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>
24 * Targeted preemption latency for CPU-bound tasks:
25 * (default: 20ms, units: nanoseconds)
27 * NOTE: this latency value is not the same as the concept of
28 * 'timeslice length' - timeslices in CFS are of variable length
29 * and have no persistent notion like in traditional, time-slice
30 * based scheduling concepts.
32 * (to see the precise effective timeslice length of your workload,
33 * run vmstat and monitor the context-switches (cs) field)
35 const_debug
unsigned int sysctl_sched_latency
= 20000000ULL;
38 * After fork, child runs first. (default) If set to 0 then
39 * parent will (try to) run first.
41 const_debug
unsigned int sysctl_sched_child_runs_first
= 1;
44 * Minimal preemption granularity for CPU-bound tasks:
45 * (default: 2 msec, units: nanoseconds)
47 const_debug
unsigned int sysctl_sched_nr_latency
= 20;
50 * sys_sched_yield() compat mode
52 * This option switches the agressive yield implementation of the
53 * old scheduler back on.
55 unsigned int __read_mostly sysctl_sched_compat_yield
;
58 * SCHED_BATCH wake-up granularity.
59 * (default: 10 msec, units: nanoseconds)
61 * This option delays the preemption effects of decoupled workloads
62 * and reduces their over-scheduling. Synchronous workloads will still
63 * have immediate wakeup/sleep latencies.
65 const_debug
unsigned int sysctl_sched_batch_wakeup_granularity
= 10000000UL;
68 * SCHED_OTHER wake-up granularity.
69 * (default: 10 msec, units: nanoseconds)
71 * This option delays the preemption effects of decoupled workloads
72 * and reduces their over-scheduling. Synchronous workloads will still
73 * have immediate wakeup/sleep latencies.
75 const_debug
unsigned int sysctl_sched_wakeup_granularity
= 10000000UL;
77 /**************************************************************
78 * CFS operations on generic schedulable entities:
81 #ifdef CONFIG_FAIR_GROUP_SCHED
83 /* cpu runqueue to which this cfs_rq is attached */
84 static inline struct rq
*rq_of(struct cfs_rq
*cfs_rq
)
89 /* An entity is a task if it doesn't "own" a runqueue */
90 #define entity_is_task(se) (!se->my_q)
92 #else /* CONFIG_FAIR_GROUP_SCHED */
94 static inline struct rq
*rq_of(struct cfs_rq
*cfs_rq
)
96 return container_of(cfs_rq
, struct rq
, cfs
);
99 #define entity_is_task(se) 1
101 #endif /* CONFIG_FAIR_GROUP_SCHED */
103 static inline struct task_struct
*task_of(struct sched_entity
*se
)
105 return container_of(se
, struct task_struct
, se
);
109 /**************************************************************
110 * Scheduling class tree data structure manipulation methods:
113 static inline u64
max_vruntime(u64 min_vruntime
, u64 vruntime
)
115 s64 delta
= (s64
)(vruntime
- min_vruntime
);
117 min_vruntime
= vruntime
;
122 static inline u64
min_vruntime(u64 min_vruntime
, u64 vruntime
)
124 s64 delta
= (s64
)(vruntime
- min_vruntime
);
126 min_vruntime
= vruntime
;
131 static inline s64
entity_key(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
133 return se
->vruntime
- cfs_rq
->min_vruntime
;
137 * Enqueue an entity into the rb-tree:
139 static void __enqueue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
141 struct rb_node
**link
= &cfs_rq
->tasks_timeline
.rb_node
;
142 struct rb_node
*parent
= NULL
;
143 struct sched_entity
*entry
;
144 s64 key
= entity_key(cfs_rq
, se
);
148 * Find the right place in the rbtree:
152 entry
= rb_entry(parent
, struct sched_entity
, run_node
);
154 * We dont care about collisions. Nodes with
155 * the same key stay together.
157 if (key
< entity_key(cfs_rq
, entry
)) {
158 link
= &parent
->rb_left
;
160 link
= &parent
->rb_right
;
166 * Maintain a cache of leftmost tree entries (it is frequently
170 cfs_rq
->rb_leftmost
= &se
->run_node
;
172 rb_link_node(&se
->run_node
, parent
, link
);
173 rb_insert_color(&se
->run_node
, &cfs_rq
->tasks_timeline
);
176 static void __dequeue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
178 if (cfs_rq
->rb_leftmost
== &se
->run_node
)
179 cfs_rq
->rb_leftmost
= rb_next(&se
->run_node
);
181 rb_erase(&se
->run_node
, &cfs_rq
->tasks_timeline
);
184 static inline struct rb_node
*first_fair(struct cfs_rq
*cfs_rq
)
186 return cfs_rq
->rb_leftmost
;
189 static struct sched_entity
*__pick_next_entity(struct cfs_rq
*cfs_rq
)
191 return rb_entry(first_fair(cfs_rq
), struct sched_entity
, run_node
);
194 static inline struct sched_entity
*__pick_last_entity(struct cfs_rq
*cfs_rq
)
196 struct rb_node
**link
= &cfs_rq
->tasks_timeline
.rb_node
;
197 struct sched_entity
*se
= NULL
;
198 struct rb_node
*parent
;
202 se
= rb_entry(parent
, struct sched_entity
, run_node
);
203 link
= &parent
->rb_right
;
209 /**************************************************************
210 * Scheduling class statistics methods:
215 * The idea is to set a period in which each task runs once.
217 * When there are too many tasks (sysctl_sched_nr_latency) we have to stretch
218 * this period because otherwise the slices get too small.
220 * p = (nr <= nl) ? l : l*nr/nl
222 static u64
__sched_period(unsigned long nr_running
)
224 u64 period
= sysctl_sched_latency
;
225 unsigned long nr_latency
= sysctl_sched_nr_latency
;
227 if (unlikely(nr_running
> nr_latency
)) {
228 period
*= nr_running
;
229 do_div(period
, nr_latency
);
236 * We calculate the wall-time slice from the period by taking a part
237 * proportional to the weight.
241 static u64
sched_slice(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
243 u64 slice
= __sched_period(cfs_rq
->nr_running
);
245 slice
*= se
->load
.weight
;
246 do_div(slice
, cfs_rq
->load
.weight
);
252 * We calculate the vruntime slice.
256 static u64
__sched_vslice(unsigned long rq_weight
, unsigned long nr_running
)
258 u64 vslice
= __sched_period(nr_running
);
260 do_div(vslice
, rq_weight
);
265 static u64
sched_vslice(struct cfs_rq
*cfs_rq
)
267 return __sched_vslice(cfs_rq
->load
.weight
, cfs_rq
->nr_running
);
270 static u64
sched_vslice_add(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
272 return __sched_vslice(cfs_rq
->load
.weight
+ se
->load
.weight
,
273 cfs_rq
->nr_running
+ 1);
277 * Update the current task's runtime statistics. Skip current tasks that
278 * are not in our scheduling class.
281 __update_curr(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
,
282 unsigned long delta_exec
)
284 unsigned long delta_exec_weighted
;
287 schedstat_set(curr
->exec_max
, max((u64
)delta_exec
, curr
->exec_max
));
289 curr
->sum_exec_runtime
+= delta_exec
;
290 schedstat_add(cfs_rq
, exec_clock
, delta_exec
);
291 delta_exec_weighted
= delta_exec
;
292 if (unlikely(curr
->load
.weight
!= NICE_0_LOAD
)) {
293 delta_exec_weighted
= calc_delta_fair(delta_exec_weighted
,
296 curr
->vruntime
+= delta_exec_weighted
;
299 * maintain cfs_rq->min_vruntime to be a monotonic increasing
300 * value tracking the leftmost vruntime in the tree.
302 if (first_fair(cfs_rq
)) {
303 vruntime
= min_vruntime(curr
->vruntime
,
304 __pick_next_entity(cfs_rq
)->vruntime
);
306 vruntime
= curr
->vruntime
;
308 cfs_rq
->min_vruntime
=
309 max_vruntime(cfs_rq
->min_vruntime
, vruntime
);
312 static void update_curr(struct cfs_rq
*cfs_rq
)
314 struct sched_entity
*curr
= cfs_rq
->curr
;
315 u64 now
= rq_of(cfs_rq
)->clock
;
316 unsigned long delta_exec
;
322 * Get the amount of time the current task was running
323 * since the last time we changed load (this cannot
324 * overflow on 32 bits):
326 delta_exec
= (unsigned long)(now
- curr
->exec_start
);
328 __update_curr(cfs_rq
, curr
, delta_exec
);
329 curr
->exec_start
= now
;
333 update_stats_wait_start(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
335 schedstat_set(se
->wait_start
, rq_of(cfs_rq
)->clock
);
339 * Task is being enqueued - update stats:
341 static void update_stats_enqueue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
344 * Are we enqueueing a waiting task? (for current tasks
345 * a dequeue/enqueue event is a NOP)
347 if (se
!= cfs_rq
->curr
)
348 update_stats_wait_start(cfs_rq
, se
);
352 update_stats_wait_end(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
354 schedstat_set(se
->wait_max
, max(se
->wait_max
,
355 rq_of(cfs_rq
)->clock
- se
->wait_start
));
356 schedstat_set(se
->wait_start
, 0);
360 update_stats_dequeue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
363 * Mark the end of the wait period if dequeueing a
366 if (se
!= cfs_rq
->curr
)
367 update_stats_wait_end(cfs_rq
, se
);
371 * We are picking a new current task - update its stats:
374 update_stats_curr_start(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
377 * We are starting a new run period:
379 se
->exec_start
= rq_of(cfs_rq
)->clock
;
383 * We are descheduling a task - update its stats:
386 update_stats_curr_end(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
391 /**************************************************
392 * Scheduling class queueing methods:
396 account_entity_enqueue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
398 update_load_add(&cfs_rq
->load
, se
->load
.weight
);
399 cfs_rq
->nr_running
++;
404 account_entity_dequeue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
406 update_load_sub(&cfs_rq
->load
, se
->load
.weight
);
407 cfs_rq
->nr_running
--;
411 static void enqueue_sleeper(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
413 #ifdef CONFIG_SCHEDSTATS
414 if (se
->sleep_start
) {
415 u64 delta
= rq_of(cfs_rq
)->clock
- se
->sleep_start
;
420 if (unlikely(delta
> se
->sleep_max
))
421 se
->sleep_max
= delta
;
424 se
->sum_sleep_runtime
+= delta
;
426 if (se
->block_start
) {
427 u64 delta
= rq_of(cfs_rq
)->clock
- se
->block_start
;
432 if (unlikely(delta
> se
->block_max
))
433 se
->block_max
= delta
;
436 se
->sum_sleep_runtime
+= delta
;
439 * Blocking time is in units of nanosecs, so shift by 20 to
440 * get a milliseconds-range estimation of the amount of
441 * time that the task spent sleeping:
443 if (unlikely(prof_on
== SLEEP_PROFILING
)) {
444 struct task_struct
*tsk
= task_of(se
);
446 profile_hits(SLEEP_PROFILING
, (void *)get_wchan(tsk
),
453 static void check_spread(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
455 #ifdef CONFIG_SCHED_DEBUG
456 s64 d
= se
->vruntime
- cfs_rq
->min_vruntime
;
461 if (d
> 3*sysctl_sched_latency
)
462 schedstat_inc(cfs_rq
, nr_spread_over
);
467 place_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int initial
)
471 vruntime
= cfs_rq
->min_vruntime
;
473 if (sched_feat(TREE_AVG
)) {
474 struct sched_entity
*last
= __pick_last_entity(cfs_rq
);
476 vruntime
+= last
->vruntime
;
479 } else if (sched_feat(APPROX_AVG
) && cfs_rq
->nr_running
)
480 vruntime
+= sched_vslice(cfs_rq
)/2;
482 if (initial
&& sched_feat(START_DEBIT
))
483 vruntime
+= sched_vslice_add(cfs_rq
, se
);
486 if (sched_feat(NEW_FAIR_SLEEPERS
) && entity_is_task(se
) &&
487 task_of(se
)->policy
!= SCHED_BATCH
)
488 vruntime
-= sysctl_sched_latency
;
490 vruntime
= max_t(s64
, vruntime
, se
->vruntime
);
493 se
->vruntime
= vruntime
;
498 enqueue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int wakeup
)
501 * Update run-time statistics of the 'current'.
506 place_entity(cfs_rq
, se
, 0);
507 enqueue_sleeper(cfs_rq
, se
);
510 update_stats_enqueue(cfs_rq
, se
);
511 check_spread(cfs_rq
, se
);
512 if (se
!= cfs_rq
->curr
)
513 __enqueue_entity(cfs_rq
, se
);
514 account_entity_enqueue(cfs_rq
, se
);
518 dequeue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int sleep
)
521 * Update run-time statistics of the 'current'.
525 update_stats_dequeue(cfs_rq
, se
);
527 se
->peer_preempt
= 0;
528 #ifdef CONFIG_SCHEDSTATS
529 if (entity_is_task(se
)) {
530 struct task_struct
*tsk
= task_of(se
);
532 if (tsk
->state
& TASK_INTERRUPTIBLE
)
533 se
->sleep_start
= rq_of(cfs_rq
)->clock
;
534 if (tsk
->state
& TASK_UNINTERRUPTIBLE
)
535 se
->block_start
= rq_of(cfs_rq
)->clock
;
540 if (se
!= cfs_rq
->curr
)
541 __dequeue_entity(cfs_rq
, se
);
542 account_entity_dequeue(cfs_rq
, se
);
546 * Preempt the current task with a newly woken task if needed:
549 check_preempt_tick(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
)
551 unsigned long ideal_runtime
, delta_exec
;
553 ideal_runtime
= sched_slice(cfs_rq
, curr
);
554 delta_exec
= curr
->sum_exec_runtime
- curr
->prev_sum_exec_runtime
;
555 if (delta_exec
> ideal_runtime
||
556 (sched_feat(PREEMPT_RESTRICT
) && curr
->peer_preempt
))
557 resched_task(rq_of(cfs_rq
)->curr
);
558 curr
->peer_preempt
= 0;
562 set_next_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
564 /* 'current' is not kept within the tree. */
567 * Any task has to be enqueued before it get to execute on
568 * a CPU. So account for the time it spent waiting on the
571 update_stats_wait_end(cfs_rq
, se
);
572 __dequeue_entity(cfs_rq
, se
);
575 update_stats_curr_start(cfs_rq
, se
);
577 #ifdef CONFIG_SCHEDSTATS
579 * Track our maximum slice length, if the CPU's load is at
580 * least twice that of our own weight (i.e. dont track it
581 * when there are only lesser-weight tasks around):
583 if (rq_of(cfs_rq
)->load
.weight
>= 2*se
->load
.weight
) {
584 se
->slice_max
= max(se
->slice_max
,
585 se
->sum_exec_runtime
- se
->prev_sum_exec_runtime
);
588 se
->prev_sum_exec_runtime
= se
->sum_exec_runtime
;
591 static struct sched_entity
*pick_next_entity(struct cfs_rq
*cfs_rq
)
593 struct sched_entity
*se
= NULL
;
595 if (first_fair(cfs_rq
)) {
596 se
= __pick_next_entity(cfs_rq
);
597 set_next_entity(cfs_rq
, se
);
603 static void put_prev_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*prev
)
606 * If still on the runqueue then deactivate_task()
607 * was not called and update_curr() has to be done:
612 update_stats_curr_end(cfs_rq
, prev
);
614 check_spread(cfs_rq
, prev
);
616 update_stats_wait_start(cfs_rq
, prev
);
617 /* Put 'current' back into the tree. */
618 __enqueue_entity(cfs_rq
, prev
);
623 static void entity_tick(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
)
626 * Update run-time statistics of the 'current'.
630 if (cfs_rq
->nr_running
> 1 || !sched_feat(WAKEUP_PREEMPT
))
631 check_preempt_tick(cfs_rq
, curr
);
634 /**************************************************
635 * CFS operations on tasks:
638 #ifdef CONFIG_FAIR_GROUP_SCHED
640 /* Walk up scheduling entities hierarchy */
641 #define for_each_sched_entity(se) \
642 for (; se; se = se->parent)
644 static inline struct cfs_rq
*task_cfs_rq(struct task_struct
*p
)
649 /* runqueue on which this entity is (to be) queued */
650 static inline struct cfs_rq
*cfs_rq_of(struct sched_entity
*se
)
655 /* runqueue "owned" by this group */
656 static inline struct cfs_rq
*group_cfs_rq(struct sched_entity
*grp
)
661 /* Given a group's cfs_rq on one cpu, return its corresponding cfs_rq on
662 * another cpu ('this_cpu')
664 static inline struct cfs_rq
*cpu_cfs_rq(struct cfs_rq
*cfs_rq
, int this_cpu
)
666 return cfs_rq
->tg
->cfs_rq
[this_cpu
];
669 /* Iterate thr' all leaf cfs_rq's on a runqueue */
670 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
671 list_for_each_entry(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
673 /* Do the two (enqueued) entities belong to the same group ? */
675 is_same_group(struct sched_entity
*se
, struct sched_entity
*pse
)
677 if (se
->cfs_rq
== pse
->cfs_rq
)
683 static inline struct sched_entity
*parent_entity(struct sched_entity
*se
)
688 #else /* CONFIG_FAIR_GROUP_SCHED */
690 #define for_each_sched_entity(se) \
691 for (; se; se = NULL)
693 static inline struct cfs_rq
*task_cfs_rq(struct task_struct
*p
)
695 return &task_rq(p
)->cfs
;
698 static inline struct cfs_rq
*cfs_rq_of(struct sched_entity
*se
)
700 struct task_struct
*p
= task_of(se
);
701 struct rq
*rq
= task_rq(p
);
706 /* runqueue "owned" by this group */
707 static inline struct cfs_rq
*group_cfs_rq(struct sched_entity
*grp
)
712 static inline struct cfs_rq
*cpu_cfs_rq(struct cfs_rq
*cfs_rq
, int this_cpu
)
714 return &cpu_rq(this_cpu
)->cfs
;
717 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
718 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
721 is_same_group(struct sched_entity
*se
, struct sched_entity
*pse
)
726 static inline struct sched_entity
*parent_entity(struct sched_entity
*se
)
731 #endif /* CONFIG_FAIR_GROUP_SCHED */
734 * The enqueue_task method is called before nr_running is
735 * increased. Here we update the fair scheduling stats and
736 * then put the task into the rbtree:
738 static void enqueue_task_fair(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
740 struct cfs_rq
*cfs_rq
;
741 struct sched_entity
*se
= &p
->se
;
743 for_each_sched_entity(se
) {
746 cfs_rq
= cfs_rq_of(se
);
747 enqueue_entity(cfs_rq
, se
, wakeup
);
753 * The dequeue_task method is called before nr_running is
754 * decreased. We remove the task from the rbtree and
755 * update the fair scheduling stats:
757 static void dequeue_task_fair(struct rq
*rq
, struct task_struct
*p
, int sleep
)
759 struct cfs_rq
*cfs_rq
;
760 struct sched_entity
*se
= &p
->se
;
762 for_each_sched_entity(se
) {
763 cfs_rq
= cfs_rq_of(se
);
764 dequeue_entity(cfs_rq
, se
, sleep
);
765 /* Don't dequeue parent if it has other entities besides us */
766 if (cfs_rq
->load
.weight
)
773 * sched_yield() support is very simple - we dequeue and enqueue.
775 * If compat_yield is turned on then we requeue to the end of the tree.
777 static void yield_task_fair(struct rq
*rq
)
779 struct cfs_rq
*cfs_rq
= task_cfs_rq(rq
->curr
);
780 struct sched_entity
*rightmost
, *se
= &rq
->curr
->se
;
783 * Are we the only task in the tree?
785 if (unlikely(cfs_rq
->nr_running
== 1))
788 if (likely(!sysctl_sched_compat_yield
)) {
789 __update_rq_clock(rq
);
791 * Update run-time statistics of the 'current'.
798 * Find the rightmost entry in the rbtree:
800 rightmost
= __pick_last_entity(cfs_rq
);
802 * Already in the rightmost position?
804 if (unlikely(rightmost
->vruntime
< se
->vruntime
))
808 * Minimally necessary key value to be last in the tree:
809 * Upon rescheduling, sched_class::put_prev_task() will place
810 * 'current' within the tree based on its new key value.
812 se
->vruntime
= rightmost
->vruntime
+ 1;
816 * Preempt the current task with a newly woken task if needed:
818 static void check_preempt_wakeup(struct rq
*rq
, struct task_struct
*p
)
820 struct task_struct
*curr
= rq
->curr
;
821 struct cfs_rq
*cfs_rq
= task_cfs_rq(curr
);
822 struct sched_entity
*se
= &curr
->se
, *pse
= &p
->se
;
825 if (unlikely(rt_prio(p
->prio
))) {
832 * Batch tasks do not preempt (their preemption is driven by
835 if (unlikely(p
->policy
== SCHED_BATCH
))
838 if (sched_feat(WAKEUP_PREEMPT
)) {
839 while (!is_same_group(se
, pse
)) {
840 se
= parent_entity(se
);
841 pse
= parent_entity(pse
);
844 delta
= se
->vruntime
- pse
->vruntime
;
845 gran
= sysctl_sched_wakeup_granularity
;
846 if (unlikely(se
->load
.weight
!= NICE_0_LOAD
))
847 gran
= calc_delta_fair(gran
, &se
->load
);
850 int now
= !sched_feat(PREEMPT_RESTRICT
);
852 if (now
|| p
->prio
< curr
->prio
|| !se
->peer_preempt
++)
858 static struct task_struct
*pick_next_task_fair(struct rq
*rq
)
860 struct cfs_rq
*cfs_rq
= &rq
->cfs
;
861 struct sched_entity
*se
;
863 if (unlikely(!cfs_rq
->nr_running
))
867 se
= pick_next_entity(cfs_rq
);
868 cfs_rq
= group_cfs_rq(se
);
875 * Account for a descheduled task:
877 static void put_prev_task_fair(struct rq
*rq
, struct task_struct
*prev
)
879 struct sched_entity
*se
= &prev
->se
;
880 struct cfs_rq
*cfs_rq
;
882 for_each_sched_entity(se
) {
883 cfs_rq
= cfs_rq_of(se
);
884 put_prev_entity(cfs_rq
, se
);
888 /**************************************************
889 * Fair scheduling class load-balancing methods:
893 * Load-balancing iterator. Note: while the runqueue stays locked
894 * during the whole iteration, the current task might be
895 * dequeued so the iterator has to be dequeue-safe. Here we
896 * achieve that by always pre-iterating before returning
899 static struct task_struct
*
900 __load_balance_iterator(struct cfs_rq
*cfs_rq
, struct rb_node
*curr
)
902 struct task_struct
*p
;
907 p
= rb_entry(curr
, struct task_struct
, se
.run_node
);
908 cfs_rq
->rb_load_balance_curr
= rb_next(curr
);
913 static struct task_struct
*load_balance_start_fair(void *arg
)
915 struct cfs_rq
*cfs_rq
= arg
;
917 return __load_balance_iterator(cfs_rq
, first_fair(cfs_rq
));
920 static struct task_struct
*load_balance_next_fair(void *arg
)
922 struct cfs_rq
*cfs_rq
= arg
;
924 return __load_balance_iterator(cfs_rq
, cfs_rq
->rb_load_balance_curr
);
927 #ifdef CONFIG_FAIR_GROUP_SCHED
928 static int cfs_rq_best_prio(struct cfs_rq
*cfs_rq
)
930 struct sched_entity
*curr
;
931 struct task_struct
*p
;
933 if (!cfs_rq
->nr_running
)
938 curr
= __pick_next_entity(cfs_rq
);
947 load_balance_fair(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
948 unsigned long max_nr_move
, unsigned long max_load_move
,
949 struct sched_domain
*sd
, enum cpu_idle_type idle
,
950 int *all_pinned
, int *this_best_prio
)
952 struct cfs_rq
*busy_cfs_rq
;
953 unsigned long load_moved
, total_nr_moved
= 0, nr_moved
;
954 long rem_load_move
= max_load_move
;
955 struct rq_iterator cfs_rq_iterator
;
957 cfs_rq_iterator
.start
= load_balance_start_fair
;
958 cfs_rq_iterator
.next
= load_balance_next_fair
;
960 for_each_leaf_cfs_rq(busiest
, busy_cfs_rq
) {
961 #ifdef CONFIG_FAIR_GROUP_SCHED
962 struct cfs_rq
*this_cfs_rq
;
964 unsigned long maxload
;
966 this_cfs_rq
= cpu_cfs_rq(busy_cfs_rq
, this_cpu
);
968 imbalance
= busy_cfs_rq
->load
.weight
- this_cfs_rq
->load
.weight
;
969 /* Don't pull if this_cfs_rq has more load than busy_cfs_rq */
973 /* Don't pull more than imbalance/2 */
975 maxload
= min(rem_load_move
, imbalance
);
977 *this_best_prio
= cfs_rq_best_prio(this_cfs_rq
);
979 # define maxload rem_load_move
981 /* pass busy_cfs_rq argument into
982 * load_balance_[start|next]_fair iterators
984 cfs_rq_iterator
.arg
= busy_cfs_rq
;
985 nr_moved
= balance_tasks(this_rq
, this_cpu
, busiest
,
986 max_nr_move
, maxload
, sd
, idle
, all_pinned
,
987 &load_moved
, this_best_prio
, &cfs_rq_iterator
);
989 total_nr_moved
+= nr_moved
;
990 max_nr_move
-= nr_moved
;
991 rem_load_move
-= load_moved
;
993 if (max_nr_move
<= 0 || rem_load_move
<= 0)
997 return max_load_move
- rem_load_move
;
1001 * scheduler tick hitting a task of our scheduling class:
1003 static void task_tick_fair(struct rq
*rq
, struct task_struct
*curr
)
1005 struct cfs_rq
*cfs_rq
;
1006 struct sched_entity
*se
= &curr
->se
;
1008 for_each_sched_entity(se
) {
1009 cfs_rq
= cfs_rq_of(se
);
1010 entity_tick(cfs_rq
, se
);
1014 #define swap(a,b) do { typeof(a) tmp = (a); (a) = (b); (b) = tmp; } while (0)
1017 * Share the fairness runtime between parent and child, thus the
1018 * total amount of pressure for CPU stays equal - new tasks
1019 * get a chance to run but frequent forkers are not allowed to
1020 * monopolize the CPU. Note: the parent runqueue is locked,
1021 * the child is not running yet.
1023 static void task_new_fair(struct rq
*rq
, struct task_struct
*p
)
1025 struct cfs_rq
*cfs_rq
= task_cfs_rq(p
);
1026 struct sched_entity
*se
= &p
->se
, *curr
= cfs_rq
->curr
;
1027 int this_cpu
= smp_processor_id();
1029 sched_info_queued(p
);
1031 update_curr(cfs_rq
);
1032 place_entity(cfs_rq
, se
, 1);
1034 if (sysctl_sched_child_runs_first
&& this_cpu
== task_cpu(p
) &&
1035 curr
->vruntime
< se
->vruntime
) {
1037 * Upon rescheduling, sched_class::put_prev_task() will place
1038 * 'current' within the tree based on its new key value.
1040 swap(curr
->vruntime
, se
->vruntime
);
1043 update_stats_enqueue(cfs_rq
, se
);
1044 check_spread(cfs_rq
, se
);
1045 check_spread(cfs_rq
, curr
);
1046 __enqueue_entity(cfs_rq
, se
);
1047 account_entity_enqueue(cfs_rq
, se
);
1048 se
->peer_preempt
= 0;
1049 resched_task(rq
->curr
);
1052 /* Account for a task changing its policy or group.
1054 * This routine is mostly called to set cfs_rq->curr field when a task
1055 * migrates between groups/classes.
1057 static void set_curr_task_fair(struct rq
*rq
)
1059 struct sched_entity
*se
= &rq
->curr
->se
;
1061 for_each_sched_entity(se
)
1062 set_next_entity(cfs_rq_of(se
), se
);
1066 * All the scheduling class methods:
1068 static const struct sched_class fair_sched_class
= {
1069 .next
= &idle_sched_class
,
1070 .enqueue_task
= enqueue_task_fair
,
1071 .dequeue_task
= dequeue_task_fair
,
1072 .yield_task
= yield_task_fair
,
1074 .check_preempt_curr
= check_preempt_wakeup
,
1076 .pick_next_task
= pick_next_task_fair
,
1077 .put_prev_task
= put_prev_task_fair
,
1079 .load_balance
= load_balance_fair
,
1081 .set_curr_task
= set_curr_task_fair
,
1082 .task_tick
= task_tick_fair
,
1083 .task_new
= task_new_fair
,
1086 #ifdef CONFIG_SCHED_DEBUG
1087 static void print_cfs_stats(struct seq_file
*m
, int cpu
)
1089 struct cfs_rq
*cfs_rq
;
1091 #ifdef CONFIG_FAIR_GROUP_SCHED
1092 print_cfs_rq(m
, cpu
, &cpu_rq(cpu
)->cfs
);
1094 for_each_leaf_cfs_rq(cpu_rq(cpu
), cfs_rq
)
1095 print_cfs_rq(m
, cpu
, cfs_rq
);