2 * Completely Fair Scheduling (CFS) Class (SCHED_NORMAL/SCHED_BATCH)
4 * Copyright (C) 2007 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
6 * Interactivity improvements by Mike Galbraith
7 * (C) 2007 Mike Galbraith <efault@gmx.de>
9 * Various enhancements by Dmitry Adamushko.
10 * (C) 2007 Dmitry Adamushko <dmitry.adamushko@gmail.com>
12 * Group scheduling enhancements by Srivatsa Vaddagiri
13 * Copyright IBM Corporation, 2007
14 * Author: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
16 * Scaled math optimizations by Thomas Gleixner
17 * Copyright (C) 2007, Thomas Gleixner <tglx@linutronix.de>
19 * Adaptive scheduling granularity, math enhancements by Peter Zijlstra
20 * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
24 * Targeted preemption latency for CPU-bound tasks:
25 * (default: 20ms, units: nanoseconds)
27 * NOTE: this latency value is not the same as the concept of
28 * 'timeslice length' - timeslices in CFS are of variable length.
29 * (to see the precise effective timeslice length of your workload,
30 * run vmstat and monitor the context-switches field)
32 * On SMP systems the value of this is multiplied by the log2 of the
33 * number of CPUs. (i.e. factor 2x on 2-way systems, 3x on 4-way
34 * systems, 4x on 8-way systems, 5x on 16-way systems, etc.)
35 * Targeted preemption latency for CPU-bound tasks:
37 const_debug
unsigned int sysctl_sched_latency
= 20000000ULL;
40 * After fork, child runs first. (default) If set to 0 then
41 * parent will (try to) run first.
43 const_debug
unsigned int sysctl_sched_child_runs_first
= 1;
46 * Minimal preemption granularity for CPU-bound tasks:
47 * (default: 2 msec, units: nanoseconds)
49 const_debug
unsigned int sysctl_sched_nr_latency
= 20;
52 * sys_sched_yield() compat mode
54 * This option switches the agressive yield implementation of the
55 * old scheduler back on.
57 unsigned int __read_mostly sysctl_sched_compat_yield
;
60 * SCHED_BATCH wake-up granularity.
61 * (default: 10 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
= 10000000UL;
70 * SCHED_OTHER wake-up granularity.
71 * (default: 10 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
= 10000000UL;
79 /**************************************************************
80 * CFS operations on generic schedulable entities:
83 #ifdef CONFIG_FAIR_GROUP_SCHED
85 /* cpu runqueue to which this cfs_rq is attached */
86 static inline struct rq
*rq_of(struct cfs_rq
*cfs_rq
)
91 /* An entity is a task if it doesn't "own" a runqueue */
92 #define entity_is_task(se) (!se->my_q)
94 #else /* CONFIG_FAIR_GROUP_SCHED */
96 static inline struct rq
*rq_of(struct cfs_rq
*cfs_rq
)
98 return container_of(cfs_rq
, struct rq
, cfs
);
101 #define entity_is_task(se) 1
103 #endif /* CONFIG_FAIR_GROUP_SCHED */
105 static inline struct task_struct
*task_of(struct sched_entity
*se
)
107 return container_of(se
, struct task_struct
, se
);
111 /**************************************************************
112 * Scheduling class tree data structure manipulation methods:
115 static inline u64
max_vruntime(u64 min_vruntime
, u64 vruntime
)
117 s64 delta
= (s64
)(vruntime
- min_vruntime
);
119 min_vruntime
= vruntime
;
124 static inline u64
min_vruntime(u64 min_vruntime
, u64 vruntime
)
126 s64 delta
= (s64
)(vruntime
- min_vruntime
);
128 min_vruntime
= vruntime
;
133 static inline s64
entity_key(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
135 return se
->vruntime
- cfs_rq
->min_vruntime
;
139 * Enqueue an entity into the rb-tree:
141 static void __enqueue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
143 struct rb_node
**link
= &cfs_rq
->tasks_timeline
.rb_node
;
144 struct rb_node
*parent
= NULL
;
145 struct sched_entity
*entry
;
146 s64 key
= entity_key(cfs_rq
, se
);
150 * Find the right place in the rbtree:
154 entry
= rb_entry(parent
, struct sched_entity
, run_node
);
156 * We dont care about collisions. Nodes with
157 * the same key stay together.
159 if (key
< entity_key(cfs_rq
, entry
)) {
160 link
= &parent
->rb_left
;
162 link
= &parent
->rb_right
;
168 * Maintain a cache of leftmost tree entries (it is frequently
172 cfs_rq
->rb_leftmost
= &se
->run_node
;
174 rb_link_node(&se
->run_node
, parent
, link
);
175 rb_insert_color(&se
->run_node
, &cfs_rq
->tasks_timeline
);
178 static void __dequeue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
180 if (cfs_rq
->rb_leftmost
== &se
->run_node
)
181 cfs_rq
->rb_leftmost
= rb_next(&se
->run_node
);
183 rb_erase(&se
->run_node
, &cfs_rq
->tasks_timeline
);
186 static inline struct rb_node
*first_fair(struct cfs_rq
*cfs_rq
)
188 return cfs_rq
->rb_leftmost
;
191 static struct sched_entity
*__pick_next_entity(struct cfs_rq
*cfs_rq
)
193 return rb_entry(first_fair(cfs_rq
), struct sched_entity
, run_node
);
196 static inline struct sched_entity
*__pick_last_entity(struct cfs_rq
*cfs_rq
)
198 struct rb_node
**link
= &cfs_rq
->tasks_timeline
.rb_node
;
199 struct sched_entity
*se
= NULL
;
200 struct rb_node
*parent
;
204 se
= rb_entry(parent
, struct sched_entity
, run_node
);
205 link
= &parent
->rb_right
;
211 /**************************************************************
212 * Scheduling class statistics methods:
217 * The idea is to set a period in which each task runs once.
219 * When there are too many tasks (sysctl_sched_nr_latency) we have to stretch
220 * this period because otherwise the slices get too small.
222 * p = (nr <= nl) ? l : l*nr/nl
224 static u64
__sched_period(unsigned long nr_running
)
226 u64 period
= sysctl_sched_latency
;
227 unsigned long nr_latency
= sysctl_sched_nr_latency
;
229 if (unlikely(nr_running
> nr_latency
)) {
230 period
*= nr_running
;
231 do_div(period
, nr_latency
);
238 * We calculate the wall-time slice from the period by taking a part
239 * proportional to the weight.
243 static u64
sched_slice(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
245 u64 slice
= __sched_period(cfs_rq
->nr_running
);
247 slice
*= se
->load
.weight
;
248 do_div(slice
, cfs_rq
->load
.weight
);
254 * We calculate the vruntime slice.
258 static u64
__sched_vslice(unsigned long rq_weight
, unsigned long nr_running
)
260 u64 vslice
= __sched_period(nr_running
);
262 do_div(vslice
, rq_weight
);
267 static u64
sched_vslice(struct cfs_rq
*cfs_rq
)
269 return __sched_vslice(cfs_rq
->load
.weight
, cfs_rq
->nr_running
);
272 static u64
sched_vslice_add(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
274 return __sched_vslice(cfs_rq
->load
.weight
+ se
->load
.weight
,
275 cfs_rq
->nr_running
+ 1);
279 * Update the current task's runtime statistics. Skip current tasks that
280 * are not in our scheduling class.
283 __update_curr(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
,
284 unsigned long delta_exec
)
286 unsigned long delta_exec_weighted
;
289 schedstat_set(curr
->exec_max
, max((u64
)delta_exec
, curr
->exec_max
));
291 curr
->sum_exec_runtime
+= delta_exec
;
292 schedstat_add(cfs_rq
, exec_clock
, delta_exec
);
293 delta_exec_weighted
= delta_exec
;
294 if (unlikely(curr
->load
.weight
!= NICE_0_LOAD
)) {
295 delta_exec_weighted
= calc_delta_fair(delta_exec_weighted
,
298 curr
->vruntime
+= delta_exec_weighted
;
301 * maintain cfs_rq->min_vruntime to be a monotonic increasing
302 * value tracking the leftmost vruntime in the tree.
304 if (first_fair(cfs_rq
)) {
305 vruntime
= min_vruntime(curr
->vruntime
,
306 __pick_next_entity(cfs_rq
)->vruntime
);
308 vruntime
= curr
->vruntime
;
310 cfs_rq
->min_vruntime
=
311 max_vruntime(cfs_rq
->min_vruntime
, vruntime
);
314 static void update_curr(struct cfs_rq
*cfs_rq
)
316 struct sched_entity
*curr
= cfs_rq
->curr
;
317 u64 now
= rq_of(cfs_rq
)->clock
;
318 unsigned long delta_exec
;
324 * Get the amount of time the current task was running
325 * since the last time we changed load (this cannot
326 * overflow on 32 bits):
328 delta_exec
= (unsigned long)(now
- curr
->exec_start
);
330 __update_curr(cfs_rq
, curr
, delta_exec
);
331 curr
->exec_start
= now
;
335 update_stats_wait_start(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
337 schedstat_set(se
->wait_start
, rq_of(cfs_rq
)->clock
);
341 * Task is being enqueued - update stats:
343 static void update_stats_enqueue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
346 * Are we enqueueing a waiting task? (for current tasks
347 * a dequeue/enqueue event is a NOP)
349 if (se
!= cfs_rq
->curr
)
350 update_stats_wait_start(cfs_rq
, se
);
354 update_stats_wait_end(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
356 schedstat_set(se
->wait_max
, max(se
->wait_max
,
357 rq_of(cfs_rq
)->clock
- se
->wait_start
));
358 schedstat_set(se
->wait_start
, 0);
362 update_stats_dequeue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
365 * Mark the end of the wait period if dequeueing a
368 if (se
!= cfs_rq
->curr
)
369 update_stats_wait_end(cfs_rq
, se
);
373 * We are picking a new current task - update its stats:
376 update_stats_curr_start(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
379 * We are starting a new run period:
381 se
->exec_start
= rq_of(cfs_rq
)->clock
;
385 * We are descheduling a task - update its stats:
388 update_stats_curr_end(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
393 /**************************************************
394 * Scheduling class queueing methods:
398 account_entity_enqueue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
400 update_load_add(&cfs_rq
->load
, se
->load
.weight
);
401 cfs_rq
->nr_running
++;
406 account_entity_dequeue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
408 update_load_sub(&cfs_rq
->load
, se
->load
.weight
);
409 cfs_rq
->nr_running
--;
413 static void enqueue_sleeper(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
415 #ifdef CONFIG_SCHEDSTATS
416 if (se
->sleep_start
) {
417 u64 delta
= rq_of(cfs_rq
)->clock
- se
->sleep_start
;
422 if (unlikely(delta
> se
->sleep_max
))
423 se
->sleep_max
= delta
;
426 se
->sum_sleep_runtime
+= delta
;
428 if (se
->block_start
) {
429 u64 delta
= rq_of(cfs_rq
)->clock
- se
->block_start
;
434 if (unlikely(delta
> se
->block_max
))
435 se
->block_max
= delta
;
438 se
->sum_sleep_runtime
+= delta
;
441 * Blocking time is in units of nanosecs, so shift by 20 to
442 * get a milliseconds-range estimation of the amount of
443 * time that the task spent sleeping:
445 if (unlikely(prof_on
== SLEEP_PROFILING
)) {
446 struct task_struct
*tsk
= task_of(se
);
448 profile_hits(SLEEP_PROFILING
, (void *)get_wchan(tsk
),
455 static void check_spread(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
457 #ifdef CONFIG_SCHED_DEBUG
458 s64 d
= se
->vruntime
- cfs_rq
->min_vruntime
;
463 if (d
> 3*sysctl_sched_latency
)
464 schedstat_inc(cfs_rq
, nr_spread_over
);
469 place_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int initial
)
473 vruntime
= cfs_rq
->min_vruntime
;
475 if (sched_feat(TREE_AVG
)) {
476 struct sched_entity
*last
= __pick_last_entity(cfs_rq
);
478 vruntime
+= last
->vruntime
;
481 } else if (sched_feat(APPROX_AVG
) && cfs_rq
->nr_running
)
482 vruntime
+= sched_vslice(cfs_rq
)/2;
484 if (initial
&& sched_feat(START_DEBIT
))
485 vruntime
+= sched_vslice_add(cfs_rq
, se
);
488 if (sched_feat(NEW_FAIR_SLEEPERS
))
489 vruntime
-= sysctl_sched_latency
;
491 vruntime
= max_t(s64
, vruntime
, se
->vruntime
);
494 se
->vruntime
= vruntime
;
499 enqueue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int wakeup
)
502 * Update run-time statistics of the 'current'.
507 place_entity(cfs_rq
, se
, 0);
508 enqueue_sleeper(cfs_rq
, se
);
511 update_stats_enqueue(cfs_rq
, se
);
512 check_spread(cfs_rq
, se
);
513 if (se
!= cfs_rq
->curr
)
514 __enqueue_entity(cfs_rq
, se
);
515 account_entity_enqueue(cfs_rq
, se
);
519 dequeue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int sleep
)
522 * Update run-time statistics of the 'current'.
526 update_stats_dequeue(cfs_rq
, se
);
528 #ifdef CONFIG_SCHEDSTATS
529 if (entity_is_task(se
)) {
530 struct task_struct
*tsk
= task_of(se
);
532 if (tsk
->state
& TASK_INTERRUPTIBLE
)
533 se
->sleep_start
= rq_of(cfs_rq
)->clock
;
534 if (tsk
->state
& TASK_UNINTERRUPTIBLE
)
535 se
->block_start
= rq_of(cfs_rq
)->clock
;
540 if (se
!= cfs_rq
->curr
)
541 __dequeue_entity(cfs_rq
, se
);
542 account_entity_dequeue(cfs_rq
, se
);
546 * Preempt the current task with a newly woken task if needed:
549 check_preempt_tick(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
)
551 unsigned long ideal_runtime
, delta_exec
;
553 ideal_runtime
= sched_slice(cfs_rq
, curr
);
554 delta_exec
= curr
->sum_exec_runtime
- curr
->prev_sum_exec_runtime
;
555 if (delta_exec
> ideal_runtime
)
556 resched_task(rq_of(cfs_rq
)->curr
);
560 set_next_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
562 /* 'current' is not kept within the tree. */
565 * Any task has to be enqueued before it get to execute on
566 * a CPU. So account for the time it spent waiting on the
569 update_stats_wait_end(cfs_rq
, se
);
570 __dequeue_entity(cfs_rq
, se
);
573 update_stats_curr_start(cfs_rq
, se
);
575 #ifdef CONFIG_SCHEDSTATS
577 * Track our maximum slice length, if the CPU's load is at
578 * least twice that of our own weight (i.e. dont track it
579 * when there are only lesser-weight tasks around):
581 if (rq_of(cfs_rq
)->load
.weight
>= 2*se
->load
.weight
) {
582 se
->slice_max
= max(se
->slice_max
,
583 se
->sum_exec_runtime
- se
->prev_sum_exec_runtime
);
586 se
->prev_sum_exec_runtime
= se
->sum_exec_runtime
;
589 static struct sched_entity
*pick_next_entity(struct cfs_rq
*cfs_rq
)
591 struct sched_entity
*se
= NULL
;
593 if (first_fair(cfs_rq
)) {
594 se
= __pick_next_entity(cfs_rq
);
595 set_next_entity(cfs_rq
, se
);
601 static void put_prev_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*prev
)
604 * If still on the runqueue then deactivate_task()
605 * was not called and update_curr() has to be done:
610 update_stats_curr_end(cfs_rq
, prev
);
612 check_spread(cfs_rq
, prev
);
614 update_stats_wait_start(cfs_rq
, prev
);
615 /* Put 'current' back into the tree. */
616 __enqueue_entity(cfs_rq
, prev
);
621 static void entity_tick(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
)
624 * Update run-time statistics of the 'current'.
628 if (cfs_rq
->nr_running
> 1)
629 check_preempt_tick(cfs_rq
, curr
);
632 /**************************************************
633 * CFS operations on tasks:
636 #ifdef CONFIG_FAIR_GROUP_SCHED
638 /* Walk up scheduling entities hierarchy */
639 #define for_each_sched_entity(se) \
640 for (; se; se = se->parent)
642 static inline struct cfs_rq
*task_cfs_rq(struct task_struct
*p
)
647 /* runqueue on which this entity is (to be) queued */
648 static inline struct cfs_rq
*cfs_rq_of(struct sched_entity
*se
)
653 /* runqueue "owned" by this group */
654 static inline struct cfs_rq
*group_cfs_rq(struct sched_entity
*grp
)
659 /* Given a group's cfs_rq on one cpu, return its corresponding cfs_rq on
660 * another cpu ('this_cpu')
662 static inline struct cfs_rq
*cpu_cfs_rq(struct cfs_rq
*cfs_rq
, int this_cpu
)
664 return cfs_rq
->tg
->cfs_rq
[this_cpu
];
667 /* Iterate thr' all leaf cfs_rq's on a runqueue */
668 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
669 list_for_each_entry(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
671 /* Do the two (enqueued) entities belong to the same group ? */
673 is_same_group(struct sched_entity
*se
, struct sched_entity
*pse
)
675 if (se
->cfs_rq
== pse
->cfs_rq
)
681 static inline struct sched_entity
*parent_entity(struct sched_entity
*se
)
686 #else /* CONFIG_FAIR_GROUP_SCHED */
688 #define for_each_sched_entity(se) \
689 for (; se; se = NULL)
691 static inline struct cfs_rq
*task_cfs_rq(struct task_struct
*p
)
693 return &task_rq(p
)->cfs
;
696 static inline struct cfs_rq
*cfs_rq_of(struct sched_entity
*se
)
698 struct task_struct
*p
= task_of(se
);
699 struct rq
*rq
= task_rq(p
);
704 /* runqueue "owned" by this group */
705 static inline struct cfs_rq
*group_cfs_rq(struct sched_entity
*grp
)
710 static inline struct cfs_rq
*cpu_cfs_rq(struct cfs_rq
*cfs_rq
, int this_cpu
)
712 return &cpu_rq(this_cpu
)->cfs
;
715 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
716 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
719 is_same_group(struct sched_entity
*se
, struct sched_entity
*pse
)
724 static inline struct sched_entity
*parent_entity(struct sched_entity
*se
)
729 #endif /* CONFIG_FAIR_GROUP_SCHED */
732 * The enqueue_task method is called before nr_running is
733 * increased. Here we update the fair scheduling stats and
734 * then put the task into the rbtree:
736 static void enqueue_task_fair(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
738 struct cfs_rq
*cfs_rq
;
739 struct sched_entity
*se
= &p
->se
;
741 for_each_sched_entity(se
) {
744 cfs_rq
= cfs_rq_of(se
);
745 enqueue_entity(cfs_rq
, se
, wakeup
);
751 * The dequeue_task method is called before nr_running is
752 * decreased. We remove the task from the rbtree and
753 * update the fair scheduling stats:
755 static void dequeue_task_fair(struct rq
*rq
, struct task_struct
*p
, int sleep
)
757 struct cfs_rq
*cfs_rq
;
758 struct sched_entity
*se
= &p
->se
;
760 for_each_sched_entity(se
) {
761 cfs_rq
= cfs_rq_of(se
);
762 dequeue_entity(cfs_rq
, se
, sleep
);
763 /* Don't dequeue parent if it has other entities besides us */
764 if (cfs_rq
->load
.weight
)
771 * sched_yield() support is very simple - we dequeue and enqueue.
773 * If compat_yield is turned on then we requeue to the end of the tree.
775 static void yield_task_fair(struct rq
*rq
)
777 struct cfs_rq
*cfs_rq
= task_cfs_rq(rq
->curr
);
778 struct sched_entity
*rightmost
, *se
= &rq
->curr
->se
;
781 * Are we the only task in the tree?
783 if (unlikely(cfs_rq
->nr_running
== 1))
786 if (likely(!sysctl_sched_compat_yield
)) {
787 __update_rq_clock(rq
);
789 * Update run-time statistics of the 'current'.
796 * Find the rightmost entry in the rbtree:
798 rightmost
= __pick_last_entity(cfs_rq
);
800 * Already in the rightmost position?
802 if (unlikely(rightmost
->vruntime
< se
->vruntime
))
806 * Minimally necessary key value to be last in the tree:
807 * Upon rescheduling, sched_class::put_prev_task() will place
808 * 'current' within the tree based on its new key value.
810 se
->vruntime
= rightmost
->vruntime
+ 1;
814 * Preempt the current task with a newly woken task if needed:
816 static void check_preempt_wakeup(struct rq
*rq
, struct task_struct
*p
)
818 struct task_struct
*curr
= rq
->curr
;
819 struct cfs_rq
*cfs_rq
= task_cfs_rq(curr
);
820 struct sched_entity
*se
= &curr
->se
, *pse
= &p
->se
;
823 if (unlikely(rt_prio(p
->prio
))) {
830 while (!is_same_group(se
, pse
)) {
831 se
= parent_entity(se
);
832 pse
= parent_entity(pse
);
835 delta
= se
->vruntime
- pse
->vruntime
;
837 if (delta
> (s64
)sysctl_sched_wakeup_granularity
)
841 static struct task_struct
*pick_next_task_fair(struct rq
*rq
)
843 struct cfs_rq
*cfs_rq
= &rq
->cfs
;
844 struct sched_entity
*se
;
846 if (unlikely(!cfs_rq
->nr_running
))
850 se
= pick_next_entity(cfs_rq
);
851 cfs_rq
= group_cfs_rq(se
);
858 * Account for a descheduled task:
860 static void put_prev_task_fair(struct rq
*rq
, struct task_struct
*prev
)
862 struct sched_entity
*se
= &prev
->se
;
863 struct cfs_rq
*cfs_rq
;
865 for_each_sched_entity(se
) {
866 cfs_rq
= cfs_rq_of(se
);
867 put_prev_entity(cfs_rq
, se
);
871 /**************************************************
872 * Fair scheduling class load-balancing methods:
876 * Load-balancing iterator. Note: while the runqueue stays locked
877 * during the whole iteration, the current task might be
878 * dequeued so the iterator has to be dequeue-safe. Here we
879 * achieve that by always pre-iterating before returning
882 static struct task_struct
*
883 __load_balance_iterator(struct cfs_rq
*cfs_rq
, struct rb_node
*curr
)
885 struct task_struct
*p
;
890 p
= rb_entry(curr
, struct task_struct
, se
.run_node
);
891 cfs_rq
->rb_load_balance_curr
= rb_next(curr
);
896 static struct task_struct
*load_balance_start_fair(void *arg
)
898 struct cfs_rq
*cfs_rq
= arg
;
900 return __load_balance_iterator(cfs_rq
, first_fair(cfs_rq
));
903 static struct task_struct
*load_balance_next_fair(void *arg
)
905 struct cfs_rq
*cfs_rq
= arg
;
907 return __load_balance_iterator(cfs_rq
, cfs_rq
->rb_load_balance_curr
);
910 #ifdef CONFIG_FAIR_GROUP_SCHED
911 static int cfs_rq_best_prio(struct cfs_rq
*cfs_rq
)
913 struct sched_entity
*curr
;
914 struct task_struct
*p
;
916 if (!cfs_rq
->nr_running
)
921 curr
= __pick_next_entity(cfs_rq
);
930 load_balance_fair(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
931 unsigned long max_nr_move
, unsigned long max_load_move
,
932 struct sched_domain
*sd
, enum cpu_idle_type idle
,
933 int *all_pinned
, int *this_best_prio
)
935 struct cfs_rq
*busy_cfs_rq
;
936 unsigned long load_moved
, total_nr_moved
= 0, nr_moved
;
937 long rem_load_move
= max_load_move
;
938 struct rq_iterator cfs_rq_iterator
;
940 cfs_rq_iterator
.start
= load_balance_start_fair
;
941 cfs_rq_iterator
.next
= load_balance_next_fair
;
943 for_each_leaf_cfs_rq(busiest
, busy_cfs_rq
) {
944 #ifdef CONFIG_FAIR_GROUP_SCHED
945 struct cfs_rq
*this_cfs_rq
;
947 unsigned long maxload
;
949 this_cfs_rq
= cpu_cfs_rq(busy_cfs_rq
, this_cpu
);
951 imbalance
= busy_cfs_rq
->load
.weight
- this_cfs_rq
->load
.weight
;
952 /* Don't pull if this_cfs_rq has more load than busy_cfs_rq */
956 /* Don't pull more than imbalance/2 */
958 maxload
= min(rem_load_move
, imbalance
);
960 *this_best_prio
= cfs_rq_best_prio(this_cfs_rq
);
962 # define maxload rem_load_move
964 /* pass busy_cfs_rq argument into
965 * load_balance_[start|next]_fair iterators
967 cfs_rq_iterator
.arg
= busy_cfs_rq
;
968 nr_moved
= balance_tasks(this_rq
, this_cpu
, busiest
,
969 max_nr_move
, maxload
, sd
, idle
, all_pinned
,
970 &load_moved
, this_best_prio
, &cfs_rq_iterator
);
972 total_nr_moved
+= nr_moved
;
973 max_nr_move
-= nr_moved
;
974 rem_load_move
-= load_moved
;
976 if (max_nr_move
<= 0 || rem_load_move
<= 0)
980 return max_load_move
- rem_load_move
;
984 * scheduler tick hitting a task of our scheduling class:
986 static void task_tick_fair(struct rq
*rq
, struct task_struct
*curr
)
988 struct cfs_rq
*cfs_rq
;
989 struct sched_entity
*se
= &curr
->se
;
991 for_each_sched_entity(se
) {
992 cfs_rq
= cfs_rq_of(se
);
993 entity_tick(cfs_rq
, se
);
997 #define swap(a,b) do { typeof(a) tmp = (a); (a) = (b); (b) = tmp; } while (0)
1000 * Share the fairness runtime between parent and child, thus the
1001 * total amount of pressure for CPU stays equal - new tasks
1002 * get a chance to run but frequent forkers are not allowed to
1003 * monopolize the CPU. Note: the parent runqueue is locked,
1004 * the child is not running yet.
1006 static void task_new_fair(struct rq
*rq
, struct task_struct
*p
)
1008 struct cfs_rq
*cfs_rq
= task_cfs_rq(p
);
1009 struct sched_entity
*se
= &p
->se
, *curr
= cfs_rq
->curr
;
1010 int this_cpu
= smp_processor_id();
1012 sched_info_queued(p
);
1014 update_curr(cfs_rq
);
1015 place_entity(cfs_rq
, se
, 1);
1017 if (sysctl_sched_child_runs_first
&& this_cpu
== task_cpu(p
) &&
1018 curr
->vruntime
< se
->vruntime
) {
1020 * Upon rescheduling, sched_class::put_prev_task() will place
1021 * 'current' within the tree based on its new key value.
1023 swap(curr
->vruntime
, se
->vruntime
);
1026 update_stats_enqueue(cfs_rq
, se
);
1027 check_spread(cfs_rq
, se
);
1028 check_spread(cfs_rq
, curr
);
1029 __enqueue_entity(cfs_rq
, se
);
1030 account_entity_enqueue(cfs_rq
, se
);
1031 resched_task(rq
->curr
);
1034 /* Account for a task changing its policy or group.
1036 * This routine is mostly called to set cfs_rq->curr field when a task
1037 * migrates between groups/classes.
1039 static void set_curr_task_fair(struct rq
*rq
)
1041 struct sched_entity
*se
= &rq
->curr
->se
;
1043 for_each_sched_entity(se
)
1044 set_next_entity(cfs_rq_of(se
), se
);
1048 * All the scheduling class methods:
1050 static const struct sched_class fair_sched_class
= {
1051 .next
= &idle_sched_class
,
1052 .enqueue_task
= enqueue_task_fair
,
1053 .dequeue_task
= dequeue_task_fair
,
1054 .yield_task
= yield_task_fair
,
1056 .check_preempt_curr
= check_preempt_wakeup
,
1058 .pick_next_task
= pick_next_task_fair
,
1059 .put_prev_task
= put_prev_task_fair
,
1061 .load_balance
= load_balance_fair
,
1063 .set_curr_task
= set_curr_task_fair
,
1064 .task_tick
= task_tick_fair
,
1065 .task_new
= task_new_fair
,
1068 #ifdef CONFIG_SCHED_DEBUG
1069 static void print_cfs_stats(struct seq_file
*m
, int cpu
)
1071 struct cfs_rq
*cfs_rq
;
1073 #ifdef CONFIG_FAIR_GROUP_SCHED
1074 print_cfs_rq(m
, cpu
, &cpu_rq(cpu
)->cfs
);
1076 for_each_leaf_cfs_rq(cpu_rq(cpu
), cfs_rq
)
1077 print_cfs_rq(m
, cpu
, cfs_rq
);