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 set_leftmost(struct cfs_rq
*cfs_rq
, struct rb_node
*leftmost
)
122 struct sched_entity
*se
;
124 cfs_rq
->rb_leftmost
= leftmost
;
126 se
= rb_entry(leftmost
, struct sched_entity
, run_node
);
127 if ((se
->vruntime
> cfs_rq
->min_vruntime
) ||
128 (cfs_rq
->min_vruntime
> (1ULL << 61) &&
129 se
->vruntime
< (1ULL << 50)))
130 cfs_rq
->min_vruntime
= se
->vruntime
;
134 s64
entity_key(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
136 return se
->fair_key
- cfs_rq
->min_vruntime
;
140 * Enqueue an entity into the rb-tree:
143 __enqueue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
145 struct rb_node
**link
= &cfs_rq
->tasks_timeline
.rb_node
;
146 struct rb_node
*parent
= NULL
;
147 struct sched_entity
*entry
;
148 s64 key
= entity_key(cfs_rq
, se
);
152 * Find the right place in the rbtree:
156 entry
= rb_entry(parent
, struct sched_entity
, run_node
);
158 * We dont care about collisions. Nodes with
159 * the same key stay together.
161 if (key
< entity_key(cfs_rq
, entry
)) {
162 link
= &parent
->rb_left
;
164 link
= &parent
->rb_right
;
170 * Maintain a cache of leftmost tree entries (it is frequently
174 set_leftmost(cfs_rq
, &se
->run_node
);
176 rb_link_node(&se
->run_node
, parent
, link
);
177 rb_insert_color(&se
->run_node
, &cfs_rq
->tasks_timeline
);
178 update_load_add(&cfs_rq
->load
, se
->load
.weight
);
179 cfs_rq
->nr_running
++;
184 __dequeue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
186 if (cfs_rq
->rb_leftmost
== &se
->run_node
)
187 set_leftmost(cfs_rq
, rb_next(&se
->run_node
));
189 rb_erase(&se
->run_node
, &cfs_rq
->tasks_timeline
);
190 update_load_sub(&cfs_rq
->load
, se
->load
.weight
);
191 cfs_rq
->nr_running
--;
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
;
258 schedstat_set(curr
->exec_max
, max((u64
)delta_exec
, curr
->exec_max
));
260 curr
->sum_exec_runtime
+= delta_exec
;
261 schedstat_add(cfs_rq
, exec_clock
, delta_exec
);
262 delta_exec_weighted
= delta_exec
;
263 if (unlikely(curr
->load
.weight
!= NICE_0_LOAD
)) {
264 delta_exec_weighted
= calc_delta_fair(delta_exec_weighted
,
267 curr
->vruntime
+= delta_exec_weighted
;
270 static void update_curr(struct cfs_rq
*cfs_rq
)
272 struct sched_entity
*curr
= cfs_rq
->curr
;
273 u64 now
= rq_of(cfs_rq
)->clock
;
274 unsigned long delta_exec
;
280 * Get the amount of time the current task was running
281 * since the last time we changed load (this cannot
282 * overflow on 32 bits):
284 delta_exec
= (unsigned long)(now
- curr
->exec_start
);
286 __update_curr(cfs_rq
, curr
, delta_exec
);
287 curr
->exec_start
= now
;
291 update_stats_wait_start(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
293 schedstat_set(se
->wait_start
, rq_of(cfs_rq
)->clock
);
296 static inline unsigned long
297 calc_weighted(unsigned long delta
, struct sched_entity
*se
)
299 unsigned long weight
= se
->load
.weight
;
301 if (unlikely(weight
!= NICE_0_LOAD
))
302 return (u64
)delta
* se
->load
.weight
>> NICE_0_SHIFT
;
308 * Task is being enqueued - update stats:
310 static void update_stats_enqueue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
313 * Are we enqueueing a waiting task? (for current tasks
314 * a dequeue/enqueue event is a NOP)
316 if (se
!= cfs_rq
->curr
)
317 update_stats_wait_start(cfs_rq
, se
);
321 se
->fair_key
= se
->vruntime
;
325 update_stats_wait_end(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
327 schedstat_set(se
->wait_max
, max(se
->wait_max
,
328 rq_of(cfs_rq
)->clock
- se
->wait_start
));
329 schedstat_set(se
->wait_start
, 0);
333 update_stats_dequeue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
337 * Mark the end of the wait period if dequeueing a
340 if (se
!= cfs_rq
->curr
)
341 update_stats_wait_end(cfs_rq
, se
);
345 * We are picking a new current task - update its stats:
348 update_stats_curr_start(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
351 * We are starting a new run period:
353 se
->exec_start
= rq_of(cfs_rq
)->clock
;
357 * We are descheduling a task - update its stats:
360 update_stats_curr_end(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
365 /**************************************************
366 * Scheduling class queueing methods:
369 static void enqueue_sleeper(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
371 #ifdef CONFIG_SCHEDSTATS
372 if (se
->sleep_start
) {
373 u64 delta
= rq_of(cfs_rq
)->clock
- se
->sleep_start
;
378 if (unlikely(delta
> se
->sleep_max
))
379 se
->sleep_max
= delta
;
382 se
->sum_sleep_runtime
+= delta
;
384 if (se
->block_start
) {
385 u64 delta
= rq_of(cfs_rq
)->clock
- se
->block_start
;
390 if (unlikely(delta
> se
->block_max
))
391 se
->block_max
= delta
;
394 se
->sum_sleep_runtime
+= delta
;
397 * Blocking time is in units of nanosecs, so shift by 20 to
398 * get a milliseconds-range estimation of the amount of
399 * time that the task spent sleeping:
401 if (unlikely(prof_on
== SLEEP_PROFILING
)) {
402 struct task_struct
*tsk
= task_of(se
);
404 profile_hits(SLEEP_PROFILING
, (void *)get_wchan(tsk
),
412 place_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int initial
)
414 u64 min_runtime
, latency
;
416 min_runtime
= cfs_rq
->min_vruntime
;
418 if (sched_feat(USE_TREE_AVG
)) {
419 struct sched_entity
*last
= __pick_last_entity(cfs_rq
);
421 min_runtime
= __pick_next_entity(cfs_rq
)->vruntime
;
422 min_runtime
+= last
->vruntime
;
425 } else if (sched_feat(APPROX_AVG
))
426 min_runtime
+= sysctl_sched_latency
/2;
428 if (initial
&& sched_feat(START_DEBIT
))
429 min_runtime
+= sched_slice(cfs_rq
, se
);
431 if (!initial
&& sched_feat(NEW_FAIR_SLEEPERS
)) {
432 latency
= sysctl_sched_latency
;
433 if (min_runtime
> latency
)
434 min_runtime
-= latency
;
439 se
->vruntime
= max(se
->vruntime
, min_runtime
);
443 enqueue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int wakeup
)
446 * Update the fair clock.
451 place_entity(cfs_rq
, se
, 0);
452 enqueue_sleeper(cfs_rq
, se
);
455 update_stats_enqueue(cfs_rq
, se
);
456 __enqueue_entity(cfs_rq
, se
);
460 dequeue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int sleep
)
462 update_stats_dequeue(cfs_rq
, se
);
463 #ifdef CONFIG_SCHEDSTATS
465 if (entity_is_task(se
)) {
466 struct task_struct
*tsk
= task_of(se
);
468 if (tsk
->state
& TASK_INTERRUPTIBLE
)
469 se
->sleep_start
= rq_of(cfs_rq
)->clock
;
470 if (tsk
->state
& TASK_UNINTERRUPTIBLE
)
471 se
->block_start
= rq_of(cfs_rq
)->clock
;
475 __dequeue_entity(cfs_rq
, se
);
479 * Preempt the current task with a newly woken task if needed:
482 check_preempt_tick(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
)
484 unsigned long ideal_runtime
, delta_exec
;
486 ideal_runtime
= sched_slice(cfs_rq
, curr
);
487 delta_exec
= curr
->sum_exec_runtime
- curr
->prev_sum_exec_runtime
;
488 if (delta_exec
> ideal_runtime
)
489 resched_task(rq_of(cfs_rq
)->curr
);
493 set_next_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
496 * Any task has to be enqueued before it get to execute on
497 * a CPU. So account for the time it spent waiting on the
500 update_stats_wait_end(cfs_rq
, se
);
501 update_stats_curr_start(cfs_rq
, se
);
503 #ifdef CONFIG_SCHEDSTATS
505 * Track our maximum slice length, if the CPU's load is at
506 * least twice that of our own weight (i.e. dont track it
507 * when there are only lesser-weight tasks around):
509 if (rq_of(cfs_rq
)->load
.weight
>= 2*se
->load
.weight
) {
510 se
->slice_max
= max(se
->slice_max
,
511 se
->sum_exec_runtime
- se
->prev_sum_exec_runtime
);
514 se
->prev_sum_exec_runtime
= se
->sum_exec_runtime
;
517 static struct sched_entity
*pick_next_entity(struct cfs_rq
*cfs_rq
)
519 struct sched_entity
*se
= __pick_next_entity(cfs_rq
);
521 set_next_entity(cfs_rq
, se
);
526 static void put_prev_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*prev
)
529 * If still on the runqueue then deactivate_task()
530 * was not called and update_curr() has to be done:
535 update_stats_curr_end(cfs_rq
, prev
);
538 update_stats_wait_start(cfs_rq
, prev
);
542 static void entity_tick(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
)
545 * Dequeue and enqueue the task to update its
546 * position within the tree:
548 dequeue_entity(cfs_rq
, curr
, 0);
549 enqueue_entity(cfs_rq
, curr
, 0);
551 if (cfs_rq
->nr_running
> 1)
552 check_preempt_tick(cfs_rq
, curr
);
555 /**************************************************
556 * CFS operations on tasks:
559 #ifdef CONFIG_FAIR_GROUP_SCHED
561 /* Walk up scheduling entities hierarchy */
562 #define for_each_sched_entity(se) \
563 for (; se; se = se->parent)
565 static inline struct cfs_rq
*task_cfs_rq(struct task_struct
*p
)
570 /* runqueue on which this entity is (to be) queued */
571 static inline struct cfs_rq
*cfs_rq_of(struct sched_entity
*se
)
576 /* runqueue "owned" by this group */
577 static inline struct cfs_rq
*group_cfs_rq(struct sched_entity
*grp
)
582 /* Given a group's cfs_rq on one cpu, return its corresponding cfs_rq on
583 * another cpu ('this_cpu')
585 static inline struct cfs_rq
*cpu_cfs_rq(struct cfs_rq
*cfs_rq
, int this_cpu
)
587 /* A later patch will take group into account */
588 return &cpu_rq(this_cpu
)->cfs
;
591 /* Iterate thr' all leaf cfs_rq's on a runqueue */
592 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
593 list_for_each_entry(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
595 /* Do the two (enqueued) tasks belong to the same group ? */
596 static inline int is_same_group(struct task_struct
*curr
, struct task_struct
*p
)
598 if (curr
->se
.cfs_rq
== p
->se
.cfs_rq
)
604 #else /* CONFIG_FAIR_GROUP_SCHED */
606 #define for_each_sched_entity(se) \
607 for (; se; se = NULL)
609 static inline struct cfs_rq
*task_cfs_rq(struct task_struct
*p
)
611 return &task_rq(p
)->cfs
;
614 static inline struct cfs_rq
*cfs_rq_of(struct sched_entity
*se
)
616 struct task_struct
*p
= task_of(se
);
617 struct rq
*rq
= task_rq(p
);
622 /* runqueue "owned" by this group */
623 static inline struct cfs_rq
*group_cfs_rq(struct sched_entity
*grp
)
628 static inline struct cfs_rq
*cpu_cfs_rq(struct cfs_rq
*cfs_rq
, int this_cpu
)
630 return &cpu_rq(this_cpu
)->cfs
;
633 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
634 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
636 static inline int is_same_group(struct task_struct
*curr
, struct task_struct
*p
)
641 #endif /* CONFIG_FAIR_GROUP_SCHED */
644 * The enqueue_task method is called before nr_running is
645 * increased. Here we update the fair scheduling stats and
646 * then put the task into the rbtree:
648 static void enqueue_task_fair(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
650 struct cfs_rq
*cfs_rq
;
651 struct sched_entity
*se
= &p
->se
;
653 for_each_sched_entity(se
) {
656 cfs_rq
= cfs_rq_of(se
);
657 enqueue_entity(cfs_rq
, se
, wakeup
);
662 * The dequeue_task method is called before nr_running is
663 * decreased. We remove the task from the rbtree and
664 * update the fair scheduling stats:
666 static void dequeue_task_fair(struct rq
*rq
, struct task_struct
*p
, int sleep
)
668 struct cfs_rq
*cfs_rq
;
669 struct sched_entity
*se
= &p
->se
;
671 for_each_sched_entity(se
) {
672 cfs_rq
= cfs_rq_of(se
);
673 dequeue_entity(cfs_rq
, se
, sleep
);
674 /* Don't dequeue parent if it has other entities besides us */
675 if (cfs_rq
->load
.weight
)
681 * sched_yield() support is very simple - we dequeue and enqueue.
683 * If compat_yield is turned on then we requeue to the end of the tree.
685 static void yield_task_fair(struct rq
*rq
, struct task_struct
*p
)
687 struct cfs_rq
*cfs_rq
= task_cfs_rq(p
);
688 struct rb_node
**link
= &cfs_rq
->tasks_timeline
.rb_node
;
689 struct sched_entity
*rightmost
, *se
= &p
->se
;
690 struct rb_node
*parent
;
693 * Are we the only task in the tree?
695 if (unlikely(cfs_rq
->nr_running
== 1))
698 if (likely(!sysctl_sched_compat_yield
)) {
699 __update_rq_clock(rq
);
701 * Dequeue and enqueue the task to update its
702 * position within the tree:
704 dequeue_entity(cfs_rq
, &p
->se
, 0);
705 enqueue_entity(cfs_rq
, &p
->se
, 0);
710 * Find the rightmost entry in the rbtree:
714 link
= &parent
->rb_right
;
717 rightmost
= rb_entry(parent
, struct sched_entity
, run_node
);
719 * Already in the rightmost position?
721 if (unlikely(rightmost
== se
))
725 * Minimally necessary key value to be last in the tree:
727 se
->fair_key
= rightmost
->fair_key
+ 1;
729 if (cfs_rq
->rb_leftmost
== &se
->run_node
)
730 cfs_rq
->rb_leftmost
= rb_next(&se
->run_node
);
732 * Relink the task to the rightmost position:
734 rb_erase(&se
->run_node
, &cfs_rq
->tasks_timeline
);
735 rb_link_node(&se
->run_node
, parent
, link
);
736 rb_insert_color(&se
->run_node
, &cfs_rq
->tasks_timeline
);
740 * Preempt the current task with a newly woken task if needed:
742 static void check_preempt_wakeup(struct rq
*rq
, struct task_struct
*p
)
744 struct task_struct
*curr
= rq
->curr
;
745 struct cfs_rq
*cfs_rq
= task_cfs_rq(curr
);
747 if (unlikely(rt_prio(p
->prio
))) {
753 if (is_same_group(curr
, p
)) {
754 s64 delta
= curr
->se
.vruntime
- p
->se
.vruntime
;
756 if (delta
> (s64
)sysctl_sched_wakeup_granularity
)
761 static struct task_struct
*pick_next_task_fair(struct rq
*rq
)
763 struct cfs_rq
*cfs_rq
= &rq
->cfs
;
764 struct sched_entity
*se
;
766 if (unlikely(!cfs_rq
->nr_running
))
770 se
= pick_next_entity(cfs_rq
);
771 cfs_rq
= group_cfs_rq(se
);
778 * Account for a descheduled task:
780 static void put_prev_task_fair(struct rq
*rq
, struct task_struct
*prev
)
782 struct sched_entity
*se
= &prev
->se
;
783 struct cfs_rq
*cfs_rq
;
785 for_each_sched_entity(se
) {
786 cfs_rq
= cfs_rq_of(se
);
787 put_prev_entity(cfs_rq
, se
);
791 /**************************************************
792 * Fair scheduling class load-balancing methods:
796 * Load-balancing iterator. Note: while the runqueue stays locked
797 * during the whole iteration, the current task might be
798 * dequeued so the iterator has to be dequeue-safe. Here we
799 * achieve that by always pre-iterating before returning
802 static inline struct task_struct
*
803 __load_balance_iterator(struct cfs_rq
*cfs_rq
, struct rb_node
*curr
)
805 struct task_struct
*p
;
810 p
= rb_entry(curr
, struct task_struct
, se
.run_node
);
811 cfs_rq
->rb_load_balance_curr
= rb_next(curr
);
816 static struct task_struct
*load_balance_start_fair(void *arg
)
818 struct cfs_rq
*cfs_rq
= arg
;
820 return __load_balance_iterator(cfs_rq
, first_fair(cfs_rq
));
823 static struct task_struct
*load_balance_next_fair(void *arg
)
825 struct cfs_rq
*cfs_rq
= arg
;
827 return __load_balance_iterator(cfs_rq
, cfs_rq
->rb_load_balance_curr
);
830 #ifdef CONFIG_FAIR_GROUP_SCHED
831 static int cfs_rq_best_prio(struct cfs_rq
*cfs_rq
)
833 struct sched_entity
*curr
;
834 struct task_struct
*p
;
836 if (!cfs_rq
->nr_running
)
839 curr
= __pick_next_entity(cfs_rq
);
847 load_balance_fair(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
848 unsigned long max_nr_move
, unsigned long max_load_move
,
849 struct sched_domain
*sd
, enum cpu_idle_type idle
,
850 int *all_pinned
, int *this_best_prio
)
852 struct cfs_rq
*busy_cfs_rq
;
853 unsigned long load_moved
, total_nr_moved
= 0, nr_moved
;
854 long rem_load_move
= max_load_move
;
855 struct rq_iterator cfs_rq_iterator
;
857 cfs_rq_iterator
.start
= load_balance_start_fair
;
858 cfs_rq_iterator
.next
= load_balance_next_fair
;
860 for_each_leaf_cfs_rq(busiest
, busy_cfs_rq
) {
861 #ifdef CONFIG_FAIR_GROUP_SCHED
862 struct cfs_rq
*this_cfs_rq
;
864 unsigned long maxload
;
866 this_cfs_rq
= cpu_cfs_rq(busy_cfs_rq
, this_cpu
);
868 imbalance
= busy_cfs_rq
->load
.weight
- this_cfs_rq
->load
.weight
;
869 /* Don't pull if this_cfs_rq has more load than busy_cfs_rq */
873 /* Don't pull more than imbalance/2 */
875 maxload
= min(rem_load_move
, imbalance
);
877 *this_best_prio
= cfs_rq_best_prio(this_cfs_rq
);
879 # define maxload rem_load_move
881 /* pass busy_cfs_rq argument into
882 * load_balance_[start|next]_fair iterators
884 cfs_rq_iterator
.arg
= busy_cfs_rq
;
885 nr_moved
= balance_tasks(this_rq
, this_cpu
, busiest
,
886 max_nr_move
, maxload
, sd
, idle
, all_pinned
,
887 &load_moved
, this_best_prio
, &cfs_rq_iterator
);
889 total_nr_moved
+= nr_moved
;
890 max_nr_move
-= nr_moved
;
891 rem_load_move
-= load_moved
;
893 if (max_nr_move
<= 0 || rem_load_move
<= 0)
897 return max_load_move
- rem_load_move
;
901 * scheduler tick hitting a task of our scheduling class:
903 static void task_tick_fair(struct rq
*rq
, struct task_struct
*curr
)
905 struct cfs_rq
*cfs_rq
;
906 struct sched_entity
*se
= &curr
->se
;
908 for_each_sched_entity(se
) {
909 cfs_rq
= cfs_rq_of(se
);
910 entity_tick(cfs_rq
, se
);
914 #define swap(a,b) do { typeof(a) tmp = (a); (a) = (b); (b) = tmp; } while (0)
917 * Share the fairness runtime between parent and child, thus the
918 * total amount of pressure for CPU stays equal - new tasks
919 * get a chance to run but frequent forkers are not allowed to
920 * monopolize the CPU. Note: the parent runqueue is locked,
921 * the child is not running yet.
923 static void task_new_fair(struct rq
*rq
, struct task_struct
*p
)
925 struct cfs_rq
*cfs_rq
= task_cfs_rq(p
);
926 struct sched_entity
*se
= &p
->se
, *curr
= cfs_rq
->curr
;
928 sched_info_queued(p
);
931 place_entity(cfs_rq
, se
, 1);
933 if (sysctl_sched_child_runs_first
&&
934 curr
->vruntime
< se
->vruntime
) {
936 dequeue_entity(cfs_rq
, curr
, 0);
937 swap(curr
->vruntime
, se
->vruntime
);
938 enqueue_entity(cfs_rq
, curr
, 0);
941 update_stats_enqueue(cfs_rq
, se
);
942 __enqueue_entity(cfs_rq
, se
);
943 resched_task(rq
->curr
);
946 #ifdef CONFIG_FAIR_GROUP_SCHED
947 /* Account for a task changing its policy or group.
949 * This routine is mostly called to set cfs_rq->curr field when a task
950 * migrates between groups/classes.
952 static void set_curr_task_fair(struct rq
*rq
)
954 struct sched_entity
*se
= &rq
->curr
->se
;
956 for_each_sched_entity(se
)
957 set_next_entity(cfs_rq_of(se
), se
);
960 static void set_curr_task_fair(struct rq
*rq
)
966 * All the scheduling class methods:
968 struct sched_class fair_sched_class __read_mostly
= {
969 .enqueue_task
= enqueue_task_fair
,
970 .dequeue_task
= dequeue_task_fair
,
971 .yield_task
= yield_task_fair
,
973 .check_preempt_curr
= check_preempt_wakeup
,
975 .pick_next_task
= pick_next_task_fair
,
976 .put_prev_task
= put_prev_task_fair
,
978 .load_balance
= load_balance_fair
,
980 .set_curr_task
= set_curr_task_fair
,
981 .task_tick
= task_tick_fair
,
982 .task_new
= task_new_fair
,
985 #ifdef CONFIG_SCHED_DEBUG
986 static void print_cfs_stats(struct seq_file
*m
, int cpu
)
988 struct cfs_rq
*cfs_rq
;
990 for_each_leaf_cfs_rq(cpu_rq(cpu
), cfs_rq
)
991 print_cfs_rq(m
, cpu
, cfs_rq
);