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 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 if ((vruntime
> min_vruntime
) ||
121 (min_vruntime
> (1ULL << 61) && vruntime
< (1ULL << 50)))
122 min_vruntime
= vruntime
;
128 set_leftmost(struct cfs_rq
*cfs_rq
, struct rb_node
*leftmost
)
130 struct sched_entity
*se
;
132 cfs_rq
->rb_leftmost
= leftmost
;
134 se
= rb_entry(leftmost
, struct sched_entity
, run_node
);
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 set_leftmost(cfs_rq
, &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 set_leftmost(cfs_rq
, 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
=
226 sysctl_sched_latency
/ sysctl_sched_min_granularity
;
228 if (unlikely(nr_running
> nr_latency
)) {
229 period
*= nr_running
;
230 do_div(period
, nr_latency
);
236 static u64
sched_slice(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
238 u64 period
= __sched_period(cfs_rq
->nr_running
);
240 period
*= se
->load
.weight
;
241 do_div(period
, cfs_rq
->load
.weight
);
246 static u64
__sched_vslice(unsigned long nr_running
)
248 u64 period
= __sched_period(nr_running
);
250 do_div(period
, nr_running
);
256 * Update the current task's runtime statistics. Skip current tasks that
257 * are not in our scheduling class.
260 __update_curr(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
,
261 unsigned long delta_exec
)
263 unsigned long delta_exec_weighted
;
264 u64 next_vruntime
, min_vruntime
;
266 schedstat_set(curr
->exec_max
, max((u64
)delta_exec
, curr
->exec_max
));
268 curr
->sum_exec_runtime
+= delta_exec
;
269 schedstat_add(cfs_rq
, exec_clock
, delta_exec
);
270 delta_exec_weighted
= delta_exec
;
271 if (unlikely(curr
->load
.weight
!= NICE_0_LOAD
)) {
272 delta_exec_weighted
= calc_delta_fair(delta_exec_weighted
,
275 curr
->vruntime
+= delta_exec_weighted
;
278 * maintain cfs_rq->min_vruntime to be a monotonic increasing
279 * value tracking the leftmost vruntime in the tree.
281 if (first_fair(cfs_rq
)) {
282 next_vruntime
= __pick_next_entity(cfs_rq
)->vruntime
;
284 /* min_vruntime() := !max_vruntime() */
285 min_vruntime
= max_vruntime(curr
->vruntime
, next_vruntime
);
286 if (min_vruntime
== next_vruntime
)
287 min_vruntime
= curr
->vruntime
;
289 min_vruntime
= next_vruntime
;
291 min_vruntime
= curr
->vruntime
;
293 cfs_rq
->min_vruntime
=
294 max_vruntime(cfs_rq
->min_vruntime
, min_vruntime
);
297 static void update_curr(struct cfs_rq
*cfs_rq
)
299 struct sched_entity
*curr
= cfs_rq
->curr
;
300 u64 now
= rq_of(cfs_rq
)->clock
;
301 unsigned long delta_exec
;
307 * Get the amount of time the current task was running
308 * since the last time we changed load (this cannot
309 * overflow on 32 bits):
311 delta_exec
= (unsigned long)(now
- curr
->exec_start
);
313 __update_curr(cfs_rq
, curr
, delta_exec
);
314 curr
->exec_start
= now
;
318 update_stats_wait_start(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
320 schedstat_set(se
->wait_start
, rq_of(cfs_rq
)->clock
);
323 static inline unsigned long
324 calc_weighted(unsigned long delta
, struct sched_entity
*se
)
326 unsigned long weight
= se
->load
.weight
;
328 if (unlikely(weight
!= NICE_0_LOAD
))
329 return (u64
)delta
* se
->load
.weight
>> NICE_0_SHIFT
;
335 * Task is being enqueued - update stats:
337 static void update_stats_enqueue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
340 * Are we enqueueing a waiting task? (for current tasks
341 * a dequeue/enqueue event is a NOP)
343 if (se
!= cfs_rq
->curr
)
344 update_stats_wait_start(cfs_rq
, se
);
348 update_stats_wait_end(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
350 schedstat_set(se
->wait_max
, max(se
->wait_max
,
351 rq_of(cfs_rq
)->clock
- se
->wait_start
));
352 schedstat_set(se
->wait_start
, 0);
356 update_stats_dequeue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
360 * Mark the end of the wait period if dequeueing a
363 if (se
!= cfs_rq
->curr
)
364 update_stats_wait_end(cfs_rq
, se
);
368 * We are picking a new current task - update its stats:
371 update_stats_curr_start(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
374 * We are starting a new run period:
376 se
->exec_start
= rq_of(cfs_rq
)->clock
;
380 * We are descheduling a task - update its stats:
383 update_stats_curr_end(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
388 /**************************************************
389 * Scheduling class queueing methods:
393 account_entity_enqueue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
395 update_load_add(&cfs_rq
->load
, se
->load
.weight
);
396 cfs_rq
->nr_running
++;
401 account_entity_dequeue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
403 update_load_sub(&cfs_rq
->load
, se
->load
.weight
);
404 cfs_rq
->nr_running
--;
408 static void enqueue_sleeper(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
410 #ifdef CONFIG_SCHEDSTATS
411 if (se
->sleep_start
) {
412 u64 delta
= rq_of(cfs_rq
)->clock
- se
->sleep_start
;
417 if (unlikely(delta
> se
->sleep_max
))
418 se
->sleep_max
= delta
;
421 se
->sum_sleep_runtime
+= delta
;
423 if (se
->block_start
) {
424 u64 delta
= rq_of(cfs_rq
)->clock
- se
->block_start
;
429 if (unlikely(delta
> se
->block_max
))
430 se
->block_max
= delta
;
433 se
->sum_sleep_runtime
+= delta
;
436 * Blocking time is in units of nanosecs, so shift by 20 to
437 * get a milliseconds-range estimation of the amount of
438 * time that the task spent sleeping:
440 if (unlikely(prof_on
== SLEEP_PROFILING
)) {
441 struct task_struct
*tsk
= task_of(se
);
443 profile_hits(SLEEP_PROFILING
, (void *)get_wchan(tsk
),
450 static void check_spread(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
452 #ifdef CONFIG_SCHED_DEBUG
453 s64 d
= se
->vruntime
- cfs_rq
->min_vruntime
;
458 if (d
> 3*sysctl_sched_latency
)
459 schedstat_inc(cfs_rq
, nr_spread_over
);
464 place_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int initial
)
468 vruntime
= cfs_rq
->min_vruntime
;
470 if (sched_feat(USE_TREE_AVG
)) {
471 struct sched_entity
*last
= __pick_last_entity(cfs_rq
);
473 vruntime
+= last
->vruntime
;
476 } else if (sched_feat(APPROX_AVG
) && cfs_rq
->nr_running
)
477 vruntime
+= __sched_vslice(cfs_rq
->nr_running
)/2;
479 if (initial
&& sched_feat(START_DEBIT
))
480 vruntime
+= __sched_vslice(cfs_rq
->nr_running
+ 1);
482 if (!initial
&& sched_feat(NEW_FAIR_SLEEPERS
)) {
483 s64 latency
= cfs_rq
->min_vruntime
- se
->last_min_vruntime
;
484 if (latency
< 0 || !cfs_rq
->nr_running
)
487 latency
= min_t(s64
, latency
, sysctl_sched_latency
);
491 se
->vruntime
= vruntime
;
496 enqueue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int wakeup
)
499 * Update the fair clock.
504 /* se->vruntime += cfs_rq->min_vruntime; */
505 place_entity(cfs_rq
, se
, 0);
506 enqueue_sleeper(cfs_rq
, se
);
509 update_stats_enqueue(cfs_rq
, se
);
510 check_spread(cfs_rq
, se
);
511 if (se
!= cfs_rq
->curr
)
512 __enqueue_entity(cfs_rq
, se
);
513 account_entity_enqueue(cfs_rq
, se
);
517 dequeue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int sleep
)
519 update_stats_dequeue(cfs_rq
, se
);
521 #ifdef CONFIG_SCHEDSTATS
522 if (entity_is_task(se
)) {
523 struct task_struct
*tsk
= task_of(se
);
525 if (tsk
->state
& TASK_INTERRUPTIBLE
)
526 se
->sleep_start
= rq_of(cfs_rq
)->clock
;
527 if (tsk
->state
& TASK_UNINTERRUPTIBLE
)
528 se
->block_start
= rq_of(cfs_rq
)->clock
;
531 /* se->vruntime = entity_key(cfs_rq, se); */
532 se
->last_min_vruntime
= cfs_rq
->min_vruntime
;
535 if (se
!= cfs_rq
->curr
)
536 __dequeue_entity(cfs_rq
, se
);
537 account_entity_dequeue(cfs_rq
, se
);
541 * Preempt the current task with a newly woken task if needed:
544 check_preempt_tick(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
)
546 unsigned long ideal_runtime
, delta_exec
;
548 ideal_runtime
= sched_slice(cfs_rq
, curr
);
549 delta_exec
= curr
->sum_exec_runtime
- curr
->prev_sum_exec_runtime
;
550 if (delta_exec
> ideal_runtime
)
551 resched_task(rq_of(cfs_rq
)->curr
);
555 set_next_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
557 /* 'current' is not kept within the tree. */
560 * Any task has to be enqueued before it get to execute on
561 * a CPU. So account for the time it spent waiting on the
564 update_stats_wait_end(cfs_rq
, se
);
565 __dequeue_entity(cfs_rq
, se
);
568 update_stats_curr_start(cfs_rq
, se
);
570 #ifdef CONFIG_SCHEDSTATS
572 * Track our maximum slice length, if the CPU's load is at
573 * least twice that of our own weight (i.e. dont track it
574 * when there are only lesser-weight tasks around):
576 if (rq_of(cfs_rq
)->load
.weight
>= 2*se
->load
.weight
) {
577 se
->slice_max
= max(se
->slice_max
,
578 se
->sum_exec_runtime
- se
->prev_sum_exec_runtime
);
581 se
->prev_sum_exec_runtime
= se
->sum_exec_runtime
;
584 static struct sched_entity
*pick_next_entity(struct cfs_rq
*cfs_rq
)
586 struct sched_entity
*se
= __pick_next_entity(cfs_rq
);
588 set_next_entity(cfs_rq
, se
);
593 static void put_prev_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*prev
)
596 * If still on the runqueue then deactivate_task()
597 * was not called and update_curr() has to be done:
602 update_stats_curr_end(cfs_rq
, prev
);
604 check_spread(cfs_rq
, prev
);
606 update_stats_wait_start(cfs_rq
, prev
);
607 /* Put 'current' back into the tree. */
608 __enqueue_entity(cfs_rq
, prev
);
613 static void entity_tick(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
)
616 * Update run-time statistics of the 'current'.
620 if (cfs_rq
->nr_running
> 1)
621 check_preempt_tick(cfs_rq
, curr
);
624 /**************************************************
625 * CFS operations on tasks:
628 #ifdef CONFIG_FAIR_GROUP_SCHED
630 /* Walk up scheduling entities hierarchy */
631 #define for_each_sched_entity(se) \
632 for (; se; se = se->parent)
634 static inline struct cfs_rq
*task_cfs_rq(struct task_struct
*p
)
639 /* runqueue on which this entity is (to be) queued */
640 static inline struct cfs_rq
*cfs_rq_of(struct sched_entity
*se
)
645 /* runqueue "owned" by this group */
646 static inline struct cfs_rq
*group_cfs_rq(struct sched_entity
*grp
)
651 /* Given a group's cfs_rq on one cpu, return its corresponding cfs_rq on
652 * another cpu ('this_cpu')
654 static inline struct cfs_rq
*cpu_cfs_rq(struct cfs_rq
*cfs_rq
, int this_cpu
)
656 return cfs_rq
->tg
->cfs_rq
[this_cpu
];
659 /* Iterate thr' all leaf cfs_rq's on a runqueue */
660 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
661 list_for_each_entry(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
663 /* Do the two (enqueued) tasks belong to the same group ? */
664 static inline int is_same_group(struct task_struct
*curr
, struct task_struct
*p
)
666 if (curr
->se
.cfs_rq
== p
->se
.cfs_rq
)
672 #else /* CONFIG_FAIR_GROUP_SCHED */
674 #define for_each_sched_entity(se) \
675 for (; se; se = NULL)
677 static inline struct cfs_rq
*task_cfs_rq(struct task_struct
*p
)
679 return &task_rq(p
)->cfs
;
682 static inline struct cfs_rq
*cfs_rq_of(struct sched_entity
*se
)
684 struct task_struct
*p
= task_of(se
);
685 struct rq
*rq
= task_rq(p
);
690 /* runqueue "owned" by this group */
691 static inline struct cfs_rq
*group_cfs_rq(struct sched_entity
*grp
)
696 static inline struct cfs_rq
*cpu_cfs_rq(struct cfs_rq
*cfs_rq
, int this_cpu
)
698 return &cpu_rq(this_cpu
)->cfs
;
701 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
702 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
704 static inline int is_same_group(struct task_struct
*curr
, struct task_struct
*p
)
709 #endif /* CONFIG_FAIR_GROUP_SCHED */
712 * The enqueue_task method is called before nr_running is
713 * increased. Here we update the fair scheduling stats and
714 * then put the task into the rbtree:
716 static void enqueue_task_fair(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
718 struct cfs_rq
*cfs_rq
;
719 struct sched_entity
*se
= &p
->se
;
721 for_each_sched_entity(se
) {
724 cfs_rq
= cfs_rq_of(se
);
725 enqueue_entity(cfs_rq
, se
, wakeup
);
730 * The dequeue_task method is called before nr_running is
731 * decreased. We remove the task from the rbtree and
732 * update the fair scheduling stats:
734 static void dequeue_task_fair(struct rq
*rq
, struct task_struct
*p
, int sleep
)
736 struct cfs_rq
*cfs_rq
;
737 struct sched_entity
*se
= &p
->se
;
739 for_each_sched_entity(se
) {
740 cfs_rq
= cfs_rq_of(se
);
741 dequeue_entity(cfs_rq
, se
, sleep
);
742 /* Don't dequeue parent if it has other entities besides us */
743 if (cfs_rq
->load
.weight
)
749 * sched_yield() support is very simple - we dequeue and enqueue.
751 * If compat_yield is turned on then we requeue to the end of the tree.
753 static void yield_task_fair(struct rq
*rq
)
755 struct cfs_rq
*cfs_rq
= task_cfs_rq(rq
->curr
);
756 struct rb_node
**link
= &cfs_rq
->tasks_timeline
.rb_node
;
757 struct sched_entity
*rightmost
, *se
= &rq
->curr
->se
;
758 struct rb_node
*parent
;
761 * Are we the only task in the tree?
763 if (unlikely(cfs_rq
->nr_running
== 1))
766 if (likely(!sysctl_sched_compat_yield
)) {
767 __update_rq_clock(rq
);
769 * Dequeue and enqueue the task to update its
770 * position within the tree:
772 dequeue_entity(cfs_rq
, se
, 0);
773 enqueue_entity(cfs_rq
, se
, 0);
778 * Find the rightmost entry in the rbtree:
782 link
= &parent
->rb_right
;
785 rightmost
= rb_entry(parent
, struct sched_entity
, run_node
);
787 * Already in the rightmost position?
789 if (unlikely(rightmost
== se
))
793 * Minimally necessary key value to be last in the tree:
795 se
->vruntime
= rightmost
->vruntime
+ 1;
797 if (cfs_rq
->rb_leftmost
== &se
->run_node
)
798 cfs_rq
->rb_leftmost
= rb_next(&se
->run_node
);
800 * Relink the task to the rightmost position:
802 rb_erase(&se
->run_node
, &cfs_rq
->tasks_timeline
);
803 rb_link_node(&se
->run_node
, parent
, link
);
804 rb_insert_color(&se
->run_node
, &cfs_rq
->tasks_timeline
);
808 * Preempt the current task with a newly woken task if needed:
810 static void check_preempt_wakeup(struct rq
*rq
, struct task_struct
*p
)
812 struct task_struct
*curr
= rq
->curr
;
813 struct cfs_rq
*cfs_rq
= task_cfs_rq(curr
);
815 if (unlikely(rt_prio(p
->prio
))) {
821 if (is_same_group(curr
, p
)) {
822 s64 delta
= curr
->se
.vruntime
- p
->se
.vruntime
;
824 if (delta
> (s64
)sysctl_sched_wakeup_granularity
)
829 static struct task_struct
*pick_next_task_fair(struct rq
*rq
)
831 struct cfs_rq
*cfs_rq
= &rq
->cfs
;
832 struct sched_entity
*se
;
834 if (unlikely(!cfs_rq
->nr_running
))
838 se
= pick_next_entity(cfs_rq
);
839 cfs_rq
= group_cfs_rq(se
);
846 * Account for a descheduled task:
848 static void put_prev_task_fair(struct rq
*rq
, struct task_struct
*prev
)
850 struct sched_entity
*se
= &prev
->se
;
851 struct cfs_rq
*cfs_rq
;
853 for_each_sched_entity(se
) {
854 cfs_rq
= cfs_rq_of(se
);
855 put_prev_entity(cfs_rq
, se
);
859 /**************************************************
860 * Fair scheduling class load-balancing methods:
864 * Load-balancing iterator. Note: while the runqueue stays locked
865 * during the whole iteration, the current task might be
866 * dequeued so the iterator has to be dequeue-safe. Here we
867 * achieve that by always pre-iterating before returning
870 static inline struct task_struct
*
871 __load_balance_iterator(struct cfs_rq
*cfs_rq
, struct rb_node
*curr
)
873 struct task_struct
*p
;
878 p
= rb_entry(curr
, struct task_struct
, se
.run_node
);
879 cfs_rq
->rb_load_balance_curr
= rb_next(curr
);
884 static struct task_struct
*load_balance_start_fair(void *arg
)
886 struct cfs_rq
*cfs_rq
= arg
;
888 return __load_balance_iterator(cfs_rq
, first_fair(cfs_rq
));
891 static struct task_struct
*load_balance_next_fair(void *arg
)
893 struct cfs_rq
*cfs_rq
= arg
;
895 return __load_balance_iterator(cfs_rq
, cfs_rq
->rb_load_balance_curr
);
898 #ifdef CONFIG_FAIR_GROUP_SCHED
899 static int cfs_rq_best_prio(struct cfs_rq
*cfs_rq
)
901 struct sched_entity
*curr
;
902 struct task_struct
*p
;
904 if (!cfs_rq
->nr_running
)
909 curr
= __pick_next_entity(cfs_rq
);
918 load_balance_fair(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
919 unsigned long max_nr_move
, unsigned long max_load_move
,
920 struct sched_domain
*sd
, enum cpu_idle_type idle
,
921 int *all_pinned
, int *this_best_prio
)
923 struct cfs_rq
*busy_cfs_rq
;
924 unsigned long load_moved
, total_nr_moved
= 0, nr_moved
;
925 long rem_load_move
= max_load_move
;
926 struct rq_iterator cfs_rq_iterator
;
928 cfs_rq_iterator
.start
= load_balance_start_fair
;
929 cfs_rq_iterator
.next
= load_balance_next_fair
;
931 for_each_leaf_cfs_rq(busiest
, busy_cfs_rq
) {
932 #ifdef CONFIG_FAIR_GROUP_SCHED
933 struct cfs_rq
*this_cfs_rq
;
935 unsigned long maxload
;
937 this_cfs_rq
= cpu_cfs_rq(busy_cfs_rq
, this_cpu
);
939 imbalance
= busy_cfs_rq
->load
.weight
- this_cfs_rq
->load
.weight
;
940 /* Don't pull if this_cfs_rq has more load than busy_cfs_rq */
944 /* Don't pull more than imbalance/2 */
946 maxload
= min(rem_load_move
, imbalance
);
948 *this_best_prio
= cfs_rq_best_prio(this_cfs_rq
);
950 # define maxload rem_load_move
952 /* pass busy_cfs_rq argument into
953 * load_balance_[start|next]_fair iterators
955 cfs_rq_iterator
.arg
= busy_cfs_rq
;
956 nr_moved
= balance_tasks(this_rq
, this_cpu
, busiest
,
957 max_nr_move
, maxload
, sd
, idle
, all_pinned
,
958 &load_moved
, this_best_prio
, &cfs_rq_iterator
);
960 total_nr_moved
+= nr_moved
;
961 max_nr_move
-= nr_moved
;
962 rem_load_move
-= load_moved
;
964 if (max_nr_move
<= 0 || rem_load_move
<= 0)
968 return max_load_move
- rem_load_move
;
972 * scheduler tick hitting a task of our scheduling class:
974 static void task_tick_fair(struct rq
*rq
, struct task_struct
*curr
)
976 struct cfs_rq
*cfs_rq
;
977 struct sched_entity
*se
= &curr
->se
;
979 for_each_sched_entity(se
) {
980 cfs_rq
= cfs_rq_of(se
);
981 entity_tick(cfs_rq
, se
);
985 #define swap(a,b) do { typeof(a) tmp = (a); (a) = (b); (b) = tmp; } while (0)
988 * Share the fairness runtime between parent and child, thus the
989 * total amount of pressure for CPU stays equal - new tasks
990 * get a chance to run but frequent forkers are not allowed to
991 * monopolize the CPU. Note: the parent runqueue is locked,
992 * the child is not running yet.
994 static void task_new_fair(struct rq
*rq
, struct task_struct
*p
)
996 struct cfs_rq
*cfs_rq
= task_cfs_rq(p
);
997 struct sched_entity
*se
= &p
->se
, *curr
= cfs_rq
->curr
;
999 sched_info_queued(p
);
1001 update_curr(cfs_rq
);
1002 place_entity(cfs_rq
, se
, 1);
1004 if (sysctl_sched_child_runs_first
&&
1005 curr
->vruntime
< se
->vruntime
) {
1007 * Upon rescheduling, sched_class::put_prev_task() will place
1008 * 'current' within the tree based on its new key value.
1010 swap(curr
->vruntime
, se
->vruntime
);
1013 update_stats_enqueue(cfs_rq
, se
);
1014 check_spread(cfs_rq
, se
);
1015 check_spread(cfs_rq
, curr
);
1016 __enqueue_entity(cfs_rq
, se
);
1017 account_entity_enqueue(cfs_rq
, se
);
1018 resched_task(rq
->curr
);
1021 /* Account for a task changing its policy or group.
1023 * This routine is mostly called to set cfs_rq->curr field when a task
1024 * migrates between groups/classes.
1026 static void set_curr_task_fair(struct rq
*rq
)
1028 struct sched_entity
*se
= &rq
->curr
->se
;
1030 for_each_sched_entity(se
)
1031 set_next_entity(cfs_rq_of(se
), se
);
1035 * All the scheduling class methods:
1037 struct sched_class fair_sched_class __read_mostly
= {
1038 .enqueue_task
= enqueue_task_fair
,
1039 .dequeue_task
= dequeue_task_fair
,
1040 .yield_task
= yield_task_fair
,
1042 .check_preempt_curr
= check_preempt_wakeup
,
1044 .pick_next_task
= pick_next_task_fair
,
1045 .put_prev_task
= put_prev_task_fair
,
1047 .load_balance
= load_balance_fair
,
1049 .set_curr_task
= set_curr_task_fair
,
1050 .task_tick
= task_tick_fair
,
1051 .task_new
= task_new_fair
,
1054 #ifdef CONFIG_SCHED_DEBUG
1055 static void print_cfs_stats(struct seq_file
*m
, int cpu
)
1057 struct cfs_rq
*cfs_rq
;
1059 #ifdef CONFIG_FAIR_GROUP_SCHED
1060 print_cfs_rq(m
, cpu
, &cpu_rq(cpu
)->cfs
);
1062 for_each_leaf_cfs_rq(cpu_rq(cpu
), cfs_rq
)
1063 print_cfs_rq(m
, cpu
, cfs_rq
);