2 * Completely Fair Scheduling (CFS) Class (SCHED_NORMAL/SCHED_BATCH)
4 * Copyright (C) 2007 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
6 * Interactivity improvements by Mike Galbraith
7 * (C) 2007 Mike Galbraith <efault@gmx.de>
9 * Various enhancements by Dmitry Adamushko.
10 * (C) 2007 Dmitry Adamushko <dmitry.adamushko@gmail.com>
12 * Group scheduling enhancements by Srivatsa Vaddagiri
13 * Copyright IBM Corporation, 2007
14 * Author: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
16 * Scaled math optimizations by Thomas Gleixner
17 * Copyright (C) 2007, Thomas Gleixner <tglx@linutronix.de>
19 * Adaptive scheduling granularity, math enhancements by Peter Zijlstra
20 * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
23 #include <linux/latencytop.h>
26 * Targeted preemption latency for CPU-bound tasks:
27 * (default: 20ms * (1 + ilog(ncpus)), units: nanoseconds)
29 * NOTE: this latency value is not the same as the concept of
30 * 'timeslice length' - timeslices in CFS are of variable length
31 * and have no persistent notion like in traditional, time-slice
32 * based scheduling concepts.
34 * (to see the precise effective timeslice length of your workload,
35 * run vmstat and monitor the context-switches (cs) field)
37 unsigned int sysctl_sched_latency
= 20000000ULL;
40 * Minimal preemption granularity for CPU-bound tasks:
41 * (default: 4 msec * (1 + ilog(ncpus)), units: nanoseconds)
43 unsigned int sysctl_sched_min_granularity
= 4000000ULL;
46 * is kept at sysctl_sched_latency / sysctl_sched_min_granularity
48 static unsigned int sched_nr_latency
= 5;
51 * After fork, child runs first. (default) If set to 0 then
52 * parent will (try to) run first.
54 const_debug
unsigned int sysctl_sched_child_runs_first
= 1;
57 * sys_sched_yield() compat mode
59 * This option switches the agressive yield implementation of the
60 * old scheduler back on.
62 unsigned int __read_mostly sysctl_sched_compat_yield
;
65 * SCHED_OTHER wake-up granularity.
66 * (default: 5 msec * (1 + ilog(ncpus)), units: nanoseconds)
68 * This option delays the preemption effects of decoupled workloads
69 * and reduces their over-scheduling. Synchronous workloads will still
70 * have immediate wakeup/sleep latencies.
72 unsigned int sysctl_sched_wakeup_granularity
= 5000000UL;
74 const_debug
unsigned int sysctl_sched_migration_cost
= 500000UL;
76 /**************************************************************
77 * CFS operations on generic schedulable entities:
80 static inline struct task_struct
*task_of(struct sched_entity
*se
)
82 return container_of(se
, struct task_struct
, se
);
85 #ifdef CONFIG_FAIR_GROUP_SCHED
87 /* cpu runqueue to which this cfs_rq is attached */
88 static inline struct rq
*rq_of(struct cfs_rq
*cfs_rq
)
93 /* An entity is a task if it doesn't "own" a runqueue */
94 #define entity_is_task(se) (!se->my_q)
96 /* Walk up scheduling entities hierarchy */
97 #define for_each_sched_entity(se) \
98 for (; se; se = se->parent)
100 static inline struct cfs_rq
*task_cfs_rq(struct task_struct
*p
)
105 /* runqueue on which this entity is (to be) queued */
106 static inline struct cfs_rq
*cfs_rq_of(struct sched_entity
*se
)
111 /* runqueue "owned" by this group */
112 static inline struct cfs_rq
*group_cfs_rq(struct sched_entity
*grp
)
117 /* Given a group's cfs_rq on one cpu, return its corresponding cfs_rq on
118 * another cpu ('this_cpu')
120 static inline struct cfs_rq
*cpu_cfs_rq(struct cfs_rq
*cfs_rq
, int this_cpu
)
122 return cfs_rq
->tg
->cfs_rq
[this_cpu
];
125 /* Iterate thr' all leaf cfs_rq's on a runqueue */
126 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
127 list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
129 /* Do the two (enqueued) entities belong to the same group ? */
131 is_same_group(struct sched_entity
*se
, struct sched_entity
*pse
)
133 if (se
->cfs_rq
== pse
->cfs_rq
)
139 static inline struct sched_entity
*parent_entity(struct sched_entity
*se
)
144 #else /* CONFIG_FAIR_GROUP_SCHED */
146 static inline struct rq
*rq_of(struct cfs_rq
*cfs_rq
)
148 return container_of(cfs_rq
, struct rq
, cfs
);
151 #define entity_is_task(se) 1
153 #define for_each_sched_entity(se) \
154 for (; se; se = NULL)
156 static inline struct cfs_rq
*task_cfs_rq(struct task_struct
*p
)
158 return &task_rq(p
)->cfs
;
161 static inline struct cfs_rq
*cfs_rq_of(struct sched_entity
*se
)
163 struct task_struct
*p
= task_of(se
);
164 struct rq
*rq
= task_rq(p
);
169 /* runqueue "owned" by this group */
170 static inline struct cfs_rq
*group_cfs_rq(struct sched_entity
*grp
)
175 static inline struct cfs_rq
*cpu_cfs_rq(struct cfs_rq
*cfs_rq
, int this_cpu
)
177 return &cpu_rq(this_cpu
)->cfs
;
180 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
181 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
184 is_same_group(struct sched_entity
*se
, struct sched_entity
*pse
)
189 static inline struct sched_entity
*parent_entity(struct sched_entity
*se
)
194 #endif /* CONFIG_FAIR_GROUP_SCHED */
197 /**************************************************************
198 * Scheduling class tree data structure manipulation methods:
201 static inline u64
max_vruntime(u64 min_vruntime
, u64 vruntime
)
203 s64 delta
= (s64
)(vruntime
- min_vruntime
);
205 min_vruntime
= vruntime
;
210 static inline u64
min_vruntime(u64 min_vruntime
, u64 vruntime
)
212 s64 delta
= (s64
)(vruntime
- min_vruntime
);
214 min_vruntime
= vruntime
;
219 static inline s64
entity_key(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
221 return se
->vruntime
- cfs_rq
->min_vruntime
;
225 * Enqueue an entity into the rb-tree:
227 static void __enqueue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
229 struct rb_node
**link
= &cfs_rq
->tasks_timeline
.rb_node
;
230 struct rb_node
*parent
= NULL
;
231 struct sched_entity
*entry
;
232 s64 key
= entity_key(cfs_rq
, se
);
236 * Find the right place in the rbtree:
240 entry
= rb_entry(parent
, struct sched_entity
, run_node
);
242 * We dont care about collisions. Nodes with
243 * the same key stay together.
245 if (key
< entity_key(cfs_rq
, entry
)) {
246 link
= &parent
->rb_left
;
248 link
= &parent
->rb_right
;
254 * Maintain a cache of leftmost tree entries (it is frequently
258 cfs_rq
->rb_leftmost
= &se
->run_node
;
260 * maintain cfs_rq->min_vruntime to be a monotonic increasing
261 * value tracking the leftmost vruntime in the tree.
263 cfs_rq
->min_vruntime
=
264 max_vruntime(cfs_rq
->min_vruntime
, se
->vruntime
);
267 rb_link_node(&se
->run_node
, parent
, link
);
268 rb_insert_color(&se
->run_node
, &cfs_rq
->tasks_timeline
);
271 static void __dequeue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
273 if (cfs_rq
->rb_leftmost
== &se
->run_node
) {
274 struct rb_node
*next_node
;
275 struct sched_entity
*next
;
277 next_node
= rb_next(&se
->run_node
);
278 cfs_rq
->rb_leftmost
= next_node
;
281 next
= rb_entry(next_node
,
282 struct sched_entity
, run_node
);
283 cfs_rq
->min_vruntime
=
284 max_vruntime(cfs_rq
->min_vruntime
,
289 if (cfs_rq
->next
== se
)
292 rb_erase(&se
->run_node
, &cfs_rq
->tasks_timeline
);
295 static inline struct rb_node
*first_fair(struct cfs_rq
*cfs_rq
)
297 return cfs_rq
->rb_leftmost
;
300 static struct sched_entity
*__pick_next_entity(struct cfs_rq
*cfs_rq
)
302 return rb_entry(first_fair(cfs_rq
), struct sched_entity
, run_node
);
305 static inline struct sched_entity
*__pick_last_entity(struct cfs_rq
*cfs_rq
)
307 struct rb_node
*last
= rb_last(&cfs_rq
->tasks_timeline
);
312 return rb_entry(last
, struct sched_entity
, run_node
);
315 /**************************************************************
316 * Scheduling class statistics methods:
319 #ifdef CONFIG_SCHED_DEBUG
320 int sched_nr_latency_handler(struct ctl_table
*table
, int write
,
321 struct file
*filp
, void __user
*buffer
, size_t *lenp
,
324 int ret
= proc_dointvec_minmax(table
, write
, filp
, buffer
, lenp
, ppos
);
329 sched_nr_latency
= DIV_ROUND_UP(sysctl_sched_latency
,
330 sysctl_sched_min_granularity
);
339 static inline unsigned long
340 calc_delta_weight(unsigned long delta
, struct sched_entity
*se
)
342 for_each_sched_entity(se
) {
343 delta
= calc_delta_mine(delta
,
344 se
->load
.weight
, &cfs_rq_of(se
)->load
);
353 static inline unsigned long
354 calc_delta_fair(unsigned long delta
, struct sched_entity
*se
)
356 for_each_sched_entity(se
) {
357 delta
= calc_delta_mine(delta
,
358 cfs_rq_of(se
)->load
.weight
, &se
->load
);
365 * The idea is to set a period in which each task runs once.
367 * When there are too many tasks (sysctl_sched_nr_latency) we have to stretch
368 * this period because otherwise the slices get too small.
370 * p = (nr <= nl) ? l : l*nr/nl
372 static u64
__sched_period(unsigned long nr_running
)
374 u64 period
= sysctl_sched_latency
;
375 unsigned long nr_latency
= sched_nr_latency
;
377 if (unlikely(nr_running
> nr_latency
)) {
378 period
= sysctl_sched_min_granularity
;
379 period
*= nr_running
;
386 * We calculate the wall-time slice from the period by taking a part
387 * proportional to the weight.
391 static u64
sched_slice(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
393 return calc_delta_weight(__sched_period(cfs_rq
->nr_running
), se
);
397 * We calculate the vruntime slice of a to be inserted task
401 static u64
sched_vslice_add(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
403 unsigned long nr_running
= cfs_rq
->nr_running
;
408 return __sched_period(nr_running
);
412 * Update the current task's runtime statistics. Skip current tasks that
413 * are not in our scheduling class.
416 __update_curr(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
,
417 unsigned long delta_exec
)
419 unsigned long delta_exec_weighted
;
421 schedstat_set(curr
->exec_max
, max((u64
)delta_exec
, curr
->exec_max
));
423 curr
->sum_exec_runtime
+= delta_exec
;
424 schedstat_add(cfs_rq
, exec_clock
, delta_exec
);
425 delta_exec_weighted
= calc_delta_fair(delta_exec
, curr
);
426 curr
->vruntime
+= delta_exec_weighted
;
429 static void update_curr(struct cfs_rq
*cfs_rq
)
431 struct sched_entity
*curr
= cfs_rq
->curr
;
432 u64 now
= rq_of(cfs_rq
)->clock
;
433 unsigned long delta_exec
;
439 * Get the amount of time the current task was running
440 * since the last time we changed load (this cannot
441 * overflow on 32 bits):
443 delta_exec
= (unsigned long)(now
- curr
->exec_start
);
445 __update_curr(cfs_rq
, curr
, delta_exec
);
446 curr
->exec_start
= now
;
448 if (entity_is_task(curr
)) {
449 struct task_struct
*curtask
= task_of(curr
);
451 cpuacct_charge(curtask
, delta_exec
);
456 update_stats_wait_start(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
458 schedstat_set(se
->wait_start
, rq_of(cfs_rq
)->clock
);
462 * Task is being enqueued - update stats:
464 static void update_stats_enqueue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
467 * Are we enqueueing a waiting task? (for current tasks
468 * a dequeue/enqueue event is a NOP)
470 if (se
!= cfs_rq
->curr
)
471 update_stats_wait_start(cfs_rq
, se
);
475 update_stats_wait_end(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
477 schedstat_set(se
->wait_max
, max(se
->wait_max
,
478 rq_of(cfs_rq
)->clock
- se
->wait_start
));
479 schedstat_set(se
->wait_count
, se
->wait_count
+ 1);
480 schedstat_set(se
->wait_sum
, se
->wait_sum
+
481 rq_of(cfs_rq
)->clock
- se
->wait_start
);
482 schedstat_set(se
->wait_start
, 0);
486 update_stats_dequeue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
489 * Mark the end of the wait period if dequeueing a
492 if (se
!= cfs_rq
->curr
)
493 update_stats_wait_end(cfs_rq
, se
);
497 * We are picking a new current task - update its stats:
500 update_stats_curr_start(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
503 * We are starting a new run period:
505 se
->exec_start
= rq_of(cfs_rq
)->clock
;
508 /**************************************************
509 * Scheduling class queueing methods:
512 #if defined CONFIG_SMP && defined CONFIG_FAIR_GROUP_SCHED
514 add_cfs_task_weight(struct cfs_rq
*cfs_rq
, unsigned long weight
)
516 cfs_rq
->task_weight
+= weight
;
520 add_cfs_task_weight(struct cfs_rq
*cfs_rq
, unsigned long weight
)
526 account_entity_enqueue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
528 update_load_add(&cfs_rq
->load
, se
->load
.weight
);
529 if (!parent_entity(se
))
530 inc_cpu_load(rq_of(cfs_rq
), se
->load
.weight
);
531 if (entity_is_task(se
))
532 add_cfs_task_weight(cfs_rq
, se
->load
.weight
);
533 cfs_rq
->nr_running
++;
535 list_add(&se
->group_node
, &cfs_rq
->tasks
);
539 account_entity_dequeue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
541 update_load_sub(&cfs_rq
->load
, se
->load
.weight
);
542 if (!parent_entity(se
))
543 dec_cpu_load(rq_of(cfs_rq
), se
->load
.weight
);
544 if (entity_is_task(se
))
545 add_cfs_task_weight(cfs_rq
, -se
->load
.weight
);
546 cfs_rq
->nr_running
--;
548 list_del_init(&se
->group_node
);
551 static void enqueue_sleeper(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
553 #ifdef CONFIG_SCHEDSTATS
554 if (se
->sleep_start
) {
555 u64 delta
= rq_of(cfs_rq
)->clock
- se
->sleep_start
;
556 struct task_struct
*tsk
= task_of(se
);
561 if (unlikely(delta
> se
->sleep_max
))
562 se
->sleep_max
= delta
;
565 se
->sum_sleep_runtime
+= delta
;
567 account_scheduler_latency(tsk
, delta
>> 10, 1);
569 if (se
->block_start
) {
570 u64 delta
= rq_of(cfs_rq
)->clock
- se
->block_start
;
571 struct task_struct
*tsk
= task_of(se
);
576 if (unlikely(delta
> se
->block_max
))
577 se
->block_max
= delta
;
580 se
->sum_sleep_runtime
+= delta
;
583 * Blocking time is in units of nanosecs, so shift by 20 to
584 * get a milliseconds-range estimation of the amount of
585 * time that the task spent sleeping:
587 if (unlikely(prof_on
== SLEEP_PROFILING
)) {
589 profile_hits(SLEEP_PROFILING
, (void *)get_wchan(tsk
),
592 account_scheduler_latency(tsk
, delta
>> 10, 0);
597 static void check_spread(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
599 #ifdef CONFIG_SCHED_DEBUG
600 s64 d
= se
->vruntime
- cfs_rq
->min_vruntime
;
605 if (d
> 3*sysctl_sched_latency
)
606 schedstat_inc(cfs_rq
, nr_spread_over
);
611 place_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int initial
)
615 if (first_fair(cfs_rq
)) {
616 vruntime
= min_vruntime(cfs_rq
->min_vruntime
,
617 __pick_next_entity(cfs_rq
)->vruntime
);
619 vruntime
= cfs_rq
->min_vruntime
;
622 * The 'current' period is already promised to the current tasks,
623 * however the extra weight of the new task will slow them down a
624 * little, place the new task so that it fits in the slot that
625 * stays open at the end.
627 if (initial
&& sched_feat(START_DEBIT
))
628 vruntime
+= sched_vslice_add(cfs_rq
, se
);
631 /* sleeps upto a single latency don't count. */
632 if (sched_feat(NEW_FAIR_SLEEPERS
)) {
633 unsigned long thresh
= sysctl_sched_latency
;
636 * convert the sleeper threshold into virtual time
638 if (sched_feat(NORMALIZED_SLEEPER
))
639 thresh
= calc_delta_fair(thresh
, se
);
644 /* ensure we never gain time by being placed backwards. */
645 vruntime
= max_vruntime(se
->vruntime
, vruntime
);
648 se
->vruntime
= vruntime
;
652 enqueue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int wakeup
)
655 * Update run-time statistics of the 'current'.
658 account_entity_enqueue(cfs_rq
, se
);
661 place_entity(cfs_rq
, se
, 0);
662 enqueue_sleeper(cfs_rq
, se
);
665 update_stats_enqueue(cfs_rq
, se
);
666 check_spread(cfs_rq
, se
);
667 if (se
!= cfs_rq
->curr
)
668 __enqueue_entity(cfs_rq
, se
);
672 dequeue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int sleep
)
675 * Update run-time statistics of the 'current'.
679 update_stats_dequeue(cfs_rq
, se
);
681 #ifdef CONFIG_SCHEDSTATS
682 if (entity_is_task(se
)) {
683 struct task_struct
*tsk
= task_of(se
);
685 if (tsk
->state
& TASK_INTERRUPTIBLE
)
686 se
->sleep_start
= rq_of(cfs_rq
)->clock
;
687 if (tsk
->state
& TASK_UNINTERRUPTIBLE
)
688 se
->block_start
= rq_of(cfs_rq
)->clock
;
693 if (se
!= cfs_rq
->curr
)
694 __dequeue_entity(cfs_rq
, se
);
695 account_entity_dequeue(cfs_rq
, se
);
699 * Preempt the current task with a newly woken task if needed:
702 check_preempt_tick(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
)
704 unsigned long ideal_runtime
, delta_exec
;
706 ideal_runtime
= sched_slice(cfs_rq
, curr
);
707 delta_exec
= curr
->sum_exec_runtime
- curr
->prev_sum_exec_runtime
;
708 if (delta_exec
> ideal_runtime
)
709 resched_task(rq_of(cfs_rq
)->curr
);
713 set_next_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
715 /* 'current' is not kept within the tree. */
718 * Any task has to be enqueued before it get to execute on
719 * a CPU. So account for the time it spent waiting on the
722 update_stats_wait_end(cfs_rq
, se
);
723 __dequeue_entity(cfs_rq
, se
);
726 update_stats_curr_start(cfs_rq
, se
);
728 #ifdef CONFIG_SCHEDSTATS
730 * Track our maximum slice length, if the CPU's load is at
731 * least twice that of our own weight (i.e. dont track it
732 * when there are only lesser-weight tasks around):
734 if (rq_of(cfs_rq
)->load
.weight
>= 2*se
->load
.weight
) {
735 se
->slice_max
= max(se
->slice_max
,
736 se
->sum_exec_runtime
- se
->prev_sum_exec_runtime
);
739 se
->prev_sum_exec_runtime
= se
->sum_exec_runtime
;
742 static struct sched_entity
*
743 pick_next(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
745 struct rq
*rq
= rq_of(cfs_rq
);
746 u64 pair_slice
= rq
->clock
- cfs_rq
->pair_start
;
748 if (!cfs_rq
->next
|| pair_slice
> sched_slice(cfs_rq
, cfs_rq
->next
)) {
749 cfs_rq
->pair_start
= rq
->clock
;
756 static struct sched_entity
*pick_next_entity(struct cfs_rq
*cfs_rq
)
758 struct sched_entity
*se
= NULL
;
760 if (first_fair(cfs_rq
)) {
761 se
= __pick_next_entity(cfs_rq
);
762 se
= pick_next(cfs_rq
, se
);
763 set_next_entity(cfs_rq
, se
);
769 static void put_prev_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*prev
)
772 * If still on the runqueue then deactivate_task()
773 * was not called and update_curr() has to be done:
778 check_spread(cfs_rq
, prev
);
780 update_stats_wait_start(cfs_rq
, prev
);
781 /* Put 'current' back into the tree. */
782 __enqueue_entity(cfs_rq
, prev
);
788 entity_tick(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
, int queued
)
791 * Update run-time statistics of the 'current'.
795 #ifdef CONFIG_SCHED_HRTICK
797 * queued ticks are scheduled to match the slice, so don't bother
798 * validating it and just reschedule.
801 resched_task(rq_of(cfs_rq
)->curr
);
805 * don't let the period tick interfere with the hrtick preemption
807 if (!sched_feat(DOUBLE_TICK
) &&
808 hrtimer_active(&rq_of(cfs_rq
)->hrtick_timer
))
812 if (cfs_rq
->nr_running
> 1 || !sched_feat(WAKEUP_PREEMPT
))
813 check_preempt_tick(cfs_rq
, curr
);
816 /**************************************************
817 * CFS operations on tasks:
820 #ifdef CONFIG_SCHED_HRTICK
821 static void hrtick_start_fair(struct rq
*rq
, struct task_struct
*p
)
823 struct sched_entity
*se
= &p
->se
;
824 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
826 WARN_ON(task_rq(p
) != rq
);
828 if (hrtick_enabled(rq
) && cfs_rq
->nr_running
> 1) {
829 u64 slice
= sched_slice(cfs_rq
, se
);
830 u64 ran
= se
->sum_exec_runtime
- se
->prev_sum_exec_runtime
;
831 s64 delta
= slice
- ran
;
840 * Don't schedule slices shorter than 10000ns, that just
841 * doesn't make sense. Rely on vruntime for fairness.
844 delta
= max_t(s64
, 10000LL, delta
);
846 hrtick_start(rq
, delta
);
849 #else /* !CONFIG_SCHED_HRTICK */
851 hrtick_start_fair(struct rq
*rq
, struct task_struct
*p
)
857 * The enqueue_task method is called before nr_running is
858 * increased. Here we update the fair scheduling stats and
859 * then put the task into the rbtree:
861 static void enqueue_task_fair(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
863 struct cfs_rq
*cfs_rq
;
864 struct sched_entity
*se
= &p
->se
;
866 for_each_sched_entity(se
) {
869 cfs_rq
= cfs_rq_of(se
);
870 enqueue_entity(cfs_rq
, se
, wakeup
);
874 hrtick_start_fair(rq
, rq
->curr
);
878 * The dequeue_task method is called before nr_running is
879 * decreased. We remove the task from the rbtree and
880 * update the fair scheduling stats:
882 static void dequeue_task_fair(struct rq
*rq
, struct task_struct
*p
, int sleep
)
884 struct cfs_rq
*cfs_rq
;
885 struct sched_entity
*se
= &p
->se
;
887 for_each_sched_entity(se
) {
888 cfs_rq
= cfs_rq_of(se
);
889 dequeue_entity(cfs_rq
, se
, sleep
);
890 /* Don't dequeue parent if it has other entities besides us */
891 if (cfs_rq
->load
.weight
)
896 hrtick_start_fair(rq
, rq
->curr
);
900 * sched_yield() support is very simple - we dequeue and enqueue.
902 * If compat_yield is turned on then we requeue to the end of the tree.
904 static void yield_task_fair(struct rq
*rq
)
906 struct task_struct
*curr
= rq
->curr
;
907 struct cfs_rq
*cfs_rq
= task_cfs_rq(curr
);
908 struct sched_entity
*rightmost
, *se
= &curr
->se
;
911 * Are we the only task in the tree?
913 if (unlikely(cfs_rq
->nr_running
== 1))
916 if (likely(!sysctl_sched_compat_yield
) && curr
->policy
!= SCHED_BATCH
) {
919 * Update run-time statistics of the 'current'.
926 * Find the rightmost entry in the rbtree:
928 rightmost
= __pick_last_entity(cfs_rq
);
930 * Already in the rightmost position?
932 if (unlikely(!rightmost
|| rightmost
->vruntime
< se
->vruntime
))
936 * Minimally necessary key value to be last in the tree:
937 * Upon rescheduling, sched_class::put_prev_task() will place
938 * 'current' within the tree based on its new key value.
940 se
->vruntime
= rightmost
->vruntime
+ 1;
944 * wake_idle() will wake a task on an idle cpu if task->cpu is
945 * not idle and an idle cpu is available. The span of cpus to
946 * search starts with cpus closest then further out as needed,
947 * so we always favor a closer, idle cpu.
948 * Domains may include CPUs that are not usable for migration,
949 * hence we need to mask them out (cpu_active_map)
951 * Returns the CPU we should wake onto.
953 #if defined(ARCH_HAS_SCHED_WAKE_IDLE)
954 static int wake_idle(int cpu
, struct task_struct
*p
)
957 struct sched_domain
*sd
;
961 * If it is idle, then it is the best cpu to run this task.
963 * This cpu is also the best, if it has more than one task already.
964 * Siblings must be also busy(in most cases) as they didn't already
965 * pickup the extra load from this cpu and hence we need not check
966 * sibling runqueue info. This will avoid the checks and cache miss
967 * penalities associated with that.
969 if (idle_cpu(cpu
) || cpu_rq(cpu
)->cfs
.nr_running
> 1)
972 for_each_domain(cpu
, sd
) {
973 if ((sd
->flags
& SD_WAKE_IDLE
)
974 || ((sd
->flags
& SD_WAKE_IDLE_FAR
)
975 && !task_hot(p
, task_rq(p
)->clock
, sd
))) {
976 cpus_and(tmp
, sd
->span
, p
->cpus_allowed
);
977 cpus_and(tmp
, tmp
, cpu_active_map
);
978 for_each_cpu_mask_nr(i
, tmp
) {
980 if (i
!= task_cpu(p
)) {
993 #else /* !ARCH_HAS_SCHED_WAKE_IDLE*/
994 static inline int wake_idle(int cpu
, struct task_struct
*p
)
1002 static const struct sched_class fair_sched_class
;
1004 #ifdef CONFIG_FAIR_GROUP_SCHED
1006 * effective_load() calculates the load change as seen from the root_task_group
1008 * Adding load to a group doesn't make a group heavier, but can cause movement
1009 * of group shares between cpus. Assuming the shares were perfectly aligned one
1010 * can calculate the shift in shares.
1012 * The problem is that perfectly aligning the shares is rather expensive, hence
1013 * we try to avoid doing that too often - see update_shares(), which ratelimits
1016 * We compensate this by not only taking the current delta into account, but
1017 * also considering the delta between when the shares were last adjusted and
1020 * We still saw a performance dip, some tracing learned us that between
1021 * cgroup:/ and cgroup:/foo balancing the number of affine wakeups increased
1022 * significantly. Therefore try to bias the error in direction of failing
1023 * the affine wakeup.
1026 static long effective_load(struct task_group
*tg
, int cpu
,
1029 struct sched_entity
*se
= tg
->se
[cpu
];
1036 * By not taking the decrease of shares on the other cpu into
1037 * account our error leans towards reducing the affine wakeups.
1039 if (!wl
&& sched_feat(ASYM_EFF_LOAD
))
1043 * Instead of using this increment, also add the difference
1044 * between when the shares were last updated and now.
1046 more_w
= se
->my_q
->load
.weight
- se
->my_q
->rq_weight
;
1050 for_each_sched_entity(se
) {
1051 #define D(n) (likely(n) ? (n) : 1)
1053 long S
, rw
, s
, a
, b
;
1055 S
= se
->my_q
->tg
->shares
;
1056 s
= se
->my_q
->shares
;
1057 rw
= se
->my_q
->rq_weight
;
1064 * Assume the group is already running and will
1065 * thus already be accounted for in the weight.
1067 * That is, moving shares between CPUs, does not
1068 * alter the group weight.
1079 static inline unsigned long effective_load(struct task_group
*tg
, int cpu
,
1080 unsigned long wl
, unsigned long wg
)
1088 wake_affine(struct rq
*rq
, struct sched_domain
*this_sd
, struct rq
*this_rq
,
1089 struct task_struct
*p
, int prev_cpu
, int this_cpu
, int sync
,
1090 int idx
, unsigned long load
, unsigned long this_load
,
1091 unsigned int imbalance
)
1093 struct task_struct
*curr
= this_rq
->curr
;
1094 struct task_group
*tg
;
1095 unsigned long tl
= this_load
;
1096 unsigned long tl_per_task
;
1097 unsigned long weight
;
1100 if (!(this_sd
->flags
& SD_WAKE_AFFINE
) || !sched_feat(AFFINE_WAKEUPS
))
1104 * If sync wakeup then subtract the (maximum possible)
1105 * effect of the currently running task from the load
1106 * of the current CPU:
1109 tg
= task_group(current
);
1110 weight
= current
->se
.load
.weight
;
1112 tl
+= effective_load(tg
, this_cpu
, -weight
, -weight
);
1113 load
+= effective_load(tg
, prev_cpu
, 0, -weight
);
1117 weight
= p
->se
.load
.weight
;
1119 balanced
= 100*(tl
+ effective_load(tg
, this_cpu
, weight
, weight
)) <=
1120 imbalance
*(load
+ effective_load(tg
, prev_cpu
, 0, weight
));
1123 * If the currently running task will sleep within
1124 * a reasonable amount of time then attract this newly
1127 if (sync
&& balanced
) {
1128 if (curr
->se
.avg_overlap
< sysctl_sched_migration_cost
&&
1129 p
->se
.avg_overlap
< sysctl_sched_migration_cost
)
1133 schedstat_inc(p
, se
.nr_wakeups_affine_attempts
);
1134 tl_per_task
= cpu_avg_load_per_task(this_cpu
);
1136 if ((tl
<= load
&& tl
+ target_load(prev_cpu
, idx
) <= tl_per_task
) ||
1139 * This domain has SD_WAKE_AFFINE and
1140 * p is cache cold in this domain, and
1141 * there is no bad imbalance.
1143 schedstat_inc(this_sd
, ttwu_move_affine
);
1144 schedstat_inc(p
, se
.nr_wakeups_affine
);
1151 static int select_task_rq_fair(struct task_struct
*p
, int sync
)
1153 struct sched_domain
*sd
, *this_sd
= NULL
;
1154 int prev_cpu
, this_cpu
, new_cpu
;
1155 unsigned long load
, this_load
;
1156 struct rq
*rq
, *this_rq
;
1157 unsigned int imbalance
;
1160 prev_cpu
= task_cpu(p
);
1162 this_cpu
= smp_processor_id();
1163 this_rq
= cpu_rq(this_cpu
);
1167 * 'this_sd' is the first domain that both
1168 * this_cpu and prev_cpu are present in:
1170 for_each_domain(this_cpu
, sd
) {
1171 if (cpu_isset(prev_cpu
, sd
->span
)) {
1177 if (unlikely(!cpu_isset(this_cpu
, p
->cpus_allowed
)))
1181 * Check for affine wakeup and passive balancing possibilities.
1186 idx
= this_sd
->wake_idx
;
1188 imbalance
= 100 + (this_sd
->imbalance_pct
- 100) / 2;
1190 load
= source_load(prev_cpu
, idx
);
1191 this_load
= target_load(this_cpu
, idx
);
1193 if (wake_affine(rq
, this_sd
, this_rq
, p
, prev_cpu
, this_cpu
, sync
, idx
,
1194 load
, this_load
, imbalance
))
1197 if (prev_cpu
== this_cpu
)
1201 * Start passive balancing when half the imbalance_pct
1204 if (this_sd
->flags
& SD_WAKE_BALANCE
) {
1205 if (imbalance
*this_load
<= 100*load
) {
1206 schedstat_inc(this_sd
, ttwu_move_balance
);
1207 schedstat_inc(p
, se
.nr_wakeups_passive
);
1213 return wake_idle(new_cpu
, p
);
1215 #endif /* CONFIG_SMP */
1217 static unsigned long wakeup_gran(struct sched_entity
*se
)
1219 unsigned long gran
= sysctl_sched_wakeup_granularity
;
1222 * More easily preempt - nice tasks, while not making it harder for
1225 if (sched_feat(ASYM_GRAN
))
1226 gran
= calc_delta_mine(gran
, NICE_0_LOAD
, &se
->load
);
1232 * Preempt the current task with a newly woken task if needed:
1234 static void check_preempt_wakeup(struct rq
*rq
, struct task_struct
*p
, int sync
)
1236 struct task_struct
*curr
= rq
->curr
;
1237 struct cfs_rq
*cfs_rq
= task_cfs_rq(curr
);
1238 struct sched_entity
*se
= &curr
->se
, *pse
= &p
->se
;
1241 if (unlikely(rt_prio(p
->prio
))) {
1242 update_rq_clock(rq
);
1243 update_curr(cfs_rq
);
1248 if (unlikely(se
== pse
))
1252 * We can come here with TIF_NEED_RESCHED already set from new task
1255 if (test_tsk_need_resched(curr
))
1258 cfs_rq_of(pse
)->next
= pse
;
1261 * Batch tasks do not preempt (their preemption is driven by
1264 if (unlikely(p
->policy
== SCHED_BATCH
))
1267 if (!sched_feat(WAKEUP_PREEMPT
))
1270 if (sched_feat(WAKEUP_OVERLAP
) && sync
&&
1271 se
->avg_overlap
< sysctl_sched_migration_cost
&&
1272 pse
->avg_overlap
< sysctl_sched_migration_cost
) {
1277 delta_exec
= se
->sum_exec_runtime
- se
->prev_sum_exec_runtime
;
1278 if (delta_exec
> wakeup_gran(pse
))
1282 static struct task_struct
*pick_next_task_fair(struct rq
*rq
)
1284 struct task_struct
*p
;
1285 struct cfs_rq
*cfs_rq
= &rq
->cfs
;
1286 struct sched_entity
*se
;
1288 if (unlikely(!cfs_rq
->nr_running
))
1292 se
= pick_next_entity(cfs_rq
);
1293 cfs_rq
= group_cfs_rq(se
);
1297 hrtick_start_fair(rq
, p
);
1303 * Account for a descheduled task:
1305 static void put_prev_task_fair(struct rq
*rq
, struct task_struct
*prev
)
1307 struct sched_entity
*se
= &prev
->se
;
1308 struct cfs_rq
*cfs_rq
;
1310 for_each_sched_entity(se
) {
1311 cfs_rq
= cfs_rq_of(se
);
1312 put_prev_entity(cfs_rq
, se
);
1317 /**************************************************
1318 * Fair scheduling class load-balancing methods:
1322 * Load-balancing iterator. Note: while the runqueue stays locked
1323 * during the whole iteration, the current task might be
1324 * dequeued so the iterator has to be dequeue-safe. Here we
1325 * achieve that by always pre-iterating before returning
1328 static struct task_struct
*
1329 __load_balance_iterator(struct cfs_rq
*cfs_rq
, struct list_head
*next
)
1331 struct task_struct
*p
= NULL
;
1332 struct sched_entity
*se
;
1334 if (next
== &cfs_rq
->tasks
)
1337 /* Skip over entities that are not tasks */
1339 se
= list_entry(next
, struct sched_entity
, group_node
);
1341 } while (next
!= &cfs_rq
->tasks
&& !entity_is_task(se
));
1343 if (next
== &cfs_rq
->tasks
&& !entity_is_task(se
))
1346 cfs_rq
->balance_iterator
= next
;
1348 if (entity_is_task(se
))
1354 static struct task_struct
*load_balance_start_fair(void *arg
)
1356 struct cfs_rq
*cfs_rq
= arg
;
1358 return __load_balance_iterator(cfs_rq
, cfs_rq
->tasks
.next
);
1361 static struct task_struct
*load_balance_next_fair(void *arg
)
1363 struct cfs_rq
*cfs_rq
= arg
;
1365 return __load_balance_iterator(cfs_rq
, cfs_rq
->balance_iterator
);
1368 static unsigned long
1369 __load_balance_fair(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1370 unsigned long max_load_move
, struct sched_domain
*sd
,
1371 enum cpu_idle_type idle
, int *all_pinned
, int *this_best_prio
,
1372 struct cfs_rq
*cfs_rq
)
1374 struct rq_iterator cfs_rq_iterator
;
1376 cfs_rq_iterator
.start
= load_balance_start_fair
;
1377 cfs_rq_iterator
.next
= load_balance_next_fair
;
1378 cfs_rq_iterator
.arg
= cfs_rq
;
1380 return balance_tasks(this_rq
, this_cpu
, busiest
,
1381 max_load_move
, sd
, idle
, all_pinned
,
1382 this_best_prio
, &cfs_rq_iterator
);
1385 #ifdef CONFIG_FAIR_GROUP_SCHED
1386 static unsigned long
1387 load_balance_fair(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1388 unsigned long max_load_move
,
1389 struct sched_domain
*sd
, enum cpu_idle_type idle
,
1390 int *all_pinned
, int *this_best_prio
)
1392 long rem_load_move
= max_load_move
;
1393 int busiest_cpu
= cpu_of(busiest
);
1394 struct task_group
*tg
;
1397 update_h_load(busiest_cpu
);
1399 list_for_each_entry_rcu(tg
, &task_groups
, list
) {
1400 struct cfs_rq
*busiest_cfs_rq
= tg
->cfs_rq
[busiest_cpu
];
1401 unsigned long busiest_h_load
= busiest_cfs_rq
->h_load
;
1402 unsigned long busiest_weight
= busiest_cfs_rq
->load
.weight
;
1403 u64 rem_load
, moved_load
;
1408 if (!busiest_cfs_rq
->task_weight
)
1411 rem_load
= (u64
)rem_load_move
* busiest_weight
;
1412 rem_load
= div_u64(rem_load
, busiest_h_load
+ 1);
1414 moved_load
= __load_balance_fair(this_rq
, this_cpu
, busiest
,
1415 rem_load
, sd
, idle
, all_pinned
, this_best_prio
,
1416 tg
->cfs_rq
[busiest_cpu
]);
1421 moved_load
*= busiest_h_load
;
1422 moved_load
= div_u64(moved_load
, busiest_weight
+ 1);
1424 rem_load_move
-= moved_load
;
1425 if (rem_load_move
< 0)
1430 return max_load_move
- rem_load_move
;
1433 static unsigned long
1434 load_balance_fair(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1435 unsigned long max_load_move
,
1436 struct sched_domain
*sd
, enum cpu_idle_type idle
,
1437 int *all_pinned
, int *this_best_prio
)
1439 return __load_balance_fair(this_rq
, this_cpu
, busiest
,
1440 max_load_move
, sd
, idle
, all_pinned
,
1441 this_best_prio
, &busiest
->cfs
);
1446 move_one_task_fair(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1447 struct sched_domain
*sd
, enum cpu_idle_type idle
)
1449 struct cfs_rq
*busy_cfs_rq
;
1450 struct rq_iterator cfs_rq_iterator
;
1452 cfs_rq_iterator
.start
= load_balance_start_fair
;
1453 cfs_rq_iterator
.next
= load_balance_next_fair
;
1455 for_each_leaf_cfs_rq(busiest
, busy_cfs_rq
) {
1457 * pass busy_cfs_rq argument into
1458 * load_balance_[start|next]_fair iterators
1460 cfs_rq_iterator
.arg
= busy_cfs_rq
;
1461 if (iter_move_one_task(this_rq
, this_cpu
, busiest
, sd
, idle
,
1468 #endif /* CONFIG_SMP */
1471 * scheduler tick hitting a task of our scheduling class:
1473 static void task_tick_fair(struct rq
*rq
, struct task_struct
*curr
, int queued
)
1475 struct cfs_rq
*cfs_rq
;
1476 struct sched_entity
*se
= &curr
->se
;
1478 for_each_sched_entity(se
) {
1479 cfs_rq
= cfs_rq_of(se
);
1480 entity_tick(cfs_rq
, se
, queued
);
1484 #define swap(a, b) do { typeof(a) tmp = (a); (a) = (b); (b) = tmp; } while (0)
1487 * Share the fairness runtime between parent and child, thus the
1488 * total amount of pressure for CPU stays equal - new tasks
1489 * get a chance to run but frequent forkers are not allowed to
1490 * monopolize the CPU. Note: the parent runqueue is locked,
1491 * the child is not running yet.
1493 static void task_new_fair(struct rq
*rq
, struct task_struct
*p
)
1495 struct cfs_rq
*cfs_rq
= task_cfs_rq(p
);
1496 struct sched_entity
*se
= &p
->se
, *curr
= cfs_rq
->curr
;
1497 int this_cpu
= smp_processor_id();
1499 sched_info_queued(p
);
1501 update_curr(cfs_rq
);
1502 place_entity(cfs_rq
, se
, 1);
1504 /* 'curr' will be NULL if the child belongs to a different group */
1505 if (sysctl_sched_child_runs_first
&& this_cpu
== task_cpu(p
) &&
1506 curr
&& curr
->vruntime
< se
->vruntime
) {
1508 * Upon rescheduling, sched_class::put_prev_task() will place
1509 * 'current' within the tree based on its new key value.
1511 swap(curr
->vruntime
, se
->vruntime
);
1512 resched_task(rq
->curr
);
1515 enqueue_task_fair(rq
, p
, 0);
1519 * Priority of the task has changed. Check to see if we preempt
1522 static void prio_changed_fair(struct rq
*rq
, struct task_struct
*p
,
1523 int oldprio
, int running
)
1526 * Reschedule if we are currently running on this runqueue and
1527 * our priority decreased, or if we are not currently running on
1528 * this runqueue and our priority is higher than the current's
1531 if (p
->prio
> oldprio
)
1532 resched_task(rq
->curr
);
1534 check_preempt_curr(rq
, p
, 0);
1538 * We switched to the sched_fair class.
1540 static void switched_to_fair(struct rq
*rq
, struct task_struct
*p
,
1544 * We were most likely switched from sched_rt, so
1545 * kick off the schedule if running, otherwise just see
1546 * if we can still preempt the current task.
1549 resched_task(rq
->curr
);
1551 check_preempt_curr(rq
, p
, 0);
1554 /* Account for a task changing its policy or group.
1556 * This routine is mostly called to set cfs_rq->curr field when a task
1557 * migrates between groups/classes.
1559 static void set_curr_task_fair(struct rq
*rq
)
1561 struct sched_entity
*se
= &rq
->curr
->se
;
1563 for_each_sched_entity(se
)
1564 set_next_entity(cfs_rq_of(se
), se
);
1567 #ifdef CONFIG_FAIR_GROUP_SCHED
1568 static void moved_group_fair(struct task_struct
*p
)
1570 struct cfs_rq
*cfs_rq
= task_cfs_rq(p
);
1572 update_curr(cfs_rq
);
1573 place_entity(cfs_rq
, &p
->se
, 1);
1578 * All the scheduling class methods:
1580 static const struct sched_class fair_sched_class
= {
1581 .next
= &idle_sched_class
,
1582 .enqueue_task
= enqueue_task_fair
,
1583 .dequeue_task
= dequeue_task_fair
,
1584 .yield_task
= yield_task_fair
,
1586 .select_task_rq
= select_task_rq_fair
,
1587 #endif /* CONFIG_SMP */
1589 .check_preempt_curr
= check_preempt_wakeup
,
1591 .pick_next_task
= pick_next_task_fair
,
1592 .put_prev_task
= put_prev_task_fair
,
1595 .load_balance
= load_balance_fair
,
1596 .move_one_task
= move_one_task_fair
,
1599 .set_curr_task
= set_curr_task_fair
,
1600 .task_tick
= task_tick_fair
,
1601 .task_new
= task_new_fair
,
1603 .prio_changed
= prio_changed_fair
,
1604 .switched_to
= switched_to_fair
,
1606 #ifdef CONFIG_FAIR_GROUP_SCHED
1607 .moved_group
= moved_group_fair
,
1611 #ifdef CONFIG_SCHED_DEBUG
1612 static void print_cfs_stats(struct seq_file
*m
, int cpu
)
1614 struct cfs_rq
*cfs_rq
;
1617 for_each_leaf_cfs_rq(cpu_rq(cpu
), cfs_rq
)
1618 print_cfs_rq(m
, cpu
, cfs_rq
);