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
++;
182 schedstat_add(cfs_rq
, wait_runtime
, se
->wait_runtime
);
186 __dequeue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
188 if (cfs_rq
->rb_leftmost
== &se
->run_node
)
189 set_leftmost(cfs_rq
, rb_next(&se
->run_node
));
191 rb_erase(&se
->run_node
, &cfs_rq
->tasks_timeline
);
192 update_load_sub(&cfs_rq
->load
, se
->load
.weight
);
193 cfs_rq
->nr_running
--;
196 schedstat_add(cfs_rq
, wait_runtime
, -se
->wait_runtime
);
199 static inline struct rb_node
*first_fair(struct cfs_rq
*cfs_rq
)
201 return cfs_rq
->rb_leftmost
;
204 static struct sched_entity
*__pick_next_entity(struct cfs_rq
*cfs_rq
)
206 return rb_entry(first_fair(cfs_rq
), struct sched_entity
, run_node
);
209 static inline struct sched_entity
*__pick_last_entity(struct cfs_rq
*cfs_rq
)
211 struct rb_node
**link
= &cfs_rq
->tasks_timeline
.rb_node
;
212 struct sched_entity
*se
= NULL
;
213 struct rb_node
*parent
;
217 se
= rb_entry(parent
, struct sched_entity
, run_node
);
218 link
= &parent
->rb_right
;
224 /**************************************************************
225 * Scheduling class statistics methods:
228 static u64
__sched_period(unsigned long nr_running
)
230 u64 period
= sysctl_sched_latency
;
231 unsigned long nr_latency
=
232 sysctl_sched_latency
/ sysctl_sched_min_granularity
;
234 if (unlikely(nr_running
> nr_latency
)) {
235 period
*= nr_running
;
236 do_div(period
, nr_latency
);
242 static u64
sched_slice(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
244 u64 period
= __sched_period(cfs_rq
->nr_running
);
246 period
*= se
->load
.weight
;
247 do_div(period
, cfs_rq
->load
.weight
);
253 limit_wait_runtime(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
255 long limit
= sysctl_sched_runtime_limit
;
258 * Niced tasks have the same history dynamic range as
261 if (unlikely(se
->wait_runtime
> limit
)) {
262 se
->wait_runtime
= limit
;
263 schedstat_inc(se
, wait_runtime_overruns
);
264 schedstat_inc(cfs_rq
, wait_runtime_overruns
);
266 if (unlikely(se
->wait_runtime
< -limit
)) {
267 se
->wait_runtime
= -limit
;
268 schedstat_inc(se
, wait_runtime_underruns
);
269 schedstat_inc(cfs_rq
, wait_runtime_underruns
);
274 __add_wait_runtime(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, long delta
)
276 se
->wait_runtime
+= delta
;
277 schedstat_add(se
, sum_wait_runtime
, delta
);
278 limit_wait_runtime(cfs_rq
, se
);
282 add_wait_runtime(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, long delta
)
284 schedstat_add(cfs_rq
, wait_runtime
, -se
->wait_runtime
);
285 __add_wait_runtime(cfs_rq
, se
, delta
);
286 schedstat_add(cfs_rq
, wait_runtime
, se
->wait_runtime
);
290 * Update the current task's runtime statistics. Skip current tasks that
291 * are not in our scheduling class.
294 __update_curr(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
,
295 unsigned long delta_exec
)
297 unsigned long delta
, delta_fair
, delta_mine
, delta_exec_weighted
;
298 struct load_weight
*lw
= &cfs_rq
->load
;
299 unsigned long load
= lw
->weight
;
301 schedstat_set(curr
->exec_max
, max((u64
)delta_exec
, curr
->exec_max
));
303 curr
->sum_exec_runtime
+= delta_exec
;
304 cfs_rq
->exec_clock
+= delta_exec
;
305 delta_exec_weighted
= delta_exec
;
306 if (unlikely(curr
->load
.weight
!= NICE_0_LOAD
)) {
307 delta_exec_weighted
= calc_delta_fair(delta_exec_weighted
,
310 curr
->vruntime
+= delta_exec_weighted
;
312 if (!sched_feat(FAIR_SLEEPERS
))
318 delta_fair
= calc_delta_fair(delta_exec
, lw
);
319 delta_mine
= calc_delta_mine(delta_exec
, curr
->load
.weight
, lw
);
321 if (cfs_rq
->sleeper_bonus
> sysctl_sched_min_granularity
) {
322 delta
= min((u64
)delta_mine
, cfs_rq
->sleeper_bonus
);
323 delta
= min(delta
, (unsigned long)(
324 (long)sysctl_sched_runtime_limit
- curr
->wait_runtime
));
325 cfs_rq
->sleeper_bonus
-= delta
;
329 cfs_rq
->fair_clock
+= delta_fair
;
331 * We executed delta_exec amount of time on the CPU,
332 * but we were only entitled to delta_mine amount of
333 * time during that period (if nr_running == 1 then
334 * the two values are equal)
335 * [Note: delta_mine - delta_exec is negative]:
337 add_wait_runtime(cfs_rq
, curr
, delta_mine
- delta_exec
);
340 static void update_curr(struct cfs_rq
*cfs_rq
)
342 struct sched_entity
*curr
= cfs_rq
->curr
;
343 u64 now
= rq_of(cfs_rq
)->clock
;
344 unsigned long delta_exec
;
350 * Get the amount of time the current task was running
351 * since the last time we changed load (this cannot
352 * overflow on 32 bits):
354 delta_exec
= (unsigned long)(now
- curr
->exec_start
);
356 __update_curr(cfs_rq
, curr
, delta_exec
);
357 curr
->exec_start
= now
;
361 update_stats_wait_start(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
363 se
->wait_start_fair
= cfs_rq
->fair_clock
;
364 schedstat_set(se
->wait_start
, rq_of(cfs_rq
)->clock
);
367 static inline unsigned long
368 calc_weighted(unsigned long delta
, struct sched_entity
*se
)
370 unsigned long weight
= se
->load
.weight
;
372 if (unlikely(weight
!= NICE_0_LOAD
))
373 return (u64
)delta
* se
->load
.weight
>> NICE_0_SHIFT
;
379 * Task is being enqueued - update stats:
381 static void update_stats_enqueue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
384 * Are we enqueueing a waiting task? (for current tasks
385 * a dequeue/enqueue event is a NOP)
387 if (se
!= cfs_rq
->curr
)
388 update_stats_wait_start(cfs_rq
, se
);
392 se
->fair_key
= se
->vruntime
;
396 * Note: must be called with a freshly updated rq->fair_clock.
399 __update_stats_wait_end(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
,
400 unsigned long delta_fair
)
402 schedstat_set(se
->wait_max
, max(se
->wait_max
,
403 rq_of(cfs_rq
)->clock
- se
->wait_start
));
405 delta_fair
= calc_weighted(delta_fair
, se
);
407 add_wait_runtime(cfs_rq
, se
, delta_fair
);
411 update_stats_wait_end(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
413 unsigned long delta_fair
;
415 if (unlikely(!se
->wait_start_fair
))
418 delta_fair
= (unsigned long)min((u64
)(2*sysctl_sched_runtime_limit
),
419 (u64
)(cfs_rq
->fair_clock
- se
->wait_start_fair
));
421 __update_stats_wait_end(cfs_rq
, se
, delta_fair
);
423 se
->wait_start_fair
= 0;
424 schedstat_set(se
->wait_start
, 0);
428 update_stats_dequeue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
432 * Mark the end of the wait period if dequeueing a
435 if (se
!= cfs_rq
->curr
)
436 update_stats_wait_end(cfs_rq
, se
);
440 * We are picking a new current task - update its stats:
443 update_stats_curr_start(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
446 * We are starting a new run period:
448 se
->exec_start
= rq_of(cfs_rq
)->clock
;
452 * We are descheduling a task - update its stats:
455 update_stats_curr_end(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
460 /**************************************************
461 * Scheduling class queueing methods:
464 static void __enqueue_sleeper(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
,
465 unsigned long delta_fair
)
467 unsigned long load
= cfs_rq
->load
.weight
;
471 * Do not boost sleepers if there's too much bonus 'in flight'
474 if (unlikely(cfs_rq
->sleeper_bonus
> sysctl_sched_runtime_limit
))
477 if (sched_feat(SLEEPER_LOAD_AVG
))
478 load
= rq_of(cfs_rq
)->cpu_load
[2];
481 * Fix up delta_fair with the effect of us running
482 * during the whole sleep period:
484 if (sched_feat(SLEEPER_AVG
))
485 delta_fair
= div64_likely32((u64
)delta_fair
* load
,
486 load
+ se
->load
.weight
);
488 delta_fair
= calc_weighted(delta_fair
, se
);
490 prev_runtime
= se
->wait_runtime
;
491 __add_wait_runtime(cfs_rq
, se
, delta_fair
);
492 delta_fair
= se
->wait_runtime
- prev_runtime
;
495 * Track the amount of bonus we've given to sleepers:
497 cfs_rq
->sleeper_bonus
+= delta_fair
;
500 static void enqueue_sleeper(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
502 struct task_struct
*tsk
= task_of(se
);
503 unsigned long delta_fair
;
505 if ((entity_is_task(se
) && tsk
->policy
== SCHED_BATCH
) ||
506 !sched_feat(FAIR_SLEEPERS
))
509 delta_fair
= (unsigned long)min((u64
)(2*sysctl_sched_runtime_limit
),
510 (u64
)(cfs_rq
->fair_clock
- se
->sleep_start_fair
));
512 __enqueue_sleeper(cfs_rq
, se
, delta_fair
);
514 se
->sleep_start_fair
= 0;
516 #ifdef CONFIG_SCHEDSTATS
517 if (se
->sleep_start
) {
518 u64 delta
= rq_of(cfs_rq
)->clock
- se
->sleep_start
;
523 if (unlikely(delta
> se
->sleep_max
))
524 se
->sleep_max
= delta
;
527 se
->sum_sleep_runtime
+= delta
;
529 if (se
->block_start
) {
530 u64 delta
= rq_of(cfs_rq
)->clock
- se
->block_start
;
535 if (unlikely(delta
> se
->block_max
))
536 se
->block_max
= delta
;
539 se
->sum_sleep_runtime
+= delta
;
542 * Blocking time is in units of nanosecs, so shift by 20 to
543 * get a milliseconds-range estimation of the amount of
544 * time that the task spent sleeping:
546 if (unlikely(prof_on
== SLEEP_PROFILING
)) {
547 profile_hits(SLEEP_PROFILING
, (void *)get_wchan(tsk
),
555 place_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int initial
)
557 u64 min_runtime
, latency
;
559 min_runtime
= cfs_rq
->min_vruntime
;
561 if (sched_feat(USE_TREE_AVG
)) {
562 struct sched_entity
*last
= __pick_last_entity(cfs_rq
);
564 min_runtime
= __pick_next_entity(cfs_rq
)->vruntime
;
565 min_runtime
+= last
->vruntime
;
568 } else if (sched_feat(APPROX_AVG
))
569 min_runtime
+= sysctl_sched_latency
/2;
571 if (initial
&& sched_feat(START_DEBIT
))
572 min_runtime
+= sched_slice(cfs_rq
, se
);
574 if (!initial
&& sched_feat(NEW_FAIR_SLEEPERS
)) {
575 latency
= sysctl_sched_latency
;
576 if (min_runtime
> latency
)
577 min_runtime
-= latency
;
582 se
->vruntime
= max(se
->vruntime
, min_runtime
);
586 enqueue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int wakeup
)
589 * Update the fair clock.
594 place_entity(cfs_rq
, se
, 0);
595 enqueue_sleeper(cfs_rq
, se
);
598 update_stats_enqueue(cfs_rq
, se
);
599 __enqueue_entity(cfs_rq
, se
);
603 dequeue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int sleep
)
605 update_stats_dequeue(cfs_rq
, se
);
607 se
->sleep_start_fair
= cfs_rq
->fair_clock
;
608 #ifdef CONFIG_SCHEDSTATS
609 if (entity_is_task(se
)) {
610 struct task_struct
*tsk
= task_of(se
);
612 if (tsk
->state
& TASK_INTERRUPTIBLE
)
613 se
->sleep_start
= rq_of(cfs_rq
)->clock
;
614 if (tsk
->state
& TASK_UNINTERRUPTIBLE
)
615 se
->block_start
= rq_of(cfs_rq
)->clock
;
619 __dequeue_entity(cfs_rq
, se
);
623 * Preempt the current task with a newly woken task if needed:
626 check_preempt_tick(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
)
628 unsigned long ideal_runtime
, delta_exec
;
630 ideal_runtime
= sched_slice(cfs_rq
, curr
);
631 delta_exec
= curr
->sum_exec_runtime
- curr
->prev_sum_exec_runtime
;
632 if (delta_exec
> ideal_runtime
)
633 resched_task(rq_of(cfs_rq
)->curr
);
637 set_next_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
640 * Any task has to be enqueued before it get to execute on
641 * a CPU. So account for the time it spent waiting on the
642 * runqueue. (note, here we rely on pick_next_task() having
643 * done a put_prev_task_fair() shortly before this, which
644 * updated rq->fair_clock - used by update_stats_wait_end())
646 update_stats_wait_end(cfs_rq
, se
);
647 update_stats_curr_start(cfs_rq
, se
);
649 #ifdef CONFIG_SCHEDSTATS
651 * Track our maximum slice length, if the CPU's load is at
652 * least twice that of our own weight (i.e. dont track it
653 * when there are only lesser-weight tasks around):
655 if (rq_of(cfs_rq
)->ls
.load
.weight
>= 2*se
->load
.weight
) {
656 se
->slice_max
= max(se
->slice_max
,
657 se
->sum_exec_runtime
- se
->prev_sum_exec_runtime
);
660 se
->prev_sum_exec_runtime
= se
->sum_exec_runtime
;
663 static struct sched_entity
*pick_next_entity(struct cfs_rq
*cfs_rq
)
665 struct sched_entity
*se
= __pick_next_entity(cfs_rq
);
667 set_next_entity(cfs_rq
, se
);
672 static void put_prev_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*prev
)
675 * If still on the runqueue then deactivate_task()
676 * was not called and update_curr() has to be done:
681 update_stats_curr_end(cfs_rq
, prev
);
684 update_stats_wait_start(cfs_rq
, prev
);
688 static void entity_tick(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
)
691 * Dequeue and enqueue the task to update its
692 * position within the tree:
694 dequeue_entity(cfs_rq
, curr
, 0);
695 enqueue_entity(cfs_rq
, curr
, 0);
697 if (cfs_rq
->nr_running
> 1)
698 check_preempt_tick(cfs_rq
, curr
);
701 /**************************************************
702 * CFS operations on tasks:
705 #ifdef CONFIG_FAIR_GROUP_SCHED
707 /* Walk up scheduling entities hierarchy */
708 #define for_each_sched_entity(se) \
709 for (; se; se = se->parent)
711 static inline struct cfs_rq
*task_cfs_rq(struct task_struct
*p
)
716 /* runqueue on which this entity is (to be) queued */
717 static inline struct cfs_rq
*cfs_rq_of(struct sched_entity
*se
)
722 /* runqueue "owned" by this group */
723 static inline struct cfs_rq
*group_cfs_rq(struct sched_entity
*grp
)
728 /* Given a group's cfs_rq on one cpu, return its corresponding cfs_rq on
729 * another cpu ('this_cpu')
731 static inline struct cfs_rq
*cpu_cfs_rq(struct cfs_rq
*cfs_rq
, int this_cpu
)
733 /* A later patch will take group into account */
734 return &cpu_rq(this_cpu
)->cfs
;
737 /* Iterate thr' all leaf cfs_rq's on a runqueue */
738 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
739 list_for_each_entry(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
741 /* Do the two (enqueued) tasks belong to the same group ? */
742 static inline int is_same_group(struct task_struct
*curr
, struct task_struct
*p
)
744 if (curr
->se
.cfs_rq
== p
->se
.cfs_rq
)
750 #else /* CONFIG_FAIR_GROUP_SCHED */
752 #define for_each_sched_entity(se) \
753 for (; se; se = NULL)
755 static inline struct cfs_rq
*task_cfs_rq(struct task_struct
*p
)
757 return &task_rq(p
)->cfs
;
760 static inline struct cfs_rq
*cfs_rq_of(struct sched_entity
*se
)
762 struct task_struct
*p
= task_of(se
);
763 struct rq
*rq
= task_rq(p
);
768 /* runqueue "owned" by this group */
769 static inline struct cfs_rq
*group_cfs_rq(struct sched_entity
*grp
)
774 static inline struct cfs_rq
*cpu_cfs_rq(struct cfs_rq
*cfs_rq
, int this_cpu
)
776 return &cpu_rq(this_cpu
)->cfs
;
779 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
780 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
782 static inline int is_same_group(struct task_struct
*curr
, struct task_struct
*p
)
787 #endif /* CONFIG_FAIR_GROUP_SCHED */
790 * The enqueue_task method is called before nr_running is
791 * increased. Here we update the fair scheduling stats and
792 * then put the task into the rbtree:
794 static void enqueue_task_fair(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
796 struct cfs_rq
*cfs_rq
;
797 struct sched_entity
*se
= &p
->se
;
799 for_each_sched_entity(se
) {
802 cfs_rq
= cfs_rq_of(se
);
803 enqueue_entity(cfs_rq
, se
, wakeup
);
808 * The dequeue_task method is called before nr_running is
809 * decreased. We remove the task from the rbtree and
810 * update the fair scheduling stats:
812 static void dequeue_task_fair(struct rq
*rq
, struct task_struct
*p
, int sleep
)
814 struct cfs_rq
*cfs_rq
;
815 struct sched_entity
*se
= &p
->se
;
817 for_each_sched_entity(se
) {
818 cfs_rq
= cfs_rq_of(se
);
819 dequeue_entity(cfs_rq
, se
, sleep
);
820 /* Don't dequeue parent if it has other entities besides us */
821 if (cfs_rq
->load
.weight
)
827 * sched_yield() support is very simple - we dequeue and enqueue.
829 * If compat_yield is turned on then we requeue to the end of the tree.
831 static void yield_task_fair(struct rq
*rq
, struct task_struct
*p
)
833 struct cfs_rq
*cfs_rq
= task_cfs_rq(p
);
834 struct rb_node
**link
= &cfs_rq
->tasks_timeline
.rb_node
;
835 struct sched_entity
*rightmost
, *se
= &p
->se
;
836 struct rb_node
*parent
;
839 * Are we the only task in the tree?
841 if (unlikely(cfs_rq
->nr_running
== 1))
844 if (likely(!sysctl_sched_compat_yield
)) {
845 __update_rq_clock(rq
);
847 * Dequeue and enqueue the task to update its
848 * position within the tree:
850 dequeue_entity(cfs_rq
, &p
->se
, 0);
851 enqueue_entity(cfs_rq
, &p
->se
, 0);
856 * Find the rightmost entry in the rbtree:
860 link
= &parent
->rb_right
;
863 rightmost
= rb_entry(parent
, struct sched_entity
, run_node
);
865 * Already in the rightmost position?
867 if (unlikely(rightmost
== se
))
871 * Minimally necessary key value to be last in the tree:
873 se
->fair_key
= rightmost
->fair_key
+ 1;
875 if (cfs_rq
->rb_leftmost
== &se
->run_node
)
876 cfs_rq
->rb_leftmost
= rb_next(&se
->run_node
);
878 * Relink the task to the rightmost position:
880 rb_erase(&se
->run_node
, &cfs_rq
->tasks_timeline
);
881 rb_link_node(&se
->run_node
, parent
, link
);
882 rb_insert_color(&se
->run_node
, &cfs_rq
->tasks_timeline
);
886 * Preempt the current task with a newly woken task if needed:
888 static void check_preempt_wakeup(struct rq
*rq
, struct task_struct
*p
)
890 struct task_struct
*curr
= rq
->curr
;
891 struct cfs_rq
*cfs_rq
= task_cfs_rq(curr
);
893 if (unlikely(rt_prio(p
->prio
))) {
899 if (is_same_group(curr
, p
)) {
900 s64 delta
= curr
->se
.vruntime
- p
->se
.vruntime
;
902 if (delta
> (s64
)sysctl_sched_wakeup_granularity
)
907 static struct task_struct
*pick_next_task_fair(struct rq
*rq
)
909 struct cfs_rq
*cfs_rq
= &rq
->cfs
;
910 struct sched_entity
*se
;
912 if (unlikely(!cfs_rq
->nr_running
))
916 se
= pick_next_entity(cfs_rq
);
917 cfs_rq
= group_cfs_rq(se
);
924 * Account for a descheduled task:
926 static void put_prev_task_fair(struct rq
*rq
, struct task_struct
*prev
)
928 struct sched_entity
*se
= &prev
->se
;
929 struct cfs_rq
*cfs_rq
;
931 for_each_sched_entity(se
) {
932 cfs_rq
= cfs_rq_of(se
);
933 put_prev_entity(cfs_rq
, se
);
937 /**************************************************
938 * Fair scheduling class load-balancing methods:
942 * Load-balancing iterator. Note: while the runqueue stays locked
943 * during the whole iteration, the current task might be
944 * dequeued so the iterator has to be dequeue-safe. Here we
945 * achieve that by always pre-iterating before returning
948 static inline struct task_struct
*
949 __load_balance_iterator(struct cfs_rq
*cfs_rq
, struct rb_node
*curr
)
951 struct task_struct
*p
;
956 p
= rb_entry(curr
, struct task_struct
, se
.run_node
);
957 cfs_rq
->rb_load_balance_curr
= rb_next(curr
);
962 static struct task_struct
*load_balance_start_fair(void *arg
)
964 struct cfs_rq
*cfs_rq
= arg
;
966 return __load_balance_iterator(cfs_rq
, first_fair(cfs_rq
));
969 static struct task_struct
*load_balance_next_fair(void *arg
)
971 struct cfs_rq
*cfs_rq
= arg
;
973 return __load_balance_iterator(cfs_rq
, cfs_rq
->rb_load_balance_curr
);
976 #ifdef CONFIG_FAIR_GROUP_SCHED
977 static int cfs_rq_best_prio(struct cfs_rq
*cfs_rq
)
979 struct sched_entity
*curr
;
980 struct task_struct
*p
;
982 if (!cfs_rq
->nr_running
)
985 curr
= __pick_next_entity(cfs_rq
);
993 load_balance_fair(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
994 unsigned long max_nr_move
, unsigned long max_load_move
,
995 struct sched_domain
*sd
, enum cpu_idle_type idle
,
996 int *all_pinned
, int *this_best_prio
)
998 struct cfs_rq
*busy_cfs_rq
;
999 unsigned long load_moved
, total_nr_moved
= 0, nr_moved
;
1000 long rem_load_move
= max_load_move
;
1001 struct rq_iterator cfs_rq_iterator
;
1003 cfs_rq_iterator
.start
= load_balance_start_fair
;
1004 cfs_rq_iterator
.next
= load_balance_next_fair
;
1006 for_each_leaf_cfs_rq(busiest
, busy_cfs_rq
) {
1007 #ifdef CONFIG_FAIR_GROUP_SCHED
1008 struct cfs_rq
*this_cfs_rq
;
1010 unsigned long maxload
;
1012 this_cfs_rq
= cpu_cfs_rq(busy_cfs_rq
, this_cpu
);
1014 imbalance
= busy_cfs_rq
->load
.weight
- this_cfs_rq
->load
.weight
;
1015 /* Don't pull if this_cfs_rq has more load than busy_cfs_rq */
1019 /* Don't pull more than imbalance/2 */
1021 maxload
= min(rem_load_move
, imbalance
);
1023 *this_best_prio
= cfs_rq_best_prio(this_cfs_rq
);
1025 # define maxload rem_load_move
1027 /* pass busy_cfs_rq argument into
1028 * load_balance_[start|next]_fair iterators
1030 cfs_rq_iterator
.arg
= busy_cfs_rq
;
1031 nr_moved
= balance_tasks(this_rq
, this_cpu
, busiest
,
1032 max_nr_move
, maxload
, sd
, idle
, all_pinned
,
1033 &load_moved
, this_best_prio
, &cfs_rq_iterator
);
1035 total_nr_moved
+= nr_moved
;
1036 max_nr_move
-= nr_moved
;
1037 rem_load_move
-= load_moved
;
1039 if (max_nr_move
<= 0 || rem_load_move
<= 0)
1043 return max_load_move
- rem_load_move
;
1047 * scheduler tick hitting a task of our scheduling class:
1049 static void task_tick_fair(struct rq
*rq
, struct task_struct
*curr
)
1051 struct cfs_rq
*cfs_rq
;
1052 struct sched_entity
*se
= &curr
->se
;
1054 for_each_sched_entity(se
) {
1055 cfs_rq
= cfs_rq_of(se
);
1056 entity_tick(cfs_rq
, se
);
1060 #define swap(a,b) do { typeof(a) tmp = (a); (a) = (b); (b) = tmp; } while (0)
1063 * Share the fairness runtime between parent and child, thus the
1064 * total amount of pressure for CPU stays equal - new tasks
1065 * get a chance to run but frequent forkers are not allowed to
1066 * monopolize the CPU. Note: the parent runqueue is locked,
1067 * the child is not running yet.
1069 static void task_new_fair(struct rq
*rq
, struct task_struct
*p
)
1071 struct cfs_rq
*cfs_rq
= task_cfs_rq(p
);
1072 struct sched_entity
*se
= &p
->se
, *curr
= cfs_rq
->curr
;
1074 sched_info_queued(p
);
1076 update_curr(cfs_rq
);
1077 place_entity(cfs_rq
, se
, 1);
1080 * The statistical average of wait_runtime is about
1081 * -granularity/2, so initialize the task with that:
1083 if (sched_feat(START_DEBIT
))
1084 se
->wait_runtime
= -(__sched_period(cfs_rq
->nr_running
+1) / 2);
1086 if (sysctl_sched_child_runs_first
&&
1087 curr
->vruntime
< se
->vruntime
) {
1089 dequeue_entity(cfs_rq
, curr
, 0);
1090 swap(curr
->vruntime
, se
->vruntime
);
1091 enqueue_entity(cfs_rq
, curr
, 0);
1094 update_stats_enqueue(cfs_rq
, se
);
1095 __enqueue_entity(cfs_rq
, se
);
1096 resched_task(rq
->curr
);
1099 #ifdef CONFIG_FAIR_GROUP_SCHED
1100 /* Account for a task changing its policy or group.
1102 * This routine is mostly called to set cfs_rq->curr field when a task
1103 * migrates between groups/classes.
1105 static void set_curr_task_fair(struct rq
*rq
)
1107 struct sched_entity
*se
= &rq
->curr
->se
;
1109 for_each_sched_entity(se
)
1110 set_next_entity(cfs_rq_of(se
), se
);
1113 static void set_curr_task_fair(struct rq
*rq
)
1119 * All the scheduling class methods:
1121 struct sched_class fair_sched_class __read_mostly
= {
1122 .enqueue_task
= enqueue_task_fair
,
1123 .dequeue_task
= dequeue_task_fair
,
1124 .yield_task
= yield_task_fair
,
1126 .check_preempt_curr
= check_preempt_wakeup
,
1128 .pick_next_task
= pick_next_task_fair
,
1129 .put_prev_task
= put_prev_task_fair
,
1131 .load_balance
= load_balance_fair
,
1133 .set_curr_task
= set_curr_task_fair
,
1134 .task_tick
= task_tick_fair
,
1135 .task_new
= task_new_fair
,
1138 #ifdef CONFIG_SCHED_DEBUG
1139 static void print_cfs_stats(struct seq_file
*m
, int cpu
)
1141 struct cfs_rq
*cfs_rq
;
1143 for_each_leaf_cfs_rq(cpu_rq(cpu
), cfs_rq
)
1144 print_cfs_rq(m
, cpu
, cfs_rq
);