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
->vruntime
- 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
);
187 __dequeue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
189 if (cfs_rq
->rb_leftmost
== &se
->run_node
)
190 set_leftmost(cfs_rq
, rb_next(&se
->run_node
));
192 rb_erase(&se
->run_node
, &cfs_rq
->tasks_timeline
);
195 static inline struct rb_node
*first_fair(struct cfs_rq
*cfs_rq
)
197 return cfs_rq
->rb_leftmost
;
200 static struct sched_entity
*__pick_next_entity(struct cfs_rq
*cfs_rq
)
202 return rb_entry(first_fair(cfs_rq
), struct sched_entity
, run_node
);
205 static inline struct sched_entity
*__pick_last_entity(struct cfs_rq
*cfs_rq
)
207 struct rb_node
**link
= &cfs_rq
->tasks_timeline
.rb_node
;
208 struct sched_entity
*se
= NULL
;
209 struct rb_node
*parent
;
213 se
= rb_entry(parent
, struct sched_entity
, run_node
);
214 link
= &parent
->rb_right
;
220 /**************************************************************
221 * Scheduling class statistics methods:
224 static u64
__sched_period(unsigned long nr_running
)
226 u64 period
= sysctl_sched_latency
;
227 unsigned long nr_latency
=
228 sysctl_sched_latency
/ sysctl_sched_min_granularity
;
230 if (unlikely(nr_running
> nr_latency
)) {
231 period
*= nr_running
;
232 do_div(period
, nr_latency
);
238 static u64
sched_slice(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
240 u64 period
= __sched_period(cfs_rq
->nr_running
);
242 period
*= se
->load
.weight
;
243 do_div(period
, cfs_rq
->load
.weight
);
249 * Update the current task's runtime statistics. Skip current tasks that
250 * are not in our scheduling class.
253 __update_curr(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
,
254 unsigned long delta_exec
)
256 unsigned long delta_exec_weighted
;
257 u64 next_vruntime
, min_vruntime
;
259 schedstat_set(curr
->exec_max
, max((u64
)delta_exec
, curr
->exec_max
));
261 curr
->sum_exec_runtime
+= delta_exec
;
262 schedstat_add(cfs_rq
, exec_clock
, delta_exec
);
263 delta_exec_weighted
= delta_exec
;
264 if (unlikely(curr
->load
.weight
!= NICE_0_LOAD
)) {
265 delta_exec_weighted
= calc_delta_fair(delta_exec_weighted
,
268 curr
->vruntime
+= delta_exec_weighted
;
271 * maintain cfs_rq->min_vruntime to be a monotonic increasing
272 * value tracking the leftmost vruntime in the tree.
274 if (first_fair(cfs_rq
)) {
275 next_vruntime
= __pick_next_entity(cfs_rq
)->vruntime
;
277 /* min_vruntime() := !max_vruntime() */
278 min_vruntime
= max_vruntime(curr
->vruntime
, next_vruntime
);
279 if (min_vruntime
== next_vruntime
)
280 min_vruntime
= curr
->vruntime
;
282 min_vruntime
= next_vruntime
;
284 min_vruntime
= curr
->vruntime
;
286 cfs_rq
->min_vruntime
=
287 max_vruntime(cfs_rq
->min_vruntime
, min_vruntime
);
290 static void update_curr(struct cfs_rq
*cfs_rq
)
292 struct sched_entity
*curr
= cfs_rq
->curr
;
293 u64 now
= rq_of(cfs_rq
)->clock
;
294 unsigned long delta_exec
;
300 * Get the amount of time the current task was running
301 * since the last time we changed load (this cannot
302 * overflow on 32 bits):
304 delta_exec
= (unsigned long)(now
- curr
->exec_start
);
306 __update_curr(cfs_rq
, curr
, delta_exec
);
307 curr
->exec_start
= now
;
311 update_stats_wait_start(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
313 schedstat_set(se
->wait_start
, rq_of(cfs_rq
)->clock
);
316 static inline unsigned long
317 calc_weighted(unsigned long delta
, struct sched_entity
*se
)
319 unsigned long weight
= se
->load
.weight
;
321 if (unlikely(weight
!= NICE_0_LOAD
))
322 return (u64
)delta
* se
->load
.weight
>> NICE_0_SHIFT
;
328 * Task is being enqueued - update stats:
330 static void update_stats_enqueue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
333 * Are we enqueueing a waiting task? (for current tasks
334 * a dequeue/enqueue event is a NOP)
336 if (se
!= cfs_rq
->curr
)
337 update_stats_wait_start(cfs_rq
, se
);
341 update_stats_wait_end(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
343 schedstat_set(se
->wait_max
, max(se
->wait_max
,
344 rq_of(cfs_rq
)->clock
- se
->wait_start
));
345 schedstat_set(se
->wait_start
, 0);
349 update_stats_dequeue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
353 * Mark the end of the wait period if dequeueing a
356 if (se
!= cfs_rq
->curr
)
357 update_stats_wait_end(cfs_rq
, se
);
361 * We are picking a new current task - update its stats:
364 update_stats_curr_start(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
367 * We are starting a new run period:
369 se
->exec_start
= rq_of(cfs_rq
)->clock
;
373 * We are descheduling a task - update its stats:
376 update_stats_curr_end(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
381 /**************************************************
382 * Scheduling class queueing methods:
386 account_entity_enqueue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
388 update_load_add(&cfs_rq
->load
, se
->load
.weight
);
389 cfs_rq
->nr_running
++;
394 account_entity_dequeue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
396 update_load_sub(&cfs_rq
->load
, se
->load
.weight
);
397 cfs_rq
->nr_running
--;
401 static void enqueue_sleeper(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
403 #ifdef CONFIG_SCHEDSTATS
404 if (se
->sleep_start
) {
405 u64 delta
= rq_of(cfs_rq
)->clock
- se
->sleep_start
;
410 if (unlikely(delta
> se
->sleep_max
))
411 se
->sleep_max
= delta
;
414 se
->sum_sleep_runtime
+= delta
;
416 if (se
->block_start
) {
417 u64 delta
= rq_of(cfs_rq
)->clock
- se
->block_start
;
422 if (unlikely(delta
> se
->block_max
))
423 se
->block_max
= delta
;
426 se
->sum_sleep_runtime
+= delta
;
429 * Blocking time is in units of nanosecs, so shift by 20 to
430 * get a milliseconds-range estimation of the amount of
431 * time that the task spent sleeping:
433 if (unlikely(prof_on
== SLEEP_PROFILING
)) {
434 struct task_struct
*tsk
= task_of(se
);
436 profile_hits(SLEEP_PROFILING
, (void *)get_wchan(tsk
),
444 place_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int initial
)
446 u64 min_runtime
, latency
;
448 min_runtime
= cfs_rq
->min_vruntime
;
450 if (sched_feat(USE_TREE_AVG
)) {
451 struct sched_entity
*last
= __pick_last_entity(cfs_rq
);
453 min_runtime
= __pick_next_entity(cfs_rq
)->vruntime
;
454 min_runtime
+= last
->vruntime
;
457 } else if (sched_feat(APPROX_AVG
))
458 min_runtime
+= sysctl_sched_latency
/2;
460 if (initial
&& sched_feat(START_DEBIT
))
461 min_runtime
+= sched_slice(cfs_rq
, se
);
463 if (!initial
&& sched_feat(NEW_FAIR_SLEEPERS
)) {
464 latency
= sysctl_sched_latency
;
465 if (min_runtime
> latency
)
466 min_runtime
-= latency
;
471 se
->vruntime
= max(se
->vruntime
, min_runtime
);
475 enqueue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
,
476 int wakeup
, int set_curr
)
479 * In case of the 'current'.
481 if (unlikely(set_curr
)) {
482 update_stats_curr_start(cfs_rq
, se
);
484 account_entity_enqueue(cfs_rq
, se
);
489 * Update the fair clock.
494 place_entity(cfs_rq
, se
, 0);
495 enqueue_sleeper(cfs_rq
, se
);
498 update_stats_enqueue(cfs_rq
, se
);
499 __enqueue_entity(cfs_rq
, se
);
500 account_entity_enqueue(cfs_rq
, se
);
504 dequeue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int sleep
)
506 update_stats_dequeue(cfs_rq
, se
);
507 #ifdef CONFIG_SCHEDSTATS
509 if (entity_is_task(se
)) {
510 struct task_struct
*tsk
= task_of(se
);
512 if (tsk
->state
& TASK_INTERRUPTIBLE
)
513 se
->sleep_start
= rq_of(cfs_rq
)->clock
;
514 if (tsk
->state
& TASK_UNINTERRUPTIBLE
)
515 se
->block_start
= rq_of(cfs_rq
)->clock
;
519 if (likely(se
!= cfs_rq
->curr
))
520 __dequeue_entity(cfs_rq
, se
);
522 update_stats_curr_end(cfs_rq
, se
);
525 account_entity_dequeue(cfs_rq
, se
);
529 * Preempt the current task with a newly woken task if needed:
532 check_preempt_tick(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
)
534 unsigned long ideal_runtime
, delta_exec
;
536 ideal_runtime
= sched_slice(cfs_rq
, curr
);
537 delta_exec
= curr
->sum_exec_runtime
- curr
->prev_sum_exec_runtime
;
538 if (delta_exec
> ideal_runtime
)
539 resched_task(rq_of(cfs_rq
)->curr
);
543 set_next_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
546 * Any task has to be enqueued before it get to execute on
547 * a CPU. So account for the time it spent waiting on the
550 update_stats_wait_end(cfs_rq
, se
);
551 update_stats_curr_start(cfs_rq
, se
);
553 #ifdef CONFIG_SCHEDSTATS
555 * Track our maximum slice length, if the CPU's load is at
556 * least twice that of our own weight (i.e. dont track it
557 * when there are only lesser-weight tasks around):
559 if (rq_of(cfs_rq
)->load
.weight
>= 2*se
->load
.weight
) {
560 se
->slice_max
= max(se
->slice_max
,
561 se
->sum_exec_runtime
- se
->prev_sum_exec_runtime
);
564 se
->prev_sum_exec_runtime
= se
->sum_exec_runtime
;
567 static struct sched_entity
*pick_next_entity(struct cfs_rq
*cfs_rq
)
569 struct sched_entity
*se
= __pick_next_entity(cfs_rq
);
571 /* 'current' is not kept within the tree. */
573 __dequeue_entity(cfs_rq
, se
);
575 set_next_entity(cfs_rq
, se
);
580 static void put_prev_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*prev
)
583 * If still on the runqueue then deactivate_task()
584 * was not called and update_curr() has to be done:
589 update_stats_curr_end(cfs_rq
, prev
);
592 update_stats_wait_start(cfs_rq
, prev
);
593 /* Put 'current' back into the tree. */
594 __enqueue_entity(cfs_rq
, prev
);
599 static void entity_tick(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
)
602 * Update run-time statistics of the 'current'.
606 if (cfs_rq
->nr_running
> 1)
607 check_preempt_tick(cfs_rq
, curr
);
610 /**************************************************
611 * CFS operations on tasks:
614 #ifdef CONFIG_FAIR_GROUP_SCHED
616 /* Walk up scheduling entities hierarchy */
617 #define for_each_sched_entity(se) \
618 for (; se; se = se->parent)
620 static inline struct cfs_rq
*task_cfs_rq(struct task_struct
*p
)
625 /* runqueue on which this entity is (to be) queued */
626 static inline struct cfs_rq
*cfs_rq_of(struct sched_entity
*se
)
631 /* runqueue "owned" by this group */
632 static inline struct cfs_rq
*group_cfs_rq(struct sched_entity
*grp
)
637 /* Given a group's cfs_rq on one cpu, return its corresponding cfs_rq on
638 * another cpu ('this_cpu')
640 static inline struct cfs_rq
*cpu_cfs_rq(struct cfs_rq
*cfs_rq
, int this_cpu
)
642 return cfs_rq
->tg
->cfs_rq
[this_cpu
];
645 /* Iterate thr' all leaf cfs_rq's on a runqueue */
646 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
647 list_for_each_entry(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
649 /* Do the two (enqueued) tasks belong to the same group ? */
650 static inline int is_same_group(struct task_struct
*curr
, struct task_struct
*p
)
652 if (curr
->se
.cfs_rq
== p
->se
.cfs_rq
)
658 #else /* CONFIG_FAIR_GROUP_SCHED */
660 #define for_each_sched_entity(se) \
661 for (; se; se = NULL)
663 static inline struct cfs_rq
*task_cfs_rq(struct task_struct
*p
)
665 return &task_rq(p
)->cfs
;
668 static inline struct cfs_rq
*cfs_rq_of(struct sched_entity
*se
)
670 struct task_struct
*p
= task_of(se
);
671 struct rq
*rq
= task_rq(p
);
676 /* runqueue "owned" by this group */
677 static inline struct cfs_rq
*group_cfs_rq(struct sched_entity
*grp
)
682 static inline struct cfs_rq
*cpu_cfs_rq(struct cfs_rq
*cfs_rq
, int this_cpu
)
684 return &cpu_rq(this_cpu
)->cfs
;
687 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
688 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
690 static inline int is_same_group(struct task_struct
*curr
, struct task_struct
*p
)
695 #endif /* CONFIG_FAIR_GROUP_SCHED */
698 * The enqueue_task method is called before nr_running is
699 * increased. Here we update the fair scheduling stats and
700 * then put the task into the rbtree:
702 static void enqueue_task_fair(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
704 struct cfs_rq
*cfs_rq
;
705 struct sched_entity
*se
= &p
->se
;
708 /* Are we enqueuing the current task? */
709 if (unlikely(task_running(rq
, p
)))
712 for_each_sched_entity(se
) {
715 cfs_rq
= cfs_rq_of(se
);
716 enqueue_entity(cfs_rq
, se
, wakeup
, set_curr
);
721 * The dequeue_task method is called before nr_running is
722 * decreased. We remove the task from the rbtree and
723 * update the fair scheduling stats:
725 static void dequeue_task_fair(struct rq
*rq
, struct task_struct
*p
, int sleep
)
727 struct cfs_rq
*cfs_rq
;
728 struct sched_entity
*se
= &p
->se
;
730 for_each_sched_entity(se
) {
731 cfs_rq
= cfs_rq_of(se
);
732 dequeue_entity(cfs_rq
, se
, sleep
);
733 /* Don't dequeue parent if it has other entities besides us */
734 if (cfs_rq
->load
.weight
)
740 * sched_yield() support is very simple - we dequeue and enqueue.
742 * If compat_yield is turned on then we requeue to the end of the tree.
744 static void yield_task_fair(struct rq
*rq
)
746 struct cfs_rq
*cfs_rq
= &rq
->cfs
;
747 struct rb_node
**link
= &cfs_rq
->tasks_timeline
.rb_node
;
748 struct sched_entity
*rightmost
, *se
= &rq
->curr
->se
;
749 struct rb_node
*parent
;
752 * Are we the only task in the tree?
754 if (unlikely(cfs_rq
->nr_running
== 1))
757 if (likely(!sysctl_sched_compat_yield
)) {
758 __update_rq_clock(rq
);
760 * Dequeue and enqueue the task to update its
761 * position within the tree:
763 dequeue_entity(cfs_rq
, se
, 0);
764 enqueue_entity(cfs_rq
, se
, 0, 1);
769 * Find the rightmost entry in the rbtree:
773 link
= &parent
->rb_right
;
776 rightmost
= rb_entry(parent
, struct sched_entity
, run_node
);
778 * Already in the rightmost position?
780 if (unlikely(rightmost
== se
))
784 * Minimally necessary key value to be last in the tree:
786 se
->vruntime
= rightmost
->vruntime
+ 1;
788 if (cfs_rq
->rb_leftmost
== &se
->run_node
)
789 cfs_rq
->rb_leftmost
= rb_next(&se
->run_node
);
791 * Relink the task to the rightmost position:
793 rb_erase(&se
->run_node
, &cfs_rq
->tasks_timeline
);
794 rb_link_node(&se
->run_node
, parent
, link
);
795 rb_insert_color(&se
->run_node
, &cfs_rq
->tasks_timeline
);
799 * Preempt the current task with a newly woken task if needed:
801 static void check_preempt_wakeup(struct rq
*rq
, struct task_struct
*p
)
803 struct task_struct
*curr
= rq
->curr
;
804 struct cfs_rq
*cfs_rq
= task_cfs_rq(curr
);
806 if (unlikely(rt_prio(p
->prio
))) {
812 if (is_same_group(curr
, p
)) {
813 s64 delta
= curr
->se
.vruntime
- p
->se
.vruntime
;
815 if (delta
> (s64
)sysctl_sched_wakeup_granularity
)
820 static struct task_struct
*pick_next_task_fair(struct rq
*rq
)
822 struct cfs_rq
*cfs_rq
= &rq
->cfs
;
823 struct sched_entity
*se
;
825 if (unlikely(!cfs_rq
->nr_running
))
829 se
= pick_next_entity(cfs_rq
);
830 cfs_rq
= group_cfs_rq(se
);
837 * Account for a descheduled task:
839 static void put_prev_task_fair(struct rq
*rq
, struct task_struct
*prev
)
841 struct sched_entity
*se
= &prev
->se
;
842 struct cfs_rq
*cfs_rq
;
844 for_each_sched_entity(se
) {
845 cfs_rq
= cfs_rq_of(se
);
846 put_prev_entity(cfs_rq
, se
);
850 /**************************************************
851 * Fair scheduling class load-balancing methods:
855 * Load-balancing iterator. Note: while the runqueue stays locked
856 * during the whole iteration, the current task might be
857 * dequeued so the iterator has to be dequeue-safe. Here we
858 * achieve that by always pre-iterating before returning
861 static inline struct task_struct
*
862 __load_balance_iterator(struct cfs_rq
*cfs_rq
, struct rb_node
*curr
)
864 struct task_struct
*p
;
869 p
= rb_entry(curr
, struct task_struct
, se
.run_node
);
870 cfs_rq
->rb_load_balance_curr
= rb_next(curr
);
875 static struct task_struct
*load_balance_start_fair(void *arg
)
877 struct cfs_rq
*cfs_rq
= arg
;
879 return __load_balance_iterator(cfs_rq
, first_fair(cfs_rq
));
882 static struct task_struct
*load_balance_next_fair(void *arg
)
884 struct cfs_rq
*cfs_rq
= arg
;
886 return __load_balance_iterator(cfs_rq
, cfs_rq
->rb_load_balance_curr
);
889 #ifdef CONFIG_FAIR_GROUP_SCHED
890 static int cfs_rq_best_prio(struct cfs_rq
*cfs_rq
)
892 struct sched_entity
*curr
;
893 struct task_struct
*p
;
895 if (!cfs_rq
->nr_running
)
898 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 __enqueue_entity(cfs_rq
, se
);
1003 account_entity_enqueue(cfs_rq
, se
);
1004 resched_task(rq
->curr
);
1008 * All the scheduling class methods:
1010 struct sched_class fair_sched_class __read_mostly
= {
1011 .enqueue_task
= enqueue_task_fair
,
1012 .dequeue_task
= dequeue_task_fair
,
1013 .yield_task
= yield_task_fair
,
1015 .check_preempt_curr
= check_preempt_wakeup
,
1017 .pick_next_task
= pick_next_task_fair
,
1018 .put_prev_task
= put_prev_task_fair
,
1020 .load_balance
= load_balance_fair
,
1022 .task_tick
= task_tick_fair
,
1023 .task_new
= task_new_fair
,
1026 #ifdef CONFIG_SCHED_DEBUG
1027 static void print_cfs_stats(struct seq_file
*m
, int cpu
)
1029 struct cfs_rq
*cfs_rq
;
1031 for_each_leaf_cfs_rq(cpu_rq(cpu
), cfs_rq
)
1032 print_cfs_rq(m
, cpu
, cfs_rq
);