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 * (to see the precise effective timeslice length of your workload,
30 * run vmstat and monitor the context-switches field)
32 * On SMP systems the value of this is multiplied by the log2 of the
33 * number of CPUs. (i.e. factor 2x on 2-way systems, 3x on 4-way
34 * systems, 4x on 8-way systems, 5x on 16-way systems, etc.)
35 * Targeted preemption latency for CPU-bound tasks:
37 const_debug
unsigned int sysctl_sched_latency
= 20000000ULL;
40 * After fork, child runs first. (default) If set to 0 then
41 * parent will (try to) run first.
43 const_debug
unsigned int sysctl_sched_child_runs_first
= 1;
46 * Minimal preemption granularity for CPU-bound tasks:
47 * (default: 2 msec, units: nanoseconds)
49 const_debug
unsigned int sysctl_sched_nr_latency
= 20;
52 * sys_sched_yield() compat mode
54 * This option switches the agressive yield implementation of the
55 * old scheduler back on.
57 unsigned int __read_mostly sysctl_sched_compat_yield
;
60 * SCHED_BATCH wake-up granularity.
61 * (default: 10 msec, units: nanoseconds)
63 * This option delays the preemption effects of decoupled workloads
64 * and reduces their over-scheduling. Synchronous workloads will still
65 * have immediate wakeup/sleep latencies.
67 const_debug
unsigned int sysctl_sched_batch_wakeup_granularity
= 10000000UL;
70 * SCHED_OTHER wake-up granularity.
71 * (default: 10 msec, units: nanoseconds)
73 * This option delays the preemption effects of decoupled workloads
74 * and reduces their over-scheduling. Synchronous workloads will still
75 * have immediate wakeup/sleep latencies.
77 const_debug
unsigned int sysctl_sched_wakeup_granularity
= 10000000UL;
79 /**************************************************************
80 * CFS operations on generic schedulable entities:
83 #ifdef CONFIG_FAIR_GROUP_SCHED
85 /* cpu runqueue to which this cfs_rq is attached */
86 static inline struct rq
*rq_of(struct cfs_rq
*cfs_rq
)
91 /* An entity is a task if it doesn't "own" a runqueue */
92 #define entity_is_task(se) (!se->my_q)
94 #else /* CONFIG_FAIR_GROUP_SCHED */
96 static inline struct rq
*rq_of(struct cfs_rq
*cfs_rq
)
98 return container_of(cfs_rq
, struct rq
, cfs
);
101 #define entity_is_task(se) 1
103 #endif /* CONFIG_FAIR_GROUP_SCHED */
105 static inline struct task_struct
*task_of(struct sched_entity
*se
)
107 return container_of(se
, struct task_struct
, se
);
111 /**************************************************************
112 * Scheduling class tree data structure manipulation methods:
116 max_vruntime(u64 min_vruntime
, u64 vruntime
)
118 s64 delta
= (s64
)(vruntime
- min_vruntime
);
120 min_vruntime
= vruntime
;
126 min_vruntime(u64 min_vruntime
, u64 vruntime
)
128 s64 delta
= (s64
)(vruntime
- min_vruntime
);
130 min_vruntime
= vruntime
;
136 entity_key(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
138 return se
->vruntime
- cfs_rq
->min_vruntime
;
142 * Enqueue an entity into the rb-tree:
145 __enqueue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
147 struct rb_node
**link
= &cfs_rq
->tasks_timeline
.rb_node
;
148 struct rb_node
*parent
= NULL
;
149 struct sched_entity
*entry
;
150 s64 key
= entity_key(cfs_rq
, se
);
154 * Find the right place in the rbtree:
158 entry
= rb_entry(parent
, struct sched_entity
, run_node
);
160 * We dont care about collisions. Nodes with
161 * the same key stay together.
163 if (key
< entity_key(cfs_rq
, entry
)) {
164 link
= &parent
->rb_left
;
166 link
= &parent
->rb_right
;
172 * Maintain a cache of leftmost tree entries (it is frequently
176 cfs_rq
->rb_leftmost
= &se
->run_node
;
178 rb_link_node(&se
->run_node
, parent
, link
);
179 rb_insert_color(&se
->run_node
, &cfs_rq
->tasks_timeline
);
183 __dequeue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
185 if (cfs_rq
->rb_leftmost
== &se
->run_node
)
186 cfs_rq
->rb_leftmost
= rb_next(&se
->run_node
);
188 rb_erase(&se
->run_node
, &cfs_rq
->tasks_timeline
);
191 static inline struct rb_node
*first_fair(struct cfs_rq
*cfs_rq
)
193 return cfs_rq
->rb_leftmost
;
196 static struct sched_entity
*__pick_next_entity(struct cfs_rq
*cfs_rq
)
198 return rb_entry(first_fair(cfs_rq
), struct sched_entity
, run_node
);
201 static inline struct sched_entity
*__pick_last_entity(struct cfs_rq
*cfs_rq
)
203 struct rb_node
**link
= &cfs_rq
->tasks_timeline
.rb_node
;
204 struct sched_entity
*se
= NULL
;
205 struct rb_node
*parent
;
209 se
= rb_entry(parent
, struct sched_entity
, run_node
);
210 link
= &parent
->rb_right
;
216 /**************************************************************
217 * Scheduling class statistics methods:
222 * The idea is to set a period in which each task runs once.
224 * When there are too many tasks (sysctl_sched_nr_latency) we have to stretch
225 * this period because otherwise the slices get too small.
227 * p = (nr <= nl) ? l : l*nr/nl
229 static u64
__sched_period(unsigned long nr_running
)
231 u64 period
= sysctl_sched_latency
;
232 unsigned long nr_latency
= sysctl_sched_nr_latency
;
234 if (unlikely(nr_running
> nr_latency
)) {
235 period
*= nr_running
;
236 do_div(period
, nr_latency
);
243 * We calculate the wall-time slice from the period by taking a part
244 * proportional to the weight.
248 static u64
sched_slice(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
250 u64 slice
= __sched_period(cfs_rq
->nr_running
);
252 slice
*= se
->load
.weight
;
253 do_div(slice
, cfs_rq
->load
.weight
);
259 * We calculate the vruntime slice.
263 static u64
__sched_vslice(unsigned long rq_weight
, unsigned long nr_running
)
265 u64 vslice
= __sched_period(nr_running
);
267 do_div(vslice
, rq_weight
);
272 static u64
sched_vslice(struct cfs_rq
*cfs_rq
)
274 return __sched_vslice(cfs_rq
->load
.weight
, cfs_rq
->nr_running
);
277 static u64
sched_vslice_add(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
279 return __sched_vslice(cfs_rq
->load
.weight
+ se
->load
.weight
,
280 cfs_rq
->nr_running
+ 1);
284 * Update the current task's runtime statistics. Skip current tasks that
285 * are not in our scheduling class.
288 __update_curr(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
,
289 unsigned long delta_exec
)
291 unsigned long delta_exec_weighted
;
294 schedstat_set(curr
->exec_max
, max((u64
)delta_exec
, curr
->exec_max
));
296 curr
->sum_exec_runtime
+= delta_exec
;
297 schedstat_add(cfs_rq
, exec_clock
, delta_exec
);
298 delta_exec_weighted
= delta_exec
;
299 if (unlikely(curr
->load
.weight
!= NICE_0_LOAD
)) {
300 delta_exec_weighted
= calc_delta_fair(delta_exec_weighted
,
303 curr
->vruntime
+= delta_exec_weighted
;
306 * maintain cfs_rq->min_vruntime to be a monotonic increasing
307 * value tracking the leftmost vruntime in the tree.
309 if (first_fair(cfs_rq
)) {
310 vruntime
= min_vruntime(curr
->vruntime
,
311 __pick_next_entity(cfs_rq
)->vruntime
);
313 vruntime
= curr
->vruntime
;
315 cfs_rq
->min_vruntime
=
316 max_vruntime(cfs_rq
->min_vruntime
, vruntime
);
319 static void update_curr(struct cfs_rq
*cfs_rq
)
321 struct sched_entity
*curr
= cfs_rq
->curr
;
322 u64 now
= rq_of(cfs_rq
)->clock
;
323 unsigned long delta_exec
;
329 * Get the amount of time the current task was running
330 * since the last time we changed load (this cannot
331 * overflow on 32 bits):
333 delta_exec
= (unsigned long)(now
- curr
->exec_start
);
335 __update_curr(cfs_rq
, curr
, delta_exec
);
336 curr
->exec_start
= now
;
340 update_stats_wait_start(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
342 schedstat_set(se
->wait_start
, rq_of(cfs_rq
)->clock
);
346 * Task is being enqueued - update stats:
348 static void update_stats_enqueue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
351 * Are we enqueueing a waiting task? (for current tasks
352 * a dequeue/enqueue event is a NOP)
354 if (se
!= cfs_rq
->curr
)
355 update_stats_wait_start(cfs_rq
, se
);
359 update_stats_wait_end(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
361 schedstat_set(se
->wait_max
, max(se
->wait_max
,
362 rq_of(cfs_rq
)->clock
- se
->wait_start
));
363 schedstat_set(se
->wait_start
, 0);
367 update_stats_dequeue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
370 * Mark the end of the wait period if dequeueing a
373 if (se
!= cfs_rq
->curr
)
374 update_stats_wait_end(cfs_rq
, se
);
378 * We are picking a new current task - update its stats:
381 update_stats_curr_start(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
384 * We are starting a new run period:
386 se
->exec_start
= rq_of(cfs_rq
)->clock
;
390 * We are descheduling a task - update its stats:
393 update_stats_curr_end(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
398 /**************************************************
399 * Scheduling class queueing methods:
403 account_entity_enqueue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
405 update_load_add(&cfs_rq
->load
, se
->load
.weight
);
406 cfs_rq
->nr_running
++;
411 account_entity_dequeue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
413 update_load_sub(&cfs_rq
->load
, se
->load
.weight
);
414 cfs_rq
->nr_running
--;
418 static void enqueue_sleeper(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
420 #ifdef CONFIG_SCHEDSTATS
421 if (se
->sleep_start
) {
422 u64 delta
= rq_of(cfs_rq
)->clock
- se
->sleep_start
;
427 if (unlikely(delta
> se
->sleep_max
))
428 se
->sleep_max
= delta
;
431 se
->sum_sleep_runtime
+= delta
;
433 if (se
->block_start
) {
434 u64 delta
= rq_of(cfs_rq
)->clock
- se
->block_start
;
439 if (unlikely(delta
> se
->block_max
))
440 se
->block_max
= delta
;
443 se
->sum_sleep_runtime
+= delta
;
446 * Blocking time is in units of nanosecs, so shift by 20 to
447 * get a milliseconds-range estimation of the amount of
448 * time that the task spent sleeping:
450 if (unlikely(prof_on
== SLEEP_PROFILING
)) {
451 struct task_struct
*tsk
= task_of(se
);
453 profile_hits(SLEEP_PROFILING
, (void *)get_wchan(tsk
),
460 static void check_spread(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
462 #ifdef CONFIG_SCHED_DEBUG
463 s64 d
= se
->vruntime
- cfs_rq
->min_vruntime
;
468 if (d
> 3*sysctl_sched_latency
)
469 schedstat_inc(cfs_rq
, nr_spread_over
);
474 place_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int initial
)
478 vruntime
= cfs_rq
->min_vruntime
;
480 if (sched_feat(USE_TREE_AVG
)) {
481 struct sched_entity
*last
= __pick_last_entity(cfs_rq
);
483 vruntime
+= last
->vruntime
;
486 } else if (sched_feat(APPROX_AVG
) && cfs_rq
->nr_running
)
487 vruntime
+= sched_vslice(cfs_rq
)/2;
489 if (initial
&& sched_feat(START_DEBIT
))
490 vruntime
+= sched_vslice_add(cfs_rq
, se
);
493 if (sched_feat(NEW_FAIR_SLEEPERS
))
494 vruntime
-= sysctl_sched_latency
;
496 vruntime
= max_t(s64
, vruntime
, se
->vruntime
);
499 se
->vruntime
= vruntime
;
504 enqueue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int wakeup
)
507 * Update run-time statistics of the 'current'.
512 place_entity(cfs_rq
, se
, 0);
513 enqueue_sleeper(cfs_rq
, se
);
516 update_stats_enqueue(cfs_rq
, se
);
517 check_spread(cfs_rq
, se
);
518 if (se
!= cfs_rq
->curr
)
519 __enqueue_entity(cfs_rq
, se
);
520 account_entity_enqueue(cfs_rq
, se
);
524 dequeue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int sleep
)
527 * Update run-time statistics of the 'current'.
531 update_stats_dequeue(cfs_rq
, se
);
533 #ifdef CONFIG_SCHEDSTATS
534 if (entity_is_task(se
)) {
535 struct task_struct
*tsk
= task_of(se
);
537 if (tsk
->state
& TASK_INTERRUPTIBLE
)
538 se
->sleep_start
= rq_of(cfs_rq
)->clock
;
539 if (tsk
->state
& TASK_UNINTERRUPTIBLE
)
540 se
->block_start
= rq_of(cfs_rq
)->clock
;
545 if (se
!= cfs_rq
->curr
)
546 __dequeue_entity(cfs_rq
, se
);
547 account_entity_dequeue(cfs_rq
, se
);
551 * Preempt the current task with a newly woken task if needed:
554 check_preempt_tick(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
)
556 unsigned long ideal_runtime
, delta_exec
;
558 ideal_runtime
= sched_slice(cfs_rq
, curr
);
559 delta_exec
= curr
->sum_exec_runtime
- curr
->prev_sum_exec_runtime
;
560 if (delta_exec
> ideal_runtime
)
561 resched_task(rq_of(cfs_rq
)->curr
);
565 set_next_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
567 /* 'current' is not kept within the tree. */
570 * Any task has to be enqueued before it get to execute on
571 * a CPU. So account for the time it spent waiting on the
574 update_stats_wait_end(cfs_rq
, se
);
575 __dequeue_entity(cfs_rq
, se
);
578 update_stats_curr_start(cfs_rq
, se
);
580 #ifdef CONFIG_SCHEDSTATS
582 * Track our maximum slice length, if the CPU's load is at
583 * least twice that of our own weight (i.e. dont track it
584 * when there are only lesser-weight tasks around):
586 if (rq_of(cfs_rq
)->load
.weight
>= 2*se
->load
.weight
) {
587 se
->slice_max
= max(se
->slice_max
,
588 se
->sum_exec_runtime
- se
->prev_sum_exec_runtime
);
591 se
->prev_sum_exec_runtime
= se
->sum_exec_runtime
;
594 static struct sched_entity
*pick_next_entity(struct cfs_rq
*cfs_rq
)
596 struct sched_entity
*se
= NULL
;
598 if (first_fair(cfs_rq
)) {
599 se
= __pick_next_entity(cfs_rq
);
600 set_next_entity(cfs_rq
, se
);
606 static void put_prev_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*prev
)
609 * If still on the runqueue then deactivate_task()
610 * was not called and update_curr() has to be done:
615 update_stats_curr_end(cfs_rq
, prev
);
617 check_spread(cfs_rq
, prev
);
619 update_stats_wait_start(cfs_rq
, prev
);
620 /* Put 'current' back into the tree. */
621 __enqueue_entity(cfs_rq
, prev
);
626 static void entity_tick(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
)
629 * Update run-time statistics of the 'current'.
633 if (cfs_rq
->nr_running
> 1)
634 check_preempt_tick(cfs_rq
, curr
);
637 /**************************************************
638 * CFS operations on tasks:
641 #ifdef CONFIG_FAIR_GROUP_SCHED
643 /* Walk up scheduling entities hierarchy */
644 #define for_each_sched_entity(se) \
645 for (; se; se = se->parent)
647 static inline struct cfs_rq
*task_cfs_rq(struct task_struct
*p
)
652 /* runqueue on which this entity is (to be) queued */
653 static inline struct cfs_rq
*cfs_rq_of(struct sched_entity
*se
)
658 /* runqueue "owned" by this group */
659 static inline struct cfs_rq
*group_cfs_rq(struct sched_entity
*grp
)
664 /* Given a group's cfs_rq on one cpu, return its corresponding cfs_rq on
665 * another cpu ('this_cpu')
667 static inline struct cfs_rq
*cpu_cfs_rq(struct cfs_rq
*cfs_rq
, int this_cpu
)
669 return cfs_rq
->tg
->cfs_rq
[this_cpu
];
672 /* Iterate thr' all leaf cfs_rq's on a runqueue */
673 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
674 list_for_each_entry(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
676 /* Do the two (enqueued) entities belong to the same group ? */
678 is_same_group(struct sched_entity
*se
, struct sched_entity
*pse
)
680 if (se
->cfs_rq
== pse
->cfs_rq
)
686 static inline struct sched_entity
*parent_entity(struct sched_entity
*se
)
691 #else /* CONFIG_FAIR_GROUP_SCHED */
693 #define for_each_sched_entity(se) \
694 for (; se; se = NULL)
696 static inline struct cfs_rq
*task_cfs_rq(struct task_struct
*p
)
698 return &task_rq(p
)->cfs
;
701 static inline struct cfs_rq
*cfs_rq_of(struct sched_entity
*se
)
703 struct task_struct
*p
= task_of(se
);
704 struct rq
*rq
= task_rq(p
);
709 /* runqueue "owned" by this group */
710 static inline struct cfs_rq
*group_cfs_rq(struct sched_entity
*grp
)
715 static inline struct cfs_rq
*cpu_cfs_rq(struct cfs_rq
*cfs_rq
, int this_cpu
)
717 return &cpu_rq(this_cpu
)->cfs
;
720 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
721 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
724 is_same_group(struct sched_entity
*se
, struct sched_entity
*pse
)
729 static inline struct sched_entity
*parent_entity(struct sched_entity
*se
)
734 #endif /* CONFIG_FAIR_GROUP_SCHED */
737 * The enqueue_task method is called before nr_running is
738 * increased. Here we update the fair scheduling stats and
739 * then put the task into the rbtree:
741 static void enqueue_task_fair(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
743 struct cfs_rq
*cfs_rq
;
744 struct sched_entity
*se
= &p
->se
;
746 for_each_sched_entity(se
) {
749 cfs_rq
= cfs_rq_of(se
);
750 enqueue_entity(cfs_rq
, se
, wakeup
);
756 * The dequeue_task method is called before nr_running is
757 * decreased. We remove the task from the rbtree and
758 * update the fair scheduling stats:
760 static void dequeue_task_fair(struct rq
*rq
, struct task_struct
*p
, int sleep
)
762 struct cfs_rq
*cfs_rq
;
763 struct sched_entity
*se
= &p
->se
;
765 for_each_sched_entity(se
) {
766 cfs_rq
= cfs_rq_of(se
);
767 dequeue_entity(cfs_rq
, se
, sleep
);
768 /* Don't dequeue parent if it has other entities besides us */
769 if (cfs_rq
->load
.weight
)
776 * sched_yield() support is very simple - we dequeue and enqueue.
778 * If compat_yield is turned on then we requeue to the end of the tree.
780 static void yield_task_fair(struct rq
*rq
)
782 struct cfs_rq
*cfs_rq
= task_cfs_rq(rq
->curr
);
783 struct sched_entity
*rightmost
, *se
= &rq
->curr
->se
;
786 * Are we the only task in the tree?
788 if (unlikely(cfs_rq
->nr_running
== 1))
791 if (likely(!sysctl_sched_compat_yield
)) {
792 __update_rq_clock(rq
);
794 * Update run-time statistics of the 'current'.
801 * Find the rightmost entry in the rbtree:
803 rightmost
= __pick_last_entity(cfs_rq
);
805 * Already in the rightmost position?
807 if (unlikely(rightmost
->vruntime
< se
->vruntime
))
811 * Minimally necessary key value to be last in the tree:
812 * Upon rescheduling, sched_class::put_prev_task() will place
813 * 'current' within the tree based on its new key value.
815 se
->vruntime
= rightmost
->vruntime
+ 1;
819 * Preempt the current task with a newly woken task if needed:
821 static void check_preempt_wakeup(struct rq
*rq
, struct task_struct
*p
)
823 struct task_struct
*curr
= rq
->curr
;
824 struct cfs_rq
*cfs_rq
= task_cfs_rq(curr
);
825 struct sched_entity
*se
= &curr
->se
, *pse
= &p
->se
;
828 if (unlikely(rt_prio(p
->prio
))) {
835 while (!is_same_group(se
, pse
)) {
836 se
= parent_entity(se
);
837 pse
= parent_entity(pse
);
840 delta
= se
->vruntime
- pse
->vruntime
;
842 if (delta
> (s64
)sysctl_sched_wakeup_granularity
)
846 static struct task_struct
*pick_next_task_fair(struct rq
*rq
)
848 struct cfs_rq
*cfs_rq
= &rq
->cfs
;
849 struct sched_entity
*se
;
851 if (unlikely(!cfs_rq
->nr_running
))
855 se
= pick_next_entity(cfs_rq
);
856 cfs_rq
= group_cfs_rq(se
);
863 * Account for a descheduled task:
865 static void put_prev_task_fair(struct rq
*rq
, struct task_struct
*prev
)
867 struct sched_entity
*se
= &prev
->se
;
868 struct cfs_rq
*cfs_rq
;
870 for_each_sched_entity(se
) {
871 cfs_rq
= cfs_rq_of(se
);
872 put_prev_entity(cfs_rq
, se
);
876 /**************************************************
877 * Fair scheduling class load-balancing methods:
881 * Load-balancing iterator. Note: while the runqueue stays locked
882 * during the whole iteration, the current task might be
883 * dequeued so the iterator has to be dequeue-safe. Here we
884 * achieve that by always pre-iterating before returning
887 static struct task_struct
*
888 __load_balance_iterator(struct cfs_rq
*cfs_rq
, struct rb_node
*curr
)
890 struct task_struct
*p
;
895 p
= rb_entry(curr
, struct task_struct
, se
.run_node
);
896 cfs_rq
->rb_load_balance_curr
= rb_next(curr
);
901 static struct task_struct
*load_balance_start_fair(void *arg
)
903 struct cfs_rq
*cfs_rq
= arg
;
905 return __load_balance_iterator(cfs_rq
, first_fair(cfs_rq
));
908 static struct task_struct
*load_balance_next_fair(void *arg
)
910 struct cfs_rq
*cfs_rq
= arg
;
912 return __load_balance_iterator(cfs_rq
, cfs_rq
->rb_load_balance_curr
);
915 #ifdef CONFIG_FAIR_GROUP_SCHED
916 static int cfs_rq_best_prio(struct cfs_rq
*cfs_rq
)
918 struct sched_entity
*curr
;
919 struct task_struct
*p
;
921 if (!cfs_rq
->nr_running
)
926 curr
= __pick_next_entity(cfs_rq
);
935 load_balance_fair(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
936 unsigned long max_nr_move
, unsigned long max_load_move
,
937 struct sched_domain
*sd
, enum cpu_idle_type idle
,
938 int *all_pinned
, int *this_best_prio
)
940 struct cfs_rq
*busy_cfs_rq
;
941 unsigned long load_moved
, total_nr_moved
= 0, nr_moved
;
942 long rem_load_move
= max_load_move
;
943 struct rq_iterator cfs_rq_iterator
;
945 cfs_rq_iterator
.start
= load_balance_start_fair
;
946 cfs_rq_iterator
.next
= load_balance_next_fair
;
948 for_each_leaf_cfs_rq(busiest
, busy_cfs_rq
) {
949 #ifdef CONFIG_FAIR_GROUP_SCHED
950 struct cfs_rq
*this_cfs_rq
;
952 unsigned long maxload
;
954 this_cfs_rq
= cpu_cfs_rq(busy_cfs_rq
, this_cpu
);
956 imbalance
= busy_cfs_rq
->load
.weight
- this_cfs_rq
->load
.weight
;
957 /* Don't pull if this_cfs_rq has more load than busy_cfs_rq */
961 /* Don't pull more than imbalance/2 */
963 maxload
= min(rem_load_move
, imbalance
);
965 *this_best_prio
= cfs_rq_best_prio(this_cfs_rq
);
967 # define maxload rem_load_move
969 /* pass busy_cfs_rq argument into
970 * load_balance_[start|next]_fair iterators
972 cfs_rq_iterator
.arg
= busy_cfs_rq
;
973 nr_moved
= balance_tasks(this_rq
, this_cpu
, busiest
,
974 max_nr_move
, maxload
, sd
, idle
, all_pinned
,
975 &load_moved
, this_best_prio
, &cfs_rq_iterator
);
977 total_nr_moved
+= nr_moved
;
978 max_nr_move
-= nr_moved
;
979 rem_load_move
-= load_moved
;
981 if (max_nr_move
<= 0 || rem_load_move
<= 0)
985 return max_load_move
- rem_load_move
;
989 * scheduler tick hitting a task of our scheduling class:
991 static void task_tick_fair(struct rq
*rq
, struct task_struct
*curr
)
993 struct cfs_rq
*cfs_rq
;
994 struct sched_entity
*se
= &curr
->se
;
996 for_each_sched_entity(se
) {
997 cfs_rq
= cfs_rq_of(se
);
998 entity_tick(cfs_rq
, se
);
1002 #define swap(a,b) do { typeof(a) tmp = (a); (a) = (b); (b) = tmp; } while (0)
1005 * Share the fairness runtime between parent and child, thus the
1006 * total amount of pressure for CPU stays equal - new tasks
1007 * get a chance to run but frequent forkers are not allowed to
1008 * monopolize the CPU. Note: the parent runqueue is locked,
1009 * the child is not running yet.
1011 static void task_new_fair(struct rq
*rq
, struct task_struct
*p
)
1013 struct cfs_rq
*cfs_rq
= task_cfs_rq(p
);
1014 struct sched_entity
*se
= &p
->se
, *curr
= cfs_rq
->curr
;
1016 sched_info_queued(p
);
1018 update_curr(cfs_rq
);
1019 place_entity(cfs_rq
, se
, 1);
1021 if (sysctl_sched_child_runs_first
&&
1022 curr
->vruntime
< se
->vruntime
) {
1024 * Upon rescheduling, sched_class::put_prev_task() will place
1025 * 'current' within the tree based on its new key value.
1027 swap(curr
->vruntime
, se
->vruntime
);
1030 update_stats_enqueue(cfs_rq
, se
);
1031 check_spread(cfs_rq
, se
);
1032 check_spread(cfs_rq
, curr
);
1033 __enqueue_entity(cfs_rq
, se
);
1034 account_entity_enqueue(cfs_rq
, se
);
1035 resched_task(rq
->curr
);
1038 /* Account for a task changing its policy or group.
1040 * This routine is mostly called to set cfs_rq->curr field when a task
1041 * migrates between groups/classes.
1043 static void set_curr_task_fair(struct rq
*rq
)
1045 struct sched_entity
*se
= &rq
->curr
->se
;
1047 for_each_sched_entity(se
)
1048 set_next_entity(cfs_rq_of(se
), se
);
1052 * All the scheduling class methods:
1054 static const struct sched_class fair_sched_class
= {
1055 .next
= &idle_sched_class
,
1056 .enqueue_task
= enqueue_task_fair
,
1057 .dequeue_task
= dequeue_task_fair
,
1058 .yield_task
= yield_task_fair
,
1060 .check_preempt_curr
= check_preempt_wakeup
,
1062 .pick_next_task
= pick_next_task_fair
,
1063 .put_prev_task
= put_prev_task_fair
,
1065 .load_balance
= load_balance_fair
,
1067 .set_curr_task
= set_curr_task_fair
,
1068 .task_tick
= task_tick_fair
,
1069 .task_new
= task_new_fair
,
1072 #ifdef CONFIG_SCHED_DEBUG
1073 static void print_cfs_stats(struct seq_file
*m
, int cpu
)
1075 struct cfs_rq
*cfs_rq
;
1077 #ifdef CONFIG_FAIR_GROUP_SCHED
1078 print_cfs_rq(m
, cpu
, &cpu_rq(cpu
)->cfs
);
1080 for_each_leaf_cfs_rq(cpu_rq(cpu
), cfs_rq
)
1081 print_cfs_rq(m
, cpu
, cfs_rq
);