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>
23 #include <linux/latencytop.h>
26 * Targeted preemption latency for CPU-bound tasks:
27 * (default: 20ms * (1 + ilog(ncpus)), units: nanoseconds)
29 * NOTE: this latency value is not the same as the concept of
30 * 'timeslice length' - timeslices in CFS are of variable length
31 * and have no persistent notion like in traditional, time-slice
32 * based scheduling concepts.
34 * (to see the precise effective timeslice length of your workload,
35 * run vmstat and monitor the context-switches (cs) field)
37 unsigned int sysctl_sched_latency
= 20000000ULL;
40 * Minimal preemption granularity for CPU-bound tasks:
41 * (default: 4 msec * (1 + ilog(ncpus)), units: nanoseconds)
43 unsigned int sysctl_sched_min_granularity
= 4000000ULL;
46 * is kept at sysctl_sched_latency / sysctl_sched_min_granularity
48 static unsigned int sched_nr_latency
= 5;
51 * After fork, child runs first. (default) If set to 0 then
52 * parent will (try to) run first.
54 const_debug
unsigned int sysctl_sched_child_runs_first
= 1;
57 * sys_sched_yield() compat mode
59 * This option switches the agressive yield implementation of the
60 * old scheduler back on.
62 unsigned int __read_mostly sysctl_sched_compat_yield
;
65 * SCHED_BATCH wake-up granularity.
66 * (default: 10 msec * (1 + ilog(ncpus)), units: nanoseconds)
68 * This option delays the preemption effects of decoupled workloads
69 * and reduces their over-scheduling. Synchronous workloads will still
70 * have immediate wakeup/sleep latencies.
72 unsigned int sysctl_sched_batch_wakeup_granularity
= 10000000UL;
75 * SCHED_OTHER wake-up granularity.
76 * (default: 10 msec * (1 + ilog(ncpus)), units: nanoseconds)
78 * This option delays the preemption effects of decoupled workloads
79 * and reduces their over-scheduling. Synchronous workloads will still
80 * have immediate wakeup/sleep latencies.
82 unsigned int sysctl_sched_wakeup_granularity
= 10000000UL;
84 const_debug
unsigned int sysctl_sched_migration_cost
= 500000UL;
86 /**************************************************************
87 * CFS operations on generic schedulable entities:
90 #ifdef CONFIG_FAIR_GROUP_SCHED
92 /* cpu runqueue to which this cfs_rq is attached */
93 static inline struct rq
*rq_of(struct cfs_rq
*cfs_rq
)
98 /* An entity is a task if it doesn't "own" a runqueue */
99 #define entity_is_task(se) (!se->my_q)
101 #else /* CONFIG_FAIR_GROUP_SCHED */
103 static inline struct rq
*rq_of(struct cfs_rq
*cfs_rq
)
105 return container_of(cfs_rq
, struct rq
, cfs
);
108 #define entity_is_task(se) 1
110 #endif /* CONFIG_FAIR_GROUP_SCHED */
112 static inline struct task_struct
*task_of(struct sched_entity
*se
)
114 return container_of(se
, struct task_struct
, se
);
118 /**************************************************************
119 * Scheduling class tree data structure manipulation methods:
122 static inline u64
max_vruntime(u64 min_vruntime
, u64 vruntime
)
124 s64 delta
= (s64
)(vruntime
- min_vruntime
);
126 min_vruntime
= vruntime
;
131 static inline u64
min_vruntime(u64 min_vruntime
, u64 vruntime
)
133 s64 delta
= (s64
)(vruntime
- min_vruntime
);
135 min_vruntime
= vruntime
;
140 static inline s64
entity_key(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
142 return se
->vruntime
- cfs_rq
->min_vruntime
;
146 * Enqueue an entity into the rb-tree:
148 static void __enqueue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
150 struct rb_node
**link
= &cfs_rq
->tasks_timeline
.rb_node
;
151 struct rb_node
*parent
= NULL
;
152 struct sched_entity
*entry
;
153 s64 key
= entity_key(cfs_rq
, se
);
157 * Find the right place in the rbtree:
161 entry
= rb_entry(parent
, struct sched_entity
, run_node
);
163 * We dont care about collisions. Nodes with
164 * the same key stay together.
166 if (key
< entity_key(cfs_rq
, entry
)) {
167 link
= &parent
->rb_left
;
169 link
= &parent
->rb_right
;
175 * Maintain a cache of leftmost tree entries (it is frequently
179 cfs_rq
->rb_leftmost
= &se
->run_node
;
181 * maintain cfs_rq->min_vruntime to be a monotonic increasing
182 * value tracking the leftmost vruntime in the tree.
184 cfs_rq
->min_vruntime
=
185 max_vruntime(cfs_rq
->min_vruntime
, se
->vruntime
);
188 rb_link_node(&se
->run_node
, parent
, link
);
189 rb_insert_color(&se
->run_node
, &cfs_rq
->tasks_timeline
);
192 static void __dequeue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
194 if (cfs_rq
->rb_leftmost
== &se
->run_node
) {
195 struct rb_node
*next_node
;
196 struct sched_entity
*next
;
198 next_node
= rb_next(&se
->run_node
);
199 cfs_rq
->rb_leftmost
= next_node
;
202 next
= rb_entry(next_node
,
203 struct sched_entity
, run_node
);
204 cfs_rq
->min_vruntime
=
205 max_vruntime(cfs_rq
->min_vruntime
,
210 rb_erase(&se
->run_node
, &cfs_rq
->tasks_timeline
);
213 static inline struct rb_node
*first_fair(struct cfs_rq
*cfs_rq
)
215 return cfs_rq
->rb_leftmost
;
218 static struct sched_entity
*__pick_next_entity(struct cfs_rq
*cfs_rq
)
220 return rb_entry(first_fair(cfs_rq
), struct sched_entity
, run_node
);
223 static inline struct sched_entity
*__pick_last_entity(struct cfs_rq
*cfs_rq
)
225 struct rb_node
*last
= rb_last(&cfs_rq
->tasks_timeline
);
230 return rb_entry(last
, struct sched_entity
, run_node
);
233 /**************************************************************
234 * Scheduling class statistics methods:
237 #ifdef CONFIG_SCHED_DEBUG
238 int sched_nr_latency_handler(struct ctl_table
*table
, int write
,
239 struct file
*filp
, void __user
*buffer
, size_t *lenp
,
242 int ret
= proc_dointvec_minmax(table
, write
, filp
, buffer
, lenp
, ppos
);
247 sched_nr_latency
= DIV_ROUND_UP(sysctl_sched_latency
,
248 sysctl_sched_min_granularity
);
255 * The idea is to set a period in which each task runs once.
257 * When there are too many tasks (sysctl_sched_nr_latency) we have to stretch
258 * this period because otherwise the slices get too small.
260 * p = (nr <= nl) ? l : l*nr/nl
262 static u64
__sched_period(unsigned long nr_running
)
264 u64 period
= sysctl_sched_latency
;
265 unsigned long nr_latency
= sched_nr_latency
;
267 if (unlikely(nr_running
> nr_latency
)) {
268 period
= sysctl_sched_min_granularity
;
269 period
*= nr_running
;
276 * We calculate the wall-time slice from the period by taking a part
277 * proportional to the weight.
281 static u64
sched_slice(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
283 u64 slice
= __sched_period(cfs_rq
->nr_running
);
285 slice
*= se
->load
.weight
;
286 do_div(slice
, cfs_rq
->load
.weight
);
292 * We calculate the vruntime slice.
296 static u64
__sched_vslice(unsigned long rq_weight
, unsigned long nr_running
)
298 u64 vslice
= __sched_period(nr_running
);
300 vslice
*= NICE_0_LOAD
;
301 do_div(vslice
, rq_weight
);
306 static u64
sched_vslice(struct cfs_rq
*cfs_rq
)
308 return __sched_vslice(cfs_rq
->load
.weight
, cfs_rq
->nr_running
);
311 static u64
sched_vslice_add(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
313 return __sched_vslice(cfs_rq
->load
.weight
+ se
->load
.weight
,
314 cfs_rq
->nr_running
+ 1);
318 * Update the current task's runtime statistics. Skip current tasks that
319 * are not in our scheduling class.
322 __update_curr(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
,
323 unsigned long delta_exec
)
325 unsigned long delta_exec_weighted
;
327 schedstat_set(curr
->exec_max
, max((u64
)delta_exec
, curr
->exec_max
));
329 curr
->sum_exec_runtime
+= delta_exec
;
330 schedstat_add(cfs_rq
, exec_clock
, delta_exec
);
331 delta_exec_weighted
= delta_exec
;
332 if (unlikely(curr
->load
.weight
!= NICE_0_LOAD
)) {
333 delta_exec_weighted
= calc_delta_fair(delta_exec_weighted
,
336 curr
->vruntime
+= delta_exec_weighted
;
339 static void update_curr(struct cfs_rq
*cfs_rq
)
341 struct sched_entity
*curr
= cfs_rq
->curr
;
342 u64 now
= rq_of(cfs_rq
)->clock
;
343 unsigned long delta_exec
;
349 * Get the amount of time the current task was running
350 * since the last time we changed load (this cannot
351 * overflow on 32 bits):
353 delta_exec
= (unsigned long)(now
- curr
->exec_start
);
355 __update_curr(cfs_rq
, curr
, delta_exec
);
356 curr
->exec_start
= now
;
358 if (entity_is_task(curr
)) {
359 struct task_struct
*curtask
= task_of(curr
);
361 cpuacct_charge(curtask
, delta_exec
);
366 update_stats_wait_start(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
368 schedstat_set(se
->wait_start
, rq_of(cfs_rq
)->clock
);
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 update_stats_wait_end(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
387 schedstat_set(se
->wait_max
, max(se
->wait_max
,
388 rq_of(cfs_rq
)->clock
- se
->wait_start
));
389 schedstat_set(se
->wait_count
, se
->wait_count
+ 1);
390 schedstat_set(se
->wait_sum
, se
->wait_sum
+
391 rq_of(cfs_rq
)->clock
- se
->wait_start
);
392 schedstat_set(se
->wait_start
, 0);
396 update_stats_dequeue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
399 * Mark the end of the wait period if dequeueing a
402 if (se
!= cfs_rq
->curr
)
403 update_stats_wait_end(cfs_rq
, se
);
407 * We are picking a new current task - update its stats:
410 update_stats_curr_start(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
413 * We are starting a new run period:
415 se
->exec_start
= rq_of(cfs_rq
)->clock
;
418 /**************************************************
419 * Scheduling class queueing methods:
423 account_entity_enqueue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
425 update_load_add(&cfs_rq
->load
, se
->load
.weight
);
426 cfs_rq
->nr_running
++;
431 account_entity_dequeue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
433 update_load_sub(&cfs_rq
->load
, se
->load
.weight
);
434 cfs_rq
->nr_running
--;
438 static void enqueue_sleeper(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
440 #ifdef CONFIG_SCHEDSTATS
441 if (se
->sleep_start
) {
442 u64 delta
= rq_of(cfs_rq
)->clock
- se
->sleep_start
;
443 struct task_struct
*tsk
= task_of(se
);
448 if (unlikely(delta
> se
->sleep_max
))
449 se
->sleep_max
= delta
;
452 se
->sum_sleep_runtime
+= delta
;
454 account_scheduler_latency(tsk
, delta
>> 10, 1);
456 if (se
->block_start
) {
457 u64 delta
= rq_of(cfs_rq
)->clock
- se
->block_start
;
458 struct task_struct
*tsk
= task_of(se
);
463 if (unlikely(delta
> se
->block_max
))
464 se
->block_max
= delta
;
467 se
->sum_sleep_runtime
+= delta
;
470 * Blocking time is in units of nanosecs, so shift by 20 to
471 * get a milliseconds-range estimation of the amount of
472 * time that the task spent sleeping:
474 if (unlikely(prof_on
== SLEEP_PROFILING
)) {
476 profile_hits(SLEEP_PROFILING
, (void *)get_wchan(tsk
),
479 account_scheduler_latency(tsk
, delta
>> 10, 0);
484 static void check_spread(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
486 #ifdef CONFIG_SCHED_DEBUG
487 s64 d
= se
->vruntime
- cfs_rq
->min_vruntime
;
492 if (d
> 3*sysctl_sched_latency
)
493 schedstat_inc(cfs_rq
, nr_spread_over
);
498 place_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int initial
)
502 if (first_fair(cfs_rq
)) {
503 vruntime
= min_vruntime(cfs_rq
->min_vruntime
,
504 __pick_next_entity(cfs_rq
)->vruntime
);
506 vruntime
= cfs_rq
->min_vruntime
;
508 if (sched_feat(TREE_AVG
)) {
509 struct sched_entity
*last
= __pick_last_entity(cfs_rq
);
511 vruntime
+= last
->vruntime
;
514 } else if (sched_feat(APPROX_AVG
) && cfs_rq
->nr_running
)
515 vruntime
+= sched_vslice(cfs_rq
)/2;
518 * The 'current' period is already promised to the current tasks,
519 * however the extra weight of the new task will slow them down a
520 * little, place the new task so that it fits in the slot that
521 * stays open at the end.
523 if (initial
&& sched_feat(START_DEBIT
))
524 vruntime
+= sched_vslice_add(cfs_rq
, se
);
527 /* sleeps upto a single latency don't count. */
528 if (sched_feat(NEW_FAIR_SLEEPERS
))
529 vruntime
-= sysctl_sched_latency
;
531 /* ensure we never gain time by being placed backwards. */
532 vruntime
= max_vruntime(se
->vruntime
, vruntime
);
535 se
->vruntime
= vruntime
;
539 enqueue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int wakeup
)
542 * Update run-time statistics of the 'current'.
547 place_entity(cfs_rq
, se
, 0);
548 enqueue_sleeper(cfs_rq
, se
);
551 update_stats_enqueue(cfs_rq
, se
);
552 check_spread(cfs_rq
, se
);
553 if (se
!= cfs_rq
->curr
)
554 __enqueue_entity(cfs_rq
, se
);
555 account_entity_enqueue(cfs_rq
, se
);
559 dequeue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int sleep
)
562 * Update run-time statistics of the 'current'.
566 update_stats_dequeue(cfs_rq
, se
);
568 #ifdef CONFIG_SCHEDSTATS
569 if (entity_is_task(se
)) {
570 struct task_struct
*tsk
= task_of(se
);
572 if (tsk
->state
& TASK_INTERRUPTIBLE
)
573 se
->sleep_start
= rq_of(cfs_rq
)->clock
;
574 if (tsk
->state
& TASK_UNINTERRUPTIBLE
)
575 se
->block_start
= rq_of(cfs_rq
)->clock
;
580 if (se
!= cfs_rq
->curr
)
581 __dequeue_entity(cfs_rq
, se
);
582 account_entity_dequeue(cfs_rq
, se
);
586 * Preempt the current task with a newly woken task if needed:
589 check_preempt_tick(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
)
591 unsigned long ideal_runtime
, delta_exec
;
593 ideal_runtime
= sched_slice(cfs_rq
, curr
);
594 delta_exec
= curr
->sum_exec_runtime
- curr
->prev_sum_exec_runtime
;
595 if (delta_exec
> ideal_runtime
)
596 resched_task(rq_of(cfs_rq
)->curr
);
600 set_next_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
602 /* 'current' is not kept within the tree. */
605 * Any task has to be enqueued before it get to execute on
606 * a CPU. So account for the time it spent waiting on the
609 update_stats_wait_end(cfs_rq
, se
);
610 __dequeue_entity(cfs_rq
, se
);
613 update_stats_curr_start(cfs_rq
, se
);
615 #ifdef CONFIG_SCHEDSTATS
617 * Track our maximum slice length, if the CPU's load is at
618 * least twice that of our own weight (i.e. dont track it
619 * when there are only lesser-weight tasks around):
621 if (rq_of(cfs_rq
)->load
.weight
>= 2*se
->load
.weight
) {
622 se
->slice_max
= max(se
->slice_max
,
623 se
->sum_exec_runtime
- se
->prev_sum_exec_runtime
);
626 se
->prev_sum_exec_runtime
= se
->sum_exec_runtime
;
629 static struct sched_entity
*pick_next_entity(struct cfs_rq
*cfs_rq
)
631 struct sched_entity
*se
= NULL
;
633 if (first_fair(cfs_rq
)) {
634 se
= __pick_next_entity(cfs_rq
);
635 set_next_entity(cfs_rq
, se
);
641 static void put_prev_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*prev
)
644 * If still on the runqueue then deactivate_task()
645 * was not called and update_curr() has to be done:
650 check_spread(cfs_rq
, prev
);
652 update_stats_wait_start(cfs_rq
, prev
);
653 /* Put 'current' back into the tree. */
654 __enqueue_entity(cfs_rq
, prev
);
660 entity_tick(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
, int queued
)
663 * Update run-time statistics of the 'current'.
667 #ifdef CONFIG_SCHED_HRTICK
669 * queued ticks are scheduled to match the slice, so don't bother
670 * validating it and just reschedule.
673 return resched_task(rq_of(cfs_rq
)->curr
);
675 * don't let the period tick interfere with the hrtick preemption
677 if (!sched_feat(DOUBLE_TICK
) &&
678 hrtimer_active(&rq_of(cfs_rq
)->hrtick_timer
))
682 if (cfs_rq
->nr_running
> 1 || !sched_feat(WAKEUP_PREEMPT
))
683 check_preempt_tick(cfs_rq
, curr
);
686 /**************************************************
687 * CFS operations on tasks:
690 #ifdef CONFIG_FAIR_GROUP_SCHED
692 /* Walk up scheduling entities hierarchy */
693 #define for_each_sched_entity(se) \
694 for (; se; se = se->parent)
696 static inline struct cfs_rq
*task_cfs_rq(struct task_struct
*p
)
701 /* runqueue on which this entity is (to be) queued */
702 static inline struct cfs_rq
*cfs_rq_of(struct sched_entity
*se
)
707 /* runqueue "owned" by this group */
708 static inline struct cfs_rq
*group_cfs_rq(struct sched_entity
*grp
)
713 /* Given a group's cfs_rq on one cpu, return its corresponding cfs_rq on
714 * another cpu ('this_cpu')
716 static inline struct cfs_rq
*cpu_cfs_rq(struct cfs_rq
*cfs_rq
, int this_cpu
)
718 return cfs_rq
->tg
->cfs_rq
[this_cpu
];
721 /* Iterate thr' all leaf cfs_rq's on a runqueue */
722 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
723 list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
725 /* Do the two (enqueued) entities belong to the same group ? */
727 is_same_group(struct sched_entity
*se
, struct sched_entity
*pse
)
729 if (se
->cfs_rq
== pse
->cfs_rq
)
735 static inline struct sched_entity
*parent_entity(struct sched_entity
*se
)
740 #else /* CONFIG_FAIR_GROUP_SCHED */
742 #define for_each_sched_entity(se) \
743 for (; se; se = NULL)
745 static inline struct cfs_rq
*task_cfs_rq(struct task_struct
*p
)
747 return &task_rq(p
)->cfs
;
750 static inline struct cfs_rq
*cfs_rq_of(struct sched_entity
*se
)
752 struct task_struct
*p
= task_of(se
);
753 struct rq
*rq
= task_rq(p
);
758 /* runqueue "owned" by this group */
759 static inline struct cfs_rq
*group_cfs_rq(struct sched_entity
*grp
)
764 static inline struct cfs_rq
*cpu_cfs_rq(struct cfs_rq
*cfs_rq
, int this_cpu
)
766 return &cpu_rq(this_cpu
)->cfs
;
769 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
770 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
773 is_same_group(struct sched_entity
*se
, struct sched_entity
*pse
)
778 static inline struct sched_entity
*parent_entity(struct sched_entity
*se
)
783 #endif /* CONFIG_FAIR_GROUP_SCHED */
785 #ifdef CONFIG_SCHED_HRTICK
786 static void hrtick_start_fair(struct rq
*rq
, struct task_struct
*p
)
788 int requeue
= rq
->curr
== p
;
789 struct sched_entity
*se
= &p
->se
;
790 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
792 WARN_ON(task_rq(p
) != rq
);
794 if (hrtick_enabled(rq
) && cfs_rq
->nr_running
> 1) {
795 u64 slice
= sched_slice(cfs_rq
, se
);
796 u64 ran
= se
->sum_exec_runtime
- se
->prev_sum_exec_runtime
;
797 s64 delta
= slice
- ran
;
806 * Don't schedule slices shorter than 10000ns, that just
807 * doesn't make sense. Rely on vruntime for fairness.
810 delta
= max(10000LL, delta
);
812 hrtick_start(rq
, delta
, requeue
);
817 hrtick_start_fair(struct rq
*rq
, struct task_struct
*p
)
823 * The enqueue_task method is called before nr_running is
824 * increased. Here we update the fair scheduling stats and
825 * then put the task into the rbtree:
827 static void enqueue_task_fair(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
829 struct cfs_rq
*cfs_rq
;
830 struct sched_entity
*se
= &p
->se
;
832 for_each_sched_entity(se
) {
835 cfs_rq
= cfs_rq_of(se
);
836 enqueue_entity(cfs_rq
, se
, wakeup
);
840 hrtick_start_fair(rq
, rq
->curr
);
844 * The dequeue_task method is called before nr_running is
845 * decreased. We remove the task from the rbtree and
846 * update the fair scheduling stats:
848 static void dequeue_task_fair(struct rq
*rq
, struct task_struct
*p
, int sleep
)
850 struct cfs_rq
*cfs_rq
;
851 struct sched_entity
*se
= &p
->se
;
853 for_each_sched_entity(se
) {
854 cfs_rq
= cfs_rq_of(se
);
855 dequeue_entity(cfs_rq
, se
, sleep
);
856 /* Don't dequeue parent if it has other entities besides us */
857 if (cfs_rq
->load
.weight
)
862 hrtick_start_fair(rq
, rq
->curr
);
866 * sched_yield() support is very simple - we dequeue and enqueue.
868 * If compat_yield is turned on then we requeue to the end of the tree.
870 static void yield_task_fair(struct rq
*rq
)
872 struct task_struct
*curr
= rq
->curr
;
873 struct cfs_rq
*cfs_rq
= task_cfs_rq(curr
);
874 struct sched_entity
*rightmost
, *se
= &curr
->se
;
877 * Are we the only task in the tree?
879 if (unlikely(cfs_rq
->nr_running
== 1))
882 if (likely(!sysctl_sched_compat_yield
) && curr
->policy
!= SCHED_BATCH
) {
883 __update_rq_clock(rq
);
885 * Update run-time statistics of the 'current'.
892 * Find the rightmost entry in the rbtree:
894 rightmost
= __pick_last_entity(cfs_rq
);
896 * Already in the rightmost position?
898 if (unlikely(rightmost
->vruntime
< se
->vruntime
))
902 * Minimally necessary key value to be last in the tree:
903 * Upon rescheduling, sched_class::put_prev_task() will place
904 * 'current' within the tree based on its new key value.
906 se
->vruntime
= rightmost
->vruntime
+ 1;
910 * wake_idle() will wake a task on an idle cpu if task->cpu is
911 * not idle and an idle cpu is available. The span of cpus to
912 * search starts with cpus closest then further out as needed,
913 * so we always favor a closer, idle cpu.
915 * Returns the CPU we should wake onto.
917 #if defined(ARCH_HAS_SCHED_WAKE_IDLE)
918 static int wake_idle(int cpu
, struct task_struct
*p
)
921 struct sched_domain
*sd
;
925 * If it is idle, then it is the best cpu to run this task.
927 * This cpu is also the best, if it has more than one task already.
928 * Siblings must be also busy(in most cases) as they didn't already
929 * pickup the extra load from this cpu and hence we need not check
930 * sibling runqueue info. This will avoid the checks and cache miss
931 * penalities associated with that.
933 if (idle_cpu(cpu
) || cpu_rq(cpu
)->nr_running
> 1)
936 for_each_domain(cpu
, sd
) {
937 if (sd
->flags
& SD_WAKE_IDLE
) {
938 cpus_and(tmp
, sd
->span
, p
->cpus_allowed
);
939 for_each_cpu_mask(i
, tmp
) {
941 if (i
!= task_cpu(p
)) {
955 static inline int wake_idle(int cpu
, struct task_struct
*p
)
962 static int select_task_rq_fair(struct task_struct
*p
, int sync
)
966 struct sched_domain
*sd
, *this_sd
= NULL
;
971 this_cpu
= smp_processor_id();
977 for_each_domain(this_cpu
, sd
) {
978 if (cpu_isset(cpu
, sd
->span
)) {
984 if (unlikely(!cpu_isset(this_cpu
, p
->cpus_allowed
)))
988 * Check for affine wakeup and passive balancing possibilities.
991 int idx
= this_sd
->wake_idx
;
992 unsigned int imbalance
;
993 unsigned long load
, this_load
;
995 imbalance
= 100 + (this_sd
->imbalance_pct
- 100) / 2;
997 load
= source_load(cpu
, idx
);
998 this_load
= target_load(this_cpu
, idx
);
1000 new_cpu
= this_cpu
; /* Wake to this CPU if we can */
1002 if (this_sd
->flags
& SD_WAKE_AFFINE
) {
1003 unsigned long tl
= this_load
;
1004 unsigned long tl_per_task
;
1007 * Attract cache-cold tasks on sync wakeups:
1009 if (sync
&& !task_hot(p
, rq
->clock
, this_sd
))
1012 schedstat_inc(p
, se
.nr_wakeups_affine_attempts
);
1013 tl_per_task
= cpu_avg_load_per_task(this_cpu
);
1016 * If sync wakeup then subtract the (maximum possible)
1017 * effect of the currently running task from the load
1018 * of the current CPU:
1021 tl
-= current
->se
.load
.weight
;
1024 tl
+ target_load(cpu
, idx
) <= tl_per_task
) ||
1025 100*(tl
+ p
->se
.load
.weight
) <= imbalance
*load
) {
1027 * This domain has SD_WAKE_AFFINE and
1028 * p is cache cold in this domain, and
1029 * there is no bad imbalance.
1031 schedstat_inc(this_sd
, ttwu_move_affine
);
1032 schedstat_inc(p
, se
.nr_wakeups_affine
);
1038 * Start passive balancing when half the imbalance_pct
1041 if (this_sd
->flags
& SD_WAKE_BALANCE
) {
1042 if (imbalance
*this_load
<= 100*load
) {
1043 schedstat_inc(this_sd
, ttwu_move_balance
);
1044 schedstat_inc(p
, se
.nr_wakeups_passive
);
1050 new_cpu
= cpu
; /* Could not wake to this_cpu. Wake to cpu instead */
1052 return wake_idle(new_cpu
, p
);
1054 #endif /* CONFIG_SMP */
1058 * Preempt the current task with a newly woken task if needed:
1060 static void check_preempt_wakeup(struct rq
*rq
, struct task_struct
*p
)
1062 struct task_struct
*curr
= rq
->curr
;
1063 struct cfs_rq
*cfs_rq
= task_cfs_rq(curr
);
1064 struct sched_entity
*se
= &curr
->se
, *pse
= &p
->se
;
1067 if (unlikely(rt_prio(p
->prio
))) {
1068 update_rq_clock(rq
);
1069 update_curr(cfs_rq
);
1074 * Batch tasks do not preempt (their preemption is driven by
1077 if (unlikely(p
->policy
== SCHED_BATCH
))
1080 if (!sched_feat(WAKEUP_PREEMPT
))
1083 while (!is_same_group(se
, pse
)) {
1084 se
= parent_entity(se
);
1085 pse
= parent_entity(pse
);
1088 gran
= sysctl_sched_wakeup_granularity
;
1090 * More easily preempt - nice tasks, while not making
1091 * it harder for + nice tasks.
1093 if (unlikely(se
->load
.weight
> NICE_0_LOAD
))
1094 gran
= calc_delta_fair(gran
, &se
->load
);
1096 if (pse
->vruntime
+ gran
< se
->vruntime
)
1100 static struct task_struct
*pick_next_task_fair(struct rq
*rq
)
1102 struct task_struct
*p
;
1103 struct cfs_rq
*cfs_rq
= &rq
->cfs
;
1104 struct sched_entity
*se
;
1106 if (unlikely(!cfs_rq
->nr_running
))
1110 se
= pick_next_entity(cfs_rq
);
1111 cfs_rq
= group_cfs_rq(se
);
1115 hrtick_start_fair(rq
, p
);
1121 * Account for a descheduled task:
1123 static void put_prev_task_fair(struct rq
*rq
, struct task_struct
*prev
)
1125 struct sched_entity
*se
= &prev
->se
;
1126 struct cfs_rq
*cfs_rq
;
1128 for_each_sched_entity(se
) {
1129 cfs_rq
= cfs_rq_of(se
);
1130 put_prev_entity(cfs_rq
, se
);
1135 /**************************************************
1136 * Fair scheduling class load-balancing methods:
1140 * Load-balancing iterator. Note: while the runqueue stays locked
1141 * during the whole iteration, the current task might be
1142 * dequeued so the iterator has to be dequeue-safe. Here we
1143 * achieve that by always pre-iterating before returning
1146 static struct task_struct
*
1147 __load_balance_iterator(struct cfs_rq
*cfs_rq
, struct rb_node
*curr
)
1149 struct task_struct
*p
;
1154 p
= rb_entry(curr
, struct task_struct
, se
.run_node
);
1155 cfs_rq
->rb_load_balance_curr
= rb_next(curr
);
1160 static struct task_struct
*load_balance_start_fair(void *arg
)
1162 struct cfs_rq
*cfs_rq
= arg
;
1164 return __load_balance_iterator(cfs_rq
, first_fair(cfs_rq
));
1167 static struct task_struct
*load_balance_next_fair(void *arg
)
1169 struct cfs_rq
*cfs_rq
= arg
;
1171 return __load_balance_iterator(cfs_rq
, cfs_rq
->rb_load_balance_curr
);
1174 #ifdef CONFIG_FAIR_GROUP_SCHED
1175 static int cfs_rq_best_prio(struct cfs_rq
*cfs_rq
)
1177 struct sched_entity
*curr
;
1178 struct task_struct
*p
;
1180 if (!cfs_rq
->nr_running
|| !first_fair(cfs_rq
))
1183 curr
= cfs_rq
->curr
;
1185 curr
= __pick_next_entity(cfs_rq
);
1193 static unsigned long
1194 load_balance_fair(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1195 unsigned long max_load_move
,
1196 struct sched_domain
*sd
, enum cpu_idle_type idle
,
1197 int *all_pinned
, int *this_best_prio
)
1199 struct cfs_rq
*busy_cfs_rq
;
1200 long rem_load_move
= max_load_move
;
1201 struct rq_iterator cfs_rq_iterator
;
1203 cfs_rq_iterator
.start
= load_balance_start_fair
;
1204 cfs_rq_iterator
.next
= load_balance_next_fair
;
1206 for_each_leaf_cfs_rq(busiest
, busy_cfs_rq
) {
1207 #ifdef CONFIG_FAIR_GROUP_SCHED
1208 struct cfs_rq
*this_cfs_rq
;
1210 unsigned long maxload
;
1212 this_cfs_rq
= cpu_cfs_rq(busy_cfs_rq
, this_cpu
);
1214 imbalance
= busy_cfs_rq
->load
.weight
- this_cfs_rq
->load
.weight
;
1215 /* Don't pull if this_cfs_rq has more load than busy_cfs_rq */
1219 /* Don't pull more than imbalance/2 */
1221 maxload
= min(rem_load_move
, imbalance
);
1223 *this_best_prio
= cfs_rq_best_prio(this_cfs_rq
);
1225 # define maxload rem_load_move
1228 * pass busy_cfs_rq argument into
1229 * load_balance_[start|next]_fair iterators
1231 cfs_rq_iterator
.arg
= busy_cfs_rq
;
1232 rem_load_move
-= balance_tasks(this_rq
, this_cpu
, busiest
,
1233 maxload
, sd
, idle
, all_pinned
,
1237 if (rem_load_move
<= 0)
1241 return max_load_move
- rem_load_move
;
1245 move_one_task_fair(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1246 struct sched_domain
*sd
, enum cpu_idle_type idle
)
1248 struct cfs_rq
*busy_cfs_rq
;
1249 struct rq_iterator cfs_rq_iterator
;
1251 cfs_rq_iterator
.start
= load_balance_start_fair
;
1252 cfs_rq_iterator
.next
= load_balance_next_fair
;
1254 for_each_leaf_cfs_rq(busiest
, busy_cfs_rq
) {
1256 * pass busy_cfs_rq argument into
1257 * load_balance_[start|next]_fair iterators
1259 cfs_rq_iterator
.arg
= busy_cfs_rq
;
1260 if (iter_move_one_task(this_rq
, this_cpu
, busiest
, sd
, idle
,
1270 * scheduler tick hitting a task of our scheduling class:
1272 static void task_tick_fair(struct rq
*rq
, struct task_struct
*curr
, int queued
)
1274 struct cfs_rq
*cfs_rq
;
1275 struct sched_entity
*se
= &curr
->se
;
1277 for_each_sched_entity(se
) {
1278 cfs_rq
= cfs_rq_of(se
);
1279 entity_tick(cfs_rq
, se
, queued
);
1283 #define swap(a, b) do { typeof(a) tmp = (a); (a) = (b); (b) = tmp; } while (0)
1286 * Share the fairness runtime between parent and child, thus the
1287 * total amount of pressure for CPU stays equal - new tasks
1288 * get a chance to run but frequent forkers are not allowed to
1289 * monopolize the CPU. Note: the parent runqueue is locked,
1290 * the child is not running yet.
1292 static void task_new_fair(struct rq
*rq
, struct task_struct
*p
)
1294 struct cfs_rq
*cfs_rq
= task_cfs_rq(p
);
1295 struct sched_entity
*se
= &p
->se
, *curr
= cfs_rq
->curr
;
1296 int this_cpu
= smp_processor_id();
1298 sched_info_queued(p
);
1300 update_curr(cfs_rq
);
1301 place_entity(cfs_rq
, se
, 1);
1303 /* 'curr' will be NULL if the child belongs to a different group */
1304 if (sysctl_sched_child_runs_first
&& this_cpu
== task_cpu(p
) &&
1305 curr
&& curr
->vruntime
< se
->vruntime
) {
1307 * Upon rescheduling, sched_class::put_prev_task() will place
1308 * 'current' within the tree based on its new key value.
1310 swap(curr
->vruntime
, se
->vruntime
);
1313 enqueue_task_fair(rq
, p
, 0);
1314 resched_task(rq
->curr
);
1318 * Priority of the task has changed. Check to see if we preempt
1321 static void prio_changed_fair(struct rq
*rq
, struct task_struct
*p
,
1322 int oldprio
, int running
)
1325 * Reschedule if we are currently running on this runqueue and
1326 * our priority decreased, or if we are not currently running on
1327 * this runqueue and our priority is higher than the current's
1330 if (p
->prio
> oldprio
)
1331 resched_task(rq
->curr
);
1333 check_preempt_curr(rq
, p
);
1337 * We switched to the sched_fair class.
1339 static void switched_to_fair(struct rq
*rq
, struct task_struct
*p
,
1343 * We were most likely switched from sched_rt, so
1344 * kick off the schedule if running, otherwise just see
1345 * if we can still preempt the current task.
1348 resched_task(rq
->curr
);
1350 check_preempt_curr(rq
, p
);
1353 /* Account for a task changing its policy or group.
1355 * This routine is mostly called to set cfs_rq->curr field when a task
1356 * migrates between groups/classes.
1358 static void set_curr_task_fair(struct rq
*rq
)
1360 struct sched_entity
*se
= &rq
->curr
->se
;
1362 for_each_sched_entity(se
)
1363 set_next_entity(cfs_rq_of(se
), se
);
1366 #ifdef CONFIG_FAIR_GROUP_SCHED
1367 static void moved_group_fair(struct task_struct
*p
)
1369 struct cfs_rq
*cfs_rq
= task_cfs_rq(p
);
1371 update_curr(cfs_rq
);
1372 place_entity(cfs_rq
, &p
->se
, 1);
1377 * All the scheduling class methods:
1379 static const struct sched_class fair_sched_class
= {
1380 .next
= &idle_sched_class
,
1381 .enqueue_task
= enqueue_task_fair
,
1382 .dequeue_task
= dequeue_task_fair
,
1383 .yield_task
= yield_task_fair
,
1385 .select_task_rq
= select_task_rq_fair
,
1386 #endif /* CONFIG_SMP */
1388 .check_preempt_curr
= check_preempt_wakeup
,
1390 .pick_next_task
= pick_next_task_fair
,
1391 .put_prev_task
= put_prev_task_fair
,
1394 .load_balance
= load_balance_fair
,
1395 .move_one_task
= move_one_task_fair
,
1398 .set_curr_task
= set_curr_task_fair
,
1399 .task_tick
= task_tick_fair
,
1400 .task_new
= task_new_fair
,
1402 .prio_changed
= prio_changed_fair
,
1403 .switched_to
= switched_to_fair
,
1405 #ifdef CONFIG_FAIR_GROUP_SCHED
1406 .moved_group
= moved_group_fair
,
1410 #ifdef CONFIG_SCHED_DEBUG
1411 static void print_cfs_stats(struct seq_file
*m
, int cpu
)
1413 struct cfs_rq
*cfs_rq
;
1415 #ifdef CONFIG_FAIR_GROUP_SCHED
1416 print_cfs_rq(m
, cpu
, &cpu_rq(cpu
)->cfs
);
1419 for_each_leaf_cfs_rq(cpu_rq(cpu
), cfs_rq
)
1420 print_cfs_rq(m
, cpu
, cfs_rq
);