2 * Completely Fair Scheduling (CFS) Class (SCHED_NORMAL/SCHED_BATCH)
4 * Copyright (C) 2007 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
6 * Interactivity improvements by Mike Galbraith
7 * (C) 2007 Mike Galbraith <efault@gmx.de>
9 * Various enhancements by Dmitry Adamushko.
10 * (C) 2007 Dmitry Adamushko <dmitry.adamushko@gmail.com>
12 * Group scheduling enhancements by Srivatsa Vaddagiri
13 * Copyright IBM Corporation, 2007
14 * Author: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
16 * Scaled math optimizations by Thomas Gleixner
17 * Copyright (C) 2007, Thomas Gleixner <tglx@linutronix.de>
19 * Adaptive scheduling granularity, math enhancements by Peter Zijlstra
20 * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
24 * Targeted preemption latency for CPU-bound tasks:
25 * (default: 20ms * (1 + ilog(ncpus)), units: nanoseconds)
27 * NOTE: this latency value is not the same as the concept of
28 * 'timeslice length' - timeslices in CFS are of variable length
29 * and have no persistent notion like in traditional, time-slice
30 * based scheduling concepts.
32 * (to see the precise effective timeslice length of your workload,
33 * run vmstat and monitor the context-switches (cs) field)
35 unsigned int sysctl_sched_latency
= 20000000ULL;
38 * Minimal preemption granularity for CPU-bound tasks:
39 * (default: 4 msec * (1 + ilog(ncpus)), units: nanoseconds)
41 unsigned int sysctl_sched_min_granularity
= 4000000ULL;
44 * is kept at sysctl_sched_latency / sysctl_sched_min_granularity
46 static unsigned int sched_nr_latency
= 5;
49 * After fork, child runs first. (default) If set to 0 then
50 * parent will (try to) run first.
52 const_debug
unsigned int sysctl_sched_child_runs_first
= 1;
55 * sys_sched_yield() compat mode
57 * This option switches the agressive yield implementation of the
58 * old scheduler back on.
60 unsigned int __read_mostly sysctl_sched_compat_yield
;
63 * SCHED_BATCH wake-up granularity.
64 * (default: 10 msec * (1 + ilog(ncpus)), units: nanoseconds)
66 * This option delays the preemption effects of decoupled workloads
67 * and reduces their over-scheduling. Synchronous workloads will still
68 * have immediate wakeup/sleep latencies.
70 unsigned int sysctl_sched_batch_wakeup_granularity
= 10000000UL;
73 * SCHED_OTHER wake-up granularity.
74 * (default: 10 msec * (1 + ilog(ncpus)), units: nanoseconds)
76 * This option delays the preemption effects of decoupled workloads
77 * and reduces their over-scheduling. Synchronous workloads will still
78 * have immediate wakeup/sleep latencies.
80 unsigned int sysctl_sched_wakeup_granularity
= 10000000UL;
82 const_debug
unsigned int sysctl_sched_migration_cost
= 500000UL;
84 /**************************************************************
85 * CFS operations on generic schedulable entities:
88 #ifdef CONFIG_FAIR_GROUP_SCHED
90 /* cpu runqueue to which this cfs_rq is attached */
91 static inline struct rq
*rq_of(struct cfs_rq
*cfs_rq
)
96 /* An entity is a task if it doesn't "own" a runqueue */
97 #define entity_is_task(se) (!se->my_q)
99 #else /* CONFIG_FAIR_GROUP_SCHED */
101 static inline struct rq
*rq_of(struct cfs_rq
*cfs_rq
)
103 return container_of(cfs_rq
, struct rq
, cfs
);
106 #define entity_is_task(se) 1
108 #endif /* CONFIG_FAIR_GROUP_SCHED */
110 static inline struct task_struct
*task_of(struct sched_entity
*se
)
112 return container_of(se
, struct task_struct
, se
);
116 /**************************************************************
117 * Scheduling class tree data structure manipulation methods:
120 static inline u64
max_vruntime(u64 min_vruntime
, u64 vruntime
)
122 s64 delta
= (s64
)(vruntime
- min_vruntime
);
124 min_vruntime
= vruntime
;
129 static inline u64
min_vruntime(u64 min_vruntime
, u64 vruntime
)
131 s64 delta
= (s64
)(vruntime
- min_vruntime
);
133 min_vruntime
= vruntime
;
138 static inline s64
entity_key(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
140 return se
->vruntime
- cfs_rq
->min_vruntime
;
144 * Enqueue an entity into the rb-tree:
146 static void __enqueue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
148 struct rb_node
**link
= &cfs_rq
->tasks_timeline
.rb_node
;
149 struct rb_node
*parent
= NULL
;
150 struct sched_entity
*entry
;
151 s64 key
= entity_key(cfs_rq
, se
);
155 * Find the right place in the rbtree:
159 entry
= rb_entry(parent
, struct sched_entity
, run_node
);
161 * We dont care about collisions. Nodes with
162 * the same key stay together.
164 if (key
< entity_key(cfs_rq
, entry
)) {
165 link
= &parent
->rb_left
;
167 link
= &parent
->rb_right
;
173 * Maintain a cache of leftmost tree entries (it is frequently
177 cfs_rq
->rb_leftmost
= &se
->run_node
;
179 rb_link_node(&se
->run_node
, parent
, link
);
180 rb_insert_color(&se
->run_node
, &cfs_rq
->tasks_timeline
);
183 static void __dequeue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
185 if (cfs_rq
->rb_leftmost
== &se
->run_node
)
186 cfs_rq
->rb_leftmost
= rb_next(&se
->run_node
);
188 rb_erase(&se
->run_node
, &cfs_rq
->tasks_timeline
);
191 static inline struct rb_node
*first_fair(struct cfs_rq
*cfs_rq
)
193 return cfs_rq
->rb_leftmost
;
196 static struct sched_entity
*__pick_next_entity(struct cfs_rq
*cfs_rq
)
198 return rb_entry(first_fair(cfs_rq
), struct sched_entity
, run_node
);
201 static inline struct sched_entity
*__pick_last_entity(struct cfs_rq
*cfs_rq
)
203 struct rb_node
**link
= &cfs_rq
->tasks_timeline
.rb_node
;
204 struct sched_entity
*se
= NULL
;
205 struct rb_node
*parent
;
209 se
= rb_entry(parent
, struct sched_entity
, run_node
);
210 link
= &parent
->rb_right
;
216 /**************************************************************
217 * Scheduling class statistics methods:
220 #ifdef CONFIG_SCHED_DEBUG
221 int sched_nr_latency_handler(struct ctl_table
*table
, int write
,
222 struct file
*filp
, void __user
*buffer
, size_t *lenp
,
225 int ret
= proc_dointvec_minmax(table
, write
, filp
, buffer
, lenp
, ppos
);
230 sched_nr_latency
= DIV_ROUND_UP(sysctl_sched_latency
,
231 sysctl_sched_min_granularity
);
238 * The idea is to set a period in which each task runs once.
240 * When there are too many tasks (sysctl_sched_nr_latency) we have to stretch
241 * this period because otherwise the slices get too small.
243 * p = (nr <= nl) ? l : l*nr/nl
245 static u64
__sched_period(unsigned long nr_running
)
247 u64 period
= sysctl_sched_latency
;
248 unsigned long nr_latency
= sched_nr_latency
;
250 if (unlikely(nr_running
> nr_latency
)) {
251 period
*= nr_running
;
252 do_div(period
, nr_latency
);
259 * We calculate the wall-time slice from the period by taking a part
260 * proportional to the weight.
264 static u64
sched_slice(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
266 u64 slice
= __sched_period(cfs_rq
->nr_running
);
268 slice
*= se
->load
.weight
;
269 do_div(slice
, cfs_rq
->load
.weight
);
275 * We calculate the vruntime slice.
279 static u64
__sched_vslice(unsigned long rq_weight
, unsigned long nr_running
)
281 u64 vslice
= __sched_period(nr_running
);
283 vslice
*= NICE_0_LOAD
;
284 do_div(vslice
, rq_weight
);
289 static u64
sched_vslice(struct cfs_rq
*cfs_rq
)
291 return __sched_vslice(cfs_rq
->load
.weight
, cfs_rq
->nr_running
);
294 static u64
sched_vslice_add(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
296 return __sched_vslice(cfs_rq
->load
.weight
+ se
->load
.weight
,
297 cfs_rq
->nr_running
+ 1);
301 * Update the current task's runtime statistics. Skip current tasks that
302 * are not in our scheduling class.
305 __update_curr(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
,
306 unsigned long delta_exec
)
308 unsigned long delta_exec_weighted
;
311 schedstat_set(curr
->exec_max
, max((u64
)delta_exec
, curr
->exec_max
));
313 curr
->sum_exec_runtime
+= delta_exec
;
314 schedstat_add(cfs_rq
, exec_clock
, delta_exec
);
315 delta_exec_weighted
= delta_exec
;
316 if (unlikely(curr
->load
.weight
!= NICE_0_LOAD
)) {
317 delta_exec_weighted
= calc_delta_fair(delta_exec_weighted
,
320 curr
->vruntime
+= delta_exec_weighted
;
323 * maintain cfs_rq->min_vruntime to be a monotonic increasing
324 * value tracking the leftmost vruntime in the tree.
326 if (first_fair(cfs_rq
)) {
327 vruntime
= min_vruntime(curr
->vruntime
,
328 __pick_next_entity(cfs_rq
)->vruntime
);
330 vruntime
= curr
->vruntime
;
332 cfs_rq
->min_vruntime
=
333 max_vruntime(cfs_rq
->min_vruntime
, vruntime
);
336 static void update_curr(struct cfs_rq
*cfs_rq
)
338 struct sched_entity
*curr
= cfs_rq
->curr
;
339 u64 now
= rq_of(cfs_rq
)->clock
;
340 unsigned long delta_exec
;
346 * Get the amount of time the current task was running
347 * since the last time we changed load (this cannot
348 * overflow on 32 bits):
350 delta_exec
= (unsigned long)(now
- curr
->exec_start
);
352 __update_curr(cfs_rq
, curr
, delta_exec
);
353 curr
->exec_start
= now
;
355 if (entity_is_task(curr
)) {
356 struct task_struct
*curtask
= task_of(curr
);
358 cpuacct_charge(curtask
, delta_exec
);
363 update_stats_wait_start(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
365 schedstat_set(se
->wait_start
, rq_of(cfs_rq
)->clock
);
369 * Task is being enqueued - update stats:
371 static void update_stats_enqueue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
374 * Are we enqueueing a waiting task? (for current tasks
375 * a dequeue/enqueue event is a NOP)
377 if (se
!= cfs_rq
->curr
)
378 update_stats_wait_start(cfs_rq
, se
);
382 update_stats_wait_end(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
384 schedstat_set(se
->wait_max
, max(se
->wait_max
,
385 rq_of(cfs_rq
)->clock
- se
->wait_start
));
386 schedstat_set(se
->wait_start
, 0);
390 update_stats_dequeue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
393 * Mark the end of the wait period if dequeueing a
396 if (se
!= cfs_rq
->curr
)
397 update_stats_wait_end(cfs_rq
, se
);
401 * We are picking a new current task - update its stats:
404 update_stats_curr_start(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
407 * We are starting a new run period:
409 se
->exec_start
= rq_of(cfs_rq
)->clock
;
412 /**************************************************
413 * Scheduling class queueing methods:
417 account_entity_enqueue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
419 update_load_add(&cfs_rq
->load
, se
->load
.weight
);
420 cfs_rq
->nr_running
++;
425 account_entity_dequeue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
427 update_load_sub(&cfs_rq
->load
, se
->load
.weight
);
428 cfs_rq
->nr_running
--;
432 static void enqueue_sleeper(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
434 #ifdef CONFIG_SCHEDSTATS
435 if (se
->sleep_start
) {
436 u64 delta
= rq_of(cfs_rq
)->clock
- se
->sleep_start
;
441 if (unlikely(delta
> se
->sleep_max
))
442 se
->sleep_max
= delta
;
445 se
->sum_sleep_runtime
+= delta
;
447 if (se
->block_start
) {
448 u64 delta
= rq_of(cfs_rq
)->clock
- se
->block_start
;
453 if (unlikely(delta
> se
->block_max
))
454 se
->block_max
= delta
;
457 se
->sum_sleep_runtime
+= delta
;
460 * Blocking time is in units of nanosecs, so shift by 20 to
461 * get a milliseconds-range estimation of the amount of
462 * time that the task spent sleeping:
464 if (unlikely(prof_on
== SLEEP_PROFILING
)) {
465 struct task_struct
*tsk
= task_of(se
);
467 profile_hits(SLEEP_PROFILING
, (void *)get_wchan(tsk
),
474 static void check_spread(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
476 #ifdef CONFIG_SCHED_DEBUG
477 s64 d
= se
->vruntime
- cfs_rq
->min_vruntime
;
482 if (d
> 3*sysctl_sched_latency
)
483 schedstat_inc(cfs_rq
, nr_spread_over
);
488 place_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int initial
)
492 vruntime
= cfs_rq
->min_vruntime
;
494 if (sched_feat(TREE_AVG
)) {
495 struct sched_entity
*last
= __pick_last_entity(cfs_rq
);
497 vruntime
+= last
->vruntime
;
500 } else if (sched_feat(APPROX_AVG
) && cfs_rq
->nr_running
)
501 vruntime
+= sched_vslice(cfs_rq
)/2;
504 * The 'current' period is already promised to the current tasks,
505 * however the extra weight of the new task will slow them down a
506 * little, place the new task so that it fits in the slot that
507 * stays open at the end.
509 if (initial
&& sched_feat(START_DEBIT
))
510 vruntime
+= sched_vslice_add(cfs_rq
, se
);
513 /* sleeps upto a single latency don't count. */
514 if (sched_feat(NEW_FAIR_SLEEPERS
) && entity_is_task(se
))
515 vruntime
-= sysctl_sched_latency
;
517 /* ensure we never gain time by being placed backwards. */
518 vruntime
= max_vruntime(se
->vruntime
, vruntime
);
521 se
->vruntime
= vruntime
;
525 enqueue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int wakeup
)
528 * Update run-time statistics of the 'current'.
533 place_entity(cfs_rq
, se
, 0);
534 enqueue_sleeper(cfs_rq
, se
);
537 update_stats_enqueue(cfs_rq
, se
);
538 check_spread(cfs_rq
, se
);
539 if (se
!= cfs_rq
->curr
)
540 __enqueue_entity(cfs_rq
, se
);
541 account_entity_enqueue(cfs_rq
, se
);
545 dequeue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int sleep
)
548 * Update run-time statistics of the 'current'.
552 update_stats_dequeue(cfs_rq
, se
);
554 #ifdef CONFIG_SCHEDSTATS
555 if (entity_is_task(se
)) {
556 struct task_struct
*tsk
= task_of(se
);
558 if (tsk
->state
& TASK_INTERRUPTIBLE
)
559 se
->sleep_start
= rq_of(cfs_rq
)->clock
;
560 if (tsk
->state
& TASK_UNINTERRUPTIBLE
)
561 se
->block_start
= rq_of(cfs_rq
)->clock
;
566 if (se
!= cfs_rq
->curr
)
567 __dequeue_entity(cfs_rq
, se
);
568 account_entity_dequeue(cfs_rq
, se
);
572 * Preempt the current task with a newly woken task if needed:
575 check_preempt_tick(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
)
577 unsigned long ideal_runtime
, delta_exec
;
579 ideal_runtime
= sched_slice(cfs_rq
, curr
);
580 delta_exec
= curr
->sum_exec_runtime
- curr
->prev_sum_exec_runtime
;
581 if (delta_exec
> ideal_runtime
)
582 resched_task(rq_of(cfs_rq
)->curr
);
586 set_next_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
588 /* 'current' is not kept within the tree. */
591 * Any task has to be enqueued before it get to execute on
592 * a CPU. So account for the time it spent waiting on the
595 update_stats_wait_end(cfs_rq
, se
);
596 __dequeue_entity(cfs_rq
, se
);
599 update_stats_curr_start(cfs_rq
, se
);
601 #ifdef CONFIG_SCHEDSTATS
603 * Track our maximum slice length, if the CPU's load is at
604 * least twice that of our own weight (i.e. dont track it
605 * when there are only lesser-weight tasks around):
607 if (rq_of(cfs_rq
)->load
.weight
>= 2*se
->load
.weight
) {
608 se
->slice_max
= max(se
->slice_max
,
609 se
->sum_exec_runtime
- se
->prev_sum_exec_runtime
);
612 se
->prev_sum_exec_runtime
= se
->sum_exec_runtime
;
615 static struct sched_entity
*pick_next_entity(struct cfs_rq
*cfs_rq
)
617 struct sched_entity
*se
= NULL
;
619 if (first_fair(cfs_rq
)) {
620 se
= __pick_next_entity(cfs_rq
);
621 set_next_entity(cfs_rq
, se
);
627 static void put_prev_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*prev
)
630 * If still on the runqueue then deactivate_task()
631 * was not called and update_curr() has to be done:
636 check_spread(cfs_rq
, prev
);
638 update_stats_wait_start(cfs_rq
, prev
);
639 /* Put 'current' back into the tree. */
640 __enqueue_entity(cfs_rq
, prev
);
645 static void entity_tick(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
)
648 * Update run-time statistics of the 'current'.
652 if (cfs_rq
->nr_running
> 1 || !sched_feat(WAKEUP_PREEMPT
))
653 check_preempt_tick(cfs_rq
, curr
);
656 /**************************************************
657 * CFS operations on tasks:
660 #ifdef CONFIG_FAIR_GROUP_SCHED
662 /* Walk up scheduling entities hierarchy */
663 #define for_each_sched_entity(se) \
664 for (; se; se = se->parent)
666 static inline struct cfs_rq
*task_cfs_rq(struct task_struct
*p
)
671 /* runqueue on which this entity is (to be) queued */
672 static inline struct cfs_rq
*cfs_rq_of(struct sched_entity
*se
)
677 /* runqueue "owned" by this group */
678 static inline struct cfs_rq
*group_cfs_rq(struct sched_entity
*grp
)
683 /* Given a group's cfs_rq on one cpu, return its corresponding cfs_rq on
684 * another cpu ('this_cpu')
686 static inline struct cfs_rq
*cpu_cfs_rq(struct cfs_rq
*cfs_rq
, int this_cpu
)
688 return cfs_rq
->tg
->cfs_rq
[this_cpu
];
691 /* Iterate thr' all leaf cfs_rq's on a runqueue */
692 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
693 list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
695 /* Do the two (enqueued) entities belong to the same group ? */
697 is_same_group(struct sched_entity
*se
, struct sched_entity
*pse
)
699 if (se
->cfs_rq
== pse
->cfs_rq
)
705 static inline struct sched_entity
*parent_entity(struct sched_entity
*se
)
710 #define GROUP_IMBALANCE_PCT 20
712 #else /* CONFIG_FAIR_GROUP_SCHED */
714 #define for_each_sched_entity(se) \
715 for (; se; se = NULL)
717 static inline struct cfs_rq
*task_cfs_rq(struct task_struct
*p
)
719 return &task_rq(p
)->cfs
;
722 static inline struct cfs_rq
*cfs_rq_of(struct sched_entity
*se
)
724 struct task_struct
*p
= task_of(se
);
725 struct rq
*rq
= task_rq(p
);
730 /* runqueue "owned" by this group */
731 static inline struct cfs_rq
*group_cfs_rq(struct sched_entity
*grp
)
736 static inline struct cfs_rq
*cpu_cfs_rq(struct cfs_rq
*cfs_rq
, int this_cpu
)
738 return &cpu_rq(this_cpu
)->cfs
;
741 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
742 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
745 is_same_group(struct sched_entity
*se
, struct sched_entity
*pse
)
750 static inline struct sched_entity
*parent_entity(struct sched_entity
*se
)
755 #endif /* CONFIG_FAIR_GROUP_SCHED */
758 * The enqueue_task method is called before nr_running is
759 * increased. Here we update the fair scheduling stats and
760 * then put the task into the rbtree:
762 static void enqueue_task_fair(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
764 struct cfs_rq
*cfs_rq
;
765 struct sched_entity
*se
= &p
->se
,
766 *topse
= NULL
; /* Highest schedulable entity */
769 for_each_sched_entity(se
) {
775 cfs_rq
= cfs_rq_of(se
);
776 enqueue_entity(cfs_rq
, se
, wakeup
);
779 /* Increment cpu load if we just enqueued the first task of a group on
780 * 'rq->cpu'. 'topse' represents the group to which task 'p' belongs
781 * at the highest grouping level.
784 inc_cpu_load(rq
, topse
->load
.weight
);
788 * The dequeue_task method is called before nr_running is
789 * decreased. We remove the task from the rbtree and
790 * update the fair scheduling stats:
792 static void dequeue_task_fair(struct rq
*rq
, struct task_struct
*p
, int sleep
)
794 struct cfs_rq
*cfs_rq
;
795 struct sched_entity
*se
= &p
->se
,
796 *topse
= NULL
; /* Highest schedulable entity */
799 for_each_sched_entity(se
) {
801 cfs_rq
= cfs_rq_of(se
);
802 dequeue_entity(cfs_rq
, se
, sleep
);
803 /* Don't dequeue parent if it has other entities besides us */
804 if (cfs_rq
->load
.weight
) {
805 if (parent_entity(se
))
811 /* Decrement cpu load if we just dequeued the last task of a group on
812 * 'rq->cpu'. 'topse' represents the group to which task 'p' belongs
813 * at the highest grouping level.
816 dec_cpu_load(rq
, topse
->load
.weight
);
820 * sched_yield() support is very simple - we dequeue and enqueue.
822 * If compat_yield is turned on then we requeue to the end of the tree.
824 static void yield_task_fair(struct rq
*rq
)
826 struct task_struct
*curr
= rq
->curr
;
827 struct cfs_rq
*cfs_rq
= task_cfs_rq(curr
);
828 struct sched_entity
*rightmost
, *se
= &curr
->se
;
831 * Are we the only task in the tree?
833 if (unlikely(cfs_rq
->nr_running
== 1))
836 if (likely(!sysctl_sched_compat_yield
) && curr
->policy
!= SCHED_BATCH
) {
837 __update_rq_clock(rq
);
839 * Update run-time statistics of the 'current'.
846 * Find the rightmost entry in the rbtree:
848 rightmost
= __pick_last_entity(cfs_rq
);
850 * Already in the rightmost position?
852 if (unlikely(rightmost
->vruntime
< se
->vruntime
))
856 * Minimally necessary key value to be last in the tree:
857 * Upon rescheduling, sched_class::put_prev_task() will place
858 * 'current' within the tree based on its new key value.
860 se
->vruntime
= rightmost
->vruntime
+ 1;
864 * wake_idle() will wake a task on an idle cpu if task->cpu is
865 * not idle and an idle cpu is available. The span of cpus to
866 * search starts with cpus closest then further out as needed,
867 * so we always favor a closer, idle cpu.
869 * Returns the CPU we should wake onto.
871 #if defined(ARCH_HAS_SCHED_WAKE_IDLE)
872 static int wake_idle(int cpu
, struct task_struct
*p
)
875 struct sched_domain
*sd
;
879 * If it is idle, then it is the best cpu to run this task.
881 * This cpu is also the best, if it has more than one task already.
882 * Siblings must be also busy(in most cases) as they didn't already
883 * pickup the extra load from this cpu and hence we need not check
884 * sibling runqueue info. This will avoid the checks and cache miss
885 * penalities associated with that.
887 if (idle_cpu(cpu
) || cpu_rq(cpu
)->nr_running
> 1)
890 for_each_domain(cpu
, sd
) {
891 if (sd
->flags
& SD_WAKE_IDLE
) {
892 cpus_and(tmp
, sd
->span
, p
->cpus_allowed
);
893 for_each_cpu_mask(i
, tmp
) {
895 if (i
!= task_cpu(p
)) {
909 static inline int wake_idle(int cpu
, struct task_struct
*p
)
916 static int select_task_rq_fair(struct task_struct
*p
, int sync
)
920 struct sched_domain
*sd
, *this_sd
= NULL
;
925 this_cpu
= smp_processor_id();
928 for_each_domain(this_cpu
, sd
) {
929 if (cpu_isset(cpu
, sd
->span
)) {
935 if (unlikely(!cpu_isset(this_cpu
, p
->cpus_allowed
)))
939 * Check for affine wakeup and passive balancing possibilities.
942 int idx
= this_sd
->wake_idx
;
943 unsigned int imbalance
;
944 unsigned long load
, this_load
;
946 imbalance
= 100 + (this_sd
->imbalance_pct
- 100) / 2;
948 load
= source_load(cpu
, idx
);
949 this_load
= target_load(this_cpu
, idx
);
951 new_cpu
= this_cpu
; /* Wake to this CPU if we can */
953 if (this_sd
->flags
& SD_WAKE_AFFINE
) {
954 unsigned long tl
= this_load
;
955 unsigned long tl_per_task
;
958 * Attract cache-cold tasks on sync wakeups:
960 if (sync
&& !task_hot(p
, rq
->clock
, this_sd
))
963 schedstat_inc(p
, se
.nr_wakeups_affine_attempts
);
964 tl_per_task
= cpu_avg_load_per_task(this_cpu
);
967 * If sync wakeup then subtract the (maximum possible)
968 * effect of the currently running task from the load
969 * of the current CPU:
972 tl
-= current
->se
.load
.weight
;
975 tl
+ target_load(cpu
, idx
) <= tl_per_task
) ||
976 100*(tl
+ p
->se
.load
.weight
) <= imbalance
*load
) {
978 * This domain has SD_WAKE_AFFINE and
979 * p is cache cold in this domain, and
980 * there is no bad imbalance.
982 schedstat_inc(this_sd
, ttwu_move_affine
);
983 schedstat_inc(p
, se
.nr_wakeups_affine
);
989 * Start passive balancing when half the imbalance_pct
992 if (this_sd
->flags
& SD_WAKE_BALANCE
) {
993 if (imbalance
*this_load
<= 100*load
) {
994 schedstat_inc(this_sd
, ttwu_move_balance
);
995 schedstat_inc(p
, se
.nr_wakeups_passive
);
1001 new_cpu
= cpu
; /* Could not wake to this_cpu. Wake to cpu instead */
1003 return wake_idle(new_cpu
, p
);
1005 #endif /* CONFIG_SMP */
1009 * Preempt the current task with a newly woken task if needed:
1011 static void check_preempt_wakeup(struct rq
*rq
, struct task_struct
*p
)
1013 struct task_struct
*curr
= rq
->curr
;
1014 struct cfs_rq
*cfs_rq
= task_cfs_rq(curr
);
1015 struct sched_entity
*se
= &curr
->se
, *pse
= &p
->se
;
1018 if (unlikely(rt_prio(p
->prio
))) {
1019 update_rq_clock(rq
);
1020 update_curr(cfs_rq
);
1025 * Batch tasks do not preempt (their preemption is driven by
1028 if (unlikely(p
->policy
== SCHED_BATCH
))
1031 if (!sched_feat(WAKEUP_PREEMPT
))
1034 while (!is_same_group(se
, pse
)) {
1035 se
= parent_entity(se
);
1036 pse
= parent_entity(pse
);
1039 gran
= sysctl_sched_wakeup_granularity
;
1040 if (unlikely(se
->load
.weight
!= NICE_0_LOAD
))
1041 gran
= calc_delta_fair(gran
, &se
->load
);
1043 if (pse
->vruntime
+ gran
< se
->vruntime
)
1047 static struct task_struct
*pick_next_task_fair(struct rq
*rq
)
1049 struct cfs_rq
*cfs_rq
= &rq
->cfs
;
1050 struct sched_entity
*se
;
1052 if (unlikely(!cfs_rq
->nr_running
))
1056 se
= pick_next_entity(cfs_rq
);
1057 cfs_rq
= group_cfs_rq(se
);
1064 * Account for a descheduled task:
1066 static void put_prev_task_fair(struct rq
*rq
, struct task_struct
*prev
)
1068 struct sched_entity
*se
= &prev
->se
;
1069 struct cfs_rq
*cfs_rq
;
1071 for_each_sched_entity(se
) {
1072 cfs_rq
= cfs_rq_of(se
);
1073 put_prev_entity(cfs_rq
, se
);
1078 /**************************************************
1079 * Fair scheduling class load-balancing methods:
1083 * Load-balancing iterator. Note: while the runqueue stays locked
1084 * during the whole iteration, the current task might be
1085 * dequeued so the iterator has to be dequeue-safe. Here we
1086 * achieve that by always pre-iterating before returning
1089 static struct task_struct
*
1090 __load_balance_iterator(struct cfs_rq
*cfs_rq
, struct rb_node
*curr
)
1092 struct task_struct
*p
;
1097 p
= rb_entry(curr
, struct task_struct
, se
.run_node
);
1098 cfs_rq
->rb_load_balance_curr
= rb_next(curr
);
1103 static struct task_struct
*load_balance_start_fair(void *arg
)
1105 struct cfs_rq
*cfs_rq
= arg
;
1107 return __load_balance_iterator(cfs_rq
, first_fair(cfs_rq
));
1110 static struct task_struct
*load_balance_next_fair(void *arg
)
1112 struct cfs_rq
*cfs_rq
= arg
;
1114 return __load_balance_iterator(cfs_rq
, cfs_rq
->rb_load_balance_curr
);
1117 static unsigned long
1118 load_balance_fair(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1119 unsigned long max_load_move
,
1120 struct sched_domain
*sd
, enum cpu_idle_type idle
,
1121 int *all_pinned
, int *this_best_prio
)
1123 struct cfs_rq
*busy_cfs_rq
;
1124 long rem_load_move
= max_load_move
;
1125 struct rq_iterator cfs_rq_iterator
;
1126 unsigned long load_moved
;
1128 cfs_rq_iterator
.start
= load_balance_start_fair
;
1129 cfs_rq_iterator
.next
= load_balance_next_fair
;
1131 for_each_leaf_cfs_rq(busiest
, busy_cfs_rq
) {
1132 #ifdef CONFIG_FAIR_GROUP_SCHED
1133 struct cfs_rq
*this_cfs_rq
= busy_cfs_rq
->tg
->cfs_rq
[this_cpu
];
1134 unsigned long maxload
, task_load
, group_weight
;
1135 unsigned long thisload
, per_task_load
;
1136 struct sched_entity
*se
= busy_cfs_rq
->tg
->se
[busiest
->cpu
];
1138 task_load
= busy_cfs_rq
->load
.weight
;
1139 group_weight
= se
->load
.weight
;
1142 * 'group_weight' is contributed by tasks of total weight
1143 * 'task_load'. To move 'rem_load_move' worth of weight only,
1144 * we need to move a maximum task load of:
1146 * maxload = (remload / group_weight) * task_load;
1148 maxload
= (rem_load_move
* task_load
) / group_weight
;
1150 if (!maxload
|| !task_load
)
1153 per_task_load
= task_load
/ busy_cfs_rq
->nr_running
;
1155 * balance_tasks will try to forcibly move atleast one task if
1156 * possible (because of SCHED_LOAD_SCALE_FUZZ). Avoid that if
1157 * maxload is less than GROUP_IMBALANCE_FUZZ% the per_task_load.
1159 if (100 * maxload
< GROUP_IMBALANCE_PCT
* per_task_load
)
1162 /* Disable priority-based load balance */
1163 *this_best_prio
= 0;
1164 thisload
= this_cfs_rq
->load
.weight
;
1166 # define maxload rem_load_move
1169 * pass busy_cfs_rq argument into
1170 * load_balance_[start|next]_fair iterators
1172 cfs_rq_iterator
.arg
= busy_cfs_rq
;
1173 load_moved
= balance_tasks(this_rq
, this_cpu
, busiest
,
1174 maxload
, sd
, idle
, all_pinned
,
1178 #ifdef CONFIG_FAIR_GROUP_SCHED
1180 * load_moved holds the task load that was moved. The
1181 * effective (group) weight moved would be:
1182 * load_moved_eff = load_moved/task_load * group_weight;
1184 load_moved
= (group_weight
* load_moved
) / task_load
;
1186 /* Adjust shares on both cpus to reflect load_moved */
1187 group_weight
-= load_moved
;
1188 set_se_shares(se
, group_weight
);
1190 se
= busy_cfs_rq
->tg
->se
[this_cpu
];
1192 group_weight
= load_moved
;
1194 group_weight
= se
->load
.weight
+ load_moved
;
1195 set_se_shares(se
, group_weight
);
1198 rem_load_move
-= load_moved
;
1200 if (rem_load_move
<= 0)
1204 return max_load_move
- rem_load_move
;
1208 move_one_task_fair(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1209 struct sched_domain
*sd
, enum cpu_idle_type idle
)
1211 struct cfs_rq
*busy_cfs_rq
;
1212 struct rq_iterator cfs_rq_iterator
;
1214 cfs_rq_iterator
.start
= load_balance_start_fair
;
1215 cfs_rq_iterator
.next
= load_balance_next_fair
;
1217 for_each_leaf_cfs_rq(busiest
, busy_cfs_rq
) {
1219 * pass busy_cfs_rq argument into
1220 * load_balance_[start|next]_fair iterators
1222 cfs_rq_iterator
.arg
= busy_cfs_rq
;
1223 if (iter_move_one_task(this_rq
, this_cpu
, busiest
, sd
, idle
,
1233 * scheduler tick hitting a task of our scheduling class:
1235 static void task_tick_fair(struct rq
*rq
, struct task_struct
*curr
)
1237 struct cfs_rq
*cfs_rq
;
1238 struct sched_entity
*se
= &curr
->se
;
1240 for_each_sched_entity(se
) {
1241 cfs_rq
= cfs_rq_of(se
);
1242 entity_tick(cfs_rq
, se
);
1246 #define swap(a, b) do { typeof(a) tmp = (a); (a) = (b); (b) = tmp; } while (0)
1249 * Share the fairness runtime between parent and child, thus the
1250 * total amount of pressure for CPU stays equal - new tasks
1251 * get a chance to run but frequent forkers are not allowed to
1252 * monopolize the CPU. Note: the parent runqueue is locked,
1253 * the child is not running yet.
1255 static void task_new_fair(struct rq
*rq
, struct task_struct
*p
)
1257 struct cfs_rq
*cfs_rq
= task_cfs_rq(p
);
1258 struct sched_entity
*se
= &p
->se
, *curr
= cfs_rq
->curr
;
1259 int this_cpu
= smp_processor_id();
1261 sched_info_queued(p
);
1263 update_curr(cfs_rq
);
1264 place_entity(cfs_rq
, se
, 1);
1266 /* 'curr' will be NULL if the child belongs to a different group */
1267 if (sysctl_sched_child_runs_first
&& this_cpu
== task_cpu(p
) &&
1268 curr
&& curr
->vruntime
< se
->vruntime
) {
1270 * Upon rescheduling, sched_class::put_prev_task() will place
1271 * 'current' within the tree based on its new key value.
1273 swap(curr
->vruntime
, se
->vruntime
);
1276 enqueue_task_fair(rq
, p
, 0);
1277 resched_task(rq
->curr
);
1280 /* Account for a task changing its policy or group.
1282 * This routine is mostly called to set cfs_rq->curr field when a task
1283 * migrates between groups/classes.
1285 static void set_curr_task_fair(struct rq
*rq
)
1287 struct sched_entity
*se
= &rq
->curr
->se
;
1289 for_each_sched_entity(se
)
1290 set_next_entity(cfs_rq_of(se
), se
);
1294 * All the scheduling class methods:
1296 static const struct sched_class fair_sched_class
= {
1297 .next
= &idle_sched_class
,
1298 .enqueue_task
= enqueue_task_fair
,
1299 .dequeue_task
= dequeue_task_fair
,
1300 .yield_task
= yield_task_fair
,
1302 .select_task_rq
= select_task_rq_fair
,
1303 #endif /* CONFIG_SMP */
1305 .check_preempt_curr
= check_preempt_wakeup
,
1307 .pick_next_task
= pick_next_task_fair
,
1308 .put_prev_task
= put_prev_task_fair
,
1311 .load_balance
= load_balance_fair
,
1312 .move_one_task
= move_one_task_fair
,
1315 .set_curr_task
= set_curr_task_fair
,
1316 .task_tick
= task_tick_fair
,
1317 .task_new
= task_new_fair
,
1320 #ifdef CONFIG_SCHED_DEBUG
1321 static void print_cfs_stats(struct seq_file
*m
, int cpu
)
1323 struct cfs_rq
*cfs_rq
;
1325 #ifdef CONFIG_FAIR_GROUP_SCHED
1326 print_cfs_rq(m
, cpu
, &cpu_rq(cpu
)->cfs
);
1328 lock_task_group_list();
1329 for_each_leaf_cfs_rq(cpu_rq(cpu
), cfs_rq
)
1330 print_cfs_rq(m
, cpu
, cfs_rq
);
1331 unlock_task_group_list();