set_cfqq(struct cfq_io_cq
*cic
, struct cfq_queue
*cfqq
,
910 cic
->cfqq
[is_sync
] = cfqq
;
913 static inline struct cfq_data
*cic_to_cfqd(struct cfq_io_cq
*cic
)
915 return cic
->icq
.q
->elevator
->elevator_data
;
919 * We regard a request as SYNC, if it's either a read or has the SYNC bit
920 * set (in which case it could also be direct WRITE).
922 static inline bool cfq_bio_sync(struct bio
*bio
)
924 return bio_data_dir(bio
) == READ
|| (bio
->bi_rw
& REQ_SYNC
);
928 * scheduler run of queue, if there are requests pending and no one in the
929 * driver that will restart queueing
931 static inline void cfq_schedule_dispatch(struct cfq_data
*cfqd
)
933 if (cfqd
->busy_queues
) {
934 cfq_log(cfqd
, "schedule dispatch");
935 kblockd_schedule_work(&cfqd
->unplug_work
);
940 * Scale schedule slice based on io priority. Use the sync time slice only
941 * if a queue is marked sync and has sync io queued. A sync queue with async
942 * io only, should not get full sync slice length.
944 static inline int cfq_prio_slice(struct cfq_data
*cfqd
, bool sync
,
947 const int base_slice
= cfqd
->cfq_slice
[sync
];
949 WARN_ON(prio
>= IOPRIO_BE_NR
);
951 return base_slice
+ (base_slice
/CFQ_SLICE_SCALE
* (4 - prio
));
955 cfq_prio_to_slice(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
957 return cfq_prio_slice(cfqd
, cfq_cfqq_sync(cfqq
), cfqq
->ioprio
);
961 * cfqg_scale_charge - scale disk time charge according to cfqg weight
962 * @charge: disk time being charged
963 * @vfraction: vfraction of the cfqg, fixed point w/ CFQ_SERVICE_SHIFT
965 * Scale @charge according to @vfraction, which is in range (0, 1]. The
966 * scaling is inversely proportional.
968 * scaled = charge / vfraction
970 * The result is also in fixed point w/ CFQ_SERVICE_SHIFT.
972 static inline u64
cfqg_scale_charge(unsigned long charge
,
973 unsigned int vfraction
)
975 u64 c
= charge
<< CFQ_SERVICE_SHIFT
; /* make it fixed point */
977 /* charge / vfraction */
978 c
<<= CFQ_SERVICE_SHIFT
;
979 do_div(c
, vfraction
);
983 static inline u64
max_vdisktime(u64 min_vdisktime
, u64 vdisktime
)
985 s64 delta
= (s64
)(vdisktime
- min_vdisktime
);
987 min_vdisktime
= vdisktime
;
989 return min_vdisktime
;
992 static inline u64
min_vdisktime(u64 min_vdisktime
, u64 vdisktime
)
994 s64 delta
= (s64
)(vdisktime
- min_vdisktime
);
996 min_vdisktime
= vdisktime
;
998 return min_vdisktime
;
1001 static void update_min_vdisktime(struct cfq_rb_root
*st
)
1003 struct cfq_group
*cfqg
;
1006 cfqg
= rb_entry_cfqg(st
->left
);
1007 st
->min_vdisktime
= max_vdisktime(st
->min_vdisktime
,
1013 * get averaged number of queues of RT/BE priority.
1014 * average is updated, with a formula that gives more weight to higher numbers,
1015 * to quickly follows sudden increases and decrease slowly
1018 static inline unsigned cfq_group_get_avg_queues(struct cfq_data
*cfqd
,
1019 struct cfq_group
*cfqg
, bool rt
)
1021 unsigned min_q
, max_q
;
1022 unsigned mult
= cfq_hist_divisor
- 1;
1023 unsigned round
= cfq_hist_divisor
/ 2;
1024 unsigned busy
= cfq_group_busy_queues_wl(rt
, cfqd
, cfqg
);
1026 min_q
= min(cfqg
->busy_queues_avg
[rt
], busy
);
1027 max_q
= max(cfqg
->busy_queues_avg
[rt
], busy
);
1028 cfqg
->busy_queues_avg
[rt
] = (mult
* max_q
+ min_q
+ round
) /
1030 return cfqg
->busy_queues_avg
[rt
];
1033 static inline unsigned
1034 cfq_group_slice(struct cfq_data
*cfqd
, struct cfq_group
*cfqg
)
1036 return cfqd
->cfq_target_latency
* cfqg
->vfraction
>> CFQ_SERVICE_SHIFT
;
1039 static inline unsigned
1040 cfq_scaled_cfqq_slice(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
1042 unsigned slice
= cfq_prio_to_slice(cfqd
, cfqq
);
1043 if (cfqd
->cfq_latency
) {
1045 * interested queues (we consider only the ones with the same
1046 * priority class in the cfq group)
1048 unsigned iq
= cfq_group_get_avg_queues(cfqd
, cfqq
->cfqg
,
1049 cfq_class_rt(cfqq
));
1050 unsigned sync_slice
= cfqd
->cfq_slice
[1];
1051 unsigned expect_latency
= sync_slice
* iq
;
1052 unsigned group_slice
= cfq_group_slice(cfqd
, cfqq
->cfqg
);
1054 if (expect_latency
> group_slice
) {
1055 unsigned base_low_slice
= 2 * cfqd
->cfq_slice_idle
;
1056 /* scale low_slice according to IO priority
1057 * and sync vs async */
1058 unsigned low_slice
=
1059 min(slice
, base_low_slice
* slice
/ sync_slice
);
1060 /* the adapted slice value is scaled to fit all iqs
1061 * into the target latency */
1062 slice
= max(slice
* group_slice
/ expect_latency
,
1070 cfq_set_prio_slice(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
1072 unsigned slice
= cfq_scaled_cfqq_slice(cfqd
, cfqq
);
1074 cfqq
->slice_start
= jiffies
;
1075 cfqq
->slice_end
= jiffies
+ slice
;
1076 cfqq
->allocated_slice
= slice
;
1077 cfq_log_cfqq(cfqd
, cfqq
, "set_slice=%lu", cfqq
->slice_end
- jiffies
);
1081 * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
1082 * isn't valid until the first request from the dispatch is activated
1083 * and the slice time set.
1085 static inline bool cfq_slice_used(struct cfq_queue
*cfqq
)
1087 if (cfq_cfqq_slice_new(cfqq
))
1089 if (time_before(jiffies
, cfqq
->slice_end
))
1096 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
1097 * We choose the request that is closest to the head right now. Distance
1098 * behind the head is penalized and only allowed to a certain extent.
1100 static struct request
*
1101 cfq_choose_req(struct cfq_data
*cfqd
, struct request
*rq1
, struct request
*rq2
, sector_t last
)
1103 sector_t s1
, s2
, d1
= 0, d2
= 0;
1104 unsigned long back_max
;
1105 #define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */
1106 #define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */
1107 unsigned wrap
= 0; /* bit mask: requests behind the disk head? */
1109 if (rq1
== NULL
|| rq1
== rq2
)
1114 if (rq_is_sync(rq1
) != rq_is_sync(rq2
))
1115 return rq_is_sync(rq1
) ? rq1
: rq2
;
1117 if ((rq1
->cmd_flags
^ rq2
->cmd_flags
) & REQ_PRIO
)
1118 return rq1
->cmd_flags
& REQ_PRIO
? rq1
: rq2
;
1120 s1
= blk_rq_pos(rq1
);
1121 s2
= blk_rq_pos(rq2
);
1124 * by definition, 1KiB is 2 sectors
1126 back_max
= cfqd
->cfq_back_max
* 2;
1129 * Strict one way elevator _except_ in the case where we allow
1130 * short backward seeks which are biased as twice the cost of a
1131 * similar forward seek.
1135 else if (s1
+ back_max
>= last
)
1136 d1
= (last
- s1
) * cfqd
->cfq_back_penalty
;
1138 wrap
|= CFQ_RQ1_WRAP
;
1142 else if (s2
+ back_max
>= last
)
1143 d2
= (last
- s2
) * cfqd
->cfq_back_penalty
;
1145 wrap
|= CFQ_RQ2_WRAP
;
1147 /* Found required data */
1150 * By doing switch() on the bit mask "wrap" we avoid having to
1151 * check two variables for all permutations: --> faster!
1154 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
1170 case (CFQ_RQ1_WRAP
|CFQ_RQ2_WRAP
): /* both rqs wrapped */
1173 * Since both rqs are wrapped,
1174 * start with the one that's further behind head
1175 * (--> only *one* back seek required),
1176 * since back seek takes more time than forward.
1186 * The below is leftmost cache rbtree addon
1188 static struct cfq_queue
*cfq_rb_first(struct cfq_rb_root
*root
)
1190 /* Service tree is empty */
1195 root
->left
= rb_first(&root
->rb
);
1198 return rb_entry(root
->left
, struct cfq_queue
, rb_node
);
1203 static struct cfq_group
*cfq_rb_first_group(struct cfq_rb_root
*root
)
1206 root
->left
= rb_first(&root
->rb
);
1209 return rb_entry_cfqg(root
->left
);
1214 static void rb_erase_init(struct rb_node
*n
, struct rb_root
*root
)
1220 static void cfq_rb_erase(struct rb_node
*n
, struct cfq_rb_root
*root
)
1222 if (root
->left
== n
)
1224 rb_erase_init(n
, &root
->rb
);
1229 * would be nice to take fifo expire time into account as well
1231 static struct request
*
1232 cfq_find_next_rq(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
1233 struct request
*last
)
1235 struct rb_node
*rbnext
= rb_next(&last
->rb_node
);
1236 struct rb_node
*rbprev
= rb_prev(&last
->rb_node
);
1237 struct request
*next
= NULL
, *prev
= NULL
;
1239 BUG_ON(RB_EMPTY_NODE(&last
->rb_node
));
1242 prev
= rb_entry_rq(rbprev
);
1245 next
= rb_entry_rq(rbnext
);
1247 rbnext
= rb_first(&cfqq
->sort_list
);
1248 if (rbnext
&& rbnext
!= &last
->rb_node
)
1249 next
= rb_entry_rq(rbnext
);
1252 return cfq_choose_req(cfqd
, next
, prev
, blk_rq_pos(last
));
1255 static unsigned long cfq_slice_offset(struct cfq_data
*cfqd
,
1256 struct cfq_queue
*cfqq
)
1259 * just an approximation, should be ok.
1261 return (cfqq
->cfqg
->nr_cfqq
- 1) * (cfq_prio_slice(cfqd
, 1, 0) -
1262 cfq_prio_slice(cfqd
, cfq_cfqq_sync(cfqq
), cfqq
->ioprio
));
1266 cfqg_key(struct cfq_rb_root
*st
, struct cfq_group
*cfqg
)
1268 return cfqg
->vdisktime
- st
->min_vdisktime
;
1272 __cfq_group_service_tree_add(struct cfq_rb_root
*st
, struct cfq_group
*cfqg
)
1274 struct rb_node
**node
= &st
->rb
.rb_node
;
1275 struct rb_node
*parent
= NULL
;
1276 struct cfq_group
*__cfqg
;
1277 s64 key
= cfqg_key(st
, cfqg
);
1280 while (*node
!= NULL
) {
1282 __cfqg
= rb_entry_cfqg(parent
);
1284 if (key
< cfqg_key(st
, __cfqg
))
1285 node
= &parent
->rb_left
;
1287 node
= &parent
->rb_right
;
1293 st
->left
= &cfqg
->rb_node
;
1295 rb_link_node(&cfqg
->rb_node
, parent
, node
);
1296 rb_insert_color(&cfqg
->rb_node
, &st
->rb
);
1300 * This has to be called only on activation of cfqg
1303 cfq_update_group_weight(struct cfq_group
*cfqg
)
1305 if (cfqg
->new_weight
) {
1306 cfqg
->weight
= cfqg
->new_weight
;
1307 cfqg
->new_weight
= 0;
1312 cfq_update_group_leaf_weight(struct cfq_group
*cfqg
)
1314 BUG_ON(!RB_EMPTY_NODE(&cfqg
->rb_node
));
1316 if (cfqg
->new_leaf_weight
) {
1317 cfqg
->leaf_weight
= cfqg
->new_leaf_weight
;
1318 cfqg
->new_leaf_weight
= 0;
1323 cfq_group_service_tree_add(struct cfq_rb_root
*st
, struct cfq_group
*cfqg
)
1325 unsigned int vfr
= 1 << CFQ_SERVICE_SHIFT
; /* start with 1 */
1326 struct cfq_group
*pos
= cfqg
;
1327 struct cfq_group
*parent
;
1330 /* add to the service tree */
1331 BUG_ON(!RB_EMPTY_NODE(&cfqg
->rb_node
));
1334 * Update leaf_weight. We cannot update weight at this point
1335 * because cfqg might already have been activated and is
1336 * contributing its current weight to the parent's child_weight.
1338 cfq_update_group_leaf_weight(cfqg
);
1339 __cfq_group_service_tree_add(st
, cfqg
);
1342 * Activate @cfqg and calculate the portion of vfraction @cfqg is
1343 * entitled to. vfraction is calculated by walking the tree
1344 * towards the root calculating the fraction it has at each level.
1345 * The compounded ratio is how much vfraction @cfqg owns.
1347 * Start with the proportion tasks in this cfqg has against active
1348 * children cfqgs - its leaf_weight against children_weight.
1350 propagate
= !pos
->nr_active
++;
1351 pos
->children_weight
+= pos
->leaf_weight
;
1352 vfr
= vfr
* pos
->leaf_weight
/ pos
->children_weight
;
1355 * Compound ->weight walking up the tree. Both activation and
1356 * vfraction calculation are done in the same loop. Propagation
1357 * stops once an already activated node is met. vfraction
1358 * calculation should always continue to the root.
1360 while ((parent
= cfqg_parent(pos
))) {
1362 cfq_update_group_weight(pos
);
1363 propagate
= !parent
->nr_active
++;
1364 parent
->children_weight
+= pos
->weight
;
1366 vfr
= vfr
* pos
->weight
/ parent
->children_weight
;
1370 cfqg
->vfraction
= max_t(unsigned, vfr
, 1);
1374 cfq_group_notify_queue_add(struct cfq_data
*cfqd
, struct cfq_group
*cfqg
)
1376 struct cfq_rb_root
*st
= &cfqd
->grp_service_tree
;
1377 struct cfq_group
*__cfqg
;
1381 if (!RB_EMPTY_NODE(&cfqg
->rb_node
))
1385 * Currently put the group at the end. Later implement something
1386 * so that groups get lesser vtime based on their weights, so that
1387 * if group does not loose all if it was not continuously backlogged.
1389 n
= rb_last(&st
->rb
);
1391 __cfqg
= rb_entry_cfqg(n
);
1392 cfqg
->vdisktime
= __cfqg
->vdisktime
+ CFQ_IDLE_DELAY
;
1394 cfqg
->vdisktime
= st
->min_vdisktime
;
1395 cfq_group_service_tree_add(st
, cfqg
);
1399 cfq_group_service_tree_del(struct cfq_rb_root
*st
, struct cfq_group
*cfqg
)
1401 struct cfq_group
*pos
= cfqg
;
1405 * Undo activation from cfq_group_service_tree_add(). Deactivate
1406 * @cfqg and propagate deactivation upwards.
1408 propagate
= !--pos
->nr_active
;
1409 pos
->children_weight
-= pos
->leaf_weight
;
1412 struct cfq_group
*parent
= cfqg_parent(pos
);
1414 /* @pos has 0 nr_active at this point */
1415 WARN_ON_ONCE(pos
->children_weight
);
1421 propagate
= !--parent
->nr_active
;
1422 parent
->children_weight
-= pos
->weight
;
1426 /* remove from the service tree */
1427 if (!RB_EMPTY_NODE(&cfqg
->rb_node
))
1428 cfq_rb_erase(&cfqg
->rb_node
, st
);
1432 cfq_group_notify_queue_del(struct cfq_data
*cfqd
, struct cfq_group
*cfqg
)
1434 struct cfq_rb_root
*st
= &cfqd
->grp_service_tree
;
1436 BUG_ON(cfqg
->nr_cfqq
< 1);
1439 /* If there are other cfq queues under this group, don't delete it */
1443 cfq_log_cfqg(cfqd
, cfqg
, "del_from_rr group");
1444 cfq_group_service_tree_del(st
, cfqg
);
1445 cfqg
->saved_wl_slice
= 0;
1446 cfqg_stats_update_dequeue(cfqg
);
1449 static inline unsigned int cfq_cfqq_slice_usage(struct cfq_queue
*cfqq
,
1450 unsigned int *unaccounted_time
)
1452 unsigned int slice_used
;
1455 * Queue got expired before even a single request completed or
1456 * got expired immediately after first request completion.
1458 if (!cfqq
->slice_start
|| cfqq
->slice_start
== jiffies
) {
1460 * Also charge the seek time incurred to the group, otherwise
1461 * if there are mutiple queues in the group, each can dispatch
1462 * a single request on seeky media and cause lots of seek time
1463 * and group will never know it.
1465 slice_used
= max_t(unsigned, (jiffies
- cfqq
->dispatch_start
),
1468 slice_used
= jiffies
- cfqq
->slice_start
;
1469 if (slice_used
> cfqq
->allocated_slice
) {
1470 *unaccounted_time
= slice_used
- cfqq
->allocated_slice
;
1471 slice_used
= cfqq
->allocated_slice
;
1473 if (time_after(cfqq
->slice_start
, cfqq
->dispatch_start
))
1474 *unaccounted_time
+= cfqq
->slice_start
-
1475 cfqq
->dispatch_start
;
1481 static void cfq_group_served(struct cfq_data
*cfqd
, struct cfq_group
*cfqg
,
1482 struct cfq_queue
*cfqq
)
1484 struct cfq_rb_root
*st
= &cfqd
->grp_service_tree
;
1485 unsigned int used_sl
, charge
, unaccounted_sl
= 0;
1486 int nr_sync
= cfqg
->nr_cfqq
- cfqg_busy_async_queues(cfqd
, cfqg
)
1487 - cfqg
->service_tree_idle
.count
;
1490 BUG_ON(nr_sync
< 0);
1491 used_sl
= charge
= cfq_cfqq_slice_usage(cfqq
, &unaccounted_sl
);
1493 if (iops_mode(cfqd
))
1494 charge
= cfqq
->slice_dispatch
;
1495 else if (!cfq_cfqq_sync(cfqq
) && !nr_sync
)
1496 charge
= cfqq
->allocated_slice
;
1499 * Can't update vdisktime while on service tree and cfqg->vfraction
1500 * is valid only while on it. Cache vfr, leave the service tree,
1501 * update vdisktime and go back on. The re-addition to the tree
1502 * will also update the weights as necessary.
1504 vfr
= cfqg
->vfraction
;
1505 cfq_group_service_tree_del(st
, cfqg
);
1506 cfqg
->vdisktime
+= cfqg_scale_charge(charge
, vfr
);
1507 cfq_group_service_tree_add(st
, cfqg
);
1509 /* This group is being expired. Save the context */
1510 if (time_after(cfqd
->workload_expires
, jiffies
)) {
1511 cfqg
->saved_wl_slice
= cfqd
->workload_expires
1513 cfqg
->saved_wl_type
= cfqd
->serving_wl_type
;
1514 cfqg
->saved_wl_class
= cfqd
->serving_wl_class
;
1516 cfqg
->saved_wl_slice
= 0;
1518 cfq_log_cfqg(cfqd
, cfqg
, "served: vt=%llu min_vt=%llu", cfqg
->vdisktime
,
1520 cfq_log_cfqq(cfqq
->cfqd
, cfqq
,
1521 "sl_used=%u disp=%u charge=%u iops=%u sect=%lu",
1522 used_sl
, cfqq
->slice_dispatch
, charge
,
1523 iops_mode(cfqd
), cfqq
->nr_sectors
);
1524 cfqg_stats_update_timeslice_used(cfqg
, used_sl
, unaccounted_sl
);
1525 cfqg_stats_set_start_empty_time(cfqg
);
1529 * cfq_init_cfqg_base - initialize base part of a cfq_group
1530 * @cfqg: cfq_group to initialize
1532 * Initialize the base part which is used whether %CONFIG_CFQ_GROUP_IOSCHED
1533 * is enabled or not.
1535 static void cfq_init_cfqg_base(struct cfq_group
*cfqg
)
1537 struct cfq_rb_root
*st
;
1540 for_each_cfqg_st(cfqg
, i
, j
, st
)
1542 RB_CLEAR_NODE(&cfqg
->rb_node
);
1544 cfqg
->ttime
.last_end_request
= jiffies
;
1547 #ifdef CONFIG_CFQ_GROUP_IOSCHED
1548 static void cfqg_stats_init(struct cfqg_stats
*stats
)
1550 blkg_rwstat_init(&stats
->service_bytes
);
1551 blkg_rwstat_init(&stats
->serviced
);
1552 blkg_rwstat_init(&stats
->merged
);
1553 blkg_rwstat_init(&stats
->service_time
);
1554 blkg_rwstat_init(&stats
->wait_time
);
1555 blkg_rwstat_init(&stats
->queued
);
1557 blkg_stat_init(&stats
->sectors
);
1558 blkg_stat_init(&stats
->time
);
1560 #ifdef CONFIG_DEBUG_BLK_CGROUP
1561 blkg_stat_init(&stats
->unaccounted_time
);
1562 blkg_stat_init(&stats
->avg_queue_size_sum
);
1563 blkg_stat_init(&stats
->avg_queue_size_samples
);
1564 blkg_stat_init(&stats
->dequeue
);
1565 blkg_stat_init(&stats
->group_wait_time
);
1566 blkg_stat_init(&stats
->idle_time
);
1567 blkg_stat_init(&stats
->empty_time
);
1571 static void cfq_cpd_init(const struct blkcg
*blkcg
)
1573 struct cfq_group_data
*cgd
=
1574 cpd_to_cfqgd(blkcg
->pd
[blkcg_policy_cfq
.plid
]);
1576 if (blkcg
== &blkcg_root
) {
1577 cgd
->weight
= 2 * CFQ_WEIGHT_DEFAULT
;
1578 cgd
->leaf_weight
= 2 * CFQ_WEIGHT_DEFAULT
;
1580 cgd
->weight
= CFQ_WEIGHT_DEFAULT
;
1581 cgd
->leaf_weight
= CFQ_WEIGHT_DEFAULT
;
1585 static void cfq_pd_init(struct blkcg_gq
*blkg
)
1587 struct cfq_group
*cfqg
= blkg_to_cfqg(blkg
);
1588 struct cfq_group_data
*cgd
= blkcg_to_cfqgd(blkg
->blkcg
);
1590 cfq_init_cfqg_base(cfqg
);
1591 cfqg
->weight
= cgd
->weight
;
1592 cfqg
->leaf_weight
= cgd
->leaf_weight
;
1593 cfqg_stats_init(&cfqg
->stats
);
1594 cfqg_stats_init(&cfqg
->dead_stats
);
1597 static void cfq_pd_offline(struct blkcg_gq
*blkg
)
1600 * @blkg is going offline and will be ignored by
1601 * blkg_[rw]stat_recursive_sum(). Transfer stats to the parent so
1602 * that they don't get lost. If IOs complete after this point, the
1603 * stats for them will be lost. Oh well...
1605 cfqg_stats_xfer_dead(blkg_to_cfqg(blkg
));
1608 /* offset delta from cfqg->stats to cfqg->dead_stats */
1609 static const int dead_stats_off_delta
= offsetof(struct cfq_group
, dead_stats
) -
1610 offsetof(struct cfq_group
, stats
);
1612 /* to be used by recursive prfill, sums live and dead stats recursively */
1613 static u64
cfqg_stat_pd_recursive_sum(struct blkg_policy_data
*pd
, int off
)
1617 sum
+= blkg_stat_recursive_sum(pd
, off
);
1618 sum
+= blkg_stat_recursive_sum(pd
, off
+ dead_stats_off_delta
);
1622 /* to be used by recursive prfill, sums live and dead rwstats recursively */
1623 static struct blkg_rwstat
cfqg_rwstat_pd_recursive_sum(struct blkg_policy_data
*pd
,
1626 struct blkg_rwstat a
, b
;
1628 a
= blkg_rwstat_recursive_sum(pd
, off
);
1629 b
= blkg_rwstat_recursive_sum(pd
, off
+ dead_stats_off_delta
);
1630 blkg_rwstat_merge(&a
, &b
);
1634 static void cfq_pd_reset_stats(struct blkcg_gq
*blkg
)
1636 struct cfq_group
*cfqg
= blkg_to_cfqg(blkg
);
1638 cfqg_stats_reset(&cfqg
->stats
);
1639 cfqg_stats_reset(&cfqg
->dead_stats
);
1643 * Search for the cfq group current task belongs to. request_queue lock must
1646 static struct cfq_group
*cfq_lookup_create_cfqg(struct cfq_data
*cfqd
,
1647 struct blkcg
*blkcg
)
1649 struct request_queue
*q
= cfqd
->queue
;
1650 struct cfq_group
*cfqg
= NULL
;
1652 /* avoid lookup for the common case where there's no blkcg */
1653 if (blkcg
== &blkcg_root
) {
1654 cfqg
= cfqd
->root_group
;
1656 struct blkcg_gq
*blkg
;
1658 blkg
= blkg_lookup_create(blkcg
, q
);
1660 cfqg
= blkg_to_cfqg(blkg
);
1666 static void cfq_link_cfqq_cfqg(struct cfq_queue
*cfqq
, struct cfq_group
*cfqg
)
1668 /* Currently, all async queues are mapped to root group */
1669 if (!cfq_cfqq_sync(cfqq
))
1670 cfqg
= cfqq
->cfqd
->root_group
;
1673 /* cfqq reference on cfqg */
1677 static u64
cfqg_prfill_weight_device(struct seq_file
*sf
,
1678 struct blkg_policy_data
*pd
, int off
)
1680 struct cfq_group
*cfqg
= pd_to_cfqg(pd
);
1682 if (!cfqg
->dev_weight
)
1684 return __blkg_prfill_u64(sf
, pd
, cfqg
->dev_weight
);
1687 static int cfqg_print_weight_device(struct seq_file
*sf
, void *v
)
1689 blkcg_print_blkgs(sf
, css_to_blkcg(seq_css(sf
)),
1690 cfqg_prfill_weight_device
, &blkcg_policy_cfq
,
1695 static u64
cfqg_prfill_leaf_weight_device(struct seq_file
*sf
,
1696 struct blkg_policy_data
*pd
, int off
)
1698 struct cfq_group
*cfqg
= pd_to_cfqg(pd
);
1700 if (!cfqg
->dev_leaf_weight
)
1702 return __blkg_prfill_u64(sf
, pd
, cfqg
->dev_leaf_weight
);
1705 static int cfqg_print_leaf_weight_device(struct seq_file
*sf
, void *v
)
1707 blkcg_print_blkgs(sf
, css_to_blkcg(seq_css(sf
)),
1708 cfqg_prfill_leaf_weight_device
, &blkcg_policy_cfq
,
1713 static int cfq_print_weight(struct seq_file
*sf
, void *v
)
1715 struct blkcg
*blkcg
= css_to_blkcg(seq_css(sf
));
1716 struct cfq_group_data
*cgd
= blkcg_to_cfqgd(blkcg
);
1717 unsigned int val
= 0;
1722 seq_printf(sf
, "%u\n", val
);
1726 static int cfq_print_leaf_weight(struct seq_file
*sf
, void *v
)
1728 struct blkcg
*blkcg
= css_to_blkcg(seq_css(sf
));
1729 struct cfq_group_data
*cgd
= blkcg_to_cfqgd(blkcg
);
1730 unsigned int val
= 0;
1733 val
= cgd
->leaf_weight
;
1735 seq_printf(sf
, "%u\n", val
);
1739 static ssize_t
__cfqg_set_weight_device(struct kernfs_open_file
*of
,
1740 char *buf
, size_t nbytes
, loff_t off
,
1741 bool is_leaf_weight
)
1743 struct blkcg
*blkcg
= css_to_blkcg(of_css(of
));
1744 struct blkg_conf_ctx ctx
;
1745 struct cfq_group
*cfqg
;
1746 struct cfq_group_data
*cfqgd
;
1749 ret
= blkg_conf_prep(blkcg
, &blkcg_policy_cfq
, buf
, &ctx
);
1754 cfqg
= blkg_to_cfqg(ctx
.blkg
);
1755 cfqgd
= blkcg_to_cfqgd(blkcg
);
1756 if (!cfqg
|| !cfqgd
)
1759 if (!ctx
.v
|| (ctx
.v
>= CFQ_WEIGHT_MIN
&& ctx
.v
<= CFQ_WEIGHT_MAX
)) {
1760 if (!is_leaf_weight
) {
1761 cfqg
->dev_weight
= ctx
.v
;
1762 cfqg
->new_weight
= ctx
.v
?: cfqgd
->weight
;
1764 cfqg
->dev_leaf_weight
= ctx
.v
;
1765 cfqg
->new_leaf_weight
= ctx
.v
?: cfqgd
->leaf_weight
;
1771 blkg_conf_finish(&ctx
);
1772 return ret
?: nbytes
;
1775 static ssize_t
cfqg_set_weight_device(struct kernfs_open_file
*of
,
1776 char *buf
, size_t nbytes
, loff_t off
)
1778 return __cfqg_set_weight_device(of
, buf
, nbytes
, off
, false);
1781 static ssize_t
cfqg_set_leaf_weight_device(struct kernfs_open_file
*of
,
1782 char *buf
, size_t nbytes
, loff_t off
)
1784 return __cfqg_set_weight_device(of
, buf
, nbytes
, off
, true);
1787 static int __cfq_set_weight(struct cgroup_subsys_state
*css
, struct cftype
*cft
,
1788 u64 val
, bool is_leaf_weight
)
1790 struct blkcg
*blkcg
= css_to_blkcg(css
);
1791 struct blkcg_gq
*blkg
;
1792 struct cfq_group_data
*cfqgd
;
1795 if (val
< CFQ_WEIGHT_MIN
|| val
> CFQ_WEIGHT_MAX
)
1798 spin_lock_irq(&blkcg
->lock
);
1799 cfqgd
= blkcg_to_cfqgd(blkcg
);
1805 if (!is_leaf_weight
)
1806 cfqgd
->weight
= val
;
1808 cfqgd
->leaf_weight
= val
;
1810 hlist_for_each_entry(blkg
, &blkcg
->blkg_list
, blkcg_node
) {
1811 struct cfq_group
*cfqg
= blkg_to_cfqg(blkg
);
1816 if (!is_leaf_weight
) {
1817 if (!cfqg
->dev_weight
)
1818 cfqg
->new_weight
= cfqgd
->weight
;
1820 if (!cfqg
->dev_leaf_weight
)
1821 cfqg
->new_leaf_weight
= cfqgd
->leaf_weight
;
1826 spin_unlock_irq(&blkcg
->lock
);
1830 static int cfq_set_weight(struct cgroup_subsys_state
*css
, struct cftype
*cft
,
1833 return __cfq_set_weight(css
, cft
, val
, false);
1836 static int cfq_set_leaf_weight(struct cgroup_subsys_state
*css
,
1837 struct cftype
*cft
, u64 val
)
1839 return __cfq_set_weight(css
, cft
, val
, true);
1842 static int cfqg_print_stat(struct seq_file
*sf
, void *v
)
1844 blkcg_print_blkgs(sf
, css_to_blkcg(seq_css(sf
)), blkg_prfill_stat
,
1845 &blkcg_policy_cfq
, seq_cft(sf
)->private, false);
1849 static int cfqg_print_rwstat(struct seq_file
*sf
, void *v
)
1851 blkcg_print_blkgs(sf
, css_to_blkcg(seq_css(sf
)), blkg_prfill_rwstat
,
1852 &blkcg_policy_cfq
, seq_cft(sf
)->private, true);
1856 static u64
cfqg_prfill_stat_recursive(struct seq_file
*sf
,
1857 struct blkg_policy_data
*pd
, int off
)
1859 u64 sum
= cfqg_stat_pd_recursive_sum(pd
, off
);
1861 return __blkg_prfill_u64(sf
, pd
, sum
);
1864 static u64
cfqg_prfill_rwstat_recursive(struct seq_file
*sf
,
1865 struct blkg_policy_data
*pd
, int off
)
1867 struct blkg_rwstat sum
= cfqg_rwstat_pd_recursive_sum(pd
, off
);
1869 return __blkg_prfill_rwstat(sf
, pd
, &sum
);
1872 static int cfqg_print_stat_recursive(struct seq_file
*sf
, void *v
)
1874 blkcg_print_blkgs(sf
, css_to_blkcg(seq_css(sf
)),
1875 cfqg_prfill_stat_recursive
, &blkcg_policy_cfq
,
1876 seq_cft(sf
)->private, false);
1880 static int cfqg_print_rwstat_recursive(struct seq_file
*sf
, void *v
)
1882 blkcg_print_blkgs(sf
, css_to_blkcg(seq_css(sf
)),
1883 cfqg_prfill_rwstat_recursive
, &blkcg_policy_cfq
,
1884 seq_cft(sf
)->private, true);
1888 #ifdef CONFIG_DEBUG_BLK_CGROUP
1889 static u64
cfqg_prfill_avg_queue_size(struct seq_file
*sf
,
1890 struct blkg_policy_data
*pd
, int off
)
1892 struct cfq_group
*cfqg
= pd_to_cfqg(pd
);
1893 u64 samples
= blkg_stat_read(&cfqg
->stats
.avg_queue_size_samples
);
1897 v
= blkg_stat_read(&cfqg
->stats
.avg_queue_size_sum
);
1898 v
= div64_u64(v
, samples
);
1900 __blkg_prfill_u64(sf
, pd
, v
);
1904 /* print avg_queue_size */
1905 static int cfqg_print_avg_queue_size(struct seq_file
*sf
, void *v
)
1907 blkcg_print_blkgs(sf
, css_to_blkcg(seq_css(sf
)),
1908 cfqg_prfill_avg_queue_size
, &blkcg_policy_cfq
,
1912 #endif /* CONFIG_DEBUG_BLK_CGROUP */
1914 static struct cftype cfq_blkcg_files
[] = {
1915 /* on root, weight is mapped to leaf_weight */
1917 .name
= "weight_device",
1918 .flags
= CFTYPE_ONLY_ON_ROOT
,
1919 .seq_show
= cfqg_print_leaf_weight_device
,
1920 .write
= cfqg_set_leaf_weight_device
,
1924 .flags
= CFTYPE_ONLY_ON_ROOT
,
1925 .seq_show
= cfq_print_leaf_weight
,
1926 .write_u64
= cfq_set_leaf_weight
,
1929 /* no such mapping necessary for !roots */
1931 .name
= "weight_device",
1932 .flags
= CFTYPE_NOT_ON_ROOT
,
1933 .seq_show
= cfqg_print_weight_device
,
1934 .write
= cfqg_set_weight_device
,
1938 .flags
= CFTYPE_NOT_ON_ROOT
,
1939 .seq_show
= cfq_print_weight
,
1940 .write_u64
= cfq_set_weight
,
1944 .name
= "leaf_weight_device",
1945 .seq_show
= cfqg_print_leaf_weight_device
,
1946 .write
= cfqg_set_leaf_weight_device
,
1949 .name
= "leaf_weight",
1950 .seq_show
= cfq_print_leaf_weight
,
1951 .write_u64
= cfq_set_leaf_weight
,
1954 /* statistics, covers only the tasks in the cfqg */
1957 .private = offsetof(struct cfq_group
, stats
.time
),
1958 .seq_show
= cfqg_print_stat
,
1962 .private = offsetof(struct cfq_group
, stats
.sectors
),
1963 .seq_show
= cfqg_print_stat
,
1966 .name
= "io_service_bytes",
1967 .private = offsetof(struct cfq_group
, stats
.service_bytes
),
1968 .seq_show
= cfqg_print_rwstat
,
1971 .name
= "io_serviced",
1972 .private = offsetof(struct cfq_group
, stats
.serviced
),
1973 .seq_show
= cfqg_print_rwstat
,
1976 .name
= "io_service_time",
1977 .private = offsetof(struct cfq_group
, stats
.service_time
),
1978 .seq_show
= cfqg_print_rwstat
,
1981 .name
= "io_wait_time",
1982 .private = offsetof(struct cfq_group
, stats
.wait_time
),
1983 .seq_show
= cfqg_print_rwstat
,
1986 .name
= "io_merged",
1987 .private = offsetof(struct cfq_group
, stats
.merged
),
1988 .seq_show
= cfqg_print_rwstat
,
1991 .name
= "io_queued",
1992 .private = offsetof(struct cfq_group
, stats
.queued
),
1993 .seq_show
= cfqg_print_rwstat
,
1996 /* the same statictics which cover the cfqg and its descendants */
1998 .name
= "time_recursive",
1999 .private = offsetof(struct cfq_group
, stats
.time
),
2000 .seq_show
= cfqg_print_stat_recursive
,
2003 .name
= "sectors_recursive",
2004 .private = offsetof(struct cfq_group
, stats
.sectors
),
2005 .seq_show
= cfqg_print_stat_recursive
,
2008 .name
= "io_service_bytes_recursive",
2009 .private = offsetof(struct cfq_group
, stats
.service_bytes
),
2010 .seq_show
= cfqg_print_rwstat_recursive
,
2013 .name
= "io_serviced_recursive",
2014 .private = offsetof(struct cfq_group
, stats
.serviced
),
2015 .seq_show
= cfqg_print_rwstat_recursive
,
2018 .name
= "io_service_time_recursive",
2019 .private = offsetof(struct cfq_group
, stats
.service_time
),
2020 .seq_show
= cfqg_print_rwstat_recursive
,
2023 .name
= "io_wait_time_recursive",
2024 .private = offsetof(struct cfq_group
, stats
.wait_time
),
2025 .seq_show
= cfqg_print_rwstat_recursive
,
2028 .name
= "io_merged_recursive",
2029 .private = offsetof(struct cfq_group
, stats
.merged
),
2030 .seq_show
= cfqg_print_rwstat_recursive
,
2033 .name
= "io_queued_recursive",
2034 .private = offsetof(struct cfq_group
, stats
.queued
),
2035 .seq_show
= cfqg_print_rwstat_recursive
,
2037 #ifdef CONFIG_DEBUG_BLK_CGROUP
2039 .name
= "avg_queue_size",
2040 .seq_show
= cfqg_print_avg_queue_size
,
2043 .name
= "group_wait_time",
2044 .private = offsetof(struct cfq_group
, stats
.group_wait_time
),
2045 .seq_show
= cfqg_print_stat
,
2048 .name
= "idle_time",
2049 .private = offsetof(struct cfq_group
, stats
.idle_time
),
2050 .seq_show
= cfqg_print_stat
,
2053 .name
= "empty_time",
2054 .private = offsetof(struct cfq_group
, stats
.empty_time
),
2055 .seq_show
= cfqg_print_stat
,
2059 .private = offsetof(struct cfq_group
, stats
.dequeue
),
2060 .seq_show
= cfqg_print_stat
,
2063 .name
= "unaccounted_time",
2064 .private = offsetof(struct cfq_group
, stats
.unaccounted_time
),
2065 .seq_show
= cfqg_print_stat
,
2067 #endif /* CONFIG_DEBUG_BLK_CGROUP */
2070 #else /* GROUP_IOSCHED */
2071 static struct cfq_group
*cfq_lookup_create_cfqg(struct cfq_data
*cfqd
,
2072 struct blkcg
*blkcg
)
2074 return cfqd
->root_group
;
2078 cfq_link_cfqq_cfqg(struct cfq_queue
*cfqq
, struct cfq_group
*cfqg
) {
2082 #endif /* GROUP_IOSCHED */
2085 * The cfqd->service_trees holds all pending cfq_queue's that have
2086 * requests waiting to be processed. It is sorted in the order that
2087 * we will service the queues.
2089 static void cfq_service_tree_add(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
2092 struct rb_node
**p
, *parent
;
2093 struct cfq_queue
*__cfqq
;
2094 unsigned long rb_key
;
2095 struct cfq_rb_root
*st
;
2099 st
= st_for(cfqq
->cfqg
, cfqq_class(cfqq
), cfqq_type(cfqq
));
2100 if (cfq_class_idle(cfqq
)) {
2101 rb_key
= CFQ_IDLE_DELAY
;
2102 parent
= rb_last(&st
->rb
);
2103 if (parent
&& parent
!= &cfqq
->rb_node
) {
2104 __cfqq
= rb_entry(parent
, struct cfq_queue
, rb_node
);
2105 rb_key
+= __cfqq
->rb_key
;
2108 } else if (!add_front
) {
2110 * Get our rb key offset. Subtract any residual slice
2111 * value carried from last service. A negative resid
2112 * count indicates slice overrun, and this should position
2113 * the next service time further away in the tree.
2115 rb_key
= cfq_slice_offset(cfqd
, cfqq
) + jiffies
;
2116 rb_key
-= cfqq
->slice_resid
;
2117 cfqq
->slice_resid
= 0;
2120 __cfqq
= cfq_rb_first(st
);
2121 rb_key
+= __cfqq
? __cfqq
->rb_key
: jiffies
;
2124 if (!RB_EMPTY_NODE(&cfqq
->rb_node
)) {
2127 * same position, nothing more to do
2129 if (rb_key
== cfqq
->rb_key
&& cfqq
->service_tree
== st
)
2132 cfq_rb_erase(&cfqq
->rb_node
, cfqq
->service_tree
);
2133 cfqq
->service_tree
= NULL
;
2138 cfqq
->service_tree
= st
;
2139 p
= &st
->rb
.rb_node
;
2142 __cfqq
= rb_entry(parent
, struct cfq_queue
, rb_node
);
2145 * sort by key, that represents service time.
2147 if (time_before(rb_key
, __cfqq
->rb_key
))
2148 p
= &parent
->rb_left
;
2150 p
= &parent
->rb_right
;
2156 st
->left
= &cfqq
->rb_node
;
2158 cfqq
->rb_key
= rb_key
;
2159 rb_link_node(&cfqq
->rb_node
, parent
, p
);
2160 rb_insert_color(&cfqq
->rb_node
, &st
->rb
);
2162 if (add_front
|| !new_cfqq
)
2164 cfq_group_notify_queue_add(cfqd
, cfqq
->cfqg
);
2167 static struct cfq_queue
*
2168 cfq_prio_tree_lookup(struct cfq_data
*cfqd
, struct rb_root
*root
,
2169 sector_t sector
, struct rb_node
**ret_parent
,
2170 struct rb_node
***rb_link
)
2172 struct rb_node
**p
, *parent
;
2173 struct cfq_queue
*cfqq
= NULL
;
2181 cfqq
= rb_entry(parent
, struct cfq_queue
, p_node
);
2184 * Sort strictly based on sector. Smallest to the left,
2185 * largest to the right.
2187 if (sector
> blk_rq_pos(cfqq
->next_rq
))
2188 n
= &(*p
)->rb_right
;
2189 else if (sector
< blk_rq_pos(cfqq
->next_rq
))
2197 *ret_parent
= parent
;
2203 static void cfq_prio_tree_add(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
2205 struct rb_node
**p
, *parent
;
2206 struct cfq_queue
*__cfqq
;
2209 rb_erase(&cfqq
->p_node
, cfqq
->p_root
);
2210 cfqq
->p_root
= NULL
;
2213 if (cfq_class_idle(cfqq
))
2218 cfqq
->p_root
= &cfqd
->prio_trees
[cfqq
->org_ioprio
];
2219 __cfqq
= cfq_prio_tree_lookup(cfqd
, cfqq
->p_root
,
2220 blk_rq_pos(cfqq
->next_rq
), &parent
, &p
);
2222 rb_link_node(&cfqq
->p_node
, parent
, p
);
2223 rb_insert_color(&cfqq
->p_node
, cfqq
->p_root
);
2225 cfqq
->p_root
= NULL
;
2229 * Update cfqq's position in the service tree.
2231 static void cfq_resort_rr_list(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
2234 * Resorting requires the cfqq to be on the RR list already.
2236 if (cfq_cfqq_on_rr(cfqq
)) {
2237 cfq_service_tree_add(cfqd
, cfqq
, 0);
2238 cfq_prio_tree_add(cfqd
, cfqq
);
2243 * add to busy list of queues for service, trying to be fair in ordering
2244 * the pending list according to last request service
2246 static void cfq_add_cfqq_rr(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
2248 cfq_log_cfqq(cfqd
, cfqq
, "add_to_rr");
2249 BUG_ON(cfq_cfqq_on_rr(cfqq
));
2250 cfq_mark_cfqq_on_rr(cfqq
);
2251 cfqd
->busy_queues
++;
2252 if (cfq_cfqq_sync(cfqq
))
2253 cfqd
->busy_sync_queues
++;
2255 cfq_resort_rr_list(cfqd
, cfqq
);
2259 * Called when the cfqq no longer has requests pending, remove it from
2262 static void cfq_del_cfqq_rr(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
2264 cfq_log_cfqq(cfqd
, cfqq
, "del_from_rr");
2265 BUG_ON(!cfq_cfqq_on_rr(cfqq
));
2266 cfq_clear_cfqq_on_rr(cfqq
);
2268 if (!RB_EMPTY_NODE(&cfqq
->rb_node
)) {
2269 cfq_rb_erase(&cfqq
->rb_node
, cfqq
->service_tree
);
2270 cfqq
->service_tree
= NULL
;
2273 rb_erase(&cfqq
->p_node
, cfqq
->p_root
);
2274 cfqq
->p_root
= NULL
;
2277 cfq_group_notify_queue_del(cfqd
, cfqq
->cfqg
);
2278 BUG_ON(!cfqd
->busy_queues
);
2279 cfqd
->busy_queues
--;
2280 if (cfq_cfqq_sync(cfqq
))
2281 cfqd
->busy_sync_queues
--;
2285 * rb tree support functions
2287 static void cfq_del_rq_rb(struct request
*rq
)
2289 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
2290 const int sync
= rq_is_sync(rq
);
2292 BUG_ON(!cfqq
->queued
[sync
]);
2293 cfqq
->queued
[sync
]--;
2295 elv_rb_del(&cfqq
->sort_list
, rq
);
2297 if (cfq_cfqq_on_rr(cfqq
) && RB_EMPTY_ROOT(&cfqq
->sort_list
)) {
2299 * Queue will be deleted from service tree when we actually
2300 * expire it later. Right now just remove it from prio tree
2304 rb_erase(&cfqq
->p_node
, cfqq
->p_root
);
2305 cfqq
->p_root
= NULL
;
2310 static void cfq_add_rq_rb(struct request
*rq
)
2312 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
2313 struct cfq_data
*cfqd
= cfqq
->cfqd
;
2314 struct request
*prev
;
2316 cfqq
->queued
[rq_is_sync(rq
)]++;
2318 elv_rb_add(&cfqq
->sort_list
, rq
);
2320 if (!cfq_cfqq_on_rr(cfqq
))
2321 cfq_add_cfqq_rr(cfqd
, cfqq
);
2324 * check if this request is a better next-serve candidate
2326 prev
= cfqq
->next_rq
;
2327 cfqq
->next_rq
= cfq_choose_req(cfqd
, cfqq
->next_rq
, rq
, cfqd
->last_position
);
2330 * adjust priority tree position, if ->next_rq changes
2332 if (prev
!= cfqq
->next_rq
)
2333 cfq_prio_tree_add(cfqd
, cfqq
);
2335 BUG_ON(!cfqq
->next_rq
);
2338 static void cfq_reposition_rq_rb(struct cfq_queue
*cfqq
, struct request
*rq
)
2340 elv_rb_del(&cfqq
->sort_list
, rq
);
2341 cfqq
->queued
[rq_is_sync(rq
)]--;
2342 cfqg_stats_update_io_remove(RQ_CFQG(rq
), rq
->cmd_flags
);
2344 cfqg_stats_update_io_add(RQ_CFQG(rq
), cfqq
->cfqd
->serving_group
,
2348 static struct request
*
2349 cfq_find_rq_fmerge(struct cfq_data
*cfqd
, struct bio
*bio
)
2351 struct task_struct
*tsk
= current
;
2352 struct cfq_io_cq
*cic
;
2353 struct cfq_queue
*cfqq
;
2355 cic
= cfq_cic_lookup(cfqd
, tsk
->io_context
);
2359 cfqq
= cic_to_cfqq(cic
, cfq_bio_sync(bio
));
2361 return elv_rb_find(&cfqq
->sort_list
, bio_end_sector(bio
));
2366 static void cfq_activate_request(struct request_queue
*q
, struct request
*rq
)
2368 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
2370 cfqd
->rq_in_driver
++;
2371 cfq_log_cfqq(cfqd
, RQ_CFQQ(rq
), "activate rq, drv=%d",
2372 cfqd
->rq_in_driver
);
2374 cfqd
->last_position
= blk_rq_pos(rq
) + blk_rq_sectors(rq
);
2377 static void cfq_deactivate_request(struct request_queue
*q
, struct request
*rq
)
2379 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
2381 WARN_ON(!cfqd
->rq_in_driver
);
2382 cfqd
->rq_in_driver
--;
2383 cfq_log_cfqq(cfqd
, RQ_CFQQ(rq
), "deactivate rq, drv=%d",
2384 cfqd
->rq_in_driver
);
2387 static void cfq_remove_request(struct request
*rq
)
2389 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
2391 if (cfqq
->next_rq
== rq
)
2392 cfqq
->next_rq
= cfq_find_next_rq(cfqq
->cfqd
, cfqq
, rq
);
2394 list_del_init(&rq
->queuelist
);
2397 cfqq
->cfqd
->rq_queued
--;
2398 cfqg_stats_update_io_remove(RQ_CFQG(rq
), rq
->cmd_flags
);
2399 if (rq
->cmd_flags
& REQ_PRIO
) {
2400 WARN_ON(!cfqq
->prio_pending
);
2401 cfqq
->prio_pending
--;
2405 static int cfq_merge(struct request_queue
*q
, struct request
**req
,
2408 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
2409 struct request
*__rq
;
2411 __rq
= cfq_find_rq_fmerge(cfqd
, bio
);
2412 if (__rq
&& elv_rq_merge_ok(__rq
, bio
)) {
2414 return ELEVATOR_FRONT_MERGE
;
2417 return ELEVATOR_NO_MERGE
;
2420 static void cfq_merged_request(struct request_queue
*q
, struct request
*req
,
2423 if (type
== ELEVATOR_FRONT_MERGE
) {
2424 struct cfq_queue
*cfqq
= RQ_CFQQ(req
);
2426 cfq_reposition_rq_rb(cfqq
, req
);
2430 static void cfq_bio_merged(struct request_queue
*q
, struct request
*req
,
2433 cfqg_stats_update_io_merged(RQ_CFQG(req
), bio
->bi_rw
);
2437 cfq_merged_requests(struct request_queue
*q
, struct request
*rq
,
2438 struct request
*next
)
2440 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
2441 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
2444 * reposition in fifo if next is older than rq
2446 if (!list_empty(&rq
->queuelist
) && !list_empty(&next
->queuelist
) &&
2447 time_before(next
->fifo_time
, rq
->fifo_time
) &&
2448 cfqq
== RQ_CFQQ(next
)) {
2449 list_move(&rq
->queuelist
, &next
->queuelist
);
2450 rq
->fifo_time
= next
->fifo_time
;
2453 if (cfqq
->next_rq
== next
)
2455 cfq_remove_request(next
);
2456 cfqg_stats_update_io_merged(RQ_CFQG(rq
), next
->cmd_flags
);
2458 cfqq
= RQ_CFQQ(next
);
2460 * all requests of this queue are merged to other queues, delete it
2461 * from the service tree. If it's the active_queue,
2462 * cfq_dispatch_requests() will choose to expire it or do idle
2464 if (cfq_cfqq_on_rr(cfqq
) && RB_EMPTY_ROOT(&cfqq
->sort_list
) &&
2465 cfqq
!= cfqd
->active_queue
)
2466 cfq_del_cfqq_rr(cfqd
, cfqq
);
2469 static int cfq_allow_merge(struct request_queue
*q
, struct request
*rq
,
2472 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
2473 struct cfq_io_cq
*cic
;
2474 struct cfq_queue
*cfqq
;
2477 * Disallow merge of a sync bio into an async request.
2479 if (cfq_bio_sync(bio
) && !rq_is_sync(rq
))
2483 * Lookup the cfqq that this bio will be queued with and allow
2484 * merge only if rq is queued there.
2486 cic
= cfq_cic_lookup(cfqd
, current
->io_context
);
2490 cfqq
= cic_to_cfqq(cic
, cfq_bio_sync(bio
));
2491 return cfqq
== RQ_CFQQ(rq
);
2494 static inline void cfq_del_timer(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
2496 del_timer(&cfqd
->idle_slice_timer
);
2497 cfqg_stats_update_idle_time(cfqq
->cfqg
);
2500 static void __cfq_set_active_queue(struct cfq_data
*cfqd
,
2501 struct cfq_queue
*cfqq
)
2504 cfq_log_cfqq(cfqd
, cfqq
, "set_active wl_class:%d wl_type:%d",
2505 cfqd
->serving_wl_class
, cfqd
->serving_wl_type
);
2506 cfqg_stats_update_avg_queue_size(cfqq
->cfqg
);
2507 cfqq
->slice_start
= 0;
2508 cfqq
->dispatch_start
= jiffies
;
2509 cfqq
->allocated_slice
= 0;
2510 cfqq
->slice_end
= 0;
2511 cfqq
->slice_dispatch
= 0;
2512 cfqq
->nr_sectors
= 0;
2514 cfq_clear_cfqq_wait_request(cfqq
);
2515 cfq_clear_cfqq_must_dispatch(cfqq
);
2516 cfq_clear_cfqq_must_alloc_slice(cfqq
);
2517 cfq_clear_cfqq_fifo_expire(cfqq
);
2518 cfq_mark_cfqq_slice_new(cfqq
);
2520 cfq_del_timer(cfqd
, cfqq
);
2523 cfqd
->active_queue
= cfqq
;
2527 * current cfqq expired its slice (or was too idle), select new one
2530 __cfq_slice_expired(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
2533 cfq_log_cfqq(cfqd
, cfqq
, "slice expired t=%d", timed_out
);
2535 if (cfq_cfqq_wait_request(cfqq
))
2536 cfq_del_timer(cfqd
, cfqq
);
2538 cfq_clear_cfqq_wait_request(cfqq
);
2539 cfq_clear_cfqq_wait_busy(cfqq
);
2542 * If this cfqq is shared between multiple processes, check to
2543 * make sure that those processes are still issuing I/Os within
2544 * the mean seek distance. If not, it may be time to break the
2545 * queues apart again.
2547 if (cfq_cfqq_coop(cfqq
) && CFQQ_SEEKY(cfqq
))
2548 cfq_mark_cfqq_split_coop(cfqq
);
2551 * store what was left of this slice, if the queue idled/timed out
2554 if (cfq_cfqq_slice_new(cfqq
))
2555 cfqq
->slice_resid
= cfq_scaled_cfqq_slice(cfqd
, cfqq
);
2557 cfqq
->slice_resid
= cfqq
->slice_end
- jiffies
;
2558 cfq_log_cfqq(cfqd
, cfqq
, "resid=%ld", cfqq
->slice_resid
);
2561 cfq_group_served(cfqd
, cfqq
->cfqg
, cfqq
);
2563 if (cfq_cfqq_on_rr(cfqq
) && RB_EMPTY_ROOT(&cfqq
->sort_list
))
2564 cfq_del_cfqq_rr(cfqd
, cfqq
);
2566 cfq_resort_rr_list(cfqd
, cfqq
);
2568 if (cfqq
== cfqd
->active_queue
)
2569 cfqd
->active_queue
= NULL
;
2571 if (cfqd
->active_cic
) {
2572 put_io_context(cfqd
->active_cic
->icq
.ioc
);
2573 cfqd
->active_cic
= NULL
;
2577 static inline void cfq_slice_expired(struct cfq_data
*cfqd
, bool timed_out
)
2579 struct cfq_queue
*cfqq
= cfqd
->active_queue
;
2582 __cfq_slice_expired(cfqd
, cfqq
, timed_out
);
2586 * Get next queue for service. Unless we have a queue preemption,
2587 * we'll simply select the first cfqq in the service tree.
2589 static struct cfq_queue
*cfq_get_next_queue(struct cfq_data
*cfqd
)
2591 struct cfq_rb_root
*st
= st_for(cfqd
->serving_group
,
2592 cfqd
->serving_wl_class
, cfqd
->serving_wl_type
);
2594 if (!cfqd
->rq_queued
)
2597 /* There is nothing to dispatch */
2600 if (RB_EMPTY_ROOT(&st
->rb
))
2602 return cfq_rb_first(st
);
2605 static struct cfq_queue
*cfq_get_next_queue_forced(struct cfq_data
*cfqd
)
2607 struct cfq_group
*cfqg
;
2608 struct cfq_queue
*cfqq
;
2610 struct cfq_rb_root
*st
;
2612 if (!cfqd
->rq_queued
)
2615 cfqg
= cfq_get_next_cfqg(cfqd
);
2619 for_each_cfqg_st(cfqg
, i
, j
, st
)
2620 if ((cfqq
= cfq_rb_first(st
)) != NULL
)
2626 * Get and set a new active queue for service.
2628 static struct cfq_queue
*cfq_set_active_queue(struct cfq_data
*cfqd
,
2629 struct cfq_queue
*cfqq
)
2632 cfqq
= cfq_get_next_queue(cfqd
);
2634 __cfq_set_active_queue(cfqd
, cfqq
);
2638 static inline sector_t
cfq_dist_from_last(struct cfq_data
*cfqd
,
2641 if (blk_rq_pos(rq
) >= cfqd
->last_position
)
2642 return blk_rq_pos(rq
) - cfqd
->last_position
;
2644 return cfqd
->last_position
- blk_rq_pos(rq
);
2647 static inline int cfq_rq_close(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
2650 return cfq_dist_from_last(cfqd
, rq
) <= CFQQ_CLOSE_THR
;
2653 static struct cfq_queue
*cfqq_close(struct cfq_data
*cfqd
,
2654 struct cfq_queue
*cur_cfqq
)
2656 struct rb_root
*root
= &cfqd
->prio_trees
[cur_cfqq
->org_ioprio
];
2657 struct rb_node
*parent
, *node
;
2658 struct cfq_queue
*__cfqq
;
2659 sector_t sector
= cfqd
->last_position
;
2661 if (RB_EMPTY_ROOT(root
))
2665 * First, if we find a request starting at the end of the last
2666 * request, choose it.
2668 __cfqq
= cfq_prio_tree_lookup(cfqd
, root
, sector
, &parent
, NULL
);
2673 * If the exact sector wasn't found, the parent of the NULL leaf
2674 * will contain the closest sector.
2676 __cfqq
= rb_entry(parent
, struct cfq_queue
, p_node
);
2677 if (cfq_rq_close(cfqd
, cur_cfqq
, __cfqq
->next_rq
))
2680 if (blk_rq_pos(__cfqq
->next_rq
) < sector
)
2681 node
= rb_next(&__cfqq
->p_node
);
2683 node
= rb_prev(&__cfqq
->p_node
);
2687 __cfqq
= rb_entry(node
, struct cfq_queue
, p_node
);
2688 if (cfq_rq_close(cfqd
, cur_cfqq
, __cfqq
->next_rq
))
2696 * cur_cfqq - passed in so that we don't decide that the current queue is
2697 * closely cooperating with itself.
2699 * So, basically we're assuming that that cur_cfqq has dispatched at least
2700 * one request, and that cfqd->last_position reflects a position on the disk
2701 * associated with the I/O issued by cur_cfqq. I'm not sure this is a valid
2704 static struct cfq_queue
*cfq_close_cooperator(struct cfq_data
*cfqd
,
2705 struct cfq_queue
*cur_cfqq
)
2707 struct cfq_queue
*cfqq
;
2709 if (cfq_class_idle(cur_cfqq
))
2711 if (!cfq_cfqq_sync(cur_cfqq
))
2713 if (CFQQ_SEEKY(cur_cfqq
))
2717 * Don't search priority tree if it's the only queue in the group.
2719 if (cur_cfqq
->cfqg
->nr_cfqq
== 1)
2723 * We should notice if some of the queues are cooperating, eg
2724 * working closely on the same area of the disk. In that case,
2725 * we can group them together and don't waste time idling.
2727 cfqq
= cfqq_close(cfqd
, cur_cfqq
);
2731 /* If new queue belongs to different cfq_group, don't choose it */
2732 if (cur_cfqq
->cfqg
!= cfqq
->cfqg
)
2736 * It only makes sense to merge sync queues.
2738 if (!cfq_cfqq_sync(cfqq
))
2740 if (CFQQ_SEEKY(cfqq
))
2744 * Do not merge queues of different priority classes
2746 if (cfq_class_rt(cfqq
) != cfq_class_rt(cur_cfqq
))
2753 * Determine whether we should enforce idle window for this queue.
2756 static bool cfq_should_idle(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
2758 enum wl_class_t wl_class
= cfqq_class(cfqq
);
2759 struct cfq_rb_root
*st
= cfqq
->service_tree
;
2764 if (!cfqd
->cfq_slice_idle
)
2767 /* We never do for idle class queues. */
2768 if (wl_class
== IDLE_WORKLOAD
)
2771 /* We do for queues that were marked with idle window flag. */
2772 if (cfq_cfqq_idle_window(cfqq
) &&
2773 !(blk_queue_nonrot(cfqd
->queue
) && cfqd
->hw_tag
))
2777 * Otherwise, we do only if they are the last ones
2778 * in their service tree.
2780 if (st
->count
== 1 && cfq_cfqq_sync(cfqq
) &&
2781 !cfq_io_thinktime_big(cfqd
, &st
->ttime
, false))
2783 cfq_log_cfqq(cfqd
, cfqq
, "Not idling. st->count:%d", st
->count
);
2787 static void cfq_arm_slice_timer(struct cfq_data
*cfqd
)
2789 struct cfq_queue
*cfqq
= cfqd
->active_queue
;
2790 struct cfq_io_cq
*cic
;
2791 unsigned long sl
, group_idle
= 0;
2794 * SSD device without seek penalty, disable idling. But only do so
2795 * for devices that support queuing, otherwise we still have a problem
2796 * with sync vs async workloads.
2798 if (blk_queue_nonrot(cfqd
->queue
) && cfqd
->hw_tag
)
2801 WARN_ON(!RB_EMPTY_ROOT(&cfqq
->sort_list
));
2802 WARN_ON(cfq_cfqq_slice_new(cfqq
));
2805 * idle is disabled, either manually or by past process history
2807 if (!cfq_should_idle(cfqd
, cfqq
)) {
2808 /* no queue idling. Check for group idling */
2809 if (cfqd
->cfq_group_idle
)
2810 group_idle
= cfqd
->cfq_group_idle
;
2816 * still active requests from this queue, don't idle
2818 if (cfqq
->dispatched
)
2822 * task has exited, don't wait
2824 cic
= cfqd
->active_cic
;
2825 if (!cic
|| !atomic_read(&cic
->icq
.ioc
->active_ref
))
2829 * If our average think time is larger than the remaining time
2830 * slice, then don't idle. This avoids overrunning the allotted
2833 if (sample_valid(cic
->ttime
.ttime_samples
) &&
2834 (cfqq
->slice_end
- jiffies
< cic
->ttime
.ttime_mean
)) {
2835 cfq_log_cfqq(cfqd
, cfqq
, "Not idling. think_time:%lu",
2836 cic
->ttime
.ttime_mean
);
2840 /* There are other queues in the group, don't do group idle */
2841 if (group_idle
&& cfqq
->cfqg
->nr_cfqq
> 1)
2844 cfq_mark_cfqq_wait_request(cfqq
);
2847 sl
= cfqd
->cfq_group_idle
;
2849 sl
= cfqd
->cfq_slice_idle
;
2851 mod_timer(&cfqd
->idle_slice_timer
, jiffies
+ sl
);
2852 cfqg_stats_set_start_idle_time(cfqq
->cfqg
);
2853 cfq_log_cfqq(cfqd
, cfqq
, "arm_idle: %lu group_idle: %d", sl
,
2854 group_idle
? 1 : 0);
2858 * Move request from internal lists to the request queue dispatch list.
2860 static void cfq_dispatch_insert(struct request_queue
*q
, struct request
*rq
)
2862 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
2863 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
2865 cfq_log_cfqq(cfqd
, cfqq
, "dispatch_insert");
2867 cfqq
->next_rq
= cfq_find_next_rq(cfqd
, cfqq
, rq
);
2868 cfq_remove_request(rq
);
2870 (RQ_CFQG(rq
))->dispatched
++;
2871 elv_dispatch_sort(q
, rq
);
2873 cfqd
->rq_in_flight
[cfq_cfqq_sync(cfqq
)]++;
2874 cfqq
->nr_sectors
+= blk_rq_sectors(rq
);
2875 cfqg_stats_update_dispatch(cfqq
->cfqg
, blk_rq_bytes(rq
), rq
->cmd_flags
);
2879 * return expired entry, or NULL to just start from scratch in rbtree
2881 static struct request
*cfq_check_fifo(struct cfq_queue
*cfqq
)
2883 struct request
*rq
= NULL
;
2885 if (cfq_cfqq_fifo_expire(cfqq
))
2888 cfq_mark_cfqq_fifo_expire(cfqq
);
2890 if (list_empty(&cfqq
->fifo
))
2893 rq
= rq_entry_fifo(cfqq
->fifo
.next
);
2894 if (time_before(jiffies
, rq
->fifo_time
))
2897 cfq_log_cfqq(cfqq
->cfqd
, cfqq
, "fifo=%p", rq
);
2902 cfq_prio_to_maxrq(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
2904 const int base_rq
= cfqd
->cfq_slice_async_rq
;
2906 WARN_ON(cfqq
->ioprio
>= IOPRIO_BE_NR
);
2908 return 2 * base_rq
* (IOPRIO_BE_NR
- cfqq
->ioprio
);
2912 * Must be called with the queue_lock held.
2914 static int cfqq_process_refs(struct cfq_queue
*cfqq
)
2916 int process_refs
, io_refs
;
2918 io_refs
= cfqq
->allocated
[READ
] + cfqq
->allocated
[WRITE
];
2919 process_refs
= cfqq
->ref
- io_refs
;
2920 BUG_ON(process_refs
< 0);
2921 return process_refs
;
2924 static void cfq_setup_merge(struct cfq_queue
*cfqq
, struct cfq_queue
*new_cfqq
)
2926 int process_refs
, new_process_refs
;
2927 struct cfq_queue
*__cfqq
;
2930 * If there are no process references on the new_cfqq, then it is
2931 * unsafe to follow the ->new_cfqq chain as other cfqq's in the
2932 * chain may have dropped their last reference (not just their
2933 * last process reference).
2935 if (!cfqq_process_refs(new_cfqq
))
2938 /* Avoid a circular list and skip interim queue merges */
2939 while ((__cfqq
= new_cfqq
->new_cfqq
)) {
2945 process_refs
= cfqq_process_refs(cfqq
);
2946 new_process_refs
= cfqq_process_refs(new_cfqq
);
2948 * If the process for the cfqq has gone away, there is no
2949 * sense in merging the queues.
2951 if (process_refs
== 0 || new_process_refs
== 0)
2955 * Merge in the direction of the lesser amount of work.
2957 if (new_process_refs
>= process_refs
) {
2958 cfqq
->new_cfqq
= new_cfqq
;
2959 new_cfqq
->ref
+= process_refs
;
2961 new_cfqq
->new_cfqq
= cfqq
;
2962 cfqq
->ref
+= new_process_refs
;
2966 static enum wl_type_t
cfq_choose_wl_type(struct cfq_data
*cfqd
,
2967 struct cfq_group
*cfqg
, enum wl_class_t wl_class
)
2969 struct cfq_queue
*queue
;
2971 bool key_valid
= false;
2972 unsigned long lowest_key
= 0;
2973 enum wl_type_t cur_best
= SYNC_NOIDLE_WORKLOAD
;
2975 for (i
= 0; i
<= SYNC_WORKLOAD
; ++i
) {
2976 /* select the one with lowest rb_key */
2977 queue
= cfq_rb_first(st_for(cfqg
, wl_class
, i
));
2979 (!key_valid
|| time_before(queue
->rb_key
, lowest_key
))) {
2980 lowest_key
= queue
->rb_key
;
2990 choose_wl_class_and_type(struct cfq_data
*cfqd
, struct cfq_group
*cfqg
)
2994 struct cfq_rb_root
*st
;
2995 unsigned group_slice
;
2996 enum wl_class_t original_class
= cfqd
->serving_wl_class
;
2998 /* Choose next priority. RT > BE > IDLE */
2999 if (cfq_group_busy_queues_wl(RT_WORKLOAD
, cfqd
, cfqg
))
3000 cfqd
->serving_wl_class
= RT_WORKLOAD
;
3001 else if (cfq_group_busy_queues_wl(BE_WORKLOAD
, cfqd
, cfqg
))
3002 cfqd
->serving_wl_class
= BE_WORKLOAD
;
3004 cfqd
->serving_wl_class
= IDLE_WORKLOAD
;
3005 cfqd
->workload_expires
= jiffies
+ 1;
3009 if (original_class
!= cfqd
->serving_wl_class
)
3013 * For RT and BE, we have to choose also the type
3014 * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
3017 st
= st_for(cfqg
, cfqd
->serving_wl_class
, cfqd
->serving_wl_type
);
3021 * check workload expiration, and that we still have other queues ready
3023 if (count
&& !time_after(jiffies
, cfqd
->workload_expires
))
3027 /* otherwise select new workload type */
3028 cfqd
->serving_wl_type
= cfq_choose_wl_type(cfqd
, cfqg
,
3029 cfqd
->serving_wl_class
);
3030 st
= st_for(cfqg
, cfqd
->serving_wl_class
, cfqd
->serving_wl_type
);
3034 * the workload slice is computed as a fraction of target latency
3035 * proportional to the number of queues in that workload, over
3036 * all the queues in the same priority class
3038 group_slice
= cfq_group_slice(cfqd
, cfqg
);
3040 slice
= group_slice
* count
/
3041 max_t(unsigned, cfqg
->busy_queues_avg
[cfqd
->serving_wl_class
],
3042 cfq_group_busy_queues_wl(cfqd
->serving_wl_class
, cfqd
,
3045 if (cfqd
->serving_wl_type
== ASYNC_WORKLOAD
) {
3049 * Async queues are currently system wide. Just taking
3050 * proportion of queues with-in same group will lead to higher
3051 * async ratio system wide as generally root group is going
3052 * to have higher weight. A more accurate thing would be to
3053 * calculate system wide asnc/sync ratio.
3055 tmp
= cfqd
->cfq_target_latency
*
3056 cfqg_busy_async_queues(cfqd
, cfqg
);
3057 tmp
= tmp
/cfqd
->busy_queues
;
3058 slice
= min_t(unsigned, slice
, tmp
);
3060 /* async workload slice is scaled down according to
3061 * the sync/async slice ratio. */
3062 slice
= slice
* cfqd
->cfq_slice
[0] / cfqd
->cfq_slice
[1];
3064 /* sync workload slice is at least 2 * cfq_slice_idle */
3065 slice
= max(slice
, 2 * cfqd
->cfq_slice_idle
);
3067 slice
= max_t(unsigned, slice
, CFQ_MIN_TT
);
3068 cfq_log(cfqd
, "workload slice:%d", slice
);
3069 cfqd
->workload_expires
= jiffies
+ slice
;
3072 static struct cfq_group
*cfq_get_next_cfqg(struct cfq_data
*cfqd
)
3074 struct cfq_rb_root
*st
= &cfqd
->grp_service_tree
;
3075 struct cfq_group
*cfqg
;
3077 if (RB_EMPTY_ROOT(&st
->rb
))
3079 cfqg
= cfq_rb_first_group(st
);
3080 update_min_vdisktime(st
);
3084 static void cfq_choose_cfqg(struct cfq_data
*cfqd
)
3086 struct cfq_group
*cfqg
= cfq_get_next_cfqg(cfqd
);
3088 cfqd
->serving_group
= cfqg
;
3090 /* Restore the workload type data */
3091 if (cfqg
->saved_wl_slice
) {
3092 cfqd
->workload_expires
= jiffies
+ cfqg
->saved_wl_slice
;
3093 cfqd
->serving_wl_type
= cfqg
->saved_wl_type
;
3094 cfqd
->serving_wl_class
= cfqg
->saved_wl_class
;
3096 cfqd
->workload_expires
= jiffies
- 1;
3098 choose_wl_class_and_type(cfqd
, cfqg
);
3102 * Select a queue for service. If we have a current active queue,
3103 * check whether to continue servicing it, or retrieve and set a new one.
3105 static struct cfq_queue
*cfq_select_queue(struct cfq_data
*cfqd
)
3107 struct cfq_queue
*cfqq
, *new_cfqq
= NULL
;
3109 cfqq
= cfqd
->active_queue
;
3113 if (!cfqd
->rq_queued
)
3117 * We were waiting for group to get backlogged. Expire the queue
3119 if (cfq_cfqq_wait_busy(cfqq
) && !RB_EMPTY_ROOT(&cfqq
->sort_list
))
3123 * The active queue has run out of time, expire it and select new.
3125 if (cfq_slice_used(cfqq
) && !cfq_cfqq_must_dispatch(cfqq
)) {
3127 * If slice had not expired at the completion of last request
3128 * we might not have turned on wait_busy flag. Don't expire
3129 * the queue yet. Allow the group to get backlogged.
3131 * The very fact that we have used the slice, that means we
3132 * have been idling all along on this queue and it should be
3133 * ok to wait for this request to complete.
3135 if (cfqq
->cfqg
->nr_cfqq
== 1 && RB_EMPTY_ROOT(&cfqq
->sort_list
)
3136 && cfqq
->dispatched
&& cfq_should_idle(cfqd
, cfqq
)) {
3140 goto check_group_idle
;
3144 * The active queue has requests and isn't expired, allow it to
3147 if (!RB_EMPTY_ROOT(&cfqq
->sort_list
))
3151 * If another queue has a request waiting within our mean seek
3152 * distance, let it run. The expire code will check for close
3153 * cooperators and put the close queue at the front of the service
3154 * tree. If possible, merge the expiring queue with the new cfqq.
3156 new_cfqq
= cfq_close_cooperator(cfqd
, cfqq
);
3158 if (!cfqq
->new_cfqq
)
3159 cfq_setup_merge(cfqq
, new_cfqq
);
3164 * No requests pending. If the active queue still has requests in
3165 * flight or is idling for a new request, allow either of these
3166 * conditions to happen (or time out) before selecting a new queue.
3168 if (timer_pending(&cfqd
->idle_slice_timer
)) {
3174 * This is a deep seek queue, but the device is much faster than
3175 * the queue can deliver, don't idle
3177 if (CFQQ_SEEKY(cfqq
) && cfq_cfqq_idle_window(cfqq
) &&
3178 (cfq_cfqq_slice_new(cfqq
) ||
3179 (cfqq
->slice_end
- jiffies
> jiffies
- cfqq
->slice_start
))) {
3180 cfq_clear_cfqq_deep(cfqq
);
3181 cfq_clear_cfqq_idle_window(cfqq
);
3184 if (cfqq
->dispatched
&& cfq_should_idle(cfqd
, cfqq
)) {
3190 * If group idle is enabled and there are requests dispatched from
3191 * this group, wait for requests to complete.
3194 if (cfqd
->cfq_group_idle
&& cfqq
->cfqg
->nr_cfqq
== 1 &&
3195 cfqq
->cfqg
->dispatched
&&
3196 !cfq_io_thinktime_big(cfqd
, &cfqq
->cfqg
->ttime
, true)) {
3202 cfq_slice_expired(cfqd
, 0);
3205 * Current queue expired. Check if we have to switch to a new
3209 cfq_choose_cfqg(cfqd
);
3211 cfqq
= cfq_set_active_queue(cfqd
, new_cfqq
);
3216 static int __cfq_forced_dispatch_cfqq(struct cfq_queue
*cfqq
)
3220 while (cfqq
->next_rq
) {
3221 cfq_dispatch_insert(cfqq
->cfqd
->queue
, cfqq
->next_rq
);
3225 BUG_ON(!list_empty(&cfqq
->fifo
));
3227 /* By default cfqq is not expired if it is empty. Do it explicitly */
3228 __cfq_slice_expired(cfqq
->cfqd
, cfqq
, 0);
3233 * Drain our current requests. Used for barriers and when switching
3234 * io schedulers on-the-fly.
3236 static int cfq_forced_dispatch(struct cfq_data
*cfqd
)
3238 struct cfq_queue
*cfqq
;
3241 /* Expire the timeslice of the current active queue first */
3242 cfq_slice_expired(cfqd
, 0);
3243 while ((cfqq
= cfq_get_next_queue_forced(cfqd
)) != NULL
) {
3244 __cfq_set_active_queue(cfqd
, cfqq
);
3245 dispatched
+= __cfq_forced_dispatch_cfqq(cfqq
);
3248 BUG_ON(cfqd
->busy_queues
);
3250 cfq_log(cfqd
, "forced_dispatch=%d", dispatched
);
3254 static inline bool cfq_slice_used_soon(struct cfq_data
*cfqd
,
3255 struct cfq_queue
*cfqq
)
3257 /* the queue hasn't finished any request, can't estimate */
3258 if (cfq_cfqq_slice_new(cfqq
))
3260 if (time_after(jiffies
+ cfqd
->cfq_slice_idle
* cfqq
->dispatched
,
3267 static bool cfq_may_dispatch(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
3269 unsigned int max_dispatch
;
3272 * Drain async requests before we start sync IO
3274 if (cfq_should_idle(cfqd
, cfqq
) && cfqd
->rq_in_flight
[BLK_RW_ASYNC
])
3278 * If this is an async queue and we have sync IO in flight, let it wait
3280 if (cfqd
->rq_in_flight
[BLK_RW_SYNC
] && !cfq_cfqq_sync(cfqq
))
3283 max_dispatch
= max_t(unsigned int, cfqd
->cfq_quantum
/ 2, 1);
3284 if (cfq_class_idle(cfqq
))
3288 * Does this cfqq already have too much IO in flight?
3290 if (cfqq
->dispatched
>= max_dispatch
) {
3291 bool promote_sync
= false;
3293 * idle queue must always only have a single IO in flight
3295 if (cfq_class_idle(cfqq
))
3299 * If there is only one sync queue
3300 * we can ignore async queue here and give the sync
3301 * queue no dispatch limit. The reason is a sync queue can
3302 * preempt async queue, limiting the sync queue doesn't make
3303 * sense. This is useful for aiostress test.
3305 if (cfq_cfqq_sync(cfqq
) && cfqd
->busy_sync_queues
== 1)
3306 promote_sync
= true;
3309 * We have other queues, don't allow more IO from this one
3311 if (cfqd
->busy_queues
> 1 && cfq_slice_used_soon(cfqd
, cfqq
) &&
3316 * Sole queue user, no limit
3318 if (cfqd
->busy_queues
== 1 || promote_sync
)
3322 * Normally we start throttling cfqq when cfq_quantum/2
3323 * requests have been dispatched. But we can drive
3324 * deeper queue depths at the beginning of slice
3325 * subjected to upper limit of cfq_quantum.
3327 max_dispatch
= cfqd
->cfq_quantum
;
3331 * Async queues must wait a bit before being allowed dispatch.
3332 * We also ramp up the dispatch depth gradually for async IO,
3333 * based on the last sync IO we serviced
3335 if (!cfq_cfqq_sync(cfqq
) && cfqd
->cfq_latency
) {
3336 unsigned long last_sync
= jiffies
- cfqd
->last_delayed_sync
;
3339 depth
= last_sync
/ cfqd
->cfq_slice
[1];
3340 if (!depth
&& !cfqq
->dispatched
)
3342 if (depth
< max_dispatch
)
3343 max_dispatch
= depth
;
3347 * If we're below the current max, allow a dispatch
3349 return cfqq
->dispatched
< max_dispatch
;
3353 * Dispatch a request from cfqq, moving them to the request queue
3356 static bool cfq_dispatch_request(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
3360 BUG_ON(RB_EMPTY_ROOT(&cfqq
->sort_list
));
3362 if (!cfq_may_dispatch(cfqd
, cfqq
))
3366 * follow expired path, else get first next available
3368 rq
= cfq_check_fifo(cfqq
);
3373 * insert request into driver dispatch list
3375 cfq_dispatch_insert(cfqd
->queue
, rq
);
3377 if (!cfqd
->active_cic
) {
3378 struct cfq_io_cq
*cic
= RQ_CIC(rq
);
3380 atomic_long_inc(&cic
->icq
.ioc
->refcount
);
3381 cfqd
->active_cic
= cic
;
3388 * Find the cfqq that we need to service and move a request from that to the
3391 static int cfq_dispatch_requests(struct request_queue
*q
, int force
)
3393 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
3394 struct cfq_queue
*cfqq
;
3396 if (!cfqd
->busy_queues
)
3399 if (unlikely(force
))
3400 return cfq_forced_dispatch(cfqd
);
3402 cfqq
= cfq_select_queue(cfqd
);
3407 * Dispatch a request from this cfqq, if it is allowed
3409 if (!cfq_dispatch_request(cfqd
, cfqq
))
3412 cfqq
->slice_dispatch
++;
3413 cfq_clear_cfqq_must_dispatch(cfqq
);
3416 * expire an async queue immediately if it has used up its slice. idle
3417 * queue always expire after 1 dispatch round.
3419 if (cfqd
->busy_queues
> 1 && ((!cfq_cfqq_sync(cfqq
) &&
3420 cfqq
->slice_dispatch
>= cfq_prio_to_maxrq(cfqd
, cfqq
)) ||
3421 cfq_class_idle(cfqq
))) {
3422 cfqq
->slice_end
= jiffies
+ 1;
3423 cfq_slice_expired(cfqd
, 0);
3426 cfq_log_cfqq(cfqd
, cfqq
, "dispatched a request");
3431 * task holds one reference to the queue, dropped when task exits. each rq
3432 * in-flight on this queue also holds a reference, dropped when rq is freed.
3434 * Each cfq queue took a reference on the parent group. Drop it now.
3435 * queue lock must be held here.
3437 static void cfq_put_queue(struct cfq_queue
*cfqq
)
3439 struct cfq_data
*cfqd
= cfqq
->cfqd
;
3440 struct cfq_group
*cfqg
;
3442 BUG_ON(cfqq
->ref
<= 0);
3448 cfq_log_cfqq(cfqd
, cfqq
, "put_queue");
3449 BUG_ON(rb_first(&cfqq
->sort_list
));
3450 BUG_ON(cfqq
->allocated
[READ
] + cfqq
->allocated
[WRITE
]);
3453 if (unlikely(cfqd
->active_queue
== cfqq
)) {
3454 __cfq_slice_expired(cfqd
, cfqq
, 0);
3455 cfq_schedule_dispatch(cfqd
);
3458 BUG_ON(cfq_cfqq_on_rr(cfqq
));
3459 kmem_cache_free(cfq_pool
, cfqq
);
3463 static void cfq_put_cooperator(struct cfq_queue
*cfqq
)
3465 struct cfq_queue
*__cfqq
, *next
;
3468 * If this queue was scheduled to merge with another queue, be
3469 * sure to drop the reference taken on that queue (and others in
3470 * the merge chain). See cfq_setup_merge and cfq_merge_cfqqs.
3472 __cfqq
= cfqq
->new_cfqq
;
3474 if (__cfqq
== cfqq
) {
3475 WARN(1, "cfqq->new_cfqq loop detected\n");
3478 next
= __cfqq
->new_cfqq
;
3479 cfq_put_queue(__cfqq
);
3484 static void cfq_exit_cfqq(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
3486 if (unlikely(cfqq
== cfqd
->active_queue
)) {
3487 __cfq_slice_expired(cfqd
, cfqq
, 0);
3488 cfq_schedule_dispatch(cfqd
);
3491 cfq_put_cooperator(cfqq
);
3493 cfq_put_queue(cfqq
);
3496 static void cfq_init_icq(struct io_cq
*icq
)
3498 struct cfq_io_cq
*cic
= icq_to_cic(icq
);
3500 cic
->ttime
.last_end_request
= jiffies
;
3503 static void cfq_exit_icq(struct io_cq
*icq
)
3505 struct cfq_io_cq
*cic
= icq_to_cic(icq
);
3506 struct cfq_data
*cfqd
= cic_to_cfqd(cic
);
3508 if (cic_to_cfqq(cic
, false)) {
3509 cfq_exit_cfqq(cfqd
, cic_to_cfqq(cic
, false));
3510 cic_set_cfqq(cic
, NULL
, false);
3513 if (cic_to_cfqq(cic
, true)) {
3514 cfq_exit_cfqq(cfqd
, cic_to_cfqq(cic
, true));
3515 cic_set_cfqq(cic
, NULL
, true);
3519 static void cfq_init_prio_data(struct cfq_queue
*cfqq
, struct cfq_io_cq
*cic
)
3521 struct task_struct
*tsk
= current
;
3524 if (!cfq_cfqq_prio_changed(cfqq
))
3527 ioprio_class
= IOPRIO_PRIO_CLASS(cic
->ioprio
);
3528 switch (ioprio_class
) {
3530 printk(KERN_ERR
"cfq: bad prio %x\n", ioprio_class
);
3531 case IOPRIO_CLASS_NONE
:
3533 * no prio set, inherit CPU scheduling settings
3535 cfqq
->ioprio
= task_nice_ioprio(tsk
);
3536 cfqq
->ioprio_class
= task_nice_ioclass(tsk
);
3538 case IOPRIO_CLASS_RT
:
3539 cfqq
->ioprio
= IOPRIO_PRIO_DATA(cic
->ioprio
);
3540 cfqq
->ioprio_class
= IOPRIO_CLASS_RT
;
3542 case IOPRIO_CLASS_BE
:
3543 cfqq
->ioprio
= IOPRIO_PRIO_DATA(cic
->ioprio
);
3544 cfqq
->ioprio_class
= IOPRIO_CLASS_BE
;
3546 case IOPRIO_CLASS_IDLE
:
3547 cfqq
->ioprio_class
= IOPRIO_CLASS_IDLE
;
3549 cfq_clear_cfqq_idle_window(cfqq
);
3554 * keep track of original prio settings in case we have to temporarily
3555 * elevate the priority of this queue
3557 cfqq
->org_ioprio
= cfqq
->ioprio
;
3558 cfq_clear_cfqq_prio_changed(cfqq
);
3561 static void check_ioprio_changed(struct cfq_io_cq
*cic
, struct bio
*bio
)
3563 int ioprio
= cic
->icq
.ioc
->ioprio
;
3564 struct cfq_data
*cfqd
= cic_to_cfqd(cic
);
3565 struct cfq_queue
*cfqq
;
3568 * Check whether ioprio has changed. The condition may trigger
3569 * spuriously on a newly created cic but there's no harm.
3571 if (unlikely(!cfqd
) || likely(cic
->ioprio
== ioprio
))
3574 cfqq
= cic_to_cfqq(cic
, false);
3576 cfq_put_queue(cfqq
);
3577 cfqq
= cfq_get_queue(cfqd
, BLK_RW_ASYNC
, cic
, bio
);
3578 cic_set_cfqq(cic
, cfqq
, false);
3581 cfqq
= cic_to_cfqq(cic
, true);
3583 cfq_mark_cfqq_prio_changed(cfqq
);
3585 cic
->ioprio
= ioprio
;
3588 static void cfq_init_cfqq(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
3589 pid_t pid
, bool is_sync
)
3591 RB_CLEAR_NODE(&cfqq
->rb_node
);
3592 RB_CLEAR_NODE(&cfqq
->p_node
);
3593 INIT_LIST_HEAD(&cfqq
->fifo
);
3598 cfq_mark_cfqq_prio_changed(cfqq
);
3601 if (!cfq_class_idle(cfqq
))
3602 cfq_mark_cfqq_idle_window(cfqq
);
3603 cfq_mark_cfqq_sync(cfqq
);
3608 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3609 static void check_blkcg_changed(struct cfq_io_cq
*cic
, struct bio
*bio
)
3611 struct cfq_data
*cfqd
= cic_to_cfqd(cic
);
3612 struct cfq_queue
*sync_cfqq
;
3616 serial_nr
= bio_blkcg(bio
)->css
.serial_nr
;
3620 * Check whether blkcg has changed. The condition may trigger
3621 * spuriously on a newly created cic but there's no harm.
3623 if (unlikely(!cfqd
) || likely(cic
->blkcg_serial_nr
== serial_nr
))
3626 sync_cfqq
= cic_to_cfqq(cic
, 1);
3629 * Drop reference to sync queue. A new sync queue will be
3630 * assigned in new group upon arrival of a fresh request.
3632 cfq_log_cfqq(cfqd
, sync_cfqq
, "changed cgroup");
3633 cic_set_cfqq(cic
, NULL
, 1);
3634 cfq_put_queue(sync_cfqq
);
3637 cic
->blkcg_serial_nr
= serial_nr
;
3640 static inline void check_blkcg_changed(struct cfq_io_cq
*cic
, struct bio
*bio
) { }
3641 #endif /* CONFIG_CFQ_GROUP_IOSCHED */
3643 static struct cfq_queue
*
3644 cfq_find_alloc_queue(struct cfq_data
*cfqd
, bool is_sync
, struct cfq_io_cq
*cic
,
3647 struct blkcg
*blkcg
;
3648 struct cfq_queue
*cfqq
;
3649 struct cfq_group
*cfqg
;
3653 blkcg
= bio_blkcg(bio
);
3654 cfqg
= cfq_lookup_create_cfqg(cfqd
, blkcg
);
3656 cfqq
= &cfqd
->oom_cfqq
;
3660 cfqq
= cic_to_cfqq(cic
, is_sync
);
3663 * Always try a new alloc if we fell back to the OOM cfqq
3664 * originally, since it should just be a temporary situation.
3666 if (!cfqq
|| cfqq
== &cfqd
->oom_cfqq
) {
3667 cfqq
= kmem_cache_alloc_node(cfq_pool
,
3668 GFP_NOWAIT
| __GFP_ZERO
,
3671 cfq_init_cfqq(cfqd
, cfqq
, current
->pid
, is_sync
);
3672 cfq_init_prio_data(cfqq
, cic
);
3673 cfq_link_cfqq_cfqg(cfqq
, cfqg
);
3674 cfq_log_cfqq(cfqd
, cfqq
, "alloced");
3676 cfqq
= &cfqd
->oom_cfqq
;
3683 static struct cfq_queue
**
3684 cfq_async_queue_prio(struct cfq_data
*cfqd
, int ioprio_class
, int ioprio
)
3686 switch (ioprio_class
) {
3687 case IOPRIO_CLASS_RT
:
3688 return &cfqd
->async_cfqq
[0][ioprio
];
3689 case IOPRIO_CLASS_NONE
:
3690 ioprio
= IOPRIO_NORM
;
3692 case IOPRIO_CLASS_BE
:
3693 return &cfqd
->async_cfqq
[1][ioprio
];
3694 case IOPRIO_CLASS_IDLE
:
3695 return &cfqd
->async_idle_cfqq
;
3701 static struct cfq_queue
*
3702 cfq_get_queue(struct cfq_data
*cfqd
, bool is_sync
, struct cfq_io_cq
*cic
,
3705 int ioprio_class
= IOPRIO_PRIO_CLASS(cic
->ioprio
);
3706 int ioprio
= IOPRIO_PRIO_DATA(cic
->ioprio
);
3707 struct cfq_queue
**async_cfqq
;
3708 struct cfq_queue
*cfqq
;
3711 if (!ioprio_valid(cic
->ioprio
)) {
3712 struct task_struct
*tsk
= current
;
3713 ioprio
= task_nice_ioprio(tsk
);
3714 ioprio_class
= task_nice_ioclass(tsk
);
3716 async_cfqq
= cfq_async_queue_prio(cfqd
, ioprio_class
, ioprio
);
3722 cfqq
= cfq_find_alloc_queue(cfqd
, is_sync
, cic
, bio
);
3725 * pin the queue now that it's allocated, scheduler exit will prune it
3727 if (!is_sync
&& cfqq
!= &cfqd
->oom_cfqq
) {
3737 __cfq_update_io_thinktime(struct cfq_ttime
*ttime
, unsigned long slice_idle
)
3739 unsigned long elapsed
= jiffies
- ttime
->last_end_request
;
3740 elapsed
= min(elapsed
, 2UL * slice_idle
);
3742 ttime
->ttime_samples
= (7*ttime
->ttime_samples
+ 256) / 8;
3743 ttime
->ttime_total
= (7*ttime
->ttime_total
+ 256*elapsed
) / 8;
3744 ttime
->ttime_mean
= (ttime
->ttime_total
+ 128) / ttime
->ttime_samples
;
3748 cfq_update_io_thinktime(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
3749 struct cfq_io_cq
*cic
)
3751 if (cfq_cfqq_sync(cfqq
)) {
3752 __cfq_update_io_thinktime(&cic
->ttime
, cfqd
->cfq_slice_idle
);
3753 __cfq_update_io_thinktime(&cfqq
->service_tree
->ttime
,
3754 cfqd
->cfq_slice_idle
);
3756 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3757 __cfq_update_io_thinktime(&cfqq
->cfqg
->ttime
, cfqd
->cfq_group_idle
);
3762 cfq_update_io_seektime(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
3766 sector_t n_sec
= blk_rq_sectors(rq
);
3767 if (cfqq
->last_request_pos
) {
3768 if (cfqq
->last_request_pos
< blk_rq_pos(rq
))
3769 sdist
= blk_rq_pos(rq
) - cfqq
->last_request_pos
;
3771 sdist
= cfqq
->last_request_pos
- blk_rq_pos(rq
);
3774 cfqq
->seek_history
<<= 1;
3775 if (blk_queue_nonrot(cfqd
->queue
))
3776 cfqq
->seek_history
|= (n_sec
< CFQQ_SECT_THR_NONROT
);
3778 cfqq
->seek_history
|= (sdist
> CFQQ_SEEK_THR
);
3782 * Disable idle window if the process thinks too long or seeks so much that
3786 cfq_update_idle_window(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
3787 struct cfq_io_cq
*cic
)
3789 int old_idle
, enable_idle
;
3792 * Don't idle for async or idle io prio class
3794 if (!cfq_cfqq_sync(cfqq
) || cfq_class_idle(cfqq
))
3797 enable_idle
= old_idle
= cfq_cfqq_idle_window(cfqq
);
3799 if (cfqq
->queued
[0] + cfqq
->queued
[1] >= 4)
3800 cfq_mark_cfqq_deep(cfqq
);
3802 if (cfqq
->next_rq
&& (cfqq
->next_rq
->cmd_flags
& REQ_NOIDLE
))
3804 else if (!atomic_read(&cic
->icq
.ioc
->active_ref
) ||
3805 !cfqd
->cfq_slice_idle
||
3806 (!cfq_cfqq_deep(cfqq
) && CFQQ_SEEKY(cfqq
)))
3808 else if (sample_valid(cic
->ttime
.ttime_samples
)) {
3809 if (cic
->ttime
.ttime_mean
> cfqd
->cfq_slice_idle
)
3815 if (old_idle
!= enable_idle
) {
3816 cfq_log_cfqq(cfqd
, cfqq
, "idle=%d", enable_idle
);
3818 cfq_mark_cfqq_idle_window(cfqq
);
3820 cfq_clear_cfqq_idle_window(cfqq
);
3825 * Check if new_cfqq should preempt the currently active queue. Return 0 for
3826 * no or if we aren't sure, a 1 will cause a preempt.
3829 cfq_should_preempt(struct cfq_data
*cfqd
, struct cfq_queue
*new_cfqq
,
3832 struct cfq_queue
*cfqq
;
3834 cfqq
= cfqd
->active_queue
;
3838 if (cfq_class_idle(new_cfqq
))
3841 if (cfq_class_idle(cfqq
))
3845 * Don't allow a non-RT request to preempt an ongoing RT cfqq timeslice.
3847 if (cfq_class_rt(cfqq
) && !cfq_class_rt(new_cfqq
))
3851 * if the new request is sync, but the currently running queue is
3852 * not, let the sync request have priority.
3854 if (rq_is_sync(rq
) && !cfq_cfqq_sync(cfqq
))
3857 if (new_cfqq
->cfqg
!= cfqq
->cfqg
)
3860 if (cfq_slice_used(cfqq
))
3863 /* Allow preemption only if we are idling on sync-noidle tree */
3864 if (cfqd
->serving_wl_type
== SYNC_NOIDLE_WORKLOAD
&&
3865 cfqq_type(new_cfqq
) == SYNC_NOIDLE_WORKLOAD
&&
3866 new_cfqq
->service_tree
->count
== 2 &&
3867 RB_EMPTY_ROOT(&cfqq
->sort_list
))
3871 * So both queues are sync. Let the new request get disk time if
3872 * it's a metadata request and the current queue is doing regular IO.
3874 if ((rq
->cmd_flags
& REQ_PRIO
) && !cfqq
->prio_pending
)
3878 * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
3880 if (cfq_class_rt(new_cfqq
) && !cfq_class_rt(cfqq
))
3883 /* An idle queue should not be idle now for some reason */
3884 if (RB_EMPTY_ROOT(&cfqq
->sort_list
) && !cfq_should_idle(cfqd
, cfqq
))
3887 if (!cfqd
->active_cic
|| !cfq_cfqq_wait_request(cfqq
))
3891 * if this request is as-good as one we would expect from the
3892 * current cfqq, let it preempt
3894 if (cfq_rq_close(cfqd
, cfqq
, rq
))
3901 * cfqq preempts the active queue. if we allowed preempt with no slice left,
3902 * let it have half of its nominal slice.
3904 static void cfq_preempt_queue(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
3906 enum wl_type_t old_type
= cfqq_type(cfqd
->active_queue
);
3908 cfq_log_cfqq(cfqd
, cfqq
, "preempt");
3909 cfq_slice_expired(cfqd
, 1);
3912 * workload type is changed, don't save slice, otherwise preempt
3915 if (old_type
!= cfqq_type(cfqq
))
3916 cfqq
->cfqg
->saved_wl_slice
= 0;
3919 * Put the new queue at the front of the of the current list,
3920 * so we know that it will be selected next.
3922 BUG_ON(!cfq_cfqq_on_rr(cfqq
));
3924 cfq_service_tree_add(cfqd
, cfqq
, 1);
3926 cfqq
->slice_end
= 0;
3927 cfq_mark_cfqq_slice_new(cfqq
);
3931 * Called when a new fs request (rq) is added (to cfqq). Check if there's
3932 * something we should do about it
3935 cfq_rq_enqueued(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
3938 struct cfq_io_cq
*cic
= RQ_CIC(rq
);
3941 if (rq
->cmd_flags
& REQ_PRIO
)
3942 cfqq
->prio_pending
++;
3944 cfq_update_io_thinktime(cfqd
, cfqq
, cic
);
3945 cfq_update_io_seektime(cfqd
, cfqq
, rq
);
3946 cfq_update_idle_window(cfqd
, cfqq
, cic
);
3948 cfqq
->last_request_pos
= blk_rq_pos(rq
) + blk_rq_sectors(rq
);
3950 if (cfqq
== cfqd
->active_queue
) {
3952 * Remember that we saw a request from this process, but
3953 * don't start queuing just yet. Otherwise we risk seeing lots
3954 * of tiny requests, because we disrupt the normal plugging
3955 * and merging. If the request is already larger than a single
3956 * page, let it rip immediately. For that case we assume that
3957 * merging is already done. Ditto for a busy system that
3958 * has other work pending, don't risk delaying until the
3959 * idle timer unplug to continue working.
3961 if (cfq_cfqq_wait_request(cfqq
)) {
3962 if (blk_rq_bytes(rq
) > PAGE_CACHE_SIZE
||
3963 cfqd
->busy_queues
> 1) {
3964 cfq_del_timer(cfqd
, cfqq
);
3965 cfq_clear_cfqq_wait_request(cfqq
);
3966 __blk_run_queue(cfqd
->queue
);
3968 cfqg_stats_update_idle_time(cfqq
->cfqg
);
3969 cfq_mark_cfqq_must_dispatch(cfqq
);
3972 } else if (cfq_should_preempt(cfqd
, cfqq
, rq
)) {
3974 * not the active queue - expire current slice if it is
3975 * idle and has expired it's mean thinktime or this new queue
3976 * has some old slice time left and is of higher priority or
3977 * this new queue is RT and the current one is BE
3979 cfq_preempt_queue(cfqd
, cfqq
);
3980 __blk_run_queue(cfqd
->queue
);
3984 static void cfq_insert_request(struct request_queue
*q
, struct request
*rq
)
3986 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
3987 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
3989 cfq_log_cfqq(cfqd
, cfqq
, "insert_request");
3990 cfq_init_prio_data(cfqq
, RQ_CIC(rq
));
3992 rq
->fifo_time
= jiffies
+ cfqd
->cfq_fifo_expire
[rq_is_sync(rq
)];
3993 list_add_tail(&rq
->queuelist
, &cfqq
->fifo
);
3995 cfqg_stats_update_io_add(RQ_CFQG(rq
), cfqd
->serving_group
,
3997 cfq_rq_enqueued(cfqd
, cfqq
, rq
);
4001 * Update hw_tag based on peak queue depth over 50 samples under
4004 static void cfq_update_hw_tag(struct cfq_data
*cfqd
)
4006 struct cfq_queue
*cfqq
= cfqd
->active_queue
;
4008 if (cfqd
->rq_in_driver
> cfqd
->hw_tag_est_depth
)
4009 cfqd
->hw_tag_est_depth
= cfqd
->rq_in_driver
;
4011 if (cfqd
->hw_tag
== 1)
4014 if (cfqd
->rq_queued
<= CFQ_HW_QUEUE_MIN
&&
4015 cfqd
->rq_in_driver
<= CFQ_HW_QUEUE_MIN
)
4019 * If active queue hasn't enough requests and can idle, cfq might not
4020 * dispatch sufficient requests to hardware. Don't zero hw_tag in this
4023 if (cfqq
&& cfq_cfqq_idle_window(cfqq
) &&
4024 cfqq
->dispatched
+ cfqq
->queued
[0] + cfqq
->queued
[1] <
4025 CFQ_HW_QUEUE_MIN
&& cfqd
->rq_in_driver
< CFQ_HW_QUEUE_MIN
)
4028 if (cfqd
->hw_tag_samples
++ < 50)
4031 if (cfqd
->hw_tag_est_depth
>= CFQ_HW_QUEUE_MIN
)
4037 static bool cfq_should_wait_busy(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
4039 struct cfq_io_cq
*cic
= cfqd
->active_cic
;
4041 /* If the queue already has requests, don't wait */
4042 if (!RB_EMPTY_ROOT(&cfqq
->sort_list
))
4045 /* If there are other queues in the group, don't wait */
4046 if (cfqq
->cfqg
->nr_cfqq
> 1)
4049 /* the only queue in the group, but think time is big */
4050 if (cfq_io_thinktime_big(cfqd
, &cfqq
->cfqg
->ttime
, true))
4053 if (cfq_slice_used(cfqq
))
4056 /* if slice left is less than think time, wait busy */
4057 if (cic
&& sample_valid(cic
->ttime
.ttime_samples
)
4058 && (cfqq
->slice_end
- jiffies
< cic
->ttime
.ttime_mean
))
4062 * If think times is less than a jiffy than ttime_mean=0 and above
4063 * will not be true. It might happen that slice has not expired yet
4064 * but will expire soon (4-5 ns) during select_queue(). To cover the
4065 * case where think time is less than a jiffy, mark the queue wait
4066 * busy if only 1 jiffy is left in the slice.
4068 if (cfqq
->slice_end
- jiffies
== 1)
4074 static void cfq_completed_request(struct request_queue
*q
, struct request
*rq
)
4076 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
4077 struct cfq_data
*cfqd
= cfqq
->cfqd
;
4078 const int sync
= rq_is_sync(rq
);
4082 cfq_log_cfqq(cfqd
, cfqq
, "complete rqnoidle %d",
4083 !!(rq
->cmd_flags
& REQ_NOIDLE
));
4085 cfq_update_hw_tag(cfqd
);
4087 WARN_ON(!cfqd
->rq_in_driver
);
4088 WARN_ON(!cfqq
->dispatched
);
4089 cfqd
->rq_in_driver
--;
4091 (RQ_CFQG(rq
))->dispatched
--;
4092 cfqg_stats_update_completion(cfqq
->cfqg
, rq_start_time_ns(rq
),
4093 rq_io_start_time_ns(rq
), rq
->cmd_flags
);
4095 cfqd
->rq_in_flight
[cfq_cfqq_sync(cfqq
)]--;
4098 struct cfq_rb_root
*st
;
4100 RQ_CIC(rq
)->ttime
.last_end_request
= now
;
4102 if (cfq_cfqq_on_rr(cfqq
))
4103 st
= cfqq
->service_tree
;
4105 st
= st_for(cfqq
->cfqg
, cfqq_class(cfqq
),
4108 st
->ttime
.last_end_request
= now
;
4109 if (!time_after(rq
->start_time
+ cfqd
->cfq_fifo_expire
[1], now
))
4110 cfqd
->last_delayed_sync
= now
;
4113 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4114 cfqq
->cfqg
->ttime
.last_end_request
= now
;
4118 * If this is the active queue, check if it needs to be expired,
4119 * or if we want to idle in case it has no pending requests.
4121 if (cfqd
->active_queue
== cfqq
) {
4122 const bool cfqq_empty
= RB_EMPTY_ROOT(&cfqq
->sort_list
);
4124 if (cfq_cfqq_slice_new(cfqq
)) {
4125 cfq_set_prio_slice(cfqd
, cfqq
);
4126 cfq_clear_cfqq_slice_new(cfqq
);
4130 * Should we wait for next request to come in before we expire
4133 if (cfq_should_wait_busy(cfqd
, cfqq
)) {
4134 unsigned long extend_sl
= cfqd
->cfq_slice_idle
;
4135 if (!cfqd
->cfq_slice_idle
)
4136 extend_sl
= cfqd
->cfq_group_idle
;
4137 cfqq
->slice_end
= jiffies
+ extend_sl
;
4138 cfq_mark_cfqq_wait_busy(cfqq
);
4139 cfq_log_cfqq(cfqd
, cfqq
, "will busy wait");
4143 * Idling is not enabled on:
4145 * - idle-priority queues
4147 * - queues with still some requests queued
4148 * - when there is a close cooperator
4150 if (cfq_slice_used(cfqq
) || cfq_class_idle(cfqq
))
4151 cfq_slice_expired(cfqd
, 1);
4152 else if (sync
&& cfqq_empty
&&
4153 !cfq_close_cooperator(cfqd
, cfqq
)) {
4154 cfq_arm_slice_timer(cfqd
);
4158 if (!cfqd
->rq_in_driver
)
4159 cfq_schedule_dispatch(cfqd
);
4162 static inline int __cfq_may_queue(struct cfq_queue
*cfqq
)
4164 if (cfq_cfqq_wait_request(cfqq
) && !cfq_cfqq_must_alloc_slice(cfqq
)) {
4165 cfq_mark_cfqq_must_alloc_slice(cfqq
);
4166 return ELV_MQUEUE_MUST
;
4169 return ELV_MQUEUE_MAY
;
4172 static int cfq_may_queue(struct request_queue
*q
, int rw
)
4174 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
4175 struct task_struct
*tsk
= current
;
4176 struct cfq_io_cq
*cic
;
4177 struct cfq_queue
*cfqq
;
4180 * don't force setup of a queue from here, as a call to may_queue
4181 * does not necessarily imply that a request actually will be queued.
4182 * so just lookup a possibly existing queue, or return 'may queue'
4185 cic
= cfq_cic_lookup(cfqd
, tsk
->io_context
);
4187 return ELV_MQUEUE_MAY
;
4189 cfqq
= cic_to_cfqq(cic
, rw_is_sync(rw
));
4191 cfq_init_prio_data(cfqq
, cic
);
4193 return __cfq_may_queue(cfqq
);
4196 return ELV_MQUEUE_MAY
;
4200 * queue lock held here
4202 static void cfq_put_request(struct request
*rq
)
4204 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
4207 const int rw
= rq_data_dir(rq
);
4209 BUG_ON(!cfqq
->allocated
[rw
]);
4210 cfqq
->allocated
[rw
]--;
4212 /* Put down rq reference on cfqg */
4213 cfqg_put(RQ_CFQG(rq
));
4214 rq
->elv
.priv
[0] = NULL
;
4215 rq
->elv
.priv
[1] = NULL
;
4217 cfq_put_queue(cfqq
);
4221 static struct cfq_queue
*
4222 cfq_merge_cfqqs(struct cfq_data
*cfqd
, struct cfq_io_cq
*cic
,
4223 struct cfq_queue
*cfqq
)
4225 cfq_log_cfqq(cfqd
, cfqq
, "merging with queue %p", cfqq
->new_cfqq
);
4226 cic_set_cfqq(cic
, cfqq
->new_cfqq
, 1);
4227 cfq_mark_cfqq_coop(cfqq
->new_cfqq
);
4228 cfq_put_queue(cfqq
);
4229 return cic_to_cfqq(cic
, 1);
4233 * Returns NULL if a new cfqq should be allocated, or the old cfqq if this
4234 * was the last process referring to said cfqq.
4236 static struct cfq_queue
*
4237 split_cfqq(struct cfq_io_cq
*cic
, struct cfq_queue
*cfqq
)
4239 if (cfqq_process_refs(cfqq
) == 1) {
4240 cfqq
->pid
= current
->pid
;
4241 cfq_clear_cfqq_coop(cfqq
);
4242 cfq_clear_cfqq_split_coop(cfqq
);
4246 cic_set_cfqq(cic
, NULL
, 1);
4248 cfq_put_cooperator(cfqq
);
4250 cfq_put_queue(cfqq
);
4254 * Allocate cfq data structures associated with this request.
4257 cfq_set_request(struct request_queue
*q
, struct request
*rq
, struct bio
*bio
,
4260 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
4261 struct cfq_io_cq
*cic
= icq_to_cic(rq
->elv
.icq
);
4262 const int rw
= rq_data_dir(rq
);
4263 const bool is_sync
= rq_is_sync(rq
);
4264 struct cfq_queue
*cfqq
;
4266 spin_lock_irq(q
->queue_lock
);
4268 check_ioprio_changed(cic
, bio
);
4269 check_blkcg_changed(cic
, bio
);
4271 cfqq
= cic_to_cfqq(cic
, is_sync
);
4272 if (!cfqq
|| cfqq
== &cfqd
->oom_cfqq
) {
4274 cfq_put_queue(cfqq
);
4275 cfqq
= cfq_get_queue(cfqd
, is_sync
, cic
, bio
);
4276 cic_set_cfqq(cic
, cfqq
, is_sync
);
4279 * If the queue was seeky for too long, break it apart.
4281 if (cfq_cfqq_coop(cfqq
) && cfq_cfqq_split_coop(cfqq
)) {
4282 cfq_log_cfqq(cfqd
, cfqq
, "breaking apart cfqq");
4283 cfqq
= split_cfqq(cic
, cfqq
);
4289 * Check to see if this queue is scheduled to merge with
4290 * another, closely cooperating queue. The merging of
4291 * queues happens here as it must be done in process context.
4292 * The reference on new_cfqq was taken in merge_cfqqs.
4295 cfqq
= cfq_merge_cfqqs(cfqd
, cic
, cfqq
);
4298 cfqq
->allocated
[rw
]++;
4301 cfqg_get(cfqq
->cfqg
);
4302 rq
->elv
.priv
[0] = cfqq
;
4303 rq
->elv
.priv
[1] = cfqq
->cfqg
;
4304 spin_unlock_irq(q
->queue_lock
);
4308 static void cfq_kick_queue(struct work_struct
*work
)
4310 struct cfq_data
*cfqd
=
4311 container_of(work
, struct cfq_data
, unplug_work
);
4312 struct request_queue
*q
= cfqd
->queue
;
4314 spin_lock_irq(q
->queue_lock
);
4315 __blk_run_queue(cfqd
->queue
);
4316 spin_unlock_irq(q
->queue_lock
);
4320 * Timer running if the active_queue is currently idling inside its time slice
4322 static void cfq_idle_slice_timer(unsigned long data
)
4324 struct cfq_data
*cfqd
= (struct cfq_data
*) data
;
4325 struct cfq_queue
*cfqq
;
4326 unsigned long flags
;
4329 cfq_log(cfqd
, "idle timer fired");
4331 spin_lock_irqsave(cfqd
->queue
->queue_lock
, flags
);
4333 cfqq
= cfqd
->active_queue
;
4338 * We saw a request before the queue expired, let it through
4340 if (cfq_cfqq_must_dispatch(cfqq
))
4346 if (cfq_slice_used(cfqq
))
4350 * only expire and reinvoke request handler, if there are
4351 * other queues with pending requests
4353 if (!cfqd
->busy_queues
)
4357 * not expired and it has a request pending, let it dispatch
4359 if (!RB_EMPTY_ROOT(&cfqq
->sort_list
))
4363 * Queue depth flag is reset only when the idle didn't succeed
4365 cfq_clear_cfqq_deep(cfqq
);
4368 cfq_slice_expired(cfqd
, timed_out
);
4370 cfq_schedule_dispatch(cfqd
);
4372 spin_unlock_irqrestore(cfqd
->queue
->queue_lock
, flags
);
4375 static void cfq_shutdown_timer_wq(struct cfq_data
*cfqd
)
4377 del_timer_sync(&cfqd
->idle_slice_timer
);
4378 cancel_work_sync(&cfqd
->unplug_work
);
4381 static void cfq_put_async_queues(struct cfq_data
*cfqd
)
4385 for (i
= 0; i
< IOPRIO_BE_NR
; i
++) {
4386 if (cfqd
->async_cfqq
[0][i
])
4387 cfq_put_queue(cfqd
->async_cfqq
[0][i
]);
4388 if (cfqd
->async_cfqq
[1][i
])
4389 cfq_put_queue(cfqd
->async_cfqq
[1][i
]);
4392 if (cfqd
->async_idle_cfqq
)
4393 cfq_put_queue(cfqd
->async_idle_cfqq
);
4396 static void cfq_exit_queue(struct elevator_queue
*e
)
4398 struct cfq_data
*cfqd
= e
->elevator_data
;
4399 struct request_queue
*q
= cfqd
->queue
;
4401 cfq_shutdown_timer_wq(cfqd
);
4403 spin_lock_irq(q
->queue_lock
);
4405 if (cfqd
->active_queue
)
4406 __cfq_slice_expired(cfqd
, cfqd
->active_queue
, 0);
4408 cfq_put_async_queues(cfqd
);
4410 spin_unlock_irq(q
->queue_lock
);
4412 cfq_shutdown_timer_wq(cfqd
);
4414 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4415 blkcg_deactivate_policy(q
, &blkcg_policy_cfq
);
4417 kfree(cfqd
->root_group
);
4422 static int cfq_init_queue(struct request_queue
*q
, struct elevator_type
*e
)
4424 struct cfq_data
*cfqd
;
4425 struct blkcg_gq
*blkg __maybe_unused
;
4427 struct elevator_queue
*eq
;
4429 eq
= elevator_alloc(q
, e
);
4433 cfqd
= kzalloc_node(sizeof(*cfqd
), GFP_KERNEL
, q
->node
);
4435 kobject_put(&eq
->kobj
);
4438 eq
->elevator_data
= cfqd
;
4441 spin_lock_irq(q
->queue_lock
);
4443 spin_unlock_irq(q
->queue_lock
);
4445 /* Init root service tree */
4446 cfqd
->grp_service_tree
= CFQ_RB_ROOT
;
4448 /* Init root group and prefer root group over other groups by default */
4449 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4450 ret
= blkcg_activate_policy(q
, &blkcg_policy_cfq
);
4454 cfqd
->root_group
= blkg_to_cfqg(q
->root_blkg
);
4457 cfqd
->root_group
= kzalloc_node(sizeof(*cfqd
->root_group
),
4458 GFP_KERNEL
, cfqd
->queue
->node
);
4459 if (!cfqd
->root_group
)
4462 cfq_init_cfqg_base(cfqd
->root_group
);
4464 cfqd
->root_group
->weight
= 2 * CFQ_WEIGHT_DEFAULT
;
4465 cfqd
->root_group
->leaf_weight
= 2 * CFQ_WEIGHT_DEFAULT
;
4468 * Not strictly needed (since RB_ROOT just clears the node and we
4469 * zeroed cfqd on alloc), but better be safe in case someone decides
4470 * to add magic to the rb code
4472 for (i
= 0; i
< CFQ_PRIO_LISTS
; i
++)
4473 cfqd
->prio_trees
[i
] = RB_ROOT
;
4476 * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
4477 * Grab a permanent reference to it, so that the normal code flow
4478 * will not attempt to free it. oom_cfqq is linked to root_group
4479 * but shouldn't hold a reference as it'll never be unlinked. Lose
4480 * the reference from linking right away.
4482 cfq_init_cfqq(cfqd
, &cfqd
->oom_cfqq
, 1, 0);
4483 cfqd
->oom_cfqq
.ref
++;
4485 spin_lock_irq(q
->queue_lock
);
4486 cfq_link_cfqq_cfqg(&cfqd
->oom_cfqq
, cfqd
->root_group
);
4487 cfqg_put(cfqd
->root_group
);
4488 spin_unlock_irq(q
->queue_lock
);
4490 init_timer(&cfqd
->idle_slice_timer
);
4491 cfqd
->idle_slice_timer
.function
= cfq_idle_slice_timer
;
4492 cfqd
->idle_slice_timer
.data
= (unsigned long) cfqd
;
4494 INIT_WORK(&cfqd
->unplug_work
, cfq_kick_queue
);
4496 cfqd
->cfq_quantum
= cfq_quantum
;
4497 cfqd
->cfq_fifo_expire
[0] = cfq_fifo_expire
[0];
4498 cfqd
->cfq_fifo_expire
[1] = cfq_fifo_expire
[1];
4499 cfqd
->cfq_back_max
= cfq_back_max
;
4500 cfqd
->cfq_back_penalty
= cfq_back_penalty
;
4501 cfqd
->cfq_slice
[0] = cfq_slice_async
;
4502 cfqd
->cfq_slice
[1] = cfq_slice_sync
;
4503 cfqd
->cfq_target_latency
= cfq_target_latency
;
4504 cfqd
->cfq_slice_async_rq
= cfq_slice_async_rq
;
4505 cfqd
->cfq_slice_idle
= cfq_slice_idle
;
4506 cfqd
->cfq_group_idle
= cfq_group_idle
;
4507 cfqd
->cfq_latency
= 1;
4510 * we optimistically start assuming sync ops weren't delayed in last
4511 * second, in order to have larger depth for async operations.
4513 cfqd
->last_delayed_sync
= jiffies
- HZ
;
4518 kobject_put(&eq
->kobj
);
4522 static void cfq_registered_queue(struct request_queue
*q
)
4524 struct elevator_queue
*e
= q
->elevator
;
4525 struct cfq_data
*cfqd
= e
->elevator_data
;
4528 * Default to IOPS mode with no idling for SSDs
4530 if (blk_queue_nonrot(q
))
4531 cfqd
->cfq_slice_idle
= 0;
4535 * sysfs parts below -->
4538 cfq_var_show(unsigned int var
, char *page
)
4540 return sprintf(page
, "%u\n", var
);
4544 cfq_var_store(unsigned int *var
, const char *page
, size_t count
)
4546 char *p
= (char *) page
;
4548 *var
= simple_strtoul(p
, &p
, 10);
4552 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
4553 static ssize_t __FUNC(struct elevator_queue *e, char *page) \
4555 struct cfq_data *cfqd = e->elevator_data; \
4556 unsigned int __data = __VAR; \
4558 __data = jiffies_to_msecs(__data); \
4559 return cfq_var_show(__data, (page)); \
4561 SHOW_FUNCTION(cfq_quantum_show
, cfqd
->cfq_quantum
, 0);
4562 SHOW_FUNCTION(cfq_fifo_expire_sync_show
, cfqd
->cfq_fifo_expire
[1], 1);
4563 SHOW_FUNCTION(cfq_fifo_expire_async_show
, cfqd
->cfq_fifo_expire
[0], 1);
4564 SHOW_FUNCTION(cfq_back_seek_max_show
, cfqd
->cfq_back_max
, 0);
4565 SHOW_FUNCTION(cfq_back_seek_penalty_show
, cfqd
->cfq_back_penalty
, 0);
4566 SHOW_FUNCTION(cfq_slice_idle_show
, cfqd
->cfq_slice_idle
, 1);
4567 SHOW_FUNCTION(cfq_group_idle_show
, cfqd
->cfq_group_idle
, 1);
4568 SHOW_FUNCTION(cfq_slice_sync_show
, cfqd
->cfq_slice
[1], 1);
4569 SHOW_FUNCTION(cfq_slice_async_show
, cfqd
->cfq_slice
[0], 1);
4570 SHOW_FUNCTION(cfq_slice_async_rq_show
, cfqd
->cfq_slice_async_rq
, 0);
4571 SHOW_FUNCTION(cfq_low_latency_show
, cfqd
->cfq_latency
, 0);
4572 SHOW_FUNCTION(cfq_target_latency_show
, cfqd
->cfq_target_latency
, 1);
4573 #undef SHOW_FUNCTION
4575 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
4576 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
4578 struct cfq_data *cfqd = e->elevator_data; \
4579 unsigned int __data; \
4580 int ret = cfq_var_store(&__data, (page), count); \
4581 if (__data < (MIN)) \
4583 else if (__data > (MAX)) \
4586 *(__PTR) = msecs_to_jiffies(__data); \
4588 *(__PTR) = __data; \
4591 STORE_FUNCTION(cfq_quantum_store
, &cfqd
->cfq_quantum
, 1, UINT_MAX
, 0);
4592 STORE_FUNCTION(cfq_fifo_expire_sync_store
, &cfqd
->cfq_fifo_expire
[1], 1,
4594 STORE_FUNCTION(cfq_fifo_expire_async_store
, &cfqd
->cfq_fifo_expire
[0], 1,
4596 STORE_FUNCTION(cfq_back_seek_max_store
, &cfqd
->cfq_back_max
, 0, UINT_MAX
, 0);
4597 STORE_FUNCTION(cfq_back_seek_penalty_store
, &cfqd
->cfq_back_penalty
, 1,
4599 STORE_FUNCTION(cfq_slice_idle_store
, &cfqd
->cfq_slice_idle
, 0, UINT_MAX
, 1);
4600 STORE_FUNCTION(cfq_group_idle_store
, &cfqd
->cfq_group_idle
, 0, UINT_MAX
, 1);
4601 STORE_FUNCTION(cfq_slice_sync_store
, &cfqd
->cfq_slice
[1], 1, UINT_MAX
, 1);
4602 STORE_FUNCTION(cfq_slice_async_store
, &cfqd
->cfq_slice
[0], 1, UINT_MAX
, 1);
4603 STORE_FUNCTION(cfq_slice_async_rq_store
, &cfqd
->cfq_slice_async_rq
, 1,
4605 STORE_FUNCTION(cfq_low_latency_store
, &cfqd
->cfq_latency
, 0, 1, 0);
4606 STORE_FUNCTION(cfq_target_latency_store
, &cfqd
->cfq_target_latency
, 1, UINT_MAX
, 1);
4607 #undef STORE_FUNCTION
4609 #define CFQ_ATTR(name) \
4610 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
4612 static struct elv_fs_entry cfq_attrs
[] = {
4614 CFQ_ATTR(fifo_expire_sync
),
4615 CFQ_ATTR(fifo_expire_async
),
4616 CFQ_ATTR(back_seek_max
),
4617 CFQ_ATTR(back_seek_penalty
),
4618 CFQ_ATTR(slice_sync
),
4619 CFQ_ATTR(slice_async
),
4620 CFQ_ATTR(slice_async_rq
),
4621 CFQ_ATTR(slice_idle
),
4622 CFQ_ATTR(group_idle
),
4623 CFQ_ATTR(low_latency
),
4624 CFQ_ATTR(target_latency
),
4628 static struct elevator_type iosched_cfq
= {
4630 .elevator_merge_fn
= cfq_merge
,
4631 .elevator_merged_fn
= cfq_merged_request
,
4632 .elevator_merge_req_fn
= cfq_merged_requests
,
4633 .elevator_allow_merge_fn
= cfq_allow_merge
,
4634 .elevator_bio_merged_fn
= cfq_bio_merged
,
4635 .elevator_dispatch_fn
= cfq_dispatch_requests
,
4636 .elevator_add_req_fn
= cfq_insert_request
,
4637 .elevator_activate_req_fn
= cfq_activate_request
,
4638 .elevator_deactivate_req_fn
= cfq_deactivate_request
,
4639 .elevator_completed_req_fn
= cfq_completed_request
,
4640 .elevator_former_req_fn
= elv_rb_former_request
,
4641 .elevator_latter_req_fn
= elv_rb_latter_request
,
4642 .elevator_init_icq_fn
= cfq_init_icq
,
4643 .elevator_exit_icq_fn
= cfq_exit_icq
,
4644 .elevator_set_req_fn
= cfq_set_request
,
4645 .elevator_put_req_fn
= cfq_put_request
,
4646 .elevator_may_queue_fn
= cfq_may_queue
,
4647 .elevator_init_fn
= cfq_init_queue
,
4648 .elevator_exit_fn
= cfq_exit_queue
,
4649 .elevator_registered_fn
= cfq_registered_queue
,
4651 .icq_size
= sizeof(struct cfq_io_cq
),
4652 .icq_align
= __alignof__(struct cfq_io_cq
),
4653 .elevator_attrs
= cfq_attrs
,
4654 .elevator_name
= "cfq",
4655 .elevator_owner
= THIS_MODULE
,
4658 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4659 static struct blkcg_policy blkcg_policy_cfq
= {
4660 .pd_size
= sizeof(struct cfq_group
),
4661 .cpd_size
= sizeof(struct cfq_group_data
),
4662 .cftypes
= cfq_blkcg_files
,
4664 .cpd_init_fn
= cfq_cpd_init
,
4665 .pd_init_fn
= cfq_pd_init
,
4666 .pd_offline_fn
= cfq_pd_offline
,
4667 .pd_reset_stats_fn
= cfq_pd_reset_stats
,
4671 static int __init
cfq_init(void)
4676 * could be 0 on HZ < 1000 setups
4678 if (!cfq_slice_async
)
4679 cfq_slice_async
= 1;
4680 if (!cfq_slice_idle
)
4683 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4684 if (!cfq_group_idle
)
4687 ret
= blkcg_policy_register(&blkcg_policy_cfq
);
4695 cfq_pool
= KMEM_CACHE(cfq_queue
, 0);
4699 ret
= elv_register(&iosched_cfq
);
4706 kmem_cache_destroy(cfq_pool
);
4708 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4709 blkcg_policy_unregister(&blkcg_policy_cfq
);
4714 static void __exit
cfq_exit(void)
4716 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4717 blkcg_policy_unregister(&blkcg_policy_cfq
);
4719 elv_unregister(&iosched_cfq
);
4720 kmem_cache_destroy(cfq_pool
);
4723 module_init(cfq_init
);
4724 module_exit(cfq_exit
);
4726 MODULE_AUTHOR("Jens Axboe");
4727 MODULE_LICENSE("GPL");
4728 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");