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: 25 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
= 25000000UL;
70 * SCHED_OTHER wake-up granularity.
71 * (default: 1 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
= 2000000UL;
79 extern struct sched_class fair_sched_class
;
81 /**************************************************************
82 * CFS operations on generic schedulable entities:
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 #else /* CONFIG_FAIR_GROUP_SCHED */
98 static inline struct rq
*rq_of(struct cfs_rq
*cfs_rq
)
100 return container_of(cfs_rq
, struct rq
, cfs
);
103 #define entity_is_task(se) 1
105 #endif /* CONFIG_FAIR_GROUP_SCHED */
107 static inline struct task_struct
*task_of(struct sched_entity
*se
)
109 return container_of(se
, struct task_struct
, se
);
113 /**************************************************************
114 * Scheduling class tree data structure manipulation methods:
118 max_vruntime(u64 min_vruntime
, u64 vruntime
)
120 s64 delta
= (s64
)(vruntime
- min_vruntime
);
122 min_vruntime
= vruntime
;
128 min_vruntime(u64 min_vruntime
, u64 vruntime
)
130 s64 delta
= (s64
)(vruntime
- min_vruntime
);
132 min_vruntime
= vruntime
;
138 entity_key(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
140 return se
->vruntime
- cfs_rq
->min_vruntime
;
144 * Enqueue an entity into the rb-tree:
147 __enqueue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
149 struct rb_node
**link
= &cfs_rq
->tasks_timeline
.rb_node
;
150 struct rb_node
*parent
= NULL
;
151 struct sched_entity
*entry
;
152 s64 key
= entity_key(cfs_rq
, se
);
156 * Find the right place in the rbtree:
160 entry
= rb_entry(parent
, struct sched_entity
, run_node
);
162 * We dont care about collisions. Nodes with
163 * the same key stay together.
165 if (key
< entity_key(cfs_rq
, entry
)) {
166 link
= &parent
->rb_left
;
168 link
= &parent
->rb_right
;
174 * Maintain a cache of leftmost tree entries (it is frequently
178 cfs_rq
->rb_leftmost
= &se
->run_node
;
180 rb_link_node(&se
->run_node
, parent
, link
);
181 rb_insert_color(&se
->run_node
, &cfs_rq
->tasks_timeline
);
185 __dequeue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
187 if (cfs_rq
->rb_leftmost
== &se
->run_node
)
188 cfs_rq
->rb_leftmost
= rb_next(&se
->run_node
);
190 rb_erase(&se
->run_node
, &cfs_rq
->tasks_timeline
);
193 static inline struct rb_node
*first_fair(struct cfs_rq
*cfs_rq
)
195 return cfs_rq
->rb_leftmost
;
198 static struct sched_entity
*__pick_next_entity(struct cfs_rq
*cfs_rq
)
200 return rb_entry(first_fair(cfs_rq
), struct sched_entity
, run_node
);
203 static inline struct sched_entity
*__pick_last_entity(struct cfs_rq
*cfs_rq
)
205 struct rb_node
**link
= &cfs_rq
->tasks_timeline
.rb_node
;
206 struct sched_entity
*se
= NULL
;
207 struct rb_node
*parent
;
211 se
= rb_entry(parent
, struct sched_entity
, run_node
);
212 link
= &parent
->rb_right
;
218 /**************************************************************
219 * Scheduling class statistics methods:
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
);
235 static u64
sched_slice(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
237 u64 period
= __sched_period(cfs_rq
->nr_running
);
239 period
*= se
->load
.weight
;
240 do_div(period
, cfs_rq
->load
.weight
);
245 static u64
__sched_vslice(unsigned long nr_running
)
247 unsigned long period
= sysctl_sched_latency
;
248 unsigned long nr_latency
= sysctl_sched_nr_latency
;
250 if (unlikely(nr_running
> nr_latency
))
251 nr_running
= nr_latency
;
253 period
/= nr_running
;
259 * Update the current task's runtime statistics. Skip current tasks that
260 * are not in our scheduling class.
263 __update_curr(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
,
264 unsigned long delta_exec
)
266 unsigned long delta_exec_weighted
;
269 schedstat_set(curr
->exec_max
, max((u64
)delta_exec
, curr
->exec_max
));
271 curr
->sum_exec_runtime
+= delta_exec
;
272 schedstat_add(cfs_rq
, exec_clock
, delta_exec
);
273 delta_exec_weighted
= delta_exec
;
274 if (unlikely(curr
->load
.weight
!= NICE_0_LOAD
)) {
275 delta_exec_weighted
= calc_delta_fair(delta_exec_weighted
,
278 curr
->vruntime
+= delta_exec_weighted
;
281 * maintain cfs_rq->min_vruntime to be a monotonic increasing
282 * value tracking the leftmost vruntime in the tree.
284 if (first_fair(cfs_rq
)) {
285 vruntime
= min_vruntime(curr
->vruntime
,
286 __pick_next_entity(cfs_rq
)->vruntime
);
288 vruntime
= curr
->vruntime
;
290 cfs_rq
->min_vruntime
=
291 max_vruntime(cfs_rq
->min_vruntime
, vruntime
);
294 static void update_curr(struct cfs_rq
*cfs_rq
)
296 struct sched_entity
*curr
= cfs_rq
->curr
;
297 u64 now
= rq_of(cfs_rq
)->clock
;
298 unsigned long delta_exec
;
304 * Get the amount of time the current task was running
305 * since the last time we changed load (this cannot
306 * overflow on 32 bits):
308 delta_exec
= (unsigned long)(now
- curr
->exec_start
);
310 __update_curr(cfs_rq
, curr
, delta_exec
);
311 curr
->exec_start
= now
;
315 update_stats_wait_start(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
317 schedstat_set(se
->wait_start
, rq_of(cfs_rq
)->clock
);
320 static inline unsigned long
321 calc_weighted(unsigned long delta
, struct sched_entity
*se
)
323 unsigned long weight
= se
->load
.weight
;
325 if (unlikely(weight
!= NICE_0_LOAD
))
326 return (u64
)delta
* se
->load
.weight
>> NICE_0_SHIFT
;
332 * Task is being enqueued - update stats:
334 static void update_stats_enqueue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
337 * Are we enqueueing a waiting task? (for current tasks
338 * a dequeue/enqueue event is a NOP)
340 if (se
!= cfs_rq
->curr
)
341 update_stats_wait_start(cfs_rq
, se
);
345 update_stats_wait_end(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
347 schedstat_set(se
->wait_max
, max(se
->wait_max
,
348 rq_of(cfs_rq
)->clock
- se
->wait_start
));
349 schedstat_set(se
->wait_start
, 0);
353 update_stats_dequeue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
357 * Mark the end of the wait period if dequeueing a
360 if (se
!= cfs_rq
->curr
)
361 update_stats_wait_end(cfs_rq
, se
);
365 * We are picking a new current task - update its stats:
368 update_stats_curr_start(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
371 * We are starting a new run period:
373 se
->exec_start
= rq_of(cfs_rq
)->clock
;
377 * We are descheduling a task - update its stats:
380 update_stats_curr_end(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
385 /**************************************************
386 * Scheduling class queueing methods:
390 account_entity_enqueue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
392 update_load_add(&cfs_rq
->load
, se
->load
.weight
);
393 cfs_rq
->nr_running
++;
398 account_entity_dequeue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
400 update_load_sub(&cfs_rq
->load
, se
->load
.weight
);
401 cfs_rq
->nr_running
--;
405 static void enqueue_sleeper(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
407 #ifdef CONFIG_SCHEDSTATS
408 if (se
->sleep_start
) {
409 u64 delta
= rq_of(cfs_rq
)->clock
- se
->sleep_start
;
414 if (unlikely(delta
> se
->sleep_max
))
415 se
->sleep_max
= delta
;
418 se
->sum_sleep_runtime
+= delta
;
420 if (se
->block_start
) {
421 u64 delta
= rq_of(cfs_rq
)->clock
- se
->block_start
;
426 if (unlikely(delta
> se
->block_max
))
427 se
->block_max
= delta
;
430 se
->sum_sleep_runtime
+= delta
;
433 * Blocking time is in units of nanosecs, so shift by 20 to
434 * get a milliseconds-range estimation of the amount of
435 * time that the task spent sleeping:
437 if (unlikely(prof_on
== SLEEP_PROFILING
)) {
438 struct task_struct
*tsk
= task_of(se
);
440 profile_hits(SLEEP_PROFILING
, (void *)get_wchan(tsk
),
447 static void check_spread(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
449 #ifdef CONFIG_SCHED_DEBUG
450 s64 d
= se
->vruntime
- cfs_rq
->min_vruntime
;
455 if (d
> 3*sysctl_sched_latency
)
456 schedstat_inc(cfs_rq
, nr_spread_over
);
461 place_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int initial
)
465 vruntime
= cfs_rq
->min_vruntime
;
467 if (sched_feat(USE_TREE_AVG
)) {
468 struct sched_entity
*last
= __pick_last_entity(cfs_rq
);
470 vruntime
+= last
->vruntime
;
473 } else if (sched_feat(APPROX_AVG
) && cfs_rq
->nr_running
)
474 vruntime
+= __sched_vslice(cfs_rq
->nr_running
)/2;
476 if (initial
&& sched_feat(START_DEBIT
))
477 vruntime
+= __sched_vslice(cfs_rq
->nr_running
+ 1);
480 if (sched_feat(NEW_FAIR_SLEEPERS
))
481 vruntime
-= sysctl_sched_latency
;
483 vruntime
= max_t(s64
, vruntime
, se
->vruntime
);
486 se
->vruntime
= vruntime
;
491 enqueue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int wakeup
)
494 * Update the fair clock.
499 place_entity(cfs_rq
, se
, 0);
500 enqueue_sleeper(cfs_rq
, se
);
503 update_stats_enqueue(cfs_rq
, se
);
504 check_spread(cfs_rq
, se
);
505 if (se
!= cfs_rq
->curr
)
506 __enqueue_entity(cfs_rq
, se
);
507 account_entity_enqueue(cfs_rq
, se
);
511 dequeue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int sleep
)
513 update_stats_dequeue(cfs_rq
, se
);
515 #ifdef CONFIG_SCHEDSTATS
516 if (entity_is_task(se
)) {
517 struct task_struct
*tsk
= task_of(se
);
519 if (tsk
->state
& TASK_INTERRUPTIBLE
)
520 se
->sleep_start
= rq_of(cfs_rq
)->clock
;
521 if (tsk
->state
& TASK_UNINTERRUPTIBLE
)
522 se
->block_start
= rq_of(cfs_rq
)->clock
;
527 if (se
!= cfs_rq
->curr
)
528 __dequeue_entity(cfs_rq
, se
);
529 account_entity_dequeue(cfs_rq
, se
);
533 * Preempt the current task with a newly woken task if needed:
536 check_preempt_tick(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
)
538 unsigned long ideal_runtime
, delta_exec
;
540 ideal_runtime
= sched_slice(cfs_rq
, curr
);
541 delta_exec
= curr
->sum_exec_runtime
- curr
->prev_sum_exec_runtime
;
542 if (delta_exec
> ideal_runtime
)
543 resched_task(rq_of(cfs_rq
)->curr
);
547 set_next_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
549 /* 'current' is not kept within the tree. */
552 * Any task has to be enqueued before it get to execute on
553 * a CPU. So account for the time it spent waiting on the
556 update_stats_wait_end(cfs_rq
, se
);
557 __dequeue_entity(cfs_rq
, se
);
560 update_stats_curr_start(cfs_rq
, se
);
562 #ifdef CONFIG_SCHEDSTATS
564 * Track our maximum slice length, if the CPU's load is at
565 * least twice that of our own weight (i.e. dont track it
566 * when there are only lesser-weight tasks around):
568 if (rq_of(cfs_rq
)->load
.weight
>= 2*se
->load
.weight
) {
569 se
->slice_max
= max(se
->slice_max
,
570 se
->sum_exec_runtime
- se
->prev_sum_exec_runtime
);
573 se
->prev_sum_exec_runtime
= se
->sum_exec_runtime
;
576 static struct sched_entity
*pick_next_entity(struct cfs_rq
*cfs_rq
)
578 struct sched_entity
*se
= __pick_next_entity(cfs_rq
);
580 set_next_entity(cfs_rq
, se
);
585 static void put_prev_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*prev
)
588 * If still on the runqueue then deactivate_task()
589 * was not called and update_curr() has to be done:
594 update_stats_curr_end(cfs_rq
, prev
);
596 check_spread(cfs_rq
, prev
);
598 update_stats_wait_start(cfs_rq
, prev
);
599 /* Put 'current' back into the tree. */
600 __enqueue_entity(cfs_rq
, prev
);
605 static void entity_tick(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
)
608 * Update run-time statistics of the 'current'.
612 if (cfs_rq
->nr_running
> 1)
613 check_preempt_tick(cfs_rq
, curr
);
616 /**************************************************
617 * CFS operations on tasks:
620 #ifdef CONFIG_FAIR_GROUP_SCHED
622 /* Walk up scheduling entities hierarchy */
623 #define for_each_sched_entity(se) \
624 for (; se; se = se->parent)
626 static inline struct cfs_rq
*task_cfs_rq(struct task_struct
*p
)
631 /* runqueue on which this entity is (to be) queued */
632 static inline struct cfs_rq
*cfs_rq_of(struct sched_entity
*se
)
637 /* runqueue "owned" by this group */
638 static inline struct cfs_rq
*group_cfs_rq(struct sched_entity
*grp
)
643 /* Given a group's cfs_rq on one cpu, return its corresponding cfs_rq on
644 * another cpu ('this_cpu')
646 static inline struct cfs_rq
*cpu_cfs_rq(struct cfs_rq
*cfs_rq
, int this_cpu
)
648 return cfs_rq
->tg
->cfs_rq
[this_cpu
];
651 /* Iterate thr' all leaf cfs_rq's on a runqueue */
652 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
653 list_for_each_entry(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
655 /* Do the two (enqueued) tasks belong to the same group ? */
656 static inline int is_same_group(struct task_struct
*curr
, struct task_struct
*p
)
658 if (curr
->se
.cfs_rq
== p
->se
.cfs_rq
)
664 #else /* CONFIG_FAIR_GROUP_SCHED */
666 #define for_each_sched_entity(se) \
667 for (; se; se = NULL)
669 static inline struct cfs_rq
*task_cfs_rq(struct task_struct
*p
)
671 return &task_rq(p
)->cfs
;
674 static inline struct cfs_rq
*cfs_rq_of(struct sched_entity
*se
)
676 struct task_struct
*p
= task_of(se
);
677 struct rq
*rq
= task_rq(p
);
682 /* runqueue "owned" by this group */
683 static inline struct cfs_rq
*group_cfs_rq(struct sched_entity
*grp
)
688 static inline struct cfs_rq
*cpu_cfs_rq(struct cfs_rq
*cfs_rq
, int this_cpu
)
690 return &cpu_rq(this_cpu
)->cfs
;
693 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
694 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
696 static inline int is_same_group(struct task_struct
*curr
, struct task_struct
*p
)
701 #endif /* CONFIG_FAIR_GROUP_SCHED */
704 * The enqueue_task method is called before nr_running is
705 * increased. Here we update the fair scheduling stats and
706 * then put the task into the rbtree:
708 static void enqueue_task_fair(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
710 struct cfs_rq
*cfs_rq
;
711 struct sched_entity
*se
= &p
->se
;
713 for_each_sched_entity(se
) {
716 cfs_rq
= cfs_rq_of(se
);
717 enqueue_entity(cfs_rq
, se
, wakeup
);
722 * The dequeue_task method is called before nr_running is
723 * decreased. We remove the task from the rbtree and
724 * update the fair scheduling stats:
726 static void dequeue_task_fair(struct rq
*rq
, struct task_struct
*p
, int sleep
)
728 struct cfs_rq
*cfs_rq
;
729 struct sched_entity
*se
= &p
->se
;
731 for_each_sched_entity(se
) {
732 cfs_rq
= cfs_rq_of(se
);
733 dequeue_entity(cfs_rq
, se
, sleep
);
734 /* Don't dequeue parent if it has other entities besides us */
735 if (cfs_rq
->load
.weight
)
741 * sched_yield() support is very simple - we dequeue and enqueue.
743 * If compat_yield is turned on then we requeue to the end of the tree.
745 static void yield_task_fair(struct rq
*rq
)
747 struct cfs_rq
*cfs_rq
= task_cfs_rq(rq
->curr
);
748 struct sched_entity
*rightmost
, *se
= &rq
->curr
->se
;
751 * Are we the only task in the tree?
753 if (unlikely(cfs_rq
->nr_running
== 1))
756 if (likely(!sysctl_sched_compat_yield
)) {
757 __update_rq_clock(rq
);
759 * Dequeue and enqueue the task to update its
760 * position within the tree:
767 * Find the rightmost entry in the rbtree:
769 rightmost
= __pick_last_entity(cfs_rq
);
771 * Already in the rightmost position?
773 if (unlikely(rightmost
->vruntime
< se
->vruntime
))
777 * Minimally necessary key value to be last in the tree:
778 * Upon rescheduling, sched_class::put_prev_task() will place
779 * 'current' within the tree based on its new key value.
781 se
->vruntime
= rightmost
->vruntime
+ 1;
785 * Preempt the current task with a newly woken task if needed:
787 static void check_preempt_wakeup(struct rq
*rq
, struct task_struct
*p
)
789 struct task_struct
*curr
= rq
->curr
;
790 struct cfs_rq
*cfs_rq
= task_cfs_rq(curr
), *pcfs_rq
;
791 struct sched_entity
*se
= &curr
->se
, *pse
= &p
->se
;
793 if (unlikely(rt_prio(p
->prio
))) {
800 for_each_sched_entity(se
) {
801 cfs_rq
= cfs_rq_of(se
);
802 pcfs_rq
= cfs_rq_of(pse
);
804 if (cfs_rq
== pcfs_rq
) {
805 s64 delta
= se
->vruntime
- pse
->vruntime
;
807 if (delta
> (s64
)sysctl_sched_wakeup_granularity
)
811 #ifdef CONFIG_FAIR_GROUP_SCHED
817 static struct task_struct
*pick_next_task_fair(struct rq
*rq
)
819 struct cfs_rq
*cfs_rq
= &rq
->cfs
;
820 struct sched_entity
*se
;
822 if (unlikely(!cfs_rq
->nr_running
))
826 se
= pick_next_entity(cfs_rq
);
827 cfs_rq
= group_cfs_rq(se
);
834 * Account for a descheduled task:
836 static void put_prev_task_fair(struct rq
*rq
, struct task_struct
*prev
)
838 struct sched_entity
*se
= &prev
->se
;
839 struct cfs_rq
*cfs_rq
;
841 for_each_sched_entity(se
) {
842 cfs_rq
= cfs_rq_of(se
);
843 put_prev_entity(cfs_rq
, se
);
847 /**************************************************
848 * Fair scheduling class load-balancing methods:
852 * Load-balancing iterator. Note: while the runqueue stays locked
853 * during the whole iteration, the current task might be
854 * dequeued so the iterator has to be dequeue-safe. Here we
855 * achieve that by always pre-iterating before returning
858 static inline struct task_struct
*
859 __load_balance_iterator(struct cfs_rq
*cfs_rq
, struct rb_node
*curr
)
861 struct task_struct
*p
;
866 p
= rb_entry(curr
, struct task_struct
, se
.run_node
);
867 cfs_rq
->rb_load_balance_curr
= rb_next(curr
);
872 static struct task_struct
*load_balance_start_fair(void *arg
)
874 struct cfs_rq
*cfs_rq
= arg
;
876 return __load_balance_iterator(cfs_rq
, first_fair(cfs_rq
));
879 static struct task_struct
*load_balance_next_fair(void *arg
)
881 struct cfs_rq
*cfs_rq
= arg
;
883 return __load_balance_iterator(cfs_rq
, cfs_rq
->rb_load_balance_curr
);
886 #ifdef CONFIG_FAIR_GROUP_SCHED
887 static int cfs_rq_best_prio(struct cfs_rq
*cfs_rq
)
889 struct sched_entity
*curr
;
890 struct task_struct
*p
;
892 if (!cfs_rq
->nr_running
)
897 curr
= __pick_next_entity(cfs_rq
);
906 load_balance_fair(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
907 unsigned long max_nr_move
, unsigned long max_load_move
,
908 struct sched_domain
*sd
, enum cpu_idle_type idle
,
909 int *all_pinned
, int *this_best_prio
)
911 struct cfs_rq
*busy_cfs_rq
;
912 unsigned long load_moved
, total_nr_moved
= 0, nr_moved
;
913 long rem_load_move
= max_load_move
;
914 struct rq_iterator cfs_rq_iterator
;
916 cfs_rq_iterator
.start
= load_balance_start_fair
;
917 cfs_rq_iterator
.next
= load_balance_next_fair
;
919 for_each_leaf_cfs_rq(busiest
, busy_cfs_rq
) {
920 #ifdef CONFIG_FAIR_GROUP_SCHED
921 struct cfs_rq
*this_cfs_rq
;
923 unsigned long maxload
;
925 this_cfs_rq
= cpu_cfs_rq(busy_cfs_rq
, this_cpu
);
927 imbalance
= busy_cfs_rq
->load
.weight
- this_cfs_rq
->load
.weight
;
928 /* Don't pull if this_cfs_rq has more load than busy_cfs_rq */
932 /* Don't pull more than imbalance/2 */
934 maxload
= min(rem_load_move
, imbalance
);
936 *this_best_prio
= cfs_rq_best_prio(this_cfs_rq
);
938 # define maxload rem_load_move
940 /* pass busy_cfs_rq argument into
941 * load_balance_[start|next]_fair iterators
943 cfs_rq_iterator
.arg
= busy_cfs_rq
;
944 nr_moved
= balance_tasks(this_rq
, this_cpu
, busiest
,
945 max_nr_move
, maxload
, sd
, idle
, all_pinned
,
946 &load_moved
, this_best_prio
, &cfs_rq_iterator
);
948 total_nr_moved
+= nr_moved
;
949 max_nr_move
-= nr_moved
;
950 rem_load_move
-= load_moved
;
952 if (max_nr_move
<= 0 || rem_load_move
<= 0)
956 return max_load_move
- rem_load_move
;
960 * scheduler tick hitting a task of our scheduling class:
962 static void task_tick_fair(struct rq
*rq
, struct task_struct
*curr
)
964 struct cfs_rq
*cfs_rq
;
965 struct sched_entity
*se
= &curr
->se
;
967 for_each_sched_entity(se
) {
968 cfs_rq
= cfs_rq_of(se
);
969 entity_tick(cfs_rq
, se
);
973 #define swap(a,b) do { typeof(a) tmp = (a); (a) = (b); (b) = tmp; } while (0)
976 * Share the fairness runtime between parent and child, thus the
977 * total amount of pressure for CPU stays equal - new tasks
978 * get a chance to run but frequent forkers are not allowed to
979 * monopolize the CPU. Note: the parent runqueue is locked,
980 * the child is not running yet.
982 static void task_new_fair(struct rq
*rq
, struct task_struct
*p
)
984 struct cfs_rq
*cfs_rq
= task_cfs_rq(p
);
985 struct sched_entity
*se
= &p
->se
, *curr
= cfs_rq
->curr
;
987 sched_info_queued(p
);
990 place_entity(cfs_rq
, se
, 1);
992 if (sysctl_sched_child_runs_first
&&
993 curr
->vruntime
< se
->vruntime
) {
995 * Upon rescheduling, sched_class::put_prev_task() will place
996 * 'current' within the tree based on its new key value.
998 swap(curr
->vruntime
, se
->vruntime
);
1001 update_stats_enqueue(cfs_rq
, se
);
1002 check_spread(cfs_rq
, se
);
1003 check_spread(cfs_rq
, curr
);
1004 __enqueue_entity(cfs_rq
, se
);
1005 account_entity_enqueue(cfs_rq
, se
);
1006 resched_task(rq
->curr
);
1009 /* Account for a task changing its policy or group.
1011 * This routine is mostly called to set cfs_rq->curr field when a task
1012 * migrates between groups/classes.
1014 static void set_curr_task_fair(struct rq
*rq
)
1016 struct sched_entity
*se
= &rq
->curr
->se
;
1018 for_each_sched_entity(se
)
1019 set_next_entity(cfs_rq_of(se
), se
);
1023 * All the scheduling class methods:
1025 struct sched_class fair_sched_class __read_mostly
= {
1026 .enqueue_task
= enqueue_task_fair
,
1027 .dequeue_task
= dequeue_task_fair
,
1028 .yield_task
= yield_task_fair
,
1030 .check_preempt_curr
= check_preempt_wakeup
,
1032 .pick_next_task
= pick_next_task_fair
,
1033 .put_prev_task
= put_prev_task_fair
,
1035 .load_balance
= load_balance_fair
,
1037 .set_curr_task
= set_curr_task_fair
,
1038 .task_tick
= task_tick_fair
,
1039 .task_new
= task_new_fair
,
1042 #ifdef CONFIG_SCHED_DEBUG
1043 static void print_cfs_stats(struct seq_file
*m
, int cpu
)
1045 struct cfs_rq
*cfs_rq
;
1047 #ifdef CONFIG_FAIR_GROUP_SCHED
1048 print_cfs_rq(m
, cpu
, &cpu_rq(cpu
)->cfs
);
1050 for_each_leaf_cfs_rq(cpu_rq(cpu
), cfs_rq
)
1051 print_cfs_rq(m
, cpu
, cfs_rq
);