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 rb_link_node(&se
->run_node
, parent
, link
);
182 rb_insert_color(&se
->run_node
, &cfs_rq
->tasks_timeline
);
185 static void __dequeue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
187 if (cfs_rq
->rb_leftmost
== &se
->run_node
)
188 cfs_rq
->rb_leftmost
= rb_next(&se
->run_node
);
190 rb_erase(&se
->run_node
, &cfs_rq
->tasks_timeline
);
193 static inline struct rb_node
*first_fair(struct cfs_rq
*cfs_rq
)
195 return cfs_rq
->rb_leftmost
;
198 static struct sched_entity
*__pick_next_entity(struct cfs_rq
*cfs_rq
)
200 return rb_entry(first_fair(cfs_rq
), struct sched_entity
, run_node
);
203 static inline struct sched_entity
*__pick_last_entity(struct cfs_rq
*cfs_rq
)
205 struct rb_node
**link
= &cfs_rq
->tasks_timeline
.rb_node
;
206 struct sched_entity
*se
= NULL
;
207 struct rb_node
*parent
;
211 se
= rb_entry(parent
, struct sched_entity
, run_node
);
212 link
= &parent
->rb_right
;
218 /**************************************************************
219 * Scheduling class statistics methods:
222 #ifdef CONFIG_SCHED_DEBUG
223 int sched_nr_latency_handler(struct ctl_table
*table
, int write
,
224 struct file
*filp
, void __user
*buffer
, size_t *lenp
,
227 int ret
= proc_dointvec_minmax(table
, write
, filp
, buffer
, lenp
, ppos
);
232 sched_nr_latency
= DIV_ROUND_UP(sysctl_sched_latency
,
233 sysctl_sched_min_granularity
);
240 * The idea is to set a period in which each task runs once.
242 * When there are too many tasks (sysctl_sched_nr_latency) we have to stretch
243 * this period because otherwise the slices get too small.
245 * p = (nr <= nl) ? l : l*nr/nl
247 static u64
__sched_period(unsigned long nr_running
)
249 u64 period
= sysctl_sched_latency
;
250 unsigned long nr_latency
= sched_nr_latency
;
252 if (unlikely(nr_running
> nr_latency
)) {
253 period
= sysctl_sched_min_granularity
;
254 period
*= nr_running
;
261 * We calculate the wall-time slice from the period by taking a part
262 * proportional to the weight.
266 static u64
sched_slice(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
268 u64 slice
= __sched_period(cfs_rq
->nr_running
);
270 slice
*= se
->load
.weight
;
271 do_div(slice
, cfs_rq
->load
.weight
);
277 * We calculate the vruntime slice.
281 static u64
__sched_vslice(unsigned long rq_weight
, unsigned long nr_running
)
283 u64 vslice
= __sched_period(nr_running
);
285 vslice
*= NICE_0_LOAD
;
286 do_div(vslice
, rq_weight
);
291 static u64
sched_vslice(struct cfs_rq
*cfs_rq
)
293 return __sched_vslice(cfs_rq
->load
.weight
, cfs_rq
->nr_running
);
296 static u64
sched_vslice_add(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
298 return __sched_vslice(cfs_rq
->load
.weight
+ se
->load
.weight
,
299 cfs_rq
->nr_running
+ 1);
303 * Update the current task's runtime statistics. Skip current tasks that
304 * are not in our scheduling class.
307 __update_curr(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
,
308 unsigned long delta_exec
)
310 unsigned long delta_exec_weighted
;
313 schedstat_set(curr
->exec_max
, max((u64
)delta_exec
, curr
->exec_max
));
315 curr
->sum_exec_runtime
+= delta_exec
;
316 schedstat_add(cfs_rq
, exec_clock
, delta_exec
);
317 delta_exec_weighted
= delta_exec
;
318 if (unlikely(curr
->load
.weight
!= NICE_0_LOAD
)) {
319 delta_exec_weighted
= calc_delta_fair(delta_exec_weighted
,
322 curr
->vruntime
+= delta_exec_weighted
;
325 * maintain cfs_rq->min_vruntime to be a monotonic increasing
326 * value tracking the leftmost vruntime in the tree.
328 if (first_fair(cfs_rq
)) {
329 vruntime
= min_vruntime(curr
->vruntime
,
330 __pick_next_entity(cfs_rq
)->vruntime
);
332 vruntime
= curr
->vruntime
;
334 cfs_rq
->min_vruntime
=
335 max_vruntime(cfs_rq
->min_vruntime
, vruntime
);
338 static void update_curr(struct cfs_rq
*cfs_rq
)
340 struct sched_entity
*curr
= cfs_rq
->curr
;
341 u64 now
= rq_of(cfs_rq
)->clock
;
342 unsigned long delta_exec
;
348 * Get the amount of time the current task was running
349 * since the last time we changed load (this cannot
350 * overflow on 32 bits):
352 delta_exec
= (unsigned long)(now
- curr
->exec_start
);
354 __update_curr(cfs_rq
, curr
, delta_exec
);
355 curr
->exec_start
= now
;
357 if (entity_is_task(curr
)) {
358 struct task_struct
*curtask
= task_of(curr
);
360 cpuacct_charge(curtask
, delta_exec
);
365 update_stats_wait_start(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
367 schedstat_set(se
->wait_start
, rq_of(cfs_rq
)->clock
);
371 * Task is being enqueued - update stats:
373 static void update_stats_enqueue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
376 * Are we enqueueing a waiting task? (for current tasks
377 * a dequeue/enqueue event is a NOP)
379 if (se
!= cfs_rq
->curr
)
380 update_stats_wait_start(cfs_rq
, se
);
384 update_stats_wait_end(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
386 schedstat_set(se
->wait_max
, max(se
->wait_max
,
387 rq_of(cfs_rq
)->clock
- se
->wait_start
));
388 schedstat_set(se
->wait_start
, 0);
392 update_stats_dequeue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
395 * Mark the end of the wait period if dequeueing a
398 if (se
!= cfs_rq
->curr
)
399 update_stats_wait_end(cfs_rq
, se
);
403 * We are picking a new current task - update its stats:
406 update_stats_curr_start(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
409 * We are starting a new run period:
411 se
->exec_start
= rq_of(cfs_rq
)->clock
;
414 /**************************************************
415 * Scheduling class queueing methods:
419 account_entity_enqueue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
421 update_load_add(&cfs_rq
->load
, se
->load
.weight
);
422 cfs_rq
->nr_running
++;
427 account_entity_dequeue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
429 update_load_sub(&cfs_rq
->load
, se
->load
.weight
);
430 cfs_rq
->nr_running
--;
434 static void enqueue_sleeper(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
436 #ifdef CONFIG_SCHEDSTATS
437 if (se
->sleep_start
) {
438 u64 delta
= rq_of(cfs_rq
)->clock
- se
->sleep_start
;
439 struct task_struct
*tsk
= task_of(se
);
444 if (unlikely(delta
> se
->sleep_max
))
445 se
->sleep_max
= delta
;
448 se
->sum_sleep_runtime
+= delta
;
450 account_scheduler_latency(tsk
, delta
>> 10, 1);
452 if (se
->block_start
) {
453 u64 delta
= rq_of(cfs_rq
)->clock
- se
->block_start
;
454 struct task_struct
*tsk
= task_of(se
);
459 if (unlikely(delta
> se
->block_max
))
460 se
->block_max
= delta
;
463 se
->sum_sleep_runtime
+= delta
;
466 * Blocking time is in units of nanosecs, so shift by 20 to
467 * get a milliseconds-range estimation of the amount of
468 * time that the task spent sleeping:
470 if (unlikely(prof_on
== SLEEP_PROFILING
)) {
472 profile_hits(SLEEP_PROFILING
, (void *)get_wchan(tsk
),
475 account_scheduler_latency(tsk
, delta
>> 10, 0);
480 static void check_spread(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
482 #ifdef CONFIG_SCHED_DEBUG
483 s64 d
= se
->vruntime
- cfs_rq
->min_vruntime
;
488 if (d
> 3*sysctl_sched_latency
)
489 schedstat_inc(cfs_rq
, nr_spread_over
);
494 place_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int initial
)
498 vruntime
= cfs_rq
->min_vruntime
;
500 if (sched_feat(TREE_AVG
)) {
501 struct sched_entity
*last
= __pick_last_entity(cfs_rq
);
503 vruntime
+= last
->vruntime
;
506 } else if (sched_feat(APPROX_AVG
) && cfs_rq
->nr_running
)
507 vruntime
+= sched_vslice(cfs_rq
)/2;
510 * The 'current' period is already promised to the current tasks,
511 * however the extra weight of the new task will slow them down a
512 * little, place the new task so that it fits in the slot that
513 * stays open at the end.
515 if (initial
&& sched_feat(START_DEBIT
))
516 vruntime
+= sched_vslice_add(cfs_rq
, se
);
519 /* sleeps upto a single latency don't count. */
520 if (sched_feat(NEW_FAIR_SLEEPERS
) && entity_is_task(se
))
521 vruntime
-= sysctl_sched_latency
;
523 /* ensure we never gain time by being placed backwards. */
524 vruntime
= max_vruntime(se
->vruntime
, vruntime
);
527 se
->vruntime
= vruntime
;
531 enqueue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int wakeup
)
534 * Update run-time statistics of the 'current'.
539 place_entity(cfs_rq
, se
, 0);
540 enqueue_sleeper(cfs_rq
, se
);
543 update_stats_enqueue(cfs_rq
, se
);
544 check_spread(cfs_rq
, se
);
545 if (se
!= cfs_rq
->curr
)
546 __enqueue_entity(cfs_rq
, se
);
547 account_entity_enqueue(cfs_rq
, se
);
551 dequeue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int sleep
)
554 * Update run-time statistics of the 'current'.
558 update_stats_dequeue(cfs_rq
, se
);
560 #ifdef CONFIG_SCHEDSTATS
561 if (entity_is_task(se
)) {
562 struct task_struct
*tsk
= task_of(se
);
564 if (tsk
->state
& TASK_INTERRUPTIBLE
)
565 se
->sleep_start
= rq_of(cfs_rq
)->clock
;
566 if (tsk
->state
& TASK_UNINTERRUPTIBLE
)
567 se
->block_start
= rq_of(cfs_rq
)->clock
;
572 if (se
!= cfs_rq
->curr
)
573 __dequeue_entity(cfs_rq
, se
);
574 account_entity_dequeue(cfs_rq
, se
);
578 * Preempt the current task with a newly woken task if needed:
581 check_preempt_tick(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
)
583 unsigned long ideal_runtime
, delta_exec
;
585 ideal_runtime
= sched_slice(cfs_rq
, curr
);
586 delta_exec
= curr
->sum_exec_runtime
- curr
->prev_sum_exec_runtime
;
587 if (delta_exec
> ideal_runtime
)
588 resched_task(rq_of(cfs_rq
)->curr
);
592 set_next_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
594 /* 'current' is not kept within the tree. */
597 * Any task has to be enqueued before it get to execute on
598 * a CPU. So account for the time it spent waiting on the
601 update_stats_wait_end(cfs_rq
, se
);
602 __dequeue_entity(cfs_rq
, se
);
605 update_stats_curr_start(cfs_rq
, se
);
607 #ifdef CONFIG_SCHEDSTATS
609 * Track our maximum slice length, if the CPU's load is at
610 * least twice that of our own weight (i.e. dont track it
611 * when there are only lesser-weight tasks around):
613 if (rq_of(cfs_rq
)->load
.weight
>= 2*se
->load
.weight
) {
614 se
->slice_max
= max(se
->slice_max
,
615 se
->sum_exec_runtime
- se
->prev_sum_exec_runtime
);
618 se
->prev_sum_exec_runtime
= se
->sum_exec_runtime
;
621 static struct sched_entity
*pick_next_entity(struct cfs_rq
*cfs_rq
)
623 struct sched_entity
*se
= NULL
;
625 if (first_fair(cfs_rq
)) {
626 se
= __pick_next_entity(cfs_rq
);
627 set_next_entity(cfs_rq
, se
);
633 static void put_prev_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*prev
)
636 * If still on the runqueue then deactivate_task()
637 * was not called and update_curr() has to be done:
642 check_spread(cfs_rq
, prev
);
644 update_stats_wait_start(cfs_rq
, prev
);
645 /* Put 'current' back into the tree. */
646 __enqueue_entity(cfs_rq
, prev
);
652 entity_tick(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
, int queued
)
655 * Update run-time statistics of the 'current'.
659 #ifdef CONFIG_SCHED_HRTICK
661 * queued ticks are scheduled to match the slice, so don't bother
662 * validating it and just reschedule.
665 return resched_task(rq_of(cfs_rq
)->curr
);
667 * don't let the period tick interfere with the hrtick preemption
669 if (!sched_feat(DOUBLE_TICK
) &&
670 hrtimer_active(&rq_of(cfs_rq
)->hrtick_timer
))
674 if (cfs_rq
->nr_running
> 1 || !sched_feat(WAKEUP_PREEMPT
))
675 check_preempt_tick(cfs_rq
, curr
);
678 /**************************************************
679 * CFS operations on tasks:
682 #ifdef CONFIG_FAIR_GROUP_SCHED
684 /* Walk up scheduling entities hierarchy */
685 #define for_each_sched_entity(se) \
686 for (; se; se = se->parent)
688 static inline struct cfs_rq
*task_cfs_rq(struct task_struct
*p
)
693 /* runqueue on which this entity is (to be) queued */
694 static inline struct cfs_rq
*cfs_rq_of(struct sched_entity
*se
)
699 /* runqueue "owned" by this group */
700 static inline struct cfs_rq
*group_cfs_rq(struct sched_entity
*grp
)
705 /* Given a group's cfs_rq on one cpu, return its corresponding cfs_rq on
706 * another cpu ('this_cpu')
708 static inline struct cfs_rq
*cpu_cfs_rq(struct cfs_rq
*cfs_rq
, int this_cpu
)
710 return cfs_rq
->tg
->cfs_rq
[this_cpu
];
713 /* Iterate thr' all leaf cfs_rq's on a runqueue */
714 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
715 list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
717 /* Do the two (enqueued) entities belong to the same group ? */
719 is_same_group(struct sched_entity
*se
, struct sched_entity
*pse
)
721 if (se
->cfs_rq
== pse
->cfs_rq
)
727 static inline struct sched_entity
*parent_entity(struct sched_entity
*se
)
732 #define GROUP_IMBALANCE_PCT 20
734 #else /* CONFIG_FAIR_GROUP_SCHED */
736 #define for_each_sched_entity(se) \
737 for (; se; se = NULL)
739 static inline struct cfs_rq
*task_cfs_rq(struct task_struct
*p
)
741 return &task_rq(p
)->cfs
;
744 static inline struct cfs_rq
*cfs_rq_of(struct sched_entity
*se
)
746 struct task_struct
*p
= task_of(se
);
747 struct rq
*rq
= task_rq(p
);
752 /* runqueue "owned" by this group */
753 static inline struct cfs_rq
*group_cfs_rq(struct sched_entity
*grp
)
758 static inline struct cfs_rq
*cpu_cfs_rq(struct cfs_rq
*cfs_rq
, int this_cpu
)
760 return &cpu_rq(this_cpu
)->cfs
;
763 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
764 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
767 is_same_group(struct sched_entity
*se
, struct sched_entity
*pse
)
772 static inline struct sched_entity
*parent_entity(struct sched_entity
*se
)
777 #endif /* CONFIG_FAIR_GROUP_SCHED */
779 #ifdef CONFIG_SCHED_HRTICK
780 static void hrtick_start_fair(struct rq
*rq
, struct task_struct
*p
)
782 int requeue
= rq
->curr
== p
;
783 struct sched_entity
*se
= &p
->se
;
784 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
786 WARN_ON(task_rq(p
) != rq
);
788 if (hrtick_enabled(rq
) && cfs_rq
->nr_running
> 1) {
789 u64 slice
= sched_slice(cfs_rq
, se
);
790 u64 ran
= se
->sum_exec_runtime
- se
->prev_sum_exec_runtime
;
791 s64 delta
= slice
- ran
;
800 * Don't schedule slices shorter than 10000ns, that just
801 * doesn't make sense. Rely on vruntime for fairness.
804 delta
= max(10000LL, delta
);
806 hrtick_start(rq
, delta
, requeue
);
811 hrtick_start_fair(struct rq
*rq
, struct task_struct
*p
)
817 * The enqueue_task method is called before nr_running is
818 * increased. Here we update the fair scheduling stats and
819 * then put the task into the rbtree:
821 static void enqueue_task_fair(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
823 struct cfs_rq
*cfs_rq
;
824 struct sched_entity
*se
= &p
->se
,
825 *topse
= NULL
; /* Highest schedulable entity */
828 for_each_sched_entity(se
) {
834 cfs_rq
= cfs_rq_of(se
);
835 enqueue_entity(cfs_rq
, se
, wakeup
);
838 /* Increment cpu load if we just enqueued the first task of a group on
839 * 'rq->cpu'. 'topse' represents the group to which task 'p' belongs
840 * at the highest grouping level.
843 inc_cpu_load(rq
, topse
->load
.weight
);
845 hrtick_start_fair(rq
, rq
->curr
);
849 * The dequeue_task method is called before nr_running is
850 * decreased. We remove the task from the rbtree and
851 * update the fair scheduling stats:
853 static void dequeue_task_fair(struct rq
*rq
, struct task_struct
*p
, int sleep
)
855 struct cfs_rq
*cfs_rq
;
856 struct sched_entity
*se
= &p
->se
,
857 *topse
= NULL
; /* Highest schedulable entity */
860 for_each_sched_entity(se
) {
862 cfs_rq
= cfs_rq_of(se
);
863 dequeue_entity(cfs_rq
, se
, sleep
);
864 /* Don't dequeue parent if it has other entities besides us */
865 if (cfs_rq
->load
.weight
) {
866 if (parent_entity(se
))
872 /* Decrement cpu load if we just dequeued the last task of a group on
873 * 'rq->cpu'. 'topse' represents the group to which task 'p' belongs
874 * at the highest grouping level.
877 dec_cpu_load(rq
, topse
->load
.weight
);
879 hrtick_start_fair(rq
, rq
->curr
);
883 * sched_yield() support is very simple - we dequeue and enqueue.
885 * If compat_yield is turned on then we requeue to the end of the tree.
887 static void yield_task_fair(struct rq
*rq
)
889 struct task_struct
*curr
= rq
->curr
;
890 struct cfs_rq
*cfs_rq
= task_cfs_rq(curr
);
891 struct sched_entity
*rightmost
, *se
= &curr
->se
;
894 * Are we the only task in the tree?
896 if (unlikely(cfs_rq
->nr_running
== 1))
899 if (likely(!sysctl_sched_compat_yield
) && curr
->policy
!= SCHED_BATCH
) {
900 __update_rq_clock(rq
);
902 * Update run-time statistics of the 'current'.
909 * Find the rightmost entry in the rbtree:
911 rightmost
= __pick_last_entity(cfs_rq
);
913 * Already in the rightmost position?
915 if (unlikely(rightmost
->vruntime
< se
->vruntime
))
919 * Minimally necessary key value to be last in the tree:
920 * Upon rescheduling, sched_class::put_prev_task() will place
921 * 'current' within the tree based on its new key value.
923 se
->vruntime
= rightmost
->vruntime
+ 1;
927 * wake_idle() will wake a task on an idle cpu if task->cpu is
928 * not idle and an idle cpu is available. The span of cpus to
929 * search starts with cpus closest then further out as needed,
930 * so we always favor a closer, idle cpu.
932 * Returns the CPU we should wake onto.
934 #if defined(ARCH_HAS_SCHED_WAKE_IDLE)
935 static int wake_idle(int cpu
, struct task_struct
*p
)
938 struct sched_domain
*sd
;
942 * If it is idle, then it is the best cpu to run this task.
944 * This cpu is also the best, if it has more than one task already.
945 * Siblings must be also busy(in most cases) as they didn't already
946 * pickup the extra load from this cpu and hence we need not check
947 * sibling runqueue info. This will avoid the checks and cache miss
948 * penalities associated with that.
950 if (idle_cpu(cpu
) || cpu_rq(cpu
)->nr_running
> 1)
953 for_each_domain(cpu
, sd
) {
954 if (sd
->flags
& SD_WAKE_IDLE
) {
955 cpus_and(tmp
, sd
->span
, p
->cpus_allowed
);
956 for_each_cpu_mask(i
, tmp
) {
958 if (i
!= task_cpu(p
)) {
972 static inline int wake_idle(int cpu
, struct task_struct
*p
)
979 static int select_task_rq_fair(struct task_struct
*p
, int sync
)
983 struct sched_domain
*sd
, *this_sd
= NULL
;
988 this_cpu
= smp_processor_id();
994 for_each_domain(this_cpu
, sd
) {
995 if (cpu_isset(cpu
, sd
->span
)) {
1001 if (unlikely(!cpu_isset(this_cpu
, p
->cpus_allowed
)))
1005 * Check for affine wakeup and passive balancing possibilities.
1008 int idx
= this_sd
->wake_idx
;
1009 unsigned int imbalance
;
1010 unsigned long load
, this_load
;
1012 imbalance
= 100 + (this_sd
->imbalance_pct
- 100) / 2;
1014 load
= source_load(cpu
, idx
);
1015 this_load
= target_load(this_cpu
, idx
);
1017 new_cpu
= this_cpu
; /* Wake to this CPU if we can */
1019 if (this_sd
->flags
& SD_WAKE_AFFINE
) {
1020 unsigned long tl
= this_load
;
1021 unsigned long tl_per_task
;
1024 * Attract cache-cold tasks on sync wakeups:
1026 if (sync
&& !task_hot(p
, rq
->clock
, this_sd
))
1029 schedstat_inc(p
, se
.nr_wakeups_affine_attempts
);
1030 tl_per_task
= cpu_avg_load_per_task(this_cpu
);
1033 * If sync wakeup then subtract the (maximum possible)
1034 * effect of the currently running task from the load
1035 * of the current CPU:
1038 tl
-= current
->se
.load
.weight
;
1041 tl
+ target_load(cpu
, idx
) <= tl_per_task
) ||
1042 100*(tl
+ p
->se
.load
.weight
) <= imbalance
*load
) {
1044 * This domain has SD_WAKE_AFFINE and
1045 * p is cache cold in this domain, and
1046 * there is no bad imbalance.
1048 schedstat_inc(this_sd
, ttwu_move_affine
);
1049 schedstat_inc(p
, se
.nr_wakeups_affine
);
1055 * Start passive balancing when half the imbalance_pct
1058 if (this_sd
->flags
& SD_WAKE_BALANCE
) {
1059 if (imbalance
*this_load
<= 100*load
) {
1060 schedstat_inc(this_sd
, ttwu_move_balance
);
1061 schedstat_inc(p
, se
.nr_wakeups_passive
);
1067 new_cpu
= cpu
; /* Could not wake to this_cpu. Wake to cpu instead */
1069 return wake_idle(new_cpu
, p
);
1071 #endif /* CONFIG_SMP */
1075 * Preempt the current task with a newly woken task if needed:
1077 static void check_preempt_wakeup(struct rq
*rq
, struct task_struct
*p
)
1079 struct task_struct
*curr
= rq
->curr
;
1080 struct cfs_rq
*cfs_rq
= task_cfs_rq(curr
);
1081 struct sched_entity
*se
= &curr
->se
, *pse
= &p
->se
;
1084 if (unlikely(rt_prio(p
->prio
))) {
1085 update_rq_clock(rq
);
1086 update_curr(cfs_rq
);
1091 * Batch tasks do not preempt (their preemption is driven by
1094 if (unlikely(p
->policy
== SCHED_BATCH
))
1097 if (!sched_feat(WAKEUP_PREEMPT
))
1100 while (!is_same_group(se
, pse
)) {
1101 se
= parent_entity(se
);
1102 pse
= parent_entity(pse
);
1105 gran
= sysctl_sched_wakeup_granularity
;
1106 if (unlikely(se
->load
.weight
!= NICE_0_LOAD
))
1107 gran
= calc_delta_fair(gran
, &se
->load
);
1109 if (pse
->vruntime
+ gran
< se
->vruntime
)
1113 static struct task_struct
*pick_next_task_fair(struct rq
*rq
)
1115 struct task_struct
*p
;
1116 struct cfs_rq
*cfs_rq
= &rq
->cfs
;
1117 struct sched_entity
*se
;
1119 if (unlikely(!cfs_rq
->nr_running
))
1123 se
= pick_next_entity(cfs_rq
);
1124 cfs_rq
= group_cfs_rq(se
);
1128 hrtick_start_fair(rq
, p
);
1134 * Account for a descheduled task:
1136 static void put_prev_task_fair(struct rq
*rq
, struct task_struct
*prev
)
1138 struct sched_entity
*se
= &prev
->se
;
1139 struct cfs_rq
*cfs_rq
;
1141 for_each_sched_entity(se
) {
1142 cfs_rq
= cfs_rq_of(se
);
1143 put_prev_entity(cfs_rq
, se
);
1148 /**************************************************
1149 * Fair scheduling class load-balancing methods:
1153 * Load-balancing iterator. Note: while the runqueue stays locked
1154 * during the whole iteration, the current task might be
1155 * dequeued so the iterator has to be dequeue-safe. Here we
1156 * achieve that by always pre-iterating before returning
1159 static struct task_struct
*
1160 __load_balance_iterator(struct cfs_rq
*cfs_rq
, struct rb_node
*curr
)
1162 struct task_struct
*p
;
1167 p
= rb_entry(curr
, struct task_struct
, se
.run_node
);
1168 cfs_rq
->rb_load_balance_curr
= rb_next(curr
);
1173 static struct task_struct
*load_balance_start_fair(void *arg
)
1175 struct cfs_rq
*cfs_rq
= arg
;
1177 return __load_balance_iterator(cfs_rq
, first_fair(cfs_rq
));
1180 static struct task_struct
*load_balance_next_fair(void *arg
)
1182 struct cfs_rq
*cfs_rq
= arg
;
1184 return __load_balance_iterator(cfs_rq
, cfs_rq
->rb_load_balance_curr
);
1187 static unsigned long
1188 load_balance_fair(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1189 unsigned long max_load_move
,
1190 struct sched_domain
*sd
, enum cpu_idle_type idle
,
1191 int *all_pinned
, int *this_best_prio
)
1193 struct cfs_rq
*busy_cfs_rq
;
1194 long rem_load_move
= max_load_move
;
1195 struct rq_iterator cfs_rq_iterator
;
1196 unsigned long load_moved
;
1198 cfs_rq_iterator
.start
= load_balance_start_fair
;
1199 cfs_rq_iterator
.next
= load_balance_next_fair
;
1201 for_each_leaf_cfs_rq(busiest
, busy_cfs_rq
) {
1202 #ifdef CONFIG_FAIR_GROUP_SCHED
1203 struct cfs_rq
*this_cfs_rq
= busy_cfs_rq
->tg
->cfs_rq
[this_cpu
];
1204 unsigned long maxload
, task_load
, group_weight
;
1205 unsigned long thisload
, per_task_load
;
1206 struct sched_entity
*se
= busy_cfs_rq
->tg
->se
[busiest
->cpu
];
1208 task_load
= busy_cfs_rq
->load
.weight
;
1209 group_weight
= se
->load
.weight
;
1212 * 'group_weight' is contributed by tasks of total weight
1213 * 'task_load'. To move 'rem_load_move' worth of weight only,
1214 * we need to move a maximum task load of:
1216 * maxload = (remload / group_weight) * task_load;
1218 maxload
= (rem_load_move
* task_load
) / group_weight
;
1220 if (!maxload
|| !task_load
)
1223 per_task_load
= task_load
/ busy_cfs_rq
->nr_running
;
1225 * balance_tasks will try to forcibly move atleast one task if
1226 * possible (because of SCHED_LOAD_SCALE_FUZZ). Avoid that if
1227 * maxload is less than GROUP_IMBALANCE_FUZZ% the per_task_load.
1229 if (100 * maxload
< GROUP_IMBALANCE_PCT
* per_task_load
)
1232 /* Disable priority-based load balance */
1233 *this_best_prio
= 0;
1234 thisload
= this_cfs_rq
->load
.weight
;
1236 # define maxload rem_load_move
1239 * pass busy_cfs_rq argument into
1240 * load_balance_[start|next]_fair iterators
1242 cfs_rq_iterator
.arg
= busy_cfs_rq
;
1243 load_moved
= balance_tasks(this_rq
, this_cpu
, busiest
,
1244 maxload
, sd
, idle
, all_pinned
,
1248 #ifdef CONFIG_FAIR_GROUP_SCHED
1250 * load_moved holds the task load that was moved. The
1251 * effective (group) weight moved would be:
1252 * load_moved_eff = load_moved/task_load * group_weight;
1254 load_moved
= (group_weight
* load_moved
) / task_load
;
1256 /* Adjust shares on both cpus to reflect load_moved */
1257 group_weight
-= load_moved
;
1258 set_se_shares(se
, group_weight
);
1260 se
= busy_cfs_rq
->tg
->se
[this_cpu
];
1262 group_weight
= load_moved
;
1264 group_weight
= se
->load
.weight
+ load_moved
;
1265 set_se_shares(se
, group_weight
);
1268 rem_load_move
-= load_moved
;
1270 if (rem_load_move
<= 0)
1274 return max_load_move
- rem_load_move
;
1278 move_one_task_fair(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1279 struct sched_domain
*sd
, enum cpu_idle_type idle
)
1281 struct cfs_rq
*busy_cfs_rq
;
1282 struct rq_iterator cfs_rq_iterator
;
1284 cfs_rq_iterator
.start
= load_balance_start_fair
;
1285 cfs_rq_iterator
.next
= load_balance_next_fair
;
1287 for_each_leaf_cfs_rq(busiest
, busy_cfs_rq
) {
1289 * pass busy_cfs_rq argument into
1290 * load_balance_[start|next]_fair iterators
1292 cfs_rq_iterator
.arg
= busy_cfs_rq
;
1293 if (iter_move_one_task(this_rq
, this_cpu
, busiest
, sd
, idle
,
1303 * scheduler tick hitting a task of our scheduling class:
1305 static void task_tick_fair(struct rq
*rq
, struct task_struct
*curr
, int queued
)
1307 struct cfs_rq
*cfs_rq
;
1308 struct sched_entity
*se
= &curr
->se
;
1310 for_each_sched_entity(se
) {
1311 cfs_rq
= cfs_rq_of(se
);
1312 entity_tick(cfs_rq
, se
, queued
);
1316 #define swap(a, b) do { typeof(a) tmp = (a); (a) = (b); (b) = tmp; } while (0)
1319 * Share the fairness runtime between parent and child, thus the
1320 * total amount of pressure for CPU stays equal - new tasks
1321 * get a chance to run but frequent forkers are not allowed to
1322 * monopolize the CPU. Note: the parent runqueue is locked,
1323 * the child is not running yet.
1325 static void task_new_fair(struct rq
*rq
, struct task_struct
*p
)
1327 struct cfs_rq
*cfs_rq
= task_cfs_rq(p
);
1328 struct sched_entity
*se
= &p
->se
, *curr
= cfs_rq
->curr
;
1329 int this_cpu
= smp_processor_id();
1331 sched_info_queued(p
);
1333 update_curr(cfs_rq
);
1334 place_entity(cfs_rq
, se
, 1);
1336 /* 'curr' will be NULL if the child belongs to a different group */
1337 if (sysctl_sched_child_runs_first
&& this_cpu
== task_cpu(p
) &&
1338 curr
&& curr
->vruntime
< se
->vruntime
) {
1340 * Upon rescheduling, sched_class::put_prev_task() will place
1341 * 'current' within the tree based on its new key value.
1343 swap(curr
->vruntime
, se
->vruntime
);
1346 enqueue_task_fair(rq
, p
, 0);
1347 resched_task(rq
->curr
);
1351 * Priority of the task has changed. Check to see if we preempt
1354 static void prio_changed_fair(struct rq
*rq
, struct task_struct
*p
,
1355 int oldprio
, int running
)
1358 * Reschedule if we are currently running on this runqueue and
1359 * our priority decreased, or if we are not currently running on
1360 * this runqueue and our priority is higher than the current's
1363 if (p
->prio
> oldprio
)
1364 resched_task(rq
->curr
);
1366 check_preempt_curr(rq
, p
);
1370 * We switched to the sched_fair class.
1372 static void switched_to_fair(struct rq
*rq
, struct task_struct
*p
,
1376 * We were most likely switched from sched_rt, so
1377 * kick off the schedule if running, otherwise just see
1378 * if we can still preempt the current task.
1381 resched_task(rq
->curr
);
1383 check_preempt_curr(rq
, p
);
1386 /* Account for a task changing its policy or group.
1388 * This routine is mostly called to set cfs_rq->curr field when a task
1389 * migrates between groups/classes.
1391 static void set_curr_task_fair(struct rq
*rq
)
1393 struct sched_entity
*se
= &rq
->curr
->se
;
1395 for_each_sched_entity(se
)
1396 set_next_entity(cfs_rq_of(se
), se
);
1400 * All the scheduling class methods:
1402 static const struct sched_class fair_sched_class
= {
1403 .next
= &idle_sched_class
,
1404 .enqueue_task
= enqueue_task_fair
,
1405 .dequeue_task
= dequeue_task_fair
,
1406 .yield_task
= yield_task_fair
,
1408 .select_task_rq
= select_task_rq_fair
,
1409 #endif /* CONFIG_SMP */
1411 .check_preempt_curr
= check_preempt_wakeup
,
1413 .pick_next_task
= pick_next_task_fair
,
1414 .put_prev_task
= put_prev_task_fair
,
1417 .load_balance
= load_balance_fair
,
1418 .move_one_task
= move_one_task_fair
,
1421 .set_curr_task
= set_curr_task_fair
,
1422 .task_tick
= task_tick_fair
,
1423 .task_new
= task_new_fair
,
1425 .prio_changed
= prio_changed_fair
,
1426 .switched_to
= switched_to_fair
,
1429 #ifdef CONFIG_SCHED_DEBUG
1430 static void print_cfs_stats(struct seq_file
*m
, int cpu
)
1432 struct cfs_rq
*cfs_rq
;
1434 #ifdef CONFIG_FAIR_GROUP_SCHED
1435 print_cfs_rq(m
, cpu
, &cpu_rq(cpu
)->cfs
);
1438 for_each_leaf_cfs_rq(cpu_rq(cpu
), cfs_rq
)
1439 print_cfs_rq(m
, cpu
, cfs_rq
);