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 cfs_rq
->min_vruntime
= max(se
->vruntime
,
128 cfs_rq
->min_vruntime
);
133 * Enqueue an entity into the rb-tree:
136 __enqueue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
138 struct rb_node
**link
= &cfs_rq
->tasks_timeline
.rb_node
;
139 struct rb_node
*parent
= NULL
;
140 struct sched_entity
*entry
;
141 s64 key
= se
->fair_key
;
145 * Find the right place in the rbtree:
149 entry
= rb_entry(parent
, struct sched_entity
, run_node
);
151 * We dont care about collisions. Nodes with
152 * the same key stay together.
154 if (key
- entry
->fair_key
< 0) {
155 link
= &parent
->rb_left
;
157 link
= &parent
->rb_right
;
163 * Maintain a cache of leftmost tree entries (it is frequently
167 set_leftmost(cfs_rq
, &se
->run_node
);
169 rb_link_node(&se
->run_node
, parent
, link
);
170 rb_insert_color(&se
->run_node
, &cfs_rq
->tasks_timeline
);
171 update_load_add(&cfs_rq
->load
, se
->load
.weight
);
172 cfs_rq
->nr_running
++;
175 schedstat_add(cfs_rq
, wait_runtime
, se
->wait_runtime
);
179 __dequeue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
181 if (cfs_rq
->rb_leftmost
== &se
->run_node
)
182 set_leftmost(cfs_rq
, rb_next(&se
->run_node
));
184 rb_erase(&se
->run_node
, &cfs_rq
->tasks_timeline
);
185 update_load_sub(&cfs_rq
->load
, se
->load
.weight
);
186 cfs_rq
->nr_running
--;
189 schedstat_add(cfs_rq
, wait_runtime
, -se
->wait_runtime
);
192 static inline struct rb_node
*first_fair(struct cfs_rq
*cfs_rq
)
194 return cfs_rq
->rb_leftmost
;
197 static struct sched_entity
*__pick_next_entity(struct cfs_rq
*cfs_rq
)
199 return rb_entry(first_fair(cfs_rq
), struct sched_entity
, run_node
);
202 static inline struct sched_entity
*__pick_last_entity(struct cfs_rq
*cfs_rq
)
204 struct rb_node
**link
= &cfs_rq
->tasks_timeline
.rb_node
;
205 struct sched_entity
*se
= NULL
;
206 struct rb_node
*parent
;
210 se
= rb_entry(parent
, struct sched_entity
, run_node
);
211 link
= &parent
->rb_right
;
217 /**************************************************************
218 * Scheduling class statistics methods:
221 static u64
__sched_period(unsigned long nr_running
)
223 u64 period
= sysctl_sched_latency
;
224 unsigned long nr_latency
=
225 sysctl_sched_latency
/ sysctl_sched_min_granularity
;
227 if (unlikely(nr_running
> nr_latency
)) {
228 period
*= nr_running
;
229 do_div(period
, nr_latency
);
235 static u64
sched_slice(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
237 u64 period
= __sched_period(cfs_rq
->nr_running
);
239 period
*= se
->load
.weight
;
240 do_div(period
, cfs_rq
->load
.weight
);
246 limit_wait_runtime(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
248 long limit
= sysctl_sched_runtime_limit
;
251 * Niced tasks have the same history dynamic range as
254 if (unlikely(se
->wait_runtime
> limit
)) {
255 se
->wait_runtime
= limit
;
256 schedstat_inc(se
, wait_runtime_overruns
);
257 schedstat_inc(cfs_rq
, wait_runtime_overruns
);
259 if (unlikely(se
->wait_runtime
< -limit
)) {
260 se
->wait_runtime
= -limit
;
261 schedstat_inc(se
, wait_runtime_underruns
);
262 schedstat_inc(cfs_rq
, wait_runtime_underruns
);
267 __add_wait_runtime(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, long delta
)
269 se
->wait_runtime
+= delta
;
270 schedstat_add(se
, sum_wait_runtime
, delta
);
271 limit_wait_runtime(cfs_rq
, se
);
275 add_wait_runtime(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, long delta
)
277 schedstat_add(cfs_rq
, wait_runtime
, -se
->wait_runtime
);
278 __add_wait_runtime(cfs_rq
, se
, delta
);
279 schedstat_add(cfs_rq
, wait_runtime
, se
->wait_runtime
);
283 * Update the current task's runtime statistics. Skip current tasks that
284 * are not in our scheduling class.
287 __update_curr(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
,
288 unsigned long delta_exec
)
290 unsigned long delta
, delta_fair
, delta_mine
, delta_exec_weighted
;
291 struct load_weight
*lw
= &cfs_rq
->load
;
292 unsigned long load
= lw
->weight
;
294 schedstat_set(curr
->exec_max
, max((u64
)delta_exec
, curr
->exec_max
));
296 curr
->sum_exec_runtime
+= delta_exec
;
297 cfs_rq
->exec_clock
+= delta_exec
;
298 delta_exec_weighted
= delta_exec
;
299 if (unlikely(curr
->load
.weight
!= NICE_0_LOAD
)) {
300 delta_exec_weighted
= calc_delta_fair(delta_exec_weighted
,
303 curr
->vruntime
+= delta_exec_weighted
;
305 if (!sched_feat(FAIR_SLEEPERS
))
311 delta_fair
= calc_delta_fair(delta_exec
, lw
);
312 delta_mine
= calc_delta_mine(delta_exec
, curr
->load
.weight
, lw
);
314 if (cfs_rq
->sleeper_bonus
> sysctl_sched_min_granularity
) {
315 delta
= min((u64
)delta_mine
, cfs_rq
->sleeper_bonus
);
316 delta
= min(delta
, (unsigned long)(
317 (long)sysctl_sched_runtime_limit
- curr
->wait_runtime
));
318 cfs_rq
->sleeper_bonus
-= delta
;
322 cfs_rq
->fair_clock
+= delta_fair
;
324 * We executed delta_exec amount of time on the CPU,
325 * but we were only entitled to delta_mine amount of
326 * time during that period (if nr_running == 1 then
327 * the two values are equal)
328 * [Note: delta_mine - delta_exec is negative]:
330 add_wait_runtime(cfs_rq
, curr
, delta_mine
- delta_exec
);
333 static void update_curr(struct cfs_rq
*cfs_rq
)
335 struct sched_entity
*curr
= cfs_rq
->curr
;
336 u64 now
= rq_of(cfs_rq
)->clock
;
337 unsigned long delta_exec
;
343 * Get the amount of time the current task was running
344 * since the last time we changed load (this cannot
345 * overflow on 32 bits):
347 delta_exec
= (unsigned long)(now
- curr
->exec_start
);
349 __update_curr(cfs_rq
, curr
, delta_exec
);
350 curr
->exec_start
= now
;
354 update_stats_wait_start(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
356 se
->wait_start_fair
= cfs_rq
->fair_clock
;
357 schedstat_set(se
->wait_start
, rq_of(cfs_rq
)->clock
);
360 static inline unsigned long
361 calc_weighted(unsigned long delta
, struct sched_entity
*se
)
363 unsigned long weight
= se
->load
.weight
;
365 if (unlikely(weight
!= NICE_0_LOAD
))
366 return (u64
)delta
* se
->load
.weight
>> NICE_0_SHIFT
;
372 * Task is being enqueued - update stats:
374 static void update_stats_enqueue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
377 * Are we enqueueing a waiting task? (for current tasks
378 * a dequeue/enqueue event is a NOP)
380 if (se
!= cfs_rq
->curr
)
381 update_stats_wait_start(cfs_rq
, se
);
385 se
->fair_key
= se
->vruntime
;
389 * Note: must be called with a freshly updated rq->fair_clock.
392 __update_stats_wait_end(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
,
393 unsigned long delta_fair
)
395 schedstat_set(se
->wait_max
, max(se
->wait_max
,
396 rq_of(cfs_rq
)->clock
- se
->wait_start
));
398 delta_fair
= calc_weighted(delta_fair
, se
);
400 add_wait_runtime(cfs_rq
, se
, delta_fair
);
404 update_stats_wait_end(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
406 unsigned long delta_fair
;
408 if (unlikely(!se
->wait_start_fair
))
411 delta_fair
= (unsigned long)min((u64
)(2*sysctl_sched_runtime_limit
),
412 (u64
)(cfs_rq
->fair_clock
- se
->wait_start_fair
));
414 __update_stats_wait_end(cfs_rq
, se
, delta_fair
);
416 se
->wait_start_fair
= 0;
417 schedstat_set(se
->wait_start
, 0);
421 update_stats_dequeue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
425 * Mark the end of the wait period if dequeueing a
428 if (se
!= cfs_rq
->curr
)
429 update_stats_wait_end(cfs_rq
, se
);
433 * We are picking a new current task - update its stats:
436 update_stats_curr_start(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
439 * We are starting a new run period:
441 se
->exec_start
= rq_of(cfs_rq
)->clock
;
445 * We are descheduling a task - update its stats:
448 update_stats_curr_end(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
453 /**************************************************
454 * Scheduling class queueing methods:
457 static void __enqueue_sleeper(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
,
458 unsigned long delta_fair
)
460 unsigned long load
= cfs_rq
->load
.weight
;
464 * Do not boost sleepers if there's too much bonus 'in flight'
467 if (unlikely(cfs_rq
->sleeper_bonus
> sysctl_sched_runtime_limit
))
470 if (sched_feat(SLEEPER_LOAD_AVG
))
471 load
= rq_of(cfs_rq
)->cpu_load
[2];
474 * Fix up delta_fair with the effect of us running
475 * during the whole sleep period:
477 if (sched_feat(SLEEPER_AVG
))
478 delta_fair
= div64_likely32((u64
)delta_fair
* load
,
479 load
+ se
->load
.weight
);
481 delta_fair
= calc_weighted(delta_fair
, se
);
483 prev_runtime
= se
->wait_runtime
;
484 __add_wait_runtime(cfs_rq
, se
, delta_fair
);
485 delta_fair
= se
->wait_runtime
- prev_runtime
;
488 * Track the amount of bonus we've given to sleepers:
490 cfs_rq
->sleeper_bonus
+= delta_fair
;
493 static void enqueue_sleeper(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
495 struct task_struct
*tsk
= task_of(se
);
496 unsigned long delta_fair
;
498 if ((entity_is_task(se
) && tsk
->policy
== SCHED_BATCH
) ||
499 !sched_feat(FAIR_SLEEPERS
))
502 delta_fair
= (unsigned long)min((u64
)(2*sysctl_sched_runtime_limit
),
503 (u64
)(cfs_rq
->fair_clock
- se
->sleep_start_fair
));
505 __enqueue_sleeper(cfs_rq
, se
, delta_fair
);
507 se
->sleep_start_fair
= 0;
509 #ifdef CONFIG_SCHEDSTATS
510 if (se
->sleep_start
) {
511 u64 delta
= rq_of(cfs_rq
)->clock
- se
->sleep_start
;
516 if (unlikely(delta
> se
->sleep_max
))
517 se
->sleep_max
= delta
;
520 se
->sum_sleep_runtime
+= delta
;
522 if (se
->block_start
) {
523 u64 delta
= rq_of(cfs_rq
)->clock
- se
->block_start
;
528 if (unlikely(delta
> se
->block_max
))
529 se
->block_max
= delta
;
532 se
->sum_sleep_runtime
+= delta
;
535 * Blocking time is in units of nanosecs, so shift by 20 to
536 * get a milliseconds-range estimation of the amount of
537 * time that the task spent sleeping:
539 if (unlikely(prof_on
== SLEEP_PROFILING
)) {
540 profile_hits(SLEEP_PROFILING
, (void *)get_wchan(tsk
),
548 place_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int initial
)
550 struct sched_entity
*last
= __pick_last_entity(cfs_rq
);
551 u64 min_runtime
, latency
;
553 min_runtime
= cfs_rq
->min_vruntime
;
555 min_runtime
+= last
->vruntime
;
557 if (initial
&& sched_feat(START_DEBIT
))
558 min_runtime
+= sysctl_sched_latency
/2;
561 if (!initial
&& sched_feat(NEW_FAIR_SLEEPERS
)) {
562 latency
= sysctl_sched_latency
;
563 if (min_runtime
> latency
)
564 min_runtime
-= latency
;
569 se
->vruntime
= max(se
->vruntime
, min_runtime
);
573 enqueue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int wakeup
)
576 * Update the fair clock.
581 place_entity(cfs_rq
, se
, 0);
582 enqueue_sleeper(cfs_rq
, se
);
585 update_stats_enqueue(cfs_rq
, se
);
586 __enqueue_entity(cfs_rq
, se
);
590 dequeue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int sleep
)
592 update_stats_dequeue(cfs_rq
, se
);
594 se
->sleep_start_fair
= cfs_rq
->fair_clock
;
595 #ifdef CONFIG_SCHEDSTATS
596 if (entity_is_task(se
)) {
597 struct task_struct
*tsk
= task_of(se
);
599 if (tsk
->state
& TASK_INTERRUPTIBLE
)
600 se
->sleep_start
= rq_of(cfs_rq
)->clock
;
601 if (tsk
->state
& TASK_UNINTERRUPTIBLE
)
602 se
->block_start
= rq_of(cfs_rq
)->clock
;
606 __dequeue_entity(cfs_rq
, se
);
610 * Preempt the current task with a newly woken task if needed:
613 check_preempt_tick(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
)
615 unsigned long ideal_runtime
, delta_exec
;
617 ideal_runtime
= sched_slice(cfs_rq
, curr
);
618 delta_exec
= curr
->sum_exec_runtime
- curr
->prev_sum_exec_runtime
;
619 if (delta_exec
> ideal_runtime
)
620 resched_task(rq_of(cfs_rq
)->curr
);
624 set_next_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
627 * Any task has to be enqueued before it get to execute on
628 * a CPU. So account for the time it spent waiting on the
629 * runqueue. (note, here we rely on pick_next_task() having
630 * done a put_prev_task_fair() shortly before this, which
631 * updated rq->fair_clock - used by update_stats_wait_end())
633 update_stats_wait_end(cfs_rq
, se
);
634 update_stats_curr_start(cfs_rq
, se
);
636 #ifdef CONFIG_SCHEDSTATS
638 * Track our maximum slice length, if the CPU's load is at
639 * least twice that of our own weight (i.e. dont track it
640 * when there are only lesser-weight tasks around):
642 if (rq_of(cfs_rq
)->ls
.load
.weight
>= 2*se
->load
.weight
) {
643 se
->slice_max
= max(se
->slice_max
,
644 se
->sum_exec_runtime
- se
->prev_sum_exec_runtime
);
647 se
->prev_sum_exec_runtime
= se
->sum_exec_runtime
;
650 static struct sched_entity
*pick_next_entity(struct cfs_rq
*cfs_rq
)
652 struct sched_entity
*se
= __pick_next_entity(cfs_rq
);
654 set_next_entity(cfs_rq
, se
);
659 static void put_prev_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*prev
)
662 * If still on the runqueue then deactivate_task()
663 * was not called and update_curr() has to be done:
668 update_stats_curr_end(cfs_rq
, prev
);
671 update_stats_wait_start(cfs_rq
, prev
);
675 static void entity_tick(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
)
678 * Dequeue and enqueue the task to update its
679 * position within the tree:
681 dequeue_entity(cfs_rq
, curr
, 0);
682 enqueue_entity(cfs_rq
, curr
, 0);
684 if (cfs_rq
->nr_running
> 1)
685 check_preempt_tick(cfs_rq
, curr
);
688 /**************************************************
689 * CFS operations on tasks:
692 #ifdef CONFIG_FAIR_GROUP_SCHED
694 /* Walk up scheduling entities hierarchy */
695 #define for_each_sched_entity(se) \
696 for (; se; se = se->parent)
698 static inline struct cfs_rq
*task_cfs_rq(struct task_struct
*p
)
703 /* runqueue on which this entity is (to be) queued */
704 static inline struct cfs_rq
*cfs_rq_of(struct sched_entity
*se
)
709 /* runqueue "owned" by this group */
710 static inline struct cfs_rq
*group_cfs_rq(struct sched_entity
*grp
)
715 /* Given a group's cfs_rq on one cpu, return its corresponding cfs_rq on
716 * another cpu ('this_cpu')
718 static inline struct cfs_rq
*cpu_cfs_rq(struct cfs_rq
*cfs_rq
, int this_cpu
)
720 /* A later patch will take group into account */
721 return &cpu_rq(this_cpu
)->cfs
;
724 /* Iterate thr' all leaf cfs_rq's on a runqueue */
725 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
726 list_for_each_entry(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
728 /* Do the two (enqueued) tasks belong to the same group ? */
729 static inline int is_same_group(struct task_struct
*curr
, struct task_struct
*p
)
731 if (curr
->se
.cfs_rq
== p
->se
.cfs_rq
)
737 #else /* CONFIG_FAIR_GROUP_SCHED */
739 #define for_each_sched_entity(se) \
740 for (; se; se = NULL)
742 static inline struct cfs_rq
*task_cfs_rq(struct task_struct
*p
)
744 return &task_rq(p
)->cfs
;
747 static inline struct cfs_rq
*cfs_rq_of(struct sched_entity
*se
)
749 struct task_struct
*p
= task_of(se
);
750 struct rq
*rq
= task_rq(p
);
755 /* runqueue "owned" by this group */
756 static inline struct cfs_rq
*group_cfs_rq(struct sched_entity
*grp
)
761 static inline struct cfs_rq
*cpu_cfs_rq(struct cfs_rq
*cfs_rq
, int this_cpu
)
763 return &cpu_rq(this_cpu
)->cfs
;
766 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
767 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
769 static inline int is_same_group(struct task_struct
*curr
, struct task_struct
*p
)
774 #endif /* CONFIG_FAIR_GROUP_SCHED */
777 * The enqueue_task method is called before nr_running is
778 * increased. Here we update the fair scheduling stats and
779 * then put the task into the rbtree:
781 static void enqueue_task_fair(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
783 struct cfs_rq
*cfs_rq
;
784 struct sched_entity
*se
= &p
->se
;
786 for_each_sched_entity(se
) {
789 cfs_rq
= cfs_rq_of(se
);
790 enqueue_entity(cfs_rq
, se
, wakeup
);
795 * The dequeue_task method is called before nr_running is
796 * decreased. We remove the task from the rbtree and
797 * update the fair scheduling stats:
799 static void dequeue_task_fair(struct rq
*rq
, struct task_struct
*p
, int sleep
)
801 struct cfs_rq
*cfs_rq
;
802 struct sched_entity
*se
= &p
->se
;
804 for_each_sched_entity(se
) {
805 cfs_rq
= cfs_rq_of(se
);
806 dequeue_entity(cfs_rq
, se
, sleep
);
807 /* Don't dequeue parent if it has other entities besides us */
808 if (cfs_rq
->load
.weight
)
814 * sched_yield() support is very simple - we dequeue and enqueue.
816 * If compat_yield is turned on then we requeue to the end of the tree.
818 static void yield_task_fair(struct rq
*rq
, struct task_struct
*p
)
820 struct cfs_rq
*cfs_rq
= task_cfs_rq(p
);
821 struct rb_node
**link
= &cfs_rq
->tasks_timeline
.rb_node
;
822 struct sched_entity
*rightmost
, *se
= &p
->se
;
823 struct rb_node
*parent
;
826 * Are we the only task in the tree?
828 if (unlikely(cfs_rq
->nr_running
== 1))
831 if (likely(!sysctl_sched_compat_yield
)) {
832 __update_rq_clock(rq
);
834 * Dequeue and enqueue the task to update its
835 * position within the tree:
837 dequeue_entity(cfs_rq
, &p
->se
, 0);
838 enqueue_entity(cfs_rq
, &p
->se
, 0);
843 * Find the rightmost entry in the rbtree:
847 link
= &parent
->rb_right
;
850 rightmost
= rb_entry(parent
, struct sched_entity
, run_node
);
852 * Already in the rightmost position?
854 if (unlikely(rightmost
== se
))
858 * Minimally necessary key value to be last in the tree:
860 se
->fair_key
= rightmost
->fair_key
+ 1;
862 if (cfs_rq
->rb_leftmost
== &se
->run_node
)
863 cfs_rq
->rb_leftmost
= rb_next(&se
->run_node
);
865 * Relink the task to the rightmost position:
867 rb_erase(&se
->run_node
, &cfs_rq
->tasks_timeline
);
868 rb_link_node(&se
->run_node
, parent
, link
);
869 rb_insert_color(&se
->run_node
, &cfs_rq
->tasks_timeline
);
873 * Preempt the current task with a newly woken task if needed:
875 static void check_preempt_wakeup(struct rq
*rq
, struct task_struct
*p
)
877 struct task_struct
*curr
= rq
->curr
;
878 struct cfs_rq
*cfs_rq
= task_cfs_rq(curr
);
880 if (unlikely(rt_prio(p
->prio
))) {
886 if (is_same_group(curr
, p
)) {
887 s64 delta
= curr
->se
.vruntime
- p
->se
.vruntime
;
889 if (delta
> (s64
)sysctl_sched_wakeup_granularity
)
894 static struct task_struct
*pick_next_task_fair(struct rq
*rq
)
896 struct cfs_rq
*cfs_rq
= &rq
->cfs
;
897 struct sched_entity
*se
;
899 if (unlikely(!cfs_rq
->nr_running
))
903 se
= pick_next_entity(cfs_rq
);
904 cfs_rq
= group_cfs_rq(se
);
911 * Account for a descheduled task:
913 static void put_prev_task_fair(struct rq
*rq
, struct task_struct
*prev
)
915 struct sched_entity
*se
= &prev
->se
;
916 struct cfs_rq
*cfs_rq
;
918 for_each_sched_entity(se
) {
919 cfs_rq
= cfs_rq_of(se
);
920 put_prev_entity(cfs_rq
, se
);
924 /**************************************************
925 * Fair scheduling class load-balancing methods:
929 * Load-balancing iterator. Note: while the runqueue stays locked
930 * during the whole iteration, the current task might be
931 * dequeued so the iterator has to be dequeue-safe. Here we
932 * achieve that by always pre-iterating before returning
935 static inline struct task_struct
*
936 __load_balance_iterator(struct cfs_rq
*cfs_rq
, struct rb_node
*curr
)
938 struct task_struct
*p
;
943 p
= rb_entry(curr
, struct task_struct
, se
.run_node
);
944 cfs_rq
->rb_load_balance_curr
= rb_next(curr
);
949 static struct task_struct
*load_balance_start_fair(void *arg
)
951 struct cfs_rq
*cfs_rq
= arg
;
953 return __load_balance_iterator(cfs_rq
, first_fair(cfs_rq
));
956 static struct task_struct
*load_balance_next_fair(void *arg
)
958 struct cfs_rq
*cfs_rq
= arg
;
960 return __load_balance_iterator(cfs_rq
, cfs_rq
->rb_load_balance_curr
);
963 #ifdef CONFIG_FAIR_GROUP_SCHED
964 static int cfs_rq_best_prio(struct cfs_rq
*cfs_rq
)
966 struct sched_entity
*curr
;
967 struct task_struct
*p
;
969 if (!cfs_rq
->nr_running
)
972 curr
= __pick_next_entity(cfs_rq
);
980 load_balance_fair(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
981 unsigned long max_nr_move
, unsigned long max_load_move
,
982 struct sched_domain
*sd
, enum cpu_idle_type idle
,
983 int *all_pinned
, int *this_best_prio
)
985 struct cfs_rq
*busy_cfs_rq
;
986 unsigned long load_moved
, total_nr_moved
= 0, nr_moved
;
987 long rem_load_move
= max_load_move
;
988 struct rq_iterator cfs_rq_iterator
;
990 cfs_rq_iterator
.start
= load_balance_start_fair
;
991 cfs_rq_iterator
.next
= load_balance_next_fair
;
993 for_each_leaf_cfs_rq(busiest
, busy_cfs_rq
) {
994 #ifdef CONFIG_FAIR_GROUP_SCHED
995 struct cfs_rq
*this_cfs_rq
;
997 unsigned long maxload
;
999 this_cfs_rq
= cpu_cfs_rq(busy_cfs_rq
, this_cpu
);
1001 imbalance
= busy_cfs_rq
->load
.weight
- this_cfs_rq
->load
.weight
;
1002 /* Don't pull if this_cfs_rq has more load than busy_cfs_rq */
1006 /* Don't pull more than imbalance/2 */
1008 maxload
= min(rem_load_move
, imbalance
);
1010 *this_best_prio
= cfs_rq_best_prio(this_cfs_rq
);
1012 # define maxload rem_load_move
1014 /* pass busy_cfs_rq argument into
1015 * load_balance_[start|next]_fair iterators
1017 cfs_rq_iterator
.arg
= busy_cfs_rq
;
1018 nr_moved
= balance_tasks(this_rq
, this_cpu
, busiest
,
1019 max_nr_move
, maxload
, sd
, idle
, all_pinned
,
1020 &load_moved
, this_best_prio
, &cfs_rq_iterator
);
1022 total_nr_moved
+= nr_moved
;
1023 max_nr_move
-= nr_moved
;
1024 rem_load_move
-= load_moved
;
1026 if (max_nr_move
<= 0 || rem_load_move
<= 0)
1030 return max_load_move
- rem_load_move
;
1034 * scheduler tick hitting a task of our scheduling class:
1036 static void task_tick_fair(struct rq
*rq
, struct task_struct
*curr
)
1038 struct cfs_rq
*cfs_rq
;
1039 struct sched_entity
*se
= &curr
->se
;
1041 for_each_sched_entity(se
) {
1042 cfs_rq
= cfs_rq_of(se
);
1043 entity_tick(cfs_rq
, se
);
1047 #define swap(a,b) do { typeof(a) tmp = (a); (a) = (b); (b) = tmp; } while (0)
1050 * Share the fairness runtime between parent and child, thus the
1051 * total amount of pressure for CPU stays equal - new tasks
1052 * get a chance to run but frequent forkers are not allowed to
1053 * monopolize the CPU. Note: the parent runqueue is locked,
1054 * the child is not running yet.
1056 static void task_new_fair(struct rq
*rq
, struct task_struct
*p
)
1058 struct cfs_rq
*cfs_rq
= task_cfs_rq(p
);
1059 struct sched_entity
*se
= &p
->se
, *curr
= cfs_rq
->curr
;
1061 sched_info_queued(p
);
1063 update_curr(cfs_rq
);
1064 place_entity(cfs_rq
, se
, 1);
1067 * The statistical average of wait_runtime is about
1068 * -granularity/2, so initialize the task with that:
1070 if (sched_feat(START_DEBIT
))
1071 se
->wait_runtime
= -(__sched_period(cfs_rq
->nr_running
+1) / 2);
1073 if (sysctl_sched_child_runs_first
&&
1074 curr
->vruntime
< se
->vruntime
) {
1076 dequeue_entity(cfs_rq
, curr
, 0);
1077 swap(curr
->vruntime
, se
->vruntime
);
1078 enqueue_entity(cfs_rq
, curr
, 0);
1081 update_stats_enqueue(cfs_rq
, se
);
1082 __enqueue_entity(cfs_rq
, se
);
1083 resched_task(rq
->curr
);
1086 #ifdef CONFIG_FAIR_GROUP_SCHED
1087 /* Account for a task changing its policy or group.
1089 * This routine is mostly called to set cfs_rq->curr field when a task
1090 * migrates between groups/classes.
1092 static void set_curr_task_fair(struct rq
*rq
)
1094 struct sched_entity
*se
= &rq
->curr
->se
;
1096 for_each_sched_entity(se
)
1097 set_next_entity(cfs_rq_of(se
), se
);
1100 static void set_curr_task_fair(struct rq
*rq
)
1106 * All the scheduling class methods:
1108 struct sched_class fair_sched_class __read_mostly
= {
1109 .enqueue_task
= enqueue_task_fair
,
1110 .dequeue_task
= dequeue_task_fair
,
1111 .yield_task
= yield_task_fair
,
1113 .check_preempt_curr
= check_preempt_wakeup
,
1115 .pick_next_task
= pick_next_task_fair
,
1116 .put_prev_task
= put_prev_task_fair
,
1118 .load_balance
= load_balance_fair
,
1120 .set_curr_task
= set_curr_task_fair
,
1121 .task_tick
= task_tick_fair
,
1122 .task_new
= task_new_fair
,
1125 #ifdef CONFIG_SCHED_DEBUG
1126 static void print_cfs_stats(struct seq_file
*m
, int cpu
)
1128 struct cfs_rq
*cfs_rq
;
1130 for_each_leaf_cfs_rq(cpu_rq(cpu
), cfs_rq
)
1131 print_cfs_rq(m
, cpu
, cfs_rq
);