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 unsigned int sysctl_sched_min_granularity __read_mostly
= 2000000ULL;
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 unsigned int sysctl_sched_runtime_limit __read_mostly
;
81 extern struct sched_class fair_sched_class
;
83 /**************************************************************
84 * CFS operations on generic schedulable entities:
87 #ifdef CONFIG_FAIR_GROUP_SCHED
89 /* cpu runqueue to which this cfs_rq is attached */
90 static inline struct rq
*rq_of(struct cfs_rq
*cfs_rq
)
95 /* An entity is a task if it doesn't "own" a runqueue */
96 #define entity_is_task(se) (!se->my_q)
98 #else /* CONFIG_FAIR_GROUP_SCHED */
100 static inline struct rq
*rq_of(struct cfs_rq
*cfs_rq
)
102 return container_of(cfs_rq
, struct rq
, cfs
);
105 #define entity_is_task(se) 1
107 #endif /* CONFIG_FAIR_GROUP_SCHED */
109 static inline struct task_struct
*task_of(struct sched_entity
*se
)
111 return container_of(se
, struct task_struct
, se
);
115 /**************************************************************
116 * Scheduling class tree data structure manipulation methods:
120 max_vruntime(u64 min_vruntime
, u64 vruntime
)
122 if ((vruntime
> min_vruntime
) ||
123 (min_vruntime
> (1ULL << 61) && vruntime
< (1ULL << 50)))
124 min_vruntime
= vruntime
;
130 set_leftmost(struct cfs_rq
*cfs_rq
, struct rb_node
*leftmost
)
132 struct sched_entity
*se
;
134 cfs_rq
->rb_leftmost
= leftmost
;
136 se
= rb_entry(leftmost
, struct sched_entity
, run_node
);
140 entity_key(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
142 return se
->fair_key
- cfs_rq
->min_vruntime
;
146 * Enqueue an entity into the rb-tree:
149 __enqueue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
151 struct rb_node
**link
= &cfs_rq
->tasks_timeline
.rb_node
;
152 struct rb_node
*parent
= NULL
;
153 struct sched_entity
*entry
;
154 s64 key
= entity_key(cfs_rq
, se
);
158 * Find the right place in the rbtree:
162 entry
= rb_entry(parent
, struct sched_entity
, run_node
);
164 * We dont care about collisions. Nodes with
165 * the same key stay together.
167 if (key
< entity_key(cfs_rq
, entry
)) {
168 link
= &parent
->rb_left
;
170 link
= &parent
->rb_right
;
176 * Maintain a cache of leftmost tree entries (it is frequently
180 set_leftmost(cfs_rq
, &se
->run_node
);
182 rb_link_node(&se
->run_node
, parent
, link
);
183 rb_insert_color(&se
->run_node
, &cfs_rq
->tasks_timeline
);
184 update_load_add(&cfs_rq
->load
, se
->load
.weight
);
185 cfs_rq
->nr_running
++;
190 __dequeue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
192 if (cfs_rq
->rb_leftmost
== &se
->run_node
)
193 set_leftmost(cfs_rq
, rb_next(&se
->run_node
));
195 rb_erase(&se
->run_node
, &cfs_rq
->tasks_timeline
);
196 update_load_sub(&cfs_rq
->load
, se
->load
.weight
);
197 cfs_rq
->nr_running
--;
201 static inline struct rb_node
*first_fair(struct cfs_rq
*cfs_rq
)
203 return cfs_rq
->rb_leftmost
;
206 static struct sched_entity
*__pick_next_entity(struct cfs_rq
*cfs_rq
)
208 return rb_entry(first_fair(cfs_rq
), struct sched_entity
, run_node
);
211 static inline struct sched_entity
*__pick_last_entity(struct cfs_rq
*cfs_rq
)
213 struct rb_node
**link
= &cfs_rq
->tasks_timeline
.rb_node
;
214 struct sched_entity
*se
= NULL
;
215 struct rb_node
*parent
;
219 se
= rb_entry(parent
, struct sched_entity
, run_node
);
220 link
= &parent
->rb_right
;
226 /**************************************************************
227 * Scheduling class statistics methods:
230 static u64
__sched_period(unsigned long nr_running
)
232 u64 period
= sysctl_sched_latency
;
233 unsigned long nr_latency
=
234 sysctl_sched_latency
/ sysctl_sched_min_granularity
;
236 if (unlikely(nr_running
> nr_latency
)) {
237 period
*= nr_running
;
238 do_div(period
, nr_latency
);
244 static u64
sched_slice(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
246 u64 period
= __sched_period(cfs_rq
->nr_running
);
248 period
*= se
->load
.weight
;
249 do_div(period
, cfs_rq
->load
.weight
);
255 * Update the current task's runtime statistics. Skip current tasks that
256 * are not in our scheduling class.
259 __update_curr(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
,
260 unsigned long delta_exec
)
262 unsigned long delta_exec_weighted
;
263 u64 next_vruntime
, min_vruntime
;
265 schedstat_set(curr
->exec_max
, max((u64
)delta_exec
, curr
->exec_max
));
267 curr
->sum_exec_runtime
+= delta_exec
;
268 schedstat_add(cfs_rq
, exec_clock
, delta_exec
);
269 delta_exec_weighted
= delta_exec
;
270 if (unlikely(curr
->load
.weight
!= NICE_0_LOAD
)) {
271 delta_exec_weighted
= calc_delta_fair(delta_exec_weighted
,
274 curr
->vruntime
+= delta_exec_weighted
;
277 * maintain cfs_rq->min_vruntime to be a monotonic increasing
278 * value tracking the leftmost vruntime in the tree.
280 if (first_fair(cfs_rq
)) {
281 next_vruntime
= __pick_next_entity(cfs_rq
)->vruntime
;
283 /* min_vruntime() := !max_vruntime() */
284 min_vruntime
= max_vruntime(curr
->vruntime
, next_vruntime
);
285 if (min_vruntime
== next_vruntime
)
286 min_vruntime
= curr
->vruntime
;
288 min_vruntime
= next_vruntime
;
290 min_vruntime
= curr
->vruntime
;
292 cfs_rq
->min_vruntime
=
293 max_vruntime(cfs_rq
->min_vruntime
, min_vruntime
);
296 static void update_curr(struct cfs_rq
*cfs_rq
)
298 struct sched_entity
*curr
= cfs_rq
->curr
;
299 u64 now
= rq_of(cfs_rq
)->clock
;
300 unsigned long delta_exec
;
306 * Get the amount of time the current task was running
307 * since the last time we changed load (this cannot
308 * overflow on 32 bits):
310 delta_exec
= (unsigned long)(now
- curr
->exec_start
);
312 __update_curr(cfs_rq
, curr
, delta_exec
);
313 curr
->exec_start
= now
;
317 update_stats_wait_start(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
319 schedstat_set(se
->wait_start
, rq_of(cfs_rq
)->clock
);
322 static inline unsigned long
323 calc_weighted(unsigned long delta
, struct sched_entity
*se
)
325 unsigned long weight
= se
->load
.weight
;
327 if (unlikely(weight
!= NICE_0_LOAD
))
328 return (u64
)delta
* se
->load
.weight
>> NICE_0_SHIFT
;
334 * Task is being enqueued - update stats:
336 static void update_stats_enqueue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
339 * Are we enqueueing a waiting task? (for current tasks
340 * a dequeue/enqueue event is a NOP)
342 if (se
!= cfs_rq
->curr
)
343 update_stats_wait_start(cfs_rq
, se
);
347 se
->fair_key
= se
->vruntime
;
351 update_stats_wait_end(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
353 schedstat_set(se
->wait_max
, max(se
->wait_max
,
354 rq_of(cfs_rq
)->clock
- se
->wait_start
));
355 schedstat_set(se
->wait_start
, 0);
359 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:
395 static void enqueue_sleeper(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
397 #ifdef CONFIG_SCHEDSTATS
398 if (se
->sleep_start
) {
399 u64 delta
= rq_of(cfs_rq
)->clock
- se
->sleep_start
;
404 if (unlikely(delta
> se
->sleep_max
))
405 se
->sleep_max
= delta
;
408 se
->sum_sleep_runtime
+= delta
;
410 if (se
->block_start
) {
411 u64 delta
= rq_of(cfs_rq
)->clock
- se
->block_start
;
416 if (unlikely(delta
> se
->block_max
))
417 se
->block_max
= delta
;
420 se
->sum_sleep_runtime
+= delta
;
423 * Blocking time is in units of nanosecs, so shift by 20 to
424 * get a milliseconds-range estimation of the amount of
425 * time that the task spent sleeping:
427 if (unlikely(prof_on
== SLEEP_PROFILING
)) {
428 struct task_struct
*tsk
= task_of(se
);
430 profile_hits(SLEEP_PROFILING
, (void *)get_wchan(tsk
),
438 place_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int initial
)
440 u64 min_runtime
, latency
;
442 min_runtime
= cfs_rq
->min_vruntime
;
444 if (sched_feat(USE_TREE_AVG
)) {
445 struct sched_entity
*last
= __pick_last_entity(cfs_rq
);
447 min_runtime
= __pick_next_entity(cfs_rq
)->vruntime
;
448 min_runtime
+= last
->vruntime
;
451 } else if (sched_feat(APPROX_AVG
))
452 min_runtime
+= sysctl_sched_latency
/2;
454 if (initial
&& sched_feat(START_DEBIT
))
455 min_runtime
+= sched_slice(cfs_rq
, se
);
457 if (!initial
&& sched_feat(NEW_FAIR_SLEEPERS
)) {
458 latency
= sysctl_sched_latency
;
459 if (min_runtime
> latency
)
460 min_runtime
-= latency
;
465 se
->vruntime
= max(se
->vruntime
, min_runtime
);
469 enqueue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int wakeup
)
472 * Update the fair clock.
477 place_entity(cfs_rq
, se
, 0);
478 enqueue_sleeper(cfs_rq
, se
);
481 update_stats_enqueue(cfs_rq
, se
);
482 __enqueue_entity(cfs_rq
, se
);
486 dequeue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int sleep
)
488 update_stats_dequeue(cfs_rq
, se
);
489 #ifdef CONFIG_SCHEDSTATS
491 if (entity_is_task(se
)) {
492 struct task_struct
*tsk
= task_of(se
);
494 if (tsk
->state
& TASK_INTERRUPTIBLE
)
495 se
->sleep_start
= rq_of(cfs_rq
)->clock
;
496 if (tsk
->state
& TASK_UNINTERRUPTIBLE
)
497 se
->block_start
= rq_of(cfs_rq
)->clock
;
501 __dequeue_entity(cfs_rq
, se
);
505 * Preempt the current task with a newly woken task if needed:
508 check_preempt_tick(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
)
510 unsigned long ideal_runtime
, delta_exec
;
512 ideal_runtime
= sched_slice(cfs_rq
, curr
);
513 delta_exec
= curr
->sum_exec_runtime
- curr
->prev_sum_exec_runtime
;
514 if (delta_exec
> ideal_runtime
)
515 resched_task(rq_of(cfs_rq
)->curr
);
519 set_next_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
522 * Any task has to be enqueued before it get to execute on
523 * a CPU. So account for the time it spent waiting on the
526 update_stats_wait_end(cfs_rq
, se
);
527 update_stats_curr_start(cfs_rq
, se
);
529 #ifdef CONFIG_SCHEDSTATS
531 * Track our maximum slice length, if the CPU's load is at
532 * least twice that of our own weight (i.e. dont track it
533 * when there are only lesser-weight tasks around):
535 if (rq_of(cfs_rq
)->load
.weight
>= 2*se
->load
.weight
) {
536 se
->slice_max
= max(se
->slice_max
,
537 se
->sum_exec_runtime
- se
->prev_sum_exec_runtime
);
540 se
->prev_sum_exec_runtime
= se
->sum_exec_runtime
;
543 static struct sched_entity
*pick_next_entity(struct cfs_rq
*cfs_rq
)
545 struct sched_entity
*se
= __pick_next_entity(cfs_rq
);
547 set_next_entity(cfs_rq
, se
);
552 static void put_prev_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*prev
)
555 * If still on the runqueue then deactivate_task()
556 * was not called and update_curr() has to be done:
561 update_stats_curr_end(cfs_rq
, prev
);
564 update_stats_wait_start(cfs_rq
, prev
);
568 static void entity_tick(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
)
571 * Dequeue and enqueue the task to update its
572 * position within the tree:
574 dequeue_entity(cfs_rq
, curr
, 0);
575 enqueue_entity(cfs_rq
, curr
, 0);
577 if (cfs_rq
->nr_running
> 1)
578 check_preempt_tick(cfs_rq
, curr
);
581 /**************************************************
582 * CFS operations on tasks:
585 #ifdef CONFIG_FAIR_GROUP_SCHED
587 /* Walk up scheduling entities hierarchy */
588 #define for_each_sched_entity(se) \
589 for (; se; se = se->parent)
591 static inline struct cfs_rq
*task_cfs_rq(struct task_struct
*p
)
596 /* runqueue on which this entity is (to be) queued */
597 static inline struct cfs_rq
*cfs_rq_of(struct sched_entity
*se
)
602 /* runqueue "owned" by this group */
603 static inline struct cfs_rq
*group_cfs_rq(struct sched_entity
*grp
)
608 /* Given a group's cfs_rq on one cpu, return its corresponding cfs_rq on
609 * another cpu ('this_cpu')
611 static inline struct cfs_rq
*cpu_cfs_rq(struct cfs_rq
*cfs_rq
, int this_cpu
)
613 /* A later patch will take group into account */
614 return &cpu_rq(this_cpu
)->cfs
;
617 /* Iterate thr' all leaf cfs_rq's on a runqueue */
618 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
619 list_for_each_entry(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
621 /* Do the two (enqueued) tasks belong to the same group ? */
622 static inline int is_same_group(struct task_struct
*curr
, struct task_struct
*p
)
624 if (curr
->se
.cfs_rq
== p
->se
.cfs_rq
)
630 #else /* CONFIG_FAIR_GROUP_SCHED */
632 #define for_each_sched_entity(se) \
633 for (; se; se = NULL)
635 static inline struct cfs_rq
*task_cfs_rq(struct task_struct
*p
)
637 return &task_rq(p
)->cfs
;
640 static inline struct cfs_rq
*cfs_rq_of(struct sched_entity
*se
)
642 struct task_struct
*p
= task_of(se
);
643 struct rq
*rq
= task_rq(p
);
648 /* runqueue "owned" by this group */
649 static inline struct cfs_rq
*group_cfs_rq(struct sched_entity
*grp
)
654 static inline struct cfs_rq
*cpu_cfs_rq(struct cfs_rq
*cfs_rq
, int this_cpu
)
656 return &cpu_rq(this_cpu
)->cfs
;
659 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
660 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
662 static inline int is_same_group(struct task_struct
*curr
, struct task_struct
*p
)
667 #endif /* CONFIG_FAIR_GROUP_SCHED */
670 * The enqueue_task method is called before nr_running is
671 * increased. Here we update the fair scheduling stats and
672 * then put the task into the rbtree:
674 static void enqueue_task_fair(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
676 struct cfs_rq
*cfs_rq
;
677 struct sched_entity
*se
= &p
->se
;
679 for_each_sched_entity(se
) {
682 cfs_rq
= cfs_rq_of(se
);
683 enqueue_entity(cfs_rq
, se
, wakeup
);
688 * The dequeue_task method is called before nr_running is
689 * decreased. We remove the task from the rbtree and
690 * update the fair scheduling stats:
692 static void dequeue_task_fair(struct rq
*rq
, struct task_struct
*p
, int sleep
)
694 struct cfs_rq
*cfs_rq
;
695 struct sched_entity
*se
= &p
->se
;
697 for_each_sched_entity(se
) {
698 cfs_rq
= cfs_rq_of(se
);
699 dequeue_entity(cfs_rq
, se
, sleep
);
700 /* Don't dequeue parent if it has other entities besides us */
701 if (cfs_rq
->load
.weight
)
707 * sched_yield() support is very simple - we dequeue and enqueue.
709 * If compat_yield is turned on then we requeue to the end of the tree.
711 static void yield_task_fair(struct rq
*rq
, struct task_struct
*p
)
713 struct cfs_rq
*cfs_rq
= task_cfs_rq(p
);
714 struct rb_node
**link
= &cfs_rq
->tasks_timeline
.rb_node
;
715 struct sched_entity
*rightmost
, *se
= &p
->se
;
716 struct rb_node
*parent
;
719 * Are we the only task in the tree?
721 if (unlikely(cfs_rq
->nr_running
== 1))
724 if (likely(!sysctl_sched_compat_yield
)) {
725 __update_rq_clock(rq
);
727 * Dequeue and enqueue the task to update its
728 * position within the tree:
730 dequeue_entity(cfs_rq
, &p
->se
, 0);
731 enqueue_entity(cfs_rq
, &p
->se
, 0);
736 * Find the rightmost entry in the rbtree:
740 link
= &parent
->rb_right
;
743 rightmost
= rb_entry(parent
, struct sched_entity
, run_node
);
745 * Already in the rightmost position?
747 if (unlikely(rightmost
== se
))
751 * Minimally necessary key value to be last in the tree:
753 se
->fair_key
= rightmost
->fair_key
+ 1;
755 if (cfs_rq
->rb_leftmost
== &se
->run_node
)
756 cfs_rq
->rb_leftmost
= rb_next(&se
->run_node
);
758 * Relink the task to the rightmost position:
760 rb_erase(&se
->run_node
, &cfs_rq
->tasks_timeline
);
761 rb_link_node(&se
->run_node
, parent
, link
);
762 rb_insert_color(&se
->run_node
, &cfs_rq
->tasks_timeline
);
766 * Preempt the current task with a newly woken task if needed:
768 static void check_preempt_wakeup(struct rq
*rq
, struct task_struct
*p
)
770 struct task_struct
*curr
= rq
->curr
;
771 struct cfs_rq
*cfs_rq
= task_cfs_rq(curr
);
773 if (unlikely(rt_prio(p
->prio
))) {
779 if (is_same_group(curr
, p
)) {
780 s64 delta
= curr
->se
.vruntime
- p
->se
.vruntime
;
782 if (delta
> (s64
)sysctl_sched_wakeup_granularity
)
787 static struct task_struct
*pick_next_task_fair(struct rq
*rq
)
789 struct cfs_rq
*cfs_rq
= &rq
->cfs
;
790 struct sched_entity
*se
;
792 if (unlikely(!cfs_rq
->nr_running
))
796 se
= pick_next_entity(cfs_rq
);
797 cfs_rq
= group_cfs_rq(se
);
804 * Account for a descheduled task:
806 static void put_prev_task_fair(struct rq
*rq
, struct task_struct
*prev
)
808 struct sched_entity
*se
= &prev
->se
;
809 struct cfs_rq
*cfs_rq
;
811 for_each_sched_entity(se
) {
812 cfs_rq
= cfs_rq_of(se
);
813 put_prev_entity(cfs_rq
, se
);
817 /**************************************************
818 * Fair scheduling class load-balancing methods:
822 * Load-balancing iterator. Note: while the runqueue stays locked
823 * during the whole iteration, the current task might be
824 * dequeued so the iterator has to be dequeue-safe. Here we
825 * achieve that by always pre-iterating before returning
828 static inline struct task_struct
*
829 __load_balance_iterator(struct cfs_rq
*cfs_rq
, struct rb_node
*curr
)
831 struct task_struct
*p
;
836 p
= rb_entry(curr
, struct task_struct
, se
.run_node
);
837 cfs_rq
->rb_load_balance_curr
= rb_next(curr
);
842 static struct task_struct
*load_balance_start_fair(void *arg
)
844 struct cfs_rq
*cfs_rq
= arg
;
846 return __load_balance_iterator(cfs_rq
, first_fair(cfs_rq
));
849 static struct task_struct
*load_balance_next_fair(void *arg
)
851 struct cfs_rq
*cfs_rq
= arg
;
853 return __load_balance_iterator(cfs_rq
, cfs_rq
->rb_load_balance_curr
);
856 #ifdef CONFIG_FAIR_GROUP_SCHED
857 static int cfs_rq_best_prio(struct cfs_rq
*cfs_rq
)
859 struct sched_entity
*curr
;
860 struct task_struct
*p
;
862 if (!cfs_rq
->nr_running
)
865 curr
= __pick_next_entity(cfs_rq
);
873 load_balance_fair(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
874 unsigned long max_nr_move
, unsigned long max_load_move
,
875 struct sched_domain
*sd
, enum cpu_idle_type idle
,
876 int *all_pinned
, int *this_best_prio
)
878 struct cfs_rq
*busy_cfs_rq
;
879 unsigned long load_moved
, total_nr_moved
= 0, nr_moved
;
880 long rem_load_move
= max_load_move
;
881 struct rq_iterator cfs_rq_iterator
;
883 cfs_rq_iterator
.start
= load_balance_start_fair
;
884 cfs_rq_iterator
.next
= load_balance_next_fair
;
886 for_each_leaf_cfs_rq(busiest
, busy_cfs_rq
) {
887 #ifdef CONFIG_FAIR_GROUP_SCHED
888 struct cfs_rq
*this_cfs_rq
;
890 unsigned long maxload
;
892 this_cfs_rq
= cpu_cfs_rq(busy_cfs_rq
, this_cpu
);
894 imbalance
= busy_cfs_rq
->load
.weight
- this_cfs_rq
->load
.weight
;
895 /* Don't pull if this_cfs_rq has more load than busy_cfs_rq */
899 /* Don't pull more than imbalance/2 */
901 maxload
= min(rem_load_move
, imbalance
);
903 *this_best_prio
= cfs_rq_best_prio(this_cfs_rq
);
905 # define maxload rem_load_move
907 /* pass busy_cfs_rq argument into
908 * load_balance_[start|next]_fair iterators
910 cfs_rq_iterator
.arg
= busy_cfs_rq
;
911 nr_moved
= balance_tasks(this_rq
, this_cpu
, busiest
,
912 max_nr_move
, maxload
, sd
, idle
, all_pinned
,
913 &load_moved
, this_best_prio
, &cfs_rq_iterator
);
915 total_nr_moved
+= nr_moved
;
916 max_nr_move
-= nr_moved
;
917 rem_load_move
-= load_moved
;
919 if (max_nr_move
<= 0 || rem_load_move
<= 0)
923 return max_load_move
- rem_load_move
;
927 * scheduler tick hitting a task of our scheduling class:
929 static void task_tick_fair(struct rq
*rq
, struct task_struct
*curr
)
931 struct cfs_rq
*cfs_rq
;
932 struct sched_entity
*se
= &curr
->se
;
934 for_each_sched_entity(se
) {
935 cfs_rq
= cfs_rq_of(se
);
936 entity_tick(cfs_rq
, se
);
940 #define swap(a,b) do { typeof(a) tmp = (a); (a) = (b); (b) = tmp; } while (0)
943 * Share the fairness runtime between parent and child, thus the
944 * total amount of pressure for CPU stays equal - new tasks
945 * get a chance to run but frequent forkers are not allowed to
946 * monopolize the CPU. Note: the parent runqueue is locked,
947 * the child is not running yet.
949 static void task_new_fair(struct rq
*rq
, struct task_struct
*p
)
951 struct cfs_rq
*cfs_rq
= task_cfs_rq(p
);
952 struct sched_entity
*se
= &p
->se
, *curr
= cfs_rq
->curr
;
954 sched_info_queued(p
);
957 place_entity(cfs_rq
, se
, 1);
959 if (sysctl_sched_child_runs_first
&&
960 curr
->vruntime
< se
->vruntime
) {
962 dequeue_entity(cfs_rq
, curr
, 0);
963 swap(curr
->vruntime
, se
->vruntime
);
964 enqueue_entity(cfs_rq
, curr
, 0);
967 update_stats_enqueue(cfs_rq
, se
);
968 __enqueue_entity(cfs_rq
, se
);
969 resched_task(rq
->curr
);
972 #ifdef CONFIG_FAIR_GROUP_SCHED
973 /* Account for a task changing its policy or group.
975 * This routine is mostly called to set cfs_rq->curr field when a task
976 * migrates between groups/classes.
978 static void set_curr_task_fair(struct rq
*rq
)
980 struct sched_entity
*se
= &rq
->curr
->se
;
982 for_each_sched_entity(se
)
983 set_next_entity(cfs_rq_of(se
), se
);
986 static void set_curr_task_fair(struct rq
*rq
)
992 * All the scheduling class methods:
994 struct sched_class fair_sched_class __read_mostly
= {
995 .enqueue_task
= enqueue_task_fair
,
996 .dequeue_task
= dequeue_task_fair
,
997 .yield_task
= yield_task_fair
,
999 .check_preempt_curr
= check_preempt_wakeup
,
1001 .pick_next_task
= pick_next_task_fair
,
1002 .put_prev_task
= put_prev_task_fair
,
1004 .load_balance
= load_balance_fair
,
1006 .set_curr_task
= set_curr_task_fair
,
1007 .task_tick
= task_tick_fair
,
1008 .task_new
= task_new_fair
,
1011 #ifdef CONFIG_SCHED_DEBUG
1012 static void print_cfs_stats(struct seq_file
*m
, int cpu
)
1014 struct cfs_rq
*cfs_rq
;
1016 for_each_leaf_cfs_rq(cpu_rq(cpu
), cfs_rq
)
1017 print_cfs_rq(m
, cpu
, cfs_rq
);