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: 5 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
= 5000000UL;
74 const_debug
unsigned int sysctl_sched_migration_cost
= 500000UL;
76 static const struct sched_class fair_sched_class
;
78 /**************************************************************
79 * CFS operations on generic schedulable entities:
82 static inline struct task_struct
*task_of(struct sched_entity
*se
)
84 return container_of(se
, struct task_struct
, se
);
87 #ifdef CONFIG_FAIR_GROUP_SCHED
89 /* cpu runqueue to which this cfs_rq is attached */
90 static inline struct rq
*rq_of(struct cfs_rq
*cfs_rq
)
95 /* An entity is a task if it doesn't "own" a runqueue */
96 #define entity_is_task(se) (!se->my_q)
98 /* Walk up scheduling entities hierarchy */
99 #define for_each_sched_entity(se) \
100 for (; se; se = se->parent)
102 static inline struct cfs_rq
*task_cfs_rq(struct task_struct
*p
)
107 /* runqueue on which this entity is (to be) queued */
108 static inline struct cfs_rq
*cfs_rq_of(struct sched_entity
*se
)
113 /* runqueue "owned" by this group */
114 static inline struct cfs_rq
*group_cfs_rq(struct sched_entity
*grp
)
119 /* Given a group's cfs_rq on one cpu, return its corresponding cfs_rq on
120 * another cpu ('this_cpu')
122 static inline struct cfs_rq
*cpu_cfs_rq(struct cfs_rq
*cfs_rq
, int this_cpu
)
124 return cfs_rq
->tg
->cfs_rq
[this_cpu
];
127 /* Iterate thr' all leaf cfs_rq's on a runqueue */
128 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
129 list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
131 /* Do the two (enqueued) entities belong to the same group ? */
133 is_same_group(struct sched_entity
*se
, struct sched_entity
*pse
)
135 if (se
->cfs_rq
== pse
->cfs_rq
)
141 static inline struct sched_entity
*parent_entity(struct sched_entity
*se
)
146 /* return depth at which a sched entity is present in the hierarchy */
147 static inline int depth_se(struct sched_entity
*se
)
151 for_each_sched_entity(se
)
158 find_matching_se(struct sched_entity
**se
, struct sched_entity
**pse
)
160 int se_depth
, pse_depth
;
163 * preemption test can be made between sibling entities who are in the
164 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
165 * both tasks until we find their ancestors who are siblings of common
169 /* First walk up until both entities are at same depth */
170 se_depth
= depth_se(*se
);
171 pse_depth
= depth_se(*pse
);
173 while (se_depth
> pse_depth
) {
175 *se
= parent_entity(*se
);
178 while (pse_depth
> se_depth
) {
180 *pse
= parent_entity(*pse
);
183 while (!is_same_group(*se
, *pse
)) {
184 *se
= parent_entity(*se
);
185 *pse
= parent_entity(*pse
);
189 #else /* CONFIG_FAIR_GROUP_SCHED */
191 static inline struct rq
*rq_of(struct cfs_rq
*cfs_rq
)
193 return container_of(cfs_rq
, struct rq
, cfs
);
196 #define entity_is_task(se) 1
198 #define for_each_sched_entity(se) \
199 for (; se; se = NULL)
201 static inline struct cfs_rq
*task_cfs_rq(struct task_struct
*p
)
203 return &task_rq(p
)->cfs
;
206 static inline struct cfs_rq
*cfs_rq_of(struct sched_entity
*se
)
208 struct task_struct
*p
= task_of(se
);
209 struct rq
*rq
= task_rq(p
);
214 /* runqueue "owned" by this group */
215 static inline struct cfs_rq
*group_cfs_rq(struct sched_entity
*grp
)
220 static inline struct cfs_rq
*cpu_cfs_rq(struct cfs_rq
*cfs_rq
, int this_cpu
)
222 return &cpu_rq(this_cpu
)->cfs
;
225 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
226 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
229 is_same_group(struct sched_entity
*se
, struct sched_entity
*pse
)
234 static inline struct sched_entity
*parent_entity(struct sched_entity
*se
)
240 find_matching_se(struct sched_entity
**se
, struct sched_entity
**pse
)
244 #endif /* CONFIG_FAIR_GROUP_SCHED */
247 /**************************************************************
248 * Scheduling class tree data structure manipulation methods:
251 static inline u64
max_vruntime(u64 min_vruntime
, u64 vruntime
)
253 s64 delta
= (s64
)(vruntime
- min_vruntime
);
255 min_vruntime
= vruntime
;
260 static inline u64
min_vruntime(u64 min_vruntime
, u64 vruntime
)
262 s64 delta
= (s64
)(vruntime
- min_vruntime
);
264 min_vruntime
= vruntime
;
269 static inline s64
entity_key(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
271 return se
->vruntime
- cfs_rq
->min_vruntime
;
274 static void update_min_vruntime(struct cfs_rq
*cfs_rq
)
276 u64 vruntime
= cfs_rq
->min_vruntime
;
279 vruntime
= cfs_rq
->curr
->vruntime
;
281 if (cfs_rq
->rb_leftmost
) {
282 struct sched_entity
*se
= rb_entry(cfs_rq
->rb_leftmost
,
286 if (vruntime
== cfs_rq
->min_vruntime
)
287 vruntime
= se
->vruntime
;
289 vruntime
= min_vruntime(vruntime
, se
->vruntime
);
292 cfs_rq
->min_vruntime
= max_vruntime(cfs_rq
->min_vruntime
, vruntime
);
296 * Enqueue an entity into the rb-tree:
298 static void __enqueue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
300 struct rb_node
**link
= &cfs_rq
->tasks_timeline
.rb_node
;
301 struct rb_node
*parent
= NULL
;
302 struct sched_entity
*entry
;
303 s64 key
= entity_key(cfs_rq
, se
);
307 * Find the right place in the rbtree:
311 entry
= rb_entry(parent
, struct sched_entity
, run_node
);
313 * We dont care about collisions. Nodes with
314 * the same key stay together.
316 if (key
< entity_key(cfs_rq
, entry
)) {
317 link
= &parent
->rb_left
;
319 link
= &parent
->rb_right
;
325 * Maintain a cache of leftmost tree entries (it is frequently
329 cfs_rq
->rb_leftmost
= &se
->run_node
;
331 rb_link_node(&se
->run_node
, parent
, link
);
332 rb_insert_color(&se
->run_node
, &cfs_rq
->tasks_timeline
);
335 static void __dequeue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
337 if (cfs_rq
->rb_leftmost
== &se
->run_node
) {
338 struct rb_node
*next_node
;
340 next_node
= rb_next(&se
->run_node
);
341 cfs_rq
->rb_leftmost
= next_node
;
344 if (cfs_rq
->next
== se
)
347 rb_erase(&se
->run_node
, &cfs_rq
->tasks_timeline
);
350 static inline struct rb_node
*first_fair(struct cfs_rq
*cfs_rq
)
352 return cfs_rq
->rb_leftmost
;
355 static struct sched_entity
*__pick_next_entity(struct cfs_rq
*cfs_rq
)
357 return rb_entry(first_fair(cfs_rq
), struct sched_entity
, run_node
);
360 static inline struct sched_entity
*__pick_last_entity(struct cfs_rq
*cfs_rq
)
362 struct rb_node
*last
= rb_last(&cfs_rq
->tasks_timeline
);
367 return rb_entry(last
, struct sched_entity
, run_node
);
370 /**************************************************************
371 * Scheduling class statistics methods:
374 #ifdef CONFIG_SCHED_DEBUG
375 int sched_nr_latency_handler(struct ctl_table
*table
, int write
,
376 struct file
*filp
, void __user
*buffer
, size_t *lenp
,
379 int ret
= proc_dointvec_minmax(table
, write
, filp
, buffer
, lenp
, ppos
);
384 sched_nr_latency
= DIV_ROUND_UP(sysctl_sched_latency
,
385 sysctl_sched_min_granularity
);
394 static inline unsigned long
395 calc_delta_weight(unsigned long delta
, struct sched_entity
*se
)
397 for_each_sched_entity(se
) {
398 delta
= calc_delta_mine(delta
,
399 se
->load
.weight
, &cfs_rq_of(se
)->load
);
408 static inline unsigned long
409 calc_delta_fair(unsigned long delta
, struct sched_entity
*se
)
411 if (unlikely(se
->load
.weight
!= NICE_0_LOAD
))
412 delta
= calc_delta_mine(delta
, NICE_0_LOAD
, &se
->load
);
418 * The idea is to set a period in which each task runs once.
420 * When there are too many tasks (sysctl_sched_nr_latency) we have to stretch
421 * this period because otherwise the slices get too small.
423 * p = (nr <= nl) ? l : l*nr/nl
425 static u64
__sched_period(unsigned long nr_running
)
427 u64 period
= sysctl_sched_latency
;
428 unsigned long nr_latency
= sched_nr_latency
;
430 if (unlikely(nr_running
> nr_latency
)) {
431 period
= sysctl_sched_min_granularity
;
432 period
*= nr_running
;
439 * We calculate the wall-time slice from the period by taking a part
440 * proportional to the weight.
444 static u64
sched_slice(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
446 unsigned long nr_running
= cfs_rq
->nr_running
;
448 if (unlikely(!se
->on_rq
))
451 return calc_delta_weight(__sched_period(nr_running
), se
);
455 * We calculate the vruntime slice of a to be inserted task
459 static u64
sched_vslice(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
461 return calc_delta_fair(sched_slice(cfs_rq
, se
), se
);
465 * Update the current task's runtime statistics. Skip current tasks that
466 * are not in our scheduling class.
469 __update_curr(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
,
470 unsigned long delta_exec
)
472 unsigned long delta_exec_weighted
;
474 schedstat_set(curr
->exec_max
, max((u64
)delta_exec
, curr
->exec_max
));
476 curr
->sum_exec_runtime
+= delta_exec
;
477 schedstat_add(cfs_rq
, exec_clock
, delta_exec
);
478 delta_exec_weighted
= calc_delta_fair(delta_exec
, curr
);
479 curr
->vruntime
+= delta_exec_weighted
;
480 update_min_vruntime(cfs_rq
);
483 static void update_curr(struct cfs_rq
*cfs_rq
)
485 struct sched_entity
*curr
= cfs_rq
->curr
;
486 u64 now
= rq_of(cfs_rq
)->clock
;
487 unsigned long delta_exec
;
493 * Get the amount of time the current task was running
494 * since the last time we changed load (this cannot
495 * overflow on 32 bits):
497 delta_exec
= (unsigned long)(now
- curr
->exec_start
);
499 __update_curr(cfs_rq
, curr
, delta_exec
);
500 curr
->exec_start
= now
;
502 if (entity_is_task(curr
)) {
503 struct task_struct
*curtask
= task_of(curr
);
505 cpuacct_charge(curtask
, delta_exec
);
506 account_group_exec_runtime(curtask
, delta_exec
);
511 update_stats_wait_start(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
513 schedstat_set(se
->wait_start
, rq_of(cfs_rq
)->clock
);
517 * Task is being enqueued - update stats:
519 static void update_stats_enqueue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
522 * Are we enqueueing a waiting task? (for current tasks
523 * a dequeue/enqueue event is a NOP)
525 if (se
!= cfs_rq
->curr
)
526 update_stats_wait_start(cfs_rq
, se
);
530 update_stats_wait_end(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
532 schedstat_set(se
->wait_max
, max(se
->wait_max
,
533 rq_of(cfs_rq
)->clock
- se
->wait_start
));
534 schedstat_set(se
->wait_count
, se
->wait_count
+ 1);
535 schedstat_set(se
->wait_sum
, se
->wait_sum
+
536 rq_of(cfs_rq
)->clock
- se
->wait_start
);
537 schedstat_set(se
->wait_start
, 0);
541 update_stats_dequeue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
544 * Mark the end of the wait period if dequeueing a
547 if (se
!= cfs_rq
->curr
)
548 update_stats_wait_end(cfs_rq
, se
);
552 * We are picking a new current task - update its stats:
555 update_stats_curr_start(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
558 * We are starting a new run period:
560 se
->exec_start
= rq_of(cfs_rq
)->clock
;
563 /**************************************************
564 * Scheduling class queueing methods:
567 #if defined CONFIG_SMP && defined CONFIG_FAIR_GROUP_SCHED
569 add_cfs_task_weight(struct cfs_rq
*cfs_rq
, unsigned long weight
)
571 cfs_rq
->task_weight
+= weight
;
575 add_cfs_task_weight(struct cfs_rq
*cfs_rq
, unsigned long weight
)
581 account_entity_enqueue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
583 update_load_add(&cfs_rq
->load
, se
->load
.weight
);
584 if (!parent_entity(se
))
585 inc_cpu_load(rq_of(cfs_rq
), se
->load
.weight
);
586 if (entity_is_task(se
)) {
587 add_cfs_task_weight(cfs_rq
, se
->load
.weight
);
588 list_add(&se
->group_node
, &cfs_rq
->tasks
);
590 cfs_rq
->nr_running
++;
595 account_entity_dequeue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
597 update_load_sub(&cfs_rq
->load
, se
->load
.weight
);
598 if (!parent_entity(se
))
599 dec_cpu_load(rq_of(cfs_rq
), se
->load
.weight
);
600 if (entity_is_task(se
)) {
601 add_cfs_task_weight(cfs_rq
, -se
->load
.weight
);
602 list_del_init(&se
->group_node
);
604 cfs_rq
->nr_running
--;
608 static void enqueue_sleeper(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
610 #ifdef CONFIG_SCHEDSTATS
611 if (se
->sleep_start
) {
612 u64 delta
= rq_of(cfs_rq
)->clock
- se
->sleep_start
;
613 struct task_struct
*tsk
= task_of(se
);
618 if (unlikely(delta
> se
->sleep_max
))
619 se
->sleep_max
= delta
;
622 se
->sum_sleep_runtime
+= delta
;
624 account_scheduler_latency(tsk
, delta
>> 10, 1);
626 if (se
->block_start
) {
627 u64 delta
= rq_of(cfs_rq
)->clock
- se
->block_start
;
628 struct task_struct
*tsk
= task_of(se
);
633 if (unlikely(delta
> se
->block_max
))
634 se
->block_max
= delta
;
637 se
->sum_sleep_runtime
+= delta
;
640 * Blocking time is in units of nanosecs, so shift by 20 to
641 * get a milliseconds-range estimation of the amount of
642 * time that the task spent sleeping:
644 if (unlikely(prof_on
== SLEEP_PROFILING
)) {
646 profile_hits(SLEEP_PROFILING
, (void *)get_wchan(tsk
),
649 account_scheduler_latency(tsk
, delta
>> 10, 0);
654 static void check_spread(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
656 #ifdef CONFIG_SCHED_DEBUG
657 s64 d
= se
->vruntime
- cfs_rq
->min_vruntime
;
662 if (d
> 3*sysctl_sched_latency
)
663 schedstat_inc(cfs_rq
, nr_spread_over
);
668 place_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int initial
)
670 u64 vruntime
= cfs_rq
->min_vruntime
;
673 * The 'current' period is already promised to the current tasks,
674 * however the extra weight of the new task will slow them down a
675 * little, place the new task so that it fits in the slot that
676 * stays open at the end.
678 if (initial
&& sched_feat(START_DEBIT
))
679 vruntime
+= sched_vslice(cfs_rq
, se
);
682 /* sleeps upto a single latency don't count. */
683 if (sched_feat(NEW_FAIR_SLEEPERS
)) {
684 unsigned long thresh
= sysctl_sched_latency
;
687 * convert the sleeper threshold into virtual time
689 if (sched_feat(NORMALIZED_SLEEPER
))
690 thresh
= calc_delta_fair(thresh
, se
);
695 /* ensure we never gain time by being placed backwards. */
696 vruntime
= max_vruntime(se
->vruntime
, vruntime
);
699 se
->vruntime
= vruntime
;
703 enqueue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int wakeup
)
706 * Update run-time statistics of the 'current'.
709 account_entity_enqueue(cfs_rq
, se
);
712 place_entity(cfs_rq
, se
, 0);
713 enqueue_sleeper(cfs_rq
, se
);
716 update_stats_enqueue(cfs_rq
, se
);
717 check_spread(cfs_rq
, se
);
718 if (se
!= cfs_rq
->curr
)
719 __enqueue_entity(cfs_rq
, se
);
723 dequeue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int sleep
)
726 * Update run-time statistics of the 'current'.
730 update_stats_dequeue(cfs_rq
, se
);
732 #ifdef CONFIG_SCHEDSTATS
733 if (entity_is_task(se
)) {
734 struct task_struct
*tsk
= task_of(se
);
736 if (tsk
->state
& TASK_INTERRUPTIBLE
)
737 se
->sleep_start
= rq_of(cfs_rq
)->clock
;
738 if (tsk
->state
& TASK_UNINTERRUPTIBLE
)
739 se
->block_start
= rq_of(cfs_rq
)->clock
;
744 if (se
!= cfs_rq
->curr
)
745 __dequeue_entity(cfs_rq
, se
);
746 account_entity_dequeue(cfs_rq
, se
);
747 update_min_vruntime(cfs_rq
);
751 * Preempt the current task with a newly woken task if needed:
754 check_preempt_tick(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
)
756 unsigned long ideal_runtime
, delta_exec
;
758 ideal_runtime
= sched_slice(cfs_rq
, curr
);
759 delta_exec
= curr
->sum_exec_runtime
- curr
->prev_sum_exec_runtime
;
760 if (delta_exec
> ideal_runtime
)
761 resched_task(rq_of(cfs_rq
)->curr
);
765 set_next_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
767 /* 'current' is not kept within the tree. */
770 * Any task has to be enqueued before it get to execute on
771 * a CPU. So account for the time it spent waiting on the
774 update_stats_wait_end(cfs_rq
, se
);
775 __dequeue_entity(cfs_rq
, se
);
778 update_stats_curr_start(cfs_rq
, se
);
780 #ifdef CONFIG_SCHEDSTATS
782 * Track our maximum slice length, if the CPU's load is at
783 * least twice that of our own weight (i.e. dont track it
784 * when there are only lesser-weight tasks around):
786 if (rq_of(cfs_rq
)->load
.weight
>= 2*se
->load
.weight
) {
787 se
->slice_max
= max(se
->slice_max
,
788 se
->sum_exec_runtime
- se
->prev_sum_exec_runtime
);
791 se
->prev_sum_exec_runtime
= se
->sum_exec_runtime
;
795 wakeup_preempt_entity(struct sched_entity
*curr
, struct sched_entity
*se
);
797 static struct sched_entity
*
798 pick_next(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
800 if (!cfs_rq
->next
|| wakeup_preempt_entity(cfs_rq
->next
, se
) == 1)
806 static struct sched_entity
*pick_next_entity(struct cfs_rq
*cfs_rq
)
808 struct sched_entity
*se
= NULL
;
810 if (first_fair(cfs_rq
)) {
811 se
= __pick_next_entity(cfs_rq
);
812 se
= pick_next(cfs_rq
, se
);
813 set_next_entity(cfs_rq
, se
);
819 static void put_prev_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*prev
)
822 * If still on the runqueue then deactivate_task()
823 * was not called and update_curr() has to be done:
828 check_spread(cfs_rq
, prev
);
830 update_stats_wait_start(cfs_rq
, prev
);
831 /* Put 'current' back into the tree. */
832 __enqueue_entity(cfs_rq
, prev
);
838 entity_tick(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
, int queued
)
841 * Update run-time statistics of the 'current'.
845 #ifdef CONFIG_SCHED_HRTICK
847 * queued ticks are scheduled to match the slice, so don't bother
848 * validating it and just reschedule.
851 resched_task(rq_of(cfs_rq
)->curr
);
855 * don't let the period tick interfere with the hrtick preemption
857 if (!sched_feat(DOUBLE_TICK
) &&
858 hrtimer_active(&rq_of(cfs_rq
)->hrtick_timer
))
862 if (cfs_rq
->nr_running
> 1 || !sched_feat(WAKEUP_PREEMPT
))
863 check_preempt_tick(cfs_rq
, curr
);
866 /**************************************************
867 * CFS operations on tasks:
870 #ifdef CONFIG_SCHED_HRTICK
871 static void hrtick_start_fair(struct rq
*rq
, struct task_struct
*p
)
873 struct sched_entity
*se
= &p
->se
;
874 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
876 WARN_ON(task_rq(p
) != rq
);
878 if (hrtick_enabled(rq
) && cfs_rq
->nr_running
> 1) {
879 u64 slice
= sched_slice(cfs_rq
, se
);
880 u64 ran
= se
->sum_exec_runtime
- se
->prev_sum_exec_runtime
;
881 s64 delta
= slice
- ran
;
890 * Don't schedule slices shorter than 10000ns, that just
891 * doesn't make sense. Rely on vruntime for fairness.
894 delta
= max_t(s64
, 10000LL, delta
);
896 hrtick_start(rq
, delta
);
901 * called from enqueue/dequeue and updates the hrtick when the
902 * current task is from our class and nr_running is low enough
905 static void hrtick_update(struct rq
*rq
)
907 struct task_struct
*curr
= rq
->curr
;
909 if (curr
->sched_class
!= &fair_sched_class
)
912 if (cfs_rq_of(&curr
->se
)->nr_running
< sched_nr_latency
)
913 hrtick_start_fair(rq
, curr
);
915 #else /* !CONFIG_SCHED_HRTICK */
917 hrtick_start_fair(struct rq
*rq
, struct task_struct
*p
)
921 static inline void hrtick_update(struct rq
*rq
)
927 * The enqueue_task method is called before nr_running is
928 * increased. Here we update the fair scheduling stats and
929 * then put the task into the rbtree:
931 static void enqueue_task_fair(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
933 struct cfs_rq
*cfs_rq
;
934 struct sched_entity
*se
= &p
->se
;
936 for_each_sched_entity(se
) {
939 cfs_rq
= cfs_rq_of(se
);
940 enqueue_entity(cfs_rq
, se
, wakeup
);
948 * The dequeue_task method is called before nr_running is
949 * decreased. We remove the task from the rbtree and
950 * update the fair scheduling stats:
952 static void dequeue_task_fair(struct rq
*rq
, struct task_struct
*p
, int sleep
)
954 struct cfs_rq
*cfs_rq
;
955 struct sched_entity
*se
= &p
->se
;
957 for_each_sched_entity(se
) {
958 cfs_rq
= cfs_rq_of(se
);
959 dequeue_entity(cfs_rq
, se
, sleep
);
960 /* Don't dequeue parent if it has other entities besides us */
961 if (cfs_rq
->load
.weight
)
970 * sched_yield() support is very simple - we dequeue and enqueue.
972 * If compat_yield is turned on then we requeue to the end of the tree.
974 static void yield_task_fair(struct rq
*rq
)
976 struct task_struct
*curr
= rq
->curr
;
977 struct cfs_rq
*cfs_rq
= task_cfs_rq(curr
);
978 struct sched_entity
*rightmost
, *se
= &curr
->se
;
981 * Are we the only task in the tree?
983 if (unlikely(cfs_rq
->nr_running
== 1))
986 if (likely(!sysctl_sched_compat_yield
) && curr
->policy
!= SCHED_BATCH
) {
989 * Update run-time statistics of the 'current'.
996 * Find the rightmost entry in the rbtree:
998 rightmost
= __pick_last_entity(cfs_rq
);
1000 * Already in the rightmost position?
1002 if (unlikely(!rightmost
|| rightmost
->vruntime
< se
->vruntime
))
1006 * Minimally necessary key value to be last in the tree:
1007 * Upon rescheduling, sched_class::put_prev_task() will place
1008 * 'current' within the tree based on its new key value.
1010 se
->vruntime
= rightmost
->vruntime
+ 1;
1014 * wake_idle() will wake a task on an idle cpu if task->cpu is
1015 * not idle and an idle cpu is available. The span of cpus to
1016 * search starts with cpus closest then further out as needed,
1017 * so we always favor a closer, idle cpu.
1018 * Domains may include CPUs that are not usable for migration,
1019 * hence we need to mask them out (cpu_active_map)
1021 * Returns the CPU we should wake onto.
1023 #if defined(ARCH_HAS_SCHED_WAKE_IDLE)
1024 static int wake_idle(int cpu
, struct task_struct
*p
)
1027 struct sched_domain
*sd
;
1031 * If it is idle, then it is the best cpu to run this task.
1033 * This cpu is also the best, if it has more than one task already.
1034 * Siblings must be also busy(in most cases) as they didn't already
1035 * pickup the extra load from this cpu and hence we need not check
1036 * sibling runqueue info. This will avoid the checks and cache miss
1037 * penalities associated with that.
1039 if (idle_cpu(cpu
) || cpu_rq(cpu
)->cfs
.nr_running
> 1)
1042 for_each_domain(cpu
, sd
) {
1043 if ((sd
->flags
& SD_WAKE_IDLE
)
1044 || ((sd
->flags
& SD_WAKE_IDLE_FAR
)
1045 && !task_hot(p
, task_rq(p
)->clock
, sd
))) {
1046 cpus_and(tmp
, sd
->span
, p
->cpus_allowed
);
1047 cpus_and(tmp
, tmp
, cpu_active_map
);
1048 for_each_cpu_mask_nr(i
, tmp
) {
1050 if (i
!= task_cpu(p
)) {
1052 se
.nr_wakeups_idle
);
1063 #else /* !ARCH_HAS_SCHED_WAKE_IDLE*/
1064 static inline int wake_idle(int cpu
, struct task_struct
*p
)
1072 #ifdef CONFIG_FAIR_GROUP_SCHED
1074 * effective_load() calculates the load change as seen from the root_task_group
1076 * Adding load to a group doesn't make a group heavier, but can cause movement
1077 * of group shares between cpus. Assuming the shares were perfectly aligned one
1078 * can calculate the shift in shares.
1080 * The problem is that perfectly aligning the shares is rather expensive, hence
1081 * we try to avoid doing that too often - see update_shares(), which ratelimits
1084 * We compensate this by not only taking the current delta into account, but
1085 * also considering the delta between when the shares were last adjusted and
1088 * We still saw a performance dip, some tracing learned us that between
1089 * cgroup:/ and cgroup:/foo balancing the number of affine wakeups increased
1090 * significantly. Therefore try to bias the error in direction of failing
1091 * the affine wakeup.
1094 static long effective_load(struct task_group
*tg
, int cpu
,
1097 struct sched_entity
*se
= tg
->se
[cpu
];
1103 * By not taking the decrease of shares on the other cpu into
1104 * account our error leans towards reducing the affine wakeups.
1106 if (!wl
&& sched_feat(ASYM_EFF_LOAD
))
1109 for_each_sched_entity(se
) {
1110 long S
, rw
, s
, a
, b
;
1114 * Instead of using this increment, also add the difference
1115 * between when the shares were last updated and now.
1117 more_w
= se
->my_q
->load
.weight
- se
->my_q
->rq_weight
;
1121 S
= se
->my_q
->tg
->shares
;
1122 s
= se
->my_q
->shares
;
1123 rw
= se
->my_q
->rq_weight
;
1134 * Assume the group is already running and will
1135 * thus already be accounted for in the weight.
1137 * That is, moving shares between CPUs, does not
1138 * alter the group weight.
1148 static inline unsigned long effective_load(struct task_group
*tg
, int cpu
,
1149 unsigned long wl
, unsigned long wg
)
1157 wake_affine(struct sched_domain
*this_sd
, struct rq
*this_rq
,
1158 struct task_struct
*p
, int prev_cpu
, int this_cpu
, int sync
,
1159 int idx
, unsigned long load
, unsigned long this_load
,
1160 unsigned int imbalance
)
1162 struct task_struct
*curr
= this_rq
->curr
;
1163 struct task_group
*tg
;
1164 unsigned long tl
= this_load
;
1165 unsigned long tl_per_task
;
1166 unsigned long weight
;
1169 if (!(this_sd
->flags
& SD_WAKE_AFFINE
) || !sched_feat(AFFINE_WAKEUPS
))
1172 if (sync
&& (curr
->se
.avg_overlap
> sysctl_sched_migration_cost
||
1173 p
->se
.avg_overlap
> sysctl_sched_migration_cost
))
1177 * If sync wakeup then subtract the (maximum possible)
1178 * effect of the currently running task from the load
1179 * of the current CPU:
1182 tg
= task_group(current
);
1183 weight
= current
->se
.load
.weight
;
1185 tl
+= effective_load(tg
, this_cpu
, -weight
, -weight
);
1186 load
+= effective_load(tg
, prev_cpu
, 0, -weight
);
1190 weight
= p
->se
.load
.weight
;
1192 balanced
= 100*(tl
+ effective_load(tg
, this_cpu
, weight
, weight
)) <=
1193 imbalance
*(load
+ effective_load(tg
, prev_cpu
, 0, weight
));
1196 * If the currently running task will sleep within
1197 * a reasonable amount of time then attract this newly
1200 if (sync
&& balanced
)
1203 schedstat_inc(p
, se
.nr_wakeups_affine_attempts
);
1204 tl_per_task
= cpu_avg_load_per_task(this_cpu
);
1206 if (balanced
|| (tl
<= load
&& tl
+ target_load(prev_cpu
, idx
) <=
1209 * This domain has SD_WAKE_AFFINE and
1210 * p is cache cold in this domain, and
1211 * there is no bad imbalance.
1213 schedstat_inc(this_sd
, ttwu_move_affine
);
1214 schedstat_inc(p
, se
.nr_wakeups_affine
);
1221 static int select_task_rq_fair(struct task_struct
*p
, int sync
)
1223 struct sched_domain
*sd
, *this_sd
= NULL
;
1224 int prev_cpu
, this_cpu
, new_cpu
;
1225 unsigned long load
, this_load
;
1227 unsigned int imbalance
;
1230 prev_cpu
= task_cpu(p
);
1231 this_cpu
= smp_processor_id();
1232 this_rq
= cpu_rq(this_cpu
);
1235 if (prev_cpu
== this_cpu
)
1238 * 'this_sd' is the first domain that both
1239 * this_cpu and prev_cpu are present in:
1241 for_each_domain(this_cpu
, sd
) {
1242 if (cpu_isset(prev_cpu
, sd
->span
)) {
1248 if (unlikely(!cpu_isset(this_cpu
, p
->cpus_allowed
)))
1252 * Check for affine wakeup and passive balancing possibilities.
1257 idx
= this_sd
->wake_idx
;
1259 imbalance
= 100 + (this_sd
->imbalance_pct
- 100) / 2;
1261 load
= source_load(prev_cpu
, idx
);
1262 this_load
= target_load(this_cpu
, idx
);
1264 if (wake_affine(this_sd
, this_rq
, p
, prev_cpu
, this_cpu
, sync
, idx
,
1265 load
, this_load
, imbalance
))
1269 * Start passive balancing when half the imbalance_pct
1272 if (this_sd
->flags
& SD_WAKE_BALANCE
) {
1273 if (imbalance
*this_load
<= 100*load
) {
1274 schedstat_inc(this_sd
, ttwu_move_balance
);
1275 schedstat_inc(p
, se
.nr_wakeups_passive
);
1281 return wake_idle(new_cpu
, p
);
1283 #endif /* CONFIG_SMP */
1285 static unsigned long wakeup_gran(struct sched_entity
*se
)
1287 unsigned long gran
= sysctl_sched_wakeup_granularity
;
1290 * More easily preempt - nice tasks, while not making it harder for
1293 if (!sched_feat(ASYM_GRAN
) || se
->load
.weight
> NICE_0_LOAD
)
1294 gran
= calc_delta_fair(sysctl_sched_wakeup_granularity
, se
);
1300 * Should 'se' preempt 'curr'.
1314 wakeup_preempt_entity(struct sched_entity
*curr
, struct sched_entity
*se
)
1316 s64 gran
, vdiff
= curr
->vruntime
- se
->vruntime
;
1321 gran
= wakeup_gran(curr
);
1329 * Preempt the current task with a newly woken task if needed:
1331 static void check_preempt_wakeup(struct rq
*rq
, struct task_struct
*p
, int sync
)
1333 struct task_struct
*curr
= rq
->curr
;
1334 struct cfs_rq
*cfs_rq
= task_cfs_rq(curr
);
1335 struct sched_entity
*se
= &curr
->se
, *pse
= &p
->se
;
1337 if (unlikely(rt_prio(p
->prio
))) {
1338 update_rq_clock(rq
);
1339 update_curr(cfs_rq
);
1344 if (unlikely(se
== pse
))
1347 cfs_rq_of(pse
)->next
= pse
;
1350 * We can come here with TIF_NEED_RESCHED already set from new task
1353 if (test_tsk_need_resched(curr
))
1357 * Batch tasks do not preempt (their preemption is driven by
1360 if (unlikely(p
->policy
== SCHED_BATCH
))
1363 if (!sched_feat(WAKEUP_PREEMPT
))
1366 if (sched_feat(WAKEUP_OVERLAP
) && (sync
||
1367 (se
->avg_overlap
< sysctl_sched_migration_cost
&&
1368 pse
->avg_overlap
< sysctl_sched_migration_cost
))) {
1373 find_matching_se(&se
, &pse
);
1378 if (wakeup_preempt_entity(se
, pse
) == 1) {
1383 se
= parent_entity(se
);
1384 pse
= parent_entity(pse
);
1388 static struct task_struct
*pick_next_task_fair(struct rq
*rq
)
1390 struct task_struct
*p
;
1391 struct cfs_rq
*cfs_rq
= &rq
->cfs
;
1392 struct sched_entity
*se
;
1394 if (unlikely(!cfs_rq
->nr_running
))
1398 se
= pick_next_entity(cfs_rq
);
1399 cfs_rq
= group_cfs_rq(se
);
1403 hrtick_start_fair(rq
, p
);
1409 * Account for a descheduled task:
1411 static void put_prev_task_fair(struct rq
*rq
, struct task_struct
*prev
)
1413 struct sched_entity
*se
= &prev
->se
;
1414 struct cfs_rq
*cfs_rq
;
1416 for_each_sched_entity(se
) {
1417 cfs_rq
= cfs_rq_of(se
);
1418 put_prev_entity(cfs_rq
, se
);
1423 /**************************************************
1424 * Fair scheduling class load-balancing methods:
1428 * Load-balancing iterator. Note: while the runqueue stays locked
1429 * during the whole iteration, the current task might be
1430 * dequeued so the iterator has to be dequeue-safe. Here we
1431 * achieve that by always pre-iterating before returning
1434 static struct task_struct
*
1435 __load_balance_iterator(struct cfs_rq
*cfs_rq
, struct list_head
*next
)
1437 struct task_struct
*p
= NULL
;
1438 struct sched_entity
*se
;
1440 if (next
== &cfs_rq
->tasks
)
1443 se
= list_entry(next
, struct sched_entity
, group_node
);
1445 cfs_rq
->balance_iterator
= next
->next
;
1450 static struct task_struct
*load_balance_start_fair(void *arg
)
1452 struct cfs_rq
*cfs_rq
= arg
;
1454 return __load_balance_iterator(cfs_rq
, cfs_rq
->tasks
.next
);
1457 static struct task_struct
*load_balance_next_fair(void *arg
)
1459 struct cfs_rq
*cfs_rq
= arg
;
1461 return __load_balance_iterator(cfs_rq
, cfs_rq
->balance_iterator
);
1464 static unsigned long
1465 __load_balance_fair(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1466 unsigned long max_load_move
, struct sched_domain
*sd
,
1467 enum cpu_idle_type idle
, int *all_pinned
, int *this_best_prio
,
1468 struct cfs_rq
*cfs_rq
)
1470 struct rq_iterator cfs_rq_iterator
;
1472 cfs_rq_iterator
.start
= load_balance_start_fair
;
1473 cfs_rq_iterator
.next
= load_balance_next_fair
;
1474 cfs_rq_iterator
.arg
= cfs_rq
;
1476 return balance_tasks(this_rq
, this_cpu
, busiest
,
1477 max_load_move
, sd
, idle
, all_pinned
,
1478 this_best_prio
, &cfs_rq_iterator
);
1481 #ifdef CONFIG_FAIR_GROUP_SCHED
1482 static unsigned long
1483 load_balance_fair(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1484 unsigned long max_load_move
,
1485 struct sched_domain
*sd
, enum cpu_idle_type idle
,
1486 int *all_pinned
, int *this_best_prio
)
1488 long rem_load_move
= max_load_move
;
1489 int busiest_cpu
= cpu_of(busiest
);
1490 struct task_group
*tg
;
1493 update_h_load(busiest_cpu
);
1495 list_for_each_entry_rcu(tg
, &task_groups
, list
) {
1496 struct cfs_rq
*busiest_cfs_rq
= tg
->cfs_rq
[busiest_cpu
];
1497 unsigned long busiest_h_load
= busiest_cfs_rq
->h_load
;
1498 unsigned long busiest_weight
= busiest_cfs_rq
->load
.weight
;
1499 u64 rem_load
, moved_load
;
1504 if (!busiest_cfs_rq
->task_weight
)
1507 rem_load
= (u64
)rem_load_move
* busiest_weight
;
1508 rem_load
= div_u64(rem_load
, busiest_h_load
+ 1);
1510 moved_load
= __load_balance_fair(this_rq
, this_cpu
, busiest
,
1511 rem_load
, sd
, idle
, all_pinned
, this_best_prio
,
1512 tg
->cfs_rq
[busiest_cpu
]);
1517 moved_load
*= busiest_h_load
;
1518 moved_load
= div_u64(moved_load
, busiest_weight
+ 1);
1520 rem_load_move
-= moved_load
;
1521 if (rem_load_move
< 0)
1526 return max_load_move
- rem_load_move
;
1529 static unsigned long
1530 load_balance_fair(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1531 unsigned long max_load_move
,
1532 struct sched_domain
*sd
, enum cpu_idle_type idle
,
1533 int *all_pinned
, int *this_best_prio
)
1535 return __load_balance_fair(this_rq
, this_cpu
, busiest
,
1536 max_load_move
, sd
, idle
, all_pinned
,
1537 this_best_prio
, &busiest
->cfs
);
1542 move_one_task_fair(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1543 struct sched_domain
*sd
, enum cpu_idle_type idle
)
1545 struct cfs_rq
*busy_cfs_rq
;
1546 struct rq_iterator cfs_rq_iterator
;
1548 cfs_rq_iterator
.start
= load_balance_start_fair
;
1549 cfs_rq_iterator
.next
= load_balance_next_fair
;
1551 for_each_leaf_cfs_rq(busiest
, busy_cfs_rq
) {
1553 * pass busy_cfs_rq argument into
1554 * load_balance_[start|next]_fair iterators
1556 cfs_rq_iterator
.arg
= busy_cfs_rq
;
1557 if (iter_move_one_task(this_rq
, this_cpu
, busiest
, sd
, idle
,
1564 #endif /* CONFIG_SMP */
1567 * scheduler tick hitting a task of our scheduling class:
1569 static void task_tick_fair(struct rq
*rq
, struct task_struct
*curr
, int queued
)
1571 struct cfs_rq
*cfs_rq
;
1572 struct sched_entity
*se
= &curr
->se
;
1574 for_each_sched_entity(se
) {
1575 cfs_rq
= cfs_rq_of(se
);
1576 entity_tick(cfs_rq
, se
, queued
);
1580 #define swap(a, b) do { typeof(a) tmp = (a); (a) = (b); (b) = tmp; } while (0)
1583 * Share the fairness runtime between parent and child, thus the
1584 * total amount of pressure for CPU stays equal - new tasks
1585 * get a chance to run but frequent forkers are not allowed to
1586 * monopolize the CPU. Note: the parent runqueue is locked,
1587 * the child is not running yet.
1589 static void task_new_fair(struct rq
*rq
, struct task_struct
*p
)
1591 struct cfs_rq
*cfs_rq
= task_cfs_rq(p
);
1592 struct sched_entity
*se
= &p
->se
, *curr
= cfs_rq
->curr
;
1593 int this_cpu
= smp_processor_id();
1595 sched_info_queued(p
);
1597 update_curr(cfs_rq
);
1598 place_entity(cfs_rq
, se
, 1);
1600 /* 'curr' will be NULL if the child belongs to a different group */
1601 if (sysctl_sched_child_runs_first
&& this_cpu
== task_cpu(p
) &&
1602 curr
&& curr
->vruntime
< se
->vruntime
) {
1604 * Upon rescheduling, sched_class::put_prev_task() will place
1605 * 'current' within the tree based on its new key value.
1607 swap(curr
->vruntime
, se
->vruntime
);
1608 resched_task(rq
->curr
);
1611 enqueue_task_fair(rq
, p
, 0);
1615 * Priority of the task has changed. Check to see if we preempt
1618 static void prio_changed_fair(struct rq
*rq
, struct task_struct
*p
,
1619 int oldprio
, int running
)
1622 * Reschedule if we are currently running on this runqueue and
1623 * our priority decreased, or if we are not currently running on
1624 * this runqueue and our priority is higher than the current's
1627 if (p
->prio
> oldprio
)
1628 resched_task(rq
->curr
);
1630 check_preempt_curr(rq
, p
, 0);
1634 * We switched to the sched_fair class.
1636 static void switched_to_fair(struct rq
*rq
, struct task_struct
*p
,
1640 * We were most likely switched from sched_rt, so
1641 * kick off the schedule if running, otherwise just see
1642 * if we can still preempt the current task.
1645 resched_task(rq
->curr
);
1647 check_preempt_curr(rq
, p
, 0);
1650 /* Account for a task changing its policy or group.
1652 * This routine is mostly called to set cfs_rq->curr field when a task
1653 * migrates between groups/classes.
1655 static void set_curr_task_fair(struct rq
*rq
)
1657 struct sched_entity
*se
= &rq
->curr
->se
;
1659 for_each_sched_entity(se
)
1660 set_next_entity(cfs_rq_of(se
), se
);
1663 #ifdef CONFIG_FAIR_GROUP_SCHED
1664 static void moved_group_fair(struct task_struct
*p
)
1666 struct cfs_rq
*cfs_rq
= task_cfs_rq(p
);
1668 update_curr(cfs_rq
);
1669 place_entity(cfs_rq
, &p
->se
, 1);
1674 * All the scheduling class methods:
1676 static const struct sched_class fair_sched_class
= {
1677 .next
= &idle_sched_class
,
1678 .enqueue_task
= enqueue_task_fair
,
1679 .dequeue_task
= dequeue_task_fair
,
1680 .yield_task
= yield_task_fair
,
1682 .check_preempt_curr
= check_preempt_wakeup
,
1684 .pick_next_task
= pick_next_task_fair
,
1685 .put_prev_task
= put_prev_task_fair
,
1688 .select_task_rq
= select_task_rq_fair
,
1690 .load_balance
= load_balance_fair
,
1691 .move_one_task
= move_one_task_fair
,
1694 .set_curr_task
= set_curr_task_fair
,
1695 .task_tick
= task_tick_fair
,
1696 .task_new
= task_new_fair
,
1698 .prio_changed
= prio_changed_fair
,
1699 .switched_to
= switched_to_fair
,
1701 #ifdef CONFIG_FAIR_GROUP_SCHED
1702 .moved_group
= moved_group_fair
,
1706 #ifdef CONFIG_SCHED_DEBUG
1707 static void print_cfs_stats(struct seq_file
*m
, int cpu
)
1709 struct cfs_rq
*cfs_rq
;
1712 for_each_leaf_cfs_rq(cpu_rq(cpu
), cfs_rq
)
1713 print_cfs_rq(m
, cpu
, cfs_rq
);