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_OTHER 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_wakeup_granularity
= 10000000UL;
74 const_debug
unsigned int sysctl_sched_migration_cost
= 500000UL;
76 /**************************************************************
77 * CFS operations on generic schedulable entities:
80 static inline struct task_struct
*task_of(struct sched_entity
*se
)
82 return container_of(se
, struct task_struct
, se
);
85 #ifdef CONFIG_FAIR_GROUP_SCHED
87 /* cpu runqueue to which this cfs_rq is attached */
88 static inline struct rq
*rq_of(struct cfs_rq
*cfs_rq
)
93 /* An entity is a task if it doesn't "own" a runqueue */
94 #define entity_is_task(se) (!se->my_q)
96 /* Walk up scheduling entities hierarchy */
97 #define for_each_sched_entity(se) \
98 for (; se; se = se->parent)
100 static inline struct cfs_rq
*task_cfs_rq(struct task_struct
*p
)
105 /* runqueue on which this entity is (to be) queued */
106 static inline struct cfs_rq
*cfs_rq_of(struct sched_entity
*se
)
111 /* runqueue "owned" by this group */
112 static inline struct cfs_rq
*group_cfs_rq(struct sched_entity
*grp
)
117 /* Given a group's cfs_rq on one cpu, return its corresponding cfs_rq on
118 * another cpu ('this_cpu')
120 static inline struct cfs_rq
*cpu_cfs_rq(struct cfs_rq
*cfs_rq
, int this_cpu
)
122 return cfs_rq
->tg
->cfs_rq
[this_cpu
];
125 /* Iterate thr' all leaf cfs_rq's on a runqueue */
126 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
127 list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
129 /* Do the two (enqueued) entities belong to the same group ? */
131 is_same_group(struct sched_entity
*se
, struct sched_entity
*pse
)
133 if (se
->cfs_rq
== pse
->cfs_rq
)
139 static inline struct sched_entity
*parent_entity(struct sched_entity
*se
)
144 #else /* CONFIG_FAIR_GROUP_SCHED */
146 static inline struct rq
*rq_of(struct cfs_rq
*cfs_rq
)
148 return container_of(cfs_rq
, struct rq
, cfs
);
151 #define entity_is_task(se) 1
153 #define for_each_sched_entity(se) \
154 for (; se; se = NULL)
156 static inline struct cfs_rq
*task_cfs_rq(struct task_struct
*p
)
158 return &task_rq(p
)->cfs
;
161 static inline struct cfs_rq
*cfs_rq_of(struct sched_entity
*se
)
163 struct task_struct
*p
= task_of(se
);
164 struct rq
*rq
= task_rq(p
);
169 /* runqueue "owned" by this group */
170 static inline struct cfs_rq
*group_cfs_rq(struct sched_entity
*grp
)
175 static inline struct cfs_rq
*cpu_cfs_rq(struct cfs_rq
*cfs_rq
, int this_cpu
)
177 return &cpu_rq(this_cpu
)->cfs
;
180 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
181 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
184 is_same_group(struct sched_entity
*se
, struct sched_entity
*pse
)
189 static inline struct sched_entity
*parent_entity(struct sched_entity
*se
)
194 #endif /* CONFIG_FAIR_GROUP_SCHED */
197 /**************************************************************
198 * Scheduling class tree data structure manipulation methods:
201 static inline u64
max_vruntime(u64 min_vruntime
, u64 vruntime
)
203 s64 delta
= (s64
)(vruntime
- min_vruntime
);
205 min_vruntime
= vruntime
;
210 static inline u64
min_vruntime(u64 min_vruntime
, u64 vruntime
)
212 s64 delta
= (s64
)(vruntime
- min_vruntime
);
214 min_vruntime
= vruntime
;
219 static inline s64
entity_key(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
221 return se
->vruntime
- cfs_rq
->min_vruntime
;
225 * Enqueue an entity into the rb-tree:
227 static void __enqueue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
229 struct rb_node
**link
= &cfs_rq
->tasks_timeline
.rb_node
;
230 struct rb_node
*parent
= NULL
;
231 struct sched_entity
*entry
;
232 s64 key
= entity_key(cfs_rq
, se
);
236 * Find the right place in the rbtree:
240 entry
= rb_entry(parent
, struct sched_entity
, run_node
);
242 * We dont care about collisions. Nodes with
243 * the same key stay together.
245 if (key
< entity_key(cfs_rq
, entry
)) {
246 link
= &parent
->rb_left
;
248 link
= &parent
->rb_right
;
254 * Maintain a cache of leftmost tree entries (it is frequently
258 cfs_rq
->rb_leftmost
= &se
->run_node
;
260 * maintain cfs_rq->min_vruntime to be a monotonic increasing
261 * value tracking the leftmost vruntime in the tree.
263 cfs_rq
->min_vruntime
=
264 max_vruntime(cfs_rq
->min_vruntime
, se
->vruntime
);
267 rb_link_node(&se
->run_node
, parent
, link
);
268 rb_insert_color(&se
->run_node
, &cfs_rq
->tasks_timeline
);
271 static void __dequeue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
273 if (cfs_rq
->rb_leftmost
== &se
->run_node
) {
274 struct rb_node
*next_node
;
275 struct sched_entity
*next
;
277 next_node
= rb_next(&se
->run_node
);
278 cfs_rq
->rb_leftmost
= next_node
;
281 next
= rb_entry(next_node
,
282 struct sched_entity
, run_node
);
283 cfs_rq
->min_vruntime
=
284 max_vruntime(cfs_rq
->min_vruntime
,
289 if (cfs_rq
->next
== se
)
292 rb_erase(&se
->run_node
, &cfs_rq
->tasks_timeline
);
295 static inline struct rb_node
*first_fair(struct cfs_rq
*cfs_rq
)
297 return cfs_rq
->rb_leftmost
;
300 static struct sched_entity
*__pick_next_entity(struct cfs_rq
*cfs_rq
)
302 return rb_entry(first_fair(cfs_rq
), struct sched_entity
, run_node
);
305 static inline struct sched_entity
*__pick_last_entity(struct cfs_rq
*cfs_rq
)
307 struct rb_node
*last
= rb_last(&cfs_rq
->tasks_timeline
);
312 return rb_entry(last
, struct sched_entity
, run_node
);
315 /**************************************************************
316 * Scheduling class statistics methods:
319 #ifdef CONFIG_SCHED_DEBUG
320 int sched_nr_latency_handler(struct ctl_table
*table
, int write
,
321 struct file
*filp
, void __user
*buffer
, size_t *lenp
,
324 int ret
= proc_dointvec_minmax(table
, write
, filp
, buffer
, lenp
, ppos
);
329 sched_nr_latency
= DIV_ROUND_UP(sysctl_sched_latency
,
330 sysctl_sched_min_granularity
);
337 * The idea is to set a period in which each task runs once.
339 * When there are too many tasks (sysctl_sched_nr_latency) we have to stretch
340 * this period because otherwise the slices get too small.
342 * p = (nr <= nl) ? l : l*nr/nl
344 static u64
__sched_period(unsigned long nr_running
)
346 u64 period
= sysctl_sched_latency
;
347 unsigned long nr_latency
= sched_nr_latency
;
349 if (unlikely(nr_running
> nr_latency
)) {
350 period
= sysctl_sched_min_granularity
;
351 period
*= nr_running
;
358 * We calculate the wall-time slice from the period by taking a part
359 * proportional to the weight.
363 static u64
sched_slice(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
365 return calc_delta_mine(__sched_period(cfs_rq
->nr_running
),
366 se
->load
.weight
, &cfs_rq
->load
);
370 * We calculate the vruntime slice.
374 static u64
__sched_vslice(unsigned long rq_weight
, unsigned long nr_running
)
376 u64 vslice
= __sched_period(nr_running
);
378 vslice
*= NICE_0_LOAD
;
379 do_div(vslice
, rq_weight
);
384 static u64
sched_vslice_add(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
386 return __sched_vslice(cfs_rq
->load
.weight
+ se
->load
.weight
,
387 cfs_rq
->nr_running
+ 1);
391 * Update the current task's runtime statistics. Skip current tasks that
392 * are not in our scheduling class.
395 __update_curr(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
,
396 unsigned long delta_exec
)
398 unsigned long delta_exec_weighted
;
400 schedstat_set(curr
->exec_max
, max((u64
)delta_exec
, curr
->exec_max
));
402 curr
->sum_exec_runtime
+= delta_exec
;
403 schedstat_add(cfs_rq
, exec_clock
, delta_exec
);
404 delta_exec_weighted
= delta_exec
;
405 if (unlikely(curr
->load
.weight
!= NICE_0_LOAD
)) {
406 delta_exec_weighted
= calc_delta_fair(delta_exec_weighted
,
409 curr
->vruntime
+= delta_exec_weighted
;
412 static void update_curr(struct cfs_rq
*cfs_rq
)
414 struct sched_entity
*curr
= cfs_rq
->curr
;
415 u64 now
= rq_of(cfs_rq
)->clock
;
416 unsigned long delta_exec
;
422 * Get the amount of time the current task was running
423 * since the last time we changed load (this cannot
424 * overflow on 32 bits):
426 delta_exec
= (unsigned long)(now
- curr
->exec_start
);
428 __update_curr(cfs_rq
, curr
, delta_exec
);
429 curr
->exec_start
= now
;
431 if (entity_is_task(curr
)) {
432 struct task_struct
*curtask
= task_of(curr
);
434 cpuacct_charge(curtask
, delta_exec
);
439 update_stats_wait_start(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
441 schedstat_set(se
->wait_start
, rq_of(cfs_rq
)->clock
);
445 * Task is being enqueued - update stats:
447 static void update_stats_enqueue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
450 * Are we enqueueing a waiting task? (for current tasks
451 * a dequeue/enqueue event is a NOP)
453 if (se
!= cfs_rq
->curr
)
454 update_stats_wait_start(cfs_rq
, se
);
458 update_stats_wait_end(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
460 schedstat_set(se
->wait_max
, max(se
->wait_max
,
461 rq_of(cfs_rq
)->clock
- se
->wait_start
));
462 schedstat_set(se
->wait_count
, se
->wait_count
+ 1);
463 schedstat_set(se
->wait_sum
, se
->wait_sum
+
464 rq_of(cfs_rq
)->clock
- se
->wait_start
);
465 schedstat_set(se
->wait_start
, 0);
469 update_stats_dequeue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
472 * Mark the end of the wait period if dequeueing a
475 if (se
!= cfs_rq
->curr
)
476 update_stats_wait_end(cfs_rq
, se
);
480 * We are picking a new current task - update its stats:
483 update_stats_curr_start(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
486 * We are starting a new run period:
488 se
->exec_start
= rq_of(cfs_rq
)->clock
;
491 /**************************************************
492 * Scheduling class queueing methods:
496 account_entity_enqueue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
498 update_load_add(&cfs_rq
->load
, se
->load
.weight
);
499 cfs_rq
->nr_running
++;
504 account_entity_dequeue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
506 update_load_sub(&cfs_rq
->load
, se
->load
.weight
);
507 cfs_rq
->nr_running
--;
511 static void enqueue_sleeper(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
513 #ifdef CONFIG_SCHEDSTATS
514 if (se
->sleep_start
) {
515 u64 delta
= rq_of(cfs_rq
)->clock
- se
->sleep_start
;
516 struct task_struct
*tsk
= task_of(se
);
521 if (unlikely(delta
> se
->sleep_max
))
522 se
->sleep_max
= delta
;
525 se
->sum_sleep_runtime
+= delta
;
527 account_scheduler_latency(tsk
, delta
>> 10, 1);
529 if (se
->block_start
) {
530 u64 delta
= rq_of(cfs_rq
)->clock
- se
->block_start
;
531 struct task_struct
*tsk
= task_of(se
);
536 if (unlikely(delta
> se
->block_max
))
537 se
->block_max
= delta
;
540 se
->sum_sleep_runtime
+= delta
;
543 * Blocking time is in units of nanosecs, so shift by 20 to
544 * get a milliseconds-range estimation of the amount of
545 * time that the task spent sleeping:
547 if (unlikely(prof_on
== SLEEP_PROFILING
)) {
549 profile_hits(SLEEP_PROFILING
, (void *)get_wchan(tsk
),
552 account_scheduler_latency(tsk
, delta
>> 10, 0);
557 static void check_spread(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
559 #ifdef CONFIG_SCHED_DEBUG
560 s64 d
= se
->vruntime
- cfs_rq
->min_vruntime
;
565 if (d
> 3*sysctl_sched_latency
)
566 schedstat_inc(cfs_rq
, nr_spread_over
);
571 place_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int initial
)
575 if (first_fair(cfs_rq
)) {
576 vruntime
= min_vruntime(cfs_rq
->min_vruntime
,
577 __pick_next_entity(cfs_rq
)->vruntime
);
579 vruntime
= cfs_rq
->min_vruntime
;
582 * The 'current' period is already promised to the current tasks,
583 * however the extra weight of the new task will slow them down a
584 * little, place the new task so that it fits in the slot that
585 * stays open at the end.
587 if (initial
&& sched_feat(START_DEBIT
))
588 vruntime
+= sched_vslice_add(cfs_rq
, se
);
591 /* sleeps upto a single latency don't count. */
592 if (sched_feat(NEW_FAIR_SLEEPERS
)) {
593 if (sched_feat(NORMALIZED_SLEEPER
))
594 vruntime
-= calc_delta_fair(sysctl_sched_latency
,
597 vruntime
-= sysctl_sched_latency
;
600 /* ensure we never gain time by being placed backwards. */
601 vruntime
= max_vruntime(se
->vruntime
, vruntime
);
604 se
->vruntime
= vruntime
;
608 enqueue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int wakeup
)
611 * Update run-time statistics of the 'current'.
616 place_entity(cfs_rq
, se
, 0);
617 enqueue_sleeper(cfs_rq
, se
);
620 update_stats_enqueue(cfs_rq
, se
);
621 check_spread(cfs_rq
, se
);
622 if (se
!= cfs_rq
->curr
)
623 __enqueue_entity(cfs_rq
, se
);
624 account_entity_enqueue(cfs_rq
, se
);
627 static void update_avg(u64
*avg
, u64 sample
)
629 s64 diff
= sample
- *avg
;
633 static void update_avg_stats(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
635 if (!se
->last_wakeup
)
638 update_avg(&se
->avg_overlap
, se
->sum_exec_runtime
- se
->last_wakeup
);
643 dequeue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int sleep
)
646 * Update run-time statistics of the 'current'.
650 update_stats_dequeue(cfs_rq
, se
);
652 update_avg_stats(cfs_rq
, se
);
653 #ifdef CONFIG_SCHEDSTATS
654 if (entity_is_task(se
)) {
655 struct task_struct
*tsk
= task_of(se
);
657 if (tsk
->state
& TASK_INTERRUPTIBLE
)
658 se
->sleep_start
= rq_of(cfs_rq
)->clock
;
659 if (tsk
->state
& TASK_UNINTERRUPTIBLE
)
660 se
->block_start
= rq_of(cfs_rq
)->clock
;
665 if (se
!= cfs_rq
->curr
)
666 __dequeue_entity(cfs_rq
, se
);
667 account_entity_dequeue(cfs_rq
, se
);
671 * Preempt the current task with a newly woken task if needed:
674 check_preempt_tick(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
)
676 unsigned long ideal_runtime
, delta_exec
;
678 ideal_runtime
= sched_slice(cfs_rq
, curr
);
679 delta_exec
= curr
->sum_exec_runtime
- curr
->prev_sum_exec_runtime
;
680 if (delta_exec
> ideal_runtime
)
681 resched_task(rq_of(cfs_rq
)->curr
);
685 set_next_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
687 /* 'current' is not kept within the tree. */
690 * Any task has to be enqueued before it get to execute on
691 * a CPU. So account for the time it spent waiting on the
694 update_stats_wait_end(cfs_rq
, se
);
695 __dequeue_entity(cfs_rq
, se
);
698 update_stats_curr_start(cfs_rq
, se
);
700 #ifdef CONFIG_SCHEDSTATS
702 * Track our maximum slice length, if the CPU's load is at
703 * least twice that of our own weight (i.e. dont track it
704 * when there are only lesser-weight tasks around):
706 if (rq_of(cfs_rq
)->load
.weight
>= 2*se
->load
.weight
) {
707 se
->slice_max
= max(se
->slice_max
,
708 se
->sum_exec_runtime
- se
->prev_sum_exec_runtime
);
711 se
->prev_sum_exec_runtime
= se
->sum_exec_runtime
;
715 wakeup_preempt_entity(struct sched_entity
*curr
, struct sched_entity
*se
);
717 static struct sched_entity
*
718 pick_next(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
723 if (wakeup_preempt_entity(cfs_rq
->next
, se
) != 0)
729 static struct sched_entity
*pick_next_entity(struct cfs_rq
*cfs_rq
)
731 struct sched_entity
*se
= NULL
;
733 if (first_fair(cfs_rq
)) {
734 se
= __pick_next_entity(cfs_rq
);
735 se
= pick_next(cfs_rq
, se
);
736 set_next_entity(cfs_rq
, se
);
742 static void put_prev_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*prev
)
745 * If still on the runqueue then deactivate_task()
746 * was not called and update_curr() has to be done:
751 check_spread(cfs_rq
, prev
);
753 update_stats_wait_start(cfs_rq
, prev
);
754 /* Put 'current' back into the tree. */
755 __enqueue_entity(cfs_rq
, prev
);
761 entity_tick(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
, int queued
)
764 * Update run-time statistics of the 'current'.
768 #ifdef CONFIG_SCHED_HRTICK
770 * queued ticks are scheduled to match the slice, so don't bother
771 * validating it and just reschedule.
774 return resched_task(rq_of(cfs_rq
)->curr
);
776 * don't let the period tick interfere with the hrtick preemption
778 if (!sched_feat(DOUBLE_TICK
) &&
779 hrtimer_active(&rq_of(cfs_rq
)->hrtick_timer
))
783 if (cfs_rq
->nr_running
> 1 || !sched_feat(WAKEUP_PREEMPT
))
784 check_preempt_tick(cfs_rq
, curr
);
787 /**************************************************
788 * CFS operations on tasks:
791 #ifdef CONFIG_SCHED_HRTICK
792 static void hrtick_start_fair(struct rq
*rq
, struct task_struct
*p
)
794 int requeue
= rq
->curr
== p
;
795 struct sched_entity
*se
= &p
->se
;
796 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
798 WARN_ON(task_rq(p
) != rq
);
800 if (hrtick_enabled(rq
) && cfs_rq
->nr_running
> 1) {
801 u64 slice
= sched_slice(cfs_rq
, se
);
802 u64 ran
= se
->sum_exec_runtime
- se
->prev_sum_exec_runtime
;
803 s64 delta
= slice
- ran
;
812 * Don't schedule slices shorter than 10000ns, that just
813 * doesn't make sense. Rely on vruntime for fairness.
816 delta
= max(10000LL, delta
);
818 hrtick_start(rq
, delta
, requeue
);
823 hrtick_start_fair(struct rq
*rq
, struct task_struct
*p
)
829 * The enqueue_task method is called before nr_running is
830 * increased. Here we update the fair scheduling stats and
831 * then put the task into the rbtree:
833 static void enqueue_task_fair(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
835 struct cfs_rq
*cfs_rq
;
836 struct sched_entity
*se
= &p
->se
;
838 for_each_sched_entity(se
) {
841 cfs_rq
= cfs_rq_of(se
);
842 enqueue_entity(cfs_rq
, se
, wakeup
);
846 hrtick_start_fair(rq
, rq
->curr
);
850 * The dequeue_task method is called before nr_running is
851 * decreased. We remove the task from the rbtree and
852 * update the fair scheduling stats:
854 static void dequeue_task_fair(struct rq
*rq
, struct task_struct
*p
, int sleep
)
856 struct cfs_rq
*cfs_rq
;
857 struct sched_entity
*se
= &p
->se
;
859 for_each_sched_entity(se
) {
860 cfs_rq
= cfs_rq_of(se
);
861 dequeue_entity(cfs_rq
, se
, sleep
);
862 /* Don't dequeue parent if it has other entities besides us */
863 if (cfs_rq
->load
.weight
)
868 hrtick_start_fair(rq
, rq
->curr
);
872 * sched_yield() support is very simple - we dequeue and enqueue.
874 * If compat_yield is turned on then we requeue to the end of the tree.
876 static void yield_task_fair(struct rq
*rq
)
878 struct task_struct
*curr
= rq
->curr
;
879 struct cfs_rq
*cfs_rq
= task_cfs_rq(curr
);
880 struct sched_entity
*rightmost
, *se
= &curr
->se
;
883 * Are we the only task in the tree?
885 if (unlikely(cfs_rq
->nr_running
== 1))
888 if (likely(!sysctl_sched_compat_yield
) && curr
->policy
!= SCHED_BATCH
) {
889 __update_rq_clock(rq
);
891 * Update run-time statistics of the 'current'.
898 * Find the rightmost entry in the rbtree:
900 rightmost
= __pick_last_entity(cfs_rq
);
902 * Already in the rightmost position?
904 if (unlikely(!rightmost
|| rightmost
->vruntime
< se
->vruntime
))
908 * Minimally necessary key value to be last in the tree:
909 * Upon rescheduling, sched_class::put_prev_task() will place
910 * 'current' within the tree based on its new key value.
912 se
->vruntime
= rightmost
->vruntime
+ 1;
916 * wake_idle() will wake a task on an idle cpu if task->cpu is
917 * not idle and an idle cpu is available. The span of cpus to
918 * search starts with cpus closest then further out as needed,
919 * so we always favor a closer, idle cpu.
921 * Returns the CPU we should wake onto.
923 #if defined(ARCH_HAS_SCHED_WAKE_IDLE)
924 static int wake_idle(int cpu
, struct task_struct
*p
)
927 struct sched_domain
*sd
;
931 * If it is idle, then it is the best cpu to run this task.
933 * This cpu is also the best, if it has more than one task already.
934 * Siblings must be also busy(in most cases) as they didn't already
935 * pickup the extra load from this cpu and hence we need not check
936 * sibling runqueue info. This will avoid the checks and cache miss
937 * penalities associated with that.
939 if (idle_cpu(cpu
) || cpu_rq(cpu
)->nr_running
> 1)
942 for_each_domain(cpu
, sd
) {
943 if ((sd
->flags
& SD_WAKE_IDLE
)
944 || ((sd
->flags
& SD_WAKE_IDLE_FAR
)
945 && !task_hot(p
, task_rq(p
)->clock
, sd
))) {
946 cpus_and(tmp
, sd
->span
, p
->cpus_allowed
);
947 for_each_cpu_mask(i
, tmp
) {
949 if (i
!= task_cpu(p
)) {
963 static inline int wake_idle(int cpu
, struct task_struct
*p
)
971 static const struct sched_class fair_sched_class
;
974 wake_affine(struct rq
*rq
, struct sched_domain
*this_sd
, struct rq
*this_rq
,
975 struct task_struct
*p
, int prev_cpu
, int this_cpu
, int sync
,
976 int idx
, unsigned long load
, unsigned long this_load
,
977 unsigned int imbalance
)
979 struct task_struct
*curr
= this_rq
->curr
;
980 unsigned long tl
= this_load
;
981 unsigned long tl_per_task
;
983 if (!(this_sd
->flags
& SD_WAKE_AFFINE
))
987 * If the currently running task will sleep within
988 * a reasonable amount of time then attract this newly
991 if (sync
&& curr
->sched_class
== &fair_sched_class
) {
992 if (curr
->se
.avg_overlap
< sysctl_sched_migration_cost
&&
993 p
->se
.avg_overlap
< sysctl_sched_migration_cost
)
997 schedstat_inc(p
, se
.nr_wakeups_affine_attempts
);
998 tl_per_task
= cpu_avg_load_per_task(this_cpu
);
1001 * If sync wakeup then subtract the (maximum possible)
1002 * effect of the currently running task from the load
1003 * of the current CPU:
1006 tl
-= current
->se
.load
.weight
;
1008 if ((tl
<= load
&& tl
+ target_load(prev_cpu
, idx
) <= tl_per_task
) ||
1009 100*(tl
+ p
->se
.load
.weight
) <= imbalance
*load
) {
1011 * This domain has SD_WAKE_AFFINE and
1012 * p is cache cold in this domain, and
1013 * there is no bad imbalance.
1015 schedstat_inc(this_sd
, ttwu_move_affine
);
1016 schedstat_inc(p
, se
.nr_wakeups_affine
);
1023 static int select_task_rq_fair(struct task_struct
*p
, int sync
)
1025 struct sched_domain
*sd
, *this_sd
= NULL
;
1026 int prev_cpu
, this_cpu
, new_cpu
;
1027 unsigned long load
, this_load
;
1028 struct rq
*rq
, *this_rq
;
1029 unsigned int imbalance
;
1032 prev_cpu
= task_cpu(p
);
1034 this_cpu
= smp_processor_id();
1035 this_rq
= cpu_rq(this_cpu
);
1039 * 'this_sd' is the first domain that both
1040 * this_cpu and prev_cpu are present in:
1042 for_each_domain(this_cpu
, sd
) {
1043 if (cpu_isset(prev_cpu
, sd
->span
)) {
1049 if (unlikely(!cpu_isset(this_cpu
, p
->cpus_allowed
)))
1053 * Check for affine wakeup and passive balancing possibilities.
1058 idx
= this_sd
->wake_idx
;
1060 imbalance
= 100 + (this_sd
->imbalance_pct
- 100) / 2;
1062 load
= source_load(prev_cpu
, idx
);
1063 this_load
= target_load(this_cpu
, idx
);
1065 if (wake_affine(rq
, this_sd
, this_rq
, p
, prev_cpu
, this_cpu
, sync
, idx
,
1066 load
, this_load
, imbalance
))
1069 if (prev_cpu
== this_cpu
)
1073 * Start passive balancing when half the imbalance_pct
1076 if (this_sd
->flags
& SD_WAKE_BALANCE
) {
1077 if (imbalance
*this_load
<= 100*load
) {
1078 schedstat_inc(this_sd
, ttwu_move_balance
);
1079 schedstat_inc(p
, se
.nr_wakeups_passive
);
1085 return wake_idle(new_cpu
, p
);
1087 #endif /* CONFIG_SMP */
1089 static unsigned long wakeup_gran(struct sched_entity
*se
)
1091 unsigned long gran
= sysctl_sched_wakeup_granularity
;
1094 * More easily preempt - nice tasks, while not making
1095 * it harder for + nice tasks.
1097 if (unlikely(se
->load
.weight
> NICE_0_LOAD
))
1098 gran
= calc_delta_fair(gran
, &se
->load
);
1104 * Should 'se' preempt 'curr'.
1118 wakeup_preempt_entity(struct sched_entity
*curr
, struct sched_entity
*se
)
1120 s64 gran
, vdiff
= curr
->vruntime
- se
->vruntime
;
1125 gran
= wakeup_gran(curr
);
1132 /* return depth at which a sched entity is present in the hierarchy */
1133 static inline int depth_se(struct sched_entity
*se
)
1137 for_each_sched_entity(se
)
1144 * Preempt the current task with a newly woken task if needed:
1146 static void check_preempt_wakeup(struct rq
*rq
, struct task_struct
*p
)
1148 struct task_struct
*curr
= rq
->curr
;
1149 struct cfs_rq
*cfs_rq
= task_cfs_rq(curr
);
1150 struct sched_entity
*se
= &curr
->se
, *pse
= &p
->se
;
1151 int se_depth
, pse_depth
;
1153 if (unlikely(rt_prio(p
->prio
))) {
1154 update_rq_clock(rq
);
1155 update_curr(cfs_rq
);
1160 se
->last_wakeup
= se
->sum_exec_runtime
;
1161 if (unlikely(se
== pse
))
1164 cfs_rq_of(pse
)->next
= pse
;
1167 * Batch tasks do not preempt (their preemption is driven by
1170 if (unlikely(p
->policy
== SCHED_BATCH
))
1173 if (!sched_feat(WAKEUP_PREEMPT
))
1177 * preemption test can be made between sibling entities who are in the
1178 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
1179 * both tasks until we find their ancestors who are siblings of common
1183 /* First walk up until both entities are at same depth */
1184 se_depth
= depth_se(se
);
1185 pse_depth
= depth_se(pse
);
1187 while (se_depth
> pse_depth
) {
1189 se
= parent_entity(se
);
1192 while (pse_depth
> se_depth
) {
1194 pse
= parent_entity(pse
);
1197 while (!is_same_group(se
, pse
)) {
1198 se
= parent_entity(se
);
1199 pse
= parent_entity(pse
);
1202 if (wakeup_preempt_entity(se
, pse
) == 1)
1206 static struct task_struct
*pick_next_task_fair(struct rq
*rq
)
1208 struct task_struct
*p
;
1209 struct cfs_rq
*cfs_rq
= &rq
->cfs
;
1210 struct sched_entity
*se
;
1212 if (unlikely(!cfs_rq
->nr_running
))
1216 se
= pick_next_entity(cfs_rq
);
1217 cfs_rq
= group_cfs_rq(se
);
1221 hrtick_start_fair(rq
, p
);
1227 * Account for a descheduled task:
1229 static void put_prev_task_fair(struct rq
*rq
, struct task_struct
*prev
)
1231 struct sched_entity
*se
= &prev
->se
;
1232 struct cfs_rq
*cfs_rq
;
1234 for_each_sched_entity(se
) {
1235 cfs_rq
= cfs_rq_of(se
);
1236 put_prev_entity(cfs_rq
, se
);
1241 /**************************************************
1242 * Fair scheduling class load-balancing methods:
1246 * Load-balancing iterator. Note: while the runqueue stays locked
1247 * during the whole iteration, the current task might be
1248 * dequeued so the iterator has to be dequeue-safe. Here we
1249 * achieve that by always pre-iterating before returning
1252 static struct task_struct
*
1253 __load_balance_iterator(struct cfs_rq
*cfs_rq
, struct rb_node
*curr
)
1255 struct task_struct
*p
= NULL
;
1256 struct sched_entity
*se
;
1261 /* Skip over entities that are not tasks */
1263 se
= rb_entry(curr
, struct sched_entity
, run_node
);
1264 curr
= rb_next(curr
);
1265 } while (curr
&& !entity_is_task(se
));
1267 cfs_rq
->rb_load_balance_curr
= curr
;
1269 if (entity_is_task(se
))
1275 static struct task_struct
*load_balance_start_fair(void *arg
)
1277 struct cfs_rq
*cfs_rq
= arg
;
1279 return __load_balance_iterator(cfs_rq
, first_fair(cfs_rq
));
1282 static struct task_struct
*load_balance_next_fair(void *arg
)
1284 struct cfs_rq
*cfs_rq
= arg
;
1286 return __load_balance_iterator(cfs_rq
, cfs_rq
->rb_load_balance_curr
);
1289 #ifdef CONFIG_FAIR_GROUP_SCHED
1290 static int cfs_rq_best_prio(struct cfs_rq
*cfs_rq
)
1292 struct sched_entity
*curr
;
1293 struct task_struct
*p
;
1295 if (!cfs_rq
->nr_running
|| !first_fair(cfs_rq
))
1298 curr
= cfs_rq
->curr
;
1300 curr
= __pick_next_entity(cfs_rq
);
1308 static unsigned long
1309 load_balance_fair(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1310 unsigned long max_load_move
,
1311 struct sched_domain
*sd
, enum cpu_idle_type idle
,
1312 int *all_pinned
, int *this_best_prio
)
1314 struct cfs_rq
*busy_cfs_rq
;
1315 long rem_load_move
= max_load_move
;
1316 struct rq_iterator cfs_rq_iterator
;
1318 cfs_rq_iterator
.start
= load_balance_start_fair
;
1319 cfs_rq_iterator
.next
= load_balance_next_fair
;
1321 for_each_leaf_cfs_rq(busiest
, busy_cfs_rq
) {
1322 #ifdef CONFIG_FAIR_GROUP_SCHED
1323 struct cfs_rq
*this_cfs_rq
;
1325 unsigned long maxload
;
1327 this_cfs_rq
= cpu_cfs_rq(busy_cfs_rq
, this_cpu
);
1329 imbalance
= busy_cfs_rq
->load
.weight
- this_cfs_rq
->load
.weight
;
1330 /* Don't pull if this_cfs_rq has more load than busy_cfs_rq */
1334 /* Don't pull more than imbalance/2 */
1336 maxload
= min(rem_load_move
, imbalance
);
1338 *this_best_prio
= cfs_rq_best_prio(this_cfs_rq
);
1340 # define maxload rem_load_move
1343 * pass busy_cfs_rq argument into
1344 * load_balance_[start|next]_fair iterators
1346 cfs_rq_iterator
.arg
= busy_cfs_rq
;
1347 rem_load_move
-= balance_tasks(this_rq
, this_cpu
, busiest
,
1348 maxload
, sd
, idle
, all_pinned
,
1352 if (rem_load_move
<= 0)
1356 return max_load_move
- rem_load_move
;
1360 move_one_task_fair(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1361 struct sched_domain
*sd
, enum cpu_idle_type idle
)
1363 struct cfs_rq
*busy_cfs_rq
;
1364 struct rq_iterator cfs_rq_iterator
;
1366 cfs_rq_iterator
.start
= load_balance_start_fair
;
1367 cfs_rq_iterator
.next
= load_balance_next_fair
;
1369 for_each_leaf_cfs_rq(busiest
, busy_cfs_rq
) {
1371 * pass busy_cfs_rq argument into
1372 * load_balance_[start|next]_fair iterators
1374 cfs_rq_iterator
.arg
= busy_cfs_rq
;
1375 if (iter_move_one_task(this_rq
, this_cpu
, busiest
, sd
, idle
,
1385 * scheduler tick hitting a task of our scheduling class:
1387 static void task_tick_fair(struct rq
*rq
, struct task_struct
*curr
, int queued
)
1389 struct cfs_rq
*cfs_rq
;
1390 struct sched_entity
*se
= &curr
->se
;
1392 for_each_sched_entity(se
) {
1393 cfs_rq
= cfs_rq_of(se
);
1394 entity_tick(cfs_rq
, se
, queued
);
1398 #define swap(a, b) do { typeof(a) tmp = (a); (a) = (b); (b) = tmp; } while (0)
1401 * Share the fairness runtime between parent and child, thus the
1402 * total amount of pressure for CPU stays equal - new tasks
1403 * get a chance to run but frequent forkers are not allowed to
1404 * monopolize the CPU. Note: the parent runqueue is locked,
1405 * the child is not running yet.
1407 static void task_new_fair(struct rq
*rq
, struct task_struct
*p
)
1409 struct cfs_rq
*cfs_rq
= task_cfs_rq(p
);
1410 struct sched_entity
*se
= &p
->se
, *curr
= cfs_rq
->curr
;
1411 int this_cpu
= smp_processor_id();
1413 sched_info_queued(p
);
1415 update_curr(cfs_rq
);
1416 place_entity(cfs_rq
, se
, 1);
1418 /* 'curr' will be NULL if the child belongs to a different group */
1419 if (sysctl_sched_child_runs_first
&& this_cpu
== task_cpu(p
) &&
1420 curr
&& curr
->vruntime
< se
->vruntime
) {
1422 * Upon rescheduling, sched_class::put_prev_task() will place
1423 * 'current' within the tree based on its new key value.
1425 swap(curr
->vruntime
, se
->vruntime
);
1428 enqueue_task_fair(rq
, p
, 0);
1429 resched_task(rq
->curr
);
1433 * Priority of the task has changed. Check to see if we preempt
1436 static void prio_changed_fair(struct rq
*rq
, struct task_struct
*p
,
1437 int oldprio
, int running
)
1440 * Reschedule if we are currently running on this runqueue and
1441 * our priority decreased, or if we are not currently running on
1442 * this runqueue and our priority is higher than the current's
1445 if (p
->prio
> oldprio
)
1446 resched_task(rq
->curr
);
1448 check_preempt_curr(rq
, p
);
1452 * We switched to the sched_fair class.
1454 static void switched_to_fair(struct rq
*rq
, struct task_struct
*p
,
1458 * We were most likely switched from sched_rt, so
1459 * kick off the schedule if running, otherwise just see
1460 * if we can still preempt the current task.
1463 resched_task(rq
->curr
);
1465 check_preempt_curr(rq
, p
);
1468 /* Account for a task changing its policy or group.
1470 * This routine is mostly called to set cfs_rq->curr field when a task
1471 * migrates between groups/classes.
1473 static void set_curr_task_fair(struct rq
*rq
)
1475 struct sched_entity
*se
= &rq
->curr
->se
;
1477 for_each_sched_entity(se
)
1478 set_next_entity(cfs_rq_of(se
), se
);
1481 #ifdef CONFIG_FAIR_GROUP_SCHED
1482 static void moved_group_fair(struct task_struct
*p
)
1484 struct cfs_rq
*cfs_rq
= task_cfs_rq(p
);
1486 update_curr(cfs_rq
);
1487 place_entity(cfs_rq
, &p
->se
, 1);
1492 * All the scheduling class methods:
1494 static const struct sched_class fair_sched_class
= {
1495 .next
= &idle_sched_class
,
1496 .enqueue_task
= enqueue_task_fair
,
1497 .dequeue_task
= dequeue_task_fair
,
1498 .yield_task
= yield_task_fair
,
1500 .select_task_rq
= select_task_rq_fair
,
1501 #endif /* CONFIG_SMP */
1503 .check_preempt_curr
= check_preempt_wakeup
,
1505 .pick_next_task
= pick_next_task_fair
,
1506 .put_prev_task
= put_prev_task_fair
,
1509 .load_balance
= load_balance_fair
,
1510 .move_one_task
= move_one_task_fair
,
1513 .set_curr_task
= set_curr_task_fair
,
1514 .task_tick
= task_tick_fair
,
1515 .task_new
= task_new_fair
,
1517 .prio_changed
= prio_changed_fair
,
1518 .switched_to
= switched_to_fair
,
1520 #ifdef CONFIG_FAIR_GROUP_SCHED
1521 .moved_group
= moved_group_fair
,
1525 #ifdef CONFIG_SCHED_DEBUG
1526 static void print_cfs_stats(struct seq_file
*m
, int cpu
)
1528 struct cfs_rq
*cfs_rq
;
1531 for_each_leaf_cfs_rq(cpu_rq(cpu
), cfs_rq
)
1532 print_cfs_rq(m
, cpu
, cfs_rq
);