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
,
889 static inline struct cfq_io_cq
*icq_to_cic(struct io_cq
*icq
)
891 /* cic->icq is the first member, %NULL will convert to %NULL */
892 return container_of(icq
, struct cfq_io_cq
, icq
);
895 static inline struct cfq_io_cq
*cfq_cic_lookup(struct cfq_data
*cfqd
,
896 struct io_context
*ioc
)
899 return icq_to_cic(ioc_lookup_icq(ioc
, cfqd
->queue
));
903 static inline struct cfq_queue
*cic_to_cfqq(struct cfq_io_cq
*cic
, bool is_sync
)
905 return cic
->cfqq
[is_sync
];
908 static inline void cic_set_cfqq(struct cfq_io_cq
*cic
, struct cfq_queue
*cfqq
,
911 cic
->cfqq
[is_sync
] = cfqq
;
914 static inline struct cfq_data
*cic_to_cfqd(struct cfq_io_cq
*cic
)
916 return cic
->icq
.q
->elevator
->elevator_data
;
920 * We regard a request as SYNC, if it's either a read or has the SYNC bit
921 * set (in which case it could also be direct WRITE).
923 static inline bool cfq_bio_sync(struct bio
*bio
)
925 return bio_data_dir(bio
) == READ
|| (bio
->bi_rw
& REQ_SYNC
);
929 * scheduler run of queue, if there are requests pending and no one in the
930 * driver that will restart queueing
932 static inline void cfq_schedule_dispatch(struct cfq_data
*cfqd
)
934 if (cfqd
->busy_queues
) {
935 cfq_log(cfqd
, "schedule dispatch");
936 kblockd_schedule_work(&cfqd
->unplug_work
);
941 * Scale schedule slice based on io priority. Use the sync time slice only
942 * if a queue is marked sync and has sync io queued. A sync queue with async
943 * io only, should not get full sync slice length.
945 static inline int cfq_prio_slice(struct cfq_data
*cfqd
, bool sync
,
948 const int base_slice
= cfqd
->cfq_slice
[sync
];
950 WARN_ON(prio
>= IOPRIO_BE_NR
);
952 return base_slice
+ (base_slice
/CFQ_SLICE_SCALE
* (4 - prio
));
956 cfq_prio_to_slice(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
958 return cfq_prio_slice(cfqd
, cfq_cfqq_sync(cfqq
), cfqq
->ioprio
);
962 * cfqg_scale_charge - scale disk time charge according to cfqg weight
963 * @charge: disk time being charged
964 * @vfraction: vfraction of the cfqg, fixed point w/ CFQ_SERVICE_SHIFT
966 * Scale @charge according to @vfraction, which is in range (0, 1]. The
967 * scaling is inversely proportional.
969 * scaled = charge / vfraction
971 * The result is also in fixed point w/ CFQ_SERVICE_SHIFT.
973 static inline u64
cfqg_scale_charge(unsigned long charge
,
974 unsigned int vfraction
)
976 u64 c
= charge
<< CFQ_SERVICE_SHIFT
; /* make it fixed point */
978 /* charge / vfraction */
979 c
<<= CFQ_SERVICE_SHIFT
;
980 do_div(c
, vfraction
);
984 static inline u64
max_vdisktime(u64 min_vdisktime
, u64 vdisktime
)
986 s64 delta
= (s64
)(vdisktime
- min_vdisktime
);
988 min_vdisktime
= vdisktime
;
990 return min_vdisktime
;
993 static inline u64
min_vdisktime(u64 min_vdisktime
, u64 vdisktime
)
995 s64 delta
= (s64
)(vdisktime
- min_vdisktime
);
997 min_vdisktime
= vdisktime
;
999 return min_vdisktime
;
1002 static void update_min_vdisktime(struct cfq_rb_root
*st
)
1004 struct cfq_group
*cfqg
;
1007 cfqg
= rb_entry_cfqg(st
->left
);
1008 st
->min_vdisktime
= max_vdisktime(st
->min_vdisktime
,
1014 * get averaged number of queues of RT/BE priority.
1015 * average is updated, with a formula that gives more weight to higher numbers,
1016 * to quickly follows sudden increases and decrease slowly
1019 static inline unsigned cfq_group_get_avg_queues(struct cfq_data
*cfqd
,
1020 struct cfq_group
*cfqg
, bool rt
)
1022 unsigned min_q
, max_q
;
1023 unsigned mult
= cfq_hist_divisor
- 1;
1024 unsigned round
= cfq_hist_divisor
/ 2;
1025 unsigned busy
= cfq_group_busy_queues_wl(rt
, cfqd
, cfqg
);
1027 min_q
= min(cfqg
->busy_queues_avg
[rt
], busy
);
1028 max_q
= max(cfqg
->busy_queues_avg
[rt
], busy
);
1029 cfqg
->busy_queues_avg
[rt
] = (mult
* max_q
+ min_q
+ round
) /
1031 return cfqg
->busy_queues_avg
[rt
];
1034 static inline unsigned
1035 cfq_group_slice(struct cfq_data
*cfqd
, struct cfq_group
*cfqg
)
1037 return cfqd
->cfq_target_latency
* cfqg
->vfraction
>> CFQ_SERVICE_SHIFT
;
1040 static inline unsigned
1041 cfq_scaled_cfqq_slice(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
1043 unsigned slice
= cfq_prio_to_slice(cfqd
, cfqq
);
1044 if (cfqd
->cfq_latency
) {
1046 * interested queues (we consider only the ones with the same
1047 * priority class in the cfq group)
1049 unsigned iq
= cfq_group_get_avg_queues(cfqd
, cfqq
->cfqg
,
1050 cfq_class_rt(cfqq
));
1051 unsigned sync_slice
= cfqd
->cfq_slice
[1];
1052 unsigned expect_latency
= sync_slice
* iq
;
1053 unsigned group_slice
= cfq_group_slice(cfqd
, cfqq
->cfqg
);
1055 if (expect_latency
> group_slice
) {
1056 unsigned base_low_slice
= 2 * cfqd
->cfq_slice_idle
;
1057 /* scale low_slice according to IO priority
1058 * and sync vs async */
1059 unsigned low_slice
=
1060 min(slice
, base_low_slice
* slice
/ sync_slice
);
1061 /* the adapted slice value is scaled to fit all iqs
1062 * into the target latency */
1063 slice
= max(slice
* group_slice
/ expect_latency
,
1071 cfq_set_prio_slice(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
1073 unsigned slice
= cfq_scaled_cfqq_slice(cfqd
, cfqq
);
1075 cfqq
->slice_start
= jiffies
;
1076 cfqq
->slice_end
= jiffies
+ slice
;
1077 cfqq
->allocated_slice
= slice
;
1078 cfq_log_cfqq(cfqd
, cfqq
, "set_slice=%lu", cfqq
->slice_end
- jiffies
);
1082 * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
1083 * isn't valid until the first request from the dispatch is activated
1084 * and the slice time set.
1086 static inline bool cfq_slice_used(struct cfq_queue
*cfqq
)
1088 if (cfq_cfqq_slice_new(cfqq
))
1090 if (time_before(jiffies
, cfqq
->slice_end
))
1097 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
1098 * We choose the request that is closest to the head right now. Distance
1099 * behind the head is penalized and only allowed to a certain extent.
1101 static struct request
*
1102 cfq_choose_req(struct cfq_data
*cfqd
, struct request
*rq1
, struct request
*rq2
, sector_t last
)
1104 sector_t s1
, s2
, d1
= 0, d2
= 0;
1105 unsigned long back_max
;
1106 #define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */
1107 #define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */
1108 unsigned wrap
= 0; /* bit mask: requests behind the disk head? */
1110 if (rq1
== NULL
|| rq1
== rq2
)
1115 if (rq_is_sync(rq1
) != rq_is_sync(rq2
))
1116 return rq_is_sync(rq1
) ? rq1
: rq2
;
1118 if ((rq1
->cmd_flags
^ rq2
->cmd_flags
) & REQ_PRIO
)
1119 return rq1
->cmd_flags
& REQ_PRIO
? rq1
: rq2
;
1121 s1
= blk_rq_pos(rq1
);
1122 s2
= blk_rq_pos(rq2
);
1125 * by definition, 1KiB is 2 sectors
1127 back_max
= cfqd
->cfq_back_max
* 2;
1130 * Strict one way elevator _except_ in the case where we allow
1131 * short backward seeks which are biased as twice the cost of a
1132 * similar forward seek.
1136 else if (s1
+ back_max
>= last
)
1137 d1
= (last
- s1
) * cfqd
->cfq_back_penalty
;
1139 wrap
|= CFQ_RQ1_WRAP
;
1143 else if (s2
+ back_max
>= last
)
1144 d2
= (last
- s2
) * cfqd
->cfq_back_penalty
;
1146 wrap
|= CFQ_RQ2_WRAP
;
1148 /* Found required data */
1151 * By doing switch() on the bit mask "wrap" we avoid having to
1152 * check two variables for all permutations: --> faster!
1155 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
1171 case (CFQ_RQ1_WRAP
|CFQ_RQ2_WRAP
): /* both rqs wrapped */
1174 * Since both rqs are wrapped,
1175 * start with the one that's further behind head
1176 * (--> only *one* back seek required),
1177 * since back seek takes more time than forward.
1187 * The below is leftmost cache rbtree addon
1189 static struct cfq_queue
*cfq_rb_first(struct cfq_rb_root
*root
)
1191 /* Service tree is empty */
1196 root
->left
= rb_first(&root
->rb
);
1199 return rb_entry(root
->left
, struct cfq_queue
, rb_node
);
1204 static struct cfq_group
*cfq_rb_first_group(struct cfq_rb_root
*root
)
1207 root
->left
= rb_first(&root
->rb
);
1210 return rb_entry_cfqg(root
->left
);
1215 static void rb_erase_init(struct rb_node
*n
, struct rb_root
*root
)
1221 static void cfq_rb_erase(struct rb_node
*n
, struct cfq_rb_root
*root
)
1223 if (root
->left
== n
)
1225 rb_erase_init(n
, &root
->rb
);
1230 * would be nice to take fifo expire time into account as well
1232 static struct request
*
1233 cfq_find_next_rq(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
1234 struct request
*last
)
1236 struct rb_node
*rbnext
= rb_next(&last
->rb_node
);
1237 struct rb_node
*rbprev
= rb_prev(&last
->rb_node
);
1238 struct request
*next
= NULL
, *prev
= NULL
;
1240 BUG_ON(RB_EMPTY_NODE(&last
->rb_node
));
1243 prev
= rb_entry_rq(rbprev
);
1246 next
= rb_entry_rq(rbnext
);
1248 rbnext
= rb_first(&cfqq
->sort_list
);
1249 if (rbnext
&& rbnext
!= &last
->rb_node
)
1250 next
= rb_entry_rq(rbnext
);
1253 return cfq_choose_req(cfqd
, next
, prev
, blk_rq_pos(last
));
1256 static unsigned long cfq_slice_offset(struct cfq_data
*cfqd
,
1257 struct cfq_queue
*cfqq
)
1260 * just an approximation, should be ok.
1262 return (cfqq
->cfqg
->nr_cfqq
- 1) * (cfq_prio_slice(cfqd
, 1, 0) -
1263 cfq_prio_slice(cfqd
, cfq_cfqq_sync(cfqq
), cfqq
->ioprio
));
1267 cfqg_key(struct cfq_rb_root
*st
, struct cfq_group
*cfqg
)
1269 return cfqg
->vdisktime
- st
->min_vdisktime
;
1273 __cfq_group_service_tree_add(struct cfq_rb_root
*st
, struct cfq_group
*cfqg
)
1275 struct rb_node
**node
= &st
->rb
.rb_node
;
1276 struct rb_node
*parent
= NULL
;
1277 struct cfq_group
*__cfqg
;
1278 s64 key
= cfqg_key(st
, cfqg
);
1281 while (*node
!= NULL
) {
1283 __cfqg
= rb_entry_cfqg(parent
);
1285 if (key
< cfqg_key(st
, __cfqg
))
1286 node
= &parent
->rb_left
;
1288 node
= &parent
->rb_right
;
1294 st
->left
= &cfqg
->rb_node
;
1296 rb_link_node(&cfqg
->rb_node
, parent
, node
);
1297 rb_insert_color(&cfqg
->rb_node
, &st
->rb
);
1301 * This has to be called only on activation of cfqg
1304 cfq_update_group_weight(struct cfq_group
*cfqg
)
1306 if (cfqg
->new_weight
) {
1307 cfqg
->weight
= cfqg
->new_weight
;
1308 cfqg
->new_weight
= 0;
1313 cfq_update_group_leaf_weight(struct cfq_group
*cfqg
)
1315 BUG_ON(!RB_EMPTY_NODE(&cfqg
->rb_node
));
1317 if (cfqg
->new_leaf_weight
) {
1318 cfqg
->leaf_weight
= cfqg
->new_leaf_weight
;
1319 cfqg
->new_leaf_weight
= 0;
1324 cfq_group_service_tree_add(struct cfq_rb_root
*st
, struct cfq_group
*cfqg
)
1326 unsigned int vfr
= 1 << CFQ_SERVICE_SHIFT
; /* start with 1 */
1327 struct cfq_group
*pos
= cfqg
;
1328 struct cfq_group
*parent
;
1331 /* add to the service tree */
1332 BUG_ON(!RB_EMPTY_NODE(&cfqg
->rb_node
));
1335 * Update leaf_weight. We cannot update weight at this point
1336 * because cfqg might already have been activated and is
1337 * contributing its current weight to the parent's child_weight.
1339 cfq_update_group_leaf_weight(cfqg
);
1340 __cfq_group_service_tree_add(st
, cfqg
);
1343 * Activate @cfqg and calculate the portion of vfraction @cfqg is
1344 * entitled to. vfraction is calculated by walking the tree
1345 * towards the root calculating the fraction it has at each level.
1346 * The compounded ratio is how much vfraction @cfqg owns.
1348 * Start with the proportion tasks in this cfqg has against active
1349 * children cfqgs - its leaf_weight against children_weight.
1351 propagate
= !pos
->nr_active
++;
1352 pos
->children_weight
+= pos
->leaf_weight
;
1353 vfr
= vfr
* pos
->leaf_weight
/ pos
->children_weight
;
1356 * Compound ->weight walking up the tree. Both activation and
1357 * vfraction calculation are done in the same loop. Propagation
1358 * stops once an already activated node is met. vfraction
1359 * calculation should always continue to the root.
1361 while ((parent
= cfqg_parent(pos
))) {
1363 cfq_update_group_weight(pos
);
1364 propagate
= !parent
->nr_active
++;
1365 parent
->children_weight
+= pos
->weight
;
1367 vfr
= vfr
* pos
->weight
/ parent
->children_weight
;
1371 cfqg
->vfraction
= max_t(unsigned, vfr
, 1);
1375 cfq_group_notify_queue_add(struct cfq_data
*cfqd
, struct cfq_group
*cfqg
)
1377 struct cfq_rb_root
*st
= &cfqd
->grp_service_tree
;
1378 struct cfq_group
*__cfqg
;
1382 if (!RB_EMPTY_NODE(&cfqg
->rb_node
))
1386 * Currently put the group at the end. Later implement something
1387 * so that groups get lesser vtime based on their weights, so that
1388 * if group does not loose all if it was not continuously backlogged.
1390 n
= rb_last(&st
->rb
);
1392 __cfqg
= rb_entry_cfqg(n
);
1393 cfqg
->vdisktime
= __cfqg
->vdisktime
+ CFQ_IDLE_DELAY
;
1395 cfqg
->vdisktime
= st
->min_vdisktime
;
1396 cfq_group_service_tree_add(st
, cfqg
);
1400 cfq_group_service_tree_del(struct cfq_rb_root
*st
, struct cfq_group
*cfqg
)
1402 struct cfq_group
*pos
= cfqg
;
1406 * Undo activation from cfq_group_service_tree_add(). Deactivate
1407 * @cfqg and propagate deactivation upwards.
1409 propagate
= !--pos
->nr_active
;
1410 pos
->children_weight
-= pos
->leaf_weight
;
1413 struct cfq_group
*parent
= cfqg_parent(pos
);
1415 /* @pos has 0 nr_active at this point */
1416 WARN_ON_ONCE(pos
->children_weight
);
1422 propagate
= !--parent
->nr_active
;
1423 parent
->children_weight
-= pos
->weight
;
1427 /* remove from the service tree */
1428 if (!RB_EMPTY_NODE(&cfqg
->rb_node
))
1429 cfq_rb_erase(&cfqg
->rb_node
, st
);
1433 cfq_group_notify_queue_del(struct cfq_data
*cfqd
, struct cfq_group
*cfqg
)
1435 struct cfq_rb_root
*st
= &cfqd
->grp_service_tree
;
1437 BUG_ON(cfqg
->nr_cfqq
< 1);
1440 /* If there are other cfq queues under this group, don't delete it */
1444 cfq_log_cfqg(cfqd
, cfqg
, "del_from_rr group");
1445 cfq_group_service_tree_del(st
, cfqg
);
1446 cfqg
->saved_wl_slice
= 0;
1447 cfqg_stats_update_dequeue(cfqg
);
1450 static inline unsigned int cfq_cfqq_slice_usage(struct cfq_queue
*cfqq
,
1451 unsigned int *unaccounted_time
)
1453 unsigned int slice_used
;
1456 * Queue got expired before even a single request completed or
1457 * got expired immediately after first request completion.
1459 if (!cfqq
->slice_start
|| cfqq
->slice_start
== jiffies
) {
1461 * Also charge the seek time incurred to the group, otherwise
1462 * if there are mutiple queues in the group, each can dispatch
1463 * a single request on seeky media and cause lots of seek time
1464 * and group will never know it.
1466 slice_used
= max_t(unsigned, (jiffies
- cfqq
->dispatch_start
),
1469 slice_used
= jiffies
- cfqq
->slice_start
;
1470 if (slice_used
> cfqq
->allocated_slice
) {
1471 *unaccounted_time
= slice_used
- cfqq
->allocated_slice
;
1472 slice_used
= cfqq
->allocated_slice
;
1474 if (time_after(cfqq
->slice_start
, cfqq
->dispatch_start
))
1475 *unaccounted_time
+= cfqq
->slice_start
-
1476 cfqq
->dispatch_start
;
1482 static void cfq_group_served(struct cfq_data
*cfqd
, struct cfq_group
*cfqg
,
1483 struct cfq_queue
*cfqq
)
1485 struct cfq_rb_root
*st
= &cfqd
->grp_service_tree
;
1486 unsigned int used_sl
, charge
, unaccounted_sl
= 0;
1487 int nr_sync
= cfqg
->nr_cfqq
- cfqg_busy_async_queues(cfqd
, cfqg
)
1488 - cfqg
->service_tree_idle
.count
;
1491 BUG_ON(nr_sync
< 0);
1492 used_sl
= charge
= cfq_cfqq_slice_usage(cfqq
, &unaccounted_sl
);
1494 if (iops_mode(cfqd
))
1495 charge
= cfqq
->slice_dispatch
;
1496 else if (!cfq_cfqq_sync(cfqq
) && !nr_sync
)
1497 charge
= cfqq
->allocated_slice
;
1500 * Can't update vdisktime while on service tree and cfqg->vfraction
1501 * is valid only while on it. Cache vfr, leave the service tree,
1502 * update vdisktime and go back on. The re-addition to the tree
1503 * will also update the weights as necessary.
1505 vfr
= cfqg
->vfraction
;
1506 cfq_group_service_tree_del(st
, cfqg
);
1507 cfqg
->vdisktime
+= cfqg_scale_charge(charge
, vfr
);
1508 cfq_group_service_tree_add(st
, cfqg
);
1510 /* This group is being expired. Save the context */
1511 if (time_after(cfqd
->workload_expires
, jiffies
)) {
1512 cfqg
->saved_wl_slice
= cfqd
->workload_expires
1514 cfqg
->saved_wl_type
= cfqd
->serving_wl_type
;
1515 cfqg
->saved_wl_class
= cfqd
->serving_wl_class
;
1517 cfqg
->saved_wl_slice
= 0;
1519 cfq_log_cfqg(cfqd
, cfqg
, "served: vt=%llu min_vt=%llu", cfqg
->vdisktime
,
1521 cfq_log_cfqq(cfqq
->cfqd
, cfqq
,
1522 "sl_used=%u disp=%u charge=%u iops=%u sect=%lu",
1523 used_sl
, cfqq
->slice_dispatch
, charge
,
1524 iops_mode(cfqd
), cfqq
->nr_sectors
);
1525 cfqg_stats_update_timeslice_used(cfqg
, used_sl
, unaccounted_sl
);
1526 cfqg_stats_set_start_empty_time(cfqg
);
1530 * cfq_init_cfqg_base - initialize base part of a cfq_group
1531 * @cfqg: cfq_group to initialize
1533 * Initialize the base part which is used whether %CONFIG_CFQ_GROUP_IOSCHED
1534 * is enabled or not.
1536 static void cfq_init_cfqg_base(struct cfq_group
*cfqg
)
1538 struct cfq_rb_root
*st
;
1541 for_each_cfqg_st(cfqg
, i
, j
, st
)
1543 RB_CLEAR_NODE(&cfqg
->rb_node
);
1545 cfqg
->ttime
.last_end_request
= jiffies
;
1548 #ifdef CONFIG_CFQ_GROUP_IOSCHED
1549 static void cfqg_stats_init(struct cfqg_stats
*stats
)
1551 blkg_rwstat_init(&stats
->service_bytes
);
1552 blkg_rwstat_init(&stats
->serviced
);
1553 blkg_rwstat_init(&stats
->merged
);
1554 blkg_rwstat_init(&stats
->service_time
);
1555 blkg_rwstat_init(&stats
->wait_time
);
1556 blkg_rwstat_init(&stats
->queued
);
1558 blkg_stat_init(&stats
->sectors
);
1559 blkg_stat_init(&stats
->time
);
1561 #ifdef CONFIG_DEBUG_BLK_CGROUP
1562 blkg_stat_init(&stats
->unaccounted_time
);
1563 blkg_stat_init(&stats
->avg_queue_size_sum
);
1564 blkg_stat_init(&stats
->avg_queue_size_samples
);
1565 blkg_stat_init(&stats
->dequeue
);
1566 blkg_stat_init(&stats
->group_wait_time
);
1567 blkg_stat_init(&stats
->idle_time
);
1568 blkg_stat_init(&stats
->empty_time
);
1572 static void cfq_cpd_init(const struct blkcg
*blkcg
)
1574 struct cfq_group_data
*cgd
=
1575 cpd_to_cfqgd(blkcg
->pd
[blkcg_policy_cfq
.plid
]);
1577 if (blkcg
== &blkcg_root
) {
1578 cgd
->weight
= 2 * CFQ_WEIGHT_DEFAULT
;
1579 cgd
->leaf_weight
= 2 * CFQ_WEIGHT_DEFAULT
;
1581 cgd
->weight
= CFQ_WEIGHT_DEFAULT
;
1582 cgd
->leaf_weight
= CFQ_WEIGHT_DEFAULT
;
1586 static void cfq_pd_init(struct blkcg_gq
*blkg
)
1588 struct cfq_group
*cfqg
= blkg_to_cfqg(blkg
);
1589 struct cfq_group_data
*cgd
= blkcg_to_cfqgd(blkg
->blkcg
);
1591 cfq_init_cfqg_base(cfqg
);
1592 cfqg
->weight
= cgd
->weight
;
1593 cfqg
->leaf_weight
= cgd
->leaf_weight
;
1594 cfqg_stats_init(&cfqg
->stats
);
1595 cfqg_stats_init(&cfqg
->dead_stats
);
1598 static void cfq_pd_offline(struct blkcg_gq
*blkg
)
1601 * @blkg is going offline and will be ignored by
1602 * blkg_[rw]stat_recursive_sum(). Transfer stats to the parent so
1603 * that they don't get lost. If IOs complete after this point, the
1604 * stats for them will be lost. Oh well...
1606 cfqg_stats_xfer_dead(blkg_to_cfqg(blkg
));
1609 /* offset delta from cfqg->stats to cfqg->dead_stats */
1610 static const int dead_stats_off_delta
= offsetof(struct cfq_group
, dead_stats
) -
1611 offsetof(struct cfq_group
, stats
);
1613 /* to be used by recursive prfill, sums live and dead stats recursively */
1614 static u64
cfqg_stat_pd_recursive_sum(struct blkg_policy_data
*pd
, int off
)
1618 sum
+= blkg_stat_recursive_sum(pd
, off
);
1619 sum
+= blkg_stat_recursive_sum(pd
, off
+ dead_stats_off_delta
);
1623 /* to be used by recursive prfill, sums live and dead rwstats recursively */
1624 static struct blkg_rwstat
cfqg_rwstat_pd_recursive_sum(struct blkg_policy_data
*pd
,
1627 struct blkg_rwstat a
, b
;
1629 a
= blkg_rwstat_recursive_sum(pd
, off
);
1630 b
= blkg_rwstat_recursive_sum(pd
, off
+ dead_stats_off_delta
);
1631 blkg_rwstat_merge(&a
, &b
);
1635 static void cfq_pd_reset_stats(struct blkcg_gq
*blkg
)
1637 struct cfq_group
*cfqg
= blkg_to_cfqg(blkg
);
1639 cfqg_stats_reset(&cfqg
->stats
);
1640 cfqg_stats_reset(&cfqg
->dead_stats
);
1644 * Search for the cfq group current task belongs to. request_queue lock must
1647 static struct cfq_group
*cfq_lookup_create_cfqg(struct cfq_data
*cfqd
,
1648 struct blkcg
*blkcg
)
1650 struct request_queue
*q
= cfqd
->queue
;
1651 struct cfq_group
*cfqg
= NULL
;
1653 /* avoid lookup for the common case where there's no blkcg */
1654 if (blkcg
== &blkcg_root
) {
1655 cfqg
= cfqd
->root_group
;
1657 struct blkcg_gq
*blkg
;
1659 blkg
= blkg_lookup_create(blkcg
, q
);
1661 cfqg
= blkg_to_cfqg(blkg
);
1667 static void cfq_link_cfqq_cfqg(struct cfq_queue
*cfqq
, struct cfq_group
*cfqg
)
1669 /* Currently, all async queues are mapped to root group */
1670 if (!cfq_cfqq_sync(cfqq
))
1671 cfqg
= cfqq
->cfqd
->root_group
;
1674 /* cfqq reference on cfqg */
1678 static u64
cfqg_prfill_weight_device(struct seq_file
*sf
,
1679 struct blkg_policy_data
*pd
, int off
)
1681 struct cfq_group
*cfqg
= pd_to_cfqg(pd
);
1683 if (!cfqg
->dev_weight
)
1685 return __blkg_prfill_u64(sf
, pd
, cfqg
->dev_weight
);
1688 static int cfqg_print_weight_device(struct seq_file
*sf
, void *v
)
1690 blkcg_print_blkgs(sf
, css_to_blkcg(seq_css(sf
)),
1691 cfqg_prfill_weight_device
, &blkcg_policy_cfq
,
1696 static u64
cfqg_prfill_leaf_weight_device(struct seq_file
*sf
,
1697 struct blkg_policy_data
*pd
, int off
)
1699 struct cfq_group
*cfqg
= pd_to_cfqg(pd
);
1701 if (!cfqg
->dev_leaf_weight
)
1703 return __blkg_prfill_u64(sf
, pd
, cfqg
->dev_leaf_weight
);
1706 static int cfqg_print_leaf_weight_device(struct seq_file
*sf
, void *v
)
1708 blkcg_print_blkgs(sf
, css_to_blkcg(seq_css(sf
)),
1709 cfqg_prfill_leaf_weight_device
, &blkcg_policy_cfq
,
1714 static int cfq_print_weight(struct seq_file
*sf
, void *v
)
1716 struct blkcg
*blkcg
= css_to_blkcg(seq_css(sf
));
1717 struct cfq_group_data
*cgd
= blkcg_to_cfqgd(blkcg
);
1718 unsigned int val
= 0;
1723 seq_printf(sf
, "%u\n", val
);
1727 static int cfq_print_leaf_weight(struct seq_file
*sf
, void *v
)
1729 struct blkcg
*blkcg
= css_to_blkcg(seq_css(sf
));
1730 struct cfq_group_data
*cgd
= blkcg_to_cfqgd(blkcg
);
1731 unsigned int val
= 0;
1734 val
= cgd
->leaf_weight
;
1736 seq_printf(sf
, "%u\n", val
);
1740 static ssize_t
__cfqg_set_weight_device(struct kernfs_open_file
*of
,
1741 char *buf
, size_t nbytes
, loff_t off
,
1742 bool is_leaf_weight
)
1744 struct blkcg
*blkcg
= css_to_blkcg(of_css(of
));
1745 struct blkg_conf_ctx ctx
;
1746 struct cfq_group
*cfqg
;
1747 struct cfq_group_data
*cfqgd
;
1750 ret
= blkg_conf_prep(blkcg
, &blkcg_policy_cfq
, buf
, &ctx
);
1755 cfqg
= blkg_to_cfqg(ctx
.blkg
);
1756 cfqgd
= blkcg_to_cfqgd(blkcg
);
1757 if (!cfqg
|| !cfqgd
)
1760 if (!ctx
.v
|| (ctx
.v
>= CFQ_WEIGHT_MIN
&& ctx
.v
<= CFQ_WEIGHT_MAX
)) {
1761 if (!is_leaf_weight
) {
1762 cfqg
->dev_weight
= ctx
.v
;
1763 cfqg
->new_weight
= ctx
.v
?: cfqgd
->weight
;
1765 cfqg
->dev_leaf_weight
= ctx
.v
;
1766 cfqg
->new_leaf_weight
= ctx
.v
?: cfqgd
->leaf_weight
;
1772 blkg_conf_finish(&ctx
);
1773 return ret
?: nbytes
;
1776 static ssize_t
cfqg_set_weight_device(struct kernfs_open_file
*of
,
1777 char *buf
, size_t nbytes
, loff_t off
)
1779 return __cfqg_set_weight_device(of
, buf
, nbytes
, off
, false);
1782 static ssize_t
cfqg_set_leaf_weight_device(struct kernfs_open_file
*of
,
1783 char *buf
, size_t nbytes
, loff_t off
)
1785 return __cfqg_set_weight_device(of
, buf
, nbytes
, off
, true);
1788 static int __cfq_set_weight(struct cgroup_subsys_state
*css
, struct cftype
*cft
,
1789 u64 val
, bool is_leaf_weight
)
1791 struct blkcg
*blkcg
= css_to_blkcg(css
);
1792 struct blkcg_gq
*blkg
;
1793 struct cfq_group_data
*cfqgd
;
1796 if (val
< CFQ_WEIGHT_MIN
|| val
> CFQ_WEIGHT_MAX
)
1799 spin_lock_irq(&blkcg
->lock
);
1800 cfqgd
= blkcg_to_cfqgd(blkcg
);
1806 if (!is_leaf_weight
)
1807 cfqgd
->weight
= val
;
1809 cfqgd
->leaf_weight
= val
;
1811 hlist_for_each_entry(blkg
, &blkcg
->blkg_list
, blkcg_node
) {
1812 struct cfq_group
*cfqg
= blkg_to_cfqg(blkg
);
1817 if (!is_leaf_weight
) {
1818 if (!cfqg
->dev_weight
)
1819 cfqg
->new_weight
= cfqgd
->weight
;
1821 if (!cfqg
->dev_leaf_weight
)
1822 cfqg
->new_leaf_weight
= cfqgd
->leaf_weight
;
1827 spin_unlock_irq(&blkcg
->lock
);
1831 static int cfq_set_weight(struct cgroup_subsys_state
*css
, struct cftype
*cft
,
1834 return __cfq_set_weight(css
, cft
, val
, false);
1837 static int cfq_set_leaf_weight(struct cgroup_subsys_state
*css
,
1838 struct cftype
*cft
, u64 val
)
1840 return __cfq_set_weight(css
, cft
, val
, true);
1843 static int cfqg_print_stat(struct seq_file
*sf
, void *v
)
1845 blkcg_print_blkgs(sf
, css_to_blkcg(seq_css(sf
)), blkg_prfill_stat
,
1846 &blkcg_policy_cfq
, seq_cft(sf
)->private, false);
1850 static int cfqg_print_rwstat(struct seq_file
*sf
, void *v
)
1852 blkcg_print_blkgs(sf
, css_to_blkcg(seq_css(sf
)), blkg_prfill_rwstat
,
1853 &blkcg_policy_cfq
, seq_cft(sf
)->private, true);
1857 static u64
cfqg_prfill_stat_recursive(struct seq_file
*sf
,
1858 struct blkg_policy_data
*pd
, int off
)
1860 u64 sum
= cfqg_stat_pd_recursive_sum(pd
, off
);
1862 return __blkg_prfill_u64(sf
, pd
, sum
);
1865 static u64
cfqg_prfill_rwstat_recursive(struct seq_file
*sf
,
1866 struct blkg_policy_data
*pd
, int off
)
1868 struct blkg_rwstat sum
= cfqg_rwstat_pd_recursive_sum(pd
, off
);
1870 return __blkg_prfill_rwstat(sf
, pd
, &sum
);
1873 static int cfqg_print_stat_recursive(struct seq_file
*sf
, void *v
)
1875 blkcg_print_blkgs(sf
, css_to_blkcg(seq_css(sf
)),
1876 cfqg_prfill_stat_recursive
, &blkcg_policy_cfq
,
1877 seq_cft(sf
)->private, false);
1881 static int cfqg_print_rwstat_recursive(struct seq_file
*sf
, void *v
)
1883 blkcg_print_blkgs(sf
, css_to_blkcg(seq_css(sf
)),
1884 cfqg_prfill_rwstat_recursive
, &blkcg_policy_cfq
,
1885 seq_cft(sf
)->private, true);
1889 #ifdef CONFIG_DEBUG_BLK_CGROUP
1890 static u64
cfqg_prfill_avg_queue_size(struct seq_file
*sf
,
1891 struct blkg_policy_data
*pd
, int off
)
1893 struct cfq_group
*cfqg
= pd_to_cfqg(pd
);
1894 u64 samples
= blkg_stat_read(&cfqg
->stats
.avg_queue_size_samples
);
1898 v
= blkg_stat_read(&cfqg
->stats
.avg_queue_size_sum
);
1899 v
= div64_u64(v
, samples
);
1901 __blkg_prfill_u64(sf
, pd
, v
);
1905 /* print avg_queue_size */
1906 static int cfqg_print_avg_queue_size(struct seq_file
*sf
, void *v
)
1908 blkcg_print_blkgs(sf
, css_to_blkcg(seq_css(sf
)),
1909 cfqg_prfill_avg_queue_size
, &blkcg_policy_cfq
,
1913 #endif /* CONFIG_DEBUG_BLK_CGROUP */
1915 static struct cftype cfq_blkcg_files
[] = {
1916 /* on root, weight is mapped to leaf_weight */
1918 .name
= "weight_device",
1919 .flags
= CFTYPE_ONLY_ON_ROOT
,
1920 .seq_show
= cfqg_print_leaf_weight_device
,
1921 .write
= cfqg_set_leaf_weight_device
,
1925 .flags
= CFTYPE_ONLY_ON_ROOT
,
1926 .seq_show
= cfq_print_leaf_weight
,
1927 .write_u64
= cfq_set_leaf_weight
,
1930 /* no such mapping necessary for !roots */
1932 .name
= "weight_device",
1933 .flags
= CFTYPE_NOT_ON_ROOT
,
1934 .seq_show
= cfqg_print_weight_device
,
1935 .write
= cfqg_set_weight_device
,
1939 .flags
= CFTYPE_NOT_ON_ROOT
,
1940 .seq_show
= cfq_print_weight
,
1941 .write_u64
= cfq_set_weight
,
1945 .name
= "leaf_weight_device",
1946 .seq_show
= cfqg_print_leaf_weight_device
,
1947 .write
= cfqg_set_leaf_weight_device
,
1950 .name
= "leaf_weight",
1951 .seq_show
= cfq_print_leaf_weight
,
1952 .write_u64
= cfq_set_leaf_weight
,
1955 /* statistics, covers only the tasks in the cfqg */
1958 .private = offsetof(struct cfq_group
, stats
.time
),
1959 .seq_show
= cfqg_print_stat
,
1963 .private = offsetof(struct cfq_group
, stats
.sectors
),
1964 .seq_show
= cfqg_print_stat
,
1967 .name
= "io_service_bytes",
1968 .private = offsetof(struct cfq_group
, stats
.service_bytes
),
1969 .seq_show
= cfqg_print_rwstat
,
1972 .name
= "io_serviced",
1973 .private = offsetof(struct cfq_group
, stats
.serviced
),
1974 .seq_show
= cfqg_print_rwstat
,
1977 .name
= "io_service_time",
1978 .private = offsetof(struct cfq_group
, stats
.service_time
),
1979 .seq_show
= cfqg_print_rwstat
,
1982 .name
= "io_wait_time",
1983 .private = offsetof(struct cfq_group
, stats
.wait_time
),
1984 .seq_show
= cfqg_print_rwstat
,
1987 .name
= "io_merged",
1988 .private = offsetof(struct cfq_group
, stats
.merged
),
1989 .seq_show
= cfqg_print_rwstat
,
1992 .name
= "io_queued",
1993 .private = offsetof(struct cfq_group
, stats
.queued
),
1994 .seq_show
= cfqg_print_rwstat
,
1997 /* the same statictics which cover the cfqg and its descendants */
1999 .name
= "time_recursive",
2000 .private = offsetof(struct cfq_group
, stats
.time
),
2001 .seq_show
= cfqg_print_stat_recursive
,
2004 .name
= "sectors_recursive",
2005 .private = offsetof(struct cfq_group
, stats
.sectors
),
2006 .seq_show
= cfqg_print_stat_recursive
,
2009 .name
= "io_service_bytes_recursive",
2010 .private = offsetof(struct cfq_group
, stats
.service_bytes
),
2011 .seq_show
= cfqg_print_rwstat_recursive
,
2014 .name
= "io_serviced_recursive",
2015 .private = offsetof(struct cfq_group
, stats
.serviced
),
2016 .seq_show
= cfqg_print_rwstat_recursive
,
2019 .name
= "io_service_time_recursive",
2020 .private = offsetof(struct cfq_group
, stats
.service_time
),
2021 .seq_show
= cfqg_print_rwstat_recursive
,
2024 .name
= "io_wait_time_recursive",
2025 .private = offsetof(struct cfq_group
, stats
.wait_time
),
2026 .seq_show
= cfqg_print_rwstat_recursive
,
2029 .name
= "io_merged_recursive",
2030 .private = offsetof(struct cfq_group
, stats
.merged
),
2031 .seq_show
= cfqg_print_rwstat_recursive
,
2034 .name
= "io_queued_recursive",
2035 .private = offsetof(struct cfq_group
, stats
.queued
),
2036 .seq_show
= cfqg_print_rwstat_recursive
,
2038 #ifdef CONFIG_DEBUG_BLK_CGROUP
2040 .name
= "avg_queue_size",
2041 .seq_show
= cfqg_print_avg_queue_size
,
2044 .name
= "group_wait_time",
2045 .private = offsetof(struct cfq_group
, stats
.group_wait_time
),
2046 .seq_show
= cfqg_print_stat
,
2049 .name
= "idle_time",
2050 .private = offsetof(struct cfq_group
, stats
.idle_time
),
2051 .seq_show
= cfqg_print_stat
,
2054 .name
= "empty_time",
2055 .private = offsetof(struct cfq_group
, stats
.empty_time
),
2056 .seq_show
= cfqg_print_stat
,
2060 .private = offsetof(struct cfq_group
, stats
.dequeue
),
2061 .seq_show
= cfqg_print_stat
,
2064 .name
= "unaccounted_time",
2065 .private = offsetof(struct cfq_group
, stats
.unaccounted_time
),
2066 .seq_show
= cfqg_print_stat
,
2068 #endif /* CONFIG_DEBUG_BLK_CGROUP */
2071 #else /* GROUP_IOSCHED */
2072 static struct cfq_group
*cfq_lookup_create_cfqg(struct cfq_data
*cfqd
,
2073 struct blkcg
*blkcg
)
2075 return cfqd
->root_group
;
2079 cfq_link_cfqq_cfqg(struct cfq_queue
*cfqq
, struct cfq_group
*cfqg
) {
2083 #endif /* GROUP_IOSCHED */
2086 * The cfqd->service_trees holds all pending cfq_queue's that have
2087 * requests waiting to be processed. It is sorted in the order that
2088 * we will service the queues.
2090 static void cfq_service_tree_add(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
2093 struct rb_node
**p
, *parent
;
2094 struct cfq_queue
*__cfqq
;
2095 unsigned long rb_key
;
2096 struct cfq_rb_root
*st
;
2100 st
= st_for(cfqq
->cfqg
, cfqq_class(cfqq
), cfqq_type(cfqq
));
2101 if (cfq_class_idle(cfqq
)) {
2102 rb_key
= CFQ_IDLE_DELAY
;
2103 parent
= rb_last(&st
->rb
);
2104 if (parent
&& parent
!= &cfqq
->rb_node
) {
2105 __cfqq
= rb_entry(parent
, struct cfq_queue
, rb_node
);
2106 rb_key
+= __cfqq
->rb_key
;
2109 } else if (!add_front
) {
2111 * Get our rb key offset. Subtract any residual slice
2112 * value carried from last service. A negative resid
2113 * count indicates slice overrun, and this should position
2114 * the next service time further away in the tree.
2116 rb_key
= cfq_slice_offset(cfqd
, cfqq
) + jiffies
;
2117 rb_key
-= cfqq
->slice_resid
;
2118 cfqq
->slice_resid
= 0;
2121 __cfqq
= cfq_rb_first(st
);
2122 rb_key
+= __cfqq
? __cfqq
->rb_key
: jiffies
;
2125 if (!RB_EMPTY_NODE(&cfqq
->rb_node
)) {
2128 * same position, nothing more to do
2130 if (rb_key
== cfqq
->rb_key
&& cfqq
->service_tree
== st
)
2133 cfq_rb_erase(&cfqq
->rb_node
, cfqq
->service_tree
);
2134 cfqq
->service_tree
= NULL
;
2139 cfqq
->service_tree
= st
;
2140 p
= &st
->rb
.rb_node
;
2143 __cfqq
= rb_entry(parent
, struct cfq_queue
, rb_node
);
2146 * sort by key, that represents service time.
2148 if (time_before(rb_key
, __cfqq
->rb_key
))
2149 p
= &parent
->rb_left
;
2151 p
= &parent
->rb_right
;
2157 st
->left
= &cfqq
->rb_node
;
2159 cfqq
->rb_key
= rb_key
;
2160 rb_link_node(&cfqq
->rb_node
, parent
, p
);
2161 rb_insert_color(&cfqq
->rb_node
, &st
->rb
);
2163 if (add_front
|| !new_cfqq
)
2165 cfq_group_notify_queue_add(cfqd
, cfqq
->cfqg
);
2168 static struct cfq_queue
*
2169 cfq_prio_tree_lookup(struct cfq_data
*cfqd
, struct rb_root
*root
,
2170 sector_t sector
, struct rb_node
**ret_parent
,
2171 struct rb_node
***rb_link
)
2173 struct rb_node
**p
, *parent
;
2174 struct cfq_queue
*cfqq
= NULL
;
2182 cfqq
= rb_entry(parent
, struct cfq_queue
, p_node
);
2185 * Sort strictly based on sector. Smallest to the left,
2186 * largest to the right.
2188 if (sector
> blk_rq_pos(cfqq
->next_rq
))
2189 n
= &(*p
)->rb_right
;
2190 else if (sector
< blk_rq_pos(cfqq
->next_rq
))
2198 *ret_parent
= parent
;
2204 static void cfq_prio_tree_add(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
2206 struct rb_node
**p
, *parent
;
2207 struct cfq_queue
*__cfqq
;
2210 rb_erase(&cfqq
->p_node
, cfqq
->p_root
);
2211 cfqq
->p_root
= NULL
;
2214 if (cfq_class_idle(cfqq
))
2219 cfqq
->p_root
= &cfqd
->prio_trees
[cfqq
->org_ioprio
];
2220 __cfqq
= cfq_prio_tree_lookup(cfqd
, cfqq
->p_root
,
2221 blk_rq_pos(cfqq
->next_rq
), &parent
, &p
);
2223 rb_link_node(&cfqq
->p_node
, parent
, p
);
2224 rb_insert_color(&cfqq
->p_node
, cfqq
->p_root
);
2226 cfqq
->p_root
= NULL
;
2230 * Update cfqq's position in the service tree.
2232 static void cfq_resort_rr_list(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
2235 * Resorting requires the cfqq to be on the RR list already.
2237 if (cfq_cfqq_on_rr(cfqq
)) {
2238 cfq_service_tree_add(cfqd
, cfqq
, 0);
2239 cfq_prio_tree_add(cfqd
, cfqq
);
2244 * add to busy list of queues for service, trying to be fair in ordering
2245 * the pending list according to last request service
2247 static void cfq_add_cfqq_rr(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
2249 cfq_log_cfqq(cfqd
, cfqq
, "add_to_rr");
2250 BUG_ON(cfq_cfqq_on_rr(cfqq
));
2251 cfq_mark_cfqq_on_rr(cfqq
);
2252 cfqd
->busy_queues
++;
2253 if (cfq_cfqq_sync(cfqq
))
2254 cfqd
->busy_sync_queues
++;
2256 cfq_resort_rr_list(cfqd
, cfqq
);
2260 * Called when the cfqq no longer has requests pending, remove it from
2263 static void cfq_del_cfqq_rr(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
2265 cfq_log_cfqq(cfqd
, cfqq
, "del_from_rr");
2266 BUG_ON(!cfq_cfqq_on_rr(cfqq
));
2267 cfq_clear_cfqq_on_rr(cfqq
);
2269 if (!RB_EMPTY_NODE(&cfqq
->rb_node
)) {
2270 cfq_rb_erase(&cfqq
->rb_node
, cfqq
->service_tree
);
2271 cfqq
->service_tree
= NULL
;
2274 rb_erase(&cfqq
->p_node
, cfqq
->p_root
);
2275 cfqq
->p_root
= NULL
;
2278 cfq_group_notify_queue_del(cfqd
, cfqq
->cfqg
);
2279 BUG_ON(!cfqd
->busy_queues
);
2280 cfqd
->busy_queues
--;
2281 if (cfq_cfqq_sync(cfqq
))
2282 cfqd
->busy_sync_queues
--;
2286 * rb tree support functions
2288 static void cfq_del_rq_rb(struct request
*rq
)
2290 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
2291 const int sync
= rq_is_sync(rq
);
2293 BUG_ON(!cfqq
->queued
[sync
]);
2294 cfqq
->queued
[sync
]--;
2296 elv_rb_del(&cfqq
->sort_list
, rq
);
2298 if (cfq_cfqq_on_rr(cfqq
) && RB_EMPTY_ROOT(&cfqq
->sort_list
)) {
2300 * Queue will be deleted from service tree when we actually
2301 * expire it later. Right now just remove it from prio tree
2305 rb_erase(&cfqq
->p_node
, cfqq
->p_root
);
2306 cfqq
->p_root
= NULL
;
2311 static void cfq_add_rq_rb(struct request
*rq
)
2313 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
2314 struct cfq_data
*cfqd
= cfqq
->cfqd
;
2315 struct request
*prev
;
2317 cfqq
->queued
[rq_is_sync(rq
)]++;
2319 elv_rb_add(&cfqq
->sort_list
, rq
);
2321 if (!cfq_cfqq_on_rr(cfqq
))
2322 cfq_add_cfqq_rr(cfqd
, cfqq
);
2325 * check if this request is a better next-serve candidate
2327 prev
= cfqq
->next_rq
;
2328 cfqq
->next_rq
= cfq_choose_req(cfqd
, cfqq
->next_rq
, rq
, cfqd
->last_position
);
2331 * adjust priority tree position, if ->next_rq changes
2333 if (prev
!= cfqq
->next_rq
)
2334 cfq_prio_tree_add(cfqd
, cfqq
);
2336 BUG_ON(!cfqq
->next_rq
);
2339 static void cfq_reposition_rq_rb(struct cfq_queue
*cfqq
, struct request
*rq
)
2341 elv_rb_del(&cfqq
->sort_list
, rq
);
2342 cfqq
->queued
[rq_is_sync(rq
)]--;
2343 cfqg_stats_update_io_remove(RQ_CFQG(rq
), rq
->cmd_flags
);
2345 cfqg_stats_update_io_add(RQ_CFQG(rq
), cfqq
->cfqd
->serving_group
,
2349 static struct request
*
2350 cfq_find_rq_fmerge(struct cfq_data
*cfqd
, struct bio
*bio
)
2352 struct task_struct
*tsk
= current
;
2353 struct cfq_io_cq
*cic
;
2354 struct cfq_queue
*cfqq
;
2356 cic
= cfq_cic_lookup(cfqd
, tsk
->io_context
);
2360 cfqq
= cic_to_cfqq(cic
, cfq_bio_sync(bio
));
2362 return elv_rb_find(&cfqq
->sort_list
, bio_end_sector(bio
));
2367 static void cfq_activate_request(struct request_queue
*q
, struct request
*rq
)
2369 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
2371 cfqd
->rq_in_driver
++;
2372 cfq_log_cfqq(cfqd
, RQ_CFQQ(rq
), "activate rq, drv=%d",
2373 cfqd
->rq_in_driver
);
2375 cfqd
->last_position
= blk_rq_pos(rq
) + blk_rq_sectors(rq
);
2378 static void cfq_deactivate_request(struct request_queue
*q
, struct request
*rq
)
2380 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
2382 WARN_ON(!cfqd
->rq_in_driver
);
2383 cfqd
->rq_in_driver
--;
2384 cfq_log_cfqq(cfqd
, RQ_CFQQ(rq
), "deactivate rq, drv=%d",
2385 cfqd
->rq_in_driver
);
2388 static void cfq_remove_request(struct request
*rq
)
2390 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
2392 if (cfqq
->next_rq
== rq
)
2393 cfqq
->next_rq
= cfq_find_next_rq(cfqq
->cfqd
, cfqq
, rq
);
2395 list_del_init(&rq
->queuelist
);
2398 cfqq
->cfqd
->rq_queued
--;
2399 cfqg_stats_update_io_remove(RQ_CFQG(rq
), rq
->cmd_flags
);
2400 if (rq
->cmd_flags
& REQ_PRIO
) {
2401 WARN_ON(!cfqq
->prio_pending
);
2402 cfqq
->prio_pending
--;
2406 static int cfq_merge(struct request_queue
*q
, struct request
**req
,
2409 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
2410 struct request
*__rq
;
2412 __rq
= cfq_find_rq_fmerge(cfqd
, bio
);
2413 if (__rq
&& elv_rq_merge_ok(__rq
, bio
)) {
2415 return ELEVATOR_FRONT_MERGE
;
2418 return ELEVATOR_NO_MERGE
;
2421 static void cfq_merged_request(struct request_queue
*q
, struct request
*req
,
2424 if (type
== ELEVATOR_FRONT_MERGE
) {
2425 struct cfq_queue
*cfqq
= RQ_CFQQ(req
);
2427 cfq_reposition_rq_rb(cfqq
, req
);
2431 static void cfq_bio_merged(struct request_queue
*q
, struct request
*req
,
2434 cfqg_stats_update_io_merged(RQ_CFQG(req
), bio
->bi_rw
);
2438 cfq_merged_requests(struct request_queue
*q
, struct request
*rq
,
2439 struct request
*next
)
2441 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
2442 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
2445 * reposition in fifo if next is older than rq
2447 if (!list_empty(&rq
->queuelist
) && !list_empty(&next
->queuelist
) &&
2448 time_before(next
->fifo_time
, rq
->fifo_time
) &&
2449 cfqq
== RQ_CFQQ(next
)) {
2450 list_move(&rq
->queuelist
, &next
->queuelist
);
2451 rq
->fifo_time
= next
->fifo_time
;
2454 if (cfqq
->next_rq
== next
)
2456 cfq_remove_request(next
);
2457 cfqg_stats_update_io_merged(RQ_CFQG(rq
), next
->cmd_flags
);
2459 cfqq
= RQ_CFQQ(next
);
2461 * all requests of this queue are merged to other queues, delete it
2462 * from the service tree. If it's the active_queue,
2463 * cfq_dispatch_requests() will choose to expire it or do idle
2465 if (cfq_cfqq_on_rr(cfqq
) && RB_EMPTY_ROOT(&cfqq
->sort_list
) &&
2466 cfqq
!= cfqd
->active_queue
)
2467 cfq_del_cfqq_rr(cfqd
, cfqq
);
2470 static int cfq_allow_merge(struct request_queue
*q
, struct request
*rq
,
2473 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
2474 struct cfq_io_cq
*cic
;
2475 struct cfq_queue
*cfqq
;
2478 * Disallow merge of a sync bio into an async request.
2480 if (cfq_bio_sync(bio
) && !rq_is_sync(rq
))
2484 * Lookup the cfqq that this bio will be queued with and allow
2485 * merge only if rq is queued there.
2487 cic
= cfq_cic_lookup(cfqd
, current
->io_context
);
2491 cfqq
= cic_to_cfqq(cic
, cfq_bio_sync(bio
));
2492 return cfqq
== RQ_CFQQ(rq
);
2495 static inline void cfq_del_timer(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
2497 del_timer(&cfqd
->idle_slice_timer
);
2498 cfqg_stats_update_idle_time(cfqq
->cfqg
);
2501 static void __cfq_set_active_queue(struct cfq_data
*cfqd
,
2502 struct cfq_queue
*cfqq
)
2505 cfq_log_cfqq(cfqd
, cfqq
, "set_active wl_class:%d wl_type:%d",
2506 cfqd
->serving_wl_class
, cfqd
->serving_wl_type
);
2507 cfqg_stats_update_avg_queue_size(cfqq
->cfqg
);
2508 cfqq
->slice_start
= 0;
2509 cfqq
->dispatch_start
= jiffies
;
2510 cfqq
->allocated_slice
= 0;
2511 cfqq
->slice_end
= 0;
2512 cfqq
->slice_dispatch
= 0;
2513 cfqq
->nr_sectors
= 0;
2515 cfq_clear_cfqq_wait_request(cfqq
);
2516 cfq_clear_cfqq_must_dispatch(cfqq
);
2517 cfq_clear_cfqq_must_alloc_slice(cfqq
);
2518 cfq_clear_cfqq_fifo_expire(cfqq
);
2519 cfq_mark_cfqq_slice_new(cfqq
);
2521 cfq_del_timer(cfqd
, cfqq
);
2524 cfqd
->active_queue
= cfqq
;
2528 * current cfqq expired its slice (or was too idle), select new one
2531 __cfq_slice_expired(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
2534 cfq_log_cfqq(cfqd
, cfqq
, "slice expired t=%d", timed_out
);
2536 if (cfq_cfqq_wait_request(cfqq
))
2537 cfq_del_timer(cfqd
, cfqq
);
2539 cfq_clear_cfqq_wait_request(cfqq
);
2540 cfq_clear_cfqq_wait_busy(cfqq
);
2543 * If this cfqq is shared between multiple processes, check to
2544 * make sure that those processes are still issuing I/Os within
2545 * the mean seek distance. If not, it may be time to break the
2546 * queues apart again.
2548 if (cfq_cfqq_coop(cfqq
) && CFQQ_SEEKY(cfqq
))
2549 cfq_mark_cfqq_split_coop(cfqq
);
2552 * store what was left of this slice, if the queue idled/timed out
2555 if (cfq_cfqq_slice_new(cfqq
))
2556 cfqq
->slice_resid
= cfq_scaled_cfqq_slice(cfqd
, cfqq
);
2558 cfqq
->slice_resid
= cfqq
->slice_end
- jiffies
;
2559 cfq_log_cfqq(cfqd
, cfqq
, "resid=%ld", cfqq
->slice_resid
);
2562 cfq_group_served(cfqd
, cfqq
->cfqg
, cfqq
);
2564 if (cfq_cfqq_on_rr(cfqq
) && RB_EMPTY_ROOT(&cfqq
->sort_list
))
2565 cfq_del_cfqq_rr(cfqd
, cfqq
);
2567 cfq_resort_rr_list(cfqd
, cfqq
);
2569 if (cfqq
== cfqd
->active_queue
)
2570 cfqd
->active_queue
= NULL
;
2572 if (cfqd
->active_cic
) {
2573 put_io_context(cfqd
->active_cic
->icq
.ioc
);
2574 cfqd
->active_cic
= NULL
;
2578 static inline void cfq_slice_expired(struct cfq_data
*cfqd
, bool timed_out
)
2580 struct cfq_queue
*cfqq
= cfqd
->active_queue
;
2583 __cfq_slice_expired(cfqd
, cfqq
, timed_out
);
2587 * Get next queue for service. Unless we have a queue preemption,
2588 * we'll simply select the first cfqq in the service tree.
2590 static struct cfq_queue
*cfq_get_next_queue(struct cfq_data
*cfqd
)
2592 struct cfq_rb_root
*st
= st_for(cfqd
->serving_group
,
2593 cfqd
->serving_wl_class
, cfqd
->serving_wl_type
);
2595 if (!cfqd
->rq_queued
)
2598 /* There is nothing to dispatch */
2601 if (RB_EMPTY_ROOT(&st
->rb
))
2603 return cfq_rb_first(st
);
2606 static struct cfq_queue
*cfq_get_next_queue_forced(struct cfq_data
*cfqd
)
2608 struct cfq_group
*cfqg
;
2609 struct cfq_queue
*cfqq
;
2611 struct cfq_rb_root
*st
;
2613 if (!cfqd
->rq_queued
)
2616 cfqg
= cfq_get_next_cfqg(cfqd
);
2620 for_each_cfqg_st(cfqg
, i
, j
, st
)
2621 if ((cfqq
= cfq_rb_first(st
)) != NULL
)
2627 * Get and set a new active queue for service.
2629 static struct cfq_queue
*cfq_set_active_queue(struct cfq_data
*cfqd
,
2630 struct cfq_queue
*cfqq
)
2633 cfqq
= cfq_get_next_queue(cfqd
);
2635 __cfq_set_active_queue(cfqd
, cfqq
);
2639 static inline sector_t
cfq_dist_from_last(struct cfq_data
*cfqd
,
2642 if (blk_rq_pos(rq
) >= cfqd
->last_position
)
2643 return blk_rq_pos(rq
) - cfqd
->last_position
;
2645 return cfqd
->last_position
- blk_rq_pos(rq
);
2648 static inline int cfq_rq_close(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
2651 return cfq_dist_from_last(cfqd
, rq
) <= CFQQ_CLOSE_THR
;
2654 static struct cfq_queue
*cfqq_close(struct cfq_data
*cfqd
,
2655 struct cfq_queue
*cur_cfqq
)
2657 struct rb_root
*root
= &cfqd
->prio_trees
[cur_cfqq
->org_ioprio
];
2658 struct rb_node
*parent
, *node
;
2659 struct cfq_queue
*__cfqq
;
2660 sector_t sector
= cfqd
->last_position
;
2662 if (RB_EMPTY_ROOT(root
))
2666 * First, if we find a request starting at the end of the last
2667 * request, choose it.
2669 __cfqq
= cfq_prio_tree_lookup(cfqd
, root
, sector
, &parent
, NULL
);
2674 * If the exact sector wasn't found, the parent of the NULL leaf
2675 * will contain the closest sector.
2677 __cfqq
= rb_entry(parent
, struct cfq_queue
, p_node
);
2678 if (cfq_rq_close(cfqd
, cur_cfqq
, __cfqq
->next_rq
))
2681 if (blk_rq_pos(__cfqq
->next_rq
) < sector
)
2682 node
= rb_next(&__cfqq
->p_node
);
2684 node
= rb_prev(&__cfqq
->p_node
);
2688 __cfqq
= rb_entry(node
, struct cfq_queue
, p_node
);
2689 if (cfq_rq_close(cfqd
, cur_cfqq
, __cfqq
->next_rq
))
2697 * cur_cfqq - passed in so that we don't decide that the current queue is
2698 * closely cooperating with itself.
2700 * So, basically we're assuming that that cur_cfqq has dispatched at least
2701 * one request, and that cfqd->last_position reflects a position on the disk
2702 * associated with the I/O issued by cur_cfqq. I'm not sure this is a valid
2705 static struct cfq_queue
*cfq_close_cooperator(struct cfq_data
*cfqd
,
2706 struct cfq_queue
*cur_cfqq
)
2708 struct cfq_queue
*cfqq
;
2710 if (cfq_class_idle(cur_cfqq
))
2712 if (!cfq_cfqq_sync(cur_cfqq
))
2714 if (CFQQ_SEEKY(cur_cfqq
))
2718 * Don't search priority tree if it's the only queue in the group.
2720 if (cur_cfqq
->cfqg
->nr_cfqq
== 1)
2724 * We should notice if some of the queues are cooperating, eg
2725 * working closely on the same area of the disk. In that case,
2726 * we can group them together and don't waste time idling.
2728 cfqq
= cfqq_close(cfqd
, cur_cfqq
);
2732 /* If new queue belongs to different cfq_group, don't choose it */
2733 if (cur_cfqq
->cfqg
!= cfqq
->cfqg
)
2737 * It only makes sense to merge sync queues.
2739 if (!cfq_cfqq_sync(cfqq
))
2741 if (CFQQ_SEEKY(cfqq
))
2745 * Do not merge queues of different priority classes
2747 if (cfq_class_rt(cfqq
) != cfq_class_rt(cur_cfqq
))
2754 * Determine whether we should enforce idle window for this queue.
2757 static bool cfq_should_idle(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
2759 enum wl_class_t wl_class
= cfqq_class(cfqq
);
2760 struct cfq_rb_root
*st
= cfqq
->service_tree
;
2765 if (!cfqd
->cfq_slice_idle
)
2768 /* We never do for idle class queues. */
2769 if (wl_class
== IDLE_WORKLOAD
)
2772 /* We do for queues that were marked with idle window flag. */
2773 if (cfq_cfqq_idle_window(cfqq
) &&
2774 !(blk_queue_nonrot(cfqd
->queue
) && cfqd
->hw_tag
))
2778 * Otherwise, we do only if they are the last ones
2779 * in their service tree.
2781 if (st
->count
== 1 && cfq_cfqq_sync(cfqq
) &&
2782 !cfq_io_thinktime_big(cfqd
, &st
->ttime
, false))
2784 cfq_log_cfqq(cfqd
, cfqq
, "Not idling. st->count:%d", st
->count
);
2788 static void cfq_arm_slice_timer(struct cfq_data
*cfqd
)
2790 struct cfq_queue
*cfqq
= cfqd
->active_queue
;
2791 struct cfq_io_cq
*cic
;
2792 unsigned long sl
, group_idle
= 0;
2795 * SSD device without seek penalty, disable idling. But only do so
2796 * for devices that support queuing, otherwise we still have a problem
2797 * with sync vs async workloads.
2799 if (blk_queue_nonrot(cfqd
->queue
) && cfqd
->hw_tag
)
2802 WARN_ON(!RB_EMPTY_ROOT(&cfqq
->sort_list
));
2803 WARN_ON(cfq_cfqq_slice_new(cfqq
));
2806 * idle is disabled, either manually or by past process history
2808 if (!cfq_should_idle(cfqd
, cfqq
)) {
2809 /* no queue idling. Check for group idling */
2810 if (cfqd
->cfq_group_idle
)
2811 group_idle
= cfqd
->cfq_group_idle
;
2817 * still active requests from this queue, don't idle
2819 if (cfqq
->dispatched
)
2823 * task has exited, don't wait
2825 cic
= cfqd
->active_cic
;
2826 if (!cic
|| !atomic_read(&cic
->icq
.ioc
->active_ref
))
2830 * If our average think time is larger than the remaining time
2831 * slice, then don't idle. This avoids overrunning the allotted
2834 if (sample_valid(cic
->ttime
.ttime_samples
) &&
2835 (cfqq
->slice_end
- jiffies
< cic
->ttime
.ttime_mean
)) {
2836 cfq_log_cfqq(cfqd
, cfqq
, "Not idling. think_time:%lu",
2837 cic
->ttime
.ttime_mean
);
2841 /* There are other queues in the group, don't do group idle */
2842 if (group_idle
&& cfqq
->cfqg
->nr_cfqq
> 1)
2845 cfq_mark_cfqq_wait_request(cfqq
);
2848 sl
= cfqd
->cfq_group_idle
;
2850 sl
= cfqd
->cfq_slice_idle
;
2852 mod_timer(&cfqd
->idle_slice_timer
, jiffies
+ sl
);
2853 cfqg_stats_set_start_idle_time(cfqq
->cfqg
);
2854 cfq_log_cfqq(cfqd
, cfqq
, "arm_idle: %lu group_idle: %d", sl
,
2855 group_idle
? 1 : 0);
2859 * Move request from internal lists to the request queue dispatch list.
2861 static void cfq_dispatch_insert(struct request_queue
*q
, struct request
*rq
)
2863 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
2864 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
2866 cfq_log_cfqq(cfqd
, cfqq
, "dispatch_insert");
2868 cfqq
->next_rq
= cfq_find_next_rq(cfqd
, cfqq
, rq
);
2869 cfq_remove_request(rq
);
2871 (RQ_CFQG(rq
))->dispatched
++;
2872 elv_dispatch_sort(q
, rq
);
2874 cfqd
->rq_in_flight
[cfq_cfqq_sync(cfqq
)]++;
2875 cfqq
->nr_sectors
+= blk_rq_sectors(rq
);
2876 cfqg_stats_update_dispatch(cfqq
->cfqg
, blk_rq_bytes(rq
), rq
->cmd_flags
);
2880 * return expired entry, or NULL to just start from scratch in rbtree
2882 static struct request
*cfq_check_fifo(struct cfq_queue
*cfqq
)
2884 struct request
*rq
= NULL
;
2886 if (cfq_cfqq_fifo_expire(cfqq
))
2889 cfq_mark_cfqq_fifo_expire(cfqq
);
2891 if (list_empty(&cfqq
->fifo
))
2894 rq
= rq_entry_fifo(cfqq
->fifo
.next
);
2895 if (time_before(jiffies
, rq
->fifo_time
))
2898 cfq_log_cfqq(cfqq
->cfqd
, cfqq
, "fifo=%p", rq
);
2903 cfq_prio_to_maxrq(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
2905 const int base_rq
= cfqd
->cfq_slice_async_rq
;
2907 WARN_ON(cfqq
->ioprio
>= IOPRIO_BE_NR
);
2909 return 2 * base_rq
* (IOPRIO_BE_NR
- cfqq
->ioprio
);
2913 * Must be called with the queue_lock held.
2915 static int cfqq_process_refs(struct cfq_queue
*cfqq
)
2917 int process_refs
, io_refs
;
2919 io_refs
= cfqq
->allocated
[READ
] + cfqq
->allocated
[WRITE
];
2920 process_refs
= cfqq
->ref
- io_refs
;
2921 BUG_ON(process_refs
< 0);
2922 return process_refs
;
2925 static void cfq_setup_merge(struct cfq_queue
*cfqq
, struct cfq_queue
*new_cfqq
)
2927 int process_refs
, new_process_refs
;
2928 struct cfq_queue
*__cfqq
;
2931 * If there are no process references on the new_cfqq, then it is
2932 * unsafe to follow the ->new_cfqq chain as other cfqq's in the
2933 * chain may have dropped their last reference (not just their
2934 * last process reference).
2936 if (!cfqq_process_refs(new_cfqq
))
2939 /* Avoid a circular list and skip interim queue merges */
2940 while ((__cfqq
= new_cfqq
->new_cfqq
)) {
2946 process_refs
= cfqq_process_refs(cfqq
);
2947 new_process_refs
= cfqq_process_refs(new_cfqq
);
2949 * If the process for the cfqq has gone away, there is no
2950 * sense in merging the queues.
2952 if (process_refs
== 0 || new_process_refs
== 0)
2956 * Merge in the direction of the lesser amount of work.
2958 if (new_process_refs
>= process_refs
) {
2959 cfqq
->new_cfqq
= new_cfqq
;
2960 new_cfqq
->ref
+= process_refs
;
2962 new_cfqq
->new_cfqq
= cfqq
;
2963 cfqq
->ref
+= new_process_refs
;
2967 static enum wl_type_t
cfq_choose_wl_type(struct cfq_data
*cfqd
,
2968 struct cfq_group
*cfqg
, enum wl_class_t wl_class
)
2970 struct cfq_queue
*queue
;
2972 bool key_valid
= false;
2973 unsigned long lowest_key
= 0;
2974 enum wl_type_t cur_best
= SYNC_NOIDLE_WORKLOAD
;
2976 for (i
= 0; i
<= SYNC_WORKLOAD
; ++i
) {
2977 /* select the one with lowest rb_key */
2978 queue
= cfq_rb_first(st_for(cfqg
, wl_class
, i
));
2980 (!key_valid
|| time_before(queue
->rb_key
, lowest_key
))) {
2981 lowest_key
= queue
->rb_key
;
2991 choose_wl_class_and_type(struct cfq_data
*cfqd
, struct cfq_group
*cfqg
)
2995 struct cfq_rb_root
*st
;
2996 unsigned group_slice
;
2997 enum wl_class_t original_class
= cfqd
->serving_wl_class
;
2999 /* Choose next priority. RT > BE > IDLE */
3000 if (cfq_group_busy_queues_wl(RT_WORKLOAD
, cfqd
, cfqg
))
3001 cfqd
->serving_wl_class
= RT_WORKLOAD
;
3002 else if (cfq_group_busy_queues_wl(BE_WORKLOAD
, cfqd
, cfqg
))
3003 cfqd
->serving_wl_class
= BE_WORKLOAD
;
3005 cfqd
->serving_wl_class
= IDLE_WORKLOAD
;
3006 cfqd
->workload_expires
= jiffies
+ 1;
3010 if (original_class
!= cfqd
->serving_wl_class
)
3014 * For RT and BE, we have to choose also the type
3015 * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
3018 st
= st_for(cfqg
, cfqd
->serving_wl_class
, cfqd
->serving_wl_type
);
3022 * check workload expiration, and that we still have other queues ready
3024 if (count
&& !time_after(jiffies
, cfqd
->workload_expires
))
3028 /* otherwise select new workload type */
3029 cfqd
->serving_wl_type
= cfq_choose_wl_type(cfqd
, cfqg
,
3030 cfqd
->serving_wl_class
);
3031 st
= st_for(cfqg
, cfqd
->serving_wl_class
, cfqd
->serving_wl_type
);
3035 * the workload slice is computed as a fraction of target latency
3036 * proportional to the number of queues in that workload, over
3037 * all the queues in the same priority class
3039 group_slice
= cfq_group_slice(cfqd
, cfqg
);
3041 slice
= group_slice
* count
/
3042 max_t(unsigned, cfqg
->busy_queues_avg
[cfqd
->serving_wl_class
],
3043 cfq_group_busy_queues_wl(cfqd
->serving_wl_class
, cfqd
,
3046 if (cfqd
->serving_wl_type
== ASYNC_WORKLOAD
) {
3050 * Async queues are currently system wide. Just taking
3051 * proportion of queues with-in same group will lead to higher
3052 * async ratio system wide as generally root group is going
3053 * to have higher weight. A more accurate thing would be to
3054 * calculate system wide asnc/sync ratio.
3056 tmp
= cfqd
->cfq_target_latency
*
3057 cfqg_busy_async_queues(cfqd
, cfqg
);
3058 tmp
= tmp
/cfqd
->busy_queues
;
3059 slice
= min_t(unsigned, slice
, tmp
);
3061 /* async workload slice is scaled down according to
3062 * the sync/async slice ratio. */
3063 slice
= slice
* cfqd
->cfq_slice
[0] / cfqd
->cfq_slice
[1];
3065 /* sync workload slice is at least 2 * cfq_slice_idle */
3066 slice
= max(slice
, 2 * cfqd
->cfq_slice_idle
);
3068 slice
= max_t(unsigned, slice
, CFQ_MIN_TT
);
3069 cfq_log(cfqd
, "workload slice:%d", slice
);
3070 cfqd
->workload_expires
= jiffies
+ slice
;
3073 static struct cfq_group
*cfq_get_next_cfqg(struct cfq_data
*cfqd
)
3075 struct cfq_rb_root
*st
= &cfqd
->grp_service_tree
;
3076 struct cfq_group
*cfqg
;
3078 if (RB_EMPTY_ROOT(&st
->rb
))
3080 cfqg
= cfq_rb_first_group(st
);
3081 update_min_vdisktime(st
);
3085 static void cfq_choose_cfqg(struct cfq_data
*cfqd
)
3087 struct cfq_group
*cfqg
= cfq_get_next_cfqg(cfqd
);
3089 cfqd
->serving_group
= cfqg
;
3091 /* Restore the workload type data */
3092 if (cfqg
->saved_wl_slice
) {
3093 cfqd
->workload_expires
= jiffies
+ cfqg
->saved_wl_slice
;
3094 cfqd
->serving_wl_type
= cfqg
->saved_wl_type
;
3095 cfqd
->serving_wl_class
= cfqg
->saved_wl_class
;
3097 cfqd
->workload_expires
= jiffies
- 1;
3099 choose_wl_class_and_type(cfqd
, cfqg
);
3103 * Select a queue for service. If we have a current active queue,
3104 * check whether to continue servicing it, or retrieve and set a new one.
3106 static struct cfq_queue
*cfq_select_queue(struct cfq_data
*cfqd
)
3108 struct cfq_queue
*cfqq
, *new_cfqq
= NULL
;
3110 cfqq
= cfqd
->active_queue
;
3114 if (!cfqd
->rq_queued
)
3118 * We were waiting for group to get backlogged. Expire the queue
3120 if (cfq_cfqq_wait_busy(cfqq
) && !RB_EMPTY_ROOT(&cfqq
->sort_list
))
3124 * The active queue has run out of time, expire it and select new.
3126 if (cfq_slice_used(cfqq
) && !cfq_cfqq_must_dispatch(cfqq
)) {
3128 * If slice had not expired at the completion of last request
3129 * we might not have turned on wait_busy flag. Don't expire
3130 * the queue yet. Allow the group to get backlogged.
3132 * The very fact that we have used the slice, that means we
3133 * have been idling all along on this queue and it should be
3134 * ok to wait for this request to complete.
3136 if (cfqq
->cfqg
->nr_cfqq
== 1 && RB_EMPTY_ROOT(&cfqq
->sort_list
)
3137 && cfqq
->dispatched
&& cfq_should_idle(cfqd
, cfqq
)) {
3141 goto check_group_idle
;
3145 * The active queue has requests and isn't expired, allow it to
3148 if (!RB_EMPTY_ROOT(&cfqq
->sort_list
))
3152 * If another queue has a request waiting within our mean seek
3153 * distance, let it run. The expire code will check for close
3154 * cooperators and put the close queue at the front of the service
3155 * tree. If possible, merge the expiring queue with the new cfqq.
3157 new_cfqq
= cfq_close_cooperator(cfqd
, cfqq
);
3159 if (!cfqq
->new_cfqq
)
3160 cfq_setup_merge(cfqq
, new_cfqq
);
3165 * No requests pending. If the active queue still has requests in
3166 * flight or is idling for a new request, allow either of these
3167 * conditions to happen (or time out) before selecting a new queue.
3169 if (timer_pending(&cfqd
->idle_slice_timer
)) {
3175 * This is a deep seek queue, but the device is much faster than
3176 * the queue can deliver, don't idle
3178 if (CFQQ_SEEKY(cfqq
) && cfq_cfqq_idle_window(cfqq
) &&
3179 (cfq_cfqq_slice_new(cfqq
) ||
3180 (cfqq
->slice_end
- jiffies
> jiffies
- cfqq
->slice_start
))) {
3181 cfq_clear_cfqq_deep(cfqq
);
3182 cfq_clear_cfqq_idle_window(cfqq
);
3185 if (cfqq
->dispatched
&& cfq_should_idle(cfqd
, cfqq
)) {
3191 * If group idle is enabled and there are requests dispatched from
3192 * this group, wait for requests to complete.
3195 if (cfqd
->cfq_group_idle
&& cfqq
->cfqg
->nr_cfqq
== 1 &&
3196 cfqq
->cfqg
->dispatched
&&
3197 !cfq_io_thinktime_big(cfqd
, &cfqq
->cfqg
->ttime
, true)) {
3203 cfq_slice_expired(cfqd
, 0);
3206 * Current queue expired. Check if we have to switch to a new
3210 cfq_choose_cfqg(cfqd
);
3212 cfqq
= cfq_set_active_queue(cfqd
, new_cfqq
);
3217 static int __cfq_forced_dispatch_cfqq(struct cfq_queue
*cfqq
)
3221 while (cfqq
->next_rq
) {
3222 cfq_dispatch_insert(cfqq
->cfqd
->queue
, cfqq
->next_rq
);
3226 BUG_ON(!list_empty(&cfqq
->fifo
));
3228 /* By default cfqq is not expired if it is empty. Do it explicitly */
3229 __cfq_slice_expired(cfqq
->cfqd
, cfqq
, 0);
3234 * Drain our current requests. Used for barriers and when switching
3235 * io schedulers on-the-fly.
3237 static int cfq_forced_dispatch(struct cfq_data
*cfqd
)
3239 struct cfq_queue
*cfqq
;
3242 /* Expire the timeslice of the current active queue first */
3243 cfq_slice_expired(cfqd
, 0);
3244 while ((cfqq
= cfq_get_next_queue_forced(cfqd
)) != NULL
) {
3245 __cfq_set_active_queue(cfqd
, cfqq
);
3246 dispatched
+= __cfq_forced_dispatch_cfqq(cfqq
);
3249 BUG_ON(cfqd
->busy_queues
);
3251 cfq_log(cfqd
, "forced_dispatch=%d", dispatched
);
3255 static inline bool cfq_slice_used_soon(struct cfq_data
*cfqd
,
3256 struct cfq_queue
*cfqq
)
3258 /* the queue hasn't finished any request, can't estimate */
3259 if (cfq_cfqq_slice_new(cfqq
))
3261 if (time_after(jiffies
+ cfqd
->cfq_slice_idle
* cfqq
->dispatched
,
3268 static bool cfq_may_dispatch(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
3270 unsigned int max_dispatch
;
3273 * Drain async requests before we start sync IO
3275 if (cfq_should_idle(cfqd
, cfqq
) && cfqd
->rq_in_flight
[BLK_RW_ASYNC
])
3279 * If this is an async queue and we have sync IO in flight, let it wait
3281 if (cfqd
->rq_in_flight
[BLK_RW_SYNC
] && !cfq_cfqq_sync(cfqq
))
3284 max_dispatch
= max_t(unsigned int, cfqd
->cfq_quantum
/ 2, 1);
3285 if (cfq_class_idle(cfqq
))
3289 * Does this cfqq already have too much IO in flight?
3291 if (cfqq
->dispatched
>= max_dispatch
) {
3292 bool promote_sync
= false;
3294 * idle queue must always only have a single IO in flight
3296 if (cfq_class_idle(cfqq
))
3300 * If there is only one sync queue
3301 * we can ignore async queue here and give the sync
3302 * queue no dispatch limit. The reason is a sync queue can
3303 * preempt async queue, limiting the sync queue doesn't make
3304 * sense. This is useful for aiostress test.
3306 if (cfq_cfqq_sync(cfqq
) && cfqd
->busy_sync_queues
== 1)
3307 promote_sync
= true;
3310 * We have other queues, don't allow more IO from this one
3312 if (cfqd
->busy_queues
> 1 && cfq_slice_used_soon(cfqd
, cfqq
) &&
3317 * Sole queue user, no limit
3319 if (cfqd
->busy_queues
== 1 || promote_sync
)
3323 * Normally we start throttling cfqq when cfq_quantum/2
3324 * requests have been dispatched. But we can drive
3325 * deeper queue depths at the beginning of slice
3326 * subjected to upper limit of cfq_quantum.
3328 max_dispatch
= cfqd
->cfq_quantum
;
3332 * Async queues must wait a bit before being allowed dispatch.
3333 * We also ramp up the dispatch depth gradually for async IO,
3334 * based on the last sync IO we serviced
3336 if (!cfq_cfqq_sync(cfqq
) && cfqd
->cfq_latency
) {
3337 unsigned long last_sync
= jiffies
- cfqd
->last_delayed_sync
;
3340 depth
= last_sync
/ cfqd
->cfq_slice
[1];
3341 if (!depth
&& !cfqq
->dispatched
)
3343 if (depth
< max_dispatch
)
3344 max_dispatch
= depth
;
3348 * If we're below the current max, allow a dispatch
3350 return cfqq
->dispatched
< max_dispatch
;
3354 * Dispatch a request from cfqq, moving them to the request queue
3357 static bool cfq_dispatch_request(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
3361 BUG_ON(RB_EMPTY_ROOT(&cfqq
->sort_list
));
3363 if (!cfq_may_dispatch(cfqd
, cfqq
))
3367 * follow expired path, else get first next available
3369 rq
= cfq_check_fifo(cfqq
);
3374 * insert request into driver dispatch list
3376 cfq_dispatch_insert(cfqd
->queue
, rq
);
3378 if (!cfqd
->active_cic
) {
3379 struct cfq_io_cq
*cic
= RQ_CIC(rq
);
3381 atomic_long_inc(&cic
->icq
.ioc
->refcount
);
3382 cfqd
->active_cic
= cic
;
3389 * Find the cfqq that we need to service and move a request from that to the
3392 static int cfq_dispatch_requests(struct request_queue
*q
, int force
)
3394 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
3395 struct cfq_queue
*cfqq
;
3397 if (!cfqd
->busy_queues
)
3400 if (unlikely(force
))
3401 return cfq_forced_dispatch(cfqd
);
3403 cfqq
= cfq_select_queue(cfqd
);
3408 * Dispatch a request from this cfqq, if it is allowed
3410 if (!cfq_dispatch_request(cfqd
, cfqq
))
3413 cfqq
->slice_dispatch
++;
3414 cfq_clear_cfqq_must_dispatch(cfqq
);
3417 * expire an async queue immediately if it has used up its slice. idle
3418 * queue always expire after 1 dispatch round.
3420 if (cfqd
->busy_queues
> 1 && ((!cfq_cfqq_sync(cfqq
) &&
3421 cfqq
->slice_dispatch
>= cfq_prio_to_maxrq(cfqd
, cfqq
)) ||
3422 cfq_class_idle(cfqq
))) {
3423 cfqq
->slice_end
= jiffies
+ 1;
3424 cfq_slice_expired(cfqd
, 0);
3427 cfq_log_cfqq(cfqd
, cfqq
, "dispatched a request");
3432 * task holds one reference to the queue, dropped when task exits. each rq
3433 * in-flight on this queue also holds a reference, dropped when rq is freed.
3435 * Each cfq queue took a reference on the parent group. Drop it now.
3436 * queue lock must be held here.
3438 static void cfq_put_queue(struct cfq_queue
*cfqq
)
3440 struct cfq_data
*cfqd
= cfqq
->cfqd
;
3441 struct cfq_group
*cfqg
;
3443 BUG_ON(cfqq
->ref
<= 0);
3449 cfq_log_cfqq(cfqd
, cfqq
, "put_queue");
3450 BUG_ON(rb_first(&cfqq
->sort_list
));
3451 BUG_ON(cfqq
->allocated
[READ
] + cfqq
->allocated
[WRITE
]);
3454 if (unlikely(cfqd
->active_queue
== cfqq
)) {
3455 __cfq_slice_expired(cfqd
, cfqq
, 0);
3456 cfq_schedule_dispatch(cfqd
);
3459 BUG_ON(cfq_cfqq_on_rr(cfqq
));
3460 kmem_cache_free(cfq_pool
, cfqq
);
3464 static void cfq_put_cooperator(struct cfq_queue
*cfqq
)
3466 struct cfq_queue
*__cfqq
, *next
;
3469 * If this queue was scheduled to merge with another queue, be
3470 * sure to drop the reference taken on that queue (and others in
3471 * the merge chain). See cfq_setup_merge and cfq_merge_cfqqs.
3473 __cfqq
= cfqq
->new_cfqq
;
3475 if (__cfqq
== cfqq
) {
3476 WARN(1, "cfqq->new_cfqq loop detected\n");
3479 next
= __cfqq
->new_cfqq
;
3480 cfq_put_queue(__cfqq
);
3485 static void cfq_exit_cfqq(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
3487 if (unlikely(cfqq
== cfqd
->active_queue
)) {
3488 __cfq_slice_expired(cfqd
, cfqq
, 0);
3489 cfq_schedule_dispatch(cfqd
);
3492 cfq_put_cooperator(cfqq
);
3494 cfq_put_queue(cfqq
);
3497 static void cfq_init_icq(struct io_cq
*icq
)
3499 struct cfq_io_cq
*cic
= icq_to_cic(icq
);
3501 cic
->ttime
.last_end_request
= jiffies
;
3504 static void cfq_exit_icq(struct io_cq
*icq
)
3506 struct cfq_io_cq
*cic
= icq_to_cic(icq
);
3507 struct cfq_data
*cfqd
= cic_to_cfqd(cic
);
3509 if (cic
->cfqq
[BLK_RW_ASYNC
]) {
3510 cfq_exit_cfqq(cfqd
, cic
->cfqq
[BLK_RW_ASYNC
]);
3511 cic
->cfqq
[BLK_RW_ASYNC
] = NULL
;
3514 if (cic
->cfqq
[BLK_RW_SYNC
]) {
3515 cfq_exit_cfqq(cfqd
, cic
->cfqq
[BLK_RW_SYNC
]);
3516 cic
->cfqq
[BLK_RW_SYNC
] = NULL
;
3520 static void cfq_init_prio_data(struct cfq_queue
*cfqq
, struct cfq_io_cq
*cic
)
3522 struct task_struct
*tsk
= current
;
3525 if (!cfq_cfqq_prio_changed(cfqq
))
3528 ioprio_class
= IOPRIO_PRIO_CLASS(cic
->ioprio
);
3529 switch (ioprio_class
) {
3531 printk(KERN_ERR
"cfq: bad prio %x\n", ioprio_class
);
3532 case IOPRIO_CLASS_NONE
:
3534 * no prio set, inherit CPU scheduling settings
3536 cfqq
->ioprio
= task_nice_ioprio(tsk
);
3537 cfqq
->ioprio_class
= task_nice_ioclass(tsk
);
3539 case IOPRIO_CLASS_RT
:
3540 cfqq
->ioprio
= IOPRIO_PRIO_DATA(cic
->ioprio
);
3541 cfqq
->ioprio_class
= IOPRIO_CLASS_RT
;
3543 case IOPRIO_CLASS_BE
:
3544 cfqq
->ioprio
= IOPRIO_PRIO_DATA(cic
->ioprio
);
3545 cfqq
->ioprio_class
= IOPRIO_CLASS_BE
;
3547 case IOPRIO_CLASS_IDLE
:
3548 cfqq
->ioprio_class
= IOPRIO_CLASS_IDLE
;
3550 cfq_clear_cfqq_idle_window(cfqq
);
3555 * keep track of original prio settings in case we have to temporarily
3556 * elevate the priority of this queue
3558 cfqq
->org_ioprio
= cfqq
->ioprio
;
3559 cfq_clear_cfqq_prio_changed(cfqq
);
3562 static void check_ioprio_changed(struct cfq_io_cq
*cic
, struct bio
*bio
)
3564 int ioprio
= cic
->icq
.ioc
->ioprio
;
3565 struct cfq_data
*cfqd
= cic_to_cfqd(cic
);
3566 struct cfq_queue
*cfqq
;
3569 * Check whether ioprio has changed. The condition may trigger
3570 * spuriously on a newly created cic but there's no harm.
3572 if (unlikely(!cfqd
) || likely(cic
->ioprio
== ioprio
))
3575 cfqq
= cic
->cfqq
[BLK_RW_ASYNC
];
3577 struct cfq_queue
*new_cfqq
;
3578 new_cfqq
= cfq_get_queue(cfqd
, BLK_RW_ASYNC
, cic
, bio
,
3581 cic
->cfqq
[BLK_RW_ASYNC
] = new_cfqq
;
3582 cfq_put_queue(cfqq
);
3586 cfqq
= cic
->cfqq
[BLK_RW_SYNC
];
3588 cfq_mark_cfqq_prio_changed(cfqq
);
3590 cic
->ioprio
= ioprio
;
3593 static void cfq_init_cfqq(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
3594 pid_t pid
, bool is_sync
)
3596 RB_CLEAR_NODE(&cfqq
->rb_node
);
3597 RB_CLEAR_NODE(&cfqq
->p_node
);
3598 INIT_LIST_HEAD(&cfqq
->fifo
);
3603 cfq_mark_cfqq_prio_changed(cfqq
);
3606 if (!cfq_class_idle(cfqq
))
3607 cfq_mark_cfqq_idle_window(cfqq
);
3608 cfq_mark_cfqq_sync(cfqq
);
3613 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3614 static void check_blkcg_changed(struct cfq_io_cq
*cic
, struct bio
*bio
)
3616 struct cfq_data
*cfqd
= cic_to_cfqd(cic
);
3617 struct cfq_queue
*sync_cfqq
;
3621 serial_nr
= bio_blkcg(bio
)->css
.serial_nr
;
3625 * Check whether blkcg has changed. The condition may trigger
3626 * spuriously on a newly created cic but there's no harm.
3628 if (unlikely(!cfqd
) || likely(cic
->blkcg_serial_nr
== serial_nr
))
3631 sync_cfqq
= cic_to_cfqq(cic
, 1);
3634 * Drop reference to sync queue. A new sync queue will be
3635 * assigned in new group upon arrival of a fresh request.
3637 cfq_log_cfqq(cfqd
, sync_cfqq
, "changed cgroup");
3638 cic_set_cfqq(cic
, NULL
, 1);
3639 cfq_put_queue(sync_cfqq
);
3642 cic
->blkcg_serial_nr
= serial_nr
;
3645 static inline void check_blkcg_changed(struct cfq_io_cq
*cic
, struct bio
*bio
) { }
3646 #endif /* CONFIG_CFQ_GROUP_IOSCHED */
3648 static struct cfq_queue
*
3649 cfq_find_alloc_queue(struct cfq_data
*cfqd
, bool is_sync
, struct cfq_io_cq
*cic
,
3650 struct bio
*bio
, gfp_t gfp_mask
)
3652 struct blkcg
*blkcg
;
3653 struct cfq_queue
*cfqq
, *new_cfqq
= NULL
;
3654 struct cfq_group
*cfqg
;
3659 blkcg
= bio_blkcg(bio
);
3660 cfqg
= cfq_lookup_create_cfqg(cfqd
, blkcg
);
3662 cfqq
= &cfqd
->oom_cfqq
;
3666 cfqq
= cic_to_cfqq(cic
, is_sync
);
3669 * Always try a new alloc if we fell back to the OOM cfqq
3670 * originally, since it should just be a temporary situation.
3672 if (!cfqq
|| cfqq
== &cfqd
->oom_cfqq
) {
3677 } else if (gfp_mask
& __GFP_WAIT
) {
3679 spin_unlock_irq(cfqd
->queue
->queue_lock
);
3680 new_cfqq
= kmem_cache_alloc_node(cfq_pool
,
3681 gfp_mask
| __GFP_ZERO
,
3683 spin_lock_irq(cfqd
->queue
->queue_lock
);
3687 return &cfqd
->oom_cfqq
;
3689 cfqq
= kmem_cache_alloc_node(cfq_pool
,
3690 gfp_mask
| __GFP_ZERO
,
3695 cfq_init_cfqq(cfqd
, cfqq
, current
->pid
, is_sync
);
3696 cfq_init_prio_data(cfqq
, cic
);
3697 cfq_link_cfqq_cfqg(cfqq
, cfqg
);
3698 cfq_log_cfqq(cfqd
, cfqq
, "alloced");
3700 cfqq
= &cfqd
->oom_cfqq
;
3704 kmem_cache_free(cfq_pool
, new_cfqq
);
3710 static struct cfq_queue
**
3711 cfq_async_queue_prio(struct cfq_data
*cfqd
, int ioprio_class
, int ioprio
)
3713 switch (ioprio_class
) {
3714 case IOPRIO_CLASS_RT
:
3715 return &cfqd
->async_cfqq
[0][ioprio
];
3716 case IOPRIO_CLASS_NONE
:
3717 ioprio
= IOPRIO_NORM
;
3719 case IOPRIO_CLASS_BE
:
3720 return &cfqd
->async_cfqq
[1][ioprio
];
3721 case IOPRIO_CLASS_IDLE
:
3722 return &cfqd
->async_idle_cfqq
;
3728 static struct cfq_queue
*
3729 cfq_get_queue(struct cfq_data
*cfqd
, bool is_sync
, struct cfq_io_cq
*cic
,
3730 struct bio
*bio
, gfp_t gfp_mask
)
3732 int ioprio_class
= IOPRIO_PRIO_CLASS(cic
->ioprio
);
3733 int ioprio
= IOPRIO_PRIO_DATA(cic
->ioprio
);
3734 struct cfq_queue
**async_cfqq
= NULL
;
3735 struct cfq_queue
*cfqq
= NULL
;
3738 if (!ioprio_valid(cic
->ioprio
)) {
3739 struct task_struct
*tsk
= current
;
3740 ioprio
= task_nice_ioprio(tsk
);
3741 ioprio_class
= task_nice_ioclass(tsk
);
3743 async_cfqq
= cfq_async_queue_prio(cfqd
, ioprio_class
, ioprio
);
3748 cfqq
= cfq_find_alloc_queue(cfqd
, is_sync
, cic
, bio
, gfp_mask
);
3751 * pin the queue now that it's allocated, scheduler exit will prune it
3753 if (!is_sync
&& !(*async_cfqq
)) {
3763 __cfq_update_io_thinktime(struct cfq_ttime
*ttime
, unsigned long slice_idle
)
3765 unsigned long elapsed
= jiffies
- ttime
->last_end_request
;
3766 elapsed
= min(elapsed
, 2UL * slice_idle
);
3768 ttime
->ttime_samples
= (7*ttime
->ttime_samples
+ 256) / 8;
3769 ttime
->ttime_total
= (7*ttime
->ttime_total
+ 256*elapsed
) / 8;
3770 ttime
->ttime_mean
= (ttime
->ttime_total
+ 128) / ttime
->ttime_samples
;
3774 cfq_update_io_thinktime(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
3775 struct cfq_io_cq
*cic
)
3777 if (cfq_cfqq_sync(cfqq
)) {
3778 __cfq_update_io_thinktime(&cic
->ttime
, cfqd
->cfq_slice_idle
);
3779 __cfq_update_io_thinktime(&cfqq
->service_tree
->ttime
,
3780 cfqd
->cfq_slice_idle
);
3782 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3783 __cfq_update_io_thinktime(&cfqq
->cfqg
->ttime
, cfqd
->cfq_group_idle
);
3788 cfq_update_io_seektime(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
3792 sector_t n_sec
= blk_rq_sectors(rq
);
3793 if (cfqq
->last_request_pos
) {
3794 if (cfqq
->last_request_pos
< blk_rq_pos(rq
))
3795 sdist
= blk_rq_pos(rq
) - cfqq
->last_request_pos
;
3797 sdist
= cfqq
->last_request_pos
- blk_rq_pos(rq
);
3800 cfqq
->seek_history
<<= 1;
3801 if (blk_queue_nonrot(cfqd
->queue
))
3802 cfqq
->seek_history
|= (n_sec
< CFQQ_SECT_THR_NONROT
);
3804 cfqq
->seek_history
|= (sdist
> CFQQ_SEEK_THR
);
3808 * Disable idle window if the process thinks too long or seeks so much that
3812 cfq_update_idle_window(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
3813 struct cfq_io_cq
*cic
)
3815 int old_idle
, enable_idle
;
3818 * Don't idle for async or idle io prio class
3820 if (!cfq_cfqq_sync(cfqq
) || cfq_class_idle(cfqq
))
3823 enable_idle
= old_idle
= cfq_cfqq_idle_window(cfqq
);
3825 if (cfqq
->queued
[0] + cfqq
->queued
[1] >= 4)
3826 cfq_mark_cfqq_deep(cfqq
);
3828 if (cfqq
->next_rq
&& (cfqq
->next_rq
->cmd_flags
& REQ_NOIDLE
))
3830 else if (!atomic_read(&cic
->icq
.ioc
->active_ref
) ||
3831 !cfqd
->cfq_slice_idle
||
3832 (!cfq_cfqq_deep(cfqq
) && CFQQ_SEEKY(cfqq
)))
3834 else if (sample_valid(cic
->ttime
.ttime_samples
)) {
3835 if (cic
->ttime
.ttime_mean
> cfqd
->cfq_slice_idle
)
3841 if (old_idle
!= enable_idle
) {
3842 cfq_log_cfqq(cfqd
, cfqq
, "idle=%d", enable_idle
);
3844 cfq_mark_cfqq_idle_window(cfqq
);
3846 cfq_clear_cfqq_idle_window(cfqq
);
3851 * Check if new_cfqq should preempt the currently active queue. Return 0 for
3852 * no or if we aren't sure, a 1 will cause a preempt.
3855 cfq_should_preempt(struct cfq_data
*cfqd
, struct cfq_queue
*new_cfqq
,
3858 struct cfq_queue
*cfqq
;
3860 cfqq
= cfqd
->active_queue
;
3864 if (cfq_class_idle(new_cfqq
))
3867 if (cfq_class_idle(cfqq
))
3871 * Don't allow a non-RT request to preempt an ongoing RT cfqq timeslice.
3873 if (cfq_class_rt(cfqq
) && !cfq_class_rt(new_cfqq
))
3877 * if the new request is sync, but the currently running queue is
3878 * not, let the sync request have priority.
3880 if (rq_is_sync(rq
) && !cfq_cfqq_sync(cfqq
))
3883 if (new_cfqq
->cfqg
!= cfqq
->cfqg
)
3886 if (cfq_slice_used(cfqq
))
3889 /* Allow preemption only if we are idling on sync-noidle tree */
3890 if (cfqd
->serving_wl_type
== SYNC_NOIDLE_WORKLOAD
&&
3891 cfqq_type(new_cfqq
) == SYNC_NOIDLE_WORKLOAD
&&
3892 new_cfqq
->service_tree
->count
== 2 &&
3893 RB_EMPTY_ROOT(&cfqq
->sort_list
))
3897 * So both queues are sync. Let the new request get disk time if
3898 * it's a metadata request and the current queue is doing regular IO.
3900 if ((rq
->cmd_flags
& REQ_PRIO
) && !cfqq
->prio_pending
)
3904 * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
3906 if (cfq_class_rt(new_cfqq
) && !cfq_class_rt(cfqq
))
3909 /* An idle queue should not be idle now for some reason */
3910 if (RB_EMPTY_ROOT(&cfqq
->sort_list
) && !cfq_should_idle(cfqd
, cfqq
))
3913 if (!cfqd
->active_cic
|| !cfq_cfqq_wait_request(cfqq
))
3917 * if this request is as-good as one we would expect from the
3918 * current cfqq, let it preempt
3920 if (cfq_rq_close(cfqd
, cfqq
, rq
))
3927 * cfqq preempts the active queue. if we allowed preempt with no slice left,
3928 * let it have half of its nominal slice.
3930 static void cfq_preempt_queue(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
3932 enum wl_type_t old_type
= cfqq_type(cfqd
->active_queue
);
3934 cfq_log_cfqq(cfqd
, cfqq
, "preempt");
3935 cfq_slice_expired(cfqd
, 1);
3938 * workload type is changed, don't save slice, otherwise preempt
3941 if (old_type
!= cfqq_type(cfqq
))
3942 cfqq
->cfqg
->saved_wl_slice
= 0;
3945 * Put the new queue at the front of the of the current list,
3946 * so we know that it will be selected next.
3948 BUG_ON(!cfq_cfqq_on_rr(cfqq
));
3950 cfq_service_tree_add(cfqd
, cfqq
, 1);
3952 cfqq
->slice_end
= 0;
3953 cfq_mark_cfqq_slice_new(cfqq
);
3957 * Called when a new fs request (rq) is added (to cfqq). Check if there's
3958 * something we should do about it
3961 cfq_rq_enqueued(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
3964 struct cfq_io_cq
*cic
= RQ_CIC(rq
);
3967 if (rq
->cmd_flags
& REQ_PRIO
)
3968 cfqq
->prio_pending
++;
3970 cfq_update_io_thinktime(cfqd
, cfqq
, cic
);
3971 cfq_update_io_seektime(cfqd
, cfqq
, rq
);
3972 cfq_update_idle_window(cfqd
, cfqq
, cic
);
3974 cfqq
->last_request_pos
= blk_rq_pos(rq
) + blk_rq_sectors(rq
);
3976 if (cfqq
== cfqd
->active_queue
) {
3978 * Remember that we saw a request from this process, but
3979 * don't start queuing just yet. Otherwise we risk seeing lots
3980 * of tiny requests, because we disrupt the normal plugging
3981 * and merging. If the request is already larger than a single
3982 * page, let it rip immediately. For that case we assume that
3983 * merging is already done. Ditto for a busy system that
3984 * has other work pending, don't risk delaying until the
3985 * idle timer unplug to continue working.
3987 if (cfq_cfqq_wait_request(cfqq
)) {
3988 if (blk_rq_bytes(rq
) > PAGE_CACHE_SIZE
||
3989 cfqd
->busy_queues
> 1) {
3990 cfq_del_timer(cfqd
, cfqq
);
3991 cfq_clear_cfqq_wait_request(cfqq
);
3992 __blk_run_queue(cfqd
->queue
);
3994 cfqg_stats_update_idle_time(cfqq
->cfqg
);
3995 cfq_mark_cfqq_must_dispatch(cfqq
);
3998 } else if (cfq_should_preempt(cfqd
, cfqq
, rq
)) {
4000 * not the active queue - expire current slice if it is
4001 * idle and has expired it's mean thinktime or this new queue
4002 * has some old slice time left and is of higher priority or
4003 * this new queue is RT and the current one is BE
4005 cfq_preempt_queue(cfqd
, cfqq
);
4006 __blk_run_queue(cfqd
->queue
);
4010 static void cfq_insert_request(struct request_queue
*q
, struct request
*rq
)
4012 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
4013 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
4015 cfq_log_cfqq(cfqd
, cfqq
, "insert_request");
4016 cfq_init_prio_data(cfqq
, RQ_CIC(rq
));
4018 rq
->fifo_time
= jiffies
+ cfqd
->cfq_fifo_expire
[rq_is_sync(rq
)];
4019 list_add_tail(&rq
->queuelist
, &cfqq
->fifo
);
4021 cfqg_stats_update_io_add(RQ_CFQG(rq
), cfqd
->serving_group
,
4023 cfq_rq_enqueued(cfqd
, cfqq
, rq
);
4027 * Update hw_tag based on peak queue depth over 50 samples under
4030 static void cfq_update_hw_tag(struct cfq_data
*cfqd
)
4032 struct cfq_queue
*cfqq
= cfqd
->active_queue
;
4034 if (cfqd
->rq_in_driver
> cfqd
->hw_tag_est_depth
)
4035 cfqd
->hw_tag_est_depth
= cfqd
->rq_in_driver
;
4037 if (cfqd
->hw_tag
== 1)
4040 if (cfqd
->rq_queued
<= CFQ_HW_QUEUE_MIN
&&
4041 cfqd
->rq_in_driver
<= CFQ_HW_QUEUE_MIN
)
4045 * If active queue hasn't enough requests and can idle, cfq might not
4046 * dispatch sufficient requests to hardware. Don't zero hw_tag in this
4049 if (cfqq
&& cfq_cfqq_idle_window(cfqq
) &&
4050 cfqq
->dispatched
+ cfqq
->queued
[0] + cfqq
->queued
[1] <
4051 CFQ_HW_QUEUE_MIN
&& cfqd
->rq_in_driver
< CFQ_HW_QUEUE_MIN
)
4054 if (cfqd
->hw_tag_samples
++ < 50)
4057 if (cfqd
->hw_tag_est_depth
>= CFQ_HW_QUEUE_MIN
)
4063 static bool cfq_should_wait_busy(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
4065 struct cfq_io_cq
*cic
= cfqd
->active_cic
;
4067 /* If the queue already has requests, don't wait */
4068 if (!RB_EMPTY_ROOT(&cfqq
->sort_list
))
4071 /* If there are other queues in the group, don't wait */
4072 if (cfqq
->cfqg
->nr_cfqq
> 1)
4075 /* the only queue in the group, but think time is big */
4076 if (cfq_io_thinktime_big(cfqd
, &cfqq
->cfqg
->ttime
, true))
4079 if (cfq_slice_used(cfqq
))
4082 /* if slice left is less than think time, wait busy */
4083 if (cic
&& sample_valid(cic
->ttime
.ttime_samples
)
4084 && (cfqq
->slice_end
- jiffies
< cic
->ttime
.ttime_mean
))
4088 * If think times is less than a jiffy than ttime_mean=0 and above
4089 * will not be true. It might happen that slice has not expired yet
4090 * but will expire soon (4-5 ns) during select_queue(). To cover the
4091 * case where think time is less than a jiffy, mark the queue wait
4092 * busy if only 1 jiffy is left in the slice.
4094 if (cfqq
->slice_end
- jiffies
== 1)
4100 static void cfq_completed_request(struct request_queue
*q
, struct request
*rq
)
4102 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
4103 struct cfq_data
*cfqd
= cfqq
->cfqd
;
4104 const int sync
= rq_is_sync(rq
);
4108 cfq_log_cfqq(cfqd
, cfqq
, "complete rqnoidle %d",
4109 !!(rq
->cmd_flags
& REQ_NOIDLE
));
4111 cfq_update_hw_tag(cfqd
);
4113 WARN_ON(!cfqd
->rq_in_driver
);
4114 WARN_ON(!cfqq
->dispatched
);
4115 cfqd
->rq_in_driver
--;
4117 (RQ_CFQG(rq
))->dispatched
--;
4118 cfqg_stats_update_completion(cfqq
->cfqg
, rq_start_time_ns(rq
),
4119 rq_io_start_time_ns(rq
), rq
->cmd_flags
);
4121 cfqd
->rq_in_flight
[cfq_cfqq_sync(cfqq
)]--;
4124 struct cfq_rb_root
*st
;
4126 RQ_CIC(rq
)->ttime
.last_end_request
= now
;
4128 if (cfq_cfqq_on_rr(cfqq
))
4129 st
= cfqq
->service_tree
;
4131 st
= st_for(cfqq
->cfqg
, cfqq_class(cfqq
),
4134 st
->ttime
.last_end_request
= now
;
4135 if (!time_after(rq
->start_time
+ cfqd
->cfq_fifo_expire
[1], now
))
4136 cfqd
->last_delayed_sync
= now
;
4139 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4140 cfqq
->cfqg
->ttime
.last_end_request
= now
;
4144 * If this is the active queue, check if it needs to be expired,
4145 * or if we want to idle in case it has no pending requests.
4147 if (cfqd
->active_queue
== cfqq
) {
4148 const bool cfqq_empty
= RB_EMPTY_ROOT(&cfqq
->sort_list
);
4150 if (cfq_cfqq_slice_new(cfqq
)) {
4151 cfq_set_prio_slice(cfqd
, cfqq
);
4152 cfq_clear_cfqq_slice_new(cfqq
);
4156 * Should we wait for next request to come in before we expire
4159 if (cfq_should_wait_busy(cfqd
, cfqq
)) {
4160 unsigned long extend_sl
= cfqd
->cfq_slice_idle
;
4161 if (!cfqd
->cfq_slice_idle
)
4162 extend_sl
= cfqd
->cfq_group_idle
;
4163 cfqq
->slice_end
= jiffies
+ extend_sl
;
4164 cfq_mark_cfqq_wait_busy(cfqq
);
4165 cfq_log_cfqq(cfqd
, cfqq
, "will busy wait");
4169 * Idling is not enabled on:
4171 * - idle-priority queues
4173 * - queues with still some requests queued
4174 * - when there is a close cooperator
4176 if (cfq_slice_used(cfqq
) || cfq_class_idle(cfqq
))
4177 cfq_slice_expired(cfqd
, 1);
4178 else if (sync
&& cfqq_empty
&&
4179 !cfq_close_cooperator(cfqd
, cfqq
)) {
4180 cfq_arm_slice_timer(cfqd
);
4184 if (!cfqd
->rq_in_driver
)
4185 cfq_schedule_dispatch(cfqd
);
4188 static inline int __cfq_may_queue(struct cfq_queue
*cfqq
)
4190 if (cfq_cfqq_wait_request(cfqq
) && !cfq_cfqq_must_alloc_slice(cfqq
)) {
4191 cfq_mark_cfqq_must_alloc_slice(cfqq
);
4192 return ELV_MQUEUE_MUST
;
4195 return ELV_MQUEUE_MAY
;
4198 static int cfq_may_queue(struct request_queue
*q
, int rw
)
4200 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
4201 struct task_struct
*tsk
= current
;
4202 struct cfq_io_cq
*cic
;
4203 struct cfq_queue
*cfqq
;
4206 * don't force setup of a queue from here, as a call to may_queue
4207 * does not necessarily imply that a request actually will be queued.
4208 * so just lookup a possibly existing queue, or return 'may queue'
4211 cic
= cfq_cic_lookup(cfqd
, tsk
->io_context
);
4213 return ELV_MQUEUE_MAY
;
4215 cfqq
= cic_to_cfqq(cic
, rw_is_sync(rw
));
4217 cfq_init_prio_data(cfqq
, cic
);
4219 return __cfq_may_queue(cfqq
);
4222 return ELV_MQUEUE_MAY
;
4226 * queue lock held here
4228 static void cfq_put_request(struct request
*rq
)
4230 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
4233 const int rw
= rq_data_dir(rq
);
4235 BUG_ON(!cfqq
->allocated
[rw
]);
4236 cfqq
->allocated
[rw
]--;
4238 /* Put down rq reference on cfqg */
4239 cfqg_put(RQ_CFQG(rq
));
4240 rq
->elv
.priv
[0] = NULL
;
4241 rq
->elv
.priv
[1] = NULL
;
4243 cfq_put_queue(cfqq
);
4247 static struct cfq_queue
*
4248 cfq_merge_cfqqs(struct cfq_data
*cfqd
, struct cfq_io_cq
*cic
,
4249 struct cfq_queue
*cfqq
)
4251 cfq_log_cfqq(cfqd
, cfqq
, "merging with queue %p", cfqq
->new_cfqq
);
4252 cic_set_cfqq(cic
, cfqq
->new_cfqq
, 1);
4253 cfq_mark_cfqq_coop(cfqq
->new_cfqq
);
4254 cfq_put_queue(cfqq
);
4255 return cic_to_cfqq(cic
, 1);
4259 * Returns NULL if a new cfqq should be allocated, or the old cfqq if this
4260 * was the last process referring to said cfqq.
4262 static struct cfq_queue
*
4263 split_cfqq(struct cfq_io_cq
*cic
, struct cfq_queue
*cfqq
)
4265 if (cfqq_process_refs(cfqq
) == 1) {
4266 cfqq
->pid
= current
->pid
;
4267 cfq_clear_cfqq_coop(cfqq
);
4268 cfq_clear_cfqq_split_coop(cfqq
);
4272 cic_set_cfqq(cic
, NULL
, 1);
4274 cfq_put_cooperator(cfqq
);
4276 cfq_put_queue(cfqq
);
4280 * Allocate cfq data structures associated with this request.
4283 cfq_set_request(struct request_queue
*q
, struct request
*rq
, struct bio
*bio
,
4286 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
4287 struct cfq_io_cq
*cic
= icq_to_cic(rq
->elv
.icq
);
4288 const int rw
= rq_data_dir(rq
);
4289 const bool is_sync
= rq_is_sync(rq
);
4290 struct cfq_queue
*cfqq
;
4292 might_sleep_if(gfp_mask
& __GFP_WAIT
);
4294 spin_lock_irq(q
->queue_lock
);
4296 check_ioprio_changed(cic
, bio
);
4297 check_blkcg_changed(cic
, bio
);
4299 cfqq
= cic_to_cfqq(cic
, is_sync
);
4300 if (!cfqq
|| cfqq
== &cfqd
->oom_cfqq
) {
4301 cfqq
= cfq_get_queue(cfqd
, is_sync
, cic
, bio
, gfp_mask
);
4302 cic_set_cfqq(cic
, cfqq
, is_sync
);
4305 * If the queue was seeky for too long, break it apart.
4307 if (cfq_cfqq_coop(cfqq
) && cfq_cfqq_split_coop(cfqq
)) {
4308 cfq_log_cfqq(cfqd
, cfqq
, "breaking apart cfqq");
4309 cfqq
= split_cfqq(cic
, cfqq
);
4315 * Check to see if this queue is scheduled to merge with
4316 * another, closely cooperating queue. The merging of
4317 * queues happens here as it must be done in process context.
4318 * The reference on new_cfqq was taken in merge_cfqqs.
4321 cfqq
= cfq_merge_cfqqs(cfqd
, cic
, cfqq
);
4324 cfqq
->allocated
[rw
]++;
4327 cfqg_get(cfqq
->cfqg
);
4328 rq
->elv
.priv
[0] = cfqq
;
4329 rq
->elv
.priv
[1] = cfqq
->cfqg
;
4330 spin_unlock_irq(q
->queue_lock
);
4334 static void cfq_kick_queue(struct work_struct
*work
)
4336 struct cfq_data
*cfqd
=
4337 container_of(work
, struct cfq_data
, unplug_work
);
4338 struct request_queue
*q
= cfqd
->queue
;
4340 spin_lock_irq(q
->queue_lock
);
4341 __blk_run_queue(cfqd
->queue
);
4342 spin_unlock_irq(q
->queue_lock
);
4346 * Timer running if the active_queue is currently idling inside its time slice
4348 static void cfq_idle_slice_timer(unsigned long data
)
4350 struct cfq_data
*cfqd
= (struct cfq_data
*) data
;
4351 struct cfq_queue
*cfqq
;
4352 unsigned long flags
;
4355 cfq_log(cfqd
, "idle timer fired");
4357 spin_lock_irqsave(cfqd
->queue
->queue_lock
, flags
);
4359 cfqq
= cfqd
->active_queue
;
4364 * We saw a request before the queue expired, let it through
4366 if (cfq_cfqq_must_dispatch(cfqq
))
4372 if (cfq_slice_used(cfqq
))
4376 * only expire and reinvoke request handler, if there are
4377 * other queues with pending requests
4379 if (!cfqd
->busy_queues
)
4383 * not expired and it has a request pending, let it dispatch
4385 if (!RB_EMPTY_ROOT(&cfqq
->sort_list
))
4389 * Queue depth flag is reset only when the idle didn't succeed
4391 cfq_clear_cfqq_deep(cfqq
);
4394 cfq_slice_expired(cfqd
, timed_out
);
4396 cfq_schedule_dispatch(cfqd
);
4398 spin_unlock_irqrestore(cfqd
->queue
->queue_lock
, flags
);
4401 static void cfq_shutdown_timer_wq(struct cfq_data
*cfqd
)
4403 del_timer_sync(&cfqd
->idle_slice_timer
);
4404 cancel_work_sync(&cfqd
->unplug_work
);
4407 static void cfq_put_async_queues(struct cfq_data
*cfqd
)
4411 for (i
= 0; i
< IOPRIO_BE_NR
; i
++) {
4412 if (cfqd
->async_cfqq
[0][i
])
4413 cfq_put_queue(cfqd
->async_cfqq
[0][i
]);
4414 if (cfqd
->async_cfqq
[1][i
])
4415 cfq_put_queue(cfqd
->async_cfqq
[1][i
]);
4418 if (cfqd
->async_idle_cfqq
)
4419 cfq_put_queue(cfqd
->async_idle_cfqq
);
4422 static void cfq_exit_queue(struct elevator_queue
*e
)
4424 struct cfq_data
*cfqd
= e
->elevator_data
;
4425 struct request_queue
*q
= cfqd
->queue
;
4427 cfq_shutdown_timer_wq(cfqd
);
4429 spin_lock_irq(q
->queue_lock
);
4431 if (cfqd
->active_queue
)
4432 __cfq_slice_expired(cfqd
, cfqd
->active_queue
, 0);
4434 cfq_put_async_queues(cfqd
);
4436 spin_unlock_irq(q
->queue_lock
);
4438 cfq_shutdown_timer_wq(cfqd
);
4440 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4441 blkcg_deactivate_policy(q
, &blkcg_policy_cfq
);
4443 kfree(cfqd
->root_group
);
4448 static int cfq_init_queue(struct request_queue
*q
, struct elevator_type
*e
)
4450 struct cfq_data
*cfqd
;
4451 struct blkcg_gq
*blkg __maybe_unused
;
4453 struct elevator_queue
*eq
;
4455 eq
= elevator_alloc(q
, e
);
4459 cfqd
= kzalloc_node(sizeof(*cfqd
), GFP_KERNEL
, q
->node
);
4461 kobject_put(&eq
->kobj
);
4464 eq
->elevator_data
= cfqd
;
4467 spin_lock_irq(q
->queue_lock
);
4469 spin_unlock_irq(q
->queue_lock
);
4471 /* Init root service tree */
4472 cfqd
->grp_service_tree
= CFQ_RB_ROOT
;
4474 /* Init root group and prefer root group over other groups by default */
4475 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4476 ret
= blkcg_activate_policy(q
, &blkcg_policy_cfq
);
4480 cfqd
->root_group
= blkg_to_cfqg(q
->root_blkg
);
4483 cfqd
->root_group
= kzalloc_node(sizeof(*cfqd
->root_group
),
4484 GFP_KERNEL
, cfqd
->queue
->node
);
4485 if (!cfqd
->root_group
)
4488 cfq_init_cfqg_base(cfqd
->root_group
);
4490 cfqd
->root_group
->weight
= 2 * CFQ_WEIGHT_DEFAULT
;
4491 cfqd
->root_group
->leaf_weight
= 2 * CFQ_WEIGHT_DEFAULT
;
4494 * Not strictly needed (since RB_ROOT just clears the node and we
4495 * zeroed cfqd on alloc), but better be safe in case someone decides
4496 * to add magic to the rb code
4498 for (i
= 0; i
< CFQ_PRIO_LISTS
; i
++)
4499 cfqd
->prio_trees
[i
] = RB_ROOT
;
4502 * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
4503 * Grab a permanent reference to it, so that the normal code flow
4504 * will not attempt to free it. oom_cfqq is linked to root_group
4505 * but shouldn't hold a reference as it'll never be unlinked. Lose
4506 * the reference from linking right away.
4508 cfq_init_cfqq(cfqd
, &cfqd
->oom_cfqq
, 1, 0);
4509 cfqd
->oom_cfqq
.ref
++;
4511 spin_lock_irq(q
->queue_lock
);
4512 cfq_link_cfqq_cfqg(&cfqd
->oom_cfqq
, cfqd
->root_group
);
4513 cfqg_put(cfqd
->root_group
);
4514 spin_unlock_irq(q
->queue_lock
);
4516 init_timer(&cfqd
->idle_slice_timer
);
4517 cfqd
->idle_slice_timer
.function
= cfq_idle_slice_timer
;
4518 cfqd
->idle_slice_timer
.data
= (unsigned long) cfqd
;
4520 INIT_WORK(&cfqd
->unplug_work
, cfq_kick_queue
);
4522 cfqd
->cfq_quantum
= cfq_quantum
;
4523 cfqd
->cfq_fifo_expire
[0] = cfq_fifo_expire
[0];
4524 cfqd
->cfq_fifo_expire
[1] = cfq_fifo_expire
[1];
4525 cfqd
->cfq_back_max
= cfq_back_max
;
4526 cfqd
->cfq_back_penalty
= cfq_back_penalty
;
4527 cfqd
->cfq_slice
[0] = cfq_slice_async
;
4528 cfqd
->cfq_slice
[1] = cfq_slice_sync
;
4529 cfqd
->cfq_target_latency
= cfq_target_latency
;
4530 cfqd
->cfq_slice_async_rq
= cfq_slice_async_rq
;
4531 cfqd
->cfq_slice_idle
= cfq_slice_idle
;
4532 cfqd
->cfq_group_idle
= cfq_group_idle
;
4533 cfqd
->cfq_latency
= 1;
4536 * we optimistically start assuming sync ops weren't delayed in last
4537 * second, in order to have larger depth for async operations.
4539 cfqd
->last_delayed_sync
= jiffies
- HZ
;
4544 kobject_put(&eq
->kobj
);
4548 static void cfq_registered_queue(struct request_queue
*q
)
4550 struct elevator_queue
*e
= q
->elevator
;
4551 struct cfq_data
*cfqd
= e
->elevator_data
;
4554 * Default to IOPS mode with no idling for SSDs
4556 if (blk_queue_nonrot(q
))
4557 cfqd
->cfq_slice_idle
= 0;
4561 * sysfs parts below -->
4564 cfq_var_show(unsigned int var
, char *page
)
4566 return sprintf(page
, "%u\n", var
);
4570 cfq_var_store(unsigned int *var
, const char *page
, size_t count
)
4572 char *p
= (char *) page
;
4574 *var
= simple_strtoul(p
, &p
, 10);
4578 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
4579 static ssize_t __FUNC(struct elevator_queue *e, char *page) \
4581 struct cfq_data *cfqd = e->elevator_data; \
4582 unsigned int __data = __VAR; \
4584 __data = jiffies_to_msecs(__data); \
4585 return cfq_var_show(__data, (page)); \
4587 SHOW_FUNCTION(cfq_quantum_show
, cfqd
->cfq_quantum
, 0);
4588 SHOW_FUNCTION(cfq_fifo_expire_sync_show
, cfqd
->cfq_fifo_expire
[1], 1);
4589 SHOW_FUNCTION(cfq_fifo_expire_async_show
, cfqd
->cfq_fifo_expire
[0], 1);
4590 SHOW_FUNCTION(cfq_back_seek_max_show
, cfqd
->cfq_back_max
, 0);
4591 SHOW_FUNCTION(cfq_back_seek_penalty_show
, cfqd
->cfq_back_penalty
, 0);
4592 SHOW_FUNCTION(cfq_slice_idle_show
, cfqd
->cfq_slice_idle
, 1);
4593 SHOW_FUNCTION(cfq_group_idle_show
, cfqd
->cfq_group_idle
, 1);
4594 SHOW_FUNCTION(cfq_slice_sync_show
, cfqd
->cfq_slice
[1], 1);
4595 SHOW_FUNCTION(cfq_slice_async_show
, cfqd
->cfq_slice
[0], 1);
4596 SHOW_FUNCTION(cfq_slice_async_rq_show
, cfqd
->cfq_slice_async_rq
, 0);
4597 SHOW_FUNCTION(cfq_low_latency_show
, cfqd
->cfq_latency
, 0);
4598 SHOW_FUNCTION(cfq_target_latency_show
, cfqd
->cfq_target_latency
, 1);
4599 #undef SHOW_FUNCTION
4601 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
4602 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
4604 struct cfq_data *cfqd = e->elevator_data; \
4605 unsigned int __data; \
4606 int ret = cfq_var_store(&__data, (page), count); \
4607 if (__data < (MIN)) \
4609 else if (__data > (MAX)) \
4612 *(__PTR) = msecs_to_jiffies(__data); \
4614 *(__PTR) = __data; \
4617 STORE_FUNCTION(cfq_quantum_store
, &cfqd
->cfq_quantum
, 1, UINT_MAX
, 0);
4618 STORE_FUNCTION(cfq_fifo_expire_sync_store
, &cfqd
->cfq_fifo_expire
[1], 1,
4620 STORE_FUNCTION(cfq_fifo_expire_async_store
, &cfqd
->cfq_fifo_expire
[0], 1,
4622 STORE_FUNCTION(cfq_back_seek_max_store
, &cfqd
->cfq_back_max
, 0, UINT_MAX
, 0);
4623 STORE_FUNCTION(cfq_back_seek_penalty_store
, &cfqd
->cfq_back_penalty
, 1,
4625 STORE_FUNCTION(cfq_slice_idle_store
, &cfqd
->cfq_slice_idle
, 0, UINT_MAX
, 1);
4626 STORE_FUNCTION(cfq_group_idle_store
, &cfqd
->cfq_group_idle
, 0, UINT_MAX
, 1);
4627 STORE_FUNCTION(cfq_slice_sync_store
, &cfqd
->cfq_slice
[1], 1, UINT_MAX
, 1);
4628 STORE_FUNCTION(cfq_slice_async_store
, &cfqd
->cfq_slice
[0], 1, UINT_MAX
, 1);
4629 STORE_FUNCTION(cfq_slice_async_rq_store
, &cfqd
->cfq_slice_async_rq
, 1,
4631 STORE_FUNCTION(cfq_low_latency_store
, &cfqd
->cfq_latency
, 0, 1, 0);
4632 STORE_FUNCTION(cfq_target_latency_store
, &cfqd
->cfq_target_latency
, 1, UINT_MAX
, 1);
4633 #undef STORE_FUNCTION
4635 #define CFQ_ATTR(name) \
4636 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
4638 static struct elv_fs_entry cfq_attrs
[] = {
4640 CFQ_ATTR(fifo_expire_sync
),
4641 CFQ_ATTR(fifo_expire_async
),
4642 CFQ_ATTR(back_seek_max
),
4643 CFQ_ATTR(back_seek_penalty
),
4644 CFQ_ATTR(slice_sync
),
4645 CFQ_ATTR(slice_async
),
4646 CFQ_ATTR(slice_async_rq
),
4647 CFQ_ATTR(slice_idle
),
4648 CFQ_ATTR(group_idle
),
4649 CFQ_ATTR(low_latency
),
4650 CFQ_ATTR(target_latency
),
4654 static struct elevator_type iosched_cfq
= {
4656 .elevator_merge_fn
= cfq_merge
,
4657 .elevator_merged_fn
= cfq_merged_request
,
4658 .elevator_merge_req_fn
= cfq_merged_requests
,
4659 .elevator_allow_merge_fn
= cfq_allow_merge
,
4660 .elevator_bio_merged_fn
= cfq_bio_merged
,
4661 .elevator_dispatch_fn
= cfq_dispatch_requests
,
4662 .elevator_add_req_fn
= cfq_insert_request
,
4663 .elevator_activate_req_fn
= cfq_activate_request
,
4664 .elevator_deactivate_req_fn
= cfq_deactivate_request
,
4665 .elevator_completed_req_fn
= cfq_completed_request
,
4666 .elevator_former_req_fn
= elv_rb_former_request
,
4667 .elevator_latter_req_fn
= elv_rb_latter_request
,
4668 .elevator_init_icq_fn
= cfq_init_icq
,
4669 .elevator_exit_icq_fn
= cfq_exit_icq
,
4670 .elevator_set_req_fn
= cfq_set_request
,
4671 .elevator_put_req_fn
= cfq_put_request
,
4672 .elevator_may_queue_fn
= cfq_may_queue
,
4673 .elevator_init_fn
= cfq_init_queue
,
4674 .elevator_exit_fn
= cfq_exit_queue
,
4675 .elevator_registered_fn
= cfq_registered_queue
,
4677 .icq_size
= sizeof(struct cfq_io_cq
),
4678 .icq_align
= __alignof__(struct cfq_io_cq
),
4679 .elevator_attrs
= cfq_attrs
,
4680 .elevator_name
= "cfq",
4681 .elevator_owner
= THIS_MODULE
,
4684 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4685 static struct blkcg_policy blkcg_policy_cfq
= {
4686 .pd_size
= sizeof(struct cfq_group
),
4687 .cpd_size
= sizeof(struct cfq_group_data
),
4688 .cftypes
= cfq_blkcg_files
,
4690 .cpd_init_fn
= cfq_cpd_init
,
4691 .pd_init_fn
= cfq_pd_init
,
4692 .pd_offline_fn
= cfq_pd_offline
,
4693 .pd_reset_stats_fn
= cfq_pd_reset_stats
,
4697 static int __init
cfq_init(void)
4702 * could be 0 on HZ < 1000 setups
4704 if (!cfq_slice_async
)
4705 cfq_slice_async
= 1;
4706 if (!cfq_slice_idle
)
4709 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4710 if (!cfq_group_idle
)
4713 ret
= blkcg_policy_register(&blkcg_policy_cfq
);
4721 cfq_pool
= KMEM_CACHE(cfq_queue
, 0);
4725 ret
= elv_register(&iosched_cfq
);
4732 kmem_cache_destroy(cfq_pool
);
4734 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4735 blkcg_policy_unregister(&blkcg_policy_cfq
);
4740 static void __exit
cfq_exit(void)
4742 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4743 blkcg_policy_unregister(&blkcg_policy_cfq
);
4745 elv_unregister(&iosched_cfq
);
4746 kmem_cache_destroy(cfq_pool
);
4749 module_init(cfq_init
);
4750 module_exit(cfq_exit
);
4752 MODULE_AUTHOR("Jens Axboe");
4753 MODULE_LICENSE("GPL");
4754 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");